]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame_incremental - kernel/cgroup.c
percpu-refcount: require percpu_ref to be exited explicitly
[mirror_ubuntu-bionic-kernel.git] / kernel / cgroup.c
... / ...
CommitLineData
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/string.h>
50#include <linux/sort.h>
51#include <linux/kmod.h>
52#include <linux/delayacct.h>
53#include <linux/cgroupstats.h>
54#include <linux/hashtable.h>
55#include <linux/pid_namespace.h>
56#include <linux/idr.h>
57#include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58#include <linux/kthread.h>
59#include <linux/delay.h>
60
61#include <linux/atomic.h>
62
63/*
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
68 */
69#define CGROUP_PIDLIST_DESTROY_DELAY HZ
70
71#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
72 MAX_CFTYPE_NAME + 2)
73
74/*
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
77 *
78 * css_set_rwsem protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
80 *
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
83 */
84#ifdef CONFIG_PROVE_RCU
85DEFINE_MUTEX(cgroup_mutex);
86DECLARE_RWSEM(css_set_rwsem);
87EXPORT_SYMBOL_GPL(cgroup_mutex);
88EXPORT_SYMBOL_GPL(css_set_rwsem);
89#else
90static DEFINE_MUTEX(cgroup_mutex);
91static DECLARE_RWSEM(css_set_rwsem);
92#endif
93
94/*
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
97 */
98static DEFINE_SPINLOCK(cgroup_idr_lock);
99
100/*
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
103 */
104static DEFINE_SPINLOCK(release_agent_path_lock);
105
106#define cgroup_assert_mutex_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
110
111/*
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
116 */
117static struct workqueue_struct *cgroup_destroy_wq;
118
119/*
120 * pidlist destructions need to be flushed on cgroup destruction. Use a
121 * separate workqueue as flush domain.
122 */
123static struct workqueue_struct *cgroup_pidlist_destroy_wq;
124
125/* generate an array of cgroup subsystem pointers */
126#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127static struct cgroup_subsys *cgroup_subsys[] = {
128#include <linux/cgroup_subsys.h>
129};
130#undef SUBSYS
131
132/* array of cgroup subsystem names */
133#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134static const char *cgroup_subsys_name[] = {
135#include <linux/cgroup_subsys.h>
136};
137#undef SUBSYS
138
139/*
140 * The default hierarchy, reserved for the subsystems that are otherwise
141 * unattached - it never has more than a single cgroup, and all tasks are
142 * part of that cgroup.
143 */
144struct cgroup_root cgrp_dfl_root;
145
146/*
147 * The default hierarchy always exists but is hidden until mounted for the
148 * first time. This is for backward compatibility.
149 */
150static bool cgrp_dfl_root_visible;
151
152/* some controllers are not supported in the default hierarchy */
153static const unsigned int cgrp_dfl_root_inhibit_ss_mask = 0
154#ifdef CONFIG_CGROUP_DEBUG
155 | (1 << debug_cgrp_id)
156#endif
157 ;
158
159/* The list of hierarchy roots */
160
161static LIST_HEAD(cgroup_roots);
162static int cgroup_root_count;
163
164/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
165static DEFINE_IDR(cgroup_hierarchy_idr);
166
167/*
168 * Assign a monotonically increasing serial number to csses. It guarantees
169 * cgroups with bigger numbers are newer than those with smaller numbers.
170 * Also, as csses are always appended to the parent's ->children list, it
171 * guarantees that sibling csses are always sorted in the ascending serial
172 * number order on the list. Protected by cgroup_mutex.
173 */
174static u64 css_serial_nr_next = 1;
175
176/* This flag indicates whether tasks in the fork and exit paths should
177 * check for fork/exit handlers to call. This avoids us having to do
178 * extra work in the fork/exit path if none of the subsystems need to
179 * be called.
180 */
181static int need_forkexit_callback __read_mostly;
182
183static struct cftype cgroup_base_files[];
184
185static void cgroup_put(struct cgroup *cgrp);
186static int rebind_subsystems(struct cgroup_root *dst_root,
187 unsigned int ss_mask);
188static int cgroup_destroy_locked(struct cgroup *cgrp);
189static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss);
190static void css_release(struct percpu_ref *ref);
191static void kill_css(struct cgroup_subsys_state *css);
192static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
193 bool is_add);
194static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
195
196/* IDR wrappers which synchronize using cgroup_idr_lock */
197static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
198 gfp_t gfp_mask)
199{
200 int ret;
201
202 idr_preload(gfp_mask);
203 spin_lock_bh(&cgroup_idr_lock);
204 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
205 spin_unlock_bh(&cgroup_idr_lock);
206 idr_preload_end();
207 return ret;
208}
209
210static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
211{
212 void *ret;
213
214 spin_lock_bh(&cgroup_idr_lock);
215 ret = idr_replace(idr, ptr, id);
216 spin_unlock_bh(&cgroup_idr_lock);
217 return ret;
218}
219
220static void cgroup_idr_remove(struct idr *idr, int id)
221{
222 spin_lock_bh(&cgroup_idr_lock);
223 idr_remove(idr, id);
224 spin_unlock_bh(&cgroup_idr_lock);
225}
226
227static struct cgroup *cgroup_parent(struct cgroup *cgrp)
228{
229 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
230
231 if (parent_css)
232 return container_of(parent_css, struct cgroup, self);
233 return NULL;
234}
235
236/**
237 * cgroup_css - obtain a cgroup's css for the specified subsystem
238 * @cgrp: the cgroup of interest
239 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
240 *
241 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
242 * function must be called either under cgroup_mutex or rcu_read_lock() and
243 * the caller is responsible for pinning the returned css if it wants to
244 * keep accessing it outside the said locks. This function may return
245 * %NULL if @cgrp doesn't have @subsys_id enabled.
246 */
247static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
248 struct cgroup_subsys *ss)
249{
250 if (ss)
251 return rcu_dereference_check(cgrp->subsys[ss->id],
252 lockdep_is_held(&cgroup_mutex));
253 else
254 return &cgrp->self;
255}
256
257/**
258 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
259 * @cgrp: the cgroup of interest
260 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
261 *
262 * Similar to cgroup_css() but returns the effctive css, which is defined
263 * as the matching css of the nearest ancestor including self which has @ss
264 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
265 * function is guaranteed to return non-NULL css.
266 */
267static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
268 struct cgroup_subsys *ss)
269{
270 lockdep_assert_held(&cgroup_mutex);
271
272 if (!ss)
273 return &cgrp->self;
274
275 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
276 return NULL;
277
278 while (cgroup_parent(cgrp) &&
279 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
280 cgrp = cgroup_parent(cgrp);
281
282 return cgroup_css(cgrp, ss);
283}
284
285/* convenient tests for these bits */
286static inline bool cgroup_is_dead(const struct cgroup *cgrp)
287{
288 return !(cgrp->self.flags & CSS_ONLINE);
289}
290
291struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
292{
293 struct cgroup *cgrp = of->kn->parent->priv;
294 struct cftype *cft = of_cft(of);
295
296 /*
297 * This is open and unprotected implementation of cgroup_css().
298 * seq_css() is only called from a kernfs file operation which has
299 * an active reference on the file. Because all the subsystem
300 * files are drained before a css is disassociated with a cgroup,
301 * the matching css from the cgroup's subsys table is guaranteed to
302 * be and stay valid until the enclosing operation is complete.
303 */
304 if (cft->ss)
305 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
306 else
307 return &cgrp->self;
308}
309EXPORT_SYMBOL_GPL(of_css);
310
311/**
312 * cgroup_is_descendant - test ancestry
313 * @cgrp: the cgroup to be tested
314 * @ancestor: possible ancestor of @cgrp
315 *
316 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
317 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
318 * and @ancestor are accessible.
319 */
320bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
321{
322 while (cgrp) {
323 if (cgrp == ancestor)
324 return true;
325 cgrp = cgroup_parent(cgrp);
326 }
327 return false;
328}
329
330static int cgroup_is_releasable(const struct cgroup *cgrp)
331{
332 const int bits =
333 (1 << CGRP_RELEASABLE) |
334 (1 << CGRP_NOTIFY_ON_RELEASE);
335 return (cgrp->flags & bits) == bits;
336}
337
338static int notify_on_release(const struct cgroup *cgrp)
339{
340 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
341}
342
343/**
344 * for_each_css - iterate all css's of a cgroup
345 * @css: the iteration cursor
346 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
347 * @cgrp: the target cgroup to iterate css's of
348 *
349 * Should be called under cgroup_[tree_]mutex.
350 */
351#define for_each_css(css, ssid, cgrp) \
352 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
353 if (!((css) = rcu_dereference_check( \
354 (cgrp)->subsys[(ssid)], \
355 lockdep_is_held(&cgroup_mutex)))) { } \
356 else
357
358/**
359 * for_each_e_css - iterate all effective css's of a cgroup
360 * @css: the iteration cursor
361 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
362 * @cgrp: the target cgroup to iterate css's of
363 *
364 * Should be called under cgroup_[tree_]mutex.
365 */
366#define for_each_e_css(css, ssid, cgrp) \
367 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
368 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
369 ; \
370 else
371
372/**
373 * for_each_subsys - iterate all enabled cgroup subsystems
374 * @ss: the iteration cursor
375 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
376 */
377#define for_each_subsys(ss, ssid) \
378 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
379 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
380
381/* iterate across the hierarchies */
382#define for_each_root(root) \
383 list_for_each_entry((root), &cgroup_roots, root_list)
384
385/* iterate over child cgrps, lock should be held throughout iteration */
386#define cgroup_for_each_live_child(child, cgrp) \
387 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
388 if (({ lockdep_assert_held(&cgroup_mutex); \
389 cgroup_is_dead(child); })) \
390 ; \
391 else
392
393/* the list of cgroups eligible for automatic release. Protected by
394 * release_list_lock */
395static LIST_HEAD(release_list);
396static DEFINE_RAW_SPINLOCK(release_list_lock);
397static void cgroup_release_agent(struct work_struct *work);
398static DECLARE_WORK(release_agent_work, cgroup_release_agent);
399static void check_for_release(struct cgroup *cgrp);
400
401/*
402 * A cgroup can be associated with multiple css_sets as different tasks may
403 * belong to different cgroups on different hierarchies. In the other
404 * direction, a css_set is naturally associated with multiple cgroups.
405 * This M:N relationship is represented by the following link structure
406 * which exists for each association and allows traversing the associations
407 * from both sides.
408 */
409struct cgrp_cset_link {
410 /* the cgroup and css_set this link associates */
411 struct cgroup *cgrp;
412 struct css_set *cset;
413
414 /* list of cgrp_cset_links anchored at cgrp->cset_links */
415 struct list_head cset_link;
416
417 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
418 struct list_head cgrp_link;
419};
420
421/*
422 * The default css_set - used by init and its children prior to any
423 * hierarchies being mounted. It contains a pointer to the root state
424 * for each subsystem. Also used to anchor the list of css_sets. Not
425 * reference-counted, to improve performance when child cgroups
426 * haven't been created.
427 */
428struct css_set init_css_set = {
429 .refcount = ATOMIC_INIT(1),
430 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
431 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
432 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
433 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
434 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
435};
436
437static int css_set_count = 1; /* 1 for init_css_set */
438
439/**
440 * cgroup_update_populated - updated populated count of a cgroup
441 * @cgrp: the target cgroup
442 * @populated: inc or dec populated count
443 *
444 * @cgrp is either getting the first task (css_set) or losing the last.
445 * Update @cgrp->populated_cnt accordingly. The count is propagated
446 * towards root so that a given cgroup's populated_cnt is zero iff the
447 * cgroup and all its descendants are empty.
448 *
449 * @cgrp's interface file "cgroup.populated" is zero if
450 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
451 * changes from or to zero, userland is notified that the content of the
452 * interface file has changed. This can be used to detect when @cgrp and
453 * its descendants become populated or empty.
454 */
455static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
456{
457 lockdep_assert_held(&css_set_rwsem);
458
459 do {
460 bool trigger;
461
462 if (populated)
463 trigger = !cgrp->populated_cnt++;
464 else
465 trigger = !--cgrp->populated_cnt;
466
467 if (!trigger)
468 break;
469
470 if (cgrp->populated_kn)
471 kernfs_notify(cgrp->populated_kn);
472 cgrp = cgroup_parent(cgrp);
473 } while (cgrp);
474}
475
476/*
477 * hash table for cgroup groups. This improves the performance to find
478 * an existing css_set. This hash doesn't (currently) take into
479 * account cgroups in empty hierarchies.
480 */
481#define CSS_SET_HASH_BITS 7
482static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
483
484static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
485{
486 unsigned long key = 0UL;
487 struct cgroup_subsys *ss;
488 int i;
489
490 for_each_subsys(ss, i)
491 key += (unsigned long)css[i];
492 key = (key >> 16) ^ key;
493
494 return key;
495}
496
497static void put_css_set_locked(struct css_set *cset, bool taskexit)
498{
499 struct cgrp_cset_link *link, *tmp_link;
500 struct cgroup_subsys *ss;
501 int ssid;
502
503 lockdep_assert_held(&css_set_rwsem);
504
505 if (!atomic_dec_and_test(&cset->refcount))
506 return;
507
508 /* This css_set is dead. unlink it and release cgroup refcounts */
509 for_each_subsys(ss, ssid)
510 list_del(&cset->e_cset_node[ssid]);
511 hash_del(&cset->hlist);
512 css_set_count--;
513
514 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
515 struct cgroup *cgrp = link->cgrp;
516
517 list_del(&link->cset_link);
518 list_del(&link->cgrp_link);
519
520 /* @cgrp can't go away while we're holding css_set_rwsem */
521 if (list_empty(&cgrp->cset_links)) {
522 cgroup_update_populated(cgrp, false);
523 if (notify_on_release(cgrp)) {
524 if (taskexit)
525 set_bit(CGRP_RELEASABLE, &cgrp->flags);
526 check_for_release(cgrp);
527 }
528 }
529
530 kfree(link);
531 }
532
533 kfree_rcu(cset, rcu_head);
534}
535
536static void put_css_set(struct css_set *cset, bool taskexit)
537{
538 /*
539 * Ensure that the refcount doesn't hit zero while any readers
540 * can see it. Similar to atomic_dec_and_lock(), but for an
541 * rwlock
542 */
543 if (atomic_add_unless(&cset->refcount, -1, 1))
544 return;
545
546 down_write(&css_set_rwsem);
547 put_css_set_locked(cset, taskexit);
548 up_write(&css_set_rwsem);
549}
550
551/*
552 * refcounted get/put for css_set objects
553 */
554static inline void get_css_set(struct css_set *cset)
555{
556 atomic_inc(&cset->refcount);
557}
558
559/**
560 * compare_css_sets - helper function for find_existing_css_set().
561 * @cset: candidate css_set being tested
562 * @old_cset: existing css_set for a task
563 * @new_cgrp: cgroup that's being entered by the task
564 * @template: desired set of css pointers in css_set (pre-calculated)
565 *
566 * Returns true if "cset" matches "old_cset" except for the hierarchy
567 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
568 */
569static bool compare_css_sets(struct css_set *cset,
570 struct css_set *old_cset,
571 struct cgroup *new_cgrp,
572 struct cgroup_subsys_state *template[])
573{
574 struct list_head *l1, *l2;
575
576 /*
577 * On the default hierarchy, there can be csets which are
578 * associated with the same set of cgroups but different csses.
579 * Let's first ensure that csses match.
580 */
581 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
582 return false;
583
584 /*
585 * Compare cgroup pointers in order to distinguish between
586 * different cgroups in hierarchies. As different cgroups may
587 * share the same effective css, this comparison is always
588 * necessary.
589 */
590 l1 = &cset->cgrp_links;
591 l2 = &old_cset->cgrp_links;
592 while (1) {
593 struct cgrp_cset_link *link1, *link2;
594 struct cgroup *cgrp1, *cgrp2;
595
596 l1 = l1->next;
597 l2 = l2->next;
598 /* See if we reached the end - both lists are equal length. */
599 if (l1 == &cset->cgrp_links) {
600 BUG_ON(l2 != &old_cset->cgrp_links);
601 break;
602 } else {
603 BUG_ON(l2 == &old_cset->cgrp_links);
604 }
605 /* Locate the cgroups associated with these links. */
606 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
607 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
608 cgrp1 = link1->cgrp;
609 cgrp2 = link2->cgrp;
610 /* Hierarchies should be linked in the same order. */
611 BUG_ON(cgrp1->root != cgrp2->root);
612
613 /*
614 * If this hierarchy is the hierarchy of the cgroup
615 * that's changing, then we need to check that this
616 * css_set points to the new cgroup; if it's any other
617 * hierarchy, then this css_set should point to the
618 * same cgroup as the old css_set.
619 */
620 if (cgrp1->root == new_cgrp->root) {
621 if (cgrp1 != new_cgrp)
622 return false;
623 } else {
624 if (cgrp1 != cgrp2)
625 return false;
626 }
627 }
628 return true;
629}
630
631/**
632 * find_existing_css_set - init css array and find the matching css_set
633 * @old_cset: the css_set that we're using before the cgroup transition
634 * @cgrp: the cgroup that we're moving into
635 * @template: out param for the new set of csses, should be clear on entry
636 */
637static struct css_set *find_existing_css_set(struct css_set *old_cset,
638 struct cgroup *cgrp,
639 struct cgroup_subsys_state *template[])
640{
641 struct cgroup_root *root = cgrp->root;
642 struct cgroup_subsys *ss;
643 struct css_set *cset;
644 unsigned long key;
645 int i;
646
647 /*
648 * Build the set of subsystem state objects that we want to see in the
649 * new css_set. while subsystems can change globally, the entries here
650 * won't change, so no need for locking.
651 */
652 for_each_subsys(ss, i) {
653 if (root->subsys_mask & (1UL << i)) {
654 /*
655 * @ss is in this hierarchy, so we want the
656 * effective css from @cgrp.
657 */
658 template[i] = cgroup_e_css(cgrp, ss);
659 } else {
660 /*
661 * @ss is not in this hierarchy, so we don't want
662 * to change the css.
663 */
664 template[i] = old_cset->subsys[i];
665 }
666 }
667
668 key = css_set_hash(template);
669 hash_for_each_possible(css_set_table, cset, hlist, key) {
670 if (!compare_css_sets(cset, old_cset, cgrp, template))
671 continue;
672
673 /* This css_set matches what we need */
674 return cset;
675 }
676
677 /* No existing cgroup group matched */
678 return NULL;
679}
680
681static void free_cgrp_cset_links(struct list_head *links_to_free)
682{
683 struct cgrp_cset_link *link, *tmp_link;
684
685 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
686 list_del(&link->cset_link);
687 kfree(link);
688 }
689}
690
691/**
692 * allocate_cgrp_cset_links - allocate cgrp_cset_links
693 * @count: the number of links to allocate
694 * @tmp_links: list_head the allocated links are put on
695 *
696 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
697 * through ->cset_link. Returns 0 on success or -errno.
698 */
699static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
700{
701 struct cgrp_cset_link *link;
702 int i;
703
704 INIT_LIST_HEAD(tmp_links);
705
706 for (i = 0; i < count; i++) {
707 link = kzalloc(sizeof(*link), GFP_KERNEL);
708 if (!link) {
709 free_cgrp_cset_links(tmp_links);
710 return -ENOMEM;
711 }
712 list_add(&link->cset_link, tmp_links);
713 }
714 return 0;
715}
716
717/**
718 * link_css_set - a helper function to link a css_set to a cgroup
719 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
720 * @cset: the css_set to be linked
721 * @cgrp: the destination cgroup
722 */
723static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
724 struct cgroup *cgrp)
725{
726 struct cgrp_cset_link *link;
727
728 BUG_ON(list_empty(tmp_links));
729
730 if (cgroup_on_dfl(cgrp))
731 cset->dfl_cgrp = cgrp;
732
733 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
734 link->cset = cset;
735 link->cgrp = cgrp;
736
737 if (list_empty(&cgrp->cset_links))
738 cgroup_update_populated(cgrp, true);
739 list_move(&link->cset_link, &cgrp->cset_links);
740
741 /*
742 * Always add links to the tail of the list so that the list
743 * is sorted by order of hierarchy creation
744 */
745 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
746}
747
748/**
749 * find_css_set - return a new css_set with one cgroup updated
750 * @old_cset: the baseline css_set
751 * @cgrp: the cgroup to be updated
752 *
753 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
754 * substituted into the appropriate hierarchy.
755 */
756static struct css_set *find_css_set(struct css_set *old_cset,
757 struct cgroup *cgrp)
758{
759 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
760 struct css_set *cset;
761 struct list_head tmp_links;
762 struct cgrp_cset_link *link;
763 struct cgroup_subsys *ss;
764 unsigned long key;
765 int ssid;
766
767 lockdep_assert_held(&cgroup_mutex);
768
769 /* First see if we already have a cgroup group that matches
770 * the desired set */
771 down_read(&css_set_rwsem);
772 cset = find_existing_css_set(old_cset, cgrp, template);
773 if (cset)
774 get_css_set(cset);
775 up_read(&css_set_rwsem);
776
777 if (cset)
778 return cset;
779
780 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
781 if (!cset)
782 return NULL;
783
784 /* Allocate all the cgrp_cset_link objects that we'll need */
785 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
786 kfree(cset);
787 return NULL;
788 }
789
790 atomic_set(&cset->refcount, 1);
791 INIT_LIST_HEAD(&cset->cgrp_links);
792 INIT_LIST_HEAD(&cset->tasks);
793 INIT_LIST_HEAD(&cset->mg_tasks);
794 INIT_LIST_HEAD(&cset->mg_preload_node);
795 INIT_LIST_HEAD(&cset->mg_node);
796 INIT_HLIST_NODE(&cset->hlist);
797
798 /* Copy the set of subsystem state objects generated in
799 * find_existing_css_set() */
800 memcpy(cset->subsys, template, sizeof(cset->subsys));
801
802 down_write(&css_set_rwsem);
803 /* Add reference counts and links from the new css_set. */
804 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
805 struct cgroup *c = link->cgrp;
806
807 if (c->root == cgrp->root)
808 c = cgrp;
809 link_css_set(&tmp_links, cset, c);
810 }
811
812 BUG_ON(!list_empty(&tmp_links));
813
814 css_set_count++;
815
816 /* Add @cset to the hash table */
817 key = css_set_hash(cset->subsys);
818 hash_add(css_set_table, &cset->hlist, key);
819
820 for_each_subsys(ss, ssid)
821 list_add_tail(&cset->e_cset_node[ssid],
822 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
823
824 up_write(&css_set_rwsem);
825
826 return cset;
827}
828
829static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
830{
831 struct cgroup *root_cgrp = kf_root->kn->priv;
832
833 return root_cgrp->root;
834}
835
836static int cgroup_init_root_id(struct cgroup_root *root)
837{
838 int id;
839
840 lockdep_assert_held(&cgroup_mutex);
841
842 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
843 if (id < 0)
844 return id;
845
846 root->hierarchy_id = id;
847 return 0;
848}
849
850static void cgroup_exit_root_id(struct cgroup_root *root)
851{
852 lockdep_assert_held(&cgroup_mutex);
853
854 if (root->hierarchy_id) {
855 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
856 root->hierarchy_id = 0;
857 }
858}
859
860static void cgroup_free_root(struct cgroup_root *root)
861{
862 if (root) {
863 /* hierarhcy ID shoulid already have been released */
864 WARN_ON_ONCE(root->hierarchy_id);
865
866 idr_destroy(&root->cgroup_idr);
867 kfree(root);
868 }
869}
870
871static void cgroup_destroy_root(struct cgroup_root *root)
872{
873 struct cgroup *cgrp = &root->cgrp;
874 struct cgrp_cset_link *link, *tmp_link;
875
876 mutex_lock(&cgroup_mutex);
877
878 BUG_ON(atomic_read(&root->nr_cgrps));
879 BUG_ON(!list_empty(&cgrp->self.children));
880
881 /* Rebind all subsystems back to the default hierarchy */
882 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
883
884 /*
885 * Release all the links from cset_links to this hierarchy's
886 * root cgroup
887 */
888 down_write(&css_set_rwsem);
889
890 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
891 list_del(&link->cset_link);
892 list_del(&link->cgrp_link);
893 kfree(link);
894 }
895 up_write(&css_set_rwsem);
896
897 if (!list_empty(&root->root_list)) {
898 list_del(&root->root_list);
899 cgroup_root_count--;
900 }
901
902 cgroup_exit_root_id(root);
903
904 mutex_unlock(&cgroup_mutex);
905
906 kernfs_destroy_root(root->kf_root);
907 cgroup_free_root(root);
908}
909
910/* look up cgroup associated with given css_set on the specified hierarchy */
911static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
912 struct cgroup_root *root)
913{
914 struct cgroup *res = NULL;
915
916 lockdep_assert_held(&cgroup_mutex);
917 lockdep_assert_held(&css_set_rwsem);
918
919 if (cset == &init_css_set) {
920 res = &root->cgrp;
921 } else {
922 struct cgrp_cset_link *link;
923
924 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
925 struct cgroup *c = link->cgrp;
926
927 if (c->root == root) {
928 res = c;
929 break;
930 }
931 }
932 }
933
934 BUG_ON(!res);
935 return res;
936}
937
938/*
939 * Return the cgroup for "task" from the given hierarchy. Must be
940 * called with cgroup_mutex and css_set_rwsem held.
941 */
942static struct cgroup *task_cgroup_from_root(struct task_struct *task,
943 struct cgroup_root *root)
944{
945 /*
946 * No need to lock the task - since we hold cgroup_mutex the
947 * task can't change groups, so the only thing that can happen
948 * is that it exits and its css is set back to init_css_set.
949 */
950 return cset_cgroup_from_root(task_css_set(task), root);
951}
952
953/*
954 * A task must hold cgroup_mutex to modify cgroups.
955 *
956 * Any task can increment and decrement the count field without lock.
957 * So in general, code holding cgroup_mutex can't rely on the count
958 * field not changing. However, if the count goes to zero, then only
959 * cgroup_attach_task() can increment it again. Because a count of zero
960 * means that no tasks are currently attached, therefore there is no
961 * way a task attached to that cgroup can fork (the other way to
962 * increment the count). So code holding cgroup_mutex can safely
963 * assume that if the count is zero, it will stay zero. Similarly, if
964 * a task holds cgroup_mutex on a cgroup with zero count, it
965 * knows that the cgroup won't be removed, as cgroup_rmdir()
966 * needs that mutex.
967 *
968 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
969 * (usually) take cgroup_mutex. These are the two most performance
970 * critical pieces of code here. The exception occurs on cgroup_exit(),
971 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
972 * is taken, and if the cgroup count is zero, a usermode call made
973 * to the release agent with the name of the cgroup (path relative to
974 * the root of cgroup file system) as the argument.
975 *
976 * A cgroup can only be deleted if both its 'count' of using tasks
977 * is zero, and its list of 'children' cgroups is empty. Since all
978 * tasks in the system use _some_ cgroup, and since there is always at
979 * least one task in the system (init, pid == 1), therefore, root cgroup
980 * always has either children cgroups and/or using tasks. So we don't
981 * need a special hack to ensure that root cgroup cannot be deleted.
982 *
983 * P.S. One more locking exception. RCU is used to guard the
984 * update of a tasks cgroup pointer by cgroup_attach_task()
985 */
986
987static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
988static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
989static const struct file_operations proc_cgroupstats_operations;
990
991static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
992 char *buf)
993{
994 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
995 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
996 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
997 cft->ss->name, cft->name);
998 else
999 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1000 return buf;
1001}
1002
1003/**
1004 * cgroup_file_mode - deduce file mode of a control file
1005 * @cft: the control file in question
1006 *
1007 * returns cft->mode if ->mode is not 0
1008 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1009 * returns S_IRUGO if it has only a read handler
1010 * returns S_IWUSR if it has only a write hander
1011 */
1012static umode_t cgroup_file_mode(const struct cftype *cft)
1013{
1014 umode_t mode = 0;
1015
1016 if (cft->mode)
1017 return cft->mode;
1018
1019 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1020 mode |= S_IRUGO;
1021
1022 if (cft->write_u64 || cft->write_s64 || cft->write)
1023 mode |= S_IWUSR;
1024
1025 return mode;
1026}
1027
1028static void cgroup_get(struct cgroup *cgrp)
1029{
1030 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1031 css_get(&cgrp->self);
1032}
1033
1034static void cgroup_put(struct cgroup *cgrp)
1035{
1036 css_put(&cgrp->self);
1037}
1038
1039/**
1040 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1041 * @kn: the kernfs_node being serviced
1042 *
1043 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1044 * the method finishes if locking succeeded. Note that once this function
1045 * returns the cgroup returned by cgroup_kn_lock_live() may become
1046 * inaccessible any time. If the caller intends to continue to access the
1047 * cgroup, it should pin it before invoking this function.
1048 */
1049static void cgroup_kn_unlock(struct kernfs_node *kn)
1050{
1051 struct cgroup *cgrp;
1052
1053 if (kernfs_type(kn) == KERNFS_DIR)
1054 cgrp = kn->priv;
1055 else
1056 cgrp = kn->parent->priv;
1057
1058 mutex_unlock(&cgroup_mutex);
1059
1060 kernfs_unbreak_active_protection(kn);
1061 cgroup_put(cgrp);
1062}
1063
1064/**
1065 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1066 * @kn: the kernfs_node being serviced
1067 *
1068 * This helper is to be used by a cgroup kernfs method currently servicing
1069 * @kn. It breaks the active protection, performs cgroup locking and
1070 * verifies that the associated cgroup is alive. Returns the cgroup if
1071 * alive; otherwise, %NULL. A successful return should be undone by a
1072 * matching cgroup_kn_unlock() invocation.
1073 *
1074 * Any cgroup kernfs method implementation which requires locking the
1075 * associated cgroup should use this helper. It avoids nesting cgroup
1076 * locking under kernfs active protection and allows all kernfs operations
1077 * including self-removal.
1078 */
1079static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1080{
1081 struct cgroup *cgrp;
1082
1083 if (kernfs_type(kn) == KERNFS_DIR)
1084 cgrp = kn->priv;
1085 else
1086 cgrp = kn->parent->priv;
1087
1088 /*
1089 * We're gonna grab cgroup_mutex which nests outside kernfs
1090 * active_ref. cgroup liveliness check alone provides enough
1091 * protection against removal. Ensure @cgrp stays accessible and
1092 * break the active_ref protection.
1093 */
1094 cgroup_get(cgrp);
1095 kernfs_break_active_protection(kn);
1096
1097 mutex_lock(&cgroup_mutex);
1098
1099 if (!cgroup_is_dead(cgrp))
1100 return cgrp;
1101
1102 cgroup_kn_unlock(kn);
1103 return NULL;
1104}
1105
1106static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1107{
1108 char name[CGROUP_FILE_NAME_MAX];
1109
1110 lockdep_assert_held(&cgroup_mutex);
1111 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1112}
1113
1114/**
1115 * cgroup_clear_dir - remove subsys files in a cgroup directory
1116 * @cgrp: target cgroup
1117 * @subsys_mask: mask of the subsystem ids whose files should be removed
1118 */
1119static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1120{
1121 struct cgroup_subsys *ss;
1122 int i;
1123
1124 for_each_subsys(ss, i) {
1125 struct cftype *cfts;
1126
1127 if (!(subsys_mask & (1 << i)))
1128 continue;
1129 list_for_each_entry(cfts, &ss->cfts, node)
1130 cgroup_addrm_files(cgrp, cfts, false);
1131 }
1132}
1133
1134static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1135{
1136 struct cgroup_subsys *ss;
1137 unsigned int tmp_ss_mask;
1138 int ssid, i, ret;
1139
1140 lockdep_assert_held(&cgroup_mutex);
1141
1142 for_each_subsys(ss, ssid) {
1143 if (!(ss_mask & (1 << ssid)))
1144 continue;
1145
1146 /* if @ss has non-root csses attached to it, can't move */
1147 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1148 return -EBUSY;
1149
1150 /* can't move between two non-dummy roots either */
1151 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1152 return -EBUSY;
1153 }
1154
1155 /* skip creating root files on dfl_root for inhibited subsystems */
1156 tmp_ss_mask = ss_mask;
1157 if (dst_root == &cgrp_dfl_root)
1158 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1159
1160 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1161 if (ret) {
1162 if (dst_root != &cgrp_dfl_root)
1163 return ret;
1164
1165 /*
1166 * Rebinding back to the default root is not allowed to
1167 * fail. Using both default and non-default roots should
1168 * be rare. Moving subsystems back and forth even more so.
1169 * Just warn about it and continue.
1170 */
1171 if (cgrp_dfl_root_visible) {
1172 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1173 ret, ss_mask);
1174 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1175 }
1176 }
1177
1178 /*
1179 * Nothing can fail from this point on. Remove files for the
1180 * removed subsystems and rebind each subsystem.
1181 */
1182 for_each_subsys(ss, ssid)
1183 if (ss_mask & (1 << ssid))
1184 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1185
1186 for_each_subsys(ss, ssid) {
1187 struct cgroup_root *src_root;
1188 struct cgroup_subsys_state *css;
1189 struct css_set *cset;
1190
1191 if (!(ss_mask & (1 << ssid)))
1192 continue;
1193
1194 src_root = ss->root;
1195 css = cgroup_css(&src_root->cgrp, ss);
1196
1197 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1198
1199 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1200 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1201 ss->root = dst_root;
1202 css->cgroup = &dst_root->cgrp;
1203
1204 down_write(&css_set_rwsem);
1205 hash_for_each(css_set_table, i, cset, hlist)
1206 list_move_tail(&cset->e_cset_node[ss->id],
1207 &dst_root->cgrp.e_csets[ss->id]);
1208 up_write(&css_set_rwsem);
1209
1210 src_root->subsys_mask &= ~(1 << ssid);
1211 src_root->cgrp.child_subsys_mask &= ~(1 << ssid);
1212
1213 /* default hierarchy doesn't enable controllers by default */
1214 dst_root->subsys_mask |= 1 << ssid;
1215 if (dst_root != &cgrp_dfl_root)
1216 dst_root->cgrp.child_subsys_mask |= 1 << ssid;
1217
1218 if (ss->bind)
1219 ss->bind(css);
1220 }
1221
1222 kernfs_activate(dst_root->cgrp.kn);
1223 return 0;
1224}
1225
1226static int cgroup_show_options(struct seq_file *seq,
1227 struct kernfs_root *kf_root)
1228{
1229 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1230 struct cgroup_subsys *ss;
1231 int ssid;
1232
1233 for_each_subsys(ss, ssid)
1234 if (root->subsys_mask & (1 << ssid))
1235 seq_printf(seq, ",%s", ss->name);
1236 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1237 seq_puts(seq, ",sane_behavior");
1238 if (root->flags & CGRP_ROOT_NOPREFIX)
1239 seq_puts(seq, ",noprefix");
1240 if (root->flags & CGRP_ROOT_XATTR)
1241 seq_puts(seq, ",xattr");
1242
1243 spin_lock(&release_agent_path_lock);
1244 if (strlen(root->release_agent_path))
1245 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1246 spin_unlock(&release_agent_path_lock);
1247
1248 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1249 seq_puts(seq, ",clone_children");
1250 if (strlen(root->name))
1251 seq_printf(seq, ",name=%s", root->name);
1252 return 0;
1253}
1254
1255struct cgroup_sb_opts {
1256 unsigned int subsys_mask;
1257 unsigned int flags;
1258 char *release_agent;
1259 bool cpuset_clone_children;
1260 char *name;
1261 /* User explicitly requested empty subsystem */
1262 bool none;
1263};
1264
1265static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1266{
1267 char *token, *o = data;
1268 bool all_ss = false, one_ss = false;
1269 unsigned int mask = -1U;
1270 struct cgroup_subsys *ss;
1271 int i;
1272
1273#ifdef CONFIG_CPUSETS
1274 mask = ~(1U << cpuset_cgrp_id);
1275#endif
1276
1277 memset(opts, 0, sizeof(*opts));
1278
1279 while ((token = strsep(&o, ",")) != NULL) {
1280 if (!*token)
1281 return -EINVAL;
1282 if (!strcmp(token, "none")) {
1283 /* Explicitly have no subsystems */
1284 opts->none = true;
1285 continue;
1286 }
1287 if (!strcmp(token, "all")) {
1288 /* Mutually exclusive option 'all' + subsystem name */
1289 if (one_ss)
1290 return -EINVAL;
1291 all_ss = true;
1292 continue;
1293 }
1294 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1295 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1296 continue;
1297 }
1298 if (!strcmp(token, "noprefix")) {
1299 opts->flags |= CGRP_ROOT_NOPREFIX;
1300 continue;
1301 }
1302 if (!strcmp(token, "clone_children")) {
1303 opts->cpuset_clone_children = true;
1304 continue;
1305 }
1306 if (!strcmp(token, "xattr")) {
1307 opts->flags |= CGRP_ROOT_XATTR;
1308 continue;
1309 }
1310 if (!strncmp(token, "release_agent=", 14)) {
1311 /* Specifying two release agents is forbidden */
1312 if (opts->release_agent)
1313 return -EINVAL;
1314 opts->release_agent =
1315 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1316 if (!opts->release_agent)
1317 return -ENOMEM;
1318 continue;
1319 }
1320 if (!strncmp(token, "name=", 5)) {
1321 const char *name = token + 5;
1322 /* Can't specify an empty name */
1323 if (!strlen(name))
1324 return -EINVAL;
1325 /* Must match [\w.-]+ */
1326 for (i = 0; i < strlen(name); i++) {
1327 char c = name[i];
1328 if (isalnum(c))
1329 continue;
1330 if ((c == '.') || (c == '-') || (c == '_'))
1331 continue;
1332 return -EINVAL;
1333 }
1334 /* Specifying two names is forbidden */
1335 if (opts->name)
1336 return -EINVAL;
1337 opts->name = kstrndup(name,
1338 MAX_CGROUP_ROOT_NAMELEN - 1,
1339 GFP_KERNEL);
1340 if (!opts->name)
1341 return -ENOMEM;
1342
1343 continue;
1344 }
1345
1346 for_each_subsys(ss, i) {
1347 if (strcmp(token, ss->name))
1348 continue;
1349 if (ss->disabled)
1350 continue;
1351
1352 /* Mutually exclusive option 'all' + subsystem name */
1353 if (all_ss)
1354 return -EINVAL;
1355 opts->subsys_mask |= (1 << i);
1356 one_ss = true;
1357
1358 break;
1359 }
1360 if (i == CGROUP_SUBSYS_COUNT)
1361 return -ENOENT;
1362 }
1363
1364 /* Consistency checks */
1365
1366 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1367 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1368
1369 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1370 opts->cpuset_clone_children || opts->release_agent ||
1371 opts->name) {
1372 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1373 return -EINVAL;
1374 }
1375 } else {
1376 /*
1377 * If the 'all' option was specified select all the
1378 * subsystems, otherwise if 'none', 'name=' and a subsystem
1379 * name options were not specified, let's default to 'all'
1380 */
1381 if (all_ss || (!one_ss && !opts->none && !opts->name))
1382 for_each_subsys(ss, i)
1383 if (!ss->disabled)
1384 opts->subsys_mask |= (1 << i);
1385
1386 /*
1387 * We either have to specify by name or by subsystems. (So
1388 * all empty hierarchies must have a name).
1389 */
1390 if (!opts->subsys_mask && !opts->name)
1391 return -EINVAL;
1392 }
1393
1394 /*
1395 * Option noprefix was introduced just for backward compatibility
1396 * with the old cpuset, so we allow noprefix only if mounting just
1397 * the cpuset subsystem.
1398 */
1399 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1400 return -EINVAL;
1401
1402
1403 /* Can't specify "none" and some subsystems */
1404 if (opts->subsys_mask && opts->none)
1405 return -EINVAL;
1406
1407 return 0;
1408}
1409
1410static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1411{
1412 int ret = 0;
1413 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1414 struct cgroup_sb_opts opts;
1415 unsigned int added_mask, removed_mask;
1416
1417 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1418 pr_err("sane_behavior: remount is not allowed\n");
1419 return -EINVAL;
1420 }
1421
1422 mutex_lock(&cgroup_mutex);
1423
1424 /* See what subsystems are wanted */
1425 ret = parse_cgroupfs_options(data, &opts);
1426 if (ret)
1427 goto out_unlock;
1428
1429 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1430 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1431 task_tgid_nr(current), current->comm);
1432
1433 added_mask = opts.subsys_mask & ~root->subsys_mask;
1434 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1435
1436 /* Don't allow flags or name to change at remount */
1437 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1438 (opts.name && strcmp(opts.name, root->name))) {
1439 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1440 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1441 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1442 ret = -EINVAL;
1443 goto out_unlock;
1444 }
1445
1446 /* remounting is not allowed for populated hierarchies */
1447 if (!list_empty(&root->cgrp.self.children)) {
1448 ret = -EBUSY;
1449 goto out_unlock;
1450 }
1451
1452 ret = rebind_subsystems(root, added_mask);
1453 if (ret)
1454 goto out_unlock;
1455
1456 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1457
1458 if (opts.release_agent) {
1459 spin_lock(&release_agent_path_lock);
1460 strcpy(root->release_agent_path, opts.release_agent);
1461 spin_unlock(&release_agent_path_lock);
1462 }
1463 out_unlock:
1464 kfree(opts.release_agent);
1465 kfree(opts.name);
1466 mutex_unlock(&cgroup_mutex);
1467 return ret;
1468}
1469
1470/*
1471 * To reduce the fork() overhead for systems that are not actually using
1472 * their cgroups capability, we don't maintain the lists running through
1473 * each css_set to its tasks until we see the list actually used - in other
1474 * words after the first mount.
1475 */
1476static bool use_task_css_set_links __read_mostly;
1477
1478static void cgroup_enable_task_cg_lists(void)
1479{
1480 struct task_struct *p, *g;
1481
1482 down_write(&css_set_rwsem);
1483
1484 if (use_task_css_set_links)
1485 goto out_unlock;
1486
1487 use_task_css_set_links = true;
1488
1489 /*
1490 * We need tasklist_lock because RCU is not safe against
1491 * while_each_thread(). Besides, a forking task that has passed
1492 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1493 * is not guaranteed to have its child immediately visible in the
1494 * tasklist if we walk through it with RCU.
1495 */
1496 read_lock(&tasklist_lock);
1497 do_each_thread(g, p) {
1498 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1499 task_css_set(p) != &init_css_set);
1500
1501 /*
1502 * We should check if the process is exiting, otherwise
1503 * it will race with cgroup_exit() in that the list
1504 * entry won't be deleted though the process has exited.
1505 * Do it while holding siglock so that we don't end up
1506 * racing against cgroup_exit().
1507 */
1508 spin_lock_irq(&p->sighand->siglock);
1509 if (!(p->flags & PF_EXITING)) {
1510 struct css_set *cset = task_css_set(p);
1511
1512 list_add(&p->cg_list, &cset->tasks);
1513 get_css_set(cset);
1514 }
1515 spin_unlock_irq(&p->sighand->siglock);
1516 } while_each_thread(g, p);
1517 read_unlock(&tasklist_lock);
1518out_unlock:
1519 up_write(&css_set_rwsem);
1520}
1521
1522static void init_cgroup_housekeeping(struct cgroup *cgrp)
1523{
1524 struct cgroup_subsys *ss;
1525 int ssid;
1526
1527 INIT_LIST_HEAD(&cgrp->self.sibling);
1528 INIT_LIST_HEAD(&cgrp->self.children);
1529 INIT_LIST_HEAD(&cgrp->cset_links);
1530 INIT_LIST_HEAD(&cgrp->release_list);
1531 INIT_LIST_HEAD(&cgrp->pidlists);
1532 mutex_init(&cgrp->pidlist_mutex);
1533 cgrp->self.cgroup = cgrp;
1534 cgrp->self.flags |= CSS_ONLINE;
1535
1536 for_each_subsys(ss, ssid)
1537 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1538
1539 init_waitqueue_head(&cgrp->offline_waitq);
1540}
1541
1542static void init_cgroup_root(struct cgroup_root *root,
1543 struct cgroup_sb_opts *opts)
1544{
1545 struct cgroup *cgrp = &root->cgrp;
1546
1547 INIT_LIST_HEAD(&root->root_list);
1548 atomic_set(&root->nr_cgrps, 1);
1549 cgrp->root = root;
1550 init_cgroup_housekeeping(cgrp);
1551 idr_init(&root->cgroup_idr);
1552
1553 root->flags = opts->flags;
1554 if (opts->release_agent)
1555 strcpy(root->release_agent_path, opts->release_agent);
1556 if (opts->name)
1557 strcpy(root->name, opts->name);
1558 if (opts->cpuset_clone_children)
1559 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1560}
1561
1562static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1563{
1564 LIST_HEAD(tmp_links);
1565 struct cgroup *root_cgrp = &root->cgrp;
1566 struct css_set *cset;
1567 int i, ret;
1568
1569 lockdep_assert_held(&cgroup_mutex);
1570
1571 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1572 if (ret < 0)
1573 goto out;
1574 root_cgrp->id = ret;
1575
1576 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release);
1577 if (ret)
1578 goto out;
1579
1580 /*
1581 * We're accessing css_set_count without locking css_set_rwsem here,
1582 * but that's OK - it can only be increased by someone holding
1583 * cgroup_lock, and that's us. The worst that can happen is that we
1584 * have some link structures left over
1585 */
1586 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1587 if (ret)
1588 goto cancel_ref;
1589
1590 ret = cgroup_init_root_id(root);
1591 if (ret)
1592 goto cancel_ref;
1593
1594 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1595 KERNFS_ROOT_CREATE_DEACTIVATED,
1596 root_cgrp);
1597 if (IS_ERR(root->kf_root)) {
1598 ret = PTR_ERR(root->kf_root);
1599 goto exit_root_id;
1600 }
1601 root_cgrp->kn = root->kf_root->kn;
1602
1603 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1604 if (ret)
1605 goto destroy_root;
1606
1607 ret = rebind_subsystems(root, ss_mask);
1608 if (ret)
1609 goto destroy_root;
1610
1611 /*
1612 * There must be no failure case after here, since rebinding takes
1613 * care of subsystems' refcounts, which are explicitly dropped in
1614 * the failure exit path.
1615 */
1616 list_add(&root->root_list, &cgroup_roots);
1617 cgroup_root_count++;
1618
1619 /*
1620 * Link the root cgroup in this hierarchy into all the css_set
1621 * objects.
1622 */
1623 down_write(&css_set_rwsem);
1624 hash_for_each(css_set_table, i, cset, hlist)
1625 link_css_set(&tmp_links, cset, root_cgrp);
1626 up_write(&css_set_rwsem);
1627
1628 BUG_ON(!list_empty(&root_cgrp->self.children));
1629 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1630
1631 kernfs_activate(root_cgrp->kn);
1632 ret = 0;
1633 goto out;
1634
1635destroy_root:
1636 kernfs_destroy_root(root->kf_root);
1637 root->kf_root = NULL;
1638exit_root_id:
1639 cgroup_exit_root_id(root);
1640cancel_ref:
1641 percpu_ref_exit(&root_cgrp->self.refcnt);
1642out:
1643 free_cgrp_cset_links(&tmp_links);
1644 return ret;
1645}
1646
1647static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1648 int flags, const char *unused_dev_name,
1649 void *data)
1650{
1651 struct cgroup_root *root;
1652 struct cgroup_sb_opts opts;
1653 struct dentry *dentry;
1654 int ret;
1655 bool new_sb;
1656
1657 /*
1658 * The first time anyone tries to mount a cgroup, enable the list
1659 * linking each css_set to its tasks and fix up all existing tasks.
1660 */
1661 if (!use_task_css_set_links)
1662 cgroup_enable_task_cg_lists();
1663
1664 mutex_lock(&cgroup_mutex);
1665
1666 /* First find the desired set of subsystems */
1667 ret = parse_cgroupfs_options(data, &opts);
1668 if (ret)
1669 goto out_unlock;
1670
1671 /* look for a matching existing root */
1672 if (!opts.subsys_mask && !opts.none && !opts.name) {
1673 cgrp_dfl_root_visible = true;
1674 root = &cgrp_dfl_root;
1675 cgroup_get(&root->cgrp);
1676 ret = 0;
1677 goto out_unlock;
1678 }
1679
1680 for_each_root(root) {
1681 bool name_match = false;
1682
1683 if (root == &cgrp_dfl_root)
1684 continue;
1685
1686 /*
1687 * If we asked for a name then it must match. Also, if
1688 * name matches but sybsys_mask doesn't, we should fail.
1689 * Remember whether name matched.
1690 */
1691 if (opts.name) {
1692 if (strcmp(opts.name, root->name))
1693 continue;
1694 name_match = true;
1695 }
1696
1697 /*
1698 * If we asked for subsystems (or explicitly for no
1699 * subsystems) then they must match.
1700 */
1701 if ((opts.subsys_mask || opts.none) &&
1702 (opts.subsys_mask != root->subsys_mask)) {
1703 if (!name_match)
1704 continue;
1705 ret = -EBUSY;
1706 goto out_unlock;
1707 }
1708
1709 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1710 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1711 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1712 ret = -EINVAL;
1713 goto out_unlock;
1714 } else {
1715 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1716 }
1717 }
1718
1719 /*
1720 * A root's lifetime is governed by its root cgroup.
1721 * tryget_live failure indicate that the root is being
1722 * destroyed. Wait for destruction to complete so that the
1723 * subsystems are free. We can use wait_queue for the wait
1724 * but this path is super cold. Let's just sleep for a bit
1725 * and retry.
1726 */
1727 if (!percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1728 mutex_unlock(&cgroup_mutex);
1729 msleep(10);
1730 ret = restart_syscall();
1731 goto out_free;
1732 }
1733
1734 ret = 0;
1735 goto out_unlock;
1736 }
1737
1738 /*
1739 * No such thing, create a new one. name= matching without subsys
1740 * specification is allowed for already existing hierarchies but we
1741 * can't create new one without subsys specification.
1742 */
1743 if (!opts.subsys_mask && !opts.none) {
1744 ret = -EINVAL;
1745 goto out_unlock;
1746 }
1747
1748 root = kzalloc(sizeof(*root), GFP_KERNEL);
1749 if (!root) {
1750 ret = -ENOMEM;
1751 goto out_unlock;
1752 }
1753
1754 init_cgroup_root(root, &opts);
1755
1756 ret = cgroup_setup_root(root, opts.subsys_mask);
1757 if (ret)
1758 cgroup_free_root(root);
1759
1760out_unlock:
1761 mutex_unlock(&cgroup_mutex);
1762out_free:
1763 kfree(opts.release_agent);
1764 kfree(opts.name);
1765
1766 if (ret)
1767 return ERR_PTR(ret);
1768
1769 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1770 CGROUP_SUPER_MAGIC, &new_sb);
1771 if (IS_ERR(dentry) || !new_sb)
1772 cgroup_put(&root->cgrp);
1773 return dentry;
1774}
1775
1776static void cgroup_kill_sb(struct super_block *sb)
1777{
1778 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1779 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1780
1781 /*
1782 * If @root doesn't have any mounts or children, start killing it.
1783 * This prevents new mounts by disabling percpu_ref_tryget_live().
1784 * cgroup_mount() may wait for @root's release.
1785 *
1786 * And don't kill the default root.
1787 */
1788 if (css_has_online_children(&root->cgrp.self) ||
1789 root == &cgrp_dfl_root)
1790 cgroup_put(&root->cgrp);
1791 else
1792 percpu_ref_kill(&root->cgrp.self.refcnt);
1793
1794 kernfs_kill_sb(sb);
1795}
1796
1797static struct file_system_type cgroup_fs_type = {
1798 .name = "cgroup",
1799 .mount = cgroup_mount,
1800 .kill_sb = cgroup_kill_sb,
1801};
1802
1803static struct kobject *cgroup_kobj;
1804
1805/**
1806 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1807 * @task: target task
1808 * @buf: the buffer to write the path into
1809 * @buflen: the length of the buffer
1810 *
1811 * Determine @task's cgroup on the first (the one with the lowest non-zero
1812 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1813 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1814 * cgroup controller callbacks.
1815 *
1816 * Return value is the same as kernfs_path().
1817 */
1818char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1819{
1820 struct cgroup_root *root;
1821 struct cgroup *cgrp;
1822 int hierarchy_id = 1;
1823 char *path = NULL;
1824
1825 mutex_lock(&cgroup_mutex);
1826 down_read(&css_set_rwsem);
1827
1828 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1829
1830 if (root) {
1831 cgrp = task_cgroup_from_root(task, root);
1832 path = cgroup_path(cgrp, buf, buflen);
1833 } else {
1834 /* if no hierarchy exists, everyone is in "/" */
1835 if (strlcpy(buf, "/", buflen) < buflen)
1836 path = buf;
1837 }
1838
1839 up_read(&css_set_rwsem);
1840 mutex_unlock(&cgroup_mutex);
1841 return path;
1842}
1843EXPORT_SYMBOL_GPL(task_cgroup_path);
1844
1845/* used to track tasks and other necessary states during migration */
1846struct cgroup_taskset {
1847 /* the src and dst cset list running through cset->mg_node */
1848 struct list_head src_csets;
1849 struct list_head dst_csets;
1850
1851 /*
1852 * Fields for cgroup_taskset_*() iteration.
1853 *
1854 * Before migration is committed, the target migration tasks are on
1855 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1856 * the csets on ->dst_csets. ->csets point to either ->src_csets
1857 * or ->dst_csets depending on whether migration is committed.
1858 *
1859 * ->cur_csets and ->cur_task point to the current task position
1860 * during iteration.
1861 */
1862 struct list_head *csets;
1863 struct css_set *cur_cset;
1864 struct task_struct *cur_task;
1865};
1866
1867/**
1868 * cgroup_taskset_first - reset taskset and return the first task
1869 * @tset: taskset of interest
1870 *
1871 * @tset iteration is initialized and the first task is returned.
1872 */
1873struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1874{
1875 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1876 tset->cur_task = NULL;
1877
1878 return cgroup_taskset_next(tset);
1879}
1880
1881/**
1882 * cgroup_taskset_next - iterate to the next task in taskset
1883 * @tset: taskset of interest
1884 *
1885 * Return the next task in @tset. Iteration must have been initialized
1886 * with cgroup_taskset_first().
1887 */
1888struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1889{
1890 struct css_set *cset = tset->cur_cset;
1891 struct task_struct *task = tset->cur_task;
1892
1893 while (&cset->mg_node != tset->csets) {
1894 if (!task)
1895 task = list_first_entry(&cset->mg_tasks,
1896 struct task_struct, cg_list);
1897 else
1898 task = list_next_entry(task, cg_list);
1899
1900 if (&task->cg_list != &cset->mg_tasks) {
1901 tset->cur_cset = cset;
1902 tset->cur_task = task;
1903 return task;
1904 }
1905
1906 cset = list_next_entry(cset, mg_node);
1907 task = NULL;
1908 }
1909
1910 return NULL;
1911}
1912
1913/**
1914 * cgroup_task_migrate - move a task from one cgroup to another.
1915 * @old_cgrp: the cgroup @tsk is being migrated from
1916 * @tsk: the task being migrated
1917 * @new_cset: the new css_set @tsk is being attached to
1918 *
1919 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1920 */
1921static void cgroup_task_migrate(struct cgroup *old_cgrp,
1922 struct task_struct *tsk,
1923 struct css_set *new_cset)
1924{
1925 struct css_set *old_cset;
1926
1927 lockdep_assert_held(&cgroup_mutex);
1928 lockdep_assert_held(&css_set_rwsem);
1929
1930 /*
1931 * We are synchronized through threadgroup_lock() against PF_EXITING
1932 * setting such that we can't race against cgroup_exit() changing the
1933 * css_set to init_css_set and dropping the old one.
1934 */
1935 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1936 old_cset = task_css_set(tsk);
1937
1938 get_css_set(new_cset);
1939 rcu_assign_pointer(tsk->cgroups, new_cset);
1940
1941 /*
1942 * Use move_tail so that cgroup_taskset_first() still returns the
1943 * leader after migration. This works because cgroup_migrate()
1944 * ensures that the dst_cset of the leader is the first on the
1945 * tset's dst_csets list.
1946 */
1947 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1948
1949 /*
1950 * We just gained a reference on old_cset by taking it from the
1951 * task. As trading it for new_cset is protected by cgroup_mutex,
1952 * we're safe to drop it here; it will be freed under RCU.
1953 */
1954 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1955 put_css_set_locked(old_cset, false);
1956}
1957
1958/**
1959 * cgroup_migrate_finish - cleanup after attach
1960 * @preloaded_csets: list of preloaded css_sets
1961 *
1962 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1963 * those functions for details.
1964 */
1965static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1966{
1967 struct css_set *cset, *tmp_cset;
1968
1969 lockdep_assert_held(&cgroup_mutex);
1970
1971 down_write(&css_set_rwsem);
1972 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1973 cset->mg_src_cgrp = NULL;
1974 cset->mg_dst_cset = NULL;
1975 list_del_init(&cset->mg_preload_node);
1976 put_css_set_locked(cset, false);
1977 }
1978 up_write(&css_set_rwsem);
1979}
1980
1981/**
1982 * cgroup_migrate_add_src - add a migration source css_set
1983 * @src_cset: the source css_set to add
1984 * @dst_cgrp: the destination cgroup
1985 * @preloaded_csets: list of preloaded css_sets
1986 *
1987 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1988 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1989 * up by cgroup_migrate_finish().
1990 *
1991 * This function may be called without holding threadgroup_lock even if the
1992 * target is a process. Threads may be created and destroyed but as long
1993 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1994 * the preloaded css_sets are guaranteed to cover all migrations.
1995 */
1996static void cgroup_migrate_add_src(struct css_set *src_cset,
1997 struct cgroup *dst_cgrp,
1998 struct list_head *preloaded_csets)
1999{
2000 struct cgroup *src_cgrp;
2001
2002 lockdep_assert_held(&cgroup_mutex);
2003 lockdep_assert_held(&css_set_rwsem);
2004
2005 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2006
2007 if (!list_empty(&src_cset->mg_preload_node))
2008 return;
2009
2010 WARN_ON(src_cset->mg_src_cgrp);
2011 WARN_ON(!list_empty(&src_cset->mg_tasks));
2012 WARN_ON(!list_empty(&src_cset->mg_node));
2013
2014 src_cset->mg_src_cgrp = src_cgrp;
2015 get_css_set(src_cset);
2016 list_add(&src_cset->mg_preload_node, preloaded_csets);
2017}
2018
2019/**
2020 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2021 * @dst_cgrp: the destination cgroup (may be %NULL)
2022 * @preloaded_csets: list of preloaded source css_sets
2023 *
2024 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2025 * have been preloaded to @preloaded_csets. This function looks up and
2026 * pins all destination css_sets, links each to its source, and append them
2027 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2028 * source css_set is assumed to be its cgroup on the default hierarchy.
2029 *
2030 * This function must be called after cgroup_migrate_add_src() has been
2031 * called on each migration source css_set. After migration is performed
2032 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2033 * @preloaded_csets.
2034 */
2035static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2036 struct list_head *preloaded_csets)
2037{
2038 LIST_HEAD(csets);
2039 struct css_set *src_cset, *tmp_cset;
2040
2041 lockdep_assert_held(&cgroup_mutex);
2042
2043 /*
2044 * Except for the root, child_subsys_mask must be zero for a cgroup
2045 * with tasks so that child cgroups don't compete against tasks.
2046 */
2047 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2048 dst_cgrp->child_subsys_mask)
2049 return -EBUSY;
2050
2051 /* look up the dst cset for each src cset and link it to src */
2052 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2053 struct css_set *dst_cset;
2054
2055 dst_cset = find_css_set(src_cset,
2056 dst_cgrp ?: src_cset->dfl_cgrp);
2057 if (!dst_cset)
2058 goto err;
2059
2060 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2061
2062 /*
2063 * If src cset equals dst, it's noop. Drop the src.
2064 * cgroup_migrate() will skip the cset too. Note that we
2065 * can't handle src == dst as some nodes are used by both.
2066 */
2067 if (src_cset == dst_cset) {
2068 src_cset->mg_src_cgrp = NULL;
2069 list_del_init(&src_cset->mg_preload_node);
2070 put_css_set(src_cset, false);
2071 put_css_set(dst_cset, false);
2072 continue;
2073 }
2074
2075 src_cset->mg_dst_cset = dst_cset;
2076
2077 if (list_empty(&dst_cset->mg_preload_node))
2078 list_add(&dst_cset->mg_preload_node, &csets);
2079 else
2080 put_css_set(dst_cset, false);
2081 }
2082
2083 list_splice_tail(&csets, preloaded_csets);
2084 return 0;
2085err:
2086 cgroup_migrate_finish(&csets);
2087 return -ENOMEM;
2088}
2089
2090/**
2091 * cgroup_migrate - migrate a process or task to a cgroup
2092 * @cgrp: the destination cgroup
2093 * @leader: the leader of the process or the task to migrate
2094 * @threadgroup: whether @leader points to the whole process or a single task
2095 *
2096 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2097 * process, the caller must be holding threadgroup_lock of @leader. The
2098 * caller is also responsible for invoking cgroup_migrate_add_src() and
2099 * cgroup_migrate_prepare_dst() on the targets before invoking this
2100 * function and following up with cgroup_migrate_finish().
2101 *
2102 * As long as a controller's ->can_attach() doesn't fail, this function is
2103 * guaranteed to succeed. This means that, excluding ->can_attach()
2104 * failure, when migrating multiple targets, the success or failure can be
2105 * decided for all targets by invoking group_migrate_prepare_dst() before
2106 * actually starting migrating.
2107 */
2108static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2109 bool threadgroup)
2110{
2111 struct cgroup_taskset tset = {
2112 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2113 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2114 .csets = &tset.src_csets,
2115 };
2116 struct cgroup_subsys_state *css, *failed_css = NULL;
2117 struct css_set *cset, *tmp_cset;
2118 struct task_struct *task, *tmp_task;
2119 int i, ret;
2120
2121 /*
2122 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2123 * already PF_EXITING could be freed from underneath us unless we
2124 * take an rcu_read_lock.
2125 */
2126 down_write(&css_set_rwsem);
2127 rcu_read_lock();
2128 task = leader;
2129 do {
2130 /* @task either already exited or can't exit until the end */
2131 if (task->flags & PF_EXITING)
2132 goto next;
2133
2134 /* leave @task alone if post_fork() hasn't linked it yet */
2135 if (list_empty(&task->cg_list))
2136 goto next;
2137
2138 cset = task_css_set(task);
2139 if (!cset->mg_src_cgrp)
2140 goto next;
2141
2142 /*
2143 * cgroup_taskset_first() must always return the leader.
2144 * Take care to avoid disturbing the ordering.
2145 */
2146 list_move_tail(&task->cg_list, &cset->mg_tasks);
2147 if (list_empty(&cset->mg_node))
2148 list_add_tail(&cset->mg_node, &tset.src_csets);
2149 if (list_empty(&cset->mg_dst_cset->mg_node))
2150 list_move_tail(&cset->mg_dst_cset->mg_node,
2151 &tset.dst_csets);
2152 next:
2153 if (!threadgroup)
2154 break;
2155 } while_each_thread(leader, task);
2156 rcu_read_unlock();
2157 up_write(&css_set_rwsem);
2158
2159 /* methods shouldn't be called if no task is actually migrating */
2160 if (list_empty(&tset.src_csets))
2161 return 0;
2162
2163 /* check that we can legitimately attach to the cgroup */
2164 for_each_e_css(css, i, cgrp) {
2165 if (css->ss->can_attach) {
2166 ret = css->ss->can_attach(css, &tset);
2167 if (ret) {
2168 failed_css = css;
2169 goto out_cancel_attach;
2170 }
2171 }
2172 }
2173
2174 /*
2175 * Now that we're guaranteed success, proceed to move all tasks to
2176 * the new cgroup. There are no failure cases after here, so this
2177 * is the commit point.
2178 */
2179 down_write(&css_set_rwsem);
2180 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2181 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2182 cgroup_task_migrate(cset->mg_src_cgrp, task,
2183 cset->mg_dst_cset);
2184 }
2185 up_write(&css_set_rwsem);
2186
2187 /*
2188 * Migration is committed, all target tasks are now on dst_csets.
2189 * Nothing is sensitive to fork() after this point. Notify
2190 * controllers that migration is complete.
2191 */
2192 tset.csets = &tset.dst_csets;
2193
2194 for_each_e_css(css, i, cgrp)
2195 if (css->ss->attach)
2196 css->ss->attach(css, &tset);
2197
2198 ret = 0;
2199 goto out_release_tset;
2200
2201out_cancel_attach:
2202 for_each_e_css(css, i, cgrp) {
2203 if (css == failed_css)
2204 break;
2205 if (css->ss->cancel_attach)
2206 css->ss->cancel_attach(css, &tset);
2207 }
2208out_release_tset:
2209 down_write(&css_set_rwsem);
2210 list_splice_init(&tset.dst_csets, &tset.src_csets);
2211 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2212 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2213 list_del_init(&cset->mg_node);
2214 }
2215 up_write(&css_set_rwsem);
2216 return ret;
2217}
2218
2219/**
2220 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2221 * @dst_cgrp: the cgroup to attach to
2222 * @leader: the task or the leader of the threadgroup to be attached
2223 * @threadgroup: attach the whole threadgroup?
2224 *
2225 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2226 */
2227static int cgroup_attach_task(struct cgroup *dst_cgrp,
2228 struct task_struct *leader, bool threadgroup)
2229{
2230 LIST_HEAD(preloaded_csets);
2231 struct task_struct *task;
2232 int ret;
2233
2234 /* look up all src csets */
2235 down_read(&css_set_rwsem);
2236 rcu_read_lock();
2237 task = leader;
2238 do {
2239 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2240 &preloaded_csets);
2241 if (!threadgroup)
2242 break;
2243 } while_each_thread(leader, task);
2244 rcu_read_unlock();
2245 up_read(&css_set_rwsem);
2246
2247 /* prepare dst csets and commit */
2248 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2249 if (!ret)
2250 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2251
2252 cgroup_migrate_finish(&preloaded_csets);
2253 return ret;
2254}
2255
2256/*
2257 * Find the task_struct of the task to attach by vpid and pass it along to the
2258 * function to attach either it or all tasks in its threadgroup. Will lock
2259 * cgroup_mutex and threadgroup.
2260 */
2261static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2262 size_t nbytes, loff_t off, bool threadgroup)
2263{
2264 struct task_struct *tsk;
2265 const struct cred *cred = current_cred(), *tcred;
2266 struct cgroup *cgrp;
2267 pid_t pid;
2268 int ret;
2269
2270 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2271 return -EINVAL;
2272
2273 cgrp = cgroup_kn_lock_live(of->kn);
2274 if (!cgrp)
2275 return -ENODEV;
2276
2277retry_find_task:
2278 rcu_read_lock();
2279 if (pid) {
2280 tsk = find_task_by_vpid(pid);
2281 if (!tsk) {
2282 rcu_read_unlock();
2283 ret = -ESRCH;
2284 goto out_unlock_cgroup;
2285 }
2286 /*
2287 * even if we're attaching all tasks in the thread group, we
2288 * only need to check permissions on one of them.
2289 */
2290 tcred = __task_cred(tsk);
2291 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2292 !uid_eq(cred->euid, tcred->uid) &&
2293 !uid_eq(cred->euid, tcred->suid)) {
2294 rcu_read_unlock();
2295 ret = -EACCES;
2296 goto out_unlock_cgroup;
2297 }
2298 } else
2299 tsk = current;
2300
2301 if (threadgroup)
2302 tsk = tsk->group_leader;
2303
2304 /*
2305 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2306 * trapped in a cpuset, or RT worker may be born in a cgroup
2307 * with no rt_runtime allocated. Just say no.
2308 */
2309 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2310 ret = -EINVAL;
2311 rcu_read_unlock();
2312 goto out_unlock_cgroup;
2313 }
2314
2315 get_task_struct(tsk);
2316 rcu_read_unlock();
2317
2318 threadgroup_lock(tsk);
2319 if (threadgroup) {
2320 if (!thread_group_leader(tsk)) {
2321 /*
2322 * a race with de_thread from another thread's exec()
2323 * may strip us of our leadership, if this happens,
2324 * there is no choice but to throw this task away and
2325 * try again; this is
2326 * "double-double-toil-and-trouble-check locking".
2327 */
2328 threadgroup_unlock(tsk);
2329 put_task_struct(tsk);
2330 goto retry_find_task;
2331 }
2332 }
2333
2334 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2335
2336 threadgroup_unlock(tsk);
2337
2338 put_task_struct(tsk);
2339out_unlock_cgroup:
2340 cgroup_kn_unlock(of->kn);
2341 return ret ?: nbytes;
2342}
2343
2344/**
2345 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2346 * @from: attach to all cgroups of a given task
2347 * @tsk: the task to be attached
2348 */
2349int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2350{
2351 struct cgroup_root *root;
2352 int retval = 0;
2353
2354 mutex_lock(&cgroup_mutex);
2355 for_each_root(root) {
2356 struct cgroup *from_cgrp;
2357
2358 if (root == &cgrp_dfl_root)
2359 continue;
2360
2361 down_read(&css_set_rwsem);
2362 from_cgrp = task_cgroup_from_root(from, root);
2363 up_read(&css_set_rwsem);
2364
2365 retval = cgroup_attach_task(from_cgrp, tsk, false);
2366 if (retval)
2367 break;
2368 }
2369 mutex_unlock(&cgroup_mutex);
2370
2371 return retval;
2372}
2373EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2374
2375static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2376 char *buf, size_t nbytes, loff_t off)
2377{
2378 return __cgroup_procs_write(of, buf, nbytes, off, false);
2379}
2380
2381static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2382 char *buf, size_t nbytes, loff_t off)
2383{
2384 return __cgroup_procs_write(of, buf, nbytes, off, true);
2385}
2386
2387static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2388 char *buf, size_t nbytes, loff_t off)
2389{
2390 struct cgroup *cgrp;
2391
2392 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2393
2394 cgrp = cgroup_kn_lock_live(of->kn);
2395 if (!cgrp)
2396 return -ENODEV;
2397 spin_lock(&release_agent_path_lock);
2398 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2399 sizeof(cgrp->root->release_agent_path));
2400 spin_unlock(&release_agent_path_lock);
2401 cgroup_kn_unlock(of->kn);
2402 return nbytes;
2403}
2404
2405static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2406{
2407 struct cgroup *cgrp = seq_css(seq)->cgroup;
2408
2409 spin_lock(&release_agent_path_lock);
2410 seq_puts(seq, cgrp->root->release_agent_path);
2411 spin_unlock(&release_agent_path_lock);
2412 seq_putc(seq, '\n');
2413 return 0;
2414}
2415
2416static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2417{
2418 struct cgroup *cgrp = seq_css(seq)->cgroup;
2419
2420 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2421 return 0;
2422}
2423
2424static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2425{
2426 struct cgroup_subsys *ss;
2427 bool printed = false;
2428 int ssid;
2429
2430 for_each_subsys(ss, ssid) {
2431 if (ss_mask & (1 << ssid)) {
2432 if (printed)
2433 seq_putc(seq, ' ');
2434 seq_printf(seq, "%s", ss->name);
2435 printed = true;
2436 }
2437 }
2438 if (printed)
2439 seq_putc(seq, '\n');
2440}
2441
2442/* show controllers which are currently attached to the default hierarchy */
2443static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2444{
2445 struct cgroup *cgrp = seq_css(seq)->cgroup;
2446
2447 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2448 ~cgrp_dfl_root_inhibit_ss_mask);
2449 return 0;
2450}
2451
2452/* show controllers which are enabled from the parent */
2453static int cgroup_controllers_show(struct seq_file *seq, void *v)
2454{
2455 struct cgroup *cgrp = seq_css(seq)->cgroup;
2456
2457 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->child_subsys_mask);
2458 return 0;
2459}
2460
2461/* show controllers which are enabled for a given cgroup's children */
2462static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2463{
2464 struct cgroup *cgrp = seq_css(seq)->cgroup;
2465
2466 cgroup_print_ss_mask(seq, cgrp->child_subsys_mask);
2467 return 0;
2468}
2469
2470/**
2471 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2472 * @cgrp: root of the subtree to update csses for
2473 *
2474 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2475 * css associations need to be updated accordingly. This function looks up
2476 * all css_sets which are attached to the subtree, creates the matching
2477 * updated css_sets and migrates the tasks to the new ones.
2478 */
2479static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2480{
2481 LIST_HEAD(preloaded_csets);
2482 struct cgroup_subsys_state *css;
2483 struct css_set *src_cset;
2484 int ret;
2485
2486 lockdep_assert_held(&cgroup_mutex);
2487
2488 /* look up all csses currently attached to @cgrp's subtree */
2489 down_read(&css_set_rwsem);
2490 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2491 struct cgrp_cset_link *link;
2492
2493 /* self is not affected by child_subsys_mask change */
2494 if (css->cgroup == cgrp)
2495 continue;
2496
2497 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2498 cgroup_migrate_add_src(link->cset, cgrp,
2499 &preloaded_csets);
2500 }
2501 up_read(&css_set_rwsem);
2502
2503 /* NULL dst indicates self on default hierarchy */
2504 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2505 if (ret)
2506 goto out_finish;
2507
2508 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2509 struct task_struct *last_task = NULL, *task;
2510
2511 /* src_csets precede dst_csets, break on the first dst_cset */
2512 if (!src_cset->mg_src_cgrp)
2513 break;
2514
2515 /*
2516 * All tasks in src_cset need to be migrated to the
2517 * matching dst_cset. Empty it process by process. We
2518 * walk tasks but migrate processes. The leader might even
2519 * belong to a different cset but such src_cset would also
2520 * be among the target src_csets because the default
2521 * hierarchy enforces per-process membership.
2522 */
2523 while (true) {
2524 down_read(&css_set_rwsem);
2525 task = list_first_entry_or_null(&src_cset->tasks,
2526 struct task_struct, cg_list);
2527 if (task) {
2528 task = task->group_leader;
2529 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2530 get_task_struct(task);
2531 }
2532 up_read(&css_set_rwsem);
2533
2534 if (!task)
2535 break;
2536
2537 /* guard against possible infinite loop */
2538 if (WARN(last_task == task,
2539 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2540 goto out_finish;
2541 last_task = task;
2542
2543 threadgroup_lock(task);
2544 /* raced against de_thread() from another thread? */
2545 if (!thread_group_leader(task)) {
2546 threadgroup_unlock(task);
2547 put_task_struct(task);
2548 continue;
2549 }
2550
2551 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2552
2553 threadgroup_unlock(task);
2554 put_task_struct(task);
2555
2556 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2557 goto out_finish;
2558 }
2559 }
2560
2561out_finish:
2562 cgroup_migrate_finish(&preloaded_csets);
2563 return ret;
2564}
2565
2566/* change the enabled child controllers for a cgroup in the default hierarchy */
2567static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2568 char *buf, size_t nbytes,
2569 loff_t off)
2570{
2571 unsigned int enable = 0, disable = 0;
2572 struct cgroup *cgrp, *child;
2573 struct cgroup_subsys *ss;
2574 char *tok;
2575 int ssid, ret;
2576
2577 /*
2578 * Parse input - space separated list of subsystem names prefixed
2579 * with either + or -.
2580 */
2581 buf = strstrip(buf);
2582 while ((tok = strsep(&buf, " "))) {
2583 if (tok[0] == '\0')
2584 continue;
2585 for_each_subsys(ss, ssid) {
2586 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2587 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2588 continue;
2589
2590 if (*tok == '+') {
2591 enable |= 1 << ssid;
2592 disable &= ~(1 << ssid);
2593 } else if (*tok == '-') {
2594 disable |= 1 << ssid;
2595 enable &= ~(1 << ssid);
2596 } else {
2597 return -EINVAL;
2598 }
2599 break;
2600 }
2601 if (ssid == CGROUP_SUBSYS_COUNT)
2602 return -EINVAL;
2603 }
2604
2605 cgrp = cgroup_kn_lock_live(of->kn);
2606 if (!cgrp)
2607 return -ENODEV;
2608
2609 for_each_subsys(ss, ssid) {
2610 if (enable & (1 << ssid)) {
2611 if (cgrp->child_subsys_mask & (1 << ssid)) {
2612 enable &= ~(1 << ssid);
2613 continue;
2614 }
2615
2616 /*
2617 * Because css offlining is asynchronous, userland
2618 * might try to re-enable the same controller while
2619 * the previous instance is still around. In such
2620 * cases, wait till it's gone using offline_waitq.
2621 */
2622 cgroup_for_each_live_child(child, cgrp) {
2623 DEFINE_WAIT(wait);
2624
2625 if (!cgroup_css(child, ss))
2626 continue;
2627
2628 cgroup_get(child);
2629 prepare_to_wait(&child->offline_waitq, &wait,
2630 TASK_UNINTERRUPTIBLE);
2631 cgroup_kn_unlock(of->kn);
2632 schedule();
2633 finish_wait(&child->offline_waitq, &wait);
2634 cgroup_put(child);
2635
2636 return restart_syscall();
2637 }
2638
2639 /* unavailable or not enabled on the parent? */
2640 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2641 (cgroup_parent(cgrp) &&
2642 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ssid)))) {
2643 ret = -ENOENT;
2644 goto out_unlock;
2645 }
2646 } else if (disable & (1 << ssid)) {
2647 if (!(cgrp->child_subsys_mask & (1 << ssid))) {
2648 disable &= ~(1 << ssid);
2649 continue;
2650 }
2651
2652 /* a child has it enabled? */
2653 cgroup_for_each_live_child(child, cgrp) {
2654 if (child->child_subsys_mask & (1 << ssid)) {
2655 ret = -EBUSY;
2656 goto out_unlock;
2657 }
2658 }
2659 }
2660 }
2661
2662 if (!enable && !disable) {
2663 ret = 0;
2664 goto out_unlock;
2665 }
2666
2667 /*
2668 * Except for the root, child_subsys_mask must be zero for a cgroup
2669 * with tasks so that child cgroups don't compete against tasks.
2670 */
2671 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2672 ret = -EBUSY;
2673 goto out_unlock;
2674 }
2675
2676 /*
2677 * Create csses for enables and update child_subsys_mask. This
2678 * changes cgroup_e_css() results which in turn makes the
2679 * subsequent cgroup_update_dfl_csses() associate all tasks in the
2680 * subtree to the updated csses.
2681 */
2682 for_each_subsys(ss, ssid) {
2683 if (!(enable & (1 << ssid)))
2684 continue;
2685
2686 cgroup_for_each_live_child(child, cgrp) {
2687 ret = create_css(child, ss);
2688 if (ret)
2689 goto err_undo_css;
2690 }
2691 }
2692
2693 cgrp->child_subsys_mask |= enable;
2694 cgrp->child_subsys_mask &= ~disable;
2695
2696 ret = cgroup_update_dfl_csses(cgrp);
2697 if (ret)
2698 goto err_undo_css;
2699
2700 /* all tasks are now migrated away from the old csses, kill them */
2701 for_each_subsys(ss, ssid) {
2702 if (!(disable & (1 << ssid)))
2703 continue;
2704
2705 cgroup_for_each_live_child(child, cgrp)
2706 kill_css(cgroup_css(child, ss));
2707 }
2708
2709 kernfs_activate(cgrp->kn);
2710 ret = 0;
2711out_unlock:
2712 cgroup_kn_unlock(of->kn);
2713 return ret ?: nbytes;
2714
2715err_undo_css:
2716 cgrp->child_subsys_mask &= ~enable;
2717 cgrp->child_subsys_mask |= disable;
2718
2719 for_each_subsys(ss, ssid) {
2720 if (!(enable & (1 << ssid)))
2721 continue;
2722
2723 cgroup_for_each_live_child(child, cgrp) {
2724 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2725 if (css)
2726 kill_css(css);
2727 }
2728 }
2729 goto out_unlock;
2730}
2731
2732static int cgroup_populated_show(struct seq_file *seq, void *v)
2733{
2734 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2735 return 0;
2736}
2737
2738static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2739 size_t nbytes, loff_t off)
2740{
2741 struct cgroup *cgrp = of->kn->parent->priv;
2742 struct cftype *cft = of->kn->priv;
2743 struct cgroup_subsys_state *css;
2744 int ret;
2745
2746 if (cft->write)
2747 return cft->write(of, buf, nbytes, off);
2748
2749 /*
2750 * kernfs guarantees that a file isn't deleted with operations in
2751 * flight, which means that the matching css is and stays alive and
2752 * doesn't need to be pinned. The RCU locking is not necessary
2753 * either. It's just for the convenience of using cgroup_css().
2754 */
2755 rcu_read_lock();
2756 css = cgroup_css(cgrp, cft->ss);
2757 rcu_read_unlock();
2758
2759 if (cft->write_u64) {
2760 unsigned long long v;
2761 ret = kstrtoull(buf, 0, &v);
2762 if (!ret)
2763 ret = cft->write_u64(css, cft, v);
2764 } else if (cft->write_s64) {
2765 long long v;
2766 ret = kstrtoll(buf, 0, &v);
2767 if (!ret)
2768 ret = cft->write_s64(css, cft, v);
2769 } else {
2770 ret = -EINVAL;
2771 }
2772
2773 return ret ?: nbytes;
2774}
2775
2776static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2777{
2778 return seq_cft(seq)->seq_start(seq, ppos);
2779}
2780
2781static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2782{
2783 return seq_cft(seq)->seq_next(seq, v, ppos);
2784}
2785
2786static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2787{
2788 seq_cft(seq)->seq_stop(seq, v);
2789}
2790
2791static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2792{
2793 struct cftype *cft = seq_cft(m);
2794 struct cgroup_subsys_state *css = seq_css(m);
2795
2796 if (cft->seq_show)
2797 return cft->seq_show(m, arg);
2798
2799 if (cft->read_u64)
2800 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2801 else if (cft->read_s64)
2802 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2803 else
2804 return -EINVAL;
2805 return 0;
2806}
2807
2808static struct kernfs_ops cgroup_kf_single_ops = {
2809 .atomic_write_len = PAGE_SIZE,
2810 .write = cgroup_file_write,
2811 .seq_show = cgroup_seqfile_show,
2812};
2813
2814static struct kernfs_ops cgroup_kf_ops = {
2815 .atomic_write_len = PAGE_SIZE,
2816 .write = cgroup_file_write,
2817 .seq_start = cgroup_seqfile_start,
2818 .seq_next = cgroup_seqfile_next,
2819 .seq_stop = cgroup_seqfile_stop,
2820 .seq_show = cgroup_seqfile_show,
2821};
2822
2823/*
2824 * cgroup_rename - Only allow simple rename of directories in place.
2825 */
2826static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2827 const char *new_name_str)
2828{
2829 struct cgroup *cgrp = kn->priv;
2830 int ret;
2831
2832 if (kernfs_type(kn) != KERNFS_DIR)
2833 return -ENOTDIR;
2834 if (kn->parent != new_parent)
2835 return -EIO;
2836
2837 /*
2838 * This isn't a proper migration and its usefulness is very
2839 * limited. Disallow if sane_behavior.
2840 */
2841 if (cgroup_sane_behavior(cgrp))
2842 return -EPERM;
2843
2844 /*
2845 * We're gonna grab cgroup_mutex which nests outside kernfs
2846 * active_ref. kernfs_rename() doesn't require active_ref
2847 * protection. Break them before grabbing cgroup_mutex.
2848 */
2849 kernfs_break_active_protection(new_parent);
2850 kernfs_break_active_protection(kn);
2851
2852 mutex_lock(&cgroup_mutex);
2853
2854 ret = kernfs_rename(kn, new_parent, new_name_str);
2855
2856 mutex_unlock(&cgroup_mutex);
2857
2858 kernfs_unbreak_active_protection(kn);
2859 kernfs_unbreak_active_protection(new_parent);
2860 return ret;
2861}
2862
2863/* set uid and gid of cgroup dirs and files to that of the creator */
2864static int cgroup_kn_set_ugid(struct kernfs_node *kn)
2865{
2866 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
2867 .ia_uid = current_fsuid(),
2868 .ia_gid = current_fsgid(), };
2869
2870 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
2871 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
2872 return 0;
2873
2874 return kernfs_setattr(kn, &iattr);
2875}
2876
2877static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2878{
2879 char name[CGROUP_FILE_NAME_MAX];
2880 struct kernfs_node *kn;
2881 struct lock_class_key *key = NULL;
2882 int ret;
2883
2884#ifdef CONFIG_DEBUG_LOCK_ALLOC
2885 key = &cft->lockdep_key;
2886#endif
2887 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2888 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2889 NULL, false, key);
2890 if (IS_ERR(kn))
2891 return PTR_ERR(kn);
2892
2893 ret = cgroup_kn_set_ugid(kn);
2894 if (ret) {
2895 kernfs_remove(kn);
2896 return ret;
2897 }
2898
2899 if (cft->seq_show == cgroup_populated_show)
2900 cgrp->populated_kn = kn;
2901 return 0;
2902}
2903
2904/**
2905 * cgroup_addrm_files - add or remove files to a cgroup directory
2906 * @cgrp: the target cgroup
2907 * @cfts: array of cftypes to be added
2908 * @is_add: whether to add or remove
2909 *
2910 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2911 * For removals, this function never fails. If addition fails, this
2912 * function doesn't remove files already added. The caller is responsible
2913 * for cleaning up.
2914 */
2915static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2916 bool is_add)
2917{
2918 struct cftype *cft;
2919 int ret;
2920
2921 lockdep_assert_held(&cgroup_mutex);
2922
2923 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2924 /* does cft->flags tell us to skip this file on @cgrp? */
2925 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2926 continue;
2927 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2928 continue;
2929 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
2930 continue;
2931 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
2932 continue;
2933
2934 if (is_add) {
2935 ret = cgroup_add_file(cgrp, cft);
2936 if (ret) {
2937 pr_warn("%s: failed to add %s, err=%d\n",
2938 __func__, cft->name, ret);
2939 return ret;
2940 }
2941 } else {
2942 cgroup_rm_file(cgrp, cft);
2943 }
2944 }
2945 return 0;
2946}
2947
2948static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2949{
2950 LIST_HEAD(pending);
2951 struct cgroup_subsys *ss = cfts[0].ss;
2952 struct cgroup *root = &ss->root->cgrp;
2953 struct cgroup_subsys_state *css;
2954 int ret = 0;
2955
2956 lockdep_assert_held(&cgroup_mutex);
2957
2958 /* add/rm files for all cgroups created before */
2959 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2960 struct cgroup *cgrp = css->cgroup;
2961
2962 if (cgroup_is_dead(cgrp))
2963 continue;
2964
2965 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2966 if (ret)
2967 break;
2968 }
2969
2970 if (is_add && !ret)
2971 kernfs_activate(root->kn);
2972 return ret;
2973}
2974
2975static void cgroup_exit_cftypes(struct cftype *cfts)
2976{
2977 struct cftype *cft;
2978
2979 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2980 /* free copy for custom atomic_write_len, see init_cftypes() */
2981 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2982 kfree(cft->kf_ops);
2983 cft->kf_ops = NULL;
2984 cft->ss = NULL;
2985 }
2986}
2987
2988static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2989{
2990 struct cftype *cft;
2991
2992 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2993 struct kernfs_ops *kf_ops;
2994
2995 WARN_ON(cft->ss || cft->kf_ops);
2996
2997 if (cft->seq_start)
2998 kf_ops = &cgroup_kf_ops;
2999 else
3000 kf_ops = &cgroup_kf_single_ops;
3001
3002 /*
3003 * Ugh... if @cft wants a custom max_write_len, we need to
3004 * make a copy of kf_ops to set its atomic_write_len.
3005 */
3006 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3007 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3008 if (!kf_ops) {
3009 cgroup_exit_cftypes(cfts);
3010 return -ENOMEM;
3011 }
3012 kf_ops->atomic_write_len = cft->max_write_len;
3013 }
3014
3015 cft->kf_ops = kf_ops;
3016 cft->ss = ss;
3017 }
3018
3019 return 0;
3020}
3021
3022static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3023{
3024 lockdep_assert_held(&cgroup_mutex);
3025
3026 if (!cfts || !cfts[0].ss)
3027 return -ENOENT;
3028
3029 list_del(&cfts->node);
3030 cgroup_apply_cftypes(cfts, false);
3031 cgroup_exit_cftypes(cfts);
3032 return 0;
3033}
3034
3035/**
3036 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3037 * @cfts: zero-length name terminated array of cftypes
3038 *
3039 * Unregister @cfts. Files described by @cfts are removed from all
3040 * existing cgroups and all future cgroups won't have them either. This
3041 * function can be called anytime whether @cfts' subsys is attached or not.
3042 *
3043 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3044 * registered.
3045 */
3046int cgroup_rm_cftypes(struct cftype *cfts)
3047{
3048 int ret;
3049
3050 mutex_lock(&cgroup_mutex);
3051 ret = cgroup_rm_cftypes_locked(cfts);
3052 mutex_unlock(&cgroup_mutex);
3053 return ret;
3054}
3055
3056/**
3057 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3058 * @ss: target cgroup subsystem
3059 * @cfts: zero-length name terminated array of cftypes
3060 *
3061 * Register @cfts to @ss. Files described by @cfts are created for all
3062 * existing cgroups to which @ss is attached and all future cgroups will
3063 * have them too. This function can be called anytime whether @ss is
3064 * attached or not.
3065 *
3066 * Returns 0 on successful registration, -errno on failure. Note that this
3067 * function currently returns 0 as long as @cfts registration is successful
3068 * even if some file creation attempts on existing cgroups fail.
3069 */
3070int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3071{
3072 int ret;
3073
3074 if (ss->disabled)
3075 return 0;
3076
3077 if (!cfts || cfts[0].name[0] == '\0')
3078 return 0;
3079
3080 ret = cgroup_init_cftypes(ss, cfts);
3081 if (ret)
3082 return ret;
3083
3084 mutex_lock(&cgroup_mutex);
3085
3086 list_add_tail(&cfts->node, &ss->cfts);
3087 ret = cgroup_apply_cftypes(cfts, true);
3088 if (ret)
3089 cgroup_rm_cftypes_locked(cfts);
3090
3091 mutex_unlock(&cgroup_mutex);
3092 return ret;
3093}
3094
3095/**
3096 * cgroup_task_count - count the number of tasks in a cgroup.
3097 * @cgrp: the cgroup in question
3098 *
3099 * Return the number of tasks in the cgroup.
3100 */
3101static int cgroup_task_count(const struct cgroup *cgrp)
3102{
3103 int count = 0;
3104 struct cgrp_cset_link *link;
3105
3106 down_read(&css_set_rwsem);
3107 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3108 count += atomic_read(&link->cset->refcount);
3109 up_read(&css_set_rwsem);
3110 return count;
3111}
3112
3113/**
3114 * css_next_child - find the next child of a given css
3115 * @pos: the current position (%NULL to initiate traversal)
3116 * @parent: css whose children to walk
3117 *
3118 * This function returns the next child of @parent and should be called
3119 * under either cgroup_mutex or RCU read lock. The only requirement is
3120 * that @parent and @pos are accessible. The next sibling is guaranteed to
3121 * be returned regardless of their states.
3122 *
3123 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3124 * css which finished ->css_online() is guaranteed to be visible in the
3125 * future iterations and will stay visible until the last reference is put.
3126 * A css which hasn't finished ->css_online() or already finished
3127 * ->css_offline() may show up during traversal. It's each subsystem's
3128 * responsibility to synchronize against on/offlining.
3129 */
3130struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3131 struct cgroup_subsys_state *parent)
3132{
3133 struct cgroup_subsys_state *next;
3134
3135 cgroup_assert_mutex_or_rcu_locked();
3136
3137 /*
3138 * @pos could already have been unlinked from the sibling list.
3139 * Once a cgroup is removed, its ->sibling.next is no longer
3140 * updated when its next sibling changes. CSS_RELEASED is set when
3141 * @pos is taken off list, at which time its next pointer is valid,
3142 * and, as releases are serialized, the one pointed to by the next
3143 * pointer is guaranteed to not have started release yet. This
3144 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3145 * critical section, the one pointed to by its next pointer is
3146 * guaranteed to not have finished its RCU grace period even if we
3147 * have dropped rcu_read_lock() inbetween iterations.
3148 *
3149 * If @pos has CSS_RELEASED set, its next pointer can't be
3150 * dereferenced; however, as each css is given a monotonically
3151 * increasing unique serial number and always appended to the
3152 * sibling list, the next one can be found by walking the parent's
3153 * children until the first css with higher serial number than
3154 * @pos's. While this path can be slower, it happens iff iteration
3155 * races against release and the race window is very small.
3156 */
3157 if (!pos) {
3158 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3159 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3160 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3161 } else {
3162 list_for_each_entry_rcu(next, &parent->children, sibling)
3163 if (next->serial_nr > pos->serial_nr)
3164 break;
3165 }
3166
3167 /*
3168 * @next, if not pointing to the head, can be dereferenced and is
3169 * the next sibling.
3170 */
3171 if (&next->sibling != &parent->children)
3172 return next;
3173 return NULL;
3174}
3175
3176/**
3177 * css_next_descendant_pre - find the next descendant for pre-order walk
3178 * @pos: the current position (%NULL to initiate traversal)
3179 * @root: css whose descendants to walk
3180 *
3181 * To be used by css_for_each_descendant_pre(). Find the next descendant
3182 * to visit for pre-order traversal of @root's descendants. @root is
3183 * included in the iteration and the first node to be visited.
3184 *
3185 * While this function requires cgroup_mutex or RCU read locking, it
3186 * doesn't require the whole traversal to be contained in a single critical
3187 * section. This function will return the correct next descendant as long
3188 * as both @pos and @root are accessible and @pos is a descendant of @root.
3189 *
3190 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3191 * css which finished ->css_online() is guaranteed to be visible in the
3192 * future iterations and will stay visible until the last reference is put.
3193 * A css which hasn't finished ->css_online() or already finished
3194 * ->css_offline() may show up during traversal. It's each subsystem's
3195 * responsibility to synchronize against on/offlining.
3196 */
3197struct cgroup_subsys_state *
3198css_next_descendant_pre(struct cgroup_subsys_state *pos,
3199 struct cgroup_subsys_state *root)
3200{
3201 struct cgroup_subsys_state *next;
3202
3203 cgroup_assert_mutex_or_rcu_locked();
3204
3205 /* if first iteration, visit @root */
3206 if (!pos)
3207 return root;
3208
3209 /* visit the first child if exists */
3210 next = css_next_child(NULL, pos);
3211 if (next)
3212 return next;
3213
3214 /* no child, visit my or the closest ancestor's next sibling */
3215 while (pos != root) {
3216 next = css_next_child(pos, pos->parent);
3217 if (next)
3218 return next;
3219 pos = pos->parent;
3220 }
3221
3222 return NULL;
3223}
3224
3225/**
3226 * css_rightmost_descendant - return the rightmost descendant of a css
3227 * @pos: css of interest
3228 *
3229 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3230 * is returned. This can be used during pre-order traversal to skip
3231 * subtree of @pos.
3232 *
3233 * While this function requires cgroup_mutex or RCU read locking, it
3234 * doesn't require the whole traversal to be contained in a single critical
3235 * section. This function will return the correct rightmost descendant as
3236 * long as @pos is accessible.
3237 */
3238struct cgroup_subsys_state *
3239css_rightmost_descendant(struct cgroup_subsys_state *pos)
3240{
3241 struct cgroup_subsys_state *last, *tmp;
3242
3243 cgroup_assert_mutex_or_rcu_locked();
3244
3245 do {
3246 last = pos;
3247 /* ->prev isn't RCU safe, walk ->next till the end */
3248 pos = NULL;
3249 css_for_each_child(tmp, last)
3250 pos = tmp;
3251 } while (pos);
3252
3253 return last;
3254}
3255
3256static struct cgroup_subsys_state *
3257css_leftmost_descendant(struct cgroup_subsys_state *pos)
3258{
3259 struct cgroup_subsys_state *last;
3260
3261 do {
3262 last = pos;
3263 pos = css_next_child(NULL, pos);
3264 } while (pos);
3265
3266 return last;
3267}
3268
3269/**
3270 * css_next_descendant_post - find the next descendant for post-order walk
3271 * @pos: the current position (%NULL to initiate traversal)
3272 * @root: css whose descendants to walk
3273 *
3274 * To be used by css_for_each_descendant_post(). Find the next descendant
3275 * to visit for post-order traversal of @root's descendants. @root is
3276 * included in the iteration and the last node to be visited.
3277 *
3278 * While this function requires cgroup_mutex or RCU read locking, it
3279 * doesn't require the whole traversal to be contained in a single critical
3280 * section. This function will return the correct next descendant as long
3281 * as both @pos and @cgroup are accessible and @pos is a descendant of
3282 * @cgroup.
3283 *
3284 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3285 * css which finished ->css_online() is guaranteed to be visible in the
3286 * future iterations and will stay visible until the last reference is put.
3287 * A css which hasn't finished ->css_online() or already finished
3288 * ->css_offline() may show up during traversal. It's each subsystem's
3289 * responsibility to synchronize against on/offlining.
3290 */
3291struct cgroup_subsys_state *
3292css_next_descendant_post(struct cgroup_subsys_state *pos,
3293 struct cgroup_subsys_state *root)
3294{
3295 struct cgroup_subsys_state *next;
3296
3297 cgroup_assert_mutex_or_rcu_locked();
3298
3299 /* if first iteration, visit leftmost descendant which may be @root */
3300 if (!pos)
3301 return css_leftmost_descendant(root);
3302
3303 /* if we visited @root, we're done */
3304 if (pos == root)
3305 return NULL;
3306
3307 /* if there's an unvisited sibling, visit its leftmost descendant */
3308 next = css_next_child(pos, pos->parent);
3309 if (next)
3310 return css_leftmost_descendant(next);
3311
3312 /* no sibling left, visit parent */
3313 return pos->parent;
3314}
3315
3316/**
3317 * css_has_online_children - does a css have online children
3318 * @css: the target css
3319 *
3320 * Returns %true if @css has any online children; otherwise, %false. This
3321 * function can be called from any context but the caller is responsible
3322 * for synchronizing against on/offlining as necessary.
3323 */
3324bool css_has_online_children(struct cgroup_subsys_state *css)
3325{
3326 struct cgroup_subsys_state *child;
3327 bool ret = false;
3328
3329 rcu_read_lock();
3330 css_for_each_child(child, css) {
3331 if (css->flags & CSS_ONLINE) {
3332 ret = true;
3333 break;
3334 }
3335 }
3336 rcu_read_unlock();
3337 return ret;
3338}
3339
3340/**
3341 * css_advance_task_iter - advance a task itererator to the next css_set
3342 * @it: the iterator to advance
3343 *
3344 * Advance @it to the next css_set to walk.
3345 */
3346static void css_advance_task_iter(struct css_task_iter *it)
3347{
3348 struct list_head *l = it->cset_pos;
3349 struct cgrp_cset_link *link;
3350 struct css_set *cset;
3351
3352 /* Advance to the next non-empty css_set */
3353 do {
3354 l = l->next;
3355 if (l == it->cset_head) {
3356 it->cset_pos = NULL;
3357 return;
3358 }
3359
3360 if (it->ss) {
3361 cset = container_of(l, struct css_set,
3362 e_cset_node[it->ss->id]);
3363 } else {
3364 link = list_entry(l, struct cgrp_cset_link, cset_link);
3365 cset = link->cset;
3366 }
3367 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3368
3369 it->cset_pos = l;
3370
3371 if (!list_empty(&cset->tasks))
3372 it->task_pos = cset->tasks.next;
3373 else
3374 it->task_pos = cset->mg_tasks.next;
3375
3376 it->tasks_head = &cset->tasks;
3377 it->mg_tasks_head = &cset->mg_tasks;
3378}
3379
3380/**
3381 * css_task_iter_start - initiate task iteration
3382 * @css: the css to walk tasks of
3383 * @it: the task iterator to use
3384 *
3385 * Initiate iteration through the tasks of @css. The caller can call
3386 * css_task_iter_next() to walk through the tasks until the function
3387 * returns NULL. On completion of iteration, css_task_iter_end() must be
3388 * called.
3389 *
3390 * Note that this function acquires a lock which is released when the
3391 * iteration finishes. The caller can't sleep while iteration is in
3392 * progress.
3393 */
3394void css_task_iter_start(struct cgroup_subsys_state *css,
3395 struct css_task_iter *it)
3396 __acquires(css_set_rwsem)
3397{
3398 /* no one should try to iterate before mounting cgroups */
3399 WARN_ON_ONCE(!use_task_css_set_links);
3400
3401 down_read(&css_set_rwsem);
3402
3403 it->ss = css->ss;
3404
3405 if (it->ss)
3406 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3407 else
3408 it->cset_pos = &css->cgroup->cset_links;
3409
3410 it->cset_head = it->cset_pos;
3411
3412 css_advance_task_iter(it);
3413}
3414
3415/**
3416 * css_task_iter_next - return the next task for the iterator
3417 * @it: the task iterator being iterated
3418 *
3419 * The "next" function for task iteration. @it should have been
3420 * initialized via css_task_iter_start(). Returns NULL when the iteration
3421 * reaches the end.
3422 */
3423struct task_struct *css_task_iter_next(struct css_task_iter *it)
3424{
3425 struct task_struct *res;
3426 struct list_head *l = it->task_pos;
3427
3428 /* If the iterator cg is NULL, we have no tasks */
3429 if (!it->cset_pos)
3430 return NULL;
3431 res = list_entry(l, struct task_struct, cg_list);
3432
3433 /*
3434 * Advance iterator to find next entry. cset->tasks is consumed
3435 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3436 * next cset.
3437 */
3438 l = l->next;
3439
3440 if (l == it->tasks_head)
3441 l = it->mg_tasks_head->next;
3442
3443 if (l == it->mg_tasks_head)
3444 css_advance_task_iter(it);
3445 else
3446 it->task_pos = l;
3447
3448 return res;
3449}
3450
3451/**
3452 * css_task_iter_end - finish task iteration
3453 * @it: the task iterator to finish
3454 *
3455 * Finish task iteration started by css_task_iter_start().
3456 */
3457void css_task_iter_end(struct css_task_iter *it)
3458 __releases(css_set_rwsem)
3459{
3460 up_read(&css_set_rwsem);
3461}
3462
3463/**
3464 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3465 * @to: cgroup to which the tasks will be moved
3466 * @from: cgroup in which the tasks currently reside
3467 *
3468 * Locking rules between cgroup_post_fork() and the migration path
3469 * guarantee that, if a task is forking while being migrated, the new child
3470 * is guaranteed to be either visible in the source cgroup after the
3471 * parent's migration is complete or put into the target cgroup. No task
3472 * can slip out of migration through forking.
3473 */
3474int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3475{
3476 LIST_HEAD(preloaded_csets);
3477 struct cgrp_cset_link *link;
3478 struct css_task_iter it;
3479 struct task_struct *task;
3480 int ret;
3481
3482 mutex_lock(&cgroup_mutex);
3483
3484 /* all tasks in @from are being moved, all csets are source */
3485 down_read(&css_set_rwsem);
3486 list_for_each_entry(link, &from->cset_links, cset_link)
3487 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3488 up_read(&css_set_rwsem);
3489
3490 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3491 if (ret)
3492 goto out_err;
3493
3494 /*
3495 * Migrate tasks one-by-one until @form is empty. This fails iff
3496 * ->can_attach() fails.
3497 */
3498 do {
3499 css_task_iter_start(&from->self, &it);
3500 task = css_task_iter_next(&it);
3501 if (task)
3502 get_task_struct(task);
3503 css_task_iter_end(&it);
3504
3505 if (task) {
3506 ret = cgroup_migrate(to, task, false);
3507 put_task_struct(task);
3508 }
3509 } while (task && !ret);
3510out_err:
3511 cgroup_migrate_finish(&preloaded_csets);
3512 mutex_unlock(&cgroup_mutex);
3513 return ret;
3514}
3515
3516/*
3517 * Stuff for reading the 'tasks'/'procs' files.
3518 *
3519 * Reading this file can return large amounts of data if a cgroup has
3520 * *lots* of attached tasks. So it may need several calls to read(),
3521 * but we cannot guarantee that the information we produce is correct
3522 * unless we produce it entirely atomically.
3523 *
3524 */
3525
3526/* which pidlist file are we talking about? */
3527enum cgroup_filetype {
3528 CGROUP_FILE_PROCS,
3529 CGROUP_FILE_TASKS,
3530};
3531
3532/*
3533 * A pidlist is a list of pids that virtually represents the contents of one
3534 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3535 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3536 * to the cgroup.
3537 */
3538struct cgroup_pidlist {
3539 /*
3540 * used to find which pidlist is wanted. doesn't change as long as
3541 * this particular list stays in the list.
3542 */
3543 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3544 /* array of xids */
3545 pid_t *list;
3546 /* how many elements the above list has */
3547 int length;
3548 /* each of these stored in a list by its cgroup */
3549 struct list_head links;
3550 /* pointer to the cgroup we belong to, for list removal purposes */
3551 struct cgroup *owner;
3552 /* for delayed destruction */
3553 struct delayed_work destroy_dwork;
3554};
3555
3556/*
3557 * The following two functions "fix" the issue where there are more pids
3558 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3559 * TODO: replace with a kernel-wide solution to this problem
3560 */
3561#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3562static void *pidlist_allocate(int count)
3563{
3564 if (PIDLIST_TOO_LARGE(count))
3565 return vmalloc(count * sizeof(pid_t));
3566 else
3567 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3568}
3569
3570static void pidlist_free(void *p)
3571{
3572 if (is_vmalloc_addr(p))
3573 vfree(p);
3574 else
3575 kfree(p);
3576}
3577
3578/*
3579 * Used to destroy all pidlists lingering waiting for destroy timer. None
3580 * should be left afterwards.
3581 */
3582static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3583{
3584 struct cgroup_pidlist *l, *tmp_l;
3585
3586 mutex_lock(&cgrp->pidlist_mutex);
3587 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3588 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3589 mutex_unlock(&cgrp->pidlist_mutex);
3590
3591 flush_workqueue(cgroup_pidlist_destroy_wq);
3592 BUG_ON(!list_empty(&cgrp->pidlists));
3593}
3594
3595static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3596{
3597 struct delayed_work *dwork = to_delayed_work(work);
3598 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3599 destroy_dwork);
3600 struct cgroup_pidlist *tofree = NULL;
3601
3602 mutex_lock(&l->owner->pidlist_mutex);
3603
3604 /*
3605 * Destroy iff we didn't get queued again. The state won't change
3606 * as destroy_dwork can only be queued while locked.
3607 */
3608 if (!delayed_work_pending(dwork)) {
3609 list_del(&l->links);
3610 pidlist_free(l->list);
3611 put_pid_ns(l->key.ns);
3612 tofree = l;
3613 }
3614
3615 mutex_unlock(&l->owner->pidlist_mutex);
3616 kfree(tofree);
3617}
3618
3619/*
3620 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3621 * Returns the number of unique elements.
3622 */
3623static int pidlist_uniq(pid_t *list, int length)
3624{
3625 int src, dest = 1;
3626
3627 /*
3628 * we presume the 0th element is unique, so i starts at 1. trivial
3629 * edge cases first; no work needs to be done for either
3630 */
3631 if (length == 0 || length == 1)
3632 return length;
3633 /* src and dest walk down the list; dest counts unique elements */
3634 for (src = 1; src < length; src++) {
3635 /* find next unique element */
3636 while (list[src] == list[src-1]) {
3637 src++;
3638 if (src == length)
3639 goto after;
3640 }
3641 /* dest always points to where the next unique element goes */
3642 list[dest] = list[src];
3643 dest++;
3644 }
3645after:
3646 return dest;
3647}
3648
3649/*
3650 * The two pid files - task and cgroup.procs - guaranteed that the result
3651 * is sorted, which forced this whole pidlist fiasco. As pid order is
3652 * different per namespace, each namespace needs differently sorted list,
3653 * making it impossible to use, for example, single rbtree of member tasks
3654 * sorted by task pointer. As pidlists can be fairly large, allocating one
3655 * per open file is dangerous, so cgroup had to implement shared pool of
3656 * pidlists keyed by cgroup and namespace.
3657 *
3658 * All this extra complexity was caused by the original implementation
3659 * committing to an entirely unnecessary property. In the long term, we
3660 * want to do away with it. Explicitly scramble sort order if
3661 * sane_behavior so that no such expectation exists in the new interface.
3662 *
3663 * Scrambling is done by swapping every two consecutive bits, which is
3664 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3665 */
3666static pid_t pid_fry(pid_t pid)
3667{
3668 unsigned a = pid & 0x55555555;
3669 unsigned b = pid & 0xAAAAAAAA;
3670
3671 return (a << 1) | (b >> 1);
3672}
3673
3674static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3675{
3676 if (cgroup_sane_behavior(cgrp))
3677 return pid_fry(pid);
3678 else
3679 return pid;
3680}
3681
3682static int cmppid(const void *a, const void *b)
3683{
3684 return *(pid_t *)a - *(pid_t *)b;
3685}
3686
3687static int fried_cmppid(const void *a, const void *b)
3688{
3689 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3690}
3691
3692static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3693 enum cgroup_filetype type)
3694{
3695 struct cgroup_pidlist *l;
3696 /* don't need task_nsproxy() if we're looking at ourself */
3697 struct pid_namespace *ns = task_active_pid_ns(current);
3698
3699 lockdep_assert_held(&cgrp->pidlist_mutex);
3700
3701 list_for_each_entry(l, &cgrp->pidlists, links)
3702 if (l->key.type == type && l->key.ns == ns)
3703 return l;
3704 return NULL;
3705}
3706
3707/*
3708 * find the appropriate pidlist for our purpose (given procs vs tasks)
3709 * returns with the lock on that pidlist already held, and takes care
3710 * of the use count, or returns NULL with no locks held if we're out of
3711 * memory.
3712 */
3713static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3714 enum cgroup_filetype type)
3715{
3716 struct cgroup_pidlist *l;
3717
3718 lockdep_assert_held(&cgrp->pidlist_mutex);
3719
3720 l = cgroup_pidlist_find(cgrp, type);
3721 if (l)
3722 return l;
3723
3724 /* entry not found; create a new one */
3725 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3726 if (!l)
3727 return l;
3728
3729 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3730 l->key.type = type;
3731 /* don't need task_nsproxy() if we're looking at ourself */
3732 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3733 l->owner = cgrp;
3734 list_add(&l->links, &cgrp->pidlists);
3735 return l;
3736}
3737
3738/*
3739 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3740 */
3741static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3742 struct cgroup_pidlist **lp)
3743{
3744 pid_t *array;
3745 int length;
3746 int pid, n = 0; /* used for populating the array */
3747 struct css_task_iter it;
3748 struct task_struct *tsk;
3749 struct cgroup_pidlist *l;
3750
3751 lockdep_assert_held(&cgrp->pidlist_mutex);
3752
3753 /*
3754 * If cgroup gets more users after we read count, we won't have
3755 * enough space - tough. This race is indistinguishable to the
3756 * caller from the case that the additional cgroup users didn't
3757 * show up until sometime later on.
3758 */
3759 length = cgroup_task_count(cgrp);
3760 array = pidlist_allocate(length);
3761 if (!array)
3762 return -ENOMEM;
3763 /* now, populate the array */
3764 css_task_iter_start(&cgrp->self, &it);
3765 while ((tsk = css_task_iter_next(&it))) {
3766 if (unlikely(n == length))
3767 break;
3768 /* get tgid or pid for procs or tasks file respectively */
3769 if (type == CGROUP_FILE_PROCS)
3770 pid = task_tgid_vnr(tsk);
3771 else
3772 pid = task_pid_vnr(tsk);
3773 if (pid > 0) /* make sure to only use valid results */
3774 array[n++] = pid;
3775 }
3776 css_task_iter_end(&it);
3777 length = n;
3778 /* now sort & (if procs) strip out duplicates */
3779 if (cgroup_sane_behavior(cgrp))
3780 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3781 else
3782 sort(array, length, sizeof(pid_t), cmppid, NULL);
3783 if (type == CGROUP_FILE_PROCS)
3784 length = pidlist_uniq(array, length);
3785
3786 l = cgroup_pidlist_find_create(cgrp, type);
3787 if (!l) {
3788 mutex_unlock(&cgrp->pidlist_mutex);
3789 pidlist_free(array);
3790 return -ENOMEM;
3791 }
3792
3793 /* store array, freeing old if necessary */
3794 pidlist_free(l->list);
3795 l->list = array;
3796 l->length = length;
3797 *lp = l;
3798 return 0;
3799}
3800
3801/**
3802 * cgroupstats_build - build and fill cgroupstats
3803 * @stats: cgroupstats to fill information into
3804 * @dentry: A dentry entry belonging to the cgroup for which stats have
3805 * been requested.
3806 *
3807 * Build and fill cgroupstats so that taskstats can export it to user
3808 * space.
3809 */
3810int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3811{
3812 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3813 struct cgroup *cgrp;
3814 struct css_task_iter it;
3815 struct task_struct *tsk;
3816
3817 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3818 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3819 kernfs_type(kn) != KERNFS_DIR)
3820 return -EINVAL;
3821
3822 mutex_lock(&cgroup_mutex);
3823
3824 /*
3825 * We aren't being called from kernfs and there's no guarantee on
3826 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
3827 * @kn->priv is RCU safe. Let's do the RCU dancing.
3828 */
3829 rcu_read_lock();
3830 cgrp = rcu_dereference(kn->priv);
3831 if (!cgrp || cgroup_is_dead(cgrp)) {
3832 rcu_read_unlock();
3833 mutex_unlock(&cgroup_mutex);
3834 return -ENOENT;
3835 }
3836 rcu_read_unlock();
3837
3838 css_task_iter_start(&cgrp->self, &it);
3839 while ((tsk = css_task_iter_next(&it))) {
3840 switch (tsk->state) {
3841 case TASK_RUNNING:
3842 stats->nr_running++;
3843 break;
3844 case TASK_INTERRUPTIBLE:
3845 stats->nr_sleeping++;
3846 break;
3847 case TASK_UNINTERRUPTIBLE:
3848 stats->nr_uninterruptible++;
3849 break;
3850 case TASK_STOPPED:
3851 stats->nr_stopped++;
3852 break;
3853 default:
3854 if (delayacct_is_task_waiting_on_io(tsk))
3855 stats->nr_io_wait++;
3856 break;
3857 }
3858 }
3859 css_task_iter_end(&it);
3860
3861 mutex_unlock(&cgroup_mutex);
3862 return 0;
3863}
3864
3865
3866/*
3867 * seq_file methods for the tasks/procs files. The seq_file position is the
3868 * next pid to display; the seq_file iterator is a pointer to the pid
3869 * in the cgroup->l->list array.
3870 */
3871
3872static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3873{
3874 /*
3875 * Initially we receive a position value that corresponds to
3876 * one more than the last pid shown (or 0 on the first call or
3877 * after a seek to the start). Use a binary-search to find the
3878 * next pid to display, if any
3879 */
3880 struct kernfs_open_file *of = s->private;
3881 struct cgroup *cgrp = seq_css(s)->cgroup;
3882 struct cgroup_pidlist *l;
3883 enum cgroup_filetype type = seq_cft(s)->private;
3884 int index = 0, pid = *pos;
3885 int *iter, ret;
3886
3887 mutex_lock(&cgrp->pidlist_mutex);
3888
3889 /*
3890 * !NULL @of->priv indicates that this isn't the first start()
3891 * after open. If the matching pidlist is around, we can use that.
3892 * Look for it. Note that @of->priv can't be used directly. It
3893 * could already have been destroyed.
3894 */
3895 if (of->priv)
3896 of->priv = cgroup_pidlist_find(cgrp, type);
3897
3898 /*
3899 * Either this is the first start() after open or the matching
3900 * pidlist has been destroyed inbetween. Create a new one.
3901 */
3902 if (!of->priv) {
3903 ret = pidlist_array_load(cgrp, type,
3904 (struct cgroup_pidlist **)&of->priv);
3905 if (ret)
3906 return ERR_PTR(ret);
3907 }
3908 l = of->priv;
3909
3910 if (pid) {
3911 int end = l->length;
3912
3913 while (index < end) {
3914 int mid = (index + end) / 2;
3915 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3916 index = mid;
3917 break;
3918 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3919 index = mid + 1;
3920 else
3921 end = mid;
3922 }
3923 }
3924 /* If we're off the end of the array, we're done */
3925 if (index >= l->length)
3926 return NULL;
3927 /* Update the abstract position to be the actual pid that we found */
3928 iter = l->list + index;
3929 *pos = cgroup_pid_fry(cgrp, *iter);
3930 return iter;
3931}
3932
3933static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3934{
3935 struct kernfs_open_file *of = s->private;
3936 struct cgroup_pidlist *l = of->priv;
3937
3938 if (l)
3939 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3940 CGROUP_PIDLIST_DESTROY_DELAY);
3941 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3942}
3943
3944static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3945{
3946 struct kernfs_open_file *of = s->private;
3947 struct cgroup_pidlist *l = of->priv;
3948 pid_t *p = v;
3949 pid_t *end = l->list + l->length;
3950 /*
3951 * Advance to the next pid in the array. If this goes off the
3952 * end, we're done
3953 */
3954 p++;
3955 if (p >= end) {
3956 return NULL;
3957 } else {
3958 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3959 return p;
3960 }
3961}
3962
3963static int cgroup_pidlist_show(struct seq_file *s, void *v)
3964{
3965 return seq_printf(s, "%d\n", *(int *)v);
3966}
3967
3968static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3969 struct cftype *cft)
3970{
3971 return notify_on_release(css->cgroup);
3972}
3973
3974static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3975 struct cftype *cft, u64 val)
3976{
3977 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3978 if (val)
3979 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3980 else
3981 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3982 return 0;
3983}
3984
3985static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3986 struct cftype *cft)
3987{
3988 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3989}
3990
3991static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3992 struct cftype *cft, u64 val)
3993{
3994 if (val)
3995 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3996 else
3997 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3998 return 0;
3999}
4000
4001static struct cftype cgroup_base_files[] = {
4002 {
4003 .name = "cgroup.procs",
4004 .seq_start = cgroup_pidlist_start,
4005 .seq_next = cgroup_pidlist_next,
4006 .seq_stop = cgroup_pidlist_stop,
4007 .seq_show = cgroup_pidlist_show,
4008 .private = CGROUP_FILE_PROCS,
4009 .write = cgroup_procs_write,
4010 .mode = S_IRUGO | S_IWUSR,
4011 },
4012 {
4013 .name = "cgroup.clone_children",
4014 .flags = CFTYPE_INSANE,
4015 .read_u64 = cgroup_clone_children_read,
4016 .write_u64 = cgroup_clone_children_write,
4017 },
4018 {
4019 .name = "cgroup.sane_behavior",
4020 .flags = CFTYPE_ONLY_ON_ROOT,
4021 .seq_show = cgroup_sane_behavior_show,
4022 },
4023 {
4024 .name = "cgroup.controllers",
4025 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_ONLY_ON_ROOT,
4026 .seq_show = cgroup_root_controllers_show,
4027 },
4028 {
4029 .name = "cgroup.controllers",
4030 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4031 .seq_show = cgroup_controllers_show,
4032 },
4033 {
4034 .name = "cgroup.subtree_control",
4035 .flags = CFTYPE_ONLY_ON_DFL,
4036 .seq_show = cgroup_subtree_control_show,
4037 .write = cgroup_subtree_control_write,
4038 },
4039 {
4040 .name = "cgroup.populated",
4041 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4042 .seq_show = cgroup_populated_show,
4043 },
4044
4045 /*
4046 * Historical crazy stuff. These don't have "cgroup." prefix and
4047 * don't exist if sane_behavior. If you're depending on these, be
4048 * prepared to be burned.
4049 */
4050 {
4051 .name = "tasks",
4052 .flags = CFTYPE_INSANE, /* use "procs" instead */
4053 .seq_start = cgroup_pidlist_start,
4054 .seq_next = cgroup_pidlist_next,
4055 .seq_stop = cgroup_pidlist_stop,
4056 .seq_show = cgroup_pidlist_show,
4057 .private = CGROUP_FILE_TASKS,
4058 .write = cgroup_tasks_write,
4059 .mode = S_IRUGO | S_IWUSR,
4060 },
4061 {
4062 .name = "notify_on_release",
4063 .flags = CFTYPE_INSANE,
4064 .read_u64 = cgroup_read_notify_on_release,
4065 .write_u64 = cgroup_write_notify_on_release,
4066 },
4067 {
4068 .name = "release_agent",
4069 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4070 .seq_show = cgroup_release_agent_show,
4071 .write = cgroup_release_agent_write,
4072 .max_write_len = PATH_MAX - 1,
4073 },
4074 { } /* terminate */
4075};
4076
4077/**
4078 * cgroup_populate_dir - create subsys files in a cgroup directory
4079 * @cgrp: target cgroup
4080 * @subsys_mask: mask of the subsystem ids whose files should be added
4081 *
4082 * On failure, no file is added.
4083 */
4084static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4085{
4086 struct cgroup_subsys *ss;
4087 int i, ret = 0;
4088
4089 /* process cftsets of each subsystem */
4090 for_each_subsys(ss, i) {
4091 struct cftype *cfts;
4092
4093 if (!(subsys_mask & (1 << i)))
4094 continue;
4095
4096 list_for_each_entry(cfts, &ss->cfts, node) {
4097 ret = cgroup_addrm_files(cgrp, cfts, true);
4098 if (ret < 0)
4099 goto err;
4100 }
4101 }
4102 return 0;
4103err:
4104 cgroup_clear_dir(cgrp, subsys_mask);
4105 return ret;
4106}
4107
4108/*
4109 * css destruction is four-stage process.
4110 *
4111 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4112 * Implemented in kill_css().
4113 *
4114 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4115 * and thus css_tryget_online() is guaranteed to fail, the css can be
4116 * offlined by invoking offline_css(). After offlining, the base ref is
4117 * put. Implemented in css_killed_work_fn().
4118 *
4119 * 3. When the percpu_ref reaches zero, the only possible remaining
4120 * accessors are inside RCU read sections. css_release() schedules the
4121 * RCU callback.
4122 *
4123 * 4. After the grace period, the css can be freed. Implemented in
4124 * css_free_work_fn().
4125 *
4126 * It is actually hairier because both step 2 and 4 require process context
4127 * and thus involve punting to css->destroy_work adding two additional
4128 * steps to the already complex sequence.
4129 */
4130static void css_free_work_fn(struct work_struct *work)
4131{
4132 struct cgroup_subsys_state *css =
4133 container_of(work, struct cgroup_subsys_state, destroy_work);
4134 struct cgroup *cgrp = css->cgroup;
4135
4136 percpu_ref_exit(&css->refcnt);
4137
4138 if (css->ss) {
4139 /* css free path */
4140 if (css->parent)
4141 css_put(css->parent);
4142
4143 css->ss->css_free(css);
4144 cgroup_put(cgrp);
4145 } else {
4146 /* cgroup free path */
4147 atomic_dec(&cgrp->root->nr_cgrps);
4148 cgroup_pidlist_destroy_all(cgrp);
4149
4150 if (cgroup_parent(cgrp)) {
4151 /*
4152 * We get a ref to the parent, and put the ref when
4153 * this cgroup is being freed, so it's guaranteed
4154 * that the parent won't be destroyed before its
4155 * children.
4156 */
4157 cgroup_put(cgroup_parent(cgrp));
4158 kernfs_put(cgrp->kn);
4159 kfree(cgrp);
4160 } else {
4161 /*
4162 * This is root cgroup's refcnt reaching zero,
4163 * which indicates that the root should be
4164 * released.
4165 */
4166 cgroup_destroy_root(cgrp->root);
4167 }
4168 }
4169}
4170
4171static void css_free_rcu_fn(struct rcu_head *rcu_head)
4172{
4173 struct cgroup_subsys_state *css =
4174 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4175
4176 INIT_WORK(&css->destroy_work, css_free_work_fn);
4177 queue_work(cgroup_destroy_wq, &css->destroy_work);
4178}
4179
4180static void css_release_work_fn(struct work_struct *work)
4181{
4182 struct cgroup_subsys_state *css =
4183 container_of(work, struct cgroup_subsys_state, destroy_work);
4184 struct cgroup_subsys *ss = css->ss;
4185 struct cgroup *cgrp = css->cgroup;
4186
4187 mutex_lock(&cgroup_mutex);
4188
4189 css->flags |= CSS_RELEASED;
4190 list_del_rcu(&css->sibling);
4191
4192 if (ss) {
4193 /* css release path */
4194 cgroup_idr_remove(&ss->css_idr, css->id);
4195 } else {
4196 /* cgroup release path */
4197 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4198 cgrp->id = -1;
4199 }
4200
4201 mutex_unlock(&cgroup_mutex);
4202
4203 call_rcu(&css->rcu_head, css_free_rcu_fn);
4204}
4205
4206static void css_release(struct percpu_ref *ref)
4207{
4208 struct cgroup_subsys_state *css =
4209 container_of(ref, struct cgroup_subsys_state, refcnt);
4210
4211 INIT_WORK(&css->destroy_work, css_release_work_fn);
4212 queue_work(cgroup_destroy_wq, &css->destroy_work);
4213}
4214
4215static void init_and_link_css(struct cgroup_subsys_state *css,
4216 struct cgroup_subsys *ss, struct cgroup *cgrp)
4217{
4218 lockdep_assert_held(&cgroup_mutex);
4219
4220 cgroup_get(cgrp);
4221
4222 memset(css, 0, sizeof(*css));
4223 css->cgroup = cgrp;
4224 css->ss = ss;
4225 INIT_LIST_HEAD(&css->sibling);
4226 INIT_LIST_HEAD(&css->children);
4227 css->serial_nr = css_serial_nr_next++;
4228
4229 if (cgroup_parent(cgrp)) {
4230 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4231 css_get(css->parent);
4232 }
4233
4234 BUG_ON(cgroup_css(cgrp, ss));
4235}
4236
4237/* invoke ->css_online() on a new CSS and mark it online if successful */
4238static int online_css(struct cgroup_subsys_state *css)
4239{
4240 struct cgroup_subsys *ss = css->ss;
4241 int ret = 0;
4242
4243 lockdep_assert_held(&cgroup_mutex);
4244
4245 if (ss->css_online)
4246 ret = ss->css_online(css);
4247 if (!ret) {
4248 css->flags |= CSS_ONLINE;
4249 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4250 }
4251 return ret;
4252}
4253
4254/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4255static void offline_css(struct cgroup_subsys_state *css)
4256{
4257 struct cgroup_subsys *ss = css->ss;
4258
4259 lockdep_assert_held(&cgroup_mutex);
4260
4261 if (!(css->flags & CSS_ONLINE))
4262 return;
4263
4264 if (ss->css_offline)
4265 ss->css_offline(css);
4266
4267 css->flags &= ~CSS_ONLINE;
4268 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4269
4270 wake_up_all(&css->cgroup->offline_waitq);
4271}
4272
4273/**
4274 * create_css - create a cgroup_subsys_state
4275 * @cgrp: the cgroup new css will be associated with
4276 * @ss: the subsys of new css
4277 *
4278 * Create a new css associated with @cgrp - @ss pair. On success, the new
4279 * css is online and installed in @cgrp with all interface files created.
4280 * Returns 0 on success, -errno on failure.
4281 */
4282static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4283{
4284 struct cgroup *parent = cgroup_parent(cgrp);
4285 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4286 struct cgroup_subsys_state *css;
4287 int err;
4288
4289 lockdep_assert_held(&cgroup_mutex);
4290
4291 css = ss->css_alloc(parent_css);
4292 if (IS_ERR(css))
4293 return PTR_ERR(css);
4294
4295 init_and_link_css(css, ss, cgrp);
4296
4297 err = percpu_ref_init(&css->refcnt, css_release);
4298 if (err)
4299 goto err_free_css;
4300
4301 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4302 if (err < 0)
4303 goto err_free_percpu_ref;
4304 css->id = err;
4305
4306 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4307 if (err)
4308 goto err_free_id;
4309
4310 /* @css is ready to be brought online now, make it visible */
4311 list_add_tail_rcu(&css->sibling, &parent_css->children);
4312 cgroup_idr_replace(&ss->css_idr, css, css->id);
4313
4314 err = online_css(css);
4315 if (err)
4316 goto err_list_del;
4317
4318 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4319 cgroup_parent(parent)) {
4320 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4321 current->comm, current->pid, ss->name);
4322 if (!strcmp(ss->name, "memory"))
4323 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4324 ss->warned_broken_hierarchy = true;
4325 }
4326
4327 return 0;
4328
4329err_list_del:
4330 list_del_rcu(&css->sibling);
4331 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4332err_free_id:
4333 cgroup_idr_remove(&ss->css_idr, css->id);
4334err_free_percpu_ref:
4335 percpu_ref_exit(&css->refcnt);
4336err_free_css:
4337 call_rcu(&css->rcu_head, css_free_rcu_fn);
4338 return err;
4339}
4340
4341static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4342 umode_t mode)
4343{
4344 struct cgroup *parent, *cgrp;
4345 struct cgroup_root *root;
4346 struct cgroup_subsys *ss;
4347 struct kernfs_node *kn;
4348 int ssid, ret;
4349
4350 parent = cgroup_kn_lock_live(parent_kn);
4351 if (!parent)
4352 return -ENODEV;
4353 root = parent->root;
4354
4355 /* allocate the cgroup and its ID, 0 is reserved for the root */
4356 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4357 if (!cgrp) {
4358 ret = -ENOMEM;
4359 goto out_unlock;
4360 }
4361
4362 ret = percpu_ref_init(&cgrp->self.refcnt, css_release);
4363 if (ret)
4364 goto out_free_cgrp;
4365
4366 /*
4367 * Temporarily set the pointer to NULL, so idr_find() won't return
4368 * a half-baked cgroup.
4369 */
4370 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4371 if (cgrp->id < 0) {
4372 ret = -ENOMEM;
4373 goto out_cancel_ref;
4374 }
4375
4376 init_cgroup_housekeeping(cgrp);
4377
4378 cgrp->self.parent = &parent->self;
4379 cgrp->root = root;
4380
4381 if (notify_on_release(parent))
4382 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4383
4384 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4385 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4386
4387 /* create the directory */
4388 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4389 if (IS_ERR(kn)) {
4390 ret = PTR_ERR(kn);
4391 goto out_free_id;
4392 }
4393 cgrp->kn = kn;
4394
4395 /*
4396 * This extra ref will be put in cgroup_free_fn() and guarantees
4397 * that @cgrp->kn is always accessible.
4398 */
4399 kernfs_get(kn);
4400
4401 cgrp->self.serial_nr = css_serial_nr_next++;
4402
4403 /* allocation complete, commit to creation */
4404 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4405 atomic_inc(&root->nr_cgrps);
4406 cgroup_get(parent);
4407
4408 /*
4409 * @cgrp is now fully operational. If something fails after this
4410 * point, it'll be released via the normal destruction path.
4411 */
4412 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4413
4414 ret = cgroup_kn_set_ugid(kn);
4415 if (ret)
4416 goto out_destroy;
4417
4418 ret = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4419 if (ret)
4420 goto out_destroy;
4421
4422 /* let's create and online css's */
4423 for_each_subsys(ss, ssid) {
4424 if (parent->child_subsys_mask & (1 << ssid)) {
4425 ret = create_css(cgrp, ss);
4426 if (ret)
4427 goto out_destroy;
4428 }
4429 }
4430
4431 /*
4432 * On the default hierarchy, a child doesn't automatically inherit
4433 * child_subsys_mask from the parent. Each is configured manually.
4434 */
4435 if (!cgroup_on_dfl(cgrp))
4436 cgrp->child_subsys_mask = parent->child_subsys_mask;
4437
4438 kernfs_activate(kn);
4439
4440 ret = 0;
4441 goto out_unlock;
4442
4443out_free_id:
4444 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4445out_cancel_ref:
4446 percpu_ref_exit(&cgrp->self.refcnt);
4447out_free_cgrp:
4448 kfree(cgrp);
4449out_unlock:
4450 cgroup_kn_unlock(parent_kn);
4451 return ret;
4452
4453out_destroy:
4454 cgroup_destroy_locked(cgrp);
4455 goto out_unlock;
4456}
4457
4458/*
4459 * This is called when the refcnt of a css is confirmed to be killed.
4460 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4461 * initate destruction and put the css ref from kill_css().
4462 */
4463static void css_killed_work_fn(struct work_struct *work)
4464{
4465 struct cgroup_subsys_state *css =
4466 container_of(work, struct cgroup_subsys_state, destroy_work);
4467
4468 mutex_lock(&cgroup_mutex);
4469 offline_css(css);
4470 mutex_unlock(&cgroup_mutex);
4471
4472 css_put(css);
4473}
4474
4475/* css kill confirmation processing requires process context, bounce */
4476static void css_killed_ref_fn(struct percpu_ref *ref)
4477{
4478 struct cgroup_subsys_state *css =
4479 container_of(ref, struct cgroup_subsys_state, refcnt);
4480
4481 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4482 queue_work(cgroup_destroy_wq, &css->destroy_work);
4483}
4484
4485/**
4486 * kill_css - destroy a css
4487 * @css: css to destroy
4488 *
4489 * This function initiates destruction of @css by removing cgroup interface
4490 * files and putting its base reference. ->css_offline() will be invoked
4491 * asynchronously once css_tryget_online() is guaranteed to fail and when
4492 * the reference count reaches zero, @css will be released.
4493 */
4494static void kill_css(struct cgroup_subsys_state *css)
4495{
4496 lockdep_assert_held(&cgroup_mutex);
4497
4498 /*
4499 * This must happen before css is disassociated with its cgroup.
4500 * See seq_css() for details.
4501 */
4502 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4503
4504 /*
4505 * Killing would put the base ref, but we need to keep it alive
4506 * until after ->css_offline().
4507 */
4508 css_get(css);
4509
4510 /*
4511 * cgroup core guarantees that, by the time ->css_offline() is
4512 * invoked, no new css reference will be given out via
4513 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4514 * proceed to offlining css's because percpu_ref_kill() doesn't
4515 * guarantee that the ref is seen as killed on all CPUs on return.
4516 *
4517 * Use percpu_ref_kill_and_confirm() to get notifications as each
4518 * css is confirmed to be seen as killed on all CPUs.
4519 */
4520 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4521}
4522
4523/**
4524 * cgroup_destroy_locked - the first stage of cgroup destruction
4525 * @cgrp: cgroup to be destroyed
4526 *
4527 * css's make use of percpu refcnts whose killing latency shouldn't be
4528 * exposed to userland and are RCU protected. Also, cgroup core needs to
4529 * guarantee that css_tryget_online() won't succeed by the time
4530 * ->css_offline() is invoked. To satisfy all the requirements,
4531 * destruction is implemented in the following two steps.
4532 *
4533 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4534 * userland visible parts and start killing the percpu refcnts of
4535 * css's. Set up so that the next stage will be kicked off once all
4536 * the percpu refcnts are confirmed to be killed.
4537 *
4538 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4539 * rest of destruction. Once all cgroup references are gone, the
4540 * cgroup is RCU-freed.
4541 *
4542 * This function implements s1. After this step, @cgrp is gone as far as
4543 * the userland is concerned and a new cgroup with the same name may be
4544 * created. As cgroup doesn't care about the names internally, this
4545 * doesn't cause any problem.
4546 */
4547static int cgroup_destroy_locked(struct cgroup *cgrp)
4548 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4549{
4550 struct cgroup_subsys_state *css;
4551 bool empty;
4552 int ssid;
4553
4554 lockdep_assert_held(&cgroup_mutex);
4555
4556 /*
4557 * css_set_rwsem synchronizes access to ->cset_links and prevents
4558 * @cgrp from being removed while put_css_set() is in progress.
4559 */
4560 down_read(&css_set_rwsem);
4561 empty = list_empty(&cgrp->cset_links);
4562 up_read(&css_set_rwsem);
4563 if (!empty)
4564 return -EBUSY;
4565
4566 /*
4567 * Make sure there's no live children. We can't test emptiness of
4568 * ->self.children as dead children linger on it while being
4569 * drained; otherwise, "rmdir parent/child parent" may fail.
4570 */
4571 if (css_has_online_children(&cgrp->self))
4572 return -EBUSY;
4573
4574 /*
4575 * Mark @cgrp dead. This prevents further task migration and child
4576 * creation by disabling cgroup_lock_live_group().
4577 */
4578 cgrp->self.flags &= ~CSS_ONLINE;
4579
4580 /* initiate massacre of all css's */
4581 for_each_css(css, ssid, cgrp)
4582 kill_css(css);
4583
4584 /* CSS_ONLINE is clear, remove from ->release_list for the last time */
4585 raw_spin_lock(&release_list_lock);
4586 if (!list_empty(&cgrp->release_list))
4587 list_del_init(&cgrp->release_list);
4588 raw_spin_unlock(&release_list_lock);
4589
4590 /*
4591 * Remove @cgrp directory along with the base files. @cgrp has an
4592 * extra ref on its kn.
4593 */
4594 kernfs_remove(cgrp->kn);
4595
4596 set_bit(CGRP_RELEASABLE, &cgroup_parent(cgrp)->flags);
4597 check_for_release(cgroup_parent(cgrp));
4598
4599 /* put the base reference */
4600 percpu_ref_kill(&cgrp->self.refcnt);
4601
4602 return 0;
4603};
4604
4605static int cgroup_rmdir(struct kernfs_node *kn)
4606{
4607 struct cgroup *cgrp;
4608 int ret = 0;
4609
4610 cgrp = cgroup_kn_lock_live(kn);
4611 if (!cgrp)
4612 return 0;
4613 cgroup_get(cgrp); /* for @kn->priv clearing */
4614
4615 ret = cgroup_destroy_locked(cgrp);
4616
4617 cgroup_kn_unlock(kn);
4618
4619 /*
4620 * There are two control paths which try to determine cgroup from
4621 * dentry without going through kernfs - cgroupstats_build() and
4622 * css_tryget_online_from_dir(). Those are supported by RCU
4623 * protecting clearing of cgrp->kn->priv backpointer, which should
4624 * happen after all files under it have been removed.
4625 */
4626 if (!ret)
4627 RCU_INIT_POINTER(*(void __rcu __force **)&kn->priv, NULL);
4628
4629 cgroup_put(cgrp);
4630 return ret;
4631}
4632
4633static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4634 .remount_fs = cgroup_remount,
4635 .show_options = cgroup_show_options,
4636 .mkdir = cgroup_mkdir,
4637 .rmdir = cgroup_rmdir,
4638 .rename = cgroup_rename,
4639};
4640
4641static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4642{
4643 struct cgroup_subsys_state *css;
4644
4645 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4646
4647 mutex_lock(&cgroup_mutex);
4648
4649 idr_init(&ss->css_idr);
4650 INIT_LIST_HEAD(&ss->cfts);
4651
4652 /* Create the root cgroup state for this subsystem */
4653 ss->root = &cgrp_dfl_root;
4654 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4655 /* We don't handle early failures gracefully */
4656 BUG_ON(IS_ERR(css));
4657 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4658
4659 /*
4660 * Root csses are never destroyed and we can't initialize
4661 * percpu_ref during early init. Disable refcnting.
4662 */
4663 css->flags |= CSS_NO_REF;
4664
4665 if (early) {
4666 /* allocation can't be done safely during early init */
4667 css->id = 1;
4668 } else {
4669 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4670 BUG_ON(css->id < 0);
4671 }
4672
4673 /* Update the init_css_set to contain a subsys
4674 * pointer to this state - since the subsystem is
4675 * newly registered, all tasks and hence the
4676 * init_css_set is in the subsystem's root cgroup. */
4677 init_css_set.subsys[ss->id] = css;
4678
4679 need_forkexit_callback |= ss->fork || ss->exit;
4680
4681 /* At system boot, before all subsystems have been
4682 * registered, no tasks have been forked, so we don't
4683 * need to invoke fork callbacks here. */
4684 BUG_ON(!list_empty(&init_task.tasks));
4685
4686 BUG_ON(online_css(css));
4687
4688 mutex_unlock(&cgroup_mutex);
4689}
4690
4691/**
4692 * cgroup_init_early - cgroup initialization at system boot
4693 *
4694 * Initialize cgroups at system boot, and initialize any
4695 * subsystems that request early init.
4696 */
4697int __init cgroup_init_early(void)
4698{
4699 static struct cgroup_sb_opts __initdata opts =
4700 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4701 struct cgroup_subsys *ss;
4702 int i;
4703
4704 init_cgroup_root(&cgrp_dfl_root, &opts);
4705 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4706
4707 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4708
4709 for_each_subsys(ss, i) {
4710 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4711 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4712 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4713 ss->id, ss->name);
4714 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4715 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4716
4717 ss->id = i;
4718 ss->name = cgroup_subsys_name[i];
4719
4720 if (ss->early_init)
4721 cgroup_init_subsys(ss, true);
4722 }
4723 return 0;
4724}
4725
4726/**
4727 * cgroup_init - cgroup initialization
4728 *
4729 * Register cgroup filesystem and /proc file, and initialize
4730 * any subsystems that didn't request early init.
4731 */
4732int __init cgroup_init(void)
4733{
4734 struct cgroup_subsys *ss;
4735 unsigned long key;
4736 int ssid, err;
4737
4738 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4739
4740 mutex_lock(&cgroup_mutex);
4741
4742 /* Add init_css_set to the hash table */
4743 key = css_set_hash(init_css_set.subsys);
4744 hash_add(css_set_table, &init_css_set.hlist, key);
4745
4746 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4747
4748 mutex_unlock(&cgroup_mutex);
4749
4750 for_each_subsys(ss, ssid) {
4751 if (ss->early_init) {
4752 struct cgroup_subsys_state *css =
4753 init_css_set.subsys[ss->id];
4754
4755 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4756 GFP_KERNEL);
4757 BUG_ON(css->id < 0);
4758 } else {
4759 cgroup_init_subsys(ss, false);
4760 }
4761
4762 list_add_tail(&init_css_set.e_cset_node[ssid],
4763 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4764
4765 /*
4766 * Setting dfl_root subsys_mask needs to consider the
4767 * disabled flag and cftype registration needs kmalloc,
4768 * both of which aren't available during early_init.
4769 */
4770 if (!ss->disabled) {
4771 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4772 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4773 }
4774 }
4775
4776 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4777 if (!cgroup_kobj)
4778 return -ENOMEM;
4779
4780 err = register_filesystem(&cgroup_fs_type);
4781 if (err < 0) {
4782 kobject_put(cgroup_kobj);
4783 return err;
4784 }
4785
4786 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4787 return 0;
4788}
4789
4790static int __init cgroup_wq_init(void)
4791{
4792 /*
4793 * There isn't much point in executing destruction path in
4794 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4795 * Use 1 for @max_active.
4796 *
4797 * We would prefer to do this in cgroup_init() above, but that
4798 * is called before init_workqueues(): so leave this until after.
4799 */
4800 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4801 BUG_ON(!cgroup_destroy_wq);
4802
4803 /*
4804 * Used to destroy pidlists and separate to serve as flush domain.
4805 * Cap @max_active to 1 too.
4806 */
4807 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4808 0, 1);
4809 BUG_ON(!cgroup_pidlist_destroy_wq);
4810
4811 return 0;
4812}
4813core_initcall(cgroup_wq_init);
4814
4815/*
4816 * proc_cgroup_show()
4817 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4818 * - Used for /proc/<pid>/cgroup.
4819 */
4820
4821/* TODO: Use a proper seq_file iterator */
4822int proc_cgroup_show(struct seq_file *m, void *v)
4823{
4824 struct pid *pid;
4825 struct task_struct *tsk;
4826 char *buf, *path;
4827 int retval;
4828 struct cgroup_root *root;
4829
4830 retval = -ENOMEM;
4831 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4832 if (!buf)
4833 goto out;
4834
4835 retval = -ESRCH;
4836 pid = m->private;
4837 tsk = get_pid_task(pid, PIDTYPE_PID);
4838 if (!tsk)
4839 goto out_free;
4840
4841 retval = 0;
4842
4843 mutex_lock(&cgroup_mutex);
4844 down_read(&css_set_rwsem);
4845
4846 for_each_root(root) {
4847 struct cgroup_subsys *ss;
4848 struct cgroup *cgrp;
4849 int ssid, count = 0;
4850
4851 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4852 continue;
4853
4854 seq_printf(m, "%d:", root->hierarchy_id);
4855 for_each_subsys(ss, ssid)
4856 if (root->subsys_mask & (1 << ssid))
4857 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4858 if (strlen(root->name))
4859 seq_printf(m, "%sname=%s", count ? "," : "",
4860 root->name);
4861 seq_putc(m, ':');
4862 cgrp = task_cgroup_from_root(tsk, root);
4863 path = cgroup_path(cgrp, buf, PATH_MAX);
4864 if (!path) {
4865 retval = -ENAMETOOLONG;
4866 goto out_unlock;
4867 }
4868 seq_puts(m, path);
4869 seq_putc(m, '\n');
4870 }
4871
4872out_unlock:
4873 up_read(&css_set_rwsem);
4874 mutex_unlock(&cgroup_mutex);
4875 put_task_struct(tsk);
4876out_free:
4877 kfree(buf);
4878out:
4879 return retval;
4880}
4881
4882/* Display information about each subsystem and each hierarchy */
4883static int proc_cgroupstats_show(struct seq_file *m, void *v)
4884{
4885 struct cgroup_subsys *ss;
4886 int i;
4887
4888 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4889 /*
4890 * ideally we don't want subsystems moving around while we do this.
4891 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4892 * subsys/hierarchy state.
4893 */
4894 mutex_lock(&cgroup_mutex);
4895
4896 for_each_subsys(ss, i)
4897 seq_printf(m, "%s\t%d\t%d\t%d\n",
4898 ss->name, ss->root->hierarchy_id,
4899 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4900
4901 mutex_unlock(&cgroup_mutex);
4902 return 0;
4903}
4904
4905static int cgroupstats_open(struct inode *inode, struct file *file)
4906{
4907 return single_open(file, proc_cgroupstats_show, NULL);
4908}
4909
4910static const struct file_operations proc_cgroupstats_operations = {
4911 .open = cgroupstats_open,
4912 .read = seq_read,
4913 .llseek = seq_lseek,
4914 .release = single_release,
4915};
4916
4917/**
4918 * cgroup_fork - initialize cgroup related fields during copy_process()
4919 * @child: pointer to task_struct of forking parent process.
4920 *
4921 * A task is associated with the init_css_set until cgroup_post_fork()
4922 * attaches it to the parent's css_set. Empty cg_list indicates that
4923 * @child isn't holding reference to its css_set.
4924 */
4925void cgroup_fork(struct task_struct *child)
4926{
4927 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4928 INIT_LIST_HEAD(&child->cg_list);
4929}
4930
4931/**
4932 * cgroup_post_fork - called on a new task after adding it to the task list
4933 * @child: the task in question
4934 *
4935 * Adds the task to the list running through its css_set if necessary and
4936 * call the subsystem fork() callbacks. Has to be after the task is
4937 * visible on the task list in case we race with the first call to
4938 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4939 * list.
4940 */
4941void cgroup_post_fork(struct task_struct *child)
4942{
4943 struct cgroup_subsys *ss;
4944 int i;
4945
4946 /*
4947 * This may race against cgroup_enable_task_cg_links(). As that
4948 * function sets use_task_css_set_links before grabbing
4949 * tasklist_lock and we just went through tasklist_lock to add
4950 * @child, it's guaranteed that either we see the set
4951 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4952 * @child during its iteration.
4953 *
4954 * If we won the race, @child is associated with %current's
4955 * css_set. Grabbing css_set_rwsem guarantees both that the
4956 * association is stable, and, on completion of the parent's
4957 * migration, @child is visible in the source of migration or
4958 * already in the destination cgroup. This guarantee is necessary
4959 * when implementing operations which need to migrate all tasks of
4960 * a cgroup to another.
4961 *
4962 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4963 * will remain in init_css_set. This is safe because all tasks are
4964 * in the init_css_set before cg_links is enabled and there's no
4965 * operation which transfers all tasks out of init_css_set.
4966 */
4967 if (use_task_css_set_links) {
4968 struct css_set *cset;
4969
4970 down_write(&css_set_rwsem);
4971 cset = task_css_set(current);
4972 if (list_empty(&child->cg_list)) {
4973 rcu_assign_pointer(child->cgroups, cset);
4974 list_add(&child->cg_list, &cset->tasks);
4975 get_css_set(cset);
4976 }
4977 up_write(&css_set_rwsem);
4978 }
4979
4980 /*
4981 * Call ss->fork(). This must happen after @child is linked on
4982 * css_set; otherwise, @child might change state between ->fork()
4983 * and addition to css_set.
4984 */
4985 if (need_forkexit_callback) {
4986 for_each_subsys(ss, i)
4987 if (ss->fork)
4988 ss->fork(child);
4989 }
4990}
4991
4992/**
4993 * cgroup_exit - detach cgroup from exiting task
4994 * @tsk: pointer to task_struct of exiting process
4995 *
4996 * Description: Detach cgroup from @tsk and release it.
4997 *
4998 * Note that cgroups marked notify_on_release force every task in
4999 * them to take the global cgroup_mutex mutex when exiting.
5000 * This could impact scaling on very large systems. Be reluctant to
5001 * use notify_on_release cgroups where very high task exit scaling
5002 * is required on large systems.
5003 *
5004 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5005 * call cgroup_exit() while the task is still competent to handle
5006 * notify_on_release(), then leave the task attached to the root cgroup in
5007 * each hierarchy for the remainder of its exit. No need to bother with
5008 * init_css_set refcnting. init_css_set never goes away and we can't race
5009 * with migration path - PF_EXITING is visible to migration path.
5010 */
5011void cgroup_exit(struct task_struct *tsk)
5012{
5013 struct cgroup_subsys *ss;
5014 struct css_set *cset;
5015 bool put_cset = false;
5016 int i;
5017
5018 /*
5019 * Unlink from @tsk from its css_set. As migration path can't race
5020 * with us, we can check cg_list without grabbing css_set_rwsem.
5021 */
5022 if (!list_empty(&tsk->cg_list)) {
5023 down_write(&css_set_rwsem);
5024 list_del_init(&tsk->cg_list);
5025 up_write(&css_set_rwsem);
5026 put_cset = true;
5027 }
5028
5029 /* Reassign the task to the init_css_set. */
5030 cset = task_css_set(tsk);
5031 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5032
5033 if (need_forkexit_callback) {
5034 /* see cgroup_post_fork() for details */
5035 for_each_subsys(ss, i) {
5036 if (ss->exit) {
5037 struct cgroup_subsys_state *old_css = cset->subsys[i];
5038 struct cgroup_subsys_state *css = task_css(tsk, i);
5039
5040 ss->exit(css, old_css, tsk);
5041 }
5042 }
5043 }
5044
5045 if (put_cset)
5046 put_css_set(cset, true);
5047}
5048
5049static void check_for_release(struct cgroup *cgrp)
5050{
5051 if (cgroup_is_releasable(cgrp) && list_empty(&cgrp->cset_links) &&
5052 !css_has_online_children(&cgrp->self)) {
5053 /*
5054 * Control Group is currently removeable. If it's not
5055 * already queued for a userspace notification, queue
5056 * it now
5057 */
5058 int need_schedule_work = 0;
5059
5060 raw_spin_lock(&release_list_lock);
5061 if (!cgroup_is_dead(cgrp) &&
5062 list_empty(&cgrp->release_list)) {
5063 list_add(&cgrp->release_list, &release_list);
5064 need_schedule_work = 1;
5065 }
5066 raw_spin_unlock(&release_list_lock);
5067 if (need_schedule_work)
5068 schedule_work(&release_agent_work);
5069 }
5070}
5071
5072/*
5073 * Notify userspace when a cgroup is released, by running the
5074 * configured release agent with the name of the cgroup (path
5075 * relative to the root of cgroup file system) as the argument.
5076 *
5077 * Most likely, this user command will try to rmdir this cgroup.
5078 *
5079 * This races with the possibility that some other task will be
5080 * attached to this cgroup before it is removed, or that some other
5081 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5082 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5083 * unused, and this cgroup will be reprieved from its death sentence,
5084 * to continue to serve a useful existence. Next time it's released,
5085 * we will get notified again, if it still has 'notify_on_release' set.
5086 *
5087 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5088 * means only wait until the task is successfully execve()'d. The
5089 * separate release agent task is forked by call_usermodehelper(),
5090 * then control in this thread returns here, without waiting for the
5091 * release agent task. We don't bother to wait because the caller of
5092 * this routine has no use for the exit status of the release agent
5093 * task, so no sense holding our caller up for that.
5094 */
5095static void cgroup_release_agent(struct work_struct *work)
5096{
5097 BUG_ON(work != &release_agent_work);
5098 mutex_lock(&cgroup_mutex);
5099 raw_spin_lock(&release_list_lock);
5100 while (!list_empty(&release_list)) {
5101 char *argv[3], *envp[3];
5102 int i;
5103 char *pathbuf = NULL, *agentbuf = NULL, *path;
5104 struct cgroup *cgrp = list_entry(release_list.next,
5105 struct cgroup,
5106 release_list);
5107 list_del_init(&cgrp->release_list);
5108 raw_spin_unlock(&release_list_lock);
5109 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5110 if (!pathbuf)
5111 goto continue_free;
5112 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5113 if (!path)
5114 goto continue_free;
5115 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5116 if (!agentbuf)
5117 goto continue_free;
5118
5119 i = 0;
5120 argv[i++] = agentbuf;
5121 argv[i++] = path;
5122 argv[i] = NULL;
5123
5124 i = 0;
5125 /* minimal command environment */
5126 envp[i++] = "HOME=/";
5127 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5128 envp[i] = NULL;
5129
5130 /* Drop the lock while we invoke the usermode helper,
5131 * since the exec could involve hitting disk and hence
5132 * be a slow process */
5133 mutex_unlock(&cgroup_mutex);
5134 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5135 mutex_lock(&cgroup_mutex);
5136 continue_free:
5137 kfree(pathbuf);
5138 kfree(agentbuf);
5139 raw_spin_lock(&release_list_lock);
5140 }
5141 raw_spin_unlock(&release_list_lock);
5142 mutex_unlock(&cgroup_mutex);
5143}
5144
5145static int __init cgroup_disable(char *str)
5146{
5147 struct cgroup_subsys *ss;
5148 char *token;
5149 int i;
5150
5151 while ((token = strsep(&str, ",")) != NULL) {
5152 if (!*token)
5153 continue;
5154
5155 for_each_subsys(ss, i) {
5156 if (!strcmp(token, ss->name)) {
5157 ss->disabled = 1;
5158 printk(KERN_INFO "Disabling %s control group"
5159 " subsystem\n", ss->name);
5160 break;
5161 }
5162 }
5163 }
5164 return 1;
5165}
5166__setup("cgroup_disable=", cgroup_disable);
5167
5168/**
5169 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5170 * @dentry: directory dentry of interest
5171 * @ss: subsystem of interest
5172 *
5173 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5174 * to get the corresponding css and return it. If such css doesn't exist
5175 * or can't be pinned, an ERR_PTR value is returned.
5176 */
5177struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5178 struct cgroup_subsys *ss)
5179{
5180 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5181 struct cgroup_subsys_state *css = NULL;
5182 struct cgroup *cgrp;
5183
5184 /* is @dentry a cgroup dir? */
5185 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5186 kernfs_type(kn) != KERNFS_DIR)
5187 return ERR_PTR(-EBADF);
5188
5189 rcu_read_lock();
5190
5191 /*
5192 * This path doesn't originate from kernfs and @kn could already
5193 * have been or be removed at any point. @kn->priv is RCU
5194 * protected for this access. See cgroup_rmdir() for details.
5195 */
5196 cgrp = rcu_dereference(kn->priv);
5197 if (cgrp)
5198 css = cgroup_css(cgrp, ss);
5199
5200 if (!css || !css_tryget_online(css))
5201 css = ERR_PTR(-ENOENT);
5202
5203 rcu_read_unlock();
5204 return css;
5205}
5206
5207/**
5208 * css_from_id - lookup css by id
5209 * @id: the cgroup id
5210 * @ss: cgroup subsys to be looked into
5211 *
5212 * Returns the css if there's valid one with @id, otherwise returns NULL.
5213 * Should be called under rcu_read_lock().
5214 */
5215struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5216{
5217 WARN_ON_ONCE(!rcu_read_lock_held());
5218 return idr_find(&ss->css_idr, id);
5219}
5220
5221#ifdef CONFIG_CGROUP_DEBUG
5222static struct cgroup_subsys_state *
5223debug_css_alloc(struct cgroup_subsys_state *parent_css)
5224{
5225 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5226
5227 if (!css)
5228 return ERR_PTR(-ENOMEM);
5229
5230 return css;
5231}
5232
5233static void debug_css_free(struct cgroup_subsys_state *css)
5234{
5235 kfree(css);
5236}
5237
5238static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5239 struct cftype *cft)
5240{
5241 return cgroup_task_count(css->cgroup);
5242}
5243
5244static u64 current_css_set_read(struct cgroup_subsys_state *css,
5245 struct cftype *cft)
5246{
5247 return (u64)(unsigned long)current->cgroups;
5248}
5249
5250static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5251 struct cftype *cft)
5252{
5253 u64 count;
5254
5255 rcu_read_lock();
5256 count = atomic_read(&task_css_set(current)->refcount);
5257 rcu_read_unlock();
5258 return count;
5259}
5260
5261static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5262{
5263 struct cgrp_cset_link *link;
5264 struct css_set *cset;
5265 char *name_buf;
5266
5267 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5268 if (!name_buf)
5269 return -ENOMEM;
5270
5271 down_read(&css_set_rwsem);
5272 rcu_read_lock();
5273 cset = rcu_dereference(current->cgroups);
5274 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5275 struct cgroup *c = link->cgrp;
5276
5277 cgroup_name(c, name_buf, NAME_MAX + 1);
5278 seq_printf(seq, "Root %d group %s\n",
5279 c->root->hierarchy_id, name_buf);
5280 }
5281 rcu_read_unlock();
5282 up_read(&css_set_rwsem);
5283 kfree(name_buf);
5284 return 0;
5285}
5286
5287#define MAX_TASKS_SHOWN_PER_CSS 25
5288static int cgroup_css_links_read(struct seq_file *seq, void *v)
5289{
5290 struct cgroup_subsys_state *css = seq_css(seq);
5291 struct cgrp_cset_link *link;
5292
5293 down_read(&css_set_rwsem);
5294 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5295 struct css_set *cset = link->cset;
5296 struct task_struct *task;
5297 int count = 0;
5298
5299 seq_printf(seq, "css_set %p\n", cset);
5300
5301 list_for_each_entry(task, &cset->tasks, cg_list) {
5302 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5303 goto overflow;
5304 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5305 }
5306
5307 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5308 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5309 goto overflow;
5310 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5311 }
5312 continue;
5313 overflow:
5314 seq_puts(seq, " ...\n");
5315 }
5316 up_read(&css_set_rwsem);
5317 return 0;
5318}
5319
5320static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5321{
5322 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5323}
5324
5325static struct cftype debug_files[] = {
5326 {
5327 .name = "taskcount",
5328 .read_u64 = debug_taskcount_read,
5329 },
5330
5331 {
5332 .name = "current_css_set",
5333 .read_u64 = current_css_set_read,
5334 },
5335
5336 {
5337 .name = "current_css_set_refcount",
5338 .read_u64 = current_css_set_refcount_read,
5339 },
5340
5341 {
5342 .name = "current_css_set_cg_links",
5343 .seq_show = current_css_set_cg_links_read,
5344 },
5345
5346 {
5347 .name = "cgroup_css_links",
5348 .seq_show = cgroup_css_links_read,
5349 },
5350
5351 {
5352 .name = "releasable",
5353 .read_u64 = releasable_read,
5354 },
5355
5356 { } /* terminate */
5357};
5358
5359struct cgroup_subsys debug_cgrp_subsys = {
5360 .css_alloc = debug_css_alloc,
5361 .css_free = debug_css_free,
5362 .base_cftypes = debug_files,
5363};
5364#endif /* CONFIG_CGROUP_DEBUG */