Loading...
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/percpu-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#include <linux/atomic.h>
61#include <linux/cpuset.h>
62#include <linux/proc_ns.h>
63#include <linux/nsproxy.h>
64#include <linux/file.h>
65#include <net/sock.h>
66
67#define CREATE_TRACE_POINTS
68#include <trace/events/cgroup.h>
69
70/*
71 * pidlists linger the following amount before being destroyed. The goal
72 * is avoiding frequent destruction in the middle of consecutive read calls
73 * Expiring in the middle is a performance problem not a correctness one.
74 * 1 sec should be enough.
75 */
76#define CGROUP_PIDLIST_DESTROY_DELAY HZ
77
78#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
79 MAX_CFTYPE_NAME + 2)
80
81/*
82 * cgroup_mutex is the master lock. Any modification to cgroup or its
83 * hierarchy must be performed while holding it.
84 *
85 * css_set_lock protects task->cgroups pointer, the list of css_set
86 * objects, and the chain of tasks off each css_set.
87 *
88 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
89 * cgroup.h can use them for lockdep annotations.
90 */
91#ifdef CONFIG_PROVE_RCU
92DEFINE_MUTEX(cgroup_mutex);
93DEFINE_SPINLOCK(css_set_lock);
94EXPORT_SYMBOL_GPL(cgroup_mutex);
95EXPORT_SYMBOL_GPL(css_set_lock);
96#else
97static DEFINE_MUTEX(cgroup_mutex);
98static DEFINE_SPINLOCK(css_set_lock);
99#endif
100
101/*
102 * Protects cgroup_idr and css_idr so that IDs can be released without
103 * grabbing cgroup_mutex.
104 */
105static DEFINE_SPINLOCK(cgroup_idr_lock);
106
107/*
108 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
109 * against file removal/re-creation across css hiding.
110 */
111static DEFINE_SPINLOCK(cgroup_file_kn_lock);
112
113/*
114 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
115 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
116 */
117static DEFINE_SPINLOCK(release_agent_path_lock);
118
119struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
120
121#define cgroup_assert_mutex_or_rcu_locked() \
122 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
123 !lockdep_is_held(&cgroup_mutex), \
124 "cgroup_mutex or RCU read lock required");
125
126/*
127 * cgroup destruction makes heavy use of work items and there can be a lot
128 * of concurrent destructions. Use a separate workqueue so that cgroup
129 * destruction work items don't end up filling up max_active of system_wq
130 * which may lead to deadlock.
131 */
132static struct workqueue_struct *cgroup_destroy_wq;
133
134/*
135 * pidlist destructions need to be flushed on cgroup destruction. Use a
136 * separate workqueue as flush domain.
137 */
138static struct workqueue_struct *cgroup_pidlist_destroy_wq;
139
140/* generate an array of cgroup subsystem pointers */
141#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
142static struct cgroup_subsys *cgroup_subsys[] = {
143#include <linux/cgroup_subsys.h>
144};
145#undef SUBSYS
146
147/* array of cgroup subsystem names */
148#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
149static const char *cgroup_subsys_name[] = {
150#include <linux/cgroup_subsys.h>
151};
152#undef SUBSYS
153
154/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
155#define SUBSYS(_x) \
156 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
157 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
158 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
159 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
160#include <linux/cgroup_subsys.h>
161#undef SUBSYS
162
163#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
164static struct static_key_true *cgroup_subsys_enabled_key[] = {
165#include <linux/cgroup_subsys.h>
166};
167#undef SUBSYS
168
169#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
170static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
171#include <linux/cgroup_subsys.h>
172};
173#undef SUBSYS
174
175/*
176 * The default hierarchy, reserved for the subsystems that are otherwise
177 * unattached - it never has more than a single cgroup, and all tasks are
178 * part of that cgroup.
179 */
180struct cgroup_root cgrp_dfl_root;
181EXPORT_SYMBOL_GPL(cgrp_dfl_root);
182
183/*
184 * The default hierarchy always exists but is hidden until mounted for the
185 * first time. This is for backward compatibility.
186 */
187static bool cgrp_dfl_visible;
188
189/* Controllers blocked by the commandline in v1 */
190static u16 cgroup_no_v1_mask;
191
192/* some controllers are not supported in the default hierarchy */
193static u16 cgrp_dfl_inhibit_ss_mask;
194
195/* some controllers are implicitly enabled on the default hierarchy */
196static unsigned long cgrp_dfl_implicit_ss_mask;
197
198/* The list of hierarchy roots */
199
200static LIST_HEAD(cgroup_roots);
201static int cgroup_root_count;
202
203/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
204static DEFINE_IDR(cgroup_hierarchy_idr);
205
206/*
207 * Assign a monotonically increasing serial number to csses. It guarantees
208 * cgroups with bigger numbers are newer than those with smaller numbers.
209 * Also, as csses are always appended to the parent's ->children list, it
210 * guarantees that sibling csses are always sorted in the ascending serial
211 * number order on the list. Protected by cgroup_mutex.
212 */
213static u64 css_serial_nr_next = 1;
214
215/*
216 * These bitmask flags indicate whether tasks in the fork and exit paths have
217 * fork/exit handlers to call. This avoids us having to do extra work in the
218 * fork/exit path to check which subsystems have fork/exit callbacks.
219 */
220static u16 have_fork_callback __read_mostly;
221static u16 have_exit_callback __read_mostly;
222static u16 have_free_callback __read_mostly;
223
224/* cgroup namespace for init task */
225struct cgroup_namespace init_cgroup_ns = {
226 .count = { .counter = 2, },
227 .user_ns = &init_user_ns,
228 .ns.ops = &cgroupns_operations,
229 .ns.inum = PROC_CGROUP_INIT_INO,
230 .root_cset = &init_css_set,
231};
232
233/* Ditto for the can_fork callback. */
234static u16 have_canfork_callback __read_mostly;
235
236static struct file_system_type cgroup2_fs_type;
237static struct cftype cgroup_dfl_base_files[];
238static struct cftype cgroup_legacy_base_files[];
239
240static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
241static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
242static int cgroup_apply_control(struct cgroup *cgrp);
243static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
244static void css_task_iter_advance(struct css_task_iter *it);
245static int cgroup_destroy_locked(struct cgroup *cgrp);
246static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
247 struct cgroup_subsys *ss);
248static void css_release(struct percpu_ref *ref);
249static void kill_css(struct cgroup_subsys_state *css);
250static int cgroup_addrm_files(struct cgroup_subsys_state *css,
251 struct cgroup *cgrp, struct cftype cfts[],
252 bool is_add);
253
254/**
255 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
256 * @ssid: subsys ID of interest
257 *
258 * cgroup_subsys_enabled() can only be used with literal subsys names which
259 * is fine for individual subsystems but unsuitable for cgroup core. This
260 * is slower static_key_enabled() based test indexed by @ssid.
261 */
262static bool cgroup_ssid_enabled(int ssid)
263{
264 if (CGROUP_SUBSYS_COUNT == 0)
265 return false;
266
267 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
268}
269
270static bool cgroup_ssid_no_v1(int ssid)
271{
272 return cgroup_no_v1_mask & (1 << ssid);
273}
274
275/**
276 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
277 * @cgrp: the cgroup of interest
278 *
279 * The default hierarchy is the v2 interface of cgroup and this function
280 * can be used to test whether a cgroup is on the default hierarchy for
281 * cases where a subsystem should behave differnetly depending on the
282 * interface version.
283 *
284 * The set of behaviors which change on the default hierarchy are still
285 * being determined and the mount option is prefixed with __DEVEL__.
286 *
287 * List of changed behaviors:
288 *
289 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
290 * and "name" are disallowed.
291 *
292 * - When mounting an existing superblock, mount options should match.
293 *
294 * - Remount is disallowed.
295 *
296 * - rename(2) is disallowed.
297 *
298 * - "tasks" is removed. Everything should be at process granularity. Use
299 * "cgroup.procs" instead.
300 *
301 * - "cgroup.procs" is not sorted. pids will be unique unless they got
302 * recycled inbetween reads.
303 *
304 * - "release_agent" and "notify_on_release" are removed. Replacement
305 * notification mechanism will be implemented.
306 *
307 * - "cgroup.clone_children" is removed.
308 *
309 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
310 * and its descendants contain no task; otherwise, 1. The file also
311 * generates kernfs notification which can be monitored through poll and
312 * [di]notify when the value of the file changes.
313 *
314 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
315 * take masks of ancestors with non-empty cpus/mems, instead of being
316 * moved to an ancestor.
317 *
318 * - cpuset: a task can be moved into an empty cpuset, and again it takes
319 * masks of ancestors.
320 *
321 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
322 * is not created.
323 *
324 * - blkcg: blk-throttle becomes properly hierarchical.
325 *
326 * - debug: disallowed on the default hierarchy.
327 */
328static bool cgroup_on_dfl(const struct cgroup *cgrp)
329{
330 return cgrp->root == &cgrp_dfl_root;
331}
332
333/* IDR wrappers which synchronize using cgroup_idr_lock */
334static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
335 gfp_t gfp_mask)
336{
337 int ret;
338
339 idr_preload(gfp_mask);
340 spin_lock_bh(&cgroup_idr_lock);
341 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
342 spin_unlock_bh(&cgroup_idr_lock);
343 idr_preload_end();
344 return ret;
345}
346
347static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
348{
349 void *ret;
350
351 spin_lock_bh(&cgroup_idr_lock);
352 ret = idr_replace(idr, ptr, id);
353 spin_unlock_bh(&cgroup_idr_lock);
354 return ret;
355}
356
357static void cgroup_idr_remove(struct idr *idr, int id)
358{
359 spin_lock_bh(&cgroup_idr_lock);
360 idr_remove(idr, id);
361 spin_unlock_bh(&cgroup_idr_lock);
362}
363
364static struct cgroup *cgroup_parent(struct cgroup *cgrp)
365{
366 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
367
368 if (parent_css)
369 return container_of(parent_css, struct cgroup, self);
370 return NULL;
371}
372
373/* subsystems visibly enabled on a cgroup */
374static u16 cgroup_control(struct cgroup *cgrp)
375{
376 struct cgroup *parent = cgroup_parent(cgrp);
377 u16 root_ss_mask = cgrp->root->subsys_mask;
378
379 if (parent)
380 return parent->subtree_control;
381
382 if (cgroup_on_dfl(cgrp))
383 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
384 cgrp_dfl_implicit_ss_mask);
385 return root_ss_mask;
386}
387
388/* subsystems enabled on a cgroup */
389static u16 cgroup_ss_mask(struct cgroup *cgrp)
390{
391 struct cgroup *parent = cgroup_parent(cgrp);
392
393 if (parent)
394 return parent->subtree_ss_mask;
395
396 return cgrp->root->subsys_mask;
397}
398
399/**
400 * cgroup_css - obtain a cgroup's css for the specified subsystem
401 * @cgrp: the cgroup of interest
402 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
403 *
404 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
405 * function must be called either under cgroup_mutex or rcu_read_lock() and
406 * the caller is responsible for pinning the returned css if it wants to
407 * keep accessing it outside the said locks. This function may return
408 * %NULL if @cgrp doesn't have @subsys_id enabled.
409 */
410static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
411 struct cgroup_subsys *ss)
412{
413 if (ss)
414 return rcu_dereference_check(cgrp->subsys[ss->id],
415 lockdep_is_held(&cgroup_mutex));
416 else
417 return &cgrp->self;
418}
419
420/**
421 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
422 * @cgrp: the cgroup of interest
423 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
424 *
425 * Similar to cgroup_css() but returns the effective css, which is defined
426 * as the matching css of the nearest ancestor including self which has @ss
427 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
428 * function is guaranteed to return non-NULL css.
429 */
430static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
431 struct cgroup_subsys *ss)
432{
433 lockdep_assert_held(&cgroup_mutex);
434
435 if (!ss)
436 return &cgrp->self;
437
438 /*
439 * This function is used while updating css associations and thus
440 * can't test the csses directly. Test ss_mask.
441 */
442 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
443 cgrp = cgroup_parent(cgrp);
444 if (!cgrp)
445 return NULL;
446 }
447
448 return cgroup_css(cgrp, ss);
449}
450
451/**
452 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
453 * @cgrp: the cgroup of interest
454 * @ss: the subsystem of interest
455 *
456 * Find and get the effective css of @cgrp for @ss. The effective css is
457 * defined as the matching css of the nearest ancestor including self which
458 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
459 * the root css is returned, so this function always returns a valid css.
460 * The returned css must be put using css_put().
461 */
462struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
463 struct cgroup_subsys *ss)
464{
465 struct cgroup_subsys_state *css;
466
467 rcu_read_lock();
468
469 do {
470 css = cgroup_css(cgrp, ss);
471
472 if (css && css_tryget_online(css))
473 goto out_unlock;
474 cgrp = cgroup_parent(cgrp);
475 } while (cgrp);
476
477 css = init_css_set.subsys[ss->id];
478 css_get(css);
479out_unlock:
480 rcu_read_unlock();
481 return css;
482}
483
484/* convenient tests for these bits */
485static inline bool cgroup_is_dead(const struct cgroup *cgrp)
486{
487 return !(cgrp->self.flags & CSS_ONLINE);
488}
489
490static void cgroup_get(struct cgroup *cgrp)
491{
492 WARN_ON_ONCE(cgroup_is_dead(cgrp));
493 css_get(&cgrp->self);
494}
495
496static bool cgroup_tryget(struct cgroup *cgrp)
497{
498 return css_tryget(&cgrp->self);
499}
500
501struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
502{
503 struct cgroup *cgrp = of->kn->parent->priv;
504 struct cftype *cft = of_cft(of);
505
506 /*
507 * This is open and unprotected implementation of cgroup_css().
508 * seq_css() is only called from a kernfs file operation which has
509 * an active reference on the file. Because all the subsystem
510 * files are drained before a css is disassociated with a cgroup,
511 * the matching css from the cgroup's subsys table is guaranteed to
512 * be and stay valid until the enclosing operation is complete.
513 */
514 if (cft->ss)
515 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
516 else
517 return &cgrp->self;
518}
519EXPORT_SYMBOL_GPL(of_css);
520
521static int notify_on_release(const struct cgroup *cgrp)
522{
523 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
524}
525
526/**
527 * for_each_css - iterate all css's of a cgroup
528 * @css: the iteration cursor
529 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
530 * @cgrp: the target cgroup to iterate css's of
531 *
532 * Should be called under cgroup_[tree_]mutex.
533 */
534#define for_each_css(css, ssid, cgrp) \
535 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
536 if (!((css) = rcu_dereference_check( \
537 (cgrp)->subsys[(ssid)], \
538 lockdep_is_held(&cgroup_mutex)))) { } \
539 else
540
541/**
542 * for_each_e_css - iterate all effective css's of a cgroup
543 * @css: the iteration cursor
544 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
545 * @cgrp: the target cgroup to iterate css's of
546 *
547 * Should be called under cgroup_[tree_]mutex.
548 */
549#define for_each_e_css(css, ssid, cgrp) \
550 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
551 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
552 ; \
553 else
554
555/**
556 * for_each_subsys - iterate all enabled cgroup subsystems
557 * @ss: the iteration cursor
558 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
559 */
560#define for_each_subsys(ss, ssid) \
561 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
562 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
563
564/**
565 * do_each_subsys_mask - filter for_each_subsys with a bitmask
566 * @ss: the iteration cursor
567 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
568 * @ss_mask: the bitmask
569 *
570 * The block will only run for cases where the ssid-th bit (1 << ssid) of
571 * @ss_mask is set.
572 */
573#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
574 unsigned long __ss_mask = (ss_mask); \
575 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
576 (ssid) = 0; \
577 break; \
578 } \
579 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
580 (ss) = cgroup_subsys[ssid]; \
581 {
582
583#define while_each_subsys_mask() \
584 } \
585 } \
586} while (false)
587
588/* iterate across the hierarchies */
589#define for_each_root(root) \
590 list_for_each_entry((root), &cgroup_roots, root_list)
591
592/* iterate over child cgrps, lock should be held throughout iteration */
593#define cgroup_for_each_live_child(child, cgrp) \
594 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
595 if (({ lockdep_assert_held(&cgroup_mutex); \
596 cgroup_is_dead(child); })) \
597 ; \
598 else
599
600/* walk live descendants in preorder */
601#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
602 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
603 if (({ lockdep_assert_held(&cgroup_mutex); \
604 (dsct) = (d_css)->cgroup; \
605 cgroup_is_dead(dsct); })) \
606 ; \
607 else
608
609/* walk live descendants in postorder */
610#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
611 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
612 if (({ lockdep_assert_held(&cgroup_mutex); \
613 (dsct) = (d_css)->cgroup; \
614 cgroup_is_dead(dsct); })) \
615 ; \
616 else
617
618static void cgroup_release_agent(struct work_struct *work);
619static void check_for_release(struct cgroup *cgrp);
620
621/*
622 * A cgroup can be associated with multiple css_sets as different tasks may
623 * belong to different cgroups on different hierarchies. In the other
624 * direction, a css_set is naturally associated with multiple cgroups.
625 * This M:N relationship is represented by the following link structure
626 * which exists for each association and allows traversing the associations
627 * from both sides.
628 */
629struct cgrp_cset_link {
630 /* the cgroup and css_set this link associates */
631 struct cgroup *cgrp;
632 struct css_set *cset;
633
634 /* list of cgrp_cset_links anchored at cgrp->cset_links */
635 struct list_head cset_link;
636
637 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
638 struct list_head cgrp_link;
639};
640
641/*
642 * The default css_set - used by init and its children prior to any
643 * hierarchies being mounted. It contains a pointer to the root state
644 * for each subsystem. Also used to anchor the list of css_sets. Not
645 * reference-counted, to improve performance when child cgroups
646 * haven't been created.
647 */
648struct css_set init_css_set = {
649 .refcount = ATOMIC_INIT(1),
650 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
651 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
652 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
653 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
654 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
655 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
656};
657
658static int css_set_count = 1; /* 1 for init_css_set */
659
660/**
661 * css_set_populated - does a css_set contain any tasks?
662 * @cset: target css_set
663 */
664static bool css_set_populated(struct css_set *cset)
665{
666 lockdep_assert_held(&css_set_lock);
667
668 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
669}
670
671/**
672 * cgroup_update_populated - updated populated count of a cgroup
673 * @cgrp: the target cgroup
674 * @populated: inc or dec populated count
675 *
676 * One of the css_sets associated with @cgrp is either getting its first
677 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
678 * count is propagated towards root so that a given cgroup's populated_cnt
679 * is zero iff the cgroup and all its descendants don't contain any tasks.
680 *
681 * @cgrp's interface file "cgroup.populated" is zero if
682 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
683 * changes from or to zero, userland is notified that the content of the
684 * interface file has changed. This can be used to detect when @cgrp and
685 * its descendants become populated or empty.
686 */
687static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
688{
689 lockdep_assert_held(&css_set_lock);
690
691 do {
692 bool trigger;
693
694 if (populated)
695 trigger = !cgrp->populated_cnt++;
696 else
697 trigger = !--cgrp->populated_cnt;
698
699 if (!trigger)
700 break;
701
702 check_for_release(cgrp);
703 cgroup_file_notify(&cgrp->events_file);
704
705 cgrp = cgroup_parent(cgrp);
706 } while (cgrp);
707}
708
709/**
710 * css_set_update_populated - update populated state of a css_set
711 * @cset: target css_set
712 * @populated: whether @cset is populated or depopulated
713 *
714 * @cset is either getting the first task or losing the last. Update the
715 * ->populated_cnt of all associated cgroups accordingly.
716 */
717static void css_set_update_populated(struct css_set *cset, bool populated)
718{
719 struct cgrp_cset_link *link;
720
721 lockdep_assert_held(&css_set_lock);
722
723 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
724 cgroup_update_populated(link->cgrp, populated);
725}
726
727/**
728 * css_set_move_task - move a task from one css_set to another
729 * @task: task being moved
730 * @from_cset: css_set @task currently belongs to (may be NULL)
731 * @to_cset: new css_set @task is being moved to (may be NULL)
732 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
733 *
734 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
735 * css_set, @from_cset can be NULL. If @task is being disassociated
736 * instead of moved, @to_cset can be NULL.
737 *
738 * This function automatically handles populated_cnt updates and
739 * css_task_iter adjustments but the caller is responsible for managing
740 * @from_cset and @to_cset's reference counts.
741 */
742static void css_set_move_task(struct task_struct *task,
743 struct css_set *from_cset, struct css_set *to_cset,
744 bool use_mg_tasks)
745{
746 lockdep_assert_held(&css_set_lock);
747
748 if (to_cset && !css_set_populated(to_cset))
749 css_set_update_populated(to_cset, true);
750
751 if (from_cset) {
752 struct css_task_iter *it, *pos;
753
754 WARN_ON_ONCE(list_empty(&task->cg_list));
755
756 /*
757 * @task is leaving, advance task iterators which are
758 * pointing to it so that they can resume at the next
759 * position. Advancing an iterator might remove it from
760 * the list, use safe walk. See css_task_iter_advance*()
761 * for details.
762 */
763 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
764 iters_node)
765 if (it->task_pos == &task->cg_list)
766 css_task_iter_advance(it);
767
768 list_del_init(&task->cg_list);
769 if (!css_set_populated(from_cset))
770 css_set_update_populated(from_cset, false);
771 } else {
772 WARN_ON_ONCE(!list_empty(&task->cg_list));
773 }
774
775 if (to_cset) {
776 /*
777 * We are synchronized through cgroup_threadgroup_rwsem
778 * against PF_EXITING setting such that we can't race
779 * against cgroup_exit() changing the css_set to
780 * init_css_set and dropping the old one.
781 */
782 WARN_ON_ONCE(task->flags & PF_EXITING);
783
784 rcu_assign_pointer(task->cgroups, to_cset);
785 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
786 &to_cset->tasks);
787 }
788}
789
790/*
791 * hash table for cgroup groups. This improves the performance to find
792 * an existing css_set. This hash doesn't (currently) take into
793 * account cgroups in empty hierarchies.
794 */
795#define CSS_SET_HASH_BITS 7
796static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
797
798static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
799{
800 unsigned long key = 0UL;
801 struct cgroup_subsys *ss;
802 int i;
803
804 for_each_subsys(ss, i)
805 key += (unsigned long)css[i];
806 key = (key >> 16) ^ key;
807
808 return key;
809}
810
811static void put_css_set_locked(struct css_set *cset)
812{
813 struct cgrp_cset_link *link, *tmp_link;
814 struct cgroup_subsys *ss;
815 int ssid;
816
817 lockdep_assert_held(&css_set_lock);
818
819 if (!atomic_dec_and_test(&cset->refcount))
820 return;
821
822 /* This css_set is dead. unlink it and release cgroup and css refs */
823 for_each_subsys(ss, ssid) {
824 list_del(&cset->e_cset_node[ssid]);
825 css_put(cset->subsys[ssid]);
826 }
827 hash_del(&cset->hlist);
828 css_set_count--;
829
830 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
831 list_del(&link->cset_link);
832 list_del(&link->cgrp_link);
833 if (cgroup_parent(link->cgrp))
834 cgroup_put(link->cgrp);
835 kfree(link);
836 }
837
838 kfree_rcu(cset, rcu_head);
839}
840
841static void put_css_set(struct css_set *cset)
842{
843 unsigned long flags;
844
845 /*
846 * Ensure that the refcount doesn't hit zero while any readers
847 * can see it. Similar to atomic_dec_and_lock(), but for an
848 * rwlock
849 */
850 if (atomic_add_unless(&cset->refcount, -1, 1))
851 return;
852
853 spin_lock_irqsave(&css_set_lock, flags);
854 put_css_set_locked(cset);
855 spin_unlock_irqrestore(&css_set_lock, flags);
856}
857
858/*
859 * refcounted get/put for css_set objects
860 */
861static inline void get_css_set(struct css_set *cset)
862{
863 atomic_inc(&cset->refcount);
864}
865
866/**
867 * compare_css_sets - helper function for find_existing_css_set().
868 * @cset: candidate css_set being tested
869 * @old_cset: existing css_set for a task
870 * @new_cgrp: cgroup that's being entered by the task
871 * @template: desired set of css pointers in css_set (pre-calculated)
872 *
873 * Returns true if "cset" matches "old_cset" except for the hierarchy
874 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
875 */
876static bool compare_css_sets(struct css_set *cset,
877 struct css_set *old_cset,
878 struct cgroup *new_cgrp,
879 struct cgroup_subsys_state *template[])
880{
881 struct list_head *l1, *l2;
882
883 /*
884 * On the default hierarchy, there can be csets which are
885 * associated with the same set of cgroups but different csses.
886 * Let's first ensure that csses match.
887 */
888 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
889 return false;
890
891 /*
892 * Compare cgroup pointers in order to distinguish between
893 * different cgroups in hierarchies. As different cgroups may
894 * share the same effective css, this comparison is always
895 * necessary.
896 */
897 l1 = &cset->cgrp_links;
898 l2 = &old_cset->cgrp_links;
899 while (1) {
900 struct cgrp_cset_link *link1, *link2;
901 struct cgroup *cgrp1, *cgrp2;
902
903 l1 = l1->next;
904 l2 = l2->next;
905 /* See if we reached the end - both lists are equal length. */
906 if (l1 == &cset->cgrp_links) {
907 BUG_ON(l2 != &old_cset->cgrp_links);
908 break;
909 } else {
910 BUG_ON(l2 == &old_cset->cgrp_links);
911 }
912 /* Locate the cgroups associated with these links. */
913 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
914 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
915 cgrp1 = link1->cgrp;
916 cgrp2 = link2->cgrp;
917 /* Hierarchies should be linked in the same order. */
918 BUG_ON(cgrp1->root != cgrp2->root);
919
920 /*
921 * If this hierarchy is the hierarchy of the cgroup
922 * that's changing, then we need to check that this
923 * css_set points to the new cgroup; if it's any other
924 * hierarchy, then this css_set should point to the
925 * same cgroup as the old css_set.
926 */
927 if (cgrp1->root == new_cgrp->root) {
928 if (cgrp1 != new_cgrp)
929 return false;
930 } else {
931 if (cgrp1 != cgrp2)
932 return false;
933 }
934 }
935 return true;
936}
937
938/**
939 * find_existing_css_set - init css array and find the matching css_set
940 * @old_cset: the css_set that we're using before the cgroup transition
941 * @cgrp: the cgroup that we're moving into
942 * @template: out param for the new set of csses, should be clear on entry
943 */
944static struct css_set *find_existing_css_set(struct css_set *old_cset,
945 struct cgroup *cgrp,
946 struct cgroup_subsys_state *template[])
947{
948 struct cgroup_root *root = cgrp->root;
949 struct cgroup_subsys *ss;
950 struct css_set *cset;
951 unsigned long key;
952 int i;
953
954 /*
955 * Build the set of subsystem state objects that we want to see in the
956 * new css_set. while subsystems can change globally, the entries here
957 * won't change, so no need for locking.
958 */
959 for_each_subsys(ss, i) {
960 if (root->subsys_mask & (1UL << i)) {
961 /*
962 * @ss is in this hierarchy, so we want the
963 * effective css from @cgrp.
964 */
965 template[i] = cgroup_e_css(cgrp, ss);
966 } else {
967 /*
968 * @ss is not in this hierarchy, so we don't want
969 * to change the css.
970 */
971 template[i] = old_cset->subsys[i];
972 }
973 }
974
975 key = css_set_hash(template);
976 hash_for_each_possible(css_set_table, cset, hlist, key) {
977 if (!compare_css_sets(cset, old_cset, cgrp, template))
978 continue;
979
980 /* This css_set matches what we need */
981 return cset;
982 }
983
984 /* No existing cgroup group matched */
985 return NULL;
986}
987
988static void free_cgrp_cset_links(struct list_head *links_to_free)
989{
990 struct cgrp_cset_link *link, *tmp_link;
991
992 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
993 list_del(&link->cset_link);
994 kfree(link);
995 }
996}
997
998/**
999 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1000 * @count: the number of links to allocate
1001 * @tmp_links: list_head the allocated links are put on
1002 *
1003 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1004 * through ->cset_link. Returns 0 on success or -errno.
1005 */
1006static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1007{
1008 struct cgrp_cset_link *link;
1009 int i;
1010
1011 INIT_LIST_HEAD(tmp_links);
1012
1013 for (i = 0; i < count; i++) {
1014 link = kzalloc(sizeof(*link), GFP_KERNEL);
1015 if (!link) {
1016 free_cgrp_cset_links(tmp_links);
1017 return -ENOMEM;
1018 }
1019 list_add(&link->cset_link, tmp_links);
1020 }
1021 return 0;
1022}
1023
1024/**
1025 * link_css_set - a helper function to link a css_set to a cgroup
1026 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1027 * @cset: the css_set to be linked
1028 * @cgrp: the destination cgroup
1029 */
1030static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1031 struct cgroup *cgrp)
1032{
1033 struct cgrp_cset_link *link;
1034
1035 BUG_ON(list_empty(tmp_links));
1036
1037 if (cgroup_on_dfl(cgrp))
1038 cset->dfl_cgrp = cgrp;
1039
1040 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1041 link->cset = cset;
1042 link->cgrp = cgrp;
1043
1044 /*
1045 * Always add links to the tail of the lists so that the lists are
1046 * in choronological order.
1047 */
1048 list_move_tail(&link->cset_link, &cgrp->cset_links);
1049 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1050
1051 if (cgroup_parent(cgrp))
1052 cgroup_get(cgrp);
1053}
1054
1055/**
1056 * find_css_set - return a new css_set with one cgroup updated
1057 * @old_cset: the baseline css_set
1058 * @cgrp: the cgroup to be updated
1059 *
1060 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1061 * substituted into the appropriate hierarchy.
1062 */
1063static struct css_set *find_css_set(struct css_set *old_cset,
1064 struct cgroup *cgrp)
1065{
1066 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1067 struct css_set *cset;
1068 struct list_head tmp_links;
1069 struct cgrp_cset_link *link;
1070 struct cgroup_subsys *ss;
1071 unsigned long key;
1072 int ssid;
1073
1074 lockdep_assert_held(&cgroup_mutex);
1075
1076 /* First see if we already have a cgroup group that matches
1077 * the desired set */
1078 spin_lock_irq(&css_set_lock);
1079 cset = find_existing_css_set(old_cset, cgrp, template);
1080 if (cset)
1081 get_css_set(cset);
1082 spin_unlock_irq(&css_set_lock);
1083
1084 if (cset)
1085 return cset;
1086
1087 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1088 if (!cset)
1089 return NULL;
1090
1091 /* Allocate all the cgrp_cset_link objects that we'll need */
1092 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1093 kfree(cset);
1094 return NULL;
1095 }
1096
1097 atomic_set(&cset->refcount, 1);
1098 INIT_LIST_HEAD(&cset->cgrp_links);
1099 INIT_LIST_HEAD(&cset->tasks);
1100 INIT_LIST_HEAD(&cset->mg_tasks);
1101 INIT_LIST_HEAD(&cset->mg_preload_node);
1102 INIT_LIST_HEAD(&cset->mg_node);
1103 INIT_LIST_HEAD(&cset->task_iters);
1104 INIT_HLIST_NODE(&cset->hlist);
1105
1106 /* Copy the set of subsystem state objects generated in
1107 * find_existing_css_set() */
1108 memcpy(cset->subsys, template, sizeof(cset->subsys));
1109
1110 spin_lock_irq(&css_set_lock);
1111 /* Add reference counts and links from the new css_set. */
1112 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1113 struct cgroup *c = link->cgrp;
1114
1115 if (c->root == cgrp->root)
1116 c = cgrp;
1117 link_css_set(&tmp_links, cset, c);
1118 }
1119
1120 BUG_ON(!list_empty(&tmp_links));
1121
1122 css_set_count++;
1123
1124 /* Add @cset to the hash table */
1125 key = css_set_hash(cset->subsys);
1126 hash_add(css_set_table, &cset->hlist, key);
1127
1128 for_each_subsys(ss, ssid) {
1129 struct cgroup_subsys_state *css = cset->subsys[ssid];
1130
1131 list_add_tail(&cset->e_cset_node[ssid],
1132 &css->cgroup->e_csets[ssid]);
1133 css_get(css);
1134 }
1135
1136 spin_unlock_irq(&css_set_lock);
1137
1138 return cset;
1139}
1140
1141static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1142{
1143 struct cgroup *root_cgrp = kf_root->kn->priv;
1144
1145 return root_cgrp->root;
1146}
1147
1148static int cgroup_init_root_id(struct cgroup_root *root)
1149{
1150 int id;
1151
1152 lockdep_assert_held(&cgroup_mutex);
1153
1154 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1155 if (id < 0)
1156 return id;
1157
1158 root->hierarchy_id = id;
1159 return 0;
1160}
1161
1162static void cgroup_exit_root_id(struct cgroup_root *root)
1163{
1164 lockdep_assert_held(&cgroup_mutex);
1165
1166 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1167}
1168
1169static void cgroup_free_root(struct cgroup_root *root)
1170{
1171 if (root) {
1172 idr_destroy(&root->cgroup_idr);
1173 kfree(root);
1174 }
1175}
1176
1177static void cgroup_destroy_root(struct cgroup_root *root)
1178{
1179 struct cgroup *cgrp = &root->cgrp;
1180 struct cgrp_cset_link *link, *tmp_link;
1181
1182 trace_cgroup_destroy_root(root);
1183
1184 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1185
1186 BUG_ON(atomic_read(&root->nr_cgrps));
1187 BUG_ON(!list_empty(&cgrp->self.children));
1188
1189 /* Rebind all subsystems back to the default hierarchy */
1190 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1191
1192 /*
1193 * Release all the links from cset_links to this hierarchy's
1194 * root cgroup
1195 */
1196 spin_lock_irq(&css_set_lock);
1197
1198 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1199 list_del(&link->cset_link);
1200 list_del(&link->cgrp_link);
1201 kfree(link);
1202 }
1203
1204 spin_unlock_irq(&css_set_lock);
1205
1206 if (!list_empty(&root->root_list)) {
1207 list_del(&root->root_list);
1208 cgroup_root_count--;
1209 }
1210
1211 cgroup_exit_root_id(root);
1212
1213 mutex_unlock(&cgroup_mutex);
1214
1215 kernfs_destroy_root(root->kf_root);
1216 cgroup_free_root(root);
1217}
1218
1219/*
1220 * look up cgroup associated with current task's cgroup namespace on the
1221 * specified hierarchy
1222 */
1223static struct cgroup *
1224current_cgns_cgroup_from_root(struct cgroup_root *root)
1225{
1226 struct cgroup *res = NULL;
1227 struct css_set *cset;
1228
1229 lockdep_assert_held(&css_set_lock);
1230
1231 rcu_read_lock();
1232
1233 cset = current->nsproxy->cgroup_ns->root_cset;
1234 if (cset == &init_css_set) {
1235 res = &root->cgrp;
1236 } else {
1237 struct cgrp_cset_link *link;
1238
1239 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1240 struct cgroup *c = link->cgrp;
1241
1242 if (c->root == root) {
1243 res = c;
1244 break;
1245 }
1246 }
1247 }
1248 rcu_read_unlock();
1249
1250 BUG_ON(!res);
1251 return res;
1252}
1253
1254/* look up cgroup associated with given css_set on the specified hierarchy */
1255static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1256 struct cgroup_root *root)
1257{
1258 struct cgroup *res = NULL;
1259
1260 lockdep_assert_held(&cgroup_mutex);
1261 lockdep_assert_held(&css_set_lock);
1262
1263 if (cset == &init_css_set) {
1264 res = &root->cgrp;
1265 } else {
1266 struct cgrp_cset_link *link;
1267
1268 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1269 struct cgroup *c = link->cgrp;
1270
1271 if (c->root == root) {
1272 res = c;
1273 break;
1274 }
1275 }
1276 }
1277
1278 BUG_ON(!res);
1279 return res;
1280}
1281
1282/*
1283 * Return the cgroup for "task" from the given hierarchy. Must be
1284 * called with cgroup_mutex and css_set_lock held.
1285 */
1286static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1287 struct cgroup_root *root)
1288{
1289 /*
1290 * No need to lock the task - since we hold cgroup_mutex the
1291 * task can't change groups, so the only thing that can happen
1292 * is that it exits and its css is set back to init_css_set.
1293 */
1294 return cset_cgroup_from_root(task_css_set(task), root);
1295}
1296
1297/*
1298 * A task must hold cgroup_mutex to modify cgroups.
1299 *
1300 * Any task can increment and decrement the count field without lock.
1301 * So in general, code holding cgroup_mutex can't rely on the count
1302 * field not changing. However, if the count goes to zero, then only
1303 * cgroup_attach_task() can increment it again. Because a count of zero
1304 * means that no tasks are currently attached, therefore there is no
1305 * way a task attached to that cgroup can fork (the other way to
1306 * increment the count). So code holding cgroup_mutex can safely
1307 * assume that if the count is zero, it will stay zero. Similarly, if
1308 * a task holds cgroup_mutex on a cgroup with zero count, it
1309 * knows that the cgroup won't be removed, as cgroup_rmdir()
1310 * needs that mutex.
1311 *
1312 * A cgroup can only be deleted if both its 'count' of using tasks
1313 * is zero, and its list of 'children' cgroups is empty. Since all
1314 * tasks in the system use _some_ cgroup, and since there is always at
1315 * least one task in the system (init, pid == 1), therefore, root cgroup
1316 * always has either children cgroups and/or using tasks. So we don't
1317 * need a special hack to ensure that root cgroup cannot be deleted.
1318 *
1319 * P.S. One more locking exception. RCU is used to guard the
1320 * update of a tasks cgroup pointer by cgroup_attach_task()
1321 */
1322
1323static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1324static const struct file_operations proc_cgroupstats_operations;
1325
1326static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1327 char *buf)
1328{
1329 struct cgroup_subsys *ss = cft->ss;
1330
1331 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1332 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1333 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1334 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1335 cft->name);
1336 else
1337 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1338 return buf;
1339}
1340
1341/**
1342 * cgroup_file_mode - deduce file mode of a control file
1343 * @cft: the control file in question
1344 *
1345 * S_IRUGO for read, S_IWUSR for write.
1346 */
1347static umode_t cgroup_file_mode(const struct cftype *cft)
1348{
1349 umode_t mode = 0;
1350
1351 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1352 mode |= S_IRUGO;
1353
1354 if (cft->write_u64 || cft->write_s64 || cft->write) {
1355 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1356 mode |= S_IWUGO;
1357 else
1358 mode |= S_IWUSR;
1359 }
1360
1361 return mode;
1362}
1363
1364/**
1365 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1366 * @subtree_control: the new subtree_control mask to consider
1367 * @this_ss_mask: available subsystems
1368 *
1369 * On the default hierarchy, a subsystem may request other subsystems to be
1370 * enabled together through its ->depends_on mask. In such cases, more
1371 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1372 *
1373 * This function calculates which subsystems need to be enabled if
1374 * @subtree_control is to be applied while restricted to @this_ss_mask.
1375 */
1376static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1377{
1378 u16 cur_ss_mask = subtree_control;
1379 struct cgroup_subsys *ss;
1380 int ssid;
1381
1382 lockdep_assert_held(&cgroup_mutex);
1383
1384 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1385
1386 while (true) {
1387 u16 new_ss_mask = cur_ss_mask;
1388
1389 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1390 new_ss_mask |= ss->depends_on;
1391 } while_each_subsys_mask();
1392
1393 /*
1394 * Mask out subsystems which aren't available. This can
1395 * happen only if some depended-upon subsystems were bound
1396 * to non-default hierarchies.
1397 */
1398 new_ss_mask &= this_ss_mask;
1399
1400 if (new_ss_mask == cur_ss_mask)
1401 break;
1402 cur_ss_mask = new_ss_mask;
1403 }
1404
1405 return cur_ss_mask;
1406}
1407
1408/**
1409 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1410 * @kn: the kernfs_node being serviced
1411 *
1412 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1413 * the method finishes if locking succeeded. Note that once this function
1414 * returns the cgroup returned by cgroup_kn_lock_live() may become
1415 * inaccessible any time. If the caller intends to continue to access the
1416 * cgroup, it should pin it before invoking this function.
1417 */
1418static void cgroup_kn_unlock(struct kernfs_node *kn)
1419{
1420 struct cgroup *cgrp;
1421
1422 if (kernfs_type(kn) == KERNFS_DIR)
1423 cgrp = kn->priv;
1424 else
1425 cgrp = kn->parent->priv;
1426
1427 mutex_unlock(&cgroup_mutex);
1428
1429 kernfs_unbreak_active_protection(kn);
1430 cgroup_put(cgrp);
1431}
1432
1433/**
1434 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1435 * @kn: the kernfs_node being serviced
1436 * @drain_offline: perform offline draining on the cgroup
1437 *
1438 * This helper is to be used by a cgroup kernfs method currently servicing
1439 * @kn. It breaks the active protection, performs cgroup locking and
1440 * verifies that the associated cgroup is alive. Returns the cgroup if
1441 * alive; otherwise, %NULL. A successful return should be undone by a
1442 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1443 * cgroup is drained of offlining csses before return.
1444 *
1445 * Any cgroup kernfs method implementation which requires locking the
1446 * associated cgroup should use this helper. It avoids nesting cgroup
1447 * locking under kernfs active protection and allows all kernfs operations
1448 * including self-removal.
1449 */
1450static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1451 bool drain_offline)
1452{
1453 struct cgroup *cgrp;
1454
1455 if (kernfs_type(kn) == KERNFS_DIR)
1456 cgrp = kn->priv;
1457 else
1458 cgrp = kn->parent->priv;
1459
1460 /*
1461 * We're gonna grab cgroup_mutex which nests outside kernfs
1462 * active_ref. cgroup liveliness check alone provides enough
1463 * protection against removal. Ensure @cgrp stays accessible and
1464 * break the active_ref protection.
1465 */
1466 if (!cgroup_tryget(cgrp))
1467 return NULL;
1468 kernfs_break_active_protection(kn);
1469
1470 if (drain_offline)
1471 cgroup_lock_and_drain_offline(cgrp);
1472 else
1473 mutex_lock(&cgroup_mutex);
1474
1475 if (!cgroup_is_dead(cgrp))
1476 return cgrp;
1477
1478 cgroup_kn_unlock(kn);
1479 return NULL;
1480}
1481
1482static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1483{
1484 char name[CGROUP_FILE_NAME_MAX];
1485
1486 lockdep_assert_held(&cgroup_mutex);
1487
1488 if (cft->file_offset) {
1489 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1490 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1491
1492 spin_lock_irq(&cgroup_file_kn_lock);
1493 cfile->kn = NULL;
1494 spin_unlock_irq(&cgroup_file_kn_lock);
1495 }
1496
1497 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1498}
1499
1500/**
1501 * css_clear_dir - remove subsys files in a cgroup directory
1502 * @css: taget css
1503 */
1504static void css_clear_dir(struct cgroup_subsys_state *css)
1505{
1506 struct cgroup *cgrp = css->cgroup;
1507 struct cftype *cfts;
1508
1509 if (!(css->flags & CSS_VISIBLE))
1510 return;
1511
1512 css->flags &= ~CSS_VISIBLE;
1513
1514 list_for_each_entry(cfts, &css->ss->cfts, node)
1515 cgroup_addrm_files(css, cgrp, cfts, false);
1516}
1517
1518/**
1519 * css_populate_dir - create subsys files in a cgroup directory
1520 * @css: target css
1521 *
1522 * On failure, no file is added.
1523 */
1524static int css_populate_dir(struct cgroup_subsys_state *css)
1525{
1526 struct cgroup *cgrp = css->cgroup;
1527 struct cftype *cfts, *failed_cfts;
1528 int ret;
1529
1530 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1531 return 0;
1532
1533 if (!css->ss) {
1534 if (cgroup_on_dfl(cgrp))
1535 cfts = cgroup_dfl_base_files;
1536 else
1537 cfts = cgroup_legacy_base_files;
1538
1539 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1540 }
1541
1542 list_for_each_entry(cfts, &css->ss->cfts, node) {
1543 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1544 if (ret < 0) {
1545 failed_cfts = cfts;
1546 goto err;
1547 }
1548 }
1549
1550 css->flags |= CSS_VISIBLE;
1551
1552 return 0;
1553err:
1554 list_for_each_entry(cfts, &css->ss->cfts, node) {
1555 if (cfts == failed_cfts)
1556 break;
1557 cgroup_addrm_files(css, cgrp, cfts, false);
1558 }
1559 return ret;
1560}
1561
1562static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1563{
1564 struct cgroup *dcgrp = &dst_root->cgrp;
1565 struct cgroup_subsys *ss;
1566 int ssid, i, ret;
1567
1568 lockdep_assert_held(&cgroup_mutex);
1569
1570 do_each_subsys_mask(ss, ssid, ss_mask) {
1571 /*
1572 * If @ss has non-root csses attached to it, can't move.
1573 * If @ss is an implicit controller, it is exempt from this
1574 * rule and can be stolen.
1575 */
1576 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1577 !ss->implicit_on_dfl)
1578 return -EBUSY;
1579
1580 /* can't move between two non-dummy roots either */
1581 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1582 return -EBUSY;
1583 } while_each_subsys_mask();
1584
1585 do_each_subsys_mask(ss, ssid, ss_mask) {
1586 struct cgroup_root *src_root = ss->root;
1587 struct cgroup *scgrp = &src_root->cgrp;
1588 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1589 struct css_set *cset;
1590
1591 WARN_ON(!css || cgroup_css(dcgrp, ss));
1592
1593 /* disable from the source */
1594 src_root->subsys_mask &= ~(1 << ssid);
1595 WARN_ON(cgroup_apply_control(scgrp));
1596 cgroup_finalize_control(scgrp, 0);
1597
1598 /* rebind */
1599 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1600 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1601 ss->root = dst_root;
1602 css->cgroup = dcgrp;
1603
1604 spin_lock_irq(&css_set_lock);
1605 hash_for_each(css_set_table, i, cset, hlist)
1606 list_move_tail(&cset->e_cset_node[ss->id],
1607 &dcgrp->e_csets[ss->id]);
1608 spin_unlock_irq(&css_set_lock);
1609
1610 /* default hierarchy doesn't enable controllers by default */
1611 dst_root->subsys_mask |= 1 << ssid;
1612 if (dst_root == &cgrp_dfl_root) {
1613 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1614 } else {
1615 dcgrp->subtree_control |= 1 << ssid;
1616 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1617 }
1618
1619 ret = cgroup_apply_control(dcgrp);
1620 if (ret)
1621 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1622 ss->name, ret);
1623
1624 if (ss->bind)
1625 ss->bind(css);
1626 } while_each_subsys_mask();
1627
1628 kernfs_activate(dcgrp->kn);
1629 return 0;
1630}
1631
1632static int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1633 struct kernfs_root *kf_root)
1634{
1635 int len = 0;
1636 char *buf = NULL;
1637 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1638 struct cgroup *ns_cgroup;
1639
1640 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1641 if (!buf)
1642 return -ENOMEM;
1643
1644 spin_lock_irq(&css_set_lock);
1645 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1646 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1647 spin_unlock_irq(&css_set_lock);
1648
1649 if (len >= PATH_MAX)
1650 len = -ERANGE;
1651 else if (len > 0) {
1652 seq_escape(sf, buf, " \t\n\\");
1653 len = 0;
1654 }
1655 kfree(buf);
1656 return len;
1657}
1658
1659static int cgroup_show_options(struct seq_file *seq,
1660 struct kernfs_root *kf_root)
1661{
1662 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1663 struct cgroup_subsys *ss;
1664 int ssid;
1665
1666 if (root != &cgrp_dfl_root)
1667 for_each_subsys(ss, ssid)
1668 if (root->subsys_mask & (1 << ssid))
1669 seq_show_option(seq, ss->legacy_name, NULL);
1670 if (root->flags & CGRP_ROOT_NOPREFIX)
1671 seq_puts(seq, ",noprefix");
1672 if (root->flags & CGRP_ROOT_XATTR)
1673 seq_puts(seq, ",xattr");
1674
1675 spin_lock(&release_agent_path_lock);
1676 if (strlen(root->release_agent_path))
1677 seq_show_option(seq, "release_agent",
1678 root->release_agent_path);
1679 spin_unlock(&release_agent_path_lock);
1680
1681 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1682 seq_puts(seq, ",clone_children");
1683 if (strlen(root->name))
1684 seq_show_option(seq, "name", root->name);
1685 return 0;
1686}
1687
1688struct cgroup_sb_opts {
1689 u16 subsys_mask;
1690 unsigned int flags;
1691 char *release_agent;
1692 bool cpuset_clone_children;
1693 char *name;
1694 /* User explicitly requested empty subsystem */
1695 bool none;
1696};
1697
1698static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1699{
1700 char *token, *o = data;
1701 bool all_ss = false, one_ss = false;
1702 u16 mask = U16_MAX;
1703 struct cgroup_subsys *ss;
1704 int nr_opts = 0;
1705 int i;
1706
1707#ifdef CONFIG_CPUSETS
1708 mask = ~((u16)1 << cpuset_cgrp_id);
1709#endif
1710
1711 memset(opts, 0, sizeof(*opts));
1712
1713 while ((token = strsep(&o, ",")) != NULL) {
1714 nr_opts++;
1715
1716 if (!*token)
1717 return -EINVAL;
1718 if (!strcmp(token, "none")) {
1719 /* Explicitly have no subsystems */
1720 opts->none = true;
1721 continue;
1722 }
1723 if (!strcmp(token, "all")) {
1724 /* Mutually exclusive option 'all' + subsystem name */
1725 if (one_ss)
1726 return -EINVAL;
1727 all_ss = true;
1728 continue;
1729 }
1730 if (!strcmp(token, "noprefix")) {
1731 opts->flags |= CGRP_ROOT_NOPREFIX;
1732 continue;
1733 }
1734 if (!strcmp(token, "clone_children")) {
1735 opts->cpuset_clone_children = true;
1736 continue;
1737 }
1738 if (!strcmp(token, "xattr")) {
1739 opts->flags |= CGRP_ROOT_XATTR;
1740 continue;
1741 }
1742 if (!strncmp(token, "release_agent=", 14)) {
1743 /* Specifying two release agents is forbidden */
1744 if (opts->release_agent)
1745 return -EINVAL;
1746 opts->release_agent =
1747 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1748 if (!opts->release_agent)
1749 return -ENOMEM;
1750 continue;
1751 }
1752 if (!strncmp(token, "name=", 5)) {
1753 const char *name = token + 5;
1754 /* Can't specify an empty name */
1755 if (!strlen(name))
1756 return -EINVAL;
1757 /* Must match [\w.-]+ */
1758 for (i = 0; i < strlen(name); i++) {
1759 char c = name[i];
1760 if (isalnum(c))
1761 continue;
1762 if ((c == '.') || (c == '-') || (c == '_'))
1763 continue;
1764 return -EINVAL;
1765 }
1766 /* Specifying two names is forbidden */
1767 if (opts->name)
1768 return -EINVAL;
1769 opts->name = kstrndup(name,
1770 MAX_CGROUP_ROOT_NAMELEN - 1,
1771 GFP_KERNEL);
1772 if (!opts->name)
1773 return -ENOMEM;
1774
1775 continue;
1776 }
1777
1778 for_each_subsys(ss, i) {
1779 if (strcmp(token, ss->legacy_name))
1780 continue;
1781 if (!cgroup_ssid_enabled(i))
1782 continue;
1783 if (cgroup_ssid_no_v1(i))
1784 continue;
1785
1786 /* Mutually exclusive option 'all' + subsystem name */
1787 if (all_ss)
1788 return -EINVAL;
1789 opts->subsys_mask |= (1 << i);
1790 one_ss = true;
1791
1792 break;
1793 }
1794 if (i == CGROUP_SUBSYS_COUNT)
1795 return -ENOENT;
1796 }
1797
1798 /*
1799 * If the 'all' option was specified select all the subsystems,
1800 * otherwise if 'none', 'name=' and a subsystem name options were
1801 * not specified, let's default to 'all'
1802 */
1803 if (all_ss || (!one_ss && !opts->none && !opts->name))
1804 for_each_subsys(ss, i)
1805 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1806 opts->subsys_mask |= (1 << i);
1807
1808 /*
1809 * We either have to specify by name or by subsystems. (So all
1810 * empty hierarchies must have a name).
1811 */
1812 if (!opts->subsys_mask && !opts->name)
1813 return -EINVAL;
1814
1815 /*
1816 * Option noprefix was introduced just for backward compatibility
1817 * with the old cpuset, so we allow noprefix only if mounting just
1818 * the cpuset subsystem.
1819 */
1820 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1821 return -EINVAL;
1822
1823 /* Can't specify "none" and some subsystems */
1824 if (opts->subsys_mask && opts->none)
1825 return -EINVAL;
1826
1827 return 0;
1828}
1829
1830static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1831{
1832 int ret = 0;
1833 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1834 struct cgroup_sb_opts opts;
1835 u16 added_mask, removed_mask;
1836
1837 if (root == &cgrp_dfl_root) {
1838 pr_err("remount is not allowed\n");
1839 return -EINVAL;
1840 }
1841
1842 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1843
1844 /* See what subsystems are wanted */
1845 ret = parse_cgroupfs_options(data, &opts);
1846 if (ret)
1847 goto out_unlock;
1848
1849 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1850 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1851 task_tgid_nr(current), current->comm);
1852
1853 added_mask = opts.subsys_mask & ~root->subsys_mask;
1854 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1855
1856 /* Don't allow flags or name to change at remount */
1857 if ((opts.flags ^ root->flags) ||
1858 (opts.name && strcmp(opts.name, root->name))) {
1859 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1860 opts.flags, opts.name ?: "", root->flags, root->name);
1861 ret = -EINVAL;
1862 goto out_unlock;
1863 }
1864
1865 /* remounting is not allowed for populated hierarchies */
1866 if (!list_empty(&root->cgrp.self.children)) {
1867 ret = -EBUSY;
1868 goto out_unlock;
1869 }
1870
1871 ret = rebind_subsystems(root, added_mask);
1872 if (ret)
1873 goto out_unlock;
1874
1875 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1876
1877 if (opts.release_agent) {
1878 spin_lock(&release_agent_path_lock);
1879 strcpy(root->release_agent_path, opts.release_agent);
1880 spin_unlock(&release_agent_path_lock);
1881 }
1882
1883 trace_cgroup_remount(root);
1884
1885 out_unlock:
1886 kfree(opts.release_agent);
1887 kfree(opts.name);
1888 mutex_unlock(&cgroup_mutex);
1889 return ret;
1890}
1891
1892/*
1893 * To reduce the fork() overhead for systems that are not actually using
1894 * their cgroups capability, we don't maintain the lists running through
1895 * each css_set to its tasks until we see the list actually used - in other
1896 * words after the first mount.
1897 */
1898static bool use_task_css_set_links __read_mostly;
1899
1900static void cgroup_enable_task_cg_lists(void)
1901{
1902 struct task_struct *p, *g;
1903
1904 spin_lock_irq(&css_set_lock);
1905
1906 if (use_task_css_set_links)
1907 goto out_unlock;
1908
1909 use_task_css_set_links = true;
1910
1911 /*
1912 * We need tasklist_lock because RCU is not safe against
1913 * while_each_thread(). Besides, a forking task that has passed
1914 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1915 * is not guaranteed to have its child immediately visible in the
1916 * tasklist if we walk through it with RCU.
1917 */
1918 read_lock(&tasklist_lock);
1919 do_each_thread(g, p) {
1920 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1921 task_css_set(p) != &init_css_set);
1922
1923 /*
1924 * We should check if the process is exiting, otherwise
1925 * it will race with cgroup_exit() in that the list
1926 * entry won't be deleted though the process has exited.
1927 * Do it while holding siglock so that we don't end up
1928 * racing against cgroup_exit().
1929 *
1930 * Interrupts were already disabled while acquiring
1931 * the css_set_lock, so we do not need to disable it
1932 * again when acquiring the sighand->siglock here.
1933 */
1934 spin_lock(&p->sighand->siglock);
1935 if (!(p->flags & PF_EXITING)) {
1936 struct css_set *cset = task_css_set(p);
1937
1938 if (!css_set_populated(cset))
1939 css_set_update_populated(cset, true);
1940 list_add_tail(&p->cg_list, &cset->tasks);
1941 get_css_set(cset);
1942 }
1943 spin_unlock(&p->sighand->siglock);
1944 } while_each_thread(g, p);
1945 read_unlock(&tasklist_lock);
1946out_unlock:
1947 spin_unlock_irq(&css_set_lock);
1948}
1949
1950static void init_cgroup_housekeeping(struct cgroup *cgrp)
1951{
1952 struct cgroup_subsys *ss;
1953 int ssid;
1954
1955 INIT_LIST_HEAD(&cgrp->self.sibling);
1956 INIT_LIST_HEAD(&cgrp->self.children);
1957 INIT_LIST_HEAD(&cgrp->cset_links);
1958 INIT_LIST_HEAD(&cgrp->pidlists);
1959 mutex_init(&cgrp->pidlist_mutex);
1960 cgrp->self.cgroup = cgrp;
1961 cgrp->self.flags |= CSS_ONLINE;
1962
1963 for_each_subsys(ss, ssid)
1964 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1965
1966 init_waitqueue_head(&cgrp->offline_waitq);
1967 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1968}
1969
1970static void init_cgroup_root(struct cgroup_root *root,
1971 struct cgroup_sb_opts *opts)
1972{
1973 struct cgroup *cgrp = &root->cgrp;
1974
1975 INIT_LIST_HEAD(&root->root_list);
1976 atomic_set(&root->nr_cgrps, 1);
1977 cgrp->root = root;
1978 init_cgroup_housekeeping(cgrp);
1979 idr_init(&root->cgroup_idr);
1980
1981 root->flags = opts->flags;
1982 if (opts->release_agent)
1983 strcpy(root->release_agent_path, opts->release_agent);
1984 if (opts->name)
1985 strcpy(root->name, opts->name);
1986 if (opts->cpuset_clone_children)
1987 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1988}
1989
1990static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1991{
1992 LIST_HEAD(tmp_links);
1993 struct cgroup *root_cgrp = &root->cgrp;
1994 struct css_set *cset;
1995 int i, ret;
1996
1997 lockdep_assert_held(&cgroup_mutex);
1998
1999 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
2000 if (ret < 0)
2001 goto out;
2002 root_cgrp->id = ret;
2003 root_cgrp->ancestor_ids[0] = ret;
2004
2005 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
2006 GFP_KERNEL);
2007 if (ret)
2008 goto out;
2009
2010 /*
2011 * We're accessing css_set_count without locking css_set_lock here,
2012 * but that's OK - it can only be increased by someone holding
2013 * cgroup_lock, and that's us. Later rebinding may disable
2014 * controllers on the default hierarchy and thus create new csets,
2015 * which can't be more than the existing ones. Allocate 2x.
2016 */
2017 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2018 if (ret)
2019 goto cancel_ref;
2020
2021 ret = cgroup_init_root_id(root);
2022 if (ret)
2023 goto cancel_ref;
2024
2025 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
2026 KERNFS_ROOT_CREATE_DEACTIVATED,
2027 root_cgrp);
2028 if (IS_ERR(root->kf_root)) {
2029 ret = PTR_ERR(root->kf_root);
2030 goto exit_root_id;
2031 }
2032 root_cgrp->kn = root->kf_root->kn;
2033
2034 ret = css_populate_dir(&root_cgrp->self);
2035 if (ret)
2036 goto destroy_root;
2037
2038 ret = rebind_subsystems(root, ss_mask);
2039 if (ret)
2040 goto destroy_root;
2041
2042 trace_cgroup_setup_root(root);
2043
2044 /*
2045 * There must be no failure case after here, since rebinding takes
2046 * care of subsystems' refcounts, which are explicitly dropped in
2047 * the failure exit path.
2048 */
2049 list_add(&root->root_list, &cgroup_roots);
2050 cgroup_root_count++;
2051
2052 /*
2053 * Link the root cgroup in this hierarchy into all the css_set
2054 * objects.
2055 */
2056 spin_lock_irq(&css_set_lock);
2057 hash_for_each(css_set_table, i, cset, hlist) {
2058 link_css_set(&tmp_links, cset, root_cgrp);
2059 if (css_set_populated(cset))
2060 cgroup_update_populated(root_cgrp, true);
2061 }
2062 spin_unlock_irq(&css_set_lock);
2063
2064 BUG_ON(!list_empty(&root_cgrp->self.children));
2065 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2066
2067 kernfs_activate(root_cgrp->kn);
2068 ret = 0;
2069 goto out;
2070
2071destroy_root:
2072 kernfs_destroy_root(root->kf_root);
2073 root->kf_root = NULL;
2074exit_root_id:
2075 cgroup_exit_root_id(root);
2076cancel_ref:
2077 percpu_ref_exit(&root_cgrp->self.refcnt);
2078out:
2079 free_cgrp_cset_links(&tmp_links);
2080 return ret;
2081}
2082
2083static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2084 int flags, const char *unused_dev_name,
2085 void *data)
2086{
2087 bool is_v2 = fs_type == &cgroup2_fs_type;
2088 struct super_block *pinned_sb = NULL;
2089 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2090 struct cgroup_subsys *ss;
2091 struct cgroup_root *root;
2092 struct cgroup_sb_opts opts;
2093 struct dentry *dentry;
2094 int ret;
2095 int i;
2096 bool new_sb;
2097
2098 get_cgroup_ns(ns);
2099
2100 /* Check if the caller has permission to mount. */
2101 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2102 put_cgroup_ns(ns);
2103 return ERR_PTR(-EPERM);
2104 }
2105
2106 /*
2107 * The first time anyone tries to mount a cgroup, enable the list
2108 * linking each css_set to its tasks and fix up all existing tasks.
2109 */
2110 if (!use_task_css_set_links)
2111 cgroup_enable_task_cg_lists();
2112
2113 if (is_v2) {
2114 if (data) {
2115 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2116 put_cgroup_ns(ns);
2117 return ERR_PTR(-EINVAL);
2118 }
2119 cgrp_dfl_visible = true;
2120 root = &cgrp_dfl_root;
2121 cgroup_get(&root->cgrp);
2122 goto out_mount;
2123 }
2124
2125 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2126
2127 /* First find the desired set of subsystems */
2128 ret = parse_cgroupfs_options(data, &opts);
2129 if (ret)
2130 goto out_unlock;
2131
2132 /*
2133 * Destruction of cgroup root is asynchronous, so subsystems may
2134 * still be dying after the previous unmount. Let's drain the
2135 * dying subsystems. We just need to ensure that the ones
2136 * unmounted previously finish dying and don't care about new ones
2137 * starting. Testing ref liveliness is good enough.
2138 */
2139 for_each_subsys(ss, i) {
2140 if (!(opts.subsys_mask & (1 << i)) ||
2141 ss->root == &cgrp_dfl_root)
2142 continue;
2143
2144 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2145 mutex_unlock(&cgroup_mutex);
2146 msleep(10);
2147 ret = restart_syscall();
2148 goto out_free;
2149 }
2150 cgroup_put(&ss->root->cgrp);
2151 }
2152
2153 for_each_root(root) {
2154 bool name_match = false;
2155
2156 if (root == &cgrp_dfl_root)
2157 continue;
2158
2159 /*
2160 * If we asked for a name then it must match. Also, if
2161 * name matches but sybsys_mask doesn't, we should fail.
2162 * Remember whether name matched.
2163 */
2164 if (opts.name) {
2165 if (strcmp(opts.name, root->name))
2166 continue;
2167 name_match = true;
2168 }
2169
2170 /*
2171 * If we asked for subsystems (or explicitly for no
2172 * subsystems) then they must match.
2173 */
2174 if ((opts.subsys_mask || opts.none) &&
2175 (opts.subsys_mask != root->subsys_mask)) {
2176 if (!name_match)
2177 continue;
2178 ret = -EBUSY;
2179 goto out_unlock;
2180 }
2181
2182 if (root->flags ^ opts.flags)
2183 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2184
2185 /*
2186 * We want to reuse @root whose lifetime is governed by its
2187 * ->cgrp. Let's check whether @root is alive and keep it
2188 * that way. As cgroup_kill_sb() can happen anytime, we
2189 * want to block it by pinning the sb so that @root doesn't
2190 * get killed before mount is complete.
2191 *
2192 * With the sb pinned, tryget_live can reliably indicate
2193 * whether @root can be reused. If it's being killed,
2194 * drain it. We can use wait_queue for the wait but this
2195 * path is super cold. Let's just sleep a bit and retry.
2196 */
2197 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2198 if (IS_ERR(pinned_sb) ||
2199 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2200 mutex_unlock(&cgroup_mutex);
2201 if (!IS_ERR_OR_NULL(pinned_sb))
2202 deactivate_super(pinned_sb);
2203 msleep(10);
2204 ret = restart_syscall();
2205 goto out_free;
2206 }
2207
2208 ret = 0;
2209 goto out_unlock;
2210 }
2211
2212 /*
2213 * No such thing, create a new one. name= matching without subsys
2214 * specification is allowed for already existing hierarchies but we
2215 * can't create new one without subsys specification.
2216 */
2217 if (!opts.subsys_mask && !opts.none) {
2218 ret = -EINVAL;
2219 goto out_unlock;
2220 }
2221
2222 /* Hierarchies may only be created in the initial cgroup namespace. */
2223 if (ns != &init_cgroup_ns) {
2224 ret = -EPERM;
2225 goto out_unlock;
2226 }
2227
2228 root = kzalloc(sizeof(*root), GFP_KERNEL);
2229 if (!root) {
2230 ret = -ENOMEM;
2231 goto out_unlock;
2232 }
2233
2234 init_cgroup_root(root, &opts);
2235
2236 ret = cgroup_setup_root(root, opts.subsys_mask);
2237 if (ret)
2238 cgroup_free_root(root);
2239
2240out_unlock:
2241 mutex_unlock(&cgroup_mutex);
2242out_free:
2243 kfree(opts.release_agent);
2244 kfree(opts.name);
2245
2246 if (ret) {
2247 put_cgroup_ns(ns);
2248 return ERR_PTR(ret);
2249 }
2250out_mount:
2251 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2252 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2253 &new_sb);
2254
2255 /*
2256 * In non-init cgroup namespace, instead of root cgroup's
2257 * dentry, we return the dentry corresponding to the
2258 * cgroupns->root_cgrp.
2259 */
2260 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2261 struct dentry *nsdentry;
2262 struct cgroup *cgrp;
2263
2264 mutex_lock(&cgroup_mutex);
2265 spin_lock_irq(&css_set_lock);
2266
2267 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2268
2269 spin_unlock_irq(&css_set_lock);
2270 mutex_unlock(&cgroup_mutex);
2271
2272 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2273 dput(dentry);
2274 dentry = nsdentry;
2275 }
2276
2277 if (IS_ERR(dentry) || !new_sb)
2278 cgroup_put(&root->cgrp);
2279
2280 /*
2281 * If @pinned_sb, we're reusing an existing root and holding an
2282 * extra ref on its sb. Mount is complete. Put the extra ref.
2283 */
2284 if (pinned_sb) {
2285 WARN_ON(new_sb);
2286 deactivate_super(pinned_sb);
2287 }
2288
2289 put_cgroup_ns(ns);
2290 return dentry;
2291}
2292
2293static void cgroup_kill_sb(struct super_block *sb)
2294{
2295 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2296 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2297
2298 /*
2299 * If @root doesn't have any mounts or children, start killing it.
2300 * This prevents new mounts by disabling percpu_ref_tryget_live().
2301 * cgroup_mount() may wait for @root's release.
2302 *
2303 * And don't kill the default root.
2304 */
2305 if (!list_empty(&root->cgrp.self.children) ||
2306 root == &cgrp_dfl_root)
2307 cgroup_put(&root->cgrp);
2308 else
2309 percpu_ref_kill(&root->cgrp.self.refcnt);
2310
2311 kernfs_kill_sb(sb);
2312}
2313
2314static struct file_system_type cgroup_fs_type = {
2315 .name = "cgroup",
2316 .mount = cgroup_mount,
2317 .kill_sb = cgroup_kill_sb,
2318 .fs_flags = FS_USERNS_MOUNT,
2319};
2320
2321static struct file_system_type cgroup2_fs_type = {
2322 .name = "cgroup2",
2323 .mount = cgroup_mount,
2324 .kill_sb = cgroup_kill_sb,
2325 .fs_flags = FS_USERNS_MOUNT,
2326};
2327
2328static int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2329 struct cgroup_namespace *ns)
2330{
2331 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2332
2333 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2334}
2335
2336int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2337 struct cgroup_namespace *ns)
2338{
2339 int ret;
2340
2341 mutex_lock(&cgroup_mutex);
2342 spin_lock_irq(&css_set_lock);
2343
2344 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2345
2346 spin_unlock_irq(&css_set_lock);
2347 mutex_unlock(&cgroup_mutex);
2348
2349 return ret;
2350}
2351EXPORT_SYMBOL_GPL(cgroup_path_ns);
2352
2353/**
2354 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2355 * @task: target task
2356 * @buf: the buffer to write the path into
2357 * @buflen: the length of the buffer
2358 *
2359 * Determine @task's cgroup on the first (the one with the lowest non-zero
2360 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2361 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2362 * cgroup controller callbacks.
2363 *
2364 * Return value is the same as kernfs_path().
2365 */
2366int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2367{
2368 struct cgroup_root *root;
2369 struct cgroup *cgrp;
2370 int hierarchy_id = 1;
2371 int ret;
2372
2373 mutex_lock(&cgroup_mutex);
2374 spin_lock_irq(&css_set_lock);
2375
2376 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2377
2378 if (root) {
2379 cgrp = task_cgroup_from_root(task, root);
2380 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2381 } else {
2382 /* if no hierarchy exists, everyone is in "/" */
2383 ret = strlcpy(buf, "/", buflen);
2384 }
2385
2386 spin_unlock_irq(&css_set_lock);
2387 mutex_unlock(&cgroup_mutex);
2388 return ret;
2389}
2390EXPORT_SYMBOL_GPL(task_cgroup_path);
2391
2392/* used to track tasks and other necessary states during migration */
2393struct cgroup_taskset {
2394 /* the src and dst cset list running through cset->mg_node */
2395 struct list_head src_csets;
2396 struct list_head dst_csets;
2397
2398 /* the subsys currently being processed */
2399 int ssid;
2400
2401 /*
2402 * Fields for cgroup_taskset_*() iteration.
2403 *
2404 * Before migration is committed, the target migration tasks are on
2405 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2406 * the csets on ->dst_csets. ->csets point to either ->src_csets
2407 * or ->dst_csets depending on whether migration is committed.
2408 *
2409 * ->cur_csets and ->cur_task point to the current task position
2410 * during iteration.
2411 */
2412 struct list_head *csets;
2413 struct css_set *cur_cset;
2414 struct task_struct *cur_task;
2415};
2416
2417#define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2418 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2419 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2420 .csets = &tset.src_csets, \
2421}
2422
2423/**
2424 * cgroup_taskset_add - try to add a migration target task to a taskset
2425 * @task: target task
2426 * @tset: target taskset
2427 *
2428 * Add @task, which is a migration target, to @tset. This function becomes
2429 * noop if @task doesn't need to be migrated. @task's css_set should have
2430 * been added as a migration source and @task->cg_list will be moved from
2431 * the css_set's tasks list to mg_tasks one.
2432 */
2433static void cgroup_taskset_add(struct task_struct *task,
2434 struct cgroup_taskset *tset)
2435{
2436 struct css_set *cset;
2437
2438 lockdep_assert_held(&css_set_lock);
2439
2440 /* @task either already exited or can't exit until the end */
2441 if (task->flags & PF_EXITING)
2442 return;
2443
2444 /* leave @task alone if post_fork() hasn't linked it yet */
2445 if (list_empty(&task->cg_list))
2446 return;
2447
2448 cset = task_css_set(task);
2449 if (!cset->mg_src_cgrp)
2450 return;
2451
2452 list_move_tail(&task->cg_list, &cset->mg_tasks);
2453 if (list_empty(&cset->mg_node))
2454 list_add_tail(&cset->mg_node, &tset->src_csets);
2455 if (list_empty(&cset->mg_dst_cset->mg_node))
2456 list_move_tail(&cset->mg_dst_cset->mg_node,
2457 &tset->dst_csets);
2458}
2459
2460/**
2461 * cgroup_taskset_first - reset taskset and return the first task
2462 * @tset: taskset of interest
2463 * @dst_cssp: output variable for the destination css
2464 *
2465 * @tset iteration is initialized and the first task is returned.
2466 */
2467struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2468 struct cgroup_subsys_state **dst_cssp)
2469{
2470 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2471 tset->cur_task = NULL;
2472
2473 return cgroup_taskset_next(tset, dst_cssp);
2474}
2475
2476/**
2477 * cgroup_taskset_next - iterate to the next task in taskset
2478 * @tset: taskset of interest
2479 * @dst_cssp: output variable for the destination css
2480 *
2481 * Return the next task in @tset. Iteration must have been initialized
2482 * with cgroup_taskset_first().
2483 */
2484struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2485 struct cgroup_subsys_state **dst_cssp)
2486{
2487 struct css_set *cset = tset->cur_cset;
2488 struct task_struct *task = tset->cur_task;
2489
2490 while (&cset->mg_node != tset->csets) {
2491 if (!task)
2492 task = list_first_entry(&cset->mg_tasks,
2493 struct task_struct, cg_list);
2494 else
2495 task = list_next_entry(task, cg_list);
2496
2497 if (&task->cg_list != &cset->mg_tasks) {
2498 tset->cur_cset = cset;
2499 tset->cur_task = task;
2500
2501 /*
2502 * This function may be called both before and
2503 * after cgroup_taskset_migrate(). The two cases
2504 * can be distinguished by looking at whether @cset
2505 * has its ->mg_dst_cset set.
2506 */
2507 if (cset->mg_dst_cset)
2508 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2509 else
2510 *dst_cssp = cset->subsys[tset->ssid];
2511
2512 return task;
2513 }
2514
2515 cset = list_next_entry(cset, mg_node);
2516 task = NULL;
2517 }
2518
2519 return NULL;
2520}
2521
2522/**
2523 * cgroup_taskset_migrate - migrate a taskset
2524 * @tset: taget taskset
2525 * @root: cgroup root the migration is taking place on
2526 *
2527 * Migrate tasks in @tset as setup by migration preparation functions.
2528 * This function fails iff one of the ->can_attach callbacks fails and
2529 * guarantees that either all or none of the tasks in @tset are migrated.
2530 * @tset is consumed regardless of success.
2531 */
2532static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2533 struct cgroup_root *root)
2534{
2535 struct cgroup_subsys *ss;
2536 struct task_struct *task, *tmp_task;
2537 struct css_set *cset, *tmp_cset;
2538 int ssid, failed_ssid, ret;
2539
2540 /* methods shouldn't be called if no task is actually migrating */
2541 if (list_empty(&tset->src_csets))
2542 return 0;
2543
2544 /* check that we can legitimately attach to the cgroup */
2545 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2546 if (ss->can_attach) {
2547 tset->ssid = ssid;
2548 ret = ss->can_attach(tset);
2549 if (ret) {
2550 failed_ssid = ssid;
2551 goto out_cancel_attach;
2552 }
2553 }
2554 } while_each_subsys_mask();
2555
2556 /*
2557 * Now that we're guaranteed success, proceed to move all tasks to
2558 * the new cgroup. There are no failure cases after here, so this
2559 * is the commit point.
2560 */
2561 spin_lock_irq(&css_set_lock);
2562 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2563 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2564 struct css_set *from_cset = task_css_set(task);
2565 struct css_set *to_cset = cset->mg_dst_cset;
2566
2567 get_css_set(to_cset);
2568 css_set_move_task(task, from_cset, to_cset, true);
2569 put_css_set_locked(from_cset);
2570 }
2571 }
2572 spin_unlock_irq(&css_set_lock);
2573
2574 /*
2575 * Migration is committed, all target tasks are now on dst_csets.
2576 * Nothing is sensitive to fork() after this point. Notify
2577 * controllers that migration is complete.
2578 */
2579 tset->csets = &tset->dst_csets;
2580
2581 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2582 if (ss->attach) {
2583 tset->ssid = ssid;
2584 ss->attach(tset);
2585 }
2586 } while_each_subsys_mask();
2587
2588 ret = 0;
2589 goto out_release_tset;
2590
2591out_cancel_attach:
2592 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2593 if (ssid == failed_ssid)
2594 break;
2595 if (ss->cancel_attach) {
2596 tset->ssid = ssid;
2597 ss->cancel_attach(tset);
2598 }
2599 } while_each_subsys_mask();
2600out_release_tset:
2601 spin_lock_irq(&css_set_lock);
2602 list_splice_init(&tset->dst_csets, &tset->src_csets);
2603 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2604 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2605 list_del_init(&cset->mg_node);
2606 }
2607 spin_unlock_irq(&css_set_lock);
2608 return ret;
2609}
2610
2611/**
2612 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2613 * @dst_cgrp: destination cgroup to test
2614 *
2615 * On the default hierarchy, except for the root, subtree_control must be
2616 * zero for migration destination cgroups with tasks so that child cgroups
2617 * don't compete against tasks.
2618 */
2619static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2620{
2621 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2622 !dst_cgrp->subtree_control;
2623}
2624
2625/**
2626 * cgroup_migrate_finish - cleanup after attach
2627 * @preloaded_csets: list of preloaded css_sets
2628 *
2629 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2630 * those functions for details.
2631 */
2632static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2633{
2634 struct css_set *cset, *tmp_cset;
2635
2636 lockdep_assert_held(&cgroup_mutex);
2637
2638 spin_lock_irq(&css_set_lock);
2639 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2640 cset->mg_src_cgrp = NULL;
2641 cset->mg_dst_cgrp = NULL;
2642 cset->mg_dst_cset = NULL;
2643 list_del_init(&cset->mg_preload_node);
2644 put_css_set_locked(cset);
2645 }
2646 spin_unlock_irq(&css_set_lock);
2647}
2648
2649/**
2650 * cgroup_migrate_add_src - add a migration source css_set
2651 * @src_cset: the source css_set to add
2652 * @dst_cgrp: the destination cgroup
2653 * @preloaded_csets: list of preloaded css_sets
2654 *
2655 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2656 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2657 * up by cgroup_migrate_finish().
2658 *
2659 * This function may be called without holding cgroup_threadgroup_rwsem
2660 * even if the target is a process. Threads may be created and destroyed
2661 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2662 * into play and the preloaded css_sets are guaranteed to cover all
2663 * migrations.
2664 */
2665static void cgroup_migrate_add_src(struct css_set *src_cset,
2666 struct cgroup *dst_cgrp,
2667 struct list_head *preloaded_csets)
2668{
2669 struct cgroup *src_cgrp;
2670
2671 lockdep_assert_held(&cgroup_mutex);
2672 lockdep_assert_held(&css_set_lock);
2673
2674 /*
2675 * If ->dead, @src_set is associated with one or more dead cgroups
2676 * and doesn't contain any migratable tasks. Ignore it early so
2677 * that the rest of migration path doesn't get confused by it.
2678 */
2679 if (src_cset->dead)
2680 return;
2681
2682 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2683
2684 if (!list_empty(&src_cset->mg_preload_node))
2685 return;
2686
2687 WARN_ON(src_cset->mg_src_cgrp);
2688 WARN_ON(src_cset->mg_dst_cgrp);
2689 WARN_ON(!list_empty(&src_cset->mg_tasks));
2690 WARN_ON(!list_empty(&src_cset->mg_node));
2691
2692 src_cset->mg_src_cgrp = src_cgrp;
2693 src_cset->mg_dst_cgrp = dst_cgrp;
2694 get_css_set(src_cset);
2695 list_add(&src_cset->mg_preload_node, preloaded_csets);
2696}
2697
2698/**
2699 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2700 * @preloaded_csets: list of preloaded source css_sets
2701 *
2702 * Tasks are about to be moved and all the source css_sets have been
2703 * preloaded to @preloaded_csets. This function looks up and pins all
2704 * destination css_sets, links each to its source, and append them to
2705 * @preloaded_csets.
2706 *
2707 * This function must be called after cgroup_migrate_add_src() has been
2708 * called on each migration source css_set. After migration is performed
2709 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2710 * @preloaded_csets.
2711 */
2712static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2713{
2714 LIST_HEAD(csets);
2715 struct css_set *src_cset, *tmp_cset;
2716
2717 lockdep_assert_held(&cgroup_mutex);
2718
2719 /* look up the dst cset for each src cset and link it to src */
2720 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2721 struct css_set *dst_cset;
2722
2723 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2724 if (!dst_cset)
2725 goto err;
2726
2727 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2728
2729 /*
2730 * If src cset equals dst, it's noop. Drop the src.
2731 * cgroup_migrate() will skip the cset too. Note that we
2732 * can't handle src == dst as some nodes are used by both.
2733 */
2734 if (src_cset == dst_cset) {
2735 src_cset->mg_src_cgrp = NULL;
2736 src_cset->mg_dst_cgrp = NULL;
2737 list_del_init(&src_cset->mg_preload_node);
2738 put_css_set(src_cset);
2739 put_css_set(dst_cset);
2740 continue;
2741 }
2742
2743 src_cset->mg_dst_cset = dst_cset;
2744
2745 if (list_empty(&dst_cset->mg_preload_node))
2746 list_add(&dst_cset->mg_preload_node, &csets);
2747 else
2748 put_css_set(dst_cset);
2749 }
2750
2751 list_splice_tail(&csets, preloaded_csets);
2752 return 0;
2753err:
2754 cgroup_migrate_finish(&csets);
2755 return -ENOMEM;
2756}
2757
2758/**
2759 * cgroup_migrate - migrate a process or task to a cgroup
2760 * @leader: the leader of the process or the task to migrate
2761 * @threadgroup: whether @leader points to the whole process or a single task
2762 * @root: cgroup root migration is taking place on
2763 *
2764 * Migrate a process or task denoted by @leader. If migrating a process,
2765 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2766 * responsible for invoking cgroup_migrate_add_src() and
2767 * cgroup_migrate_prepare_dst() on the targets before invoking this
2768 * function and following up with cgroup_migrate_finish().
2769 *
2770 * As long as a controller's ->can_attach() doesn't fail, this function is
2771 * guaranteed to succeed. This means that, excluding ->can_attach()
2772 * failure, when migrating multiple targets, the success or failure can be
2773 * decided for all targets by invoking group_migrate_prepare_dst() before
2774 * actually starting migrating.
2775 */
2776static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2777 struct cgroup_root *root)
2778{
2779 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2780 struct task_struct *task;
2781
2782 /*
2783 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2784 * already PF_EXITING could be freed from underneath us unless we
2785 * take an rcu_read_lock.
2786 */
2787 spin_lock_irq(&css_set_lock);
2788 rcu_read_lock();
2789 task = leader;
2790 do {
2791 cgroup_taskset_add(task, &tset);
2792 if (!threadgroup)
2793 break;
2794 } while_each_thread(leader, task);
2795 rcu_read_unlock();
2796 spin_unlock_irq(&css_set_lock);
2797
2798 return cgroup_taskset_migrate(&tset, root);
2799}
2800
2801/**
2802 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2803 * @dst_cgrp: the cgroup to attach to
2804 * @leader: the task or the leader of the threadgroup to be attached
2805 * @threadgroup: attach the whole threadgroup?
2806 *
2807 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2808 */
2809static int cgroup_attach_task(struct cgroup *dst_cgrp,
2810 struct task_struct *leader, bool threadgroup)
2811{
2812 LIST_HEAD(preloaded_csets);
2813 struct task_struct *task;
2814 int ret;
2815
2816 if (!cgroup_may_migrate_to(dst_cgrp))
2817 return -EBUSY;
2818
2819 /* look up all src csets */
2820 spin_lock_irq(&css_set_lock);
2821 rcu_read_lock();
2822 task = leader;
2823 do {
2824 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2825 &preloaded_csets);
2826 if (!threadgroup)
2827 break;
2828 } while_each_thread(leader, task);
2829 rcu_read_unlock();
2830 spin_unlock_irq(&css_set_lock);
2831
2832 /* prepare dst csets and commit */
2833 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2834 if (!ret)
2835 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2836
2837 cgroup_migrate_finish(&preloaded_csets);
2838
2839 if (!ret)
2840 trace_cgroup_attach_task(dst_cgrp, leader, threadgroup);
2841
2842 return ret;
2843}
2844
2845static int cgroup_procs_write_permission(struct task_struct *task,
2846 struct cgroup *dst_cgrp,
2847 struct kernfs_open_file *of)
2848{
2849 const struct cred *cred = current_cred();
2850 const struct cred *tcred = get_task_cred(task);
2851 int ret = 0;
2852
2853 /*
2854 * even if we're attaching all tasks in the thread group, we only
2855 * need to check permissions on one of them.
2856 */
2857 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2858 !uid_eq(cred->euid, tcred->uid) &&
2859 !uid_eq(cred->euid, tcred->suid))
2860 ret = -EACCES;
2861
2862 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2863 struct super_block *sb = of->file->f_path.dentry->d_sb;
2864 struct cgroup *cgrp;
2865 struct inode *inode;
2866
2867 spin_lock_irq(&css_set_lock);
2868 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2869 spin_unlock_irq(&css_set_lock);
2870
2871 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2872 cgrp = cgroup_parent(cgrp);
2873
2874 ret = -ENOMEM;
2875 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2876 if (inode) {
2877 ret = inode_permission(inode, MAY_WRITE);
2878 iput(inode);
2879 }
2880 }
2881
2882 put_cred(tcred);
2883 return ret;
2884}
2885
2886/*
2887 * Find the task_struct of the task to attach by vpid and pass it along to the
2888 * function to attach either it or all tasks in its threadgroup. Will lock
2889 * cgroup_mutex and threadgroup.
2890 */
2891static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2892 size_t nbytes, loff_t off, bool threadgroup)
2893{
2894 struct task_struct *tsk;
2895 struct cgroup_subsys *ss;
2896 struct cgroup *cgrp;
2897 pid_t pid;
2898 int ssid, ret;
2899
2900 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2901 return -EINVAL;
2902
2903 cgrp = cgroup_kn_lock_live(of->kn, false);
2904 if (!cgrp)
2905 return -ENODEV;
2906
2907 percpu_down_write(&cgroup_threadgroup_rwsem);
2908 rcu_read_lock();
2909 if (pid) {
2910 tsk = find_task_by_vpid(pid);
2911 if (!tsk) {
2912 ret = -ESRCH;
2913 goto out_unlock_rcu;
2914 }
2915 } else {
2916 tsk = current;
2917 }
2918
2919 if (threadgroup)
2920 tsk = tsk->group_leader;
2921
2922 /*
2923 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2924 * trapped in a cpuset, or RT worker may be born in a cgroup
2925 * with no rt_runtime allocated. Just say no.
2926 */
2927 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2928 ret = -EINVAL;
2929 goto out_unlock_rcu;
2930 }
2931
2932 get_task_struct(tsk);
2933 rcu_read_unlock();
2934
2935 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2936 if (!ret)
2937 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2938
2939 put_task_struct(tsk);
2940 goto out_unlock_threadgroup;
2941
2942out_unlock_rcu:
2943 rcu_read_unlock();
2944out_unlock_threadgroup:
2945 percpu_up_write(&cgroup_threadgroup_rwsem);
2946 for_each_subsys(ss, ssid)
2947 if (ss->post_attach)
2948 ss->post_attach();
2949 cgroup_kn_unlock(of->kn);
2950 return ret ?: nbytes;
2951}
2952
2953/**
2954 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2955 * @from: attach to all cgroups of a given task
2956 * @tsk: the task to be attached
2957 */
2958int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2959{
2960 struct cgroup_root *root;
2961 int retval = 0;
2962
2963 mutex_lock(&cgroup_mutex);
2964 percpu_down_write(&cgroup_threadgroup_rwsem);
2965 for_each_root(root) {
2966 struct cgroup *from_cgrp;
2967
2968 if (root == &cgrp_dfl_root)
2969 continue;
2970
2971 spin_lock_irq(&css_set_lock);
2972 from_cgrp = task_cgroup_from_root(from, root);
2973 spin_unlock_irq(&css_set_lock);
2974
2975 retval = cgroup_attach_task(from_cgrp, tsk, false);
2976 if (retval)
2977 break;
2978 }
2979 percpu_up_write(&cgroup_threadgroup_rwsem);
2980 mutex_unlock(&cgroup_mutex);
2981
2982 return retval;
2983}
2984EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2985
2986static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2987 char *buf, size_t nbytes, loff_t off)
2988{
2989 return __cgroup_procs_write(of, buf, nbytes, off, false);
2990}
2991
2992static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2993 char *buf, size_t nbytes, loff_t off)
2994{
2995 return __cgroup_procs_write(of, buf, nbytes, off, true);
2996}
2997
2998static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2999 char *buf, size_t nbytes, loff_t off)
3000{
3001 struct cgroup *cgrp;
3002
3003 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
3004
3005 cgrp = cgroup_kn_lock_live(of->kn, false);
3006 if (!cgrp)
3007 return -ENODEV;
3008 spin_lock(&release_agent_path_lock);
3009 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
3010 sizeof(cgrp->root->release_agent_path));
3011 spin_unlock(&release_agent_path_lock);
3012 cgroup_kn_unlock(of->kn);
3013 return nbytes;
3014}
3015
3016static int cgroup_release_agent_show(struct seq_file *seq, void *v)
3017{
3018 struct cgroup *cgrp = seq_css(seq)->cgroup;
3019
3020 spin_lock(&release_agent_path_lock);
3021 seq_puts(seq, cgrp->root->release_agent_path);
3022 spin_unlock(&release_agent_path_lock);
3023 seq_putc(seq, '\n');
3024 return 0;
3025}
3026
3027static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
3028{
3029 seq_puts(seq, "0\n");
3030 return 0;
3031}
3032
3033static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
3034{
3035 struct cgroup_subsys *ss;
3036 bool printed = false;
3037 int ssid;
3038
3039 do_each_subsys_mask(ss, ssid, ss_mask) {
3040 if (printed)
3041 seq_putc(seq, ' ');
3042 seq_printf(seq, "%s", ss->name);
3043 printed = true;
3044 } while_each_subsys_mask();
3045 if (printed)
3046 seq_putc(seq, '\n');
3047}
3048
3049/* show controllers which are enabled from the parent */
3050static int cgroup_controllers_show(struct seq_file *seq, void *v)
3051{
3052 struct cgroup *cgrp = seq_css(seq)->cgroup;
3053
3054 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3055 return 0;
3056}
3057
3058/* show controllers which are enabled for a given cgroup's children */
3059static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3060{
3061 struct cgroup *cgrp = seq_css(seq)->cgroup;
3062
3063 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3064 return 0;
3065}
3066
3067/**
3068 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3069 * @cgrp: root of the subtree to update csses for
3070 *
3071 * @cgrp's control masks have changed and its subtree's css associations
3072 * need to be updated accordingly. This function looks up all css_sets
3073 * which are attached to the subtree, creates the matching updated css_sets
3074 * and migrates the tasks to the new ones.
3075 */
3076static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3077{
3078 LIST_HEAD(preloaded_csets);
3079 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
3080 struct cgroup_subsys_state *d_css;
3081 struct cgroup *dsct;
3082 struct css_set *src_cset;
3083 int ret;
3084
3085 lockdep_assert_held(&cgroup_mutex);
3086
3087 percpu_down_write(&cgroup_threadgroup_rwsem);
3088
3089 /* look up all csses currently attached to @cgrp's subtree */
3090 spin_lock_irq(&css_set_lock);
3091 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3092 struct cgrp_cset_link *link;
3093
3094 list_for_each_entry(link, &dsct->cset_links, cset_link)
3095 cgroup_migrate_add_src(link->cset, dsct,
3096 &preloaded_csets);
3097 }
3098 spin_unlock_irq(&css_set_lock);
3099
3100 /* NULL dst indicates self on default hierarchy */
3101 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
3102 if (ret)
3103 goto out_finish;
3104
3105 spin_lock_irq(&css_set_lock);
3106 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3107 struct task_struct *task, *ntask;
3108
3109 /* src_csets precede dst_csets, break on the first dst_cset */
3110 if (!src_cset->mg_src_cgrp)
3111 break;
3112
3113 /* all tasks in src_csets need to be migrated */
3114 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3115 cgroup_taskset_add(task, &tset);
3116 }
3117 spin_unlock_irq(&css_set_lock);
3118
3119 ret = cgroup_taskset_migrate(&tset, cgrp->root);
3120out_finish:
3121 cgroup_migrate_finish(&preloaded_csets);
3122 percpu_up_write(&cgroup_threadgroup_rwsem);
3123 return ret;
3124}
3125
3126/**
3127 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3128 * @cgrp: root of the target subtree
3129 *
3130 * Because css offlining is asynchronous, userland may try to re-enable a
3131 * controller while the previous css is still around. This function grabs
3132 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3133 */
3134static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3135 __acquires(&cgroup_mutex)
3136{
3137 struct cgroup *dsct;
3138 struct cgroup_subsys_state *d_css;
3139 struct cgroup_subsys *ss;
3140 int ssid;
3141
3142restart:
3143 mutex_lock(&cgroup_mutex);
3144
3145 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3146 for_each_subsys(ss, ssid) {
3147 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3148 DEFINE_WAIT(wait);
3149
3150 if (!css || !percpu_ref_is_dying(&css->refcnt))
3151 continue;
3152
3153 cgroup_get(dsct);
3154 prepare_to_wait(&dsct->offline_waitq, &wait,
3155 TASK_UNINTERRUPTIBLE);
3156
3157 mutex_unlock(&cgroup_mutex);
3158 schedule();
3159 finish_wait(&dsct->offline_waitq, &wait);
3160
3161 cgroup_put(dsct);
3162 goto restart;
3163 }
3164 }
3165}
3166
3167/**
3168 * cgroup_save_control - save control masks of a subtree
3169 * @cgrp: root of the target subtree
3170 *
3171 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3172 * prefixed fields for @cgrp's subtree including @cgrp itself.
3173 */
3174static void cgroup_save_control(struct cgroup *cgrp)
3175{
3176 struct cgroup *dsct;
3177 struct cgroup_subsys_state *d_css;
3178
3179 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3180 dsct->old_subtree_control = dsct->subtree_control;
3181 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3182 }
3183}
3184
3185/**
3186 * cgroup_propagate_control - refresh control masks of a subtree
3187 * @cgrp: root of the target subtree
3188 *
3189 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3190 * ->subtree_control and propagate controller availability through the
3191 * subtree so that descendants don't have unavailable controllers enabled.
3192 */
3193static void cgroup_propagate_control(struct cgroup *cgrp)
3194{
3195 struct cgroup *dsct;
3196 struct cgroup_subsys_state *d_css;
3197
3198 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3199 dsct->subtree_control &= cgroup_control(dsct);
3200 dsct->subtree_ss_mask =
3201 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3202 cgroup_ss_mask(dsct));
3203 }
3204}
3205
3206/**
3207 * cgroup_restore_control - restore control masks of a subtree
3208 * @cgrp: root of the target subtree
3209 *
3210 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3211 * prefixed fields for @cgrp's subtree including @cgrp itself.
3212 */
3213static void cgroup_restore_control(struct cgroup *cgrp)
3214{
3215 struct cgroup *dsct;
3216 struct cgroup_subsys_state *d_css;
3217
3218 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3219 dsct->subtree_control = dsct->old_subtree_control;
3220 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3221 }
3222}
3223
3224static bool css_visible(struct cgroup_subsys_state *css)
3225{
3226 struct cgroup_subsys *ss = css->ss;
3227 struct cgroup *cgrp = css->cgroup;
3228
3229 if (cgroup_control(cgrp) & (1 << ss->id))
3230 return true;
3231 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3232 return false;
3233 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3234}
3235
3236/**
3237 * cgroup_apply_control_enable - enable or show csses according to control
3238 * @cgrp: root of the target subtree
3239 *
3240 * Walk @cgrp's subtree and create new csses or make the existing ones
3241 * visible. A css is created invisible if it's being implicitly enabled
3242 * through dependency. An invisible css is made visible when the userland
3243 * explicitly enables it.
3244 *
3245 * Returns 0 on success, -errno on failure. On failure, csses which have
3246 * been processed already aren't cleaned up. The caller is responsible for
3247 * cleaning up with cgroup_apply_control_disble().
3248 */
3249static int cgroup_apply_control_enable(struct cgroup *cgrp)
3250{
3251 struct cgroup *dsct;
3252 struct cgroup_subsys_state *d_css;
3253 struct cgroup_subsys *ss;
3254 int ssid, ret;
3255
3256 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3257 for_each_subsys(ss, ssid) {
3258 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3259
3260 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3261
3262 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3263 continue;
3264
3265 if (!css) {
3266 css = css_create(dsct, ss);
3267 if (IS_ERR(css))
3268 return PTR_ERR(css);
3269 }
3270
3271 if (css_visible(css)) {
3272 ret = css_populate_dir(css);
3273 if (ret)
3274 return ret;
3275 }
3276 }
3277 }
3278
3279 return 0;
3280}
3281
3282/**
3283 * cgroup_apply_control_disable - kill or hide csses according to control
3284 * @cgrp: root of the target subtree
3285 *
3286 * Walk @cgrp's subtree and kill and hide csses so that they match
3287 * cgroup_ss_mask() and cgroup_visible_mask().
3288 *
3289 * A css is hidden when the userland requests it to be disabled while other
3290 * subsystems are still depending on it. The css must not actively control
3291 * resources and be in the vanilla state if it's made visible again later.
3292 * Controllers which may be depended upon should provide ->css_reset() for
3293 * this purpose.
3294 */
3295static void cgroup_apply_control_disable(struct cgroup *cgrp)
3296{
3297 struct cgroup *dsct;
3298 struct cgroup_subsys_state *d_css;
3299 struct cgroup_subsys *ss;
3300 int ssid;
3301
3302 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3303 for_each_subsys(ss, ssid) {
3304 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3305
3306 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3307
3308 if (!css)
3309 continue;
3310
3311 if (css->parent &&
3312 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3313 kill_css(css);
3314 } else if (!css_visible(css)) {
3315 css_clear_dir(css);
3316 if (ss->css_reset)
3317 ss->css_reset(css);
3318 }
3319 }
3320 }
3321}
3322
3323/**
3324 * cgroup_apply_control - apply control mask updates to the subtree
3325 * @cgrp: root of the target subtree
3326 *
3327 * subsystems can be enabled and disabled in a subtree using the following
3328 * steps.
3329 *
3330 * 1. Call cgroup_save_control() to stash the current state.
3331 * 2. Update ->subtree_control masks in the subtree as desired.
3332 * 3. Call cgroup_apply_control() to apply the changes.
3333 * 4. Optionally perform other related operations.
3334 * 5. Call cgroup_finalize_control() to finish up.
3335 *
3336 * This function implements step 3 and propagates the mask changes
3337 * throughout @cgrp's subtree, updates csses accordingly and perform
3338 * process migrations.
3339 */
3340static int cgroup_apply_control(struct cgroup *cgrp)
3341{
3342 int ret;
3343
3344 cgroup_propagate_control(cgrp);
3345
3346 ret = cgroup_apply_control_enable(cgrp);
3347 if (ret)
3348 return ret;
3349
3350 /*
3351 * At this point, cgroup_e_css() results reflect the new csses
3352 * making the following cgroup_update_dfl_csses() properly update
3353 * css associations of all tasks in the subtree.
3354 */
3355 ret = cgroup_update_dfl_csses(cgrp);
3356 if (ret)
3357 return ret;
3358
3359 return 0;
3360}
3361
3362/**
3363 * cgroup_finalize_control - finalize control mask update
3364 * @cgrp: root of the target subtree
3365 * @ret: the result of the update
3366 *
3367 * Finalize control mask update. See cgroup_apply_control() for more info.
3368 */
3369static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3370{
3371 if (ret) {
3372 cgroup_restore_control(cgrp);
3373 cgroup_propagate_control(cgrp);
3374 }
3375
3376 cgroup_apply_control_disable(cgrp);
3377}
3378
3379/* change the enabled child controllers for a cgroup in the default hierarchy */
3380static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3381 char *buf, size_t nbytes,
3382 loff_t off)
3383{
3384 u16 enable = 0, disable = 0;
3385 struct cgroup *cgrp, *child;
3386 struct cgroup_subsys *ss;
3387 char *tok;
3388 int ssid, ret;
3389
3390 /*
3391 * Parse input - space separated list of subsystem names prefixed
3392 * with either + or -.
3393 */
3394 buf = strstrip(buf);
3395 while ((tok = strsep(&buf, " "))) {
3396 if (tok[0] == '\0')
3397 continue;
3398 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3399 if (!cgroup_ssid_enabled(ssid) ||
3400 strcmp(tok + 1, ss->name))
3401 continue;
3402
3403 if (*tok == '+') {
3404 enable |= 1 << ssid;
3405 disable &= ~(1 << ssid);
3406 } else if (*tok == '-') {
3407 disable |= 1 << ssid;
3408 enable &= ~(1 << ssid);
3409 } else {
3410 return -EINVAL;
3411 }
3412 break;
3413 } while_each_subsys_mask();
3414 if (ssid == CGROUP_SUBSYS_COUNT)
3415 return -EINVAL;
3416 }
3417
3418 cgrp = cgroup_kn_lock_live(of->kn, true);
3419 if (!cgrp)
3420 return -ENODEV;
3421
3422 for_each_subsys(ss, ssid) {
3423 if (enable & (1 << ssid)) {
3424 if (cgrp->subtree_control & (1 << ssid)) {
3425 enable &= ~(1 << ssid);
3426 continue;
3427 }
3428
3429 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3430 ret = -ENOENT;
3431 goto out_unlock;
3432 }
3433 } else if (disable & (1 << ssid)) {
3434 if (!(cgrp->subtree_control & (1 << ssid))) {
3435 disable &= ~(1 << ssid);
3436 continue;
3437 }
3438
3439 /* a child has it enabled? */
3440 cgroup_for_each_live_child(child, cgrp) {
3441 if (child->subtree_control & (1 << ssid)) {
3442 ret = -EBUSY;
3443 goto out_unlock;
3444 }
3445 }
3446 }
3447 }
3448
3449 if (!enable && !disable) {
3450 ret = 0;
3451 goto out_unlock;
3452 }
3453
3454 /*
3455 * Except for the root, subtree_control must be zero for a cgroup
3456 * with tasks so that child cgroups don't compete against tasks.
3457 */
3458 if (enable && cgroup_parent(cgrp)) {
3459 struct cgrp_cset_link *link;
3460
3461 /*
3462 * Because namespaces pin csets too, @cgrp->cset_links
3463 * might not be empty even when @cgrp is empty. Walk and
3464 * verify each cset.
3465 */
3466 spin_lock_irq(&css_set_lock);
3467
3468 ret = 0;
3469 list_for_each_entry(link, &cgrp->cset_links, cset_link) {
3470 if (css_set_populated(link->cset)) {
3471 ret = -EBUSY;
3472 break;
3473 }
3474 }
3475
3476 spin_unlock_irq(&css_set_lock);
3477
3478 if (ret)
3479 goto out_unlock;
3480 }
3481
3482 /* save and update control masks and prepare csses */
3483 cgroup_save_control(cgrp);
3484
3485 cgrp->subtree_control |= enable;
3486 cgrp->subtree_control &= ~disable;
3487
3488 ret = cgroup_apply_control(cgrp);
3489
3490 cgroup_finalize_control(cgrp, ret);
3491
3492 kernfs_activate(cgrp->kn);
3493 ret = 0;
3494out_unlock:
3495 cgroup_kn_unlock(of->kn);
3496 return ret ?: nbytes;
3497}
3498
3499static int cgroup_events_show(struct seq_file *seq, void *v)
3500{
3501 seq_printf(seq, "populated %d\n",
3502 cgroup_is_populated(seq_css(seq)->cgroup));
3503 return 0;
3504}
3505
3506static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3507 size_t nbytes, loff_t off)
3508{
3509 struct cgroup *cgrp = of->kn->parent->priv;
3510 struct cftype *cft = of->kn->priv;
3511 struct cgroup_subsys_state *css;
3512 int ret;
3513
3514 if (cft->write)
3515 return cft->write(of, buf, nbytes, off);
3516
3517 /*
3518 * kernfs guarantees that a file isn't deleted with operations in
3519 * flight, which means that the matching css is and stays alive and
3520 * doesn't need to be pinned. The RCU locking is not necessary
3521 * either. It's just for the convenience of using cgroup_css().
3522 */
3523 rcu_read_lock();
3524 css = cgroup_css(cgrp, cft->ss);
3525 rcu_read_unlock();
3526
3527 if (cft->write_u64) {
3528 unsigned long long v;
3529 ret = kstrtoull(buf, 0, &v);
3530 if (!ret)
3531 ret = cft->write_u64(css, cft, v);
3532 } else if (cft->write_s64) {
3533 long long v;
3534 ret = kstrtoll(buf, 0, &v);
3535 if (!ret)
3536 ret = cft->write_s64(css, cft, v);
3537 } else {
3538 ret = -EINVAL;
3539 }
3540
3541 return ret ?: nbytes;
3542}
3543
3544static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3545{
3546 return seq_cft(seq)->seq_start(seq, ppos);
3547}
3548
3549static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3550{
3551 return seq_cft(seq)->seq_next(seq, v, ppos);
3552}
3553
3554static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3555{
3556 seq_cft(seq)->seq_stop(seq, v);
3557}
3558
3559static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3560{
3561 struct cftype *cft = seq_cft(m);
3562 struct cgroup_subsys_state *css = seq_css(m);
3563
3564 if (cft->seq_show)
3565 return cft->seq_show(m, arg);
3566
3567 if (cft->read_u64)
3568 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3569 else if (cft->read_s64)
3570 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3571 else
3572 return -EINVAL;
3573 return 0;
3574}
3575
3576static struct kernfs_ops cgroup_kf_single_ops = {
3577 .atomic_write_len = PAGE_SIZE,
3578 .write = cgroup_file_write,
3579 .seq_show = cgroup_seqfile_show,
3580};
3581
3582static struct kernfs_ops cgroup_kf_ops = {
3583 .atomic_write_len = PAGE_SIZE,
3584 .write = cgroup_file_write,
3585 .seq_start = cgroup_seqfile_start,
3586 .seq_next = cgroup_seqfile_next,
3587 .seq_stop = cgroup_seqfile_stop,
3588 .seq_show = cgroup_seqfile_show,
3589};
3590
3591/*
3592 * cgroup_rename - Only allow simple rename of directories in place.
3593 */
3594static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3595 const char *new_name_str)
3596{
3597 struct cgroup *cgrp = kn->priv;
3598 int ret;
3599
3600 if (kernfs_type(kn) != KERNFS_DIR)
3601 return -ENOTDIR;
3602 if (kn->parent != new_parent)
3603 return -EIO;
3604
3605 /*
3606 * This isn't a proper migration and its usefulness is very
3607 * limited. Disallow on the default hierarchy.
3608 */
3609 if (cgroup_on_dfl(cgrp))
3610 return -EPERM;
3611
3612 /*
3613 * We're gonna grab cgroup_mutex which nests outside kernfs
3614 * active_ref. kernfs_rename() doesn't require active_ref
3615 * protection. Break them before grabbing cgroup_mutex.
3616 */
3617 kernfs_break_active_protection(new_parent);
3618 kernfs_break_active_protection(kn);
3619
3620 mutex_lock(&cgroup_mutex);
3621
3622 ret = kernfs_rename(kn, new_parent, new_name_str);
3623 if (!ret)
3624 trace_cgroup_rename(cgrp);
3625
3626 mutex_unlock(&cgroup_mutex);
3627
3628 kernfs_unbreak_active_protection(kn);
3629 kernfs_unbreak_active_protection(new_parent);
3630 return ret;
3631}
3632
3633/* set uid and gid of cgroup dirs and files to that of the creator */
3634static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3635{
3636 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3637 .ia_uid = current_fsuid(),
3638 .ia_gid = current_fsgid(), };
3639
3640 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3641 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3642 return 0;
3643
3644 return kernfs_setattr(kn, &iattr);
3645}
3646
3647static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3648 struct cftype *cft)
3649{
3650 char name[CGROUP_FILE_NAME_MAX];
3651 struct kernfs_node *kn;
3652 struct lock_class_key *key = NULL;
3653 int ret;
3654
3655#ifdef CONFIG_DEBUG_LOCK_ALLOC
3656 key = &cft->lockdep_key;
3657#endif
3658 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3659 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3660 NULL, key);
3661 if (IS_ERR(kn))
3662 return PTR_ERR(kn);
3663
3664 ret = cgroup_kn_set_ugid(kn);
3665 if (ret) {
3666 kernfs_remove(kn);
3667 return ret;
3668 }
3669
3670 if (cft->file_offset) {
3671 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3672
3673 spin_lock_irq(&cgroup_file_kn_lock);
3674 cfile->kn = kn;
3675 spin_unlock_irq(&cgroup_file_kn_lock);
3676 }
3677
3678 return 0;
3679}
3680
3681/**
3682 * cgroup_addrm_files - add or remove files to a cgroup directory
3683 * @css: the target css
3684 * @cgrp: the target cgroup (usually css->cgroup)
3685 * @cfts: array of cftypes to be added
3686 * @is_add: whether to add or remove
3687 *
3688 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3689 * For removals, this function never fails.
3690 */
3691static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3692 struct cgroup *cgrp, struct cftype cfts[],
3693 bool is_add)
3694{
3695 struct cftype *cft, *cft_end = NULL;
3696 int ret = 0;
3697
3698 lockdep_assert_held(&cgroup_mutex);
3699
3700restart:
3701 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3702 /* does cft->flags tell us to skip this file on @cgrp? */
3703 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3704 continue;
3705 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3706 continue;
3707 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3708 continue;
3709 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3710 continue;
3711
3712 if (is_add) {
3713 ret = cgroup_add_file(css, cgrp, cft);
3714 if (ret) {
3715 pr_warn("%s: failed to add %s, err=%d\n",
3716 __func__, cft->name, ret);
3717 cft_end = cft;
3718 is_add = false;
3719 goto restart;
3720 }
3721 } else {
3722 cgroup_rm_file(cgrp, cft);
3723 }
3724 }
3725 return ret;
3726}
3727
3728static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3729{
3730 LIST_HEAD(pending);
3731 struct cgroup_subsys *ss = cfts[0].ss;
3732 struct cgroup *root = &ss->root->cgrp;
3733 struct cgroup_subsys_state *css;
3734 int ret = 0;
3735
3736 lockdep_assert_held(&cgroup_mutex);
3737
3738 /* add/rm files for all cgroups created before */
3739 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3740 struct cgroup *cgrp = css->cgroup;
3741
3742 if (!(css->flags & CSS_VISIBLE))
3743 continue;
3744
3745 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3746 if (ret)
3747 break;
3748 }
3749
3750 if (is_add && !ret)
3751 kernfs_activate(root->kn);
3752 return ret;
3753}
3754
3755static void cgroup_exit_cftypes(struct cftype *cfts)
3756{
3757 struct cftype *cft;
3758
3759 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3760 /* free copy for custom atomic_write_len, see init_cftypes() */
3761 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3762 kfree(cft->kf_ops);
3763 cft->kf_ops = NULL;
3764 cft->ss = NULL;
3765
3766 /* revert flags set by cgroup core while adding @cfts */
3767 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3768 }
3769}
3770
3771static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3772{
3773 struct cftype *cft;
3774
3775 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3776 struct kernfs_ops *kf_ops;
3777
3778 WARN_ON(cft->ss || cft->kf_ops);
3779
3780 if (cft->seq_start)
3781 kf_ops = &cgroup_kf_ops;
3782 else
3783 kf_ops = &cgroup_kf_single_ops;
3784
3785 /*
3786 * Ugh... if @cft wants a custom max_write_len, we need to
3787 * make a copy of kf_ops to set its atomic_write_len.
3788 */
3789 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3790 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3791 if (!kf_ops) {
3792 cgroup_exit_cftypes(cfts);
3793 return -ENOMEM;
3794 }
3795 kf_ops->atomic_write_len = cft->max_write_len;
3796 }
3797
3798 cft->kf_ops = kf_ops;
3799 cft->ss = ss;
3800 }
3801
3802 return 0;
3803}
3804
3805static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3806{
3807 lockdep_assert_held(&cgroup_mutex);
3808
3809 if (!cfts || !cfts[0].ss)
3810 return -ENOENT;
3811
3812 list_del(&cfts->node);
3813 cgroup_apply_cftypes(cfts, false);
3814 cgroup_exit_cftypes(cfts);
3815 return 0;
3816}
3817
3818/**
3819 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3820 * @cfts: zero-length name terminated array of cftypes
3821 *
3822 * Unregister @cfts. Files described by @cfts are removed from all
3823 * existing cgroups and all future cgroups won't have them either. This
3824 * function can be called anytime whether @cfts' subsys is attached or not.
3825 *
3826 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3827 * registered.
3828 */
3829int cgroup_rm_cftypes(struct cftype *cfts)
3830{
3831 int ret;
3832
3833 mutex_lock(&cgroup_mutex);
3834 ret = cgroup_rm_cftypes_locked(cfts);
3835 mutex_unlock(&cgroup_mutex);
3836 return ret;
3837}
3838
3839/**
3840 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3841 * @ss: target cgroup subsystem
3842 * @cfts: zero-length name terminated array of cftypes
3843 *
3844 * Register @cfts to @ss. Files described by @cfts are created for all
3845 * existing cgroups to which @ss is attached and all future cgroups will
3846 * have them too. This function can be called anytime whether @ss is
3847 * attached or not.
3848 *
3849 * Returns 0 on successful registration, -errno on failure. Note that this
3850 * function currently returns 0 as long as @cfts registration is successful
3851 * even if some file creation attempts on existing cgroups fail.
3852 */
3853static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3854{
3855 int ret;
3856
3857 if (!cgroup_ssid_enabled(ss->id))
3858 return 0;
3859
3860 if (!cfts || cfts[0].name[0] == '\0')
3861 return 0;
3862
3863 ret = cgroup_init_cftypes(ss, cfts);
3864 if (ret)
3865 return ret;
3866
3867 mutex_lock(&cgroup_mutex);
3868
3869 list_add_tail(&cfts->node, &ss->cfts);
3870 ret = cgroup_apply_cftypes(cfts, true);
3871 if (ret)
3872 cgroup_rm_cftypes_locked(cfts);
3873
3874 mutex_unlock(&cgroup_mutex);
3875 return ret;
3876}
3877
3878/**
3879 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3880 * @ss: target cgroup subsystem
3881 * @cfts: zero-length name terminated array of cftypes
3882 *
3883 * Similar to cgroup_add_cftypes() but the added files are only used for
3884 * the default hierarchy.
3885 */
3886int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3887{
3888 struct cftype *cft;
3889
3890 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3891 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3892 return cgroup_add_cftypes(ss, cfts);
3893}
3894
3895/**
3896 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3897 * @ss: target cgroup subsystem
3898 * @cfts: zero-length name terminated array of cftypes
3899 *
3900 * Similar to cgroup_add_cftypes() but the added files are only used for
3901 * the legacy hierarchies.
3902 */
3903int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3904{
3905 struct cftype *cft;
3906
3907 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3908 cft->flags |= __CFTYPE_NOT_ON_DFL;
3909 return cgroup_add_cftypes(ss, cfts);
3910}
3911
3912/**
3913 * cgroup_file_notify - generate a file modified event for a cgroup_file
3914 * @cfile: target cgroup_file
3915 *
3916 * @cfile must have been obtained by setting cftype->file_offset.
3917 */
3918void cgroup_file_notify(struct cgroup_file *cfile)
3919{
3920 unsigned long flags;
3921
3922 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3923 if (cfile->kn)
3924 kernfs_notify(cfile->kn);
3925 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3926}
3927
3928/**
3929 * cgroup_task_count - count the number of tasks in a cgroup.
3930 * @cgrp: the cgroup in question
3931 *
3932 * Return the number of tasks in the cgroup. The returned number can be
3933 * higher than the actual number of tasks due to css_set references from
3934 * namespace roots and temporary usages.
3935 */
3936static int cgroup_task_count(const struct cgroup *cgrp)
3937{
3938 int count = 0;
3939 struct cgrp_cset_link *link;
3940
3941 spin_lock_irq(&css_set_lock);
3942 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3943 count += atomic_read(&link->cset->refcount);
3944 spin_unlock_irq(&css_set_lock);
3945 return count;
3946}
3947
3948/**
3949 * css_next_child - find the next child of a given css
3950 * @pos: the current position (%NULL to initiate traversal)
3951 * @parent: css whose children to walk
3952 *
3953 * This function returns the next child of @parent and should be called
3954 * under either cgroup_mutex or RCU read lock. The only requirement is
3955 * that @parent and @pos are accessible. The next sibling is guaranteed to
3956 * be returned regardless of their states.
3957 *
3958 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3959 * css which finished ->css_online() is guaranteed to be visible in the
3960 * future iterations and will stay visible until the last reference is put.
3961 * A css which hasn't finished ->css_online() or already finished
3962 * ->css_offline() may show up during traversal. It's each subsystem's
3963 * responsibility to synchronize against on/offlining.
3964 */
3965struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3966 struct cgroup_subsys_state *parent)
3967{
3968 struct cgroup_subsys_state *next;
3969
3970 cgroup_assert_mutex_or_rcu_locked();
3971
3972 /*
3973 * @pos could already have been unlinked from the sibling list.
3974 * Once a cgroup is removed, its ->sibling.next is no longer
3975 * updated when its next sibling changes. CSS_RELEASED is set when
3976 * @pos is taken off list, at which time its next pointer is valid,
3977 * and, as releases are serialized, the one pointed to by the next
3978 * pointer is guaranteed to not have started release yet. This
3979 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3980 * critical section, the one pointed to by its next pointer is
3981 * guaranteed to not have finished its RCU grace period even if we
3982 * have dropped rcu_read_lock() inbetween iterations.
3983 *
3984 * If @pos has CSS_RELEASED set, its next pointer can't be
3985 * dereferenced; however, as each css is given a monotonically
3986 * increasing unique serial number and always appended to the
3987 * sibling list, the next one can be found by walking the parent's
3988 * children until the first css with higher serial number than
3989 * @pos's. While this path can be slower, it happens iff iteration
3990 * races against release and the race window is very small.
3991 */
3992 if (!pos) {
3993 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3994 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3995 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3996 } else {
3997 list_for_each_entry_rcu(next, &parent->children, sibling)
3998 if (next->serial_nr > pos->serial_nr)
3999 break;
4000 }
4001
4002 /*
4003 * @next, if not pointing to the head, can be dereferenced and is
4004 * the next sibling.
4005 */
4006 if (&next->sibling != &parent->children)
4007 return next;
4008 return NULL;
4009}
4010
4011/**
4012 * css_next_descendant_pre - find the next descendant for pre-order walk
4013 * @pos: the current position (%NULL to initiate traversal)
4014 * @root: css whose descendants to walk
4015 *
4016 * To be used by css_for_each_descendant_pre(). Find the next descendant
4017 * to visit for pre-order traversal of @root's descendants. @root is
4018 * included in the iteration and the first node to be visited.
4019 *
4020 * While this function requires cgroup_mutex or RCU read locking, it
4021 * doesn't require the whole traversal to be contained in a single critical
4022 * section. This function will return the correct next descendant as long
4023 * as both @pos and @root are accessible and @pos is a descendant of @root.
4024 *
4025 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4026 * css which finished ->css_online() is guaranteed to be visible in the
4027 * future iterations and will stay visible until the last reference is put.
4028 * A css which hasn't finished ->css_online() or already finished
4029 * ->css_offline() may show up during traversal. It's each subsystem's
4030 * responsibility to synchronize against on/offlining.
4031 */
4032struct cgroup_subsys_state *
4033css_next_descendant_pre(struct cgroup_subsys_state *pos,
4034 struct cgroup_subsys_state *root)
4035{
4036 struct cgroup_subsys_state *next;
4037
4038 cgroup_assert_mutex_or_rcu_locked();
4039
4040 /* if first iteration, visit @root */
4041 if (!pos)
4042 return root;
4043
4044 /* visit the first child if exists */
4045 next = css_next_child(NULL, pos);
4046 if (next)
4047 return next;
4048
4049 /* no child, visit my or the closest ancestor's next sibling */
4050 while (pos != root) {
4051 next = css_next_child(pos, pos->parent);
4052 if (next)
4053 return next;
4054 pos = pos->parent;
4055 }
4056
4057 return NULL;
4058}
4059
4060/**
4061 * css_rightmost_descendant - return the rightmost descendant of a css
4062 * @pos: css of interest
4063 *
4064 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4065 * is returned. This can be used during pre-order traversal to skip
4066 * subtree of @pos.
4067 *
4068 * While this function requires cgroup_mutex or RCU read locking, it
4069 * doesn't require the whole traversal to be contained in a single critical
4070 * section. This function will return the correct rightmost descendant as
4071 * long as @pos is accessible.
4072 */
4073struct cgroup_subsys_state *
4074css_rightmost_descendant(struct cgroup_subsys_state *pos)
4075{
4076 struct cgroup_subsys_state *last, *tmp;
4077
4078 cgroup_assert_mutex_or_rcu_locked();
4079
4080 do {
4081 last = pos;
4082 /* ->prev isn't RCU safe, walk ->next till the end */
4083 pos = NULL;
4084 css_for_each_child(tmp, last)
4085 pos = tmp;
4086 } while (pos);
4087
4088 return last;
4089}
4090
4091static struct cgroup_subsys_state *
4092css_leftmost_descendant(struct cgroup_subsys_state *pos)
4093{
4094 struct cgroup_subsys_state *last;
4095
4096 do {
4097 last = pos;
4098 pos = css_next_child(NULL, pos);
4099 } while (pos);
4100
4101 return last;
4102}
4103
4104/**
4105 * css_next_descendant_post - find the next descendant for post-order walk
4106 * @pos: the current position (%NULL to initiate traversal)
4107 * @root: css whose descendants to walk
4108 *
4109 * To be used by css_for_each_descendant_post(). Find the next descendant
4110 * to visit for post-order traversal of @root's descendants. @root is
4111 * included in the iteration and the last node to be visited.
4112 *
4113 * While this function requires cgroup_mutex or RCU read locking, it
4114 * doesn't require the whole traversal to be contained in a single critical
4115 * section. This function will return the correct next descendant as long
4116 * as both @pos and @cgroup are accessible and @pos is a descendant of
4117 * @cgroup.
4118 *
4119 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4120 * css which finished ->css_online() is guaranteed to be visible in the
4121 * future iterations and will stay visible until the last reference is put.
4122 * A css which hasn't finished ->css_online() or already finished
4123 * ->css_offline() may show up during traversal. It's each subsystem's
4124 * responsibility to synchronize against on/offlining.
4125 */
4126struct cgroup_subsys_state *
4127css_next_descendant_post(struct cgroup_subsys_state *pos,
4128 struct cgroup_subsys_state *root)
4129{
4130 struct cgroup_subsys_state *next;
4131
4132 cgroup_assert_mutex_or_rcu_locked();
4133
4134 /* if first iteration, visit leftmost descendant which may be @root */
4135 if (!pos)
4136 return css_leftmost_descendant(root);
4137
4138 /* if we visited @root, we're done */
4139 if (pos == root)
4140 return NULL;
4141
4142 /* if there's an unvisited sibling, visit its leftmost descendant */
4143 next = css_next_child(pos, pos->parent);
4144 if (next)
4145 return css_leftmost_descendant(next);
4146
4147 /* no sibling left, visit parent */
4148 return pos->parent;
4149}
4150
4151/**
4152 * css_has_online_children - does a css have online children
4153 * @css: the target css
4154 *
4155 * Returns %true if @css has any online children; otherwise, %false. This
4156 * function can be called from any context but the caller is responsible
4157 * for synchronizing against on/offlining as necessary.
4158 */
4159bool css_has_online_children(struct cgroup_subsys_state *css)
4160{
4161 struct cgroup_subsys_state *child;
4162 bool ret = false;
4163
4164 rcu_read_lock();
4165 css_for_each_child(child, css) {
4166 if (child->flags & CSS_ONLINE) {
4167 ret = true;
4168 break;
4169 }
4170 }
4171 rcu_read_unlock();
4172 return ret;
4173}
4174
4175/**
4176 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4177 * @it: the iterator to advance
4178 *
4179 * Advance @it to the next css_set to walk.
4180 */
4181static void css_task_iter_advance_css_set(struct css_task_iter *it)
4182{
4183 struct list_head *l = it->cset_pos;
4184 struct cgrp_cset_link *link;
4185 struct css_set *cset;
4186
4187 lockdep_assert_held(&css_set_lock);
4188
4189 /* Advance to the next non-empty css_set */
4190 do {
4191 l = l->next;
4192 if (l == it->cset_head) {
4193 it->cset_pos = NULL;
4194 it->task_pos = NULL;
4195 return;
4196 }
4197
4198 if (it->ss) {
4199 cset = container_of(l, struct css_set,
4200 e_cset_node[it->ss->id]);
4201 } else {
4202 link = list_entry(l, struct cgrp_cset_link, cset_link);
4203 cset = link->cset;
4204 }
4205 } while (!css_set_populated(cset));
4206
4207 it->cset_pos = l;
4208
4209 if (!list_empty(&cset->tasks))
4210 it->task_pos = cset->tasks.next;
4211 else
4212 it->task_pos = cset->mg_tasks.next;
4213
4214 it->tasks_head = &cset->tasks;
4215 it->mg_tasks_head = &cset->mg_tasks;
4216
4217 /*
4218 * We don't keep css_sets locked across iteration steps and thus
4219 * need to take steps to ensure that iteration can be resumed after
4220 * the lock is re-acquired. Iteration is performed at two levels -
4221 * css_sets and tasks in them.
4222 *
4223 * Once created, a css_set never leaves its cgroup lists, so a
4224 * pinned css_set is guaranteed to stay put and we can resume
4225 * iteration afterwards.
4226 *
4227 * Tasks may leave @cset across iteration steps. This is resolved
4228 * by registering each iterator with the css_set currently being
4229 * walked and making css_set_move_task() advance iterators whose
4230 * next task is leaving.
4231 */
4232 if (it->cur_cset) {
4233 list_del(&it->iters_node);
4234 put_css_set_locked(it->cur_cset);
4235 }
4236 get_css_set(cset);
4237 it->cur_cset = cset;
4238 list_add(&it->iters_node, &cset->task_iters);
4239}
4240
4241static void css_task_iter_advance(struct css_task_iter *it)
4242{
4243 struct list_head *l = it->task_pos;
4244
4245 lockdep_assert_held(&css_set_lock);
4246 WARN_ON_ONCE(!l);
4247
4248 /*
4249 * Advance iterator to find next entry. cset->tasks is consumed
4250 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4251 * next cset.
4252 */
4253 l = l->next;
4254
4255 if (l == it->tasks_head)
4256 l = it->mg_tasks_head->next;
4257
4258 if (l == it->mg_tasks_head)
4259 css_task_iter_advance_css_set(it);
4260 else
4261 it->task_pos = l;
4262}
4263
4264/**
4265 * css_task_iter_start - initiate task iteration
4266 * @css: the css to walk tasks of
4267 * @it: the task iterator to use
4268 *
4269 * Initiate iteration through the tasks of @css. The caller can call
4270 * css_task_iter_next() to walk through the tasks until the function
4271 * returns NULL. On completion of iteration, css_task_iter_end() must be
4272 * called.
4273 */
4274void css_task_iter_start(struct cgroup_subsys_state *css,
4275 struct css_task_iter *it)
4276{
4277 /* no one should try to iterate before mounting cgroups */
4278 WARN_ON_ONCE(!use_task_css_set_links);
4279
4280 memset(it, 0, sizeof(*it));
4281
4282 spin_lock_irq(&css_set_lock);
4283
4284 it->ss = css->ss;
4285
4286 if (it->ss)
4287 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4288 else
4289 it->cset_pos = &css->cgroup->cset_links;
4290
4291 it->cset_head = it->cset_pos;
4292
4293 css_task_iter_advance_css_set(it);
4294
4295 spin_unlock_irq(&css_set_lock);
4296}
4297
4298/**
4299 * css_task_iter_next - return the next task for the iterator
4300 * @it: the task iterator being iterated
4301 *
4302 * The "next" function for task iteration. @it should have been
4303 * initialized via css_task_iter_start(). Returns NULL when the iteration
4304 * reaches the end.
4305 */
4306struct task_struct *css_task_iter_next(struct css_task_iter *it)
4307{
4308 if (it->cur_task) {
4309 put_task_struct(it->cur_task);
4310 it->cur_task = NULL;
4311 }
4312
4313 spin_lock_irq(&css_set_lock);
4314
4315 if (it->task_pos) {
4316 it->cur_task = list_entry(it->task_pos, struct task_struct,
4317 cg_list);
4318 get_task_struct(it->cur_task);
4319 css_task_iter_advance(it);
4320 }
4321
4322 spin_unlock_irq(&css_set_lock);
4323
4324 return it->cur_task;
4325}
4326
4327/**
4328 * css_task_iter_end - finish task iteration
4329 * @it: the task iterator to finish
4330 *
4331 * Finish task iteration started by css_task_iter_start().
4332 */
4333void css_task_iter_end(struct css_task_iter *it)
4334{
4335 if (it->cur_cset) {
4336 spin_lock_irq(&css_set_lock);
4337 list_del(&it->iters_node);
4338 put_css_set_locked(it->cur_cset);
4339 spin_unlock_irq(&css_set_lock);
4340 }
4341
4342 if (it->cur_task)
4343 put_task_struct(it->cur_task);
4344}
4345
4346/**
4347 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4348 * @to: cgroup to which the tasks will be moved
4349 * @from: cgroup in which the tasks currently reside
4350 *
4351 * Locking rules between cgroup_post_fork() and the migration path
4352 * guarantee that, if a task is forking while being migrated, the new child
4353 * is guaranteed to be either visible in the source cgroup after the
4354 * parent's migration is complete or put into the target cgroup. No task
4355 * can slip out of migration through forking.
4356 */
4357int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4358{
4359 LIST_HEAD(preloaded_csets);
4360 struct cgrp_cset_link *link;
4361 struct css_task_iter it;
4362 struct task_struct *task;
4363 int ret;
4364
4365 if (!cgroup_may_migrate_to(to))
4366 return -EBUSY;
4367
4368 mutex_lock(&cgroup_mutex);
4369
4370 percpu_down_write(&cgroup_threadgroup_rwsem);
4371
4372 /* all tasks in @from are being moved, all csets are source */
4373 spin_lock_irq(&css_set_lock);
4374 list_for_each_entry(link, &from->cset_links, cset_link)
4375 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4376 spin_unlock_irq(&css_set_lock);
4377
4378 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4379 if (ret)
4380 goto out_err;
4381
4382 /*
4383 * Migrate tasks one-by-one until @from is empty. This fails iff
4384 * ->can_attach() fails.
4385 */
4386 do {
4387 css_task_iter_start(&from->self, &it);
4388 task = css_task_iter_next(&it);
4389 if (task)
4390 get_task_struct(task);
4391 css_task_iter_end(&it);
4392
4393 if (task) {
4394 ret = cgroup_migrate(task, false, to->root);
4395 if (!ret)
4396 trace_cgroup_transfer_tasks(to, task, false);
4397 put_task_struct(task);
4398 }
4399 } while (task && !ret);
4400out_err:
4401 cgroup_migrate_finish(&preloaded_csets);
4402 percpu_up_write(&cgroup_threadgroup_rwsem);
4403 mutex_unlock(&cgroup_mutex);
4404 return ret;
4405}
4406
4407/*
4408 * Stuff for reading the 'tasks'/'procs' files.
4409 *
4410 * Reading this file can return large amounts of data if a cgroup has
4411 * *lots* of attached tasks. So it may need several calls to read(),
4412 * but we cannot guarantee that the information we produce is correct
4413 * unless we produce it entirely atomically.
4414 *
4415 */
4416
4417/* which pidlist file are we talking about? */
4418enum cgroup_filetype {
4419 CGROUP_FILE_PROCS,
4420 CGROUP_FILE_TASKS,
4421};
4422
4423/*
4424 * A pidlist is a list of pids that virtually represents the contents of one
4425 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4426 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4427 * to the cgroup.
4428 */
4429struct cgroup_pidlist {
4430 /*
4431 * used to find which pidlist is wanted. doesn't change as long as
4432 * this particular list stays in the list.
4433 */
4434 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4435 /* array of xids */
4436 pid_t *list;
4437 /* how many elements the above list has */
4438 int length;
4439 /* each of these stored in a list by its cgroup */
4440 struct list_head links;
4441 /* pointer to the cgroup we belong to, for list removal purposes */
4442 struct cgroup *owner;
4443 /* for delayed destruction */
4444 struct delayed_work destroy_dwork;
4445};
4446
4447/*
4448 * The following two functions "fix" the issue where there are more pids
4449 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4450 * TODO: replace with a kernel-wide solution to this problem
4451 */
4452#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4453static void *pidlist_allocate(int count)
4454{
4455 if (PIDLIST_TOO_LARGE(count))
4456 return vmalloc(count * sizeof(pid_t));
4457 else
4458 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4459}
4460
4461static void pidlist_free(void *p)
4462{
4463 kvfree(p);
4464}
4465
4466/*
4467 * Used to destroy all pidlists lingering waiting for destroy timer. None
4468 * should be left afterwards.
4469 */
4470static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4471{
4472 struct cgroup_pidlist *l, *tmp_l;
4473
4474 mutex_lock(&cgrp->pidlist_mutex);
4475 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4476 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4477 mutex_unlock(&cgrp->pidlist_mutex);
4478
4479 flush_workqueue(cgroup_pidlist_destroy_wq);
4480 BUG_ON(!list_empty(&cgrp->pidlists));
4481}
4482
4483static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4484{
4485 struct delayed_work *dwork = to_delayed_work(work);
4486 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4487 destroy_dwork);
4488 struct cgroup_pidlist *tofree = NULL;
4489
4490 mutex_lock(&l->owner->pidlist_mutex);
4491
4492 /*
4493 * Destroy iff we didn't get queued again. The state won't change
4494 * as destroy_dwork can only be queued while locked.
4495 */
4496 if (!delayed_work_pending(dwork)) {
4497 list_del(&l->links);
4498 pidlist_free(l->list);
4499 put_pid_ns(l->key.ns);
4500 tofree = l;
4501 }
4502
4503 mutex_unlock(&l->owner->pidlist_mutex);
4504 kfree(tofree);
4505}
4506
4507/*
4508 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4509 * Returns the number of unique elements.
4510 */
4511static int pidlist_uniq(pid_t *list, int length)
4512{
4513 int src, dest = 1;
4514
4515 /*
4516 * we presume the 0th element is unique, so i starts at 1. trivial
4517 * edge cases first; no work needs to be done for either
4518 */
4519 if (length == 0 || length == 1)
4520 return length;
4521 /* src and dest walk down the list; dest counts unique elements */
4522 for (src = 1; src < length; src++) {
4523 /* find next unique element */
4524 while (list[src] == list[src-1]) {
4525 src++;
4526 if (src == length)
4527 goto after;
4528 }
4529 /* dest always points to where the next unique element goes */
4530 list[dest] = list[src];
4531 dest++;
4532 }
4533after:
4534 return dest;
4535}
4536
4537/*
4538 * The two pid files - task and cgroup.procs - guaranteed that the result
4539 * is sorted, which forced this whole pidlist fiasco. As pid order is
4540 * different per namespace, each namespace needs differently sorted list,
4541 * making it impossible to use, for example, single rbtree of member tasks
4542 * sorted by task pointer. As pidlists can be fairly large, allocating one
4543 * per open file is dangerous, so cgroup had to implement shared pool of
4544 * pidlists keyed by cgroup and namespace.
4545 *
4546 * All this extra complexity was caused by the original implementation
4547 * committing to an entirely unnecessary property. In the long term, we
4548 * want to do away with it. Explicitly scramble sort order if on the
4549 * default hierarchy so that no such expectation exists in the new
4550 * interface.
4551 *
4552 * Scrambling is done by swapping every two consecutive bits, which is
4553 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4554 */
4555static pid_t pid_fry(pid_t pid)
4556{
4557 unsigned a = pid & 0x55555555;
4558 unsigned b = pid & 0xAAAAAAAA;
4559
4560 return (a << 1) | (b >> 1);
4561}
4562
4563static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4564{
4565 if (cgroup_on_dfl(cgrp))
4566 return pid_fry(pid);
4567 else
4568 return pid;
4569}
4570
4571static int cmppid(const void *a, const void *b)
4572{
4573 return *(pid_t *)a - *(pid_t *)b;
4574}
4575
4576static int fried_cmppid(const void *a, const void *b)
4577{
4578 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4579}
4580
4581static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4582 enum cgroup_filetype type)
4583{
4584 struct cgroup_pidlist *l;
4585 /* don't need task_nsproxy() if we're looking at ourself */
4586 struct pid_namespace *ns = task_active_pid_ns(current);
4587
4588 lockdep_assert_held(&cgrp->pidlist_mutex);
4589
4590 list_for_each_entry(l, &cgrp->pidlists, links)
4591 if (l->key.type == type && l->key.ns == ns)
4592 return l;
4593 return NULL;
4594}
4595
4596/*
4597 * find the appropriate pidlist for our purpose (given procs vs tasks)
4598 * returns with the lock on that pidlist already held, and takes care
4599 * of the use count, or returns NULL with no locks held if we're out of
4600 * memory.
4601 */
4602static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4603 enum cgroup_filetype type)
4604{
4605 struct cgroup_pidlist *l;
4606
4607 lockdep_assert_held(&cgrp->pidlist_mutex);
4608
4609 l = cgroup_pidlist_find(cgrp, type);
4610 if (l)
4611 return l;
4612
4613 /* entry not found; create a new one */
4614 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4615 if (!l)
4616 return l;
4617
4618 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4619 l->key.type = type;
4620 /* don't need task_nsproxy() if we're looking at ourself */
4621 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4622 l->owner = cgrp;
4623 list_add(&l->links, &cgrp->pidlists);
4624 return l;
4625}
4626
4627/*
4628 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4629 */
4630static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4631 struct cgroup_pidlist **lp)
4632{
4633 pid_t *array;
4634 int length;
4635 int pid, n = 0; /* used for populating the array */
4636 struct css_task_iter it;
4637 struct task_struct *tsk;
4638 struct cgroup_pidlist *l;
4639
4640 lockdep_assert_held(&cgrp->pidlist_mutex);
4641
4642 /*
4643 * If cgroup gets more users after we read count, we won't have
4644 * enough space - tough. This race is indistinguishable to the
4645 * caller from the case that the additional cgroup users didn't
4646 * show up until sometime later on.
4647 */
4648 length = cgroup_task_count(cgrp);
4649 array = pidlist_allocate(length);
4650 if (!array)
4651 return -ENOMEM;
4652 /* now, populate the array */
4653 css_task_iter_start(&cgrp->self, &it);
4654 while ((tsk = css_task_iter_next(&it))) {
4655 if (unlikely(n == length))
4656 break;
4657 /* get tgid or pid for procs or tasks file respectively */
4658 if (type == CGROUP_FILE_PROCS)
4659 pid = task_tgid_vnr(tsk);
4660 else
4661 pid = task_pid_vnr(tsk);
4662 if (pid > 0) /* make sure to only use valid results */
4663 array[n++] = pid;
4664 }
4665 css_task_iter_end(&it);
4666 length = n;
4667 /* now sort & (if procs) strip out duplicates */
4668 if (cgroup_on_dfl(cgrp))
4669 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4670 else
4671 sort(array, length, sizeof(pid_t), cmppid, NULL);
4672 if (type == CGROUP_FILE_PROCS)
4673 length = pidlist_uniq(array, length);
4674
4675 l = cgroup_pidlist_find_create(cgrp, type);
4676 if (!l) {
4677 pidlist_free(array);
4678 return -ENOMEM;
4679 }
4680
4681 /* store array, freeing old if necessary */
4682 pidlist_free(l->list);
4683 l->list = array;
4684 l->length = length;
4685 *lp = l;
4686 return 0;
4687}
4688
4689/**
4690 * cgroupstats_build - build and fill cgroupstats
4691 * @stats: cgroupstats to fill information into
4692 * @dentry: A dentry entry belonging to the cgroup for which stats have
4693 * been requested.
4694 *
4695 * Build and fill cgroupstats so that taskstats can export it to user
4696 * space.
4697 */
4698int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4699{
4700 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4701 struct cgroup *cgrp;
4702 struct css_task_iter it;
4703 struct task_struct *tsk;
4704
4705 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4706 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4707 kernfs_type(kn) != KERNFS_DIR)
4708 return -EINVAL;
4709
4710 mutex_lock(&cgroup_mutex);
4711
4712 /*
4713 * We aren't being called from kernfs and there's no guarantee on
4714 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4715 * @kn->priv is RCU safe. Let's do the RCU dancing.
4716 */
4717 rcu_read_lock();
4718 cgrp = rcu_dereference(kn->priv);
4719 if (!cgrp || cgroup_is_dead(cgrp)) {
4720 rcu_read_unlock();
4721 mutex_unlock(&cgroup_mutex);
4722 return -ENOENT;
4723 }
4724 rcu_read_unlock();
4725
4726 css_task_iter_start(&cgrp->self, &it);
4727 while ((tsk = css_task_iter_next(&it))) {
4728 switch (tsk->state) {
4729 case TASK_RUNNING:
4730 stats->nr_running++;
4731 break;
4732 case TASK_INTERRUPTIBLE:
4733 stats->nr_sleeping++;
4734 break;
4735 case TASK_UNINTERRUPTIBLE:
4736 stats->nr_uninterruptible++;
4737 break;
4738 case TASK_STOPPED:
4739 stats->nr_stopped++;
4740 break;
4741 default:
4742 if (delayacct_is_task_waiting_on_io(tsk))
4743 stats->nr_io_wait++;
4744 break;
4745 }
4746 }
4747 css_task_iter_end(&it);
4748
4749 mutex_unlock(&cgroup_mutex);
4750 return 0;
4751}
4752
4753
4754/*
4755 * seq_file methods for the tasks/procs files. The seq_file position is the
4756 * next pid to display; the seq_file iterator is a pointer to the pid
4757 * in the cgroup->l->list array.
4758 */
4759
4760static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4761{
4762 /*
4763 * Initially we receive a position value that corresponds to
4764 * one more than the last pid shown (or 0 on the first call or
4765 * after a seek to the start). Use a binary-search to find the
4766 * next pid to display, if any
4767 */
4768 struct kernfs_open_file *of = s->private;
4769 struct cgroup *cgrp = seq_css(s)->cgroup;
4770 struct cgroup_pidlist *l;
4771 enum cgroup_filetype type = seq_cft(s)->private;
4772 int index = 0, pid = *pos;
4773 int *iter, ret;
4774
4775 mutex_lock(&cgrp->pidlist_mutex);
4776
4777 /*
4778 * !NULL @of->priv indicates that this isn't the first start()
4779 * after open. If the matching pidlist is around, we can use that.
4780 * Look for it. Note that @of->priv can't be used directly. It
4781 * could already have been destroyed.
4782 */
4783 if (of->priv)
4784 of->priv = cgroup_pidlist_find(cgrp, type);
4785
4786 /*
4787 * Either this is the first start() after open or the matching
4788 * pidlist has been destroyed inbetween. Create a new one.
4789 */
4790 if (!of->priv) {
4791 ret = pidlist_array_load(cgrp, type,
4792 (struct cgroup_pidlist **)&of->priv);
4793 if (ret)
4794 return ERR_PTR(ret);
4795 }
4796 l = of->priv;
4797
4798 if (pid) {
4799 int end = l->length;
4800
4801 while (index < end) {
4802 int mid = (index + end) / 2;
4803 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4804 index = mid;
4805 break;
4806 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4807 index = mid + 1;
4808 else
4809 end = mid;
4810 }
4811 }
4812 /* If we're off the end of the array, we're done */
4813 if (index >= l->length)
4814 return NULL;
4815 /* Update the abstract position to be the actual pid that we found */
4816 iter = l->list + index;
4817 *pos = cgroup_pid_fry(cgrp, *iter);
4818 return iter;
4819}
4820
4821static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4822{
4823 struct kernfs_open_file *of = s->private;
4824 struct cgroup_pidlist *l = of->priv;
4825
4826 if (l)
4827 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4828 CGROUP_PIDLIST_DESTROY_DELAY);
4829 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4830}
4831
4832static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4833{
4834 struct kernfs_open_file *of = s->private;
4835 struct cgroup_pidlist *l = of->priv;
4836 pid_t *p = v;
4837 pid_t *end = l->list + l->length;
4838 /*
4839 * Advance to the next pid in the array. If this goes off the
4840 * end, we're done
4841 */
4842 p++;
4843 if (p >= end) {
4844 return NULL;
4845 } else {
4846 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4847 return p;
4848 }
4849}
4850
4851static int cgroup_pidlist_show(struct seq_file *s, void *v)
4852{
4853 seq_printf(s, "%d\n", *(int *)v);
4854
4855 return 0;
4856}
4857
4858static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4859 struct cftype *cft)
4860{
4861 return notify_on_release(css->cgroup);
4862}
4863
4864static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4865 struct cftype *cft, u64 val)
4866{
4867 if (val)
4868 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4869 else
4870 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4871 return 0;
4872}
4873
4874static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4875 struct cftype *cft)
4876{
4877 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4878}
4879
4880static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4881 struct cftype *cft, u64 val)
4882{
4883 if (val)
4884 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4885 else
4886 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4887 return 0;
4888}
4889
4890/* cgroup core interface files for the default hierarchy */
4891static struct cftype cgroup_dfl_base_files[] = {
4892 {
4893 .name = "cgroup.procs",
4894 .file_offset = offsetof(struct cgroup, procs_file),
4895 .seq_start = cgroup_pidlist_start,
4896 .seq_next = cgroup_pidlist_next,
4897 .seq_stop = cgroup_pidlist_stop,
4898 .seq_show = cgroup_pidlist_show,
4899 .private = CGROUP_FILE_PROCS,
4900 .write = cgroup_procs_write,
4901 },
4902 {
4903 .name = "cgroup.controllers",
4904 .seq_show = cgroup_controllers_show,
4905 },
4906 {
4907 .name = "cgroup.subtree_control",
4908 .seq_show = cgroup_subtree_control_show,
4909 .write = cgroup_subtree_control_write,
4910 },
4911 {
4912 .name = "cgroup.events",
4913 .flags = CFTYPE_NOT_ON_ROOT,
4914 .file_offset = offsetof(struct cgroup, events_file),
4915 .seq_show = cgroup_events_show,
4916 },
4917 { } /* terminate */
4918};
4919
4920/* cgroup core interface files for the legacy hierarchies */
4921static struct cftype cgroup_legacy_base_files[] = {
4922 {
4923 .name = "cgroup.procs",
4924 .seq_start = cgroup_pidlist_start,
4925 .seq_next = cgroup_pidlist_next,
4926 .seq_stop = cgroup_pidlist_stop,
4927 .seq_show = cgroup_pidlist_show,
4928 .private = CGROUP_FILE_PROCS,
4929 .write = cgroup_procs_write,
4930 },
4931 {
4932 .name = "cgroup.clone_children",
4933 .read_u64 = cgroup_clone_children_read,
4934 .write_u64 = cgroup_clone_children_write,
4935 },
4936 {
4937 .name = "cgroup.sane_behavior",
4938 .flags = CFTYPE_ONLY_ON_ROOT,
4939 .seq_show = cgroup_sane_behavior_show,
4940 },
4941 {
4942 .name = "tasks",
4943 .seq_start = cgroup_pidlist_start,
4944 .seq_next = cgroup_pidlist_next,
4945 .seq_stop = cgroup_pidlist_stop,
4946 .seq_show = cgroup_pidlist_show,
4947 .private = CGROUP_FILE_TASKS,
4948 .write = cgroup_tasks_write,
4949 },
4950 {
4951 .name = "notify_on_release",
4952 .read_u64 = cgroup_read_notify_on_release,
4953 .write_u64 = cgroup_write_notify_on_release,
4954 },
4955 {
4956 .name = "release_agent",
4957 .flags = CFTYPE_ONLY_ON_ROOT,
4958 .seq_show = cgroup_release_agent_show,
4959 .write = cgroup_release_agent_write,
4960 .max_write_len = PATH_MAX - 1,
4961 },
4962 { } /* terminate */
4963};
4964
4965/*
4966 * css destruction is four-stage process.
4967 *
4968 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4969 * Implemented in kill_css().
4970 *
4971 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4972 * and thus css_tryget_online() is guaranteed to fail, the css can be
4973 * offlined by invoking offline_css(). After offlining, the base ref is
4974 * put. Implemented in css_killed_work_fn().
4975 *
4976 * 3. When the percpu_ref reaches zero, the only possible remaining
4977 * accessors are inside RCU read sections. css_release() schedules the
4978 * RCU callback.
4979 *
4980 * 4. After the grace period, the css can be freed. Implemented in
4981 * css_free_work_fn().
4982 *
4983 * It is actually hairier because both step 2 and 4 require process context
4984 * and thus involve punting to css->destroy_work adding two additional
4985 * steps to the already complex sequence.
4986 */
4987static void css_free_work_fn(struct work_struct *work)
4988{
4989 struct cgroup_subsys_state *css =
4990 container_of(work, struct cgroup_subsys_state, destroy_work);
4991 struct cgroup_subsys *ss = css->ss;
4992 struct cgroup *cgrp = css->cgroup;
4993
4994 percpu_ref_exit(&css->refcnt);
4995
4996 if (ss) {
4997 /* css free path */
4998 struct cgroup_subsys_state *parent = css->parent;
4999 int id = css->id;
5000
5001 ss->css_free(css);
5002 cgroup_idr_remove(&ss->css_idr, id);
5003 cgroup_put(cgrp);
5004
5005 if (parent)
5006 css_put(parent);
5007 } else {
5008 /* cgroup free path */
5009 atomic_dec(&cgrp->root->nr_cgrps);
5010 cgroup_pidlist_destroy_all(cgrp);
5011 cancel_work_sync(&cgrp->release_agent_work);
5012
5013 if (cgroup_parent(cgrp)) {
5014 /*
5015 * We get a ref to the parent, and put the ref when
5016 * this cgroup is being freed, so it's guaranteed
5017 * that the parent won't be destroyed before its
5018 * children.
5019 */
5020 cgroup_put(cgroup_parent(cgrp));
5021 kernfs_put(cgrp->kn);
5022 kfree(cgrp);
5023 } else {
5024 /*
5025 * This is root cgroup's refcnt reaching zero,
5026 * which indicates that the root should be
5027 * released.
5028 */
5029 cgroup_destroy_root(cgrp->root);
5030 }
5031 }
5032}
5033
5034static void css_free_rcu_fn(struct rcu_head *rcu_head)
5035{
5036 struct cgroup_subsys_state *css =
5037 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
5038
5039 INIT_WORK(&css->destroy_work, css_free_work_fn);
5040 queue_work(cgroup_destroy_wq, &css->destroy_work);
5041}
5042
5043static void css_release_work_fn(struct work_struct *work)
5044{
5045 struct cgroup_subsys_state *css =
5046 container_of(work, struct cgroup_subsys_state, destroy_work);
5047 struct cgroup_subsys *ss = css->ss;
5048 struct cgroup *cgrp = css->cgroup;
5049
5050 mutex_lock(&cgroup_mutex);
5051
5052 css->flags |= CSS_RELEASED;
5053 list_del_rcu(&css->sibling);
5054
5055 if (ss) {
5056 /* css release path */
5057 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5058 if (ss->css_released)
5059 ss->css_released(css);
5060 } else {
5061 /* cgroup release path */
5062 trace_cgroup_release(cgrp);
5063
5064 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
5065 cgrp->id = -1;
5066
5067 /*
5068 * There are two control paths which try to determine
5069 * cgroup from dentry without going through kernfs -
5070 * cgroupstats_build() and css_tryget_online_from_dir().
5071 * Those are supported by RCU protecting clearing of
5072 * cgrp->kn->priv backpointer.
5073 */
5074 if (cgrp->kn)
5075 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5076 NULL);
5077
5078 cgroup_bpf_put(cgrp);
5079 }
5080
5081 mutex_unlock(&cgroup_mutex);
5082
5083 call_rcu(&css->rcu_head, css_free_rcu_fn);
5084}
5085
5086static void css_release(struct percpu_ref *ref)
5087{
5088 struct cgroup_subsys_state *css =
5089 container_of(ref, struct cgroup_subsys_state, refcnt);
5090
5091 INIT_WORK(&css->destroy_work, css_release_work_fn);
5092 queue_work(cgroup_destroy_wq, &css->destroy_work);
5093}
5094
5095static void init_and_link_css(struct cgroup_subsys_state *css,
5096 struct cgroup_subsys *ss, struct cgroup *cgrp)
5097{
5098 lockdep_assert_held(&cgroup_mutex);
5099
5100 cgroup_get(cgrp);
5101
5102 memset(css, 0, sizeof(*css));
5103 css->cgroup = cgrp;
5104 css->ss = ss;
5105 css->id = -1;
5106 INIT_LIST_HEAD(&css->sibling);
5107 INIT_LIST_HEAD(&css->children);
5108 css->serial_nr = css_serial_nr_next++;
5109 atomic_set(&css->online_cnt, 0);
5110
5111 if (cgroup_parent(cgrp)) {
5112 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5113 css_get(css->parent);
5114 }
5115
5116 BUG_ON(cgroup_css(cgrp, ss));
5117}
5118
5119/* invoke ->css_online() on a new CSS and mark it online if successful */
5120static int online_css(struct cgroup_subsys_state *css)
5121{
5122 struct cgroup_subsys *ss = css->ss;
5123 int ret = 0;
5124
5125 lockdep_assert_held(&cgroup_mutex);
5126
5127 if (ss->css_online)
5128 ret = ss->css_online(css);
5129 if (!ret) {
5130 css->flags |= CSS_ONLINE;
5131 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5132
5133 atomic_inc(&css->online_cnt);
5134 if (css->parent)
5135 atomic_inc(&css->parent->online_cnt);
5136 }
5137 return ret;
5138}
5139
5140/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5141static void offline_css(struct cgroup_subsys_state *css)
5142{
5143 struct cgroup_subsys *ss = css->ss;
5144
5145 lockdep_assert_held(&cgroup_mutex);
5146
5147 if (!(css->flags & CSS_ONLINE))
5148 return;
5149
5150 if (ss->css_reset)
5151 ss->css_reset(css);
5152
5153 if (ss->css_offline)
5154 ss->css_offline(css);
5155
5156 css->flags &= ~CSS_ONLINE;
5157 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5158
5159 wake_up_all(&css->cgroup->offline_waitq);
5160}
5161
5162/**
5163 * css_create - create a cgroup_subsys_state
5164 * @cgrp: the cgroup new css will be associated with
5165 * @ss: the subsys of new css
5166 *
5167 * Create a new css associated with @cgrp - @ss pair. On success, the new
5168 * css is online and installed in @cgrp. This function doesn't create the
5169 * interface files. Returns 0 on success, -errno on failure.
5170 */
5171static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5172 struct cgroup_subsys *ss)
5173{
5174 struct cgroup *parent = cgroup_parent(cgrp);
5175 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5176 struct cgroup_subsys_state *css;
5177 int err;
5178
5179 lockdep_assert_held(&cgroup_mutex);
5180
5181 css = ss->css_alloc(parent_css);
5182 if (!css)
5183 css = ERR_PTR(-ENOMEM);
5184 if (IS_ERR(css))
5185 return css;
5186
5187 init_and_link_css(css, ss, cgrp);
5188
5189 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5190 if (err)
5191 goto err_free_css;
5192
5193 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5194 if (err < 0)
5195 goto err_free_css;
5196 css->id = err;
5197
5198 /* @css is ready to be brought online now, make it visible */
5199 list_add_tail_rcu(&css->sibling, &parent_css->children);
5200 cgroup_idr_replace(&ss->css_idr, css, css->id);
5201
5202 err = online_css(css);
5203 if (err)
5204 goto err_list_del;
5205
5206 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5207 cgroup_parent(parent)) {
5208 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5209 current->comm, current->pid, ss->name);
5210 if (!strcmp(ss->name, "memory"))
5211 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5212 ss->warned_broken_hierarchy = true;
5213 }
5214
5215 return css;
5216
5217err_list_del:
5218 list_del_rcu(&css->sibling);
5219err_free_css:
5220 call_rcu(&css->rcu_head, css_free_rcu_fn);
5221 return ERR_PTR(err);
5222}
5223
5224/*
5225 * The returned cgroup is fully initialized including its control mask, but
5226 * it isn't associated with its kernfs_node and doesn't have the control
5227 * mask applied.
5228 */
5229static struct cgroup *cgroup_create(struct cgroup *parent)
5230{
5231 struct cgroup_root *root = parent->root;
5232 struct cgroup *cgrp, *tcgrp;
5233 int level = parent->level + 1;
5234 int ret;
5235
5236 /* allocate the cgroup and its ID, 0 is reserved for the root */
5237 cgrp = kzalloc(sizeof(*cgrp) +
5238 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5239 if (!cgrp)
5240 return ERR_PTR(-ENOMEM);
5241
5242 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5243 if (ret)
5244 goto out_free_cgrp;
5245
5246 /*
5247 * Temporarily set the pointer to NULL, so idr_find() won't return
5248 * a half-baked cgroup.
5249 */
5250 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5251 if (cgrp->id < 0) {
5252 ret = -ENOMEM;
5253 goto out_cancel_ref;
5254 }
5255
5256 init_cgroup_housekeeping(cgrp);
5257
5258 cgrp->self.parent = &parent->self;
5259 cgrp->root = root;
5260 cgrp->level = level;
5261
5262 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5263 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5264
5265 if (notify_on_release(parent))
5266 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5267
5268 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5269 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5270
5271 cgrp->self.serial_nr = css_serial_nr_next++;
5272
5273 /* allocation complete, commit to creation */
5274 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5275 atomic_inc(&root->nr_cgrps);
5276 cgroup_get(parent);
5277
5278 /*
5279 * @cgrp is now fully operational. If something fails after this
5280 * point, it'll be released via the normal destruction path.
5281 */
5282 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5283
5284 /*
5285 * On the default hierarchy, a child doesn't automatically inherit
5286 * subtree_control from the parent. Each is configured manually.
5287 */
5288 if (!cgroup_on_dfl(cgrp))
5289 cgrp->subtree_control = cgroup_control(cgrp);
5290
5291 if (parent)
5292 cgroup_bpf_inherit(cgrp, parent);
5293
5294 cgroup_propagate_control(cgrp);
5295
5296 return cgrp;
5297
5298out_cancel_ref:
5299 percpu_ref_exit(&cgrp->self.refcnt);
5300out_free_cgrp:
5301 kfree(cgrp);
5302 return ERR_PTR(ret);
5303}
5304
5305static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5306 umode_t mode)
5307{
5308 struct cgroup *parent, *cgrp;
5309 struct kernfs_node *kn;
5310 int ret;
5311
5312 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5313 if (strchr(name, '\n'))
5314 return -EINVAL;
5315
5316 parent = cgroup_kn_lock_live(parent_kn, false);
5317 if (!parent)
5318 return -ENODEV;
5319
5320 cgrp = cgroup_create(parent);
5321 if (IS_ERR(cgrp)) {
5322 ret = PTR_ERR(cgrp);
5323 goto out_unlock;
5324 }
5325
5326 /* create the directory */
5327 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5328 if (IS_ERR(kn)) {
5329 ret = PTR_ERR(kn);
5330 goto out_destroy;
5331 }
5332 cgrp->kn = kn;
5333
5334 /*
5335 * This extra ref will be put in cgroup_free_fn() and guarantees
5336 * that @cgrp->kn is always accessible.
5337 */
5338 kernfs_get(kn);
5339
5340 ret = cgroup_kn_set_ugid(kn);
5341 if (ret)
5342 goto out_destroy;
5343
5344 ret = css_populate_dir(&cgrp->self);
5345 if (ret)
5346 goto out_destroy;
5347
5348 ret = cgroup_apply_control_enable(cgrp);
5349 if (ret)
5350 goto out_destroy;
5351
5352 trace_cgroup_mkdir(cgrp);
5353
5354 /* let's create and online css's */
5355 kernfs_activate(kn);
5356
5357 ret = 0;
5358 goto out_unlock;
5359
5360out_destroy:
5361 cgroup_destroy_locked(cgrp);
5362out_unlock:
5363 cgroup_kn_unlock(parent_kn);
5364 return ret;
5365}
5366
5367/*
5368 * This is called when the refcnt of a css is confirmed to be killed.
5369 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5370 * initate destruction and put the css ref from kill_css().
5371 */
5372static void css_killed_work_fn(struct work_struct *work)
5373{
5374 struct cgroup_subsys_state *css =
5375 container_of(work, struct cgroup_subsys_state, destroy_work);
5376
5377 mutex_lock(&cgroup_mutex);
5378
5379 do {
5380 offline_css(css);
5381 css_put(css);
5382 /* @css can't go away while we're holding cgroup_mutex */
5383 css = css->parent;
5384 } while (css && atomic_dec_and_test(&css->online_cnt));
5385
5386 mutex_unlock(&cgroup_mutex);
5387}
5388
5389/* css kill confirmation processing requires process context, bounce */
5390static void css_killed_ref_fn(struct percpu_ref *ref)
5391{
5392 struct cgroup_subsys_state *css =
5393 container_of(ref, struct cgroup_subsys_state, refcnt);
5394
5395 if (atomic_dec_and_test(&css->online_cnt)) {
5396 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5397 queue_work(cgroup_destroy_wq, &css->destroy_work);
5398 }
5399}
5400
5401/**
5402 * kill_css - destroy a css
5403 * @css: css to destroy
5404 *
5405 * This function initiates destruction of @css by removing cgroup interface
5406 * files and putting its base reference. ->css_offline() will be invoked
5407 * asynchronously once css_tryget_online() is guaranteed to fail and when
5408 * the reference count reaches zero, @css will be released.
5409 */
5410static void kill_css(struct cgroup_subsys_state *css)
5411{
5412 lockdep_assert_held(&cgroup_mutex);
5413
5414 /*
5415 * This must happen before css is disassociated with its cgroup.
5416 * See seq_css() for details.
5417 */
5418 css_clear_dir(css);
5419
5420 /*
5421 * Killing would put the base ref, but we need to keep it alive
5422 * until after ->css_offline().
5423 */
5424 css_get(css);
5425
5426 /*
5427 * cgroup core guarantees that, by the time ->css_offline() is
5428 * invoked, no new css reference will be given out via
5429 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5430 * proceed to offlining css's because percpu_ref_kill() doesn't
5431 * guarantee that the ref is seen as killed on all CPUs on return.
5432 *
5433 * Use percpu_ref_kill_and_confirm() to get notifications as each
5434 * css is confirmed to be seen as killed on all CPUs.
5435 */
5436 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5437}
5438
5439/**
5440 * cgroup_destroy_locked - the first stage of cgroup destruction
5441 * @cgrp: cgroup to be destroyed
5442 *
5443 * css's make use of percpu refcnts whose killing latency shouldn't be
5444 * exposed to userland and are RCU protected. Also, cgroup core needs to
5445 * guarantee that css_tryget_online() won't succeed by the time
5446 * ->css_offline() is invoked. To satisfy all the requirements,
5447 * destruction is implemented in the following two steps.
5448 *
5449 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5450 * userland visible parts and start killing the percpu refcnts of
5451 * css's. Set up so that the next stage will be kicked off once all
5452 * the percpu refcnts are confirmed to be killed.
5453 *
5454 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5455 * rest of destruction. Once all cgroup references are gone, the
5456 * cgroup is RCU-freed.
5457 *
5458 * This function implements s1. After this step, @cgrp is gone as far as
5459 * the userland is concerned and a new cgroup with the same name may be
5460 * created. As cgroup doesn't care about the names internally, this
5461 * doesn't cause any problem.
5462 */
5463static int cgroup_destroy_locked(struct cgroup *cgrp)
5464 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5465{
5466 struct cgroup_subsys_state *css;
5467 struct cgrp_cset_link *link;
5468 int ssid;
5469
5470 lockdep_assert_held(&cgroup_mutex);
5471
5472 /*
5473 * Only migration can raise populated from zero and we're already
5474 * holding cgroup_mutex.
5475 */
5476 if (cgroup_is_populated(cgrp))
5477 return -EBUSY;
5478
5479 /*
5480 * Make sure there's no live children. We can't test emptiness of
5481 * ->self.children as dead children linger on it while being
5482 * drained; otherwise, "rmdir parent/child parent" may fail.
5483 */
5484 if (css_has_online_children(&cgrp->self))
5485 return -EBUSY;
5486
5487 /*
5488 * Mark @cgrp and the associated csets dead. The former prevents
5489 * further task migration and child creation by disabling
5490 * cgroup_lock_live_group(). The latter makes the csets ignored by
5491 * the migration path.
5492 */
5493 cgrp->self.flags &= ~CSS_ONLINE;
5494
5495 spin_lock_irq(&css_set_lock);
5496 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5497 link->cset->dead = true;
5498 spin_unlock_irq(&css_set_lock);
5499
5500 /* initiate massacre of all css's */
5501 for_each_css(css, ssid, cgrp)
5502 kill_css(css);
5503
5504 /*
5505 * Remove @cgrp directory along with the base files. @cgrp has an
5506 * extra ref on its kn.
5507 */
5508 kernfs_remove(cgrp->kn);
5509
5510 check_for_release(cgroup_parent(cgrp));
5511
5512 /* put the base reference */
5513 percpu_ref_kill(&cgrp->self.refcnt);
5514
5515 return 0;
5516};
5517
5518static int cgroup_rmdir(struct kernfs_node *kn)
5519{
5520 struct cgroup *cgrp;
5521 int ret = 0;
5522
5523 cgrp = cgroup_kn_lock_live(kn, false);
5524 if (!cgrp)
5525 return 0;
5526
5527 ret = cgroup_destroy_locked(cgrp);
5528
5529 if (!ret)
5530 trace_cgroup_rmdir(cgrp);
5531
5532 cgroup_kn_unlock(kn);
5533 return ret;
5534}
5535
5536static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5537 .remount_fs = cgroup_remount,
5538 .show_options = cgroup_show_options,
5539 .mkdir = cgroup_mkdir,
5540 .rmdir = cgroup_rmdir,
5541 .rename = cgroup_rename,
5542 .show_path = cgroup_show_path,
5543};
5544
5545static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5546{
5547 struct cgroup_subsys_state *css;
5548
5549 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5550
5551 mutex_lock(&cgroup_mutex);
5552
5553 idr_init(&ss->css_idr);
5554 INIT_LIST_HEAD(&ss->cfts);
5555
5556 /* Create the root cgroup state for this subsystem */
5557 ss->root = &cgrp_dfl_root;
5558 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5559 /* We don't handle early failures gracefully */
5560 BUG_ON(IS_ERR(css));
5561 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5562
5563 /*
5564 * Root csses are never destroyed and we can't initialize
5565 * percpu_ref during early init. Disable refcnting.
5566 */
5567 css->flags |= CSS_NO_REF;
5568
5569 if (early) {
5570 /* allocation can't be done safely during early init */
5571 css->id = 1;
5572 } else {
5573 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5574 BUG_ON(css->id < 0);
5575 }
5576
5577 /* Update the init_css_set to contain a subsys
5578 * pointer to this state - since the subsystem is
5579 * newly registered, all tasks and hence the
5580 * init_css_set is in the subsystem's root cgroup. */
5581 init_css_set.subsys[ss->id] = css;
5582
5583 have_fork_callback |= (bool)ss->fork << ss->id;
5584 have_exit_callback |= (bool)ss->exit << ss->id;
5585 have_free_callback |= (bool)ss->free << ss->id;
5586 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5587
5588 /* At system boot, before all subsystems have been
5589 * registered, no tasks have been forked, so we don't
5590 * need to invoke fork callbacks here. */
5591 BUG_ON(!list_empty(&init_task.tasks));
5592
5593 BUG_ON(online_css(css));
5594
5595 mutex_unlock(&cgroup_mutex);
5596}
5597
5598/**
5599 * cgroup_init_early - cgroup initialization at system boot
5600 *
5601 * Initialize cgroups at system boot, and initialize any
5602 * subsystems that request early init.
5603 */
5604int __init cgroup_init_early(void)
5605{
5606 static struct cgroup_sb_opts __initdata opts;
5607 struct cgroup_subsys *ss;
5608 int i;
5609
5610 init_cgroup_root(&cgrp_dfl_root, &opts);
5611 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5612
5613 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5614
5615 for_each_subsys(ss, i) {
5616 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5617 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5618 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5619 ss->id, ss->name);
5620 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5621 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5622
5623 ss->id = i;
5624 ss->name = cgroup_subsys_name[i];
5625 if (!ss->legacy_name)
5626 ss->legacy_name = cgroup_subsys_name[i];
5627
5628 if (ss->early_init)
5629 cgroup_init_subsys(ss, true);
5630 }
5631 return 0;
5632}
5633
5634static u16 cgroup_disable_mask __initdata;
5635
5636/**
5637 * cgroup_init - cgroup initialization
5638 *
5639 * Register cgroup filesystem and /proc file, and initialize
5640 * any subsystems that didn't request early init.
5641 */
5642int __init cgroup_init(void)
5643{
5644 struct cgroup_subsys *ss;
5645 int ssid;
5646
5647 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5648 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5649 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5650 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5651
5652 /*
5653 * The latency of the synchronize_sched() is too high for cgroups,
5654 * avoid it at the cost of forcing all readers into the slow path.
5655 */
5656 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5657
5658 get_user_ns(init_cgroup_ns.user_ns);
5659
5660 mutex_lock(&cgroup_mutex);
5661
5662 /*
5663 * Add init_css_set to the hash table so that dfl_root can link to
5664 * it during init.
5665 */
5666 hash_add(css_set_table, &init_css_set.hlist,
5667 css_set_hash(init_css_set.subsys));
5668
5669 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5670
5671 mutex_unlock(&cgroup_mutex);
5672
5673 for_each_subsys(ss, ssid) {
5674 if (ss->early_init) {
5675 struct cgroup_subsys_state *css =
5676 init_css_set.subsys[ss->id];
5677
5678 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5679 GFP_KERNEL);
5680 BUG_ON(css->id < 0);
5681 } else {
5682 cgroup_init_subsys(ss, false);
5683 }
5684
5685 list_add_tail(&init_css_set.e_cset_node[ssid],
5686 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5687
5688 /*
5689 * Setting dfl_root subsys_mask needs to consider the
5690 * disabled flag and cftype registration needs kmalloc,
5691 * both of which aren't available during early_init.
5692 */
5693 if (cgroup_disable_mask & (1 << ssid)) {
5694 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5695 printk(KERN_INFO "Disabling %s control group subsystem\n",
5696 ss->name);
5697 continue;
5698 }
5699
5700 if (cgroup_ssid_no_v1(ssid))
5701 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5702 ss->name);
5703
5704 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5705
5706 if (ss->implicit_on_dfl)
5707 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5708 else if (!ss->dfl_cftypes)
5709 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5710
5711 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5712 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5713 } else {
5714 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5715 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5716 }
5717
5718 if (ss->bind)
5719 ss->bind(init_css_set.subsys[ssid]);
5720 }
5721
5722 /* init_css_set.subsys[] has been updated, re-hash */
5723 hash_del(&init_css_set.hlist);
5724 hash_add(css_set_table, &init_css_set.hlist,
5725 css_set_hash(init_css_set.subsys));
5726
5727 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5728 WARN_ON(register_filesystem(&cgroup_fs_type));
5729 WARN_ON(register_filesystem(&cgroup2_fs_type));
5730 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5731
5732 return 0;
5733}
5734
5735static int __init cgroup_wq_init(void)
5736{
5737 /*
5738 * There isn't much point in executing destruction path in
5739 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5740 * Use 1 for @max_active.
5741 *
5742 * We would prefer to do this in cgroup_init() above, but that
5743 * is called before init_workqueues(): so leave this until after.
5744 */
5745 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5746 BUG_ON(!cgroup_destroy_wq);
5747
5748 /*
5749 * Used to destroy pidlists and separate to serve as flush domain.
5750 * Cap @max_active to 1 too.
5751 */
5752 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5753 0, 1);
5754 BUG_ON(!cgroup_pidlist_destroy_wq);
5755
5756 return 0;
5757}
5758core_initcall(cgroup_wq_init);
5759
5760/*
5761 * proc_cgroup_show()
5762 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5763 * - Used for /proc/<pid>/cgroup.
5764 */
5765int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5766 struct pid *pid, struct task_struct *tsk)
5767{
5768 char *buf;
5769 int retval;
5770 struct cgroup_root *root;
5771
5772 retval = -ENOMEM;
5773 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5774 if (!buf)
5775 goto out;
5776
5777 mutex_lock(&cgroup_mutex);
5778 spin_lock_irq(&css_set_lock);
5779
5780 for_each_root(root) {
5781 struct cgroup_subsys *ss;
5782 struct cgroup *cgrp;
5783 int ssid, count = 0;
5784
5785 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5786 continue;
5787
5788 seq_printf(m, "%d:", root->hierarchy_id);
5789 if (root != &cgrp_dfl_root)
5790 for_each_subsys(ss, ssid)
5791 if (root->subsys_mask & (1 << ssid))
5792 seq_printf(m, "%s%s", count++ ? "," : "",
5793 ss->legacy_name);
5794 if (strlen(root->name))
5795 seq_printf(m, "%sname=%s", count ? "," : "",
5796 root->name);
5797 seq_putc(m, ':');
5798
5799 cgrp = task_cgroup_from_root(tsk, root);
5800
5801 /*
5802 * On traditional hierarchies, all zombie tasks show up as
5803 * belonging to the root cgroup. On the default hierarchy,
5804 * while a zombie doesn't show up in "cgroup.procs" and
5805 * thus can't be migrated, its /proc/PID/cgroup keeps
5806 * reporting the cgroup it belonged to before exiting. If
5807 * the cgroup is removed before the zombie is reaped,
5808 * " (deleted)" is appended to the cgroup path.
5809 */
5810 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5811 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5812 current->nsproxy->cgroup_ns);
5813 if (retval >= PATH_MAX)
5814 retval = -ENAMETOOLONG;
5815 if (retval < 0)
5816 goto out_unlock;
5817
5818 seq_puts(m, buf);
5819 } else {
5820 seq_puts(m, "/");
5821 }
5822
5823 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5824 seq_puts(m, " (deleted)\n");
5825 else
5826 seq_putc(m, '\n');
5827 }
5828
5829 retval = 0;
5830out_unlock:
5831 spin_unlock_irq(&css_set_lock);
5832 mutex_unlock(&cgroup_mutex);
5833 kfree(buf);
5834out:
5835 return retval;
5836}
5837
5838/* Display information about each subsystem and each hierarchy */
5839static int proc_cgroupstats_show(struct seq_file *m, void *v)
5840{
5841 struct cgroup_subsys *ss;
5842 int i;
5843
5844 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5845 /*
5846 * ideally we don't want subsystems moving around while we do this.
5847 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5848 * subsys/hierarchy state.
5849 */
5850 mutex_lock(&cgroup_mutex);
5851
5852 for_each_subsys(ss, i)
5853 seq_printf(m, "%s\t%d\t%d\t%d\n",
5854 ss->legacy_name, ss->root->hierarchy_id,
5855 atomic_read(&ss->root->nr_cgrps),
5856 cgroup_ssid_enabled(i));
5857
5858 mutex_unlock(&cgroup_mutex);
5859 return 0;
5860}
5861
5862static int cgroupstats_open(struct inode *inode, struct file *file)
5863{
5864 return single_open(file, proc_cgroupstats_show, NULL);
5865}
5866
5867static const struct file_operations proc_cgroupstats_operations = {
5868 .open = cgroupstats_open,
5869 .read = seq_read,
5870 .llseek = seq_lseek,
5871 .release = single_release,
5872};
5873
5874/**
5875 * cgroup_fork - initialize cgroup related fields during copy_process()
5876 * @child: pointer to task_struct of forking parent process.
5877 *
5878 * A task is associated with the init_css_set until cgroup_post_fork()
5879 * attaches it to the parent's css_set. Empty cg_list indicates that
5880 * @child isn't holding reference to its css_set.
5881 */
5882void cgroup_fork(struct task_struct *child)
5883{
5884 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5885 INIT_LIST_HEAD(&child->cg_list);
5886}
5887
5888/**
5889 * cgroup_can_fork - called on a new task before the process is exposed
5890 * @child: the task in question.
5891 *
5892 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5893 * returns an error, the fork aborts with that error code. This allows for
5894 * a cgroup subsystem to conditionally allow or deny new forks.
5895 */
5896int cgroup_can_fork(struct task_struct *child)
5897{
5898 struct cgroup_subsys *ss;
5899 int i, j, ret;
5900
5901 do_each_subsys_mask(ss, i, have_canfork_callback) {
5902 ret = ss->can_fork(child);
5903 if (ret)
5904 goto out_revert;
5905 } while_each_subsys_mask();
5906
5907 return 0;
5908
5909out_revert:
5910 for_each_subsys(ss, j) {
5911 if (j >= i)
5912 break;
5913 if (ss->cancel_fork)
5914 ss->cancel_fork(child);
5915 }
5916
5917 return ret;
5918}
5919
5920/**
5921 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5922 * @child: the task in question
5923 *
5924 * This calls the cancel_fork() callbacks if a fork failed *after*
5925 * cgroup_can_fork() succeded.
5926 */
5927void cgroup_cancel_fork(struct task_struct *child)
5928{
5929 struct cgroup_subsys *ss;
5930 int i;
5931
5932 for_each_subsys(ss, i)
5933 if (ss->cancel_fork)
5934 ss->cancel_fork(child);
5935}
5936
5937/**
5938 * cgroup_post_fork - called on a new task after adding it to the task list
5939 * @child: the task in question
5940 *
5941 * Adds the task to the list running through its css_set if necessary and
5942 * call the subsystem fork() callbacks. Has to be after the task is
5943 * visible on the task list in case we race with the first call to
5944 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5945 * list.
5946 */
5947void cgroup_post_fork(struct task_struct *child)
5948{
5949 struct cgroup_subsys *ss;
5950 int i;
5951
5952 /*
5953 * This may race against cgroup_enable_task_cg_lists(). As that
5954 * function sets use_task_css_set_links before grabbing
5955 * tasklist_lock and we just went through tasklist_lock to add
5956 * @child, it's guaranteed that either we see the set
5957 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5958 * @child during its iteration.
5959 *
5960 * If we won the race, @child is associated with %current's
5961 * css_set. Grabbing css_set_lock guarantees both that the
5962 * association is stable, and, on completion of the parent's
5963 * migration, @child is visible in the source of migration or
5964 * already in the destination cgroup. This guarantee is necessary
5965 * when implementing operations which need to migrate all tasks of
5966 * a cgroup to another.
5967 *
5968 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5969 * will remain in init_css_set. This is safe because all tasks are
5970 * in the init_css_set before cg_links is enabled and there's no
5971 * operation which transfers all tasks out of init_css_set.
5972 */
5973 if (use_task_css_set_links) {
5974 struct css_set *cset;
5975
5976 spin_lock_irq(&css_set_lock);
5977 cset = task_css_set(current);
5978 if (list_empty(&child->cg_list)) {
5979 get_css_set(cset);
5980 css_set_move_task(child, NULL, cset, false);
5981 }
5982 spin_unlock_irq(&css_set_lock);
5983 }
5984
5985 /*
5986 * Call ss->fork(). This must happen after @child is linked on
5987 * css_set; otherwise, @child might change state between ->fork()
5988 * and addition to css_set.
5989 */
5990 do_each_subsys_mask(ss, i, have_fork_callback) {
5991 ss->fork(child);
5992 } while_each_subsys_mask();
5993}
5994
5995/**
5996 * cgroup_exit - detach cgroup from exiting task
5997 * @tsk: pointer to task_struct of exiting process
5998 *
5999 * Description: Detach cgroup from @tsk and release it.
6000 *
6001 * Note that cgroups marked notify_on_release force every task in
6002 * them to take the global cgroup_mutex mutex when exiting.
6003 * This could impact scaling on very large systems. Be reluctant to
6004 * use notify_on_release cgroups where very high task exit scaling
6005 * is required on large systems.
6006 *
6007 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
6008 * call cgroup_exit() while the task is still competent to handle
6009 * notify_on_release(), then leave the task attached to the root cgroup in
6010 * each hierarchy for the remainder of its exit. No need to bother with
6011 * init_css_set refcnting. init_css_set never goes away and we can't race
6012 * with migration path - PF_EXITING is visible to migration path.
6013 */
6014void cgroup_exit(struct task_struct *tsk)
6015{
6016 struct cgroup_subsys *ss;
6017 struct css_set *cset;
6018 int i;
6019
6020 /*
6021 * Unlink from @tsk from its css_set. As migration path can't race
6022 * with us, we can check css_set and cg_list without synchronization.
6023 */
6024 cset = task_css_set(tsk);
6025
6026 if (!list_empty(&tsk->cg_list)) {
6027 spin_lock_irq(&css_set_lock);
6028 css_set_move_task(tsk, cset, NULL, false);
6029 spin_unlock_irq(&css_set_lock);
6030 } else {
6031 get_css_set(cset);
6032 }
6033
6034 /* see cgroup_post_fork() for details */
6035 do_each_subsys_mask(ss, i, have_exit_callback) {
6036 ss->exit(tsk);
6037 } while_each_subsys_mask();
6038}
6039
6040void cgroup_free(struct task_struct *task)
6041{
6042 struct css_set *cset = task_css_set(task);
6043 struct cgroup_subsys *ss;
6044 int ssid;
6045
6046 do_each_subsys_mask(ss, ssid, have_free_callback) {
6047 ss->free(task);
6048 } while_each_subsys_mask();
6049
6050 put_css_set(cset);
6051}
6052
6053static void check_for_release(struct cgroup *cgrp)
6054{
6055 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
6056 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
6057 schedule_work(&cgrp->release_agent_work);
6058}
6059
6060/*
6061 * Notify userspace when a cgroup is released, by running the
6062 * configured release agent with the name of the cgroup (path
6063 * relative to the root of cgroup file system) as the argument.
6064 *
6065 * Most likely, this user command will try to rmdir this cgroup.
6066 *
6067 * This races with the possibility that some other task will be
6068 * attached to this cgroup before it is removed, or that some other
6069 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
6070 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
6071 * unused, and this cgroup will be reprieved from its death sentence,
6072 * to continue to serve a useful existence. Next time it's released,
6073 * we will get notified again, if it still has 'notify_on_release' set.
6074 *
6075 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
6076 * means only wait until the task is successfully execve()'d. The
6077 * separate release agent task is forked by call_usermodehelper(),
6078 * then control in this thread returns here, without waiting for the
6079 * release agent task. We don't bother to wait because the caller of
6080 * this routine has no use for the exit status of the release agent
6081 * task, so no sense holding our caller up for that.
6082 */
6083static void cgroup_release_agent(struct work_struct *work)
6084{
6085 struct cgroup *cgrp =
6086 container_of(work, struct cgroup, release_agent_work);
6087 char *pathbuf = NULL, *agentbuf = NULL;
6088 char *argv[3], *envp[3];
6089 int ret;
6090
6091 mutex_lock(&cgroup_mutex);
6092
6093 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
6094 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
6095 if (!pathbuf || !agentbuf)
6096 goto out;
6097
6098 spin_lock_irq(&css_set_lock);
6099 ret = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
6100 spin_unlock_irq(&css_set_lock);
6101 if (ret < 0 || ret >= PATH_MAX)
6102 goto out;
6103
6104 argv[0] = agentbuf;
6105 argv[1] = pathbuf;
6106 argv[2] = NULL;
6107
6108 /* minimal command environment */
6109 envp[0] = "HOME=/";
6110 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
6111 envp[2] = NULL;
6112
6113 mutex_unlock(&cgroup_mutex);
6114 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
6115 goto out_free;
6116out:
6117 mutex_unlock(&cgroup_mutex);
6118out_free:
6119 kfree(agentbuf);
6120 kfree(pathbuf);
6121}
6122
6123static int __init cgroup_disable(char *str)
6124{
6125 struct cgroup_subsys *ss;
6126 char *token;
6127 int i;
6128
6129 while ((token = strsep(&str, ",")) != NULL) {
6130 if (!*token)
6131 continue;
6132
6133 for_each_subsys(ss, i) {
6134 if (strcmp(token, ss->name) &&
6135 strcmp(token, ss->legacy_name))
6136 continue;
6137 cgroup_disable_mask |= 1 << i;
6138 }
6139 }
6140 return 1;
6141}
6142__setup("cgroup_disable=", cgroup_disable);
6143
6144static int __init cgroup_no_v1(char *str)
6145{
6146 struct cgroup_subsys *ss;
6147 char *token;
6148 int i;
6149
6150 while ((token = strsep(&str, ",")) != NULL) {
6151 if (!*token)
6152 continue;
6153
6154 if (!strcmp(token, "all")) {
6155 cgroup_no_v1_mask = U16_MAX;
6156 break;
6157 }
6158
6159 for_each_subsys(ss, i) {
6160 if (strcmp(token, ss->name) &&
6161 strcmp(token, ss->legacy_name))
6162 continue;
6163
6164 cgroup_no_v1_mask |= 1 << i;
6165 }
6166 }
6167 return 1;
6168}
6169__setup("cgroup_no_v1=", cgroup_no_v1);
6170
6171/**
6172 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6173 * @dentry: directory dentry of interest
6174 * @ss: subsystem of interest
6175 *
6176 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6177 * to get the corresponding css and return it. If such css doesn't exist
6178 * or can't be pinned, an ERR_PTR value is returned.
6179 */
6180struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6181 struct cgroup_subsys *ss)
6182{
6183 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6184 struct file_system_type *s_type = dentry->d_sb->s_type;
6185 struct cgroup_subsys_state *css = NULL;
6186 struct cgroup *cgrp;
6187
6188 /* is @dentry a cgroup dir? */
6189 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6190 !kn || kernfs_type(kn) != KERNFS_DIR)
6191 return ERR_PTR(-EBADF);
6192
6193 rcu_read_lock();
6194
6195 /*
6196 * This path doesn't originate from kernfs and @kn could already
6197 * have been or be removed at any point. @kn->priv is RCU
6198 * protected for this access. See css_release_work_fn() for details.
6199 */
6200 cgrp = rcu_dereference(kn->priv);
6201 if (cgrp)
6202 css = cgroup_css(cgrp, ss);
6203
6204 if (!css || !css_tryget_online(css))
6205 css = ERR_PTR(-ENOENT);
6206
6207 rcu_read_unlock();
6208 return css;
6209}
6210
6211/**
6212 * css_from_id - lookup css by id
6213 * @id: the cgroup id
6214 * @ss: cgroup subsys to be looked into
6215 *
6216 * Returns the css if there's valid one with @id, otherwise returns NULL.
6217 * Should be called under rcu_read_lock().
6218 */
6219struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6220{
6221 WARN_ON_ONCE(!rcu_read_lock_held());
6222 return idr_find(&ss->css_idr, id);
6223}
6224
6225/**
6226 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6227 * @path: path on the default hierarchy
6228 *
6229 * Find the cgroup at @path on the default hierarchy, increment its
6230 * reference count and return it. Returns pointer to the found cgroup on
6231 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6232 * if @path points to a non-directory.
6233 */
6234struct cgroup *cgroup_get_from_path(const char *path)
6235{
6236 struct kernfs_node *kn;
6237 struct cgroup *cgrp;
6238
6239 mutex_lock(&cgroup_mutex);
6240
6241 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6242 if (kn) {
6243 if (kernfs_type(kn) == KERNFS_DIR) {
6244 cgrp = kn->priv;
6245 cgroup_get(cgrp);
6246 } else {
6247 cgrp = ERR_PTR(-ENOTDIR);
6248 }
6249 kernfs_put(kn);
6250 } else {
6251 cgrp = ERR_PTR(-ENOENT);
6252 }
6253
6254 mutex_unlock(&cgroup_mutex);
6255 return cgrp;
6256}
6257EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6258
6259/**
6260 * cgroup_get_from_fd - get a cgroup pointer from a fd
6261 * @fd: fd obtained by open(cgroup2_dir)
6262 *
6263 * Find the cgroup from a fd which should be obtained
6264 * by opening a cgroup directory. Returns a pointer to the
6265 * cgroup on success. ERR_PTR is returned if the cgroup
6266 * cannot be found.
6267 */
6268struct cgroup *cgroup_get_from_fd(int fd)
6269{
6270 struct cgroup_subsys_state *css;
6271 struct cgroup *cgrp;
6272 struct file *f;
6273
6274 f = fget_raw(fd);
6275 if (!f)
6276 return ERR_PTR(-EBADF);
6277
6278 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6279 fput(f);
6280 if (IS_ERR(css))
6281 return ERR_CAST(css);
6282
6283 cgrp = css->cgroup;
6284 if (!cgroup_on_dfl(cgrp)) {
6285 cgroup_put(cgrp);
6286 return ERR_PTR(-EBADF);
6287 }
6288
6289 return cgrp;
6290}
6291EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6292
6293/*
6294 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6295 * definition in cgroup-defs.h.
6296 */
6297#ifdef CONFIG_SOCK_CGROUP_DATA
6298
6299#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6300
6301DEFINE_SPINLOCK(cgroup_sk_update_lock);
6302static bool cgroup_sk_alloc_disabled __read_mostly;
6303
6304void cgroup_sk_alloc_disable(void)
6305{
6306 if (cgroup_sk_alloc_disabled)
6307 return;
6308 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6309 cgroup_sk_alloc_disabled = true;
6310}
6311
6312#else
6313
6314#define cgroup_sk_alloc_disabled false
6315
6316#endif
6317
6318void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6319{
6320 if (cgroup_sk_alloc_disabled)
6321 return;
6322
6323 /* Socket clone path */
6324 if (skcd->val) {
6325 cgroup_get(sock_cgroup_ptr(skcd));
6326 return;
6327 }
6328
6329 rcu_read_lock();
6330
6331 while (true) {
6332 struct css_set *cset;
6333
6334 cset = task_css_set(current);
6335 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6336 skcd->val = (unsigned long)cset->dfl_cgrp;
6337 break;
6338 }
6339 cpu_relax();
6340 }
6341
6342 rcu_read_unlock();
6343}
6344
6345void cgroup_sk_free(struct sock_cgroup_data *skcd)
6346{
6347 cgroup_put(sock_cgroup_ptr(skcd));
6348}
6349
6350#endif /* CONFIG_SOCK_CGROUP_DATA */
6351
6352/* cgroup namespaces */
6353
6354static struct ucounts *inc_cgroup_namespaces(struct user_namespace *ns)
6355{
6356 return inc_ucount(ns, current_euid(), UCOUNT_CGROUP_NAMESPACES);
6357}
6358
6359static void dec_cgroup_namespaces(struct ucounts *ucounts)
6360{
6361 dec_ucount(ucounts, UCOUNT_CGROUP_NAMESPACES);
6362}
6363
6364static struct cgroup_namespace *alloc_cgroup_ns(void)
6365{
6366 struct cgroup_namespace *new_ns;
6367 int ret;
6368
6369 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6370 if (!new_ns)
6371 return ERR_PTR(-ENOMEM);
6372 ret = ns_alloc_inum(&new_ns->ns);
6373 if (ret) {
6374 kfree(new_ns);
6375 return ERR_PTR(ret);
6376 }
6377 atomic_set(&new_ns->count, 1);
6378 new_ns->ns.ops = &cgroupns_operations;
6379 return new_ns;
6380}
6381
6382void free_cgroup_ns(struct cgroup_namespace *ns)
6383{
6384 put_css_set(ns->root_cset);
6385 dec_cgroup_namespaces(ns->ucounts);
6386 put_user_ns(ns->user_ns);
6387 ns_free_inum(&ns->ns);
6388 kfree(ns);
6389}
6390EXPORT_SYMBOL(free_cgroup_ns);
6391
6392struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6393 struct user_namespace *user_ns,
6394 struct cgroup_namespace *old_ns)
6395{
6396 struct cgroup_namespace *new_ns;
6397 struct ucounts *ucounts;
6398 struct css_set *cset;
6399
6400 BUG_ON(!old_ns);
6401
6402 if (!(flags & CLONE_NEWCGROUP)) {
6403 get_cgroup_ns(old_ns);
6404 return old_ns;
6405 }
6406
6407 /* Allow only sysadmin to create cgroup namespace. */
6408 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6409 return ERR_PTR(-EPERM);
6410
6411 ucounts = inc_cgroup_namespaces(user_ns);
6412 if (!ucounts)
6413 return ERR_PTR(-ENOSPC);
6414
6415 /* It is not safe to take cgroup_mutex here */
6416 spin_lock_irq(&css_set_lock);
6417 cset = task_css_set(current);
6418 get_css_set(cset);
6419 spin_unlock_irq(&css_set_lock);
6420
6421 new_ns = alloc_cgroup_ns();
6422 if (IS_ERR(new_ns)) {
6423 put_css_set(cset);
6424 dec_cgroup_namespaces(ucounts);
6425 return new_ns;
6426 }
6427
6428 new_ns->user_ns = get_user_ns(user_ns);
6429 new_ns->ucounts = ucounts;
6430 new_ns->root_cset = cset;
6431
6432 return new_ns;
6433}
6434
6435static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6436{
6437 return container_of(ns, struct cgroup_namespace, ns);
6438}
6439
6440static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6441{
6442 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6443
6444 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6445 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6446 return -EPERM;
6447
6448 /* Don't need to do anything if we are attaching to our own cgroupns. */
6449 if (cgroup_ns == nsproxy->cgroup_ns)
6450 return 0;
6451
6452 get_cgroup_ns(cgroup_ns);
6453 put_cgroup_ns(nsproxy->cgroup_ns);
6454 nsproxy->cgroup_ns = cgroup_ns;
6455
6456 return 0;
6457}
6458
6459static struct ns_common *cgroupns_get(struct task_struct *task)
6460{
6461 struct cgroup_namespace *ns = NULL;
6462 struct nsproxy *nsproxy;
6463
6464 task_lock(task);
6465 nsproxy = task->nsproxy;
6466 if (nsproxy) {
6467 ns = nsproxy->cgroup_ns;
6468 get_cgroup_ns(ns);
6469 }
6470 task_unlock(task);
6471
6472 return ns ? &ns->ns : NULL;
6473}
6474
6475static void cgroupns_put(struct ns_common *ns)
6476{
6477 put_cgroup_ns(to_cg_ns(ns));
6478}
6479
6480static struct user_namespace *cgroupns_owner(struct ns_common *ns)
6481{
6482 return to_cg_ns(ns)->user_ns;
6483}
6484
6485const struct proc_ns_operations cgroupns_operations = {
6486 .name = "cgroup",
6487 .type = CLONE_NEWCGROUP,
6488 .get = cgroupns_get,
6489 .put = cgroupns_put,
6490 .install = cgroupns_install,
6491 .owner = cgroupns_owner,
6492};
6493
6494static __init int cgroup_namespaces_init(void)
6495{
6496 return 0;
6497}
6498subsys_initcall(cgroup_namespaces_init);
6499
6500#ifdef CONFIG_CGROUP_BPF
6501int cgroup_bpf_update(struct cgroup *cgrp, struct bpf_prog *prog,
6502 enum bpf_attach_type type, bool overridable)
6503{
6504 struct cgroup *parent = cgroup_parent(cgrp);
6505 int ret;
6506
6507 mutex_lock(&cgroup_mutex);
6508 ret = __cgroup_bpf_update(cgrp, parent, prog, type, overridable);
6509 mutex_unlock(&cgroup_mutex);
6510 return ret;
6511}
6512#endif /* CONFIG_CGROUP_BPF */
6513
6514#ifdef CONFIG_CGROUP_DEBUG
6515static struct cgroup_subsys_state *
6516debug_css_alloc(struct cgroup_subsys_state *parent_css)
6517{
6518 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6519
6520 if (!css)
6521 return ERR_PTR(-ENOMEM);
6522
6523 return css;
6524}
6525
6526static void debug_css_free(struct cgroup_subsys_state *css)
6527{
6528 kfree(css);
6529}
6530
6531static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6532 struct cftype *cft)
6533{
6534 return cgroup_task_count(css->cgroup);
6535}
6536
6537static u64 current_css_set_read(struct cgroup_subsys_state *css,
6538 struct cftype *cft)
6539{
6540 return (u64)(unsigned long)current->cgroups;
6541}
6542
6543static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6544 struct cftype *cft)
6545{
6546 u64 count;
6547
6548 rcu_read_lock();
6549 count = atomic_read(&task_css_set(current)->refcount);
6550 rcu_read_unlock();
6551 return count;
6552}
6553
6554static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6555{
6556 struct cgrp_cset_link *link;
6557 struct css_set *cset;
6558 char *name_buf;
6559
6560 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6561 if (!name_buf)
6562 return -ENOMEM;
6563
6564 spin_lock_irq(&css_set_lock);
6565 rcu_read_lock();
6566 cset = rcu_dereference(current->cgroups);
6567 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6568 struct cgroup *c = link->cgrp;
6569
6570 cgroup_name(c, name_buf, NAME_MAX + 1);
6571 seq_printf(seq, "Root %d group %s\n",
6572 c->root->hierarchy_id, name_buf);
6573 }
6574 rcu_read_unlock();
6575 spin_unlock_irq(&css_set_lock);
6576 kfree(name_buf);
6577 return 0;
6578}
6579
6580#define MAX_TASKS_SHOWN_PER_CSS 25
6581static int cgroup_css_links_read(struct seq_file *seq, void *v)
6582{
6583 struct cgroup_subsys_state *css = seq_css(seq);
6584 struct cgrp_cset_link *link;
6585
6586 spin_lock_irq(&css_set_lock);
6587 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6588 struct css_set *cset = link->cset;
6589 struct task_struct *task;
6590 int count = 0;
6591
6592 seq_printf(seq, "css_set %p\n", cset);
6593
6594 list_for_each_entry(task, &cset->tasks, cg_list) {
6595 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6596 goto overflow;
6597 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6598 }
6599
6600 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6601 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6602 goto overflow;
6603 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6604 }
6605 continue;
6606 overflow:
6607 seq_puts(seq, " ...\n");
6608 }
6609 spin_unlock_irq(&css_set_lock);
6610 return 0;
6611}
6612
6613static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6614{
6615 return (!cgroup_is_populated(css->cgroup) &&
6616 !css_has_online_children(&css->cgroup->self));
6617}
6618
6619static struct cftype debug_files[] = {
6620 {
6621 .name = "taskcount",
6622 .read_u64 = debug_taskcount_read,
6623 },
6624
6625 {
6626 .name = "current_css_set",
6627 .read_u64 = current_css_set_read,
6628 },
6629
6630 {
6631 .name = "current_css_set_refcount",
6632 .read_u64 = current_css_set_refcount_read,
6633 },
6634
6635 {
6636 .name = "current_css_set_cg_links",
6637 .seq_show = current_css_set_cg_links_read,
6638 },
6639
6640 {
6641 .name = "cgroup_css_links",
6642 .seq_show = cgroup_css_links_read,
6643 },
6644
6645 {
6646 .name = "releasable",
6647 .read_u64 = releasable_read,
6648 },
6649
6650 { } /* terminate */
6651};
6652
6653struct cgroup_subsys debug_cgrp_subsys = {
6654 .css_alloc = debug_css_alloc,
6655 .css_free = debug_css_free,
6656 .legacy_cftypes = debug_files,
6657};
6658#endif /* CONFIG_CGROUP_DEBUG */
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#include <linux/cgroup.h>
30#include <linux/cred.h>
31#include <linux/ctype.h>
32#include <linux/errno.h>
33#include <linux/fs.h>
34#include <linux/init_task.h>
35#include <linux/kernel.h>
36#include <linux/list.h>
37#include <linux/mm.h>
38#include <linux/mutex.h>
39#include <linux/mount.h>
40#include <linux/pagemap.h>
41#include <linux/proc_fs.h>
42#include <linux/rcupdate.h>
43#include <linux/sched.h>
44#include <linux/backing-dev.h>
45#include <linux/seq_file.h>
46#include <linux/slab.h>
47#include <linux/magic.h>
48#include <linux/spinlock.h>
49#include <linux/string.h>
50#include <linux/sort.h>
51#include <linux/kmod.h>
52#include <linux/module.h>
53#include <linux/delayacct.h>
54#include <linux/cgroupstats.h>
55#include <linux/hash.h>
56#include <linux/namei.h>
57#include <linux/pid_namespace.h>
58#include <linux/idr.h>
59#include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
60#include <linux/eventfd.h>
61#include <linux/poll.h>
62#include <linux/flex_array.h> /* used in cgroup_attach_proc */
63#include <linux/kthread.h>
64
65#include <linux/atomic.h>
66
67/* css deactivation bias, makes css->refcnt negative to deny new trygets */
68#define CSS_DEACT_BIAS INT_MIN
69
70/*
71 * cgroup_mutex is the master lock. Any modification to cgroup or its
72 * hierarchy must be performed while holding it.
73 *
74 * cgroup_root_mutex nests inside cgroup_mutex and should be held to modify
75 * cgroupfs_root of any cgroup hierarchy - subsys list, flags,
76 * release_agent_path and so on. Modifying requires both cgroup_mutex and
77 * cgroup_root_mutex. Readers can acquire either of the two. This is to
78 * break the following locking order cycle.
79 *
80 * A. cgroup_mutex -> cred_guard_mutex -> s_type->i_mutex_key -> namespace_sem
81 * B. namespace_sem -> cgroup_mutex
82 *
83 * B happens only through cgroup_show_options() and using cgroup_root_mutex
84 * breaks it.
85 */
86static DEFINE_MUTEX(cgroup_mutex);
87static DEFINE_MUTEX(cgroup_root_mutex);
88
89/*
90 * Generate an array of cgroup subsystem pointers. At boot time, this is
91 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
92 * registered after that. The mutable section of this array is protected by
93 * cgroup_mutex.
94 */
95#define SUBSYS(_x) &_x ## _subsys,
96static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
97#include <linux/cgroup_subsys.h>
98};
99
100#define MAX_CGROUP_ROOT_NAMELEN 64
101
102/*
103 * A cgroupfs_root represents the root of a cgroup hierarchy,
104 * and may be associated with a superblock to form an active
105 * hierarchy
106 */
107struct cgroupfs_root {
108 struct super_block *sb;
109
110 /*
111 * The bitmask of subsystems intended to be attached to this
112 * hierarchy
113 */
114 unsigned long subsys_bits;
115
116 /* Unique id for this hierarchy. */
117 int hierarchy_id;
118
119 /* The bitmask of subsystems currently attached to this hierarchy */
120 unsigned long actual_subsys_bits;
121
122 /* A list running through the attached subsystems */
123 struct list_head subsys_list;
124
125 /* The root cgroup for this hierarchy */
126 struct cgroup top_cgroup;
127
128 /* Tracks how many cgroups are currently defined in hierarchy.*/
129 int number_of_cgroups;
130
131 /* A list running through the active hierarchies */
132 struct list_head root_list;
133
134 /* All cgroups on this root, cgroup_mutex protected */
135 struct list_head allcg_list;
136
137 /* Hierarchy-specific flags */
138 unsigned long flags;
139
140 /* The path to use for release notifications. */
141 char release_agent_path[PATH_MAX];
142
143 /* The name for this hierarchy - may be empty */
144 char name[MAX_CGROUP_ROOT_NAMELEN];
145};
146
147/*
148 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
149 * subsystems that are otherwise unattached - it never has more than a
150 * single cgroup, and all tasks are part of that cgroup.
151 */
152static struct cgroupfs_root rootnode;
153
154/*
155 * cgroupfs file entry, pointed to from leaf dentry->d_fsdata.
156 */
157struct cfent {
158 struct list_head node;
159 struct dentry *dentry;
160 struct cftype *type;
161};
162
163/*
164 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
165 * cgroup_subsys->use_id != 0.
166 */
167#define CSS_ID_MAX (65535)
168struct css_id {
169 /*
170 * The css to which this ID points. This pointer is set to valid value
171 * after cgroup is populated. If cgroup is removed, this will be NULL.
172 * This pointer is expected to be RCU-safe because destroy()
173 * is called after synchronize_rcu(). But for safe use, css_is_removed()
174 * css_tryget() should be used for avoiding race.
175 */
176 struct cgroup_subsys_state __rcu *css;
177 /*
178 * ID of this css.
179 */
180 unsigned short id;
181 /*
182 * Depth in hierarchy which this ID belongs to.
183 */
184 unsigned short depth;
185 /*
186 * ID is freed by RCU. (and lookup routine is RCU safe.)
187 */
188 struct rcu_head rcu_head;
189 /*
190 * Hierarchy of CSS ID belongs to.
191 */
192 unsigned short stack[0]; /* Array of Length (depth+1) */
193};
194
195/*
196 * cgroup_event represents events which userspace want to receive.
197 */
198struct cgroup_event {
199 /*
200 * Cgroup which the event belongs to.
201 */
202 struct cgroup *cgrp;
203 /*
204 * Control file which the event associated.
205 */
206 struct cftype *cft;
207 /*
208 * eventfd to signal userspace about the event.
209 */
210 struct eventfd_ctx *eventfd;
211 /*
212 * Each of these stored in a list by the cgroup.
213 */
214 struct list_head list;
215 /*
216 * All fields below needed to unregister event when
217 * userspace closes eventfd.
218 */
219 poll_table pt;
220 wait_queue_head_t *wqh;
221 wait_queue_t wait;
222 struct work_struct remove;
223};
224
225/* The list of hierarchy roots */
226
227static LIST_HEAD(roots);
228static int root_count;
229
230static DEFINE_IDA(hierarchy_ida);
231static int next_hierarchy_id;
232static DEFINE_SPINLOCK(hierarchy_id_lock);
233
234/* dummytop is a shorthand for the dummy hierarchy's top cgroup */
235#define dummytop (&rootnode.top_cgroup)
236
237/* This flag indicates whether tasks in the fork and exit paths should
238 * check for fork/exit handlers to call. This avoids us having to do
239 * extra work in the fork/exit path if none of the subsystems need to
240 * be called.
241 */
242static int need_forkexit_callback __read_mostly;
243
244#ifdef CONFIG_PROVE_LOCKING
245int cgroup_lock_is_held(void)
246{
247 return lockdep_is_held(&cgroup_mutex);
248}
249#else /* #ifdef CONFIG_PROVE_LOCKING */
250int cgroup_lock_is_held(void)
251{
252 return mutex_is_locked(&cgroup_mutex);
253}
254#endif /* #else #ifdef CONFIG_PROVE_LOCKING */
255
256EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
257
258static int css_unbias_refcnt(int refcnt)
259{
260 return refcnt >= 0 ? refcnt : refcnt - CSS_DEACT_BIAS;
261}
262
263/* the current nr of refs, always >= 0 whether @css is deactivated or not */
264static int css_refcnt(struct cgroup_subsys_state *css)
265{
266 int v = atomic_read(&css->refcnt);
267
268 return css_unbias_refcnt(v);
269}
270
271/* convenient tests for these bits */
272inline int cgroup_is_removed(const struct cgroup *cgrp)
273{
274 return test_bit(CGRP_REMOVED, &cgrp->flags);
275}
276
277/* bits in struct cgroupfs_root flags field */
278enum {
279 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
280};
281
282static int cgroup_is_releasable(const struct cgroup *cgrp)
283{
284 const int bits =
285 (1 << CGRP_RELEASABLE) |
286 (1 << CGRP_NOTIFY_ON_RELEASE);
287 return (cgrp->flags & bits) == bits;
288}
289
290static int notify_on_release(const struct cgroup *cgrp)
291{
292 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
293}
294
295static int clone_children(const struct cgroup *cgrp)
296{
297 return test_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
298}
299
300/*
301 * for_each_subsys() allows you to iterate on each subsystem attached to
302 * an active hierarchy
303 */
304#define for_each_subsys(_root, _ss) \
305list_for_each_entry(_ss, &_root->subsys_list, sibling)
306
307/* for_each_active_root() allows you to iterate across the active hierarchies */
308#define for_each_active_root(_root) \
309list_for_each_entry(_root, &roots, root_list)
310
311static inline struct cgroup *__d_cgrp(struct dentry *dentry)
312{
313 return dentry->d_fsdata;
314}
315
316static inline struct cfent *__d_cfe(struct dentry *dentry)
317{
318 return dentry->d_fsdata;
319}
320
321static inline struct cftype *__d_cft(struct dentry *dentry)
322{
323 return __d_cfe(dentry)->type;
324}
325
326/* the list of cgroups eligible for automatic release. Protected by
327 * release_list_lock */
328static LIST_HEAD(release_list);
329static DEFINE_RAW_SPINLOCK(release_list_lock);
330static void cgroup_release_agent(struct work_struct *work);
331static DECLARE_WORK(release_agent_work, cgroup_release_agent);
332static void check_for_release(struct cgroup *cgrp);
333
334/* Link structure for associating css_set objects with cgroups */
335struct cg_cgroup_link {
336 /*
337 * List running through cg_cgroup_links associated with a
338 * cgroup, anchored on cgroup->css_sets
339 */
340 struct list_head cgrp_link_list;
341 struct cgroup *cgrp;
342 /*
343 * List running through cg_cgroup_links pointing at a
344 * single css_set object, anchored on css_set->cg_links
345 */
346 struct list_head cg_link_list;
347 struct css_set *cg;
348};
349
350/* The default css_set - used by init and its children prior to any
351 * hierarchies being mounted. It contains a pointer to the root state
352 * for each subsystem. Also used to anchor the list of css_sets. Not
353 * reference-counted, to improve performance when child cgroups
354 * haven't been created.
355 */
356
357static struct css_set init_css_set;
358static struct cg_cgroup_link init_css_set_link;
359
360static int cgroup_init_idr(struct cgroup_subsys *ss,
361 struct cgroup_subsys_state *css);
362
363/* css_set_lock protects the list of css_set objects, and the
364 * chain of tasks off each css_set. Nests outside task->alloc_lock
365 * due to cgroup_iter_start() */
366static DEFINE_RWLOCK(css_set_lock);
367static int css_set_count;
368
369/*
370 * hash table for cgroup groups. This improves the performance to find
371 * an existing css_set. This hash doesn't (currently) take into
372 * account cgroups in empty hierarchies.
373 */
374#define CSS_SET_HASH_BITS 7
375#define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
376static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
377
378static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
379{
380 int i;
381 int index;
382 unsigned long tmp = 0UL;
383
384 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
385 tmp += (unsigned long)css[i];
386 tmp = (tmp >> 16) ^ tmp;
387
388 index = hash_long(tmp, CSS_SET_HASH_BITS);
389
390 return &css_set_table[index];
391}
392
393/* We don't maintain the lists running through each css_set to its
394 * task until after the first call to cgroup_iter_start(). This
395 * reduces the fork()/exit() overhead for people who have cgroups
396 * compiled into their kernel but not actually in use */
397static int use_task_css_set_links __read_mostly;
398
399static void __put_css_set(struct css_set *cg, int taskexit)
400{
401 struct cg_cgroup_link *link;
402 struct cg_cgroup_link *saved_link;
403 /*
404 * Ensure that the refcount doesn't hit zero while any readers
405 * can see it. Similar to atomic_dec_and_lock(), but for an
406 * rwlock
407 */
408 if (atomic_add_unless(&cg->refcount, -1, 1))
409 return;
410 write_lock(&css_set_lock);
411 if (!atomic_dec_and_test(&cg->refcount)) {
412 write_unlock(&css_set_lock);
413 return;
414 }
415
416 /* This css_set is dead. unlink it and release cgroup refcounts */
417 hlist_del(&cg->hlist);
418 css_set_count--;
419
420 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
421 cg_link_list) {
422 struct cgroup *cgrp = link->cgrp;
423 list_del(&link->cg_link_list);
424 list_del(&link->cgrp_link_list);
425 if (atomic_dec_and_test(&cgrp->count) &&
426 notify_on_release(cgrp)) {
427 if (taskexit)
428 set_bit(CGRP_RELEASABLE, &cgrp->flags);
429 check_for_release(cgrp);
430 }
431
432 kfree(link);
433 }
434
435 write_unlock(&css_set_lock);
436 kfree_rcu(cg, rcu_head);
437}
438
439/*
440 * refcounted get/put for css_set objects
441 */
442static inline void get_css_set(struct css_set *cg)
443{
444 atomic_inc(&cg->refcount);
445}
446
447static inline void put_css_set(struct css_set *cg)
448{
449 __put_css_set(cg, 0);
450}
451
452static inline void put_css_set_taskexit(struct css_set *cg)
453{
454 __put_css_set(cg, 1);
455}
456
457/*
458 * compare_css_sets - helper function for find_existing_css_set().
459 * @cg: candidate css_set being tested
460 * @old_cg: existing css_set for a task
461 * @new_cgrp: cgroup that's being entered by the task
462 * @template: desired set of css pointers in css_set (pre-calculated)
463 *
464 * Returns true if "cg" matches "old_cg" except for the hierarchy
465 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
466 */
467static bool compare_css_sets(struct css_set *cg,
468 struct css_set *old_cg,
469 struct cgroup *new_cgrp,
470 struct cgroup_subsys_state *template[])
471{
472 struct list_head *l1, *l2;
473
474 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
475 /* Not all subsystems matched */
476 return false;
477 }
478
479 /*
480 * Compare cgroup pointers in order to distinguish between
481 * different cgroups in heirarchies with no subsystems. We
482 * could get by with just this check alone (and skip the
483 * memcmp above) but on most setups the memcmp check will
484 * avoid the need for this more expensive check on almost all
485 * candidates.
486 */
487
488 l1 = &cg->cg_links;
489 l2 = &old_cg->cg_links;
490 while (1) {
491 struct cg_cgroup_link *cgl1, *cgl2;
492 struct cgroup *cg1, *cg2;
493
494 l1 = l1->next;
495 l2 = l2->next;
496 /* See if we reached the end - both lists are equal length. */
497 if (l1 == &cg->cg_links) {
498 BUG_ON(l2 != &old_cg->cg_links);
499 break;
500 } else {
501 BUG_ON(l2 == &old_cg->cg_links);
502 }
503 /* Locate the cgroups associated with these links. */
504 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
505 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
506 cg1 = cgl1->cgrp;
507 cg2 = cgl2->cgrp;
508 /* Hierarchies should be linked in the same order. */
509 BUG_ON(cg1->root != cg2->root);
510
511 /*
512 * If this hierarchy is the hierarchy of the cgroup
513 * that's changing, then we need to check that this
514 * css_set points to the new cgroup; if it's any other
515 * hierarchy, then this css_set should point to the
516 * same cgroup as the old css_set.
517 */
518 if (cg1->root == new_cgrp->root) {
519 if (cg1 != new_cgrp)
520 return false;
521 } else {
522 if (cg1 != cg2)
523 return false;
524 }
525 }
526 return true;
527}
528
529/*
530 * find_existing_css_set() is a helper for
531 * find_css_set(), and checks to see whether an existing
532 * css_set is suitable.
533 *
534 * oldcg: the cgroup group that we're using before the cgroup
535 * transition
536 *
537 * cgrp: the cgroup that we're moving into
538 *
539 * template: location in which to build the desired set of subsystem
540 * state objects for the new cgroup group
541 */
542static struct css_set *find_existing_css_set(
543 struct css_set *oldcg,
544 struct cgroup *cgrp,
545 struct cgroup_subsys_state *template[])
546{
547 int i;
548 struct cgroupfs_root *root = cgrp->root;
549 struct hlist_head *hhead;
550 struct hlist_node *node;
551 struct css_set *cg;
552
553 /*
554 * Build the set of subsystem state objects that we want to see in the
555 * new css_set. while subsystems can change globally, the entries here
556 * won't change, so no need for locking.
557 */
558 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
559 if (root->subsys_bits & (1UL << i)) {
560 /* Subsystem is in this hierarchy. So we want
561 * the subsystem state from the new
562 * cgroup */
563 template[i] = cgrp->subsys[i];
564 } else {
565 /* Subsystem is not in this hierarchy, so we
566 * don't want to change the subsystem state */
567 template[i] = oldcg->subsys[i];
568 }
569 }
570
571 hhead = css_set_hash(template);
572 hlist_for_each_entry(cg, node, hhead, hlist) {
573 if (!compare_css_sets(cg, oldcg, cgrp, template))
574 continue;
575
576 /* This css_set matches what we need */
577 return cg;
578 }
579
580 /* No existing cgroup group matched */
581 return NULL;
582}
583
584static void free_cg_links(struct list_head *tmp)
585{
586 struct cg_cgroup_link *link;
587 struct cg_cgroup_link *saved_link;
588
589 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
590 list_del(&link->cgrp_link_list);
591 kfree(link);
592 }
593}
594
595/*
596 * allocate_cg_links() allocates "count" cg_cgroup_link structures
597 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
598 * success or a negative error
599 */
600static int allocate_cg_links(int count, struct list_head *tmp)
601{
602 struct cg_cgroup_link *link;
603 int i;
604 INIT_LIST_HEAD(tmp);
605 for (i = 0; i < count; i++) {
606 link = kmalloc(sizeof(*link), GFP_KERNEL);
607 if (!link) {
608 free_cg_links(tmp);
609 return -ENOMEM;
610 }
611 list_add(&link->cgrp_link_list, tmp);
612 }
613 return 0;
614}
615
616/**
617 * link_css_set - a helper function to link a css_set to a cgroup
618 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
619 * @cg: the css_set to be linked
620 * @cgrp: the destination cgroup
621 */
622static void link_css_set(struct list_head *tmp_cg_links,
623 struct css_set *cg, struct cgroup *cgrp)
624{
625 struct cg_cgroup_link *link;
626
627 BUG_ON(list_empty(tmp_cg_links));
628 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
629 cgrp_link_list);
630 link->cg = cg;
631 link->cgrp = cgrp;
632 atomic_inc(&cgrp->count);
633 list_move(&link->cgrp_link_list, &cgrp->css_sets);
634 /*
635 * Always add links to the tail of the list so that the list
636 * is sorted by order of hierarchy creation
637 */
638 list_add_tail(&link->cg_link_list, &cg->cg_links);
639}
640
641/*
642 * find_css_set() takes an existing cgroup group and a
643 * cgroup object, and returns a css_set object that's
644 * equivalent to the old group, but with the given cgroup
645 * substituted into the appropriate hierarchy. Must be called with
646 * cgroup_mutex held
647 */
648static struct css_set *find_css_set(
649 struct css_set *oldcg, struct cgroup *cgrp)
650{
651 struct css_set *res;
652 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
653
654 struct list_head tmp_cg_links;
655
656 struct hlist_head *hhead;
657 struct cg_cgroup_link *link;
658
659 /* First see if we already have a cgroup group that matches
660 * the desired set */
661 read_lock(&css_set_lock);
662 res = find_existing_css_set(oldcg, cgrp, template);
663 if (res)
664 get_css_set(res);
665 read_unlock(&css_set_lock);
666
667 if (res)
668 return res;
669
670 res = kmalloc(sizeof(*res), GFP_KERNEL);
671 if (!res)
672 return NULL;
673
674 /* Allocate all the cg_cgroup_link objects that we'll need */
675 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
676 kfree(res);
677 return NULL;
678 }
679
680 atomic_set(&res->refcount, 1);
681 INIT_LIST_HEAD(&res->cg_links);
682 INIT_LIST_HEAD(&res->tasks);
683 INIT_HLIST_NODE(&res->hlist);
684
685 /* Copy the set of subsystem state objects generated in
686 * find_existing_css_set() */
687 memcpy(res->subsys, template, sizeof(res->subsys));
688
689 write_lock(&css_set_lock);
690 /* Add reference counts and links from the new css_set. */
691 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
692 struct cgroup *c = link->cgrp;
693 if (c->root == cgrp->root)
694 c = cgrp;
695 link_css_set(&tmp_cg_links, res, c);
696 }
697
698 BUG_ON(!list_empty(&tmp_cg_links));
699
700 css_set_count++;
701
702 /* Add this cgroup group to the hash table */
703 hhead = css_set_hash(res->subsys);
704 hlist_add_head(&res->hlist, hhead);
705
706 write_unlock(&css_set_lock);
707
708 return res;
709}
710
711/*
712 * Return the cgroup for "task" from the given hierarchy. Must be
713 * called with cgroup_mutex held.
714 */
715static struct cgroup *task_cgroup_from_root(struct task_struct *task,
716 struct cgroupfs_root *root)
717{
718 struct css_set *css;
719 struct cgroup *res = NULL;
720
721 BUG_ON(!mutex_is_locked(&cgroup_mutex));
722 read_lock(&css_set_lock);
723 /*
724 * No need to lock the task - since we hold cgroup_mutex the
725 * task can't change groups, so the only thing that can happen
726 * is that it exits and its css is set back to init_css_set.
727 */
728 css = task->cgroups;
729 if (css == &init_css_set) {
730 res = &root->top_cgroup;
731 } else {
732 struct cg_cgroup_link *link;
733 list_for_each_entry(link, &css->cg_links, cg_link_list) {
734 struct cgroup *c = link->cgrp;
735 if (c->root == root) {
736 res = c;
737 break;
738 }
739 }
740 }
741 read_unlock(&css_set_lock);
742 BUG_ON(!res);
743 return res;
744}
745
746/*
747 * There is one global cgroup mutex. We also require taking
748 * task_lock() when dereferencing a task's cgroup subsys pointers.
749 * See "The task_lock() exception", at the end of this comment.
750 *
751 * A task must hold cgroup_mutex to modify cgroups.
752 *
753 * Any task can increment and decrement the count field without lock.
754 * So in general, code holding cgroup_mutex can't rely on the count
755 * field not changing. However, if the count goes to zero, then only
756 * cgroup_attach_task() can increment it again. Because a count of zero
757 * means that no tasks are currently attached, therefore there is no
758 * way a task attached to that cgroup can fork (the other way to
759 * increment the count). So code holding cgroup_mutex can safely
760 * assume that if the count is zero, it will stay zero. Similarly, if
761 * a task holds cgroup_mutex on a cgroup with zero count, it
762 * knows that the cgroup won't be removed, as cgroup_rmdir()
763 * needs that mutex.
764 *
765 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
766 * (usually) take cgroup_mutex. These are the two most performance
767 * critical pieces of code here. The exception occurs on cgroup_exit(),
768 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
769 * is taken, and if the cgroup count is zero, a usermode call made
770 * to the release agent with the name of the cgroup (path relative to
771 * the root of cgroup file system) as the argument.
772 *
773 * A cgroup can only be deleted if both its 'count' of using tasks
774 * is zero, and its list of 'children' cgroups is empty. Since all
775 * tasks in the system use _some_ cgroup, and since there is always at
776 * least one task in the system (init, pid == 1), therefore, top_cgroup
777 * always has either children cgroups and/or using tasks. So we don't
778 * need a special hack to ensure that top_cgroup cannot be deleted.
779 *
780 * The task_lock() exception
781 *
782 * The need for this exception arises from the action of
783 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
784 * another. It does so using cgroup_mutex, however there are
785 * several performance critical places that need to reference
786 * task->cgroup without the expense of grabbing a system global
787 * mutex. Therefore except as noted below, when dereferencing or, as
788 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
789 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
790 * the task_struct routinely used for such matters.
791 *
792 * P.S. One more locking exception. RCU is used to guard the
793 * update of a tasks cgroup pointer by cgroup_attach_task()
794 */
795
796/**
797 * cgroup_lock - lock out any changes to cgroup structures
798 *
799 */
800void cgroup_lock(void)
801{
802 mutex_lock(&cgroup_mutex);
803}
804EXPORT_SYMBOL_GPL(cgroup_lock);
805
806/**
807 * cgroup_unlock - release lock on cgroup changes
808 *
809 * Undo the lock taken in a previous cgroup_lock() call.
810 */
811void cgroup_unlock(void)
812{
813 mutex_unlock(&cgroup_mutex);
814}
815EXPORT_SYMBOL_GPL(cgroup_unlock);
816
817/*
818 * A couple of forward declarations required, due to cyclic reference loop:
819 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
820 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
821 * -> cgroup_mkdir.
822 */
823
824static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
825static struct dentry *cgroup_lookup(struct inode *, struct dentry *, struct nameidata *);
826static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
827static int cgroup_populate_dir(struct cgroup *cgrp);
828static const struct inode_operations cgroup_dir_inode_operations;
829static const struct file_operations proc_cgroupstats_operations;
830
831static struct backing_dev_info cgroup_backing_dev_info = {
832 .name = "cgroup",
833 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
834};
835
836static int alloc_css_id(struct cgroup_subsys *ss,
837 struct cgroup *parent, struct cgroup *child);
838
839static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
840{
841 struct inode *inode = new_inode(sb);
842
843 if (inode) {
844 inode->i_ino = get_next_ino();
845 inode->i_mode = mode;
846 inode->i_uid = current_fsuid();
847 inode->i_gid = current_fsgid();
848 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
849 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
850 }
851 return inode;
852}
853
854/*
855 * Call subsys's pre_destroy handler.
856 * This is called before css refcnt check.
857 */
858static int cgroup_call_pre_destroy(struct cgroup *cgrp)
859{
860 struct cgroup_subsys *ss;
861 int ret = 0;
862
863 for_each_subsys(cgrp->root, ss) {
864 if (!ss->pre_destroy)
865 continue;
866
867 ret = ss->pre_destroy(cgrp);
868 if (ret) {
869 /* ->pre_destroy() failure is being deprecated */
870 WARN_ON_ONCE(!ss->__DEPRECATED_clear_css_refs);
871 break;
872 }
873 }
874
875 return ret;
876}
877
878static void cgroup_diput(struct dentry *dentry, struct inode *inode)
879{
880 /* is dentry a directory ? if so, kfree() associated cgroup */
881 if (S_ISDIR(inode->i_mode)) {
882 struct cgroup *cgrp = dentry->d_fsdata;
883 struct cgroup_subsys *ss;
884 BUG_ON(!(cgroup_is_removed(cgrp)));
885 /* It's possible for external users to be holding css
886 * reference counts on a cgroup; css_put() needs to
887 * be able to access the cgroup after decrementing
888 * the reference count in order to know if it needs to
889 * queue the cgroup to be handled by the release
890 * agent */
891 synchronize_rcu();
892
893 mutex_lock(&cgroup_mutex);
894 /*
895 * Release the subsystem state objects.
896 */
897 for_each_subsys(cgrp->root, ss)
898 ss->destroy(cgrp);
899
900 cgrp->root->number_of_cgroups--;
901 mutex_unlock(&cgroup_mutex);
902
903 /*
904 * Drop the active superblock reference that we took when we
905 * created the cgroup
906 */
907 deactivate_super(cgrp->root->sb);
908
909 /*
910 * if we're getting rid of the cgroup, refcount should ensure
911 * that there are no pidlists left.
912 */
913 BUG_ON(!list_empty(&cgrp->pidlists));
914
915 kfree_rcu(cgrp, rcu_head);
916 } else {
917 struct cfent *cfe = __d_cfe(dentry);
918 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
919
920 WARN_ONCE(!list_empty(&cfe->node) &&
921 cgrp != &cgrp->root->top_cgroup,
922 "cfe still linked for %s\n", cfe->type->name);
923 kfree(cfe);
924 }
925 iput(inode);
926}
927
928static int cgroup_delete(const struct dentry *d)
929{
930 return 1;
931}
932
933static void remove_dir(struct dentry *d)
934{
935 struct dentry *parent = dget(d->d_parent);
936
937 d_delete(d);
938 simple_rmdir(parent->d_inode, d);
939 dput(parent);
940}
941
942static int cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
943{
944 struct cfent *cfe;
945
946 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
947 lockdep_assert_held(&cgroup_mutex);
948
949 list_for_each_entry(cfe, &cgrp->files, node) {
950 struct dentry *d = cfe->dentry;
951
952 if (cft && cfe->type != cft)
953 continue;
954
955 dget(d);
956 d_delete(d);
957 simple_unlink(cgrp->dentry->d_inode, d);
958 list_del_init(&cfe->node);
959 dput(d);
960
961 return 0;
962 }
963 return -ENOENT;
964}
965
966static void cgroup_clear_directory(struct dentry *dir)
967{
968 struct cgroup *cgrp = __d_cgrp(dir);
969
970 while (!list_empty(&cgrp->files))
971 cgroup_rm_file(cgrp, NULL);
972}
973
974/*
975 * NOTE : the dentry must have been dget()'ed
976 */
977static void cgroup_d_remove_dir(struct dentry *dentry)
978{
979 struct dentry *parent;
980
981 cgroup_clear_directory(dentry);
982
983 parent = dentry->d_parent;
984 spin_lock(&parent->d_lock);
985 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
986 list_del_init(&dentry->d_u.d_child);
987 spin_unlock(&dentry->d_lock);
988 spin_unlock(&parent->d_lock);
989 remove_dir(dentry);
990}
991
992/*
993 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
994 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
995 * reference to css->refcnt. In general, this refcnt is expected to goes down
996 * to zero, soon.
997 *
998 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
999 */
1000static DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
1001
1002static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
1003{
1004 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
1005 wake_up_all(&cgroup_rmdir_waitq);
1006}
1007
1008void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
1009{
1010 css_get(css);
1011}
1012
1013void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
1014{
1015 cgroup_wakeup_rmdir_waiter(css->cgroup);
1016 css_put(css);
1017}
1018
1019/*
1020 * Call with cgroup_mutex held. Drops reference counts on modules, including
1021 * any duplicate ones that parse_cgroupfs_options took. If this function
1022 * returns an error, no reference counts are touched.
1023 */
1024static int rebind_subsystems(struct cgroupfs_root *root,
1025 unsigned long final_bits)
1026{
1027 unsigned long added_bits, removed_bits;
1028 struct cgroup *cgrp = &root->top_cgroup;
1029 int i;
1030
1031 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1032 BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
1033
1034 removed_bits = root->actual_subsys_bits & ~final_bits;
1035 added_bits = final_bits & ~root->actual_subsys_bits;
1036 /* Check that any added subsystems are currently free */
1037 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1038 unsigned long bit = 1UL << i;
1039 struct cgroup_subsys *ss = subsys[i];
1040 if (!(bit & added_bits))
1041 continue;
1042 /*
1043 * Nobody should tell us to do a subsys that doesn't exist:
1044 * parse_cgroupfs_options should catch that case and refcounts
1045 * ensure that subsystems won't disappear once selected.
1046 */
1047 BUG_ON(ss == NULL);
1048 if (ss->root != &rootnode) {
1049 /* Subsystem isn't free */
1050 return -EBUSY;
1051 }
1052 }
1053
1054 /* Currently we don't handle adding/removing subsystems when
1055 * any child cgroups exist. This is theoretically supportable
1056 * but involves complex error handling, so it's being left until
1057 * later */
1058 if (root->number_of_cgroups > 1)
1059 return -EBUSY;
1060
1061 /* Process each subsystem */
1062 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1063 struct cgroup_subsys *ss = subsys[i];
1064 unsigned long bit = 1UL << i;
1065 if (bit & added_bits) {
1066 /* We're binding this subsystem to this hierarchy */
1067 BUG_ON(ss == NULL);
1068 BUG_ON(cgrp->subsys[i]);
1069 BUG_ON(!dummytop->subsys[i]);
1070 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
1071 mutex_lock(&ss->hierarchy_mutex);
1072 cgrp->subsys[i] = dummytop->subsys[i];
1073 cgrp->subsys[i]->cgroup = cgrp;
1074 list_move(&ss->sibling, &root->subsys_list);
1075 ss->root = root;
1076 if (ss->bind)
1077 ss->bind(cgrp);
1078 mutex_unlock(&ss->hierarchy_mutex);
1079 /* refcount was already taken, and we're keeping it */
1080 } else if (bit & removed_bits) {
1081 /* We're removing this subsystem */
1082 BUG_ON(ss == NULL);
1083 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
1084 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1085 mutex_lock(&ss->hierarchy_mutex);
1086 if (ss->bind)
1087 ss->bind(dummytop);
1088 dummytop->subsys[i]->cgroup = dummytop;
1089 cgrp->subsys[i] = NULL;
1090 subsys[i]->root = &rootnode;
1091 list_move(&ss->sibling, &rootnode.subsys_list);
1092 mutex_unlock(&ss->hierarchy_mutex);
1093 /* subsystem is now free - drop reference on module */
1094 module_put(ss->module);
1095 } else if (bit & final_bits) {
1096 /* Subsystem state should already exist */
1097 BUG_ON(ss == NULL);
1098 BUG_ON(!cgrp->subsys[i]);
1099 /*
1100 * a refcount was taken, but we already had one, so
1101 * drop the extra reference.
1102 */
1103 module_put(ss->module);
1104#ifdef CONFIG_MODULE_UNLOAD
1105 BUG_ON(ss->module && !module_refcount(ss->module));
1106#endif
1107 } else {
1108 /* Subsystem state shouldn't exist */
1109 BUG_ON(cgrp->subsys[i]);
1110 }
1111 }
1112 root->subsys_bits = root->actual_subsys_bits = final_bits;
1113 synchronize_rcu();
1114
1115 return 0;
1116}
1117
1118static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1119{
1120 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1121 struct cgroup_subsys *ss;
1122
1123 mutex_lock(&cgroup_root_mutex);
1124 for_each_subsys(root, ss)
1125 seq_printf(seq, ",%s", ss->name);
1126 if (test_bit(ROOT_NOPREFIX, &root->flags))
1127 seq_puts(seq, ",noprefix");
1128 if (strlen(root->release_agent_path))
1129 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1130 if (clone_children(&root->top_cgroup))
1131 seq_puts(seq, ",clone_children");
1132 if (strlen(root->name))
1133 seq_printf(seq, ",name=%s", root->name);
1134 mutex_unlock(&cgroup_root_mutex);
1135 return 0;
1136}
1137
1138struct cgroup_sb_opts {
1139 unsigned long subsys_bits;
1140 unsigned long flags;
1141 char *release_agent;
1142 bool clone_children;
1143 char *name;
1144 /* User explicitly requested empty subsystem */
1145 bool none;
1146
1147 struct cgroupfs_root *new_root;
1148
1149};
1150
1151/*
1152 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1153 * with cgroup_mutex held to protect the subsys[] array. This function takes
1154 * refcounts on subsystems to be used, unless it returns error, in which case
1155 * no refcounts are taken.
1156 */
1157static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1158{
1159 char *token, *o = data;
1160 bool all_ss = false, one_ss = false;
1161 unsigned long mask = (unsigned long)-1;
1162 int i;
1163 bool module_pin_failed = false;
1164
1165 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1166
1167#ifdef CONFIG_CPUSETS
1168 mask = ~(1UL << cpuset_subsys_id);
1169#endif
1170
1171 memset(opts, 0, sizeof(*opts));
1172
1173 while ((token = strsep(&o, ",")) != NULL) {
1174 if (!*token)
1175 return -EINVAL;
1176 if (!strcmp(token, "none")) {
1177 /* Explicitly have no subsystems */
1178 opts->none = true;
1179 continue;
1180 }
1181 if (!strcmp(token, "all")) {
1182 /* Mutually exclusive option 'all' + subsystem name */
1183 if (one_ss)
1184 return -EINVAL;
1185 all_ss = true;
1186 continue;
1187 }
1188 if (!strcmp(token, "noprefix")) {
1189 set_bit(ROOT_NOPREFIX, &opts->flags);
1190 continue;
1191 }
1192 if (!strcmp(token, "clone_children")) {
1193 opts->clone_children = true;
1194 continue;
1195 }
1196 if (!strncmp(token, "release_agent=", 14)) {
1197 /* Specifying two release agents is forbidden */
1198 if (opts->release_agent)
1199 return -EINVAL;
1200 opts->release_agent =
1201 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1202 if (!opts->release_agent)
1203 return -ENOMEM;
1204 continue;
1205 }
1206 if (!strncmp(token, "name=", 5)) {
1207 const char *name = token + 5;
1208 /* Can't specify an empty name */
1209 if (!strlen(name))
1210 return -EINVAL;
1211 /* Must match [\w.-]+ */
1212 for (i = 0; i < strlen(name); i++) {
1213 char c = name[i];
1214 if (isalnum(c))
1215 continue;
1216 if ((c == '.') || (c == '-') || (c == '_'))
1217 continue;
1218 return -EINVAL;
1219 }
1220 /* Specifying two names is forbidden */
1221 if (opts->name)
1222 return -EINVAL;
1223 opts->name = kstrndup(name,
1224 MAX_CGROUP_ROOT_NAMELEN - 1,
1225 GFP_KERNEL);
1226 if (!opts->name)
1227 return -ENOMEM;
1228
1229 continue;
1230 }
1231
1232 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1233 struct cgroup_subsys *ss = subsys[i];
1234 if (ss == NULL)
1235 continue;
1236 if (strcmp(token, ss->name))
1237 continue;
1238 if (ss->disabled)
1239 continue;
1240
1241 /* Mutually exclusive option 'all' + subsystem name */
1242 if (all_ss)
1243 return -EINVAL;
1244 set_bit(i, &opts->subsys_bits);
1245 one_ss = true;
1246
1247 break;
1248 }
1249 if (i == CGROUP_SUBSYS_COUNT)
1250 return -ENOENT;
1251 }
1252
1253 /*
1254 * If the 'all' option was specified select all the subsystems,
1255 * otherwise if 'none', 'name=' and a subsystem name options
1256 * were not specified, let's default to 'all'
1257 */
1258 if (all_ss || (!one_ss && !opts->none && !opts->name)) {
1259 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1260 struct cgroup_subsys *ss = subsys[i];
1261 if (ss == NULL)
1262 continue;
1263 if (ss->disabled)
1264 continue;
1265 set_bit(i, &opts->subsys_bits);
1266 }
1267 }
1268
1269 /* Consistency checks */
1270
1271 /*
1272 * Option noprefix was introduced just for backward compatibility
1273 * with the old cpuset, so we allow noprefix only if mounting just
1274 * the cpuset subsystem.
1275 */
1276 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1277 (opts->subsys_bits & mask))
1278 return -EINVAL;
1279
1280
1281 /* Can't specify "none" and some subsystems */
1282 if (opts->subsys_bits && opts->none)
1283 return -EINVAL;
1284
1285 /*
1286 * We either have to specify by name or by subsystems. (So all
1287 * empty hierarchies must have a name).
1288 */
1289 if (!opts->subsys_bits && !opts->name)
1290 return -EINVAL;
1291
1292 /*
1293 * Grab references on all the modules we'll need, so the subsystems
1294 * don't dance around before rebind_subsystems attaches them. This may
1295 * take duplicate reference counts on a subsystem that's already used,
1296 * but rebind_subsystems handles this case.
1297 */
1298 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1299 unsigned long bit = 1UL << i;
1300
1301 if (!(bit & opts->subsys_bits))
1302 continue;
1303 if (!try_module_get(subsys[i]->module)) {
1304 module_pin_failed = true;
1305 break;
1306 }
1307 }
1308 if (module_pin_failed) {
1309 /*
1310 * oops, one of the modules was going away. this means that we
1311 * raced with a module_delete call, and to the user this is
1312 * essentially a "subsystem doesn't exist" case.
1313 */
1314 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1315 /* drop refcounts only on the ones we took */
1316 unsigned long bit = 1UL << i;
1317
1318 if (!(bit & opts->subsys_bits))
1319 continue;
1320 module_put(subsys[i]->module);
1321 }
1322 return -ENOENT;
1323 }
1324
1325 return 0;
1326}
1327
1328static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1329{
1330 int i;
1331 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1332 unsigned long bit = 1UL << i;
1333
1334 if (!(bit & subsys_bits))
1335 continue;
1336 module_put(subsys[i]->module);
1337 }
1338}
1339
1340static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1341{
1342 int ret = 0;
1343 struct cgroupfs_root *root = sb->s_fs_info;
1344 struct cgroup *cgrp = &root->top_cgroup;
1345 struct cgroup_sb_opts opts;
1346
1347 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1348 mutex_lock(&cgroup_mutex);
1349 mutex_lock(&cgroup_root_mutex);
1350
1351 /* See what subsystems are wanted */
1352 ret = parse_cgroupfs_options(data, &opts);
1353 if (ret)
1354 goto out_unlock;
1355
1356 /* See feature-removal-schedule.txt */
1357 if (opts.subsys_bits != root->actual_subsys_bits || opts.release_agent)
1358 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1359 task_tgid_nr(current), current->comm);
1360
1361 /* Don't allow flags or name to change at remount */
1362 if (opts.flags != root->flags ||
1363 (opts.name && strcmp(opts.name, root->name))) {
1364 ret = -EINVAL;
1365 drop_parsed_module_refcounts(opts.subsys_bits);
1366 goto out_unlock;
1367 }
1368
1369 ret = rebind_subsystems(root, opts.subsys_bits);
1370 if (ret) {
1371 drop_parsed_module_refcounts(opts.subsys_bits);
1372 goto out_unlock;
1373 }
1374
1375 /* clear out any existing files and repopulate subsystem files */
1376 cgroup_clear_directory(cgrp->dentry);
1377 cgroup_populate_dir(cgrp);
1378
1379 if (opts.release_agent)
1380 strcpy(root->release_agent_path, opts.release_agent);
1381 out_unlock:
1382 kfree(opts.release_agent);
1383 kfree(opts.name);
1384 mutex_unlock(&cgroup_root_mutex);
1385 mutex_unlock(&cgroup_mutex);
1386 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1387 return ret;
1388}
1389
1390static const struct super_operations cgroup_ops = {
1391 .statfs = simple_statfs,
1392 .drop_inode = generic_delete_inode,
1393 .show_options = cgroup_show_options,
1394 .remount_fs = cgroup_remount,
1395};
1396
1397static void init_cgroup_housekeeping(struct cgroup *cgrp)
1398{
1399 INIT_LIST_HEAD(&cgrp->sibling);
1400 INIT_LIST_HEAD(&cgrp->children);
1401 INIT_LIST_HEAD(&cgrp->files);
1402 INIT_LIST_HEAD(&cgrp->css_sets);
1403 INIT_LIST_HEAD(&cgrp->release_list);
1404 INIT_LIST_HEAD(&cgrp->pidlists);
1405 mutex_init(&cgrp->pidlist_mutex);
1406 INIT_LIST_HEAD(&cgrp->event_list);
1407 spin_lock_init(&cgrp->event_list_lock);
1408}
1409
1410static void init_cgroup_root(struct cgroupfs_root *root)
1411{
1412 struct cgroup *cgrp = &root->top_cgroup;
1413
1414 INIT_LIST_HEAD(&root->subsys_list);
1415 INIT_LIST_HEAD(&root->root_list);
1416 INIT_LIST_HEAD(&root->allcg_list);
1417 root->number_of_cgroups = 1;
1418 cgrp->root = root;
1419 cgrp->top_cgroup = cgrp;
1420 list_add_tail(&cgrp->allcg_node, &root->allcg_list);
1421 init_cgroup_housekeeping(cgrp);
1422}
1423
1424static bool init_root_id(struct cgroupfs_root *root)
1425{
1426 int ret = 0;
1427
1428 do {
1429 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1430 return false;
1431 spin_lock(&hierarchy_id_lock);
1432 /* Try to allocate the next unused ID */
1433 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1434 &root->hierarchy_id);
1435 if (ret == -ENOSPC)
1436 /* Try again starting from 0 */
1437 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1438 if (!ret) {
1439 next_hierarchy_id = root->hierarchy_id + 1;
1440 } else if (ret != -EAGAIN) {
1441 /* Can only get here if the 31-bit IDR is full ... */
1442 BUG_ON(ret);
1443 }
1444 spin_unlock(&hierarchy_id_lock);
1445 } while (ret);
1446 return true;
1447}
1448
1449static int cgroup_test_super(struct super_block *sb, void *data)
1450{
1451 struct cgroup_sb_opts *opts = data;
1452 struct cgroupfs_root *root = sb->s_fs_info;
1453
1454 /* If we asked for a name then it must match */
1455 if (opts->name && strcmp(opts->name, root->name))
1456 return 0;
1457
1458 /*
1459 * If we asked for subsystems (or explicitly for no
1460 * subsystems) then they must match
1461 */
1462 if ((opts->subsys_bits || opts->none)
1463 && (opts->subsys_bits != root->subsys_bits))
1464 return 0;
1465
1466 return 1;
1467}
1468
1469static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1470{
1471 struct cgroupfs_root *root;
1472
1473 if (!opts->subsys_bits && !opts->none)
1474 return NULL;
1475
1476 root = kzalloc(sizeof(*root), GFP_KERNEL);
1477 if (!root)
1478 return ERR_PTR(-ENOMEM);
1479
1480 if (!init_root_id(root)) {
1481 kfree(root);
1482 return ERR_PTR(-ENOMEM);
1483 }
1484 init_cgroup_root(root);
1485
1486 root->subsys_bits = opts->subsys_bits;
1487 root->flags = opts->flags;
1488 if (opts->release_agent)
1489 strcpy(root->release_agent_path, opts->release_agent);
1490 if (opts->name)
1491 strcpy(root->name, opts->name);
1492 if (opts->clone_children)
1493 set_bit(CGRP_CLONE_CHILDREN, &root->top_cgroup.flags);
1494 return root;
1495}
1496
1497static void cgroup_drop_root(struct cgroupfs_root *root)
1498{
1499 if (!root)
1500 return;
1501
1502 BUG_ON(!root->hierarchy_id);
1503 spin_lock(&hierarchy_id_lock);
1504 ida_remove(&hierarchy_ida, root->hierarchy_id);
1505 spin_unlock(&hierarchy_id_lock);
1506 kfree(root);
1507}
1508
1509static int cgroup_set_super(struct super_block *sb, void *data)
1510{
1511 int ret;
1512 struct cgroup_sb_opts *opts = data;
1513
1514 /* If we don't have a new root, we can't set up a new sb */
1515 if (!opts->new_root)
1516 return -EINVAL;
1517
1518 BUG_ON(!opts->subsys_bits && !opts->none);
1519
1520 ret = set_anon_super(sb, NULL);
1521 if (ret)
1522 return ret;
1523
1524 sb->s_fs_info = opts->new_root;
1525 opts->new_root->sb = sb;
1526
1527 sb->s_blocksize = PAGE_CACHE_SIZE;
1528 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1529 sb->s_magic = CGROUP_SUPER_MAGIC;
1530 sb->s_op = &cgroup_ops;
1531
1532 return 0;
1533}
1534
1535static int cgroup_get_rootdir(struct super_block *sb)
1536{
1537 static const struct dentry_operations cgroup_dops = {
1538 .d_iput = cgroup_diput,
1539 .d_delete = cgroup_delete,
1540 };
1541
1542 struct inode *inode =
1543 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1544
1545 if (!inode)
1546 return -ENOMEM;
1547
1548 inode->i_fop = &simple_dir_operations;
1549 inode->i_op = &cgroup_dir_inode_operations;
1550 /* directories start off with i_nlink == 2 (for "." entry) */
1551 inc_nlink(inode);
1552 sb->s_root = d_make_root(inode);
1553 if (!sb->s_root)
1554 return -ENOMEM;
1555 /* for everything else we want ->d_op set */
1556 sb->s_d_op = &cgroup_dops;
1557 return 0;
1558}
1559
1560static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1561 int flags, const char *unused_dev_name,
1562 void *data)
1563{
1564 struct cgroup_sb_opts opts;
1565 struct cgroupfs_root *root;
1566 int ret = 0;
1567 struct super_block *sb;
1568 struct cgroupfs_root *new_root;
1569 struct inode *inode;
1570
1571 /* First find the desired set of subsystems */
1572 mutex_lock(&cgroup_mutex);
1573 ret = parse_cgroupfs_options(data, &opts);
1574 mutex_unlock(&cgroup_mutex);
1575 if (ret)
1576 goto out_err;
1577
1578 /*
1579 * Allocate a new cgroup root. We may not need it if we're
1580 * reusing an existing hierarchy.
1581 */
1582 new_root = cgroup_root_from_opts(&opts);
1583 if (IS_ERR(new_root)) {
1584 ret = PTR_ERR(new_root);
1585 goto drop_modules;
1586 }
1587 opts.new_root = new_root;
1588
1589 /* Locate an existing or new sb for this hierarchy */
1590 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1591 if (IS_ERR(sb)) {
1592 ret = PTR_ERR(sb);
1593 cgroup_drop_root(opts.new_root);
1594 goto drop_modules;
1595 }
1596
1597 root = sb->s_fs_info;
1598 BUG_ON(!root);
1599 if (root == opts.new_root) {
1600 /* We used the new root structure, so this is a new hierarchy */
1601 struct list_head tmp_cg_links;
1602 struct cgroup *root_cgrp = &root->top_cgroup;
1603 struct cgroupfs_root *existing_root;
1604 const struct cred *cred;
1605 int i;
1606
1607 BUG_ON(sb->s_root != NULL);
1608
1609 ret = cgroup_get_rootdir(sb);
1610 if (ret)
1611 goto drop_new_super;
1612 inode = sb->s_root->d_inode;
1613
1614 mutex_lock(&inode->i_mutex);
1615 mutex_lock(&cgroup_mutex);
1616 mutex_lock(&cgroup_root_mutex);
1617
1618 /* Check for name clashes with existing mounts */
1619 ret = -EBUSY;
1620 if (strlen(root->name))
1621 for_each_active_root(existing_root)
1622 if (!strcmp(existing_root->name, root->name))
1623 goto unlock_drop;
1624
1625 /*
1626 * We're accessing css_set_count without locking
1627 * css_set_lock here, but that's OK - it can only be
1628 * increased by someone holding cgroup_lock, and
1629 * that's us. The worst that can happen is that we
1630 * have some link structures left over
1631 */
1632 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1633 if (ret)
1634 goto unlock_drop;
1635
1636 ret = rebind_subsystems(root, root->subsys_bits);
1637 if (ret == -EBUSY) {
1638 free_cg_links(&tmp_cg_links);
1639 goto unlock_drop;
1640 }
1641 /*
1642 * There must be no failure case after here, since rebinding
1643 * takes care of subsystems' refcounts, which are explicitly
1644 * dropped in the failure exit path.
1645 */
1646
1647 /* EBUSY should be the only error here */
1648 BUG_ON(ret);
1649
1650 list_add(&root->root_list, &roots);
1651 root_count++;
1652
1653 sb->s_root->d_fsdata = root_cgrp;
1654 root->top_cgroup.dentry = sb->s_root;
1655
1656 /* Link the top cgroup in this hierarchy into all
1657 * the css_set objects */
1658 write_lock(&css_set_lock);
1659 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1660 struct hlist_head *hhead = &css_set_table[i];
1661 struct hlist_node *node;
1662 struct css_set *cg;
1663
1664 hlist_for_each_entry(cg, node, hhead, hlist)
1665 link_css_set(&tmp_cg_links, cg, root_cgrp);
1666 }
1667 write_unlock(&css_set_lock);
1668
1669 free_cg_links(&tmp_cg_links);
1670
1671 BUG_ON(!list_empty(&root_cgrp->sibling));
1672 BUG_ON(!list_empty(&root_cgrp->children));
1673 BUG_ON(root->number_of_cgroups != 1);
1674
1675 cred = override_creds(&init_cred);
1676 cgroup_populate_dir(root_cgrp);
1677 revert_creds(cred);
1678 mutex_unlock(&cgroup_root_mutex);
1679 mutex_unlock(&cgroup_mutex);
1680 mutex_unlock(&inode->i_mutex);
1681 } else {
1682 /*
1683 * We re-used an existing hierarchy - the new root (if
1684 * any) is not needed
1685 */
1686 cgroup_drop_root(opts.new_root);
1687 /* no subsys rebinding, so refcounts don't change */
1688 drop_parsed_module_refcounts(opts.subsys_bits);
1689 }
1690
1691 kfree(opts.release_agent);
1692 kfree(opts.name);
1693 return dget(sb->s_root);
1694
1695 unlock_drop:
1696 mutex_unlock(&cgroup_root_mutex);
1697 mutex_unlock(&cgroup_mutex);
1698 mutex_unlock(&inode->i_mutex);
1699 drop_new_super:
1700 deactivate_locked_super(sb);
1701 drop_modules:
1702 drop_parsed_module_refcounts(opts.subsys_bits);
1703 out_err:
1704 kfree(opts.release_agent);
1705 kfree(opts.name);
1706 return ERR_PTR(ret);
1707}
1708
1709static void cgroup_kill_sb(struct super_block *sb) {
1710 struct cgroupfs_root *root = sb->s_fs_info;
1711 struct cgroup *cgrp = &root->top_cgroup;
1712 int ret;
1713 struct cg_cgroup_link *link;
1714 struct cg_cgroup_link *saved_link;
1715
1716 BUG_ON(!root);
1717
1718 BUG_ON(root->number_of_cgroups != 1);
1719 BUG_ON(!list_empty(&cgrp->children));
1720 BUG_ON(!list_empty(&cgrp->sibling));
1721
1722 mutex_lock(&cgroup_mutex);
1723 mutex_lock(&cgroup_root_mutex);
1724
1725 /* Rebind all subsystems back to the default hierarchy */
1726 ret = rebind_subsystems(root, 0);
1727 /* Shouldn't be able to fail ... */
1728 BUG_ON(ret);
1729
1730 /*
1731 * Release all the links from css_sets to this hierarchy's
1732 * root cgroup
1733 */
1734 write_lock(&css_set_lock);
1735
1736 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1737 cgrp_link_list) {
1738 list_del(&link->cg_link_list);
1739 list_del(&link->cgrp_link_list);
1740 kfree(link);
1741 }
1742 write_unlock(&css_set_lock);
1743
1744 if (!list_empty(&root->root_list)) {
1745 list_del(&root->root_list);
1746 root_count--;
1747 }
1748
1749 mutex_unlock(&cgroup_root_mutex);
1750 mutex_unlock(&cgroup_mutex);
1751
1752 kill_litter_super(sb);
1753 cgroup_drop_root(root);
1754}
1755
1756static struct file_system_type cgroup_fs_type = {
1757 .name = "cgroup",
1758 .mount = cgroup_mount,
1759 .kill_sb = cgroup_kill_sb,
1760};
1761
1762static struct kobject *cgroup_kobj;
1763
1764/**
1765 * cgroup_path - generate the path of a cgroup
1766 * @cgrp: the cgroup in question
1767 * @buf: the buffer to write the path into
1768 * @buflen: the length of the buffer
1769 *
1770 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1771 * reference. Writes path of cgroup into buf. Returns 0 on success,
1772 * -errno on error.
1773 */
1774int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1775{
1776 char *start;
1777 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1778 cgroup_lock_is_held());
1779
1780 if (!dentry || cgrp == dummytop) {
1781 /*
1782 * Inactive subsystems have no dentry for their root
1783 * cgroup
1784 */
1785 strcpy(buf, "/");
1786 return 0;
1787 }
1788
1789 start = buf + buflen;
1790
1791 *--start = '\0';
1792 for (;;) {
1793 int len = dentry->d_name.len;
1794
1795 if ((start -= len) < buf)
1796 return -ENAMETOOLONG;
1797 memcpy(start, dentry->d_name.name, len);
1798 cgrp = cgrp->parent;
1799 if (!cgrp)
1800 break;
1801
1802 dentry = rcu_dereference_check(cgrp->dentry,
1803 cgroup_lock_is_held());
1804 if (!cgrp->parent)
1805 continue;
1806 if (--start < buf)
1807 return -ENAMETOOLONG;
1808 *start = '/';
1809 }
1810 memmove(buf, start, buf + buflen - start);
1811 return 0;
1812}
1813EXPORT_SYMBOL_GPL(cgroup_path);
1814
1815/*
1816 * Control Group taskset
1817 */
1818struct task_and_cgroup {
1819 struct task_struct *task;
1820 struct cgroup *cgrp;
1821 struct css_set *cg;
1822};
1823
1824struct cgroup_taskset {
1825 struct task_and_cgroup single;
1826 struct flex_array *tc_array;
1827 int tc_array_len;
1828 int idx;
1829 struct cgroup *cur_cgrp;
1830};
1831
1832/**
1833 * cgroup_taskset_first - reset taskset and return the first task
1834 * @tset: taskset of interest
1835 *
1836 * @tset iteration is initialized and the first task is returned.
1837 */
1838struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1839{
1840 if (tset->tc_array) {
1841 tset->idx = 0;
1842 return cgroup_taskset_next(tset);
1843 } else {
1844 tset->cur_cgrp = tset->single.cgrp;
1845 return tset->single.task;
1846 }
1847}
1848EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1849
1850/**
1851 * cgroup_taskset_next - iterate to the next task in taskset
1852 * @tset: taskset of interest
1853 *
1854 * Return the next task in @tset. Iteration must have been initialized
1855 * with cgroup_taskset_first().
1856 */
1857struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1858{
1859 struct task_and_cgroup *tc;
1860
1861 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1862 return NULL;
1863
1864 tc = flex_array_get(tset->tc_array, tset->idx++);
1865 tset->cur_cgrp = tc->cgrp;
1866 return tc->task;
1867}
1868EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1869
1870/**
1871 * cgroup_taskset_cur_cgroup - return the matching cgroup for the current task
1872 * @tset: taskset of interest
1873 *
1874 * Return the cgroup for the current (last returned) task of @tset. This
1875 * function must be preceded by either cgroup_taskset_first() or
1876 * cgroup_taskset_next().
1877 */
1878struct cgroup *cgroup_taskset_cur_cgroup(struct cgroup_taskset *tset)
1879{
1880 return tset->cur_cgrp;
1881}
1882EXPORT_SYMBOL_GPL(cgroup_taskset_cur_cgroup);
1883
1884/**
1885 * cgroup_taskset_size - return the number of tasks in taskset
1886 * @tset: taskset of interest
1887 */
1888int cgroup_taskset_size(struct cgroup_taskset *tset)
1889{
1890 return tset->tc_array ? tset->tc_array_len : 1;
1891}
1892EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1893
1894
1895/*
1896 * cgroup_task_migrate - move a task from one cgroup to another.
1897 *
1898 * 'guarantee' is set if the caller promises that a new css_set for the task
1899 * will already exist. If not set, this function might sleep, and can fail with
1900 * -ENOMEM. Must be called with cgroup_mutex and threadgroup locked.
1901 */
1902static void cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
1903 struct task_struct *tsk, struct css_set *newcg)
1904{
1905 struct css_set *oldcg;
1906
1907 /*
1908 * We are synchronized through threadgroup_lock() against PF_EXITING
1909 * setting such that we can't race against cgroup_exit() changing the
1910 * css_set to init_css_set and dropping the old one.
1911 */
1912 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1913 oldcg = tsk->cgroups;
1914
1915 task_lock(tsk);
1916 rcu_assign_pointer(tsk->cgroups, newcg);
1917 task_unlock(tsk);
1918
1919 /* Update the css_set linked lists if we're using them */
1920 write_lock(&css_set_lock);
1921 if (!list_empty(&tsk->cg_list))
1922 list_move(&tsk->cg_list, &newcg->tasks);
1923 write_unlock(&css_set_lock);
1924
1925 /*
1926 * We just gained a reference on oldcg by taking it from the task. As
1927 * trading it for newcg is protected by cgroup_mutex, we're safe to drop
1928 * it here; it will be freed under RCU.
1929 */
1930 put_css_set(oldcg);
1931
1932 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1933}
1934
1935/**
1936 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1937 * @cgrp: the cgroup the task is attaching to
1938 * @tsk: the task to be attached
1939 *
1940 * Call with cgroup_mutex and threadgroup locked. May take task_lock of
1941 * @tsk during call.
1942 */
1943int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1944{
1945 int retval = 0;
1946 struct cgroup_subsys *ss, *failed_ss = NULL;
1947 struct cgroup *oldcgrp;
1948 struct cgroupfs_root *root = cgrp->root;
1949 struct cgroup_taskset tset = { };
1950 struct css_set *newcg;
1951
1952 /* @tsk either already exited or can't exit until the end */
1953 if (tsk->flags & PF_EXITING)
1954 return -ESRCH;
1955
1956 /* Nothing to do if the task is already in that cgroup */
1957 oldcgrp = task_cgroup_from_root(tsk, root);
1958 if (cgrp == oldcgrp)
1959 return 0;
1960
1961 tset.single.task = tsk;
1962 tset.single.cgrp = oldcgrp;
1963
1964 for_each_subsys(root, ss) {
1965 if (ss->can_attach) {
1966 retval = ss->can_attach(cgrp, &tset);
1967 if (retval) {
1968 /*
1969 * Remember on which subsystem the can_attach()
1970 * failed, so that we only call cancel_attach()
1971 * against the subsystems whose can_attach()
1972 * succeeded. (See below)
1973 */
1974 failed_ss = ss;
1975 goto out;
1976 }
1977 }
1978 }
1979
1980 newcg = find_css_set(tsk->cgroups, cgrp);
1981 if (!newcg) {
1982 retval = -ENOMEM;
1983 goto out;
1984 }
1985
1986 cgroup_task_migrate(cgrp, oldcgrp, tsk, newcg);
1987
1988 for_each_subsys(root, ss) {
1989 if (ss->attach)
1990 ss->attach(cgrp, &tset);
1991 }
1992
1993 synchronize_rcu();
1994
1995 /*
1996 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1997 * is no longer empty.
1998 */
1999 cgroup_wakeup_rmdir_waiter(cgrp);
2000out:
2001 if (retval) {
2002 for_each_subsys(root, ss) {
2003 if (ss == failed_ss)
2004 /*
2005 * This subsystem was the one that failed the
2006 * can_attach() check earlier, so we don't need
2007 * to call cancel_attach() against it or any
2008 * remaining subsystems.
2009 */
2010 break;
2011 if (ss->cancel_attach)
2012 ss->cancel_attach(cgrp, &tset);
2013 }
2014 }
2015 return retval;
2016}
2017
2018/**
2019 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2020 * @from: attach to all cgroups of a given task
2021 * @tsk: the task to be attached
2022 */
2023int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2024{
2025 struct cgroupfs_root *root;
2026 int retval = 0;
2027
2028 cgroup_lock();
2029 for_each_active_root(root) {
2030 struct cgroup *from_cg = task_cgroup_from_root(from, root);
2031
2032 retval = cgroup_attach_task(from_cg, tsk);
2033 if (retval)
2034 break;
2035 }
2036 cgroup_unlock();
2037
2038 return retval;
2039}
2040EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2041
2042/**
2043 * cgroup_attach_proc - attach all threads in a threadgroup to a cgroup
2044 * @cgrp: the cgroup to attach to
2045 * @leader: the threadgroup leader task_struct of the group to be attached
2046 *
2047 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
2048 * task_lock of each thread in leader's threadgroup individually in turn.
2049 */
2050static int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
2051{
2052 int retval, i, group_size;
2053 struct cgroup_subsys *ss, *failed_ss = NULL;
2054 /* guaranteed to be initialized later, but the compiler needs this */
2055 struct cgroupfs_root *root = cgrp->root;
2056 /* threadgroup list cursor and array */
2057 struct task_struct *tsk;
2058 struct task_and_cgroup *tc;
2059 struct flex_array *group;
2060 struct cgroup_taskset tset = { };
2061
2062 /*
2063 * step 0: in order to do expensive, possibly blocking operations for
2064 * every thread, we cannot iterate the thread group list, since it needs
2065 * rcu or tasklist locked. instead, build an array of all threads in the
2066 * group - group_rwsem prevents new threads from appearing, and if
2067 * threads exit, this will just be an over-estimate.
2068 */
2069 group_size = get_nr_threads(leader);
2070 /* flex_array supports very large thread-groups better than kmalloc. */
2071 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
2072 if (!group)
2073 return -ENOMEM;
2074 /* pre-allocate to guarantee space while iterating in rcu read-side. */
2075 retval = flex_array_prealloc(group, 0, group_size - 1, GFP_KERNEL);
2076 if (retval)
2077 goto out_free_group_list;
2078
2079 tsk = leader;
2080 i = 0;
2081 /*
2082 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2083 * already PF_EXITING could be freed from underneath us unless we
2084 * take an rcu_read_lock.
2085 */
2086 rcu_read_lock();
2087 do {
2088 struct task_and_cgroup ent;
2089
2090 /* @tsk either already exited or can't exit until the end */
2091 if (tsk->flags & PF_EXITING)
2092 continue;
2093
2094 /* as per above, nr_threads may decrease, but not increase. */
2095 BUG_ON(i >= group_size);
2096 ent.task = tsk;
2097 ent.cgrp = task_cgroup_from_root(tsk, root);
2098 /* nothing to do if this task is already in the cgroup */
2099 if (ent.cgrp == cgrp)
2100 continue;
2101 /*
2102 * saying GFP_ATOMIC has no effect here because we did prealloc
2103 * earlier, but it's good form to communicate our expectations.
2104 */
2105 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2106 BUG_ON(retval != 0);
2107 i++;
2108 } while_each_thread(leader, tsk);
2109 rcu_read_unlock();
2110 /* remember the number of threads in the array for later. */
2111 group_size = i;
2112 tset.tc_array = group;
2113 tset.tc_array_len = group_size;
2114
2115 /* methods shouldn't be called if no task is actually migrating */
2116 retval = 0;
2117 if (!group_size)
2118 goto out_free_group_list;
2119
2120 /*
2121 * step 1: check that we can legitimately attach to the cgroup.
2122 */
2123 for_each_subsys(root, ss) {
2124 if (ss->can_attach) {
2125 retval = ss->can_attach(cgrp, &tset);
2126 if (retval) {
2127 failed_ss = ss;
2128 goto out_cancel_attach;
2129 }
2130 }
2131 }
2132
2133 /*
2134 * step 2: make sure css_sets exist for all threads to be migrated.
2135 * we use find_css_set, which allocates a new one if necessary.
2136 */
2137 for (i = 0; i < group_size; i++) {
2138 tc = flex_array_get(group, i);
2139 tc->cg = find_css_set(tc->task->cgroups, cgrp);
2140 if (!tc->cg) {
2141 retval = -ENOMEM;
2142 goto out_put_css_set_refs;
2143 }
2144 }
2145
2146 /*
2147 * step 3: now that we're guaranteed success wrt the css_sets,
2148 * proceed to move all tasks to the new cgroup. There are no
2149 * failure cases after here, so this is the commit point.
2150 */
2151 for (i = 0; i < group_size; i++) {
2152 tc = flex_array_get(group, i);
2153 cgroup_task_migrate(cgrp, tc->cgrp, tc->task, tc->cg);
2154 }
2155 /* nothing is sensitive to fork() after this point. */
2156
2157 /*
2158 * step 4: do subsystem attach callbacks.
2159 */
2160 for_each_subsys(root, ss) {
2161 if (ss->attach)
2162 ss->attach(cgrp, &tset);
2163 }
2164
2165 /*
2166 * step 5: success! and cleanup
2167 */
2168 synchronize_rcu();
2169 cgroup_wakeup_rmdir_waiter(cgrp);
2170 retval = 0;
2171out_put_css_set_refs:
2172 if (retval) {
2173 for (i = 0; i < group_size; i++) {
2174 tc = flex_array_get(group, i);
2175 if (!tc->cg)
2176 break;
2177 put_css_set(tc->cg);
2178 }
2179 }
2180out_cancel_attach:
2181 if (retval) {
2182 for_each_subsys(root, ss) {
2183 if (ss == failed_ss)
2184 break;
2185 if (ss->cancel_attach)
2186 ss->cancel_attach(cgrp, &tset);
2187 }
2188 }
2189out_free_group_list:
2190 flex_array_free(group);
2191 return retval;
2192}
2193
2194/*
2195 * Find the task_struct of the task to attach by vpid and pass it along to the
2196 * function to attach either it or all tasks in its threadgroup. Will lock
2197 * cgroup_mutex and threadgroup; may take task_lock of task.
2198 */
2199static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2200{
2201 struct task_struct *tsk;
2202 const struct cred *cred = current_cred(), *tcred;
2203 int ret;
2204
2205 if (!cgroup_lock_live_group(cgrp))
2206 return -ENODEV;
2207
2208retry_find_task:
2209 rcu_read_lock();
2210 if (pid) {
2211 tsk = find_task_by_vpid(pid);
2212 if (!tsk) {
2213 rcu_read_unlock();
2214 ret= -ESRCH;
2215 goto out_unlock_cgroup;
2216 }
2217 /*
2218 * even if we're attaching all tasks in the thread group, we
2219 * only need to check permissions on one of them.
2220 */
2221 tcred = __task_cred(tsk);
2222 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2223 !uid_eq(cred->euid, tcred->uid) &&
2224 !uid_eq(cred->euid, tcred->suid)) {
2225 rcu_read_unlock();
2226 ret = -EACCES;
2227 goto out_unlock_cgroup;
2228 }
2229 } else
2230 tsk = current;
2231
2232 if (threadgroup)
2233 tsk = tsk->group_leader;
2234
2235 /*
2236 * Workqueue threads may acquire PF_THREAD_BOUND and become
2237 * trapped in a cpuset, or RT worker may be born in a cgroup
2238 * with no rt_runtime allocated. Just say no.
2239 */
2240 if (tsk == kthreadd_task || (tsk->flags & PF_THREAD_BOUND)) {
2241 ret = -EINVAL;
2242 rcu_read_unlock();
2243 goto out_unlock_cgroup;
2244 }
2245
2246 get_task_struct(tsk);
2247 rcu_read_unlock();
2248
2249 threadgroup_lock(tsk);
2250 if (threadgroup) {
2251 if (!thread_group_leader(tsk)) {
2252 /*
2253 * a race with de_thread from another thread's exec()
2254 * may strip us of our leadership, if this happens,
2255 * there is no choice but to throw this task away and
2256 * try again; this is
2257 * "double-double-toil-and-trouble-check locking".
2258 */
2259 threadgroup_unlock(tsk);
2260 put_task_struct(tsk);
2261 goto retry_find_task;
2262 }
2263 ret = cgroup_attach_proc(cgrp, tsk);
2264 } else
2265 ret = cgroup_attach_task(cgrp, tsk);
2266 threadgroup_unlock(tsk);
2267
2268 put_task_struct(tsk);
2269out_unlock_cgroup:
2270 cgroup_unlock();
2271 return ret;
2272}
2273
2274static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
2275{
2276 return attach_task_by_pid(cgrp, pid, false);
2277}
2278
2279static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
2280{
2281 return attach_task_by_pid(cgrp, tgid, true);
2282}
2283
2284/**
2285 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
2286 * @cgrp: the cgroup to be checked for liveness
2287 *
2288 * On success, returns true; the lock should be later released with
2289 * cgroup_unlock(). On failure returns false with no lock held.
2290 */
2291bool cgroup_lock_live_group(struct cgroup *cgrp)
2292{
2293 mutex_lock(&cgroup_mutex);
2294 if (cgroup_is_removed(cgrp)) {
2295 mutex_unlock(&cgroup_mutex);
2296 return false;
2297 }
2298 return true;
2299}
2300EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
2301
2302static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
2303 const char *buffer)
2304{
2305 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2306 if (strlen(buffer) >= PATH_MAX)
2307 return -EINVAL;
2308 if (!cgroup_lock_live_group(cgrp))
2309 return -ENODEV;
2310 mutex_lock(&cgroup_root_mutex);
2311 strcpy(cgrp->root->release_agent_path, buffer);
2312 mutex_unlock(&cgroup_root_mutex);
2313 cgroup_unlock();
2314 return 0;
2315}
2316
2317static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
2318 struct seq_file *seq)
2319{
2320 if (!cgroup_lock_live_group(cgrp))
2321 return -ENODEV;
2322 seq_puts(seq, cgrp->root->release_agent_path);
2323 seq_putc(seq, '\n');
2324 cgroup_unlock();
2325 return 0;
2326}
2327
2328/* A buffer size big enough for numbers or short strings */
2329#define CGROUP_LOCAL_BUFFER_SIZE 64
2330
2331static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
2332 struct file *file,
2333 const char __user *userbuf,
2334 size_t nbytes, loff_t *unused_ppos)
2335{
2336 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
2337 int retval = 0;
2338 char *end;
2339
2340 if (!nbytes)
2341 return -EINVAL;
2342 if (nbytes >= sizeof(buffer))
2343 return -E2BIG;
2344 if (copy_from_user(buffer, userbuf, nbytes))
2345 return -EFAULT;
2346
2347 buffer[nbytes] = 0; /* nul-terminate */
2348 if (cft->write_u64) {
2349 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
2350 if (*end)
2351 return -EINVAL;
2352 retval = cft->write_u64(cgrp, cft, val);
2353 } else {
2354 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
2355 if (*end)
2356 return -EINVAL;
2357 retval = cft->write_s64(cgrp, cft, val);
2358 }
2359 if (!retval)
2360 retval = nbytes;
2361 return retval;
2362}
2363
2364static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
2365 struct file *file,
2366 const char __user *userbuf,
2367 size_t nbytes, loff_t *unused_ppos)
2368{
2369 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2370 int retval = 0;
2371 size_t max_bytes = cft->max_write_len;
2372 char *buffer = local_buffer;
2373
2374 if (!max_bytes)
2375 max_bytes = sizeof(local_buffer) - 1;
2376 if (nbytes >= max_bytes)
2377 return -E2BIG;
2378 /* Allocate a dynamic buffer if we need one */
2379 if (nbytes >= sizeof(local_buffer)) {
2380 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2381 if (buffer == NULL)
2382 return -ENOMEM;
2383 }
2384 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2385 retval = -EFAULT;
2386 goto out;
2387 }
2388
2389 buffer[nbytes] = 0; /* nul-terminate */
2390 retval = cft->write_string(cgrp, cft, strstrip(buffer));
2391 if (!retval)
2392 retval = nbytes;
2393out:
2394 if (buffer != local_buffer)
2395 kfree(buffer);
2396 return retval;
2397}
2398
2399static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2400 size_t nbytes, loff_t *ppos)
2401{
2402 struct cftype *cft = __d_cft(file->f_dentry);
2403 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2404
2405 if (cgroup_is_removed(cgrp))
2406 return -ENODEV;
2407 if (cft->write)
2408 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2409 if (cft->write_u64 || cft->write_s64)
2410 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2411 if (cft->write_string)
2412 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2413 if (cft->trigger) {
2414 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2415 return ret ? ret : nbytes;
2416 }
2417 return -EINVAL;
2418}
2419
2420static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2421 struct file *file,
2422 char __user *buf, size_t nbytes,
2423 loff_t *ppos)
2424{
2425 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2426 u64 val = cft->read_u64(cgrp, cft);
2427 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2428
2429 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2430}
2431
2432static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2433 struct file *file,
2434 char __user *buf, size_t nbytes,
2435 loff_t *ppos)
2436{
2437 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2438 s64 val = cft->read_s64(cgrp, cft);
2439 int len = sprintf(tmp, "%lld\n", (long long) val);
2440
2441 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2442}
2443
2444static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2445 size_t nbytes, loff_t *ppos)
2446{
2447 struct cftype *cft = __d_cft(file->f_dentry);
2448 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2449
2450 if (cgroup_is_removed(cgrp))
2451 return -ENODEV;
2452
2453 if (cft->read)
2454 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2455 if (cft->read_u64)
2456 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2457 if (cft->read_s64)
2458 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2459 return -EINVAL;
2460}
2461
2462/*
2463 * seqfile ops/methods for returning structured data. Currently just
2464 * supports string->u64 maps, but can be extended in future.
2465 */
2466
2467struct cgroup_seqfile_state {
2468 struct cftype *cft;
2469 struct cgroup *cgroup;
2470};
2471
2472static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2473{
2474 struct seq_file *sf = cb->state;
2475 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2476}
2477
2478static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2479{
2480 struct cgroup_seqfile_state *state = m->private;
2481 struct cftype *cft = state->cft;
2482 if (cft->read_map) {
2483 struct cgroup_map_cb cb = {
2484 .fill = cgroup_map_add,
2485 .state = m,
2486 };
2487 return cft->read_map(state->cgroup, cft, &cb);
2488 }
2489 return cft->read_seq_string(state->cgroup, cft, m);
2490}
2491
2492static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2493{
2494 struct seq_file *seq = file->private_data;
2495 kfree(seq->private);
2496 return single_release(inode, file);
2497}
2498
2499static const struct file_operations cgroup_seqfile_operations = {
2500 .read = seq_read,
2501 .write = cgroup_file_write,
2502 .llseek = seq_lseek,
2503 .release = cgroup_seqfile_release,
2504};
2505
2506static int cgroup_file_open(struct inode *inode, struct file *file)
2507{
2508 int err;
2509 struct cftype *cft;
2510
2511 err = generic_file_open(inode, file);
2512 if (err)
2513 return err;
2514 cft = __d_cft(file->f_dentry);
2515
2516 if (cft->read_map || cft->read_seq_string) {
2517 struct cgroup_seqfile_state *state =
2518 kzalloc(sizeof(*state), GFP_USER);
2519 if (!state)
2520 return -ENOMEM;
2521 state->cft = cft;
2522 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2523 file->f_op = &cgroup_seqfile_operations;
2524 err = single_open(file, cgroup_seqfile_show, state);
2525 if (err < 0)
2526 kfree(state);
2527 } else if (cft->open)
2528 err = cft->open(inode, file);
2529 else
2530 err = 0;
2531
2532 return err;
2533}
2534
2535static int cgroup_file_release(struct inode *inode, struct file *file)
2536{
2537 struct cftype *cft = __d_cft(file->f_dentry);
2538 if (cft->release)
2539 return cft->release(inode, file);
2540 return 0;
2541}
2542
2543/*
2544 * cgroup_rename - Only allow simple rename of directories in place.
2545 */
2546static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2547 struct inode *new_dir, struct dentry *new_dentry)
2548{
2549 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2550 return -ENOTDIR;
2551 if (new_dentry->d_inode)
2552 return -EEXIST;
2553 if (old_dir != new_dir)
2554 return -EIO;
2555 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2556}
2557
2558static const struct file_operations cgroup_file_operations = {
2559 .read = cgroup_file_read,
2560 .write = cgroup_file_write,
2561 .llseek = generic_file_llseek,
2562 .open = cgroup_file_open,
2563 .release = cgroup_file_release,
2564};
2565
2566static const struct inode_operations cgroup_dir_inode_operations = {
2567 .lookup = cgroup_lookup,
2568 .mkdir = cgroup_mkdir,
2569 .rmdir = cgroup_rmdir,
2570 .rename = cgroup_rename,
2571};
2572
2573static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, struct nameidata *nd)
2574{
2575 if (dentry->d_name.len > NAME_MAX)
2576 return ERR_PTR(-ENAMETOOLONG);
2577 d_add(dentry, NULL);
2578 return NULL;
2579}
2580
2581/*
2582 * Check if a file is a control file
2583 */
2584static inline struct cftype *__file_cft(struct file *file)
2585{
2586 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2587 return ERR_PTR(-EINVAL);
2588 return __d_cft(file->f_dentry);
2589}
2590
2591static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2592 struct super_block *sb)
2593{
2594 struct inode *inode;
2595
2596 if (!dentry)
2597 return -ENOENT;
2598 if (dentry->d_inode)
2599 return -EEXIST;
2600
2601 inode = cgroup_new_inode(mode, sb);
2602 if (!inode)
2603 return -ENOMEM;
2604
2605 if (S_ISDIR(mode)) {
2606 inode->i_op = &cgroup_dir_inode_operations;
2607 inode->i_fop = &simple_dir_operations;
2608
2609 /* start off with i_nlink == 2 (for "." entry) */
2610 inc_nlink(inode);
2611
2612 /* start with the directory inode held, so that we can
2613 * populate it without racing with another mkdir */
2614 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2615 } else if (S_ISREG(mode)) {
2616 inode->i_size = 0;
2617 inode->i_fop = &cgroup_file_operations;
2618 }
2619 d_instantiate(dentry, inode);
2620 dget(dentry); /* Extra count - pin the dentry in core */
2621 return 0;
2622}
2623
2624/*
2625 * cgroup_create_dir - create a directory for an object.
2626 * @cgrp: the cgroup we create the directory for. It must have a valid
2627 * ->parent field. And we are going to fill its ->dentry field.
2628 * @dentry: dentry of the new cgroup
2629 * @mode: mode to set on new directory.
2630 */
2631static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2632 umode_t mode)
2633{
2634 struct dentry *parent;
2635 int error = 0;
2636
2637 parent = cgrp->parent->dentry;
2638 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2639 if (!error) {
2640 dentry->d_fsdata = cgrp;
2641 inc_nlink(parent->d_inode);
2642 rcu_assign_pointer(cgrp->dentry, dentry);
2643 dget(dentry);
2644 }
2645 dput(dentry);
2646
2647 return error;
2648}
2649
2650/**
2651 * cgroup_file_mode - deduce file mode of a control file
2652 * @cft: the control file in question
2653 *
2654 * returns cft->mode if ->mode is not 0
2655 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2656 * returns S_IRUGO if it has only a read handler
2657 * returns S_IWUSR if it has only a write hander
2658 */
2659static umode_t cgroup_file_mode(const struct cftype *cft)
2660{
2661 umode_t mode = 0;
2662
2663 if (cft->mode)
2664 return cft->mode;
2665
2666 if (cft->read || cft->read_u64 || cft->read_s64 ||
2667 cft->read_map || cft->read_seq_string)
2668 mode |= S_IRUGO;
2669
2670 if (cft->write || cft->write_u64 || cft->write_s64 ||
2671 cft->write_string || cft->trigger)
2672 mode |= S_IWUSR;
2673
2674 return mode;
2675}
2676
2677static int cgroup_add_file(struct cgroup *cgrp, struct cgroup_subsys *subsys,
2678 const struct cftype *cft)
2679{
2680 struct dentry *dir = cgrp->dentry;
2681 struct cgroup *parent = __d_cgrp(dir);
2682 struct dentry *dentry;
2683 struct cfent *cfe;
2684 int error;
2685 umode_t mode;
2686 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2687
2688 /* does @cft->flags tell us to skip creation on @cgrp? */
2689 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2690 return 0;
2691 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2692 return 0;
2693
2694 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2695 strcpy(name, subsys->name);
2696 strcat(name, ".");
2697 }
2698 strcat(name, cft->name);
2699
2700 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2701
2702 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2703 if (!cfe)
2704 return -ENOMEM;
2705
2706 dentry = lookup_one_len(name, dir, strlen(name));
2707 if (IS_ERR(dentry)) {
2708 error = PTR_ERR(dentry);
2709 goto out;
2710 }
2711
2712 mode = cgroup_file_mode(cft);
2713 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2714 if (!error) {
2715 cfe->type = (void *)cft;
2716 cfe->dentry = dentry;
2717 dentry->d_fsdata = cfe;
2718 list_add_tail(&cfe->node, &parent->files);
2719 cfe = NULL;
2720 }
2721 dput(dentry);
2722out:
2723 kfree(cfe);
2724 return error;
2725}
2726
2727static int cgroup_addrm_files(struct cgroup *cgrp, struct cgroup_subsys *subsys,
2728 const struct cftype cfts[], bool is_add)
2729{
2730 const struct cftype *cft;
2731 int err, ret = 0;
2732
2733 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2734 if (is_add)
2735 err = cgroup_add_file(cgrp, subsys, cft);
2736 else
2737 err = cgroup_rm_file(cgrp, cft);
2738 if (err) {
2739 pr_warning("cgroup_addrm_files: failed to %s %s, err=%d\n",
2740 is_add ? "add" : "remove", cft->name, err);
2741 ret = err;
2742 }
2743 }
2744 return ret;
2745}
2746
2747static DEFINE_MUTEX(cgroup_cft_mutex);
2748
2749static void cgroup_cfts_prepare(void)
2750 __acquires(&cgroup_cft_mutex) __acquires(&cgroup_mutex)
2751{
2752 /*
2753 * Thanks to the entanglement with vfs inode locking, we can't walk
2754 * the existing cgroups under cgroup_mutex and create files.
2755 * Instead, we increment reference on all cgroups and build list of
2756 * them using @cgrp->cft_q_node. Grab cgroup_cft_mutex to ensure
2757 * exclusive access to the field.
2758 */
2759 mutex_lock(&cgroup_cft_mutex);
2760 mutex_lock(&cgroup_mutex);
2761}
2762
2763static void cgroup_cfts_commit(struct cgroup_subsys *ss,
2764 const struct cftype *cfts, bool is_add)
2765 __releases(&cgroup_mutex) __releases(&cgroup_cft_mutex)
2766{
2767 LIST_HEAD(pending);
2768 struct cgroup *cgrp, *n;
2769
2770 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2771 if (cfts && ss->root != &rootnode) {
2772 list_for_each_entry(cgrp, &ss->root->allcg_list, allcg_node) {
2773 dget(cgrp->dentry);
2774 list_add_tail(&cgrp->cft_q_node, &pending);
2775 }
2776 }
2777
2778 mutex_unlock(&cgroup_mutex);
2779
2780 /*
2781 * All new cgroups will see @cfts update on @ss->cftsets. Add/rm
2782 * files for all cgroups which were created before.
2783 */
2784 list_for_each_entry_safe(cgrp, n, &pending, cft_q_node) {
2785 struct inode *inode = cgrp->dentry->d_inode;
2786
2787 mutex_lock(&inode->i_mutex);
2788 mutex_lock(&cgroup_mutex);
2789 if (!cgroup_is_removed(cgrp))
2790 cgroup_addrm_files(cgrp, ss, cfts, is_add);
2791 mutex_unlock(&cgroup_mutex);
2792 mutex_unlock(&inode->i_mutex);
2793
2794 list_del_init(&cgrp->cft_q_node);
2795 dput(cgrp->dentry);
2796 }
2797
2798 mutex_unlock(&cgroup_cft_mutex);
2799}
2800
2801/**
2802 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2803 * @ss: target cgroup subsystem
2804 * @cfts: zero-length name terminated array of cftypes
2805 *
2806 * Register @cfts to @ss. Files described by @cfts are created for all
2807 * existing cgroups to which @ss is attached and all future cgroups will
2808 * have them too. This function can be called anytime whether @ss is
2809 * attached or not.
2810 *
2811 * Returns 0 on successful registration, -errno on failure. Note that this
2812 * function currently returns 0 as long as @cfts registration is successful
2813 * even if some file creation attempts on existing cgroups fail.
2814 */
2815int cgroup_add_cftypes(struct cgroup_subsys *ss, const struct cftype *cfts)
2816{
2817 struct cftype_set *set;
2818
2819 set = kzalloc(sizeof(*set), GFP_KERNEL);
2820 if (!set)
2821 return -ENOMEM;
2822
2823 cgroup_cfts_prepare();
2824 set->cfts = cfts;
2825 list_add_tail(&set->node, &ss->cftsets);
2826 cgroup_cfts_commit(ss, cfts, true);
2827
2828 return 0;
2829}
2830EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2831
2832/**
2833 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2834 * @ss: target cgroup subsystem
2835 * @cfts: zero-length name terminated array of cftypes
2836 *
2837 * Unregister @cfts from @ss. Files described by @cfts are removed from
2838 * all existing cgroups to which @ss is attached and all future cgroups
2839 * won't have them either. This function can be called anytime whether @ss
2840 * is attached or not.
2841 *
2842 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2843 * registered with @ss.
2844 */
2845int cgroup_rm_cftypes(struct cgroup_subsys *ss, const struct cftype *cfts)
2846{
2847 struct cftype_set *set;
2848
2849 cgroup_cfts_prepare();
2850
2851 list_for_each_entry(set, &ss->cftsets, node) {
2852 if (set->cfts == cfts) {
2853 list_del_init(&set->node);
2854 cgroup_cfts_commit(ss, cfts, false);
2855 return 0;
2856 }
2857 }
2858
2859 cgroup_cfts_commit(ss, NULL, false);
2860 return -ENOENT;
2861}
2862
2863/**
2864 * cgroup_task_count - count the number of tasks in a cgroup.
2865 * @cgrp: the cgroup in question
2866 *
2867 * Return the number of tasks in the cgroup.
2868 */
2869int cgroup_task_count(const struct cgroup *cgrp)
2870{
2871 int count = 0;
2872 struct cg_cgroup_link *link;
2873
2874 read_lock(&css_set_lock);
2875 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2876 count += atomic_read(&link->cg->refcount);
2877 }
2878 read_unlock(&css_set_lock);
2879 return count;
2880}
2881
2882/*
2883 * Advance a list_head iterator. The iterator should be positioned at
2884 * the start of a css_set
2885 */
2886static void cgroup_advance_iter(struct cgroup *cgrp,
2887 struct cgroup_iter *it)
2888{
2889 struct list_head *l = it->cg_link;
2890 struct cg_cgroup_link *link;
2891 struct css_set *cg;
2892
2893 /* Advance to the next non-empty css_set */
2894 do {
2895 l = l->next;
2896 if (l == &cgrp->css_sets) {
2897 it->cg_link = NULL;
2898 return;
2899 }
2900 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2901 cg = link->cg;
2902 } while (list_empty(&cg->tasks));
2903 it->cg_link = l;
2904 it->task = cg->tasks.next;
2905}
2906
2907/*
2908 * To reduce the fork() overhead for systems that are not actually
2909 * using their cgroups capability, we don't maintain the lists running
2910 * through each css_set to its tasks until we see the list actually
2911 * used - in other words after the first call to cgroup_iter_start().
2912 */
2913static void cgroup_enable_task_cg_lists(void)
2914{
2915 struct task_struct *p, *g;
2916 write_lock(&css_set_lock);
2917 use_task_css_set_links = 1;
2918 /*
2919 * We need tasklist_lock because RCU is not safe against
2920 * while_each_thread(). Besides, a forking task that has passed
2921 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2922 * is not guaranteed to have its child immediately visible in the
2923 * tasklist if we walk through it with RCU.
2924 */
2925 read_lock(&tasklist_lock);
2926 do_each_thread(g, p) {
2927 task_lock(p);
2928 /*
2929 * We should check if the process is exiting, otherwise
2930 * it will race with cgroup_exit() in that the list
2931 * entry won't be deleted though the process has exited.
2932 */
2933 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2934 list_add(&p->cg_list, &p->cgroups->tasks);
2935 task_unlock(p);
2936 } while_each_thread(g, p);
2937 read_unlock(&tasklist_lock);
2938 write_unlock(&css_set_lock);
2939}
2940
2941void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2942 __acquires(css_set_lock)
2943{
2944 /*
2945 * The first time anyone tries to iterate across a cgroup,
2946 * we need to enable the list linking each css_set to its
2947 * tasks, and fix up all existing tasks.
2948 */
2949 if (!use_task_css_set_links)
2950 cgroup_enable_task_cg_lists();
2951
2952 read_lock(&css_set_lock);
2953 it->cg_link = &cgrp->css_sets;
2954 cgroup_advance_iter(cgrp, it);
2955}
2956
2957struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2958 struct cgroup_iter *it)
2959{
2960 struct task_struct *res;
2961 struct list_head *l = it->task;
2962 struct cg_cgroup_link *link;
2963
2964 /* If the iterator cg is NULL, we have no tasks */
2965 if (!it->cg_link)
2966 return NULL;
2967 res = list_entry(l, struct task_struct, cg_list);
2968 /* Advance iterator to find next entry */
2969 l = l->next;
2970 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2971 if (l == &link->cg->tasks) {
2972 /* We reached the end of this task list - move on to
2973 * the next cg_cgroup_link */
2974 cgroup_advance_iter(cgrp, it);
2975 } else {
2976 it->task = l;
2977 }
2978 return res;
2979}
2980
2981void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2982 __releases(css_set_lock)
2983{
2984 read_unlock(&css_set_lock);
2985}
2986
2987static inline int started_after_time(struct task_struct *t1,
2988 struct timespec *time,
2989 struct task_struct *t2)
2990{
2991 int start_diff = timespec_compare(&t1->start_time, time);
2992 if (start_diff > 0) {
2993 return 1;
2994 } else if (start_diff < 0) {
2995 return 0;
2996 } else {
2997 /*
2998 * Arbitrarily, if two processes started at the same
2999 * time, we'll say that the lower pointer value
3000 * started first. Note that t2 may have exited by now
3001 * so this may not be a valid pointer any longer, but
3002 * that's fine - it still serves to distinguish
3003 * between two tasks started (effectively) simultaneously.
3004 */
3005 return t1 > t2;
3006 }
3007}
3008
3009/*
3010 * This function is a callback from heap_insert() and is used to order
3011 * the heap.
3012 * In this case we order the heap in descending task start time.
3013 */
3014static inline int started_after(void *p1, void *p2)
3015{
3016 struct task_struct *t1 = p1;
3017 struct task_struct *t2 = p2;
3018 return started_after_time(t1, &t2->start_time, t2);
3019}
3020
3021/**
3022 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
3023 * @scan: struct cgroup_scanner containing arguments for the scan
3024 *
3025 * Arguments include pointers to callback functions test_task() and
3026 * process_task().
3027 * Iterate through all the tasks in a cgroup, calling test_task() for each,
3028 * and if it returns true, call process_task() for it also.
3029 * The test_task pointer may be NULL, meaning always true (select all tasks).
3030 * Effectively duplicates cgroup_iter_{start,next,end}()
3031 * but does not lock css_set_lock for the call to process_task().
3032 * The struct cgroup_scanner may be embedded in any structure of the caller's
3033 * creation.
3034 * It is guaranteed that process_task() will act on every task that
3035 * is a member of the cgroup for the duration of this call. This
3036 * function may or may not call process_task() for tasks that exit
3037 * or move to a different cgroup during the call, or are forked or
3038 * move into the cgroup during the call.
3039 *
3040 * Note that test_task() may be called with locks held, and may in some
3041 * situations be called multiple times for the same task, so it should
3042 * be cheap.
3043 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
3044 * pre-allocated and will be used for heap operations (and its "gt" member will
3045 * be overwritten), else a temporary heap will be used (allocation of which
3046 * may cause this function to fail).
3047 */
3048int cgroup_scan_tasks(struct cgroup_scanner *scan)
3049{
3050 int retval, i;
3051 struct cgroup_iter it;
3052 struct task_struct *p, *dropped;
3053 /* Never dereference latest_task, since it's not refcounted */
3054 struct task_struct *latest_task = NULL;
3055 struct ptr_heap tmp_heap;
3056 struct ptr_heap *heap;
3057 struct timespec latest_time = { 0, 0 };
3058
3059 if (scan->heap) {
3060 /* The caller supplied our heap and pre-allocated its memory */
3061 heap = scan->heap;
3062 heap->gt = &started_after;
3063 } else {
3064 /* We need to allocate our own heap memory */
3065 heap = &tmp_heap;
3066 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3067 if (retval)
3068 /* cannot allocate the heap */
3069 return retval;
3070 }
3071
3072 again:
3073 /*
3074 * Scan tasks in the cgroup, using the scanner's "test_task" callback
3075 * to determine which are of interest, and using the scanner's
3076 * "process_task" callback to process any of them that need an update.
3077 * Since we don't want to hold any locks during the task updates,
3078 * gather tasks to be processed in a heap structure.
3079 * The heap is sorted by descending task start time.
3080 * If the statically-sized heap fills up, we overflow tasks that
3081 * started later, and in future iterations only consider tasks that
3082 * started after the latest task in the previous pass. This
3083 * guarantees forward progress and that we don't miss any tasks.
3084 */
3085 heap->size = 0;
3086 cgroup_iter_start(scan->cg, &it);
3087 while ((p = cgroup_iter_next(scan->cg, &it))) {
3088 /*
3089 * Only affect tasks that qualify per the caller's callback,
3090 * if he provided one
3091 */
3092 if (scan->test_task && !scan->test_task(p, scan))
3093 continue;
3094 /*
3095 * Only process tasks that started after the last task
3096 * we processed
3097 */
3098 if (!started_after_time(p, &latest_time, latest_task))
3099 continue;
3100 dropped = heap_insert(heap, p);
3101 if (dropped == NULL) {
3102 /*
3103 * The new task was inserted; the heap wasn't
3104 * previously full
3105 */
3106 get_task_struct(p);
3107 } else if (dropped != p) {
3108 /*
3109 * The new task was inserted, and pushed out a
3110 * different task
3111 */
3112 get_task_struct(p);
3113 put_task_struct(dropped);
3114 }
3115 /*
3116 * Else the new task was newer than anything already in
3117 * the heap and wasn't inserted
3118 */
3119 }
3120 cgroup_iter_end(scan->cg, &it);
3121
3122 if (heap->size) {
3123 for (i = 0; i < heap->size; i++) {
3124 struct task_struct *q = heap->ptrs[i];
3125 if (i == 0) {
3126 latest_time = q->start_time;
3127 latest_task = q;
3128 }
3129 /* Process the task per the caller's callback */
3130 scan->process_task(q, scan);
3131 put_task_struct(q);
3132 }
3133 /*
3134 * If we had to process any tasks at all, scan again
3135 * in case some of them were in the middle of forking
3136 * children that didn't get processed.
3137 * Not the most efficient way to do it, but it avoids
3138 * having to take callback_mutex in the fork path
3139 */
3140 goto again;
3141 }
3142 if (heap == &tmp_heap)
3143 heap_free(&tmp_heap);
3144 return 0;
3145}
3146
3147/*
3148 * Stuff for reading the 'tasks'/'procs' files.
3149 *
3150 * Reading this file can return large amounts of data if a cgroup has
3151 * *lots* of attached tasks. So it may need several calls to read(),
3152 * but we cannot guarantee that the information we produce is correct
3153 * unless we produce it entirely atomically.
3154 *
3155 */
3156
3157/* which pidlist file are we talking about? */
3158enum cgroup_filetype {
3159 CGROUP_FILE_PROCS,
3160 CGROUP_FILE_TASKS,
3161};
3162
3163/*
3164 * A pidlist is a list of pids that virtually represents the contents of one
3165 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3166 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3167 * to the cgroup.
3168 */
3169struct cgroup_pidlist {
3170 /*
3171 * used to find which pidlist is wanted. doesn't change as long as
3172 * this particular list stays in the list.
3173 */
3174 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3175 /* array of xids */
3176 pid_t *list;
3177 /* how many elements the above list has */
3178 int length;
3179 /* how many files are using the current array */
3180 int use_count;
3181 /* each of these stored in a list by its cgroup */
3182 struct list_head links;
3183 /* pointer to the cgroup we belong to, for list removal purposes */
3184 struct cgroup *owner;
3185 /* protects the other fields */
3186 struct rw_semaphore mutex;
3187};
3188
3189/*
3190 * The following two functions "fix" the issue where there are more pids
3191 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3192 * TODO: replace with a kernel-wide solution to this problem
3193 */
3194#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3195static void *pidlist_allocate(int count)
3196{
3197 if (PIDLIST_TOO_LARGE(count))
3198 return vmalloc(count * sizeof(pid_t));
3199 else
3200 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3201}
3202static void pidlist_free(void *p)
3203{
3204 if (is_vmalloc_addr(p))
3205 vfree(p);
3206 else
3207 kfree(p);
3208}
3209static void *pidlist_resize(void *p, int newcount)
3210{
3211 void *newlist;
3212 /* note: if new alloc fails, old p will still be valid either way */
3213 if (is_vmalloc_addr(p)) {
3214 newlist = vmalloc(newcount * sizeof(pid_t));
3215 if (!newlist)
3216 return NULL;
3217 memcpy(newlist, p, newcount * sizeof(pid_t));
3218 vfree(p);
3219 } else {
3220 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
3221 }
3222 return newlist;
3223}
3224
3225/*
3226 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3227 * If the new stripped list is sufficiently smaller and there's enough memory
3228 * to allocate a new buffer, will let go of the unneeded memory. Returns the
3229 * number of unique elements.
3230 */
3231/* is the size difference enough that we should re-allocate the array? */
3232#define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
3233static int pidlist_uniq(pid_t **p, int length)
3234{
3235 int src, dest = 1;
3236 pid_t *list = *p;
3237 pid_t *newlist;
3238
3239 /*
3240 * we presume the 0th element is unique, so i starts at 1. trivial
3241 * edge cases first; no work needs to be done for either
3242 */
3243 if (length == 0 || length == 1)
3244 return length;
3245 /* src and dest walk down the list; dest counts unique elements */
3246 for (src = 1; src < length; src++) {
3247 /* find next unique element */
3248 while (list[src] == list[src-1]) {
3249 src++;
3250 if (src == length)
3251 goto after;
3252 }
3253 /* dest always points to where the next unique element goes */
3254 list[dest] = list[src];
3255 dest++;
3256 }
3257after:
3258 /*
3259 * if the length difference is large enough, we want to allocate a
3260 * smaller buffer to save memory. if this fails due to out of memory,
3261 * we'll just stay with what we've got.
3262 */
3263 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
3264 newlist = pidlist_resize(list, dest);
3265 if (newlist)
3266 *p = newlist;
3267 }
3268 return dest;
3269}
3270
3271static int cmppid(const void *a, const void *b)
3272{
3273 return *(pid_t *)a - *(pid_t *)b;
3274}
3275
3276/*
3277 * find the appropriate pidlist for our purpose (given procs vs tasks)
3278 * returns with the lock on that pidlist already held, and takes care
3279 * of the use count, or returns NULL with no locks held if we're out of
3280 * memory.
3281 */
3282static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3283 enum cgroup_filetype type)
3284{
3285 struct cgroup_pidlist *l;
3286 /* don't need task_nsproxy() if we're looking at ourself */
3287 struct pid_namespace *ns = current->nsproxy->pid_ns;
3288
3289 /*
3290 * We can't drop the pidlist_mutex before taking the l->mutex in case
3291 * the last ref-holder is trying to remove l from the list at the same
3292 * time. Holding the pidlist_mutex precludes somebody taking whichever
3293 * list we find out from under us - compare release_pid_array().
3294 */
3295 mutex_lock(&cgrp->pidlist_mutex);
3296 list_for_each_entry(l, &cgrp->pidlists, links) {
3297 if (l->key.type == type && l->key.ns == ns) {
3298 /* make sure l doesn't vanish out from under us */
3299 down_write(&l->mutex);
3300 mutex_unlock(&cgrp->pidlist_mutex);
3301 return l;
3302 }
3303 }
3304 /* entry not found; create a new one */
3305 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3306 if (!l) {
3307 mutex_unlock(&cgrp->pidlist_mutex);
3308 return l;
3309 }
3310 init_rwsem(&l->mutex);
3311 down_write(&l->mutex);
3312 l->key.type = type;
3313 l->key.ns = get_pid_ns(ns);
3314 l->use_count = 0; /* don't increment here */
3315 l->list = NULL;
3316 l->owner = cgrp;
3317 list_add(&l->links, &cgrp->pidlists);
3318 mutex_unlock(&cgrp->pidlist_mutex);
3319 return l;
3320}
3321
3322/*
3323 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3324 */
3325static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3326 struct cgroup_pidlist **lp)
3327{
3328 pid_t *array;
3329 int length;
3330 int pid, n = 0; /* used for populating the array */
3331 struct cgroup_iter it;
3332 struct task_struct *tsk;
3333 struct cgroup_pidlist *l;
3334
3335 /*
3336 * If cgroup gets more users after we read count, we won't have
3337 * enough space - tough. This race is indistinguishable to the
3338 * caller from the case that the additional cgroup users didn't
3339 * show up until sometime later on.
3340 */
3341 length = cgroup_task_count(cgrp);
3342 array = pidlist_allocate(length);
3343 if (!array)
3344 return -ENOMEM;
3345 /* now, populate the array */
3346 cgroup_iter_start(cgrp, &it);
3347 while ((tsk = cgroup_iter_next(cgrp, &it))) {
3348 if (unlikely(n == length))
3349 break;
3350 /* get tgid or pid for procs or tasks file respectively */
3351 if (type == CGROUP_FILE_PROCS)
3352 pid = task_tgid_vnr(tsk);
3353 else
3354 pid = task_pid_vnr(tsk);
3355 if (pid > 0) /* make sure to only use valid results */
3356 array[n++] = pid;
3357 }
3358 cgroup_iter_end(cgrp, &it);
3359 length = n;
3360 /* now sort & (if procs) strip out duplicates */
3361 sort(array, length, sizeof(pid_t), cmppid, NULL);
3362 if (type == CGROUP_FILE_PROCS)
3363 length = pidlist_uniq(&array, length);
3364 l = cgroup_pidlist_find(cgrp, type);
3365 if (!l) {
3366 pidlist_free(array);
3367 return -ENOMEM;
3368 }
3369 /* store array, freeing old if necessary - lock already held */
3370 pidlist_free(l->list);
3371 l->list = array;
3372 l->length = length;
3373 l->use_count++;
3374 up_write(&l->mutex);
3375 *lp = l;
3376 return 0;
3377}
3378
3379/**
3380 * cgroupstats_build - build and fill cgroupstats
3381 * @stats: cgroupstats to fill information into
3382 * @dentry: A dentry entry belonging to the cgroup for which stats have
3383 * been requested.
3384 *
3385 * Build and fill cgroupstats so that taskstats can export it to user
3386 * space.
3387 */
3388int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3389{
3390 int ret = -EINVAL;
3391 struct cgroup *cgrp;
3392 struct cgroup_iter it;
3393 struct task_struct *tsk;
3394
3395 /*
3396 * Validate dentry by checking the superblock operations,
3397 * and make sure it's a directory.
3398 */
3399 if (dentry->d_sb->s_op != &cgroup_ops ||
3400 !S_ISDIR(dentry->d_inode->i_mode))
3401 goto err;
3402
3403 ret = 0;
3404 cgrp = dentry->d_fsdata;
3405
3406 cgroup_iter_start(cgrp, &it);
3407 while ((tsk = cgroup_iter_next(cgrp, &it))) {
3408 switch (tsk->state) {
3409 case TASK_RUNNING:
3410 stats->nr_running++;
3411 break;
3412 case TASK_INTERRUPTIBLE:
3413 stats->nr_sleeping++;
3414 break;
3415 case TASK_UNINTERRUPTIBLE:
3416 stats->nr_uninterruptible++;
3417 break;
3418 case TASK_STOPPED:
3419 stats->nr_stopped++;
3420 break;
3421 default:
3422 if (delayacct_is_task_waiting_on_io(tsk))
3423 stats->nr_io_wait++;
3424 break;
3425 }
3426 }
3427 cgroup_iter_end(cgrp, &it);
3428
3429err:
3430 return ret;
3431}
3432
3433
3434/*
3435 * seq_file methods for the tasks/procs files. The seq_file position is the
3436 * next pid to display; the seq_file iterator is a pointer to the pid
3437 * in the cgroup->l->list array.
3438 */
3439
3440static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3441{
3442 /*
3443 * Initially we receive a position value that corresponds to
3444 * one more than the last pid shown (or 0 on the first call or
3445 * after a seek to the start). Use a binary-search to find the
3446 * next pid to display, if any
3447 */
3448 struct cgroup_pidlist *l = s->private;
3449 int index = 0, pid = *pos;
3450 int *iter;
3451
3452 down_read(&l->mutex);
3453 if (pid) {
3454 int end = l->length;
3455
3456 while (index < end) {
3457 int mid = (index + end) / 2;
3458 if (l->list[mid] == pid) {
3459 index = mid;
3460 break;
3461 } else if (l->list[mid] <= pid)
3462 index = mid + 1;
3463 else
3464 end = mid;
3465 }
3466 }
3467 /* If we're off the end of the array, we're done */
3468 if (index >= l->length)
3469 return NULL;
3470 /* Update the abstract position to be the actual pid that we found */
3471 iter = l->list + index;
3472 *pos = *iter;
3473 return iter;
3474}
3475
3476static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3477{
3478 struct cgroup_pidlist *l = s->private;
3479 up_read(&l->mutex);
3480}
3481
3482static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3483{
3484 struct cgroup_pidlist *l = s->private;
3485 pid_t *p = v;
3486 pid_t *end = l->list + l->length;
3487 /*
3488 * Advance to the next pid in the array. If this goes off the
3489 * end, we're done
3490 */
3491 p++;
3492 if (p >= end) {
3493 return NULL;
3494 } else {
3495 *pos = *p;
3496 return p;
3497 }
3498}
3499
3500static int cgroup_pidlist_show(struct seq_file *s, void *v)
3501{
3502 return seq_printf(s, "%d\n", *(int *)v);
3503}
3504
3505/*
3506 * seq_operations functions for iterating on pidlists through seq_file -
3507 * independent of whether it's tasks or procs
3508 */
3509static const struct seq_operations cgroup_pidlist_seq_operations = {
3510 .start = cgroup_pidlist_start,
3511 .stop = cgroup_pidlist_stop,
3512 .next = cgroup_pidlist_next,
3513 .show = cgroup_pidlist_show,
3514};
3515
3516static void cgroup_release_pid_array(struct cgroup_pidlist *l)
3517{
3518 /*
3519 * the case where we're the last user of this particular pidlist will
3520 * have us remove it from the cgroup's list, which entails taking the
3521 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
3522 * pidlist_mutex, we have to take pidlist_mutex first.
3523 */
3524 mutex_lock(&l->owner->pidlist_mutex);
3525 down_write(&l->mutex);
3526 BUG_ON(!l->use_count);
3527 if (!--l->use_count) {
3528 /* we're the last user if refcount is 0; remove and free */
3529 list_del(&l->links);
3530 mutex_unlock(&l->owner->pidlist_mutex);
3531 pidlist_free(l->list);
3532 put_pid_ns(l->key.ns);
3533 up_write(&l->mutex);
3534 kfree(l);
3535 return;
3536 }
3537 mutex_unlock(&l->owner->pidlist_mutex);
3538 up_write(&l->mutex);
3539}
3540
3541static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3542{
3543 struct cgroup_pidlist *l;
3544 if (!(file->f_mode & FMODE_READ))
3545 return 0;
3546 /*
3547 * the seq_file will only be initialized if the file was opened for
3548 * reading; hence we check if it's not null only in that case.
3549 */
3550 l = ((struct seq_file *)file->private_data)->private;
3551 cgroup_release_pid_array(l);
3552 return seq_release(inode, file);
3553}
3554
3555static const struct file_operations cgroup_pidlist_operations = {
3556 .read = seq_read,
3557 .llseek = seq_lseek,
3558 .write = cgroup_file_write,
3559 .release = cgroup_pidlist_release,
3560};
3561
3562/*
3563 * The following functions handle opens on a file that displays a pidlist
3564 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3565 * in the cgroup.
3566 */
3567/* helper function for the two below it */
3568static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3569{
3570 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3571 struct cgroup_pidlist *l;
3572 int retval;
3573
3574 /* Nothing to do for write-only files */
3575 if (!(file->f_mode & FMODE_READ))
3576 return 0;
3577
3578 /* have the array populated */
3579 retval = pidlist_array_load(cgrp, type, &l);
3580 if (retval)
3581 return retval;
3582 /* configure file information */
3583 file->f_op = &cgroup_pidlist_operations;
3584
3585 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3586 if (retval) {
3587 cgroup_release_pid_array(l);
3588 return retval;
3589 }
3590 ((struct seq_file *)file->private_data)->private = l;
3591 return 0;
3592}
3593static int cgroup_tasks_open(struct inode *unused, struct file *file)
3594{
3595 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3596}
3597static int cgroup_procs_open(struct inode *unused, struct file *file)
3598{
3599 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3600}
3601
3602static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3603 struct cftype *cft)
3604{
3605 return notify_on_release(cgrp);
3606}
3607
3608static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3609 struct cftype *cft,
3610 u64 val)
3611{
3612 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3613 if (val)
3614 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3615 else
3616 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3617 return 0;
3618}
3619
3620/*
3621 * Unregister event and free resources.
3622 *
3623 * Gets called from workqueue.
3624 */
3625static void cgroup_event_remove(struct work_struct *work)
3626{
3627 struct cgroup_event *event = container_of(work, struct cgroup_event,
3628 remove);
3629 struct cgroup *cgrp = event->cgrp;
3630
3631 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3632
3633 eventfd_ctx_put(event->eventfd);
3634 kfree(event);
3635 dput(cgrp->dentry);
3636}
3637
3638/*
3639 * Gets called on POLLHUP on eventfd when user closes it.
3640 *
3641 * Called with wqh->lock held and interrupts disabled.
3642 */
3643static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3644 int sync, void *key)
3645{
3646 struct cgroup_event *event = container_of(wait,
3647 struct cgroup_event, wait);
3648 struct cgroup *cgrp = event->cgrp;
3649 unsigned long flags = (unsigned long)key;
3650
3651 if (flags & POLLHUP) {
3652 __remove_wait_queue(event->wqh, &event->wait);
3653 spin_lock(&cgrp->event_list_lock);
3654 list_del(&event->list);
3655 spin_unlock(&cgrp->event_list_lock);
3656 /*
3657 * We are in atomic context, but cgroup_event_remove() may
3658 * sleep, so we have to call it in workqueue.
3659 */
3660 schedule_work(&event->remove);
3661 }
3662
3663 return 0;
3664}
3665
3666static void cgroup_event_ptable_queue_proc(struct file *file,
3667 wait_queue_head_t *wqh, poll_table *pt)
3668{
3669 struct cgroup_event *event = container_of(pt,
3670 struct cgroup_event, pt);
3671
3672 event->wqh = wqh;
3673 add_wait_queue(wqh, &event->wait);
3674}
3675
3676/*
3677 * Parse input and register new cgroup event handler.
3678 *
3679 * Input must be in format '<event_fd> <control_fd> <args>'.
3680 * Interpretation of args is defined by control file implementation.
3681 */
3682static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3683 const char *buffer)
3684{
3685 struct cgroup_event *event = NULL;
3686 unsigned int efd, cfd;
3687 struct file *efile = NULL;
3688 struct file *cfile = NULL;
3689 char *endp;
3690 int ret;
3691
3692 efd = simple_strtoul(buffer, &endp, 10);
3693 if (*endp != ' ')
3694 return -EINVAL;
3695 buffer = endp + 1;
3696
3697 cfd = simple_strtoul(buffer, &endp, 10);
3698 if ((*endp != ' ') && (*endp != '\0'))
3699 return -EINVAL;
3700 buffer = endp + 1;
3701
3702 event = kzalloc(sizeof(*event), GFP_KERNEL);
3703 if (!event)
3704 return -ENOMEM;
3705 event->cgrp = cgrp;
3706 INIT_LIST_HEAD(&event->list);
3707 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3708 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3709 INIT_WORK(&event->remove, cgroup_event_remove);
3710
3711 efile = eventfd_fget(efd);
3712 if (IS_ERR(efile)) {
3713 ret = PTR_ERR(efile);
3714 goto fail;
3715 }
3716
3717 event->eventfd = eventfd_ctx_fileget(efile);
3718 if (IS_ERR(event->eventfd)) {
3719 ret = PTR_ERR(event->eventfd);
3720 goto fail;
3721 }
3722
3723 cfile = fget(cfd);
3724 if (!cfile) {
3725 ret = -EBADF;
3726 goto fail;
3727 }
3728
3729 /* the process need read permission on control file */
3730 /* AV: shouldn't we check that it's been opened for read instead? */
3731 ret = inode_permission(cfile->f_path.dentry->d_inode, MAY_READ);
3732 if (ret < 0)
3733 goto fail;
3734
3735 event->cft = __file_cft(cfile);
3736 if (IS_ERR(event->cft)) {
3737 ret = PTR_ERR(event->cft);
3738 goto fail;
3739 }
3740
3741 if (!event->cft->register_event || !event->cft->unregister_event) {
3742 ret = -EINVAL;
3743 goto fail;
3744 }
3745
3746 ret = event->cft->register_event(cgrp, event->cft,
3747 event->eventfd, buffer);
3748 if (ret)
3749 goto fail;
3750
3751 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3752 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3753 ret = 0;
3754 goto fail;
3755 }
3756
3757 /*
3758 * Events should be removed after rmdir of cgroup directory, but before
3759 * destroying subsystem state objects. Let's take reference to cgroup
3760 * directory dentry to do that.
3761 */
3762 dget(cgrp->dentry);
3763
3764 spin_lock(&cgrp->event_list_lock);
3765 list_add(&event->list, &cgrp->event_list);
3766 spin_unlock(&cgrp->event_list_lock);
3767
3768 fput(cfile);
3769 fput(efile);
3770
3771 return 0;
3772
3773fail:
3774 if (cfile)
3775 fput(cfile);
3776
3777 if (event && event->eventfd && !IS_ERR(event->eventfd))
3778 eventfd_ctx_put(event->eventfd);
3779
3780 if (!IS_ERR_OR_NULL(efile))
3781 fput(efile);
3782
3783 kfree(event);
3784
3785 return ret;
3786}
3787
3788static u64 cgroup_clone_children_read(struct cgroup *cgrp,
3789 struct cftype *cft)
3790{
3791 return clone_children(cgrp);
3792}
3793
3794static int cgroup_clone_children_write(struct cgroup *cgrp,
3795 struct cftype *cft,
3796 u64 val)
3797{
3798 if (val)
3799 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3800 else
3801 clear_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3802 return 0;
3803}
3804
3805/*
3806 * for the common functions, 'private' gives the type of file
3807 */
3808/* for hysterical raisins, we can't put this on the older files */
3809#define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3810static struct cftype files[] = {
3811 {
3812 .name = "tasks",
3813 .open = cgroup_tasks_open,
3814 .write_u64 = cgroup_tasks_write,
3815 .release = cgroup_pidlist_release,
3816 .mode = S_IRUGO | S_IWUSR,
3817 },
3818 {
3819 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3820 .open = cgroup_procs_open,
3821 .write_u64 = cgroup_procs_write,
3822 .release = cgroup_pidlist_release,
3823 .mode = S_IRUGO | S_IWUSR,
3824 },
3825 {
3826 .name = "notify_on_release",
3827 .read_u64 = cgroup_read_notify_on_release,
3828 .write_u64 = cgroup_write_notify_on_release,
3829 },
3830 {
3831 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3832 .write_string = cgroup_write_event_control,
3833 .mode = S_IWUGO,
3834 },
3835 {
3836 .name = "cgroup.clone_children",
3837 .read_u64 = cgroup_clone_children_read,
3838 .write_u64 = cgroup_clone_children_write,
3839 },
3840 {
3841 .name = "release_agent",
3842 .flags = CFTYPE_ONLY_ON_ROOT,
3843 .read_seq_string = cgroup_release_agent_show,
3844 .write_string = cgroup_release_agent_write,
3845 .max_write_len = PATH_MAX,
3846 },
3847 { } /* terminate */
3848};
3849
3850static int cgroup_populate_dir(struct cgroup *cgrp)
3851{
3852 int err;
3853 struct cgroup_subsys *ss;
3854
3855 err = cgroup_addrm_files(cgrp, NULL, files, true);
3856 if (err < 0)
3857 return err;
3858
3859 /* process cftsets of each subsystem */
3860 for_each_subsys(cgrp->root, ss) {
3861 struct cftype_set *set;
3862
3863 list_for_each_entry(set, &ss->cftsets, node)
3864 cgroup_addrm_files(cgrp, ss, set->cfts, true);
3865 }
3866
3867 /* This cgroup is ready now */
3868 for_each_subsys(cgrp->root, ss) {
3869 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3870 /*
3871 * Update id->css pointer and make this css visible from
3872 * CSS ID functions. This pointer will be dereferened
3873 * from RCU-read-side without locks.
3874 */
3875 if (css->id)
3876 rcu_assign_pointer(css->id->css, css);
3877 }
3878
3879 return 0;
3880}
3881
3882static void css_dput_fn(struct work_struct *work)
3883{
3884 struct cgroup_subsys_state *css =
3885 container_of(work, struct cgroup_subsys_state, dput_work);
3886 struct dentry *dentry = css->cgroup->dentry;
3887 struct super_block *sb = dentry->d_sb;
3888
3889 atomic_inc(&sb->s_active);
3890 dput(dentry);
3891 deactivate_super(sb);
3892}
3893
3894static void init_cgroup_css(struct cgroup_subsys_state *css,
3895 struct cgroup_subsys *ss,
3896 struct cgroup *cgrp)
3897{
3898 css->cgroup = cgrp;
3899 atomic_set(&css->refcnt, 1);
3900 css->flags = 0;
3901 css->id = NULL;
3902 if (cgrp == dummytop)
3903 set_bit(CSS_ROOT, &css->flags);
3904 BUG_ON(cgrp->subsys[ss->subsys_id]);
3905 cgrp->subsys[ss->subsys_id] = css;
3906
3907 /*
3908 * If !clear_css_refs, css holds an extra ref to @cgrp->dentry
3909 * which is put on the last css_put(). dput() requires process
3910 * context, which css_put() may be called without. @css->dput_work
3911 * will be used to invoke dput() asynchronously from css_put().
3912 */
3913 INIT_WORK(&css->dput_work, css_dput_fn);
3914 if (ss->__DEPRECATED_clear_css_refs)
3915 set_bit(CSS_CLEAR_CSS_REFS, &css->flags);
3916}
3917
3918static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3919{
3920 /* We need to take each hierarchy_mutex in a consistent order */
3921 int i;
3922
3923 /*
3924 * No worry about a race with rebind_subsystems that might mess up the
3925 * locking order, since both parties are under cgroup_mutex.
3926 */
3927 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3928 struct cgroup_subsys *ss = subsys[i];
3929 if (ss == NULL)
3930 continue;
3931 if (ss->root == root)
3932 mutex_lock(&ss->hierarchy_mutex);
3933 }
3934}
3935
3936static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3937{
3938 int i;
3939
3940 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3941 struct cgroup_subsys *ss = subsys[i];
3942 if (ss == NULL)
3943 continue;
3944 if (ss->root == root)
3945 mutex_unlock(&ss->hierarchy_mutex);
3946 }
3947}
3948
3949/*
3950 * cgroup_create - create a cgroup
3951 * @parent: cgroup that will be parent of the new cgroup
3952 * @dentry: dentry of the new cgroup
3953 * @mode: mode to set on new inode
3954 *
3955 * Must be called with the mutex on the parent inode held
3956 */
3957static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3958 umode_t mode)
3959{
3960 struct cgroup *cgrp;
3961 struct cgroupfs_root *root = parent->root;
3962 int err = 0;
3963 struct cgroup_subsys *ss;
3964 struct super_block *sb = root->sb;
3965
3966 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3967 if (!cgrp)
3968 return -ENOMEM;
3969
3970 /* Grab a reference on the superblock so the hierarchy doesn't
3971 * get deleted on unmount if there are child cgroups. This
3972 * can be done outside cgroup_mutex, since the sb can't
3973 * disappear while someone has an open control file on the
3974 * fs */
3975 atomic_inc(&sb->s_active);
3976
3977 mutex_lock(&cgroup_mutex);
3978
3979 init_cgroup_housekeeping(cgrp);
3980
3981 cgrp->parent = parent;
3982 cgrp->root = parent->root;
3983 cgrp->top_cgroup = parent->top_cgroup;
3984
3985 if (notify_on_release(parent))
3986 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3987
3988 if (clone_children(parent))
3989 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3990
3991 for_each_subsys(root, ss) {
3992 struct cgroup_subsys_state *css = ss->create(cgrp);
3993
3994 if (IS_ERR(css)) {
3995 err = PTR_ERR(css);
3996 goto err_destroy;
3997 }
3998 init_cgroup_css(css, ss, cgrp);
3999 if (ss->use_id) {
4000 err = alloc_css_id(ss, parent, cgrp);
4001 if (err)
4002 goto err_destroy;
4003 }
4004 /* At error, ->destroy() callback has to free assigned ID. */
4005 if (clone_children(parent) && ss->post_clone)
4006 ss->post_clone(cgrp);
4007 }
4008
4009 cgroup_lock_hierarchy(root);
4010 list_add(&cgrp->sibling, &cgrp->parent->children);
4011 cgroup_unlock_hierarchy(root);
4012 root->number_of_cgroups++;
4013
4014 err = cgroup_create_dir(cgrp, dentry, mode);
4015 if (err < 0)
4016 goto err_remove;
4017
4018 /* If !clear_css_refs, each css holds a ref to the cgroup's dentry */
4019 for_each_subsys(root, ss)
4020 if (!ss->__DEPRECATED_clear_css_refs)
4021 dget(dentry);
4022
4023 /* The cgroup directory was pre-locked for us */
4024 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
4025
4026 list_add_tail(&cgrp->allcg_node, &root->allcg_list);
4027
4028 err = cgroup_populate_dir(cgrp);
4029 /* If err < 0, we have a half-filled directory - oh well ;) */
4030
4031 mutex_unlock(&cgroup_mutex);
4032 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4033
4034 return 0;
4035
4036 err_remove:
4037
4038 cgroup_lock_hierarchy(root);
4039 list_del(&cgrp->sibling);
4040 cgroup_unlock_hierarchy(root);
4041 root->number_of_cgroups--;
4042
4043 err_destroy:
4044
4045 for_each_subsys(root, ss) {
4046 if (cgrp->subsys[ss->subsys_id])
4047 ss->destroy(cgrp);
4048 }
4049
4050 mutex_unlock(&cgroup_mutex);
4051
4052 /* Release the reference count that we took on the superblock */
4053 deactivate_super(sb);
4054
4055 kfree(cgrp);
4056 return err;
4057}
4058
4059static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4060{
4061 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4062
4063 /* the vfs holds inode->i_mutex already */
4064 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4065}
4066
4067/*
4068 * Check the reference count on each subsystem. Since we already
4069 * established that there are no tasks in the cgroup, if the css refcount
4070 * is also 1, then there should be no outstanding references, so the
4071 * subsystem is safe to destroy. We scan across all subsystems rather than
4072 * using the per-hierarchy linked list of mounted subsystems since we can
4073 * be called via check_for_release() with no synchronization other than
4074 * RCU, and the subsystem linked list isn't RCU-safe.
4075 */
4076static int cgroup_has_css_refs(struct cgroup *cgrp)
4077{
4078 int i;
4079
4080 /*
4081 * We won't need to lock the subsys array, because the subsystems
4082 * we're concerned about aren't going anywhere since our cgroup root
4083 * has a reference on them.
4084 */
4085 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4086 struct cgroup_subsys *ss = subsys[i];
4087 struct cgroup_subsys_state *css;
4088
4089 /* Skip subsystems not present or not in this hierarchy */
4090 if (ss == NULL || ss->root != cgrp->root)
4091 continue;
4092
4093 css = cgrp->subsys[ss->subsys_id];
4094 /*
4095 * When called from check_for_release() it's possible
4096 * that by this point the cgroup has been removed
4097 * and the css deleted. But a false-positive doesn't
4098 * matter, since it can only happen if the cgroup
4099 * has been deleted and hence no longer needs the
4100 * release agent to be called anyway.
4101 */
4102 if (css && css_refcnt(css) > 1)
4103 return 1;
4104 }
4105 return 0;
4106}
4107
4108/*
4109 * Atomically mark all (or else none) of the cgroup's CSS objects as
4110 * CSS_REMOVED. Return true on success, or false if the cgroup has
4111 * busy subsystems. Call with cgroup_mutex held
4112 *
4113 * Depending on whether a subsys has __DEPRECATED_clear_css_refs set or
4114 * not, cgroup removal behaves differently.
4115 *
4116 * If clear is set, css refcnt for the subsystem should be zero before
4117 * cgroup removal can be committed. This is implemented by
4118 * CGRP_WAIT_ON_RMDIR and retry logic around ->pre_destroy(), which may be
4119 * called multiple times until all css refcnts reach zero and is allowed to
4120 * veto removal on any invocation. This behavior is deprecated and will be
4121 * removed as soon as the existing user (memcg) is updated.
4122 *
4123 * If clear is not set, each css holds an extra reference to the cgroup's
4124 * dentry and cgroup removal proceeds regardless of css refs.
4125 * ->pre_destroy() will be called at least once and is not allowed to fail.
4126 * On the last put of each css, whenever that may be, the extra dentry ref
4127 * is put so that dentry destruction happens only after all css's are
4128 * released.
4129 */
4130static int cgroup_clear_css_refs(struct cgroup *cgrp)
4131{
4132 struct cgroup_subsys *ss;
4133 unsigned long flags;
4134 bool failed = false;
4135
4136 local_irq_save(flags);
4137
4138 /*
4139 * Block new css_tryget() by deactivating refcnt. If all refcnts
4140 * for subsystems w/ clear_css_refs set were 1 at the moment of
4141 * deactivation, we succeeded.
4142 */
4143 for_each_subsys(cgrp->root, ss) {
4144 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4145
4146 WARN_ON(atomic_read(&css->refcnt) < 0);
4147 atomic_add(CSS_DEACT_BIAS, &css->refcnt);
4148
4149 if (ss->__DEPRECATED_clear_css_refs)
4150 failed |= css_refcnt(css) != 1;
4151 }
4152
4153 /*
4154 * If succeeded, set REMOVED and put all the base refs; otherwise,
4155 * restore refcnts to positive values. Either way, all in-progress
4156 * css_tryget() will be released.
4157 */
4158 for_each_subsys(cgrp->root, ss) {
4159 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4160
4161 if (!failed) {
4162 set_bit(CSS_REMOVED, &css->flags);
4163 css_put(css);
4164 } else {
4165 atomic_sub(CSS_DEACT_BIAS, &css->refcnt);
4166 }
4167 }
4168
4169 local_irq_restore(flags);
4170 return !failed;
4171}
4172
4173static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4174{
4175 struct cgroup *cgrp = dentry->d_fsdata;
4176 struct dentry *d;
4177 struct cgroup *parent;
4178 DEFINE_WAIT(wait);
4179 struct cgroup_event *event, *tmp;
4180 int ret;
4181
4182 /* the vfs holds both inode->i_mutex already */
4183again:
4184 mutex_lock(&cgroup_mutex);
4185 if (atomic_read(&cgrp->count) != 0) {
4186 mutex_unlock(&cgroup_mutex);
4187 return -EBUSY;
4188 }
4189 if (!list_empty(&cgrp->children)) {
4190 mutex_unlock(&cgroup_mutex);
4191 return -EBUSY;
4192 }
4193 mutex_unlock(&cgroup_mutex);
4194
4195 /*
4196 * In general, subsystem has no css->refcnt after pre_destroy(). But
4197 * in racy cases, subsystem may have to get css->refcnt after
4198 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
4199 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
4200 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
4201 * and subsystem's reference count handling. Please see css_get/put
4202 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
4203 */
4204 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4205
4206 /*
4207 * Call pre_destroy handlers of subsys. Notify subsystems
4208 * that rmdir() request comes.
4209 */
4210 ret = cgroup_call_pre_destroy(cgrp);
4211 if (ret) {
4212 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4213 return ret;
4214 }
4215
4216 mutex_lock(&cgroup_mutex);
4217 parent = cgrp->parent;
4218 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
4219 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4220 mutex_unlock(&cgroup_mutex);
4221 return -EBUSY;
4222 }
4223 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
4224 if (!cgroup_clear_css_refs(cgrp)) {
4225 mutex_unlock(&cgroup_mutex);
4226 /*
4227 * Because someone may call cgroup_wakeup_rmdir_waiter() before
4228 * prepare_to_wait(), we need to check this flag.
4229 */
4230 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
4231 schedule();
4232 finish_wait(&cgroup_rmdir_waitq, &wait);
4233 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4234 if (signal_pending(current))
4235 return -EINTR;
4236 goto again;
4237 }
4238 /* NO css_tryget() can success after here. */
4239 finish_wait(&cgroup_rmdir_waitq, &wait);
4240 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4241
4242 raw_spin_lock(&release_list_lock);
4243 set_bit(CGRP_REMOVED, &cgrp->flags);
4244 if (!list_empty(&cgrp->release_list))
4245 list_del_init(&cgrp->release_list);
4246 raw_spin_unlock(&release_list_lock);
4247
4248 cgroup_lock_hierarchy(cgrp->root);
4249 /* delete this cgroup from parent->children */
4250 list_del_init(&cgrp->sibling);
4251 cgroup_unlock_hierarchy(cgrp->root);
4252
4253 list_del_init(&cgrp->allcg_node);
4254
4255 d = dget(cgrp->dentry);
4256
4257 cgroup_d_remove_dir(d);
4258 dput(d);
4259
4260 set_bit(CGRP_RELEASABLE, &parent->flags);
4261 check_for_release(parent);
4262
4263 /*
4264 * Unregister events and notify userspace.
4265 * Notify userspace about cgroup removing only after rmdir of cgroup
4266 * directory to avoid race between userspace and kernelspace
4267 */
4268 spin_lock(&cgrp->event_list_lock);
4269 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
4270 list_del(&event->list);
4271 remove_wait_queue(event->wqh, &event->wait);
4272 eventfd_signal(event->eventfd, 1);
4273 schedule_work(&event->remove);
4274 }
4275 spin_unlock(&cgrp->event_list_lock);
4276
4277 mutex_unlock(&cgroup_mutex);
4278 return 0;
4279}
4280
4281static void __init_or_module cgroup_init_cftsets(struct cgroup_subsys *ss)
4282{
4283 INIT_LIST_HEAD(&ss->cftsets);
4284
4285 /*
4286 * base_cftset is embedded in subsys itself, no need to worry about
4287 * deregistration.
4288 */
4289 if (ss->base_cftypes) {
4290 ss->base_cftset.cfts = ss->base_cftypes;
4291 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4292 }
4293}
4294
4295static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4296{
4297 struct cgroup_subsys_state *css;
4298
4299 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4300
4301 /* init base cftset */
4302 cgroup_init_cftsets(ss);
4303
4304 /* Create the top cgroup state for this subsystem */
4305 list_add(&ss->sibling, &rootnode.subsys_list);
4306 ss->root = &rootnode;
4307 css = ss->create(dummytop);
4308 /* We don't handle early failures gracefully */
4309 BUG_ON(IS_ERR(css));
4310 init_cgroup_css(css, ss, dummytop);
4311
4312 /* Update the init_css_set to contain a subsys
4313 * pointer to this state - since the subsystem is
4314 * newly registered, all tasks and hence the
4315 * init_css_set is in the subsystem's top cgroup. */
4316 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
4317
4318 need_forkexit_callback |= ss->fork || ss->exit;
4319
4320 /* At system boot, before all subsystems have been
4321 * registered, no tasks have been forked, so we don't
4322 * need to invoke fork callbacks here. */
4323 BUG_ON(!list_empty(&init_task.tasks));
4324
4325 mutex_init(&ss->hierarchy_mutex);
4326 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
4327 ss->active = 1;
4328
4329 /* this function shouldn't be used with modular subsystems, since they
4330 * need to register a subsys_id, among other things */
4331 BUG_ON(ss->module);
4332}
4333
4334/**
4335 * cgroup_load_subsys: load and register a modular subsystem at runtime
4336 * @ss: the subsystem to load
4337 *
4338 * This function should be called in a modular subsystem's initcall. If the
4339 * subsystem is built as a module, it will be assigned a new subsys_id and set
4340 * up for use. If the subsystem is built-in anyway, work is delegated to the
4341 * simpler cgroup_init_subsys.
4342 */
4343int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4344{
4345 int i;
4346 struct cgroup_subsys_state *css;
4347
4348 /* check name and function validity */
4349 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4350 ss->create == NULL || ss->destroy == NULL)
4351 return -EINVAL;
4352
4353 /*
4354 * we don't support callbacks in modular subsystems. this check is
4355 * before the ss->module check for consistency; a subsystem that could
4356 * be a module should still have no callbacks even if the user isn't
4357 * compiling it as one.
4358 */
4359 if (ss->fork || ss->exit)
4360 return -EINVAL;
4361
4362 /*
4363 * an optionally modular subsystem is built-in: we want to do nothing,
4364 * since cgroup_init_subsys will have already taken care of it.
4365 */
4366 if (ss->module == NULL) {
4367 /* a few sanity checks */
4368 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
4369 BUG_ON(subsys[ss->subsys_id] != ss);
4370 return 0;
4371 }
4372
4373 /* init base cftset */
4374 cgroup_init_cftsets(ss);
4375
4376 /*
4377 * need to register a subsys id before anything else - for example,
4378 * init_cgroup_css needs it.
4379 */
4380 mutex_lock(&cgroup_mutex);
4381 /* find the first empty slot in the array */
4382 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
4383 if (subsys[i] == NULL)
4384 break;
4385 }
4386 if (i == CGROUP_SUBSYS_COUNT) {
4387 /* maximum number of subsystems already registered! */
4388 mutex_unlock(&cgroup_mutex);
4389 return -EBUSY;
4390 }
4391 /* assign ourselves the subsys_id */
4392 ss->subsys_id = i;
4393 subsys[i] = ss;
4394
4395 /*
4396 * no ss->create seems to need anything important in the ss struct, so
4397 * this can happen first (i.e. before the rootnode attachment).
4398 */
4399 css = ss->create(dummytop);
4400 if (IS_ERR(css)) {
4401 /* failure case - need to deassign the subsys[] slot. */
4402 subsys[i] = NULL;
4403 mutex_unlock(&cgroup_mutex);
4404 return PTR_ERR(css);
4405 }
4406
4407 list_add(&ss->sibling, &rootnode.subsys_list);
4408 ss->root = &rootnode;
4409
4410 /* our new subsystem will be attached to the dummy hierarchy. */
4411 init_cgroup_css(css, ss, dummytop);
4412 /* init_idr must be after init_cgroup_css because it sets css->id. */
4413 if (ss->use_id) {
4414 int ret = cgroup_init_idr(ss, css);
4415 if (ret) {
4416 dummytop->subsys[ss->subsys_id] = NULL;
4417 ss->destroy(dummytop);
4418 subsys[i] = NULL;
4419 mutex_unlock(&cgroup_mutex);
4420 return ret;
4421 }
4422 }
4423
4424 /*
4425 * Now we need to entangle the css into the existing css_sets. unlike
4426 * in cgroup_init_subsys, there are now multiple css_sets, so each one
4427 * will need a new pointer to it; done by iterating the css_set_table.
4428 * furthermore, modifying the existing css_sets will corrupt the hash
4429 * table state, so each changed css_set will need its hash recomputed.
4430 * this is all done under the css_set_lock.
4431 */
4432 write_lock(&css_set_lock);
4433 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
4434 struct css_set *cg;
4435 struct hlist_node *node, *tmp;
4436 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
4437
4438 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
4439 /* skip entries that we already rehashed */
4440 if (cg->subsys[ss->subsys_id])
4441 continue;
4442 /* remove existing entry */
4443 hlist_del(&cg->hlist);
4444 /* set new value */
4445 cg->subsys[ss->subsys_id] = css;
4446 /* recompute hash and restore entry */
4447 new_bucket = css_set_hash(cg->subsys);
4448 hlist_add_head(&cg->hlist, new_bucket);
4449 }
4450 }
4451 write_unlock(&css_set_lock);
4452
4453 mutex_init(&ss->hierarchy_mutex);
4454 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
4455 ss->active = 1;
4456
4457 /* success! */
4458 mutex_unlock(&cgroup_mutex);
4459 return 0;
4460}
4461EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4462
4463/**
4464 * cgroup_unload_subsys: unload a modular subsystem
4465 * @ss: the subsystem to unload
4466 *
4467 * This function should be called in a modular subsystem's exitcall. When this
4468 * function is invoked, the refcount on the subsystem's module will be 0, so
4469 * the subsystem will not be attached to any hierarchy.
4470 */
4471void cgroup_unload_subsys(struct cgroup_subsys *ss)
4472{
4473 struct cg_cgroup_link *link;
4474 struct hlist_head *hhead;
4475
4476 BUG_ON(ss->module == NULL);
4477
4478 /*
4479 * we shouldn't be called if the subsystem is in use, and the use of
4480 * try_module_get in parse_cgroupfs_options should ensure that it
4481 * doesn't start being used while we're killing it off.
4482 */
4483 BUG_ON(ss->root != &rootnode);
4484
4485 mutex_lock(&cgroup_mutex);
4486 /* deassign the subsys_id */
4487 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
4488 subsys[ss->subsys_id] = NULL;
4489
4490 /* remove subsystem from rootnode's list of subsystems */
4491 list_del_init(&ss->sibling);
4492
4493 /*
4494 * disentangle the css from all css_sets attached to the dummytop. as
4495 * in loading, we need to pay our respects to the hashtable gods.
4496 */
4497 write_lock(&css_set_lock);
4498 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
4499 struct css_set *cg = link->cg;
4500
4501 hlist_del(&cg->hlist);
4502 BUG_ON(!cg->subsys[ss->subsys_id]);
4503 cg->subsys[ss->subsys_id] = NULL;
4504 hhead = css_set_hash(cg->subsys);
4505 hlist_add_head(&cg->hlist, hhead);
4506 }
4507 write_unlock(&css_set_lock);
4508
4509 /*
4510 * remove subsystem's css from the dummytop and free it - need to free
4511 * before marking as null because ss->destroy needs the cgrp->subsys
4512 * pointer to find their state. note that this also takes care of
4513 * freeing the css_id.
4514 */
4515 ss->destroy(dummytop);
4516 dummytop->subsys[ss->subsys_id] = NULL;
4517
4518 mutex_unlock(&cgroup_mutex);
4519}
4520EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4521
4522/**
4523 * cgroup_init_early - cgroup initialization at system boot
4524 *
4525 * Initialize cgroups at system boot, and initialize any
4526 * subsystems that request early init.
4527 */
4528int __init cgroup_init_early(void)
4529{
4530 int i;
4531 atomic_set(&init_css_set.refcount, 1);
4532 INIT_LIST_HEAD(&init_css_set.cg_links);
4533 INIT_LIST_HEAD(&init_css_set.tasks);
4534 INIT_HLIST_NODE(&init_css_set.hlist);
4535 css_set_count = 1;
4536 init_cgroup_root(&rootnode);
4537 root_count = 1;
4538 init_task.cgroups = &init_css_set;
4539
4540 init_css_set_link.cg = &init_css_set;
4541 init_css_set_link.cgrp = dummytop;
4542 list_add(&init_css_set_link.cgrp_link_list,
4543 &rootnode.top_cgroup.css_sets);
4544 list_add(&init_css_set_link.cg_link_list,
4545 &init_css_set.cg_links);
4546
4547 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
4548 INIT_HLIST_HEAD(&css_set_table[i]);
4549
4550 /* at bootup time, we don't worry about modular subsystems */
4551 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4552 struct cgroup_subsys *ss = subsys[i];
4553
4554 BUG_ON(!ss->name);
4555 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4556 BUG_ON(!ss->create);
4557 BUG_ON(!ss->destroy);
4558 if (ss->subsys_id != i) {
4559 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4560 ss->name, ss->subsys_id);
4561 BUG();
4562 }
4563
4564 if (ss->early_init)
4565 cgroup_init_subsys(ss);
4566 }
4567 return 0;
4568}
4569
4570/**
4571 * cgroup_init - cgroup initialization
4572 *
4573 * Register cgroup filesystem and /proc file, and initialize
4574 * any subsystems that didn't request early init.
4575 */
4576int __init cgroup_init(void)
4577{
4578 int err;
4579 int i;
4580 struct hlist_head *hhead;
4581
4582 err = bdi_init(&cgroup_backing_dev_info);
4583 if (err)
4584 return err;
4585
4586 /* at bootup time, we don't worry about modular subsystems */
4587 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4588 struct cgroup_subsys *ss = subsys[i];
4589 if (!ss->early_init)
4590 cgroup_init_subsys(ss);
4591 if (ss->use_id)
4592 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
4593 }
4594
4595 /* Add init_css_set to the hash table */
4596 hhead = css_set_hash(init_css_set.subsys);
4597 hlist_add_head(&init_css_set.hlist, hhead);
4598 BUG_ON(!init_root_id(&rootnode));
4599
4600 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4601 if (!cgroup_kobj) {
4602 err = -ENOMEM;
4603 goto out;
4604 }
4605
4606 err = register_filesystem(&cgroup_fs_type);
4607 if (err < 0) {
4608 kobject_put(cgroup_kobj);
4609 goto out;
4610 }
4611
4612 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4613
4614out:
4615 if (err)
4616 bdi_destroy(&cgroup_backing_dev_info);
4617
4618 return err;
4619}
4620
4621/*
4622 * proc_cgroup_show()
4623 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4624 * - Used for /proc/<pid>/cgroup.
4625 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4626 * doesn't really matter if tsk->cgroup changes after we read it,
4627 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4628 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4629 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4630 * cgroup to top_cgroup.
4631 */
4632
4633/* TODO: Use a proper seq_file iterator */
4634static int proc_cgroup_show(struct seq_file *m, void *v)
4635{
4636 struct pid *pid;
4637 struct task_struct *tsk;
4638 char *buf;
4639 int retval;
4640 struct cgroupfs_root *root;
4641
4642 retval = -ENOMEM;
4643 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4644 if (!buf)
4645 goto out;
4646
4647 retval = -ESRCH;
4648 pid = m->private;
4649 tsk = get_pid_task(pid, PIDTYPE_PID);
4650 if (!tsk)
4651 goto out_free;
4652
4653 retval = 0;
4654
4655 mutex_lock(&cgroup_mutex);
4656
4657 for_each_active_root(root) {
4658 struct cgroup_subsys *ss;
4659 struct cgroup *cgrp;
4660 int count = 0;
4661
4662 seq_printf(m, "%d:", root->hierarchy_id);
4663 for_each_subsys(root, ss)
4664 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4665 if (strlen(root->name))
4666 seq_printf(m, "%sname=%s", count ? "," : "",
4667 root->name);
4668 seq_putc(m, ':');
4669 cgrp = task_cgroup_from_root(tsk, root);
4670 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4671 if (retval < 0)
4672 goto out_unlock;
4673 seq_puts(m, buf);
4674 seq_putc(m, '\n');
4675 }
4676
4677out_unlock:
4678 mutex_unlock(&cgroup_mutex);
4679 put_task_struct(tsk);
4680out_free:
4681 kfree(buf);
4682out:
4683 return retval;
4684}
4685
4686static int cgroup_open(struct inode *inode, struct file *file)
4687{
4688 struct pid *pid = PROC_I(inode)->pid;
4689 return single_open(file, proc_cgroup_show, pid);
4690}
4691
4692const struct file_operations proc_cgroup_operations = {
4693 .open = cgroup_open,
4694 .read = seq_read,
4695 .llseek = seq_lseek,
4696 .release = single_release,
4697};
4698
4699/* Display information about each subsystem and each hierarchy */
4700static int proc_cgroupstats_show(struct seq_file *m, void *v)
4701{
4702 int i;
4703
4704 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4705 /*
4706 * ideally we don't want subsystems moving around while we do this.
4707 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4708 * subsys/hierarchy state.
4709 */
4710 mutex_lock(&cgroup_mutex);
4711 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4712 struct cgroup_subsys *ss = subsys[i];
4713 if (ss == NULL)
4714 continue;
4715 seq_printf(m, "%s\t%d\t%d\t%d\n",
4716 ss->name, ss->root->hierarchy_id,
4717 ss->root->number_of_cgroups, !ss->disabled);
4718 }
4719 mutex_unlock(&cgroup_mutex);
4720 return 0;
4721}
4722
4723static int cgroupstats_open(struct inode *inode, struct file *file)
4724{
4725 return single_open(file, proc_cgroupstats_show, NULL);
4726}
4727
4728static const struct file_operations proc_cgroupstats_operations = {
4729 .open = cgroupstats_open,
4730 .read = seq_read,
4731 .llseek = seq_lseek,
4732 .release = single_release,
4733};
4734
4735/**
4736 * cgroup_fork - attach newly forked task to its parents cgroup.
4737 * @child: pointer to task_struct of forking parent process.
4738 *
4739 * Description: A task inherits its parent's cgroup at fork().
4740 *
4741 * A pointer to the shared css_set was automatically copied in
4742 * fork.c by dup_task_struct(). However, we ignore that copy, since
4743 * it was not made under the protection of RCU, cgroup_mutex or
4744 * threadgroup_change_begin(), so it might no longer be a valid
4745 * cgroup pointer. cgroup_attach_task() might have already changed
4746 * current->cgroups, allowing the previously referenced cgroup
4747 * group to be removed and freed.
4748 *
4749 * Outside the pointer validity we also need to process the css_set
4750 * inheritance between threadgoup_change_begin() and
4751 * threadgoup_change_end(), this way there is no leak in any process
4752 * wide migration performed by cgroup_attach_proc() that could otherwise
4753 * miss a thread because it is too early or too late in the fork stage.
4754 *
4755 * At the point that cgroup_fork() is called, 'current' is the parent
4756 * task, and the passed argument 'child' points to the child task.
4757 */
4758void cgroup_fork(struct task_struct *child)
4759{
4760 /*
4761 * We don't need to task_lock() current because current->cgroups
4762 * can't be changed concurrently here. The parent obviously hasn't
4763 * exited and called cgroup_exit(), and we are synchronized against
4764 * cgroup migration through threadgroup_change_begin().
4765 */
4766 child->cgroups = current->cgroups;
4767 get_css_set(child->cgroups);
4768 INIT_LIST_HEAD(&child->cg_list);
4769}
4770
4771/**
4772 * cgroup_fork_callbacks - run fork callbacks
4773 * @child: the new task
4774 *
4775 * Called on a new task very soon before adding it to the
4776 * tasklist. No need to take any locks since no-one can
4777 * be operating on this task.
4778 */
4779void cgroup_fork_callbacks(struct task_struct *child)
4780{
4781 if (need_forkexit_callback) {
4782 int i;
4783 /*
4784 * forkexit callbacks are only supported for builtin
4785 * subsystems, and the builtin section of the subsys array is
4786 * immutable, so we don't need to lock the subsys array here.
4787 */
4788 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4789 struct cgroup_subsys *ss = subsys[i];
4790 if (ss->fork)
4791 ss->fork(child);
4792 }
4793 }
4794}
4795
4796/**
4797 * cgroup_post_fork - called on a new task after adding it to the task list
4798 * @child: the task in question
4799 *
4800 * Adds the task to the list running through its css_set if necessary.
4801 * Has to be after the task is visible on the task list in case we race
4802 * with the first call to cgroup_iter_start() - to guarantee that the
4803 * new task ends up on its list.
4804 */
4805void cgroup_post_fork(struct task_struct *child)
4806{
4807 /*
4808 * use_task_css_set_links is set to 1 before we walk the tasklist
4809 * under the tasklist_lock and we read it here after we added the child
4810 * to the tasklist under the tasklist_lock as well. If the child wasn't
4811 * yet in the tasklist when we walked through it from
4812 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4813 * should be visible now due to the paired locking and barriers implied
4814 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4815 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4816 * lock on fork.
4817 */
4818 if (use_task_css_set_links) {
4819 write_lock(&css_set_lock);
4820 if (list_empty(&child->cg_list)) {
4821 /*
4822 * It's safe to use child->cgroups without task_lock()
4823 * here because we are protected through
4824 * threadgroup_change_begin() against concurrent
4825 * css_set change in cgroup_task_migrate(). Also
4826 * the task can't exit at that point until
4827 * wake_up_new_task() is called, so we are protected
4828 * against cgroup_exit() setting child->cgroup to
4829 * init_css_set.
4830 */
4831 list_add(&child->cg_list, &child->cgroups->tasks);
4832 }
4833 write_unlock(&css_set_lock);
4834 }
4835}
4836/**
4837 * cgroup_exit - detach cgroup from exiting task
4838 * @tsk: pointer to task_struct of exiting process
4839 * @run_callback: run exit callbacks?
4840 *
4841 * Description: Detach cgroup from @tsk and release it.
4842 *
4843 * Note that cgroups marked notify_on_release force every task in
4844 * them to take the global cgroup_mutex mutex when exiting.
4845 * This could impact scaling on very large systems. Be reluctant to
4846 * use notify_on_release cgroups where very high task exit scaling
4847 * is required on large systems.
4848 *
4849 * the_top_cgroup_hack:
4850 *
4851 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4852 *
4853 * We call cgroup_exit() while the task is still competent to
4854 * handle notify_on_release(), then leave the task attached to the
4855 * root cgroup in each hierarchy for the remainder of its exit.
4856 *
4857 * To do this properly, we would increment the reference count on
4858 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4859 * code we would add a second cgroup function call, to drop that
4860 * reference. This would just create an unnecessary hot spot on
4861 * the top_cgroup reference count, to no avail.
4862 *
4863 * Normally, holding a reference to a cgroup without bumping its
4864 * count is unsafe. The cgroup could go away, or someone could
4865 * attach us to a different cgroup, decrementing the count on
4866 * the first cgroup that we never incremented. But in this case,
4867 * top_cgroup isn't going away, and either task has PF_EXITING set,
4868 * which wards off any cgroup_attach_task() attempts, or task is a failed
4869 * fork, never visible to cgroup_attach_task.
4870 */
4871void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4872{
4873 struct css_set *cg;
4874 int i;
4875
4876 /*
4877 * Unlink from the css_set task list if necessary.
4878 * Optimistically check cg_list before taking
4879 * css_set_lock
4880 */
4881 if (!list_empty(&tsk->cg_list)) {
4882 write_lock(&css_set_lock);
4883 if (!list_empty(&tsk->cg_list))
4884 list_del_init(&tsk->cg_list);
4885 write_unlock(&css_set_lock);
4886 }
4887
4888 /* Reassign the task to the init_css_set. */
4889 task_lock(tsk);
4890 cg = tsk->cgroups;
4891 tsk->cgroups = &init_css_set;
4892
4893 if (run_callbacks && need_forkexit_callback) {
4894 /*
4895 * modular subsystems can't use callbacks, so no need to lock
4896 * the subsys array
4897 */
4898 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4899 struct cgroup_subsys *ss = subsys[i];
4900 if (ss->exit) {
4901 struct cgroup *old_cgrp =
4902 rcu_dereference_raw(cg->subsys[i])->cgroup;
4903 struct cgroup *cgrp = task_cgroup(tsk, i);
4904 ss->exit(cgrp, old_cgrp, tsk);
4905 }
4906 }
4907 }
4908 task_unlock(tsk);
4909
4910 if (cg)
4911 put_css_set_taskexit(cg);
4912}
4913
4914/**
4915 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4916 * @cgrp: the cgroup in question
4917 * @task: the task in question
4918 *
4919 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4920 * hierarchy.
4921 *
4922 * If we are sending in dummytop, then presumably we are creating
4923 * the top cgroup in the subsystem.
4924 *
4925 * Called only by the ns (nsproxy) cgroup.
4926 */
4927int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4928{
4929 int ret;
4930 struct cgroup *target;
4931
4932 if (cgrp == dummytop)
4933 return 1;
4934
4935 target = task_cgroup_from_root(task, cgrp->root);
4936 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4937 cgrp = cgrp->parent;
4938 ret = (cgrp == target);
4939 return ret;
4940}
4941
4942static void check_for_release(struct cgroup *cgrp)
4943{
4944 /* All of these checks rely on RCU to keep the cgroup
4945 * structure alive */
4946 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4947 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4948 /* Control Group is currently removeable. If it's not
4949 * already queued for a userspace notification, queue
4950 * it now */
4951 int need_schedule_work = 0;
4952 raw_spin_lock(&release_list_lock);
4953 if (!cgroup_is_removed(cgrp) &&
4954 list_empty(&cgrp->release_list)) {
4955 list_add(&cgrp->release_list, &release_list);
4956 need_schedule_work = 1;
4957 }
4958 raw_spin_unlock(&release_list_lock);
4959 if (need_schedule_work)
4960 schedule_work(&release_agent_work);
4961 }
4962}
4963
4964/* Caller must verify that the css is not for root cgroup */
4965bool __css_tryget(struct cgroup_subsys_state *css)
4966{
4967 do {
4968 int v = css_refcnt(css);
4969
4970 if (atomic_cmpxchg(&css->refcnt, v, v + 1) == v)
4971 return true;
4972 cpu_relax();
4973 } while (!test_bit(CSS_REMOVED, &css->flags));
4974
4975 return false;
4976}
4977EXPORT_SYMBOL_GPL(__css_tryget);
4978
4979/* Caller must verify that the css is not for root cgroup */
4980void __css_put(struct cgroup_subsys_state *css)
4981{
4982 struct cgroup *cgrp = css->cgroup;
4983 int v;
4984
4985 rcu_read_lock();
4986 v = css_unbias_refcnt(atomic_dec_return(&css->refcnt));
4987
4988 switch (v) {
4989 case 1:
4990 if (notify_on_release(cgrp)) {
4991 set_bit(CGRP_RELEASABLE, &cgrp->flags);
4992 check_for_release(cgrp);
4993 }
4994 cgroup_wakeup_rmdir_waiter(cgrp);
4995 break;
4996 case 0:
4997 if (!test_bit(CSS_CLEAR_CSS_REFS, &css->flags))
4998 schedule_work(&css->dput_work);
4999 break;
5000 }
5001 rcu_read_unlock();
5002}
5003EXPORT_SYMBOL_GPL(__css_put);
5004
5005/*
5006 * Notify userspace when a cgroup is released, by running the
5007 * configured release agent with the name of the cgroup (path
5008 * relative to the root of cgroup file system) as the argument.
5009 *
5010 * Most likely, this user command will try to rmdir this cgroup.
5011 *
5012 * This races with the possibility that some other task will be
5013 * attached to this cgroup before it is removed, or that some other
5014 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5015 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5016 * unused, and this cgroup will be reprieved from its death sentence,
5017 * to continue to serve a useful existence. Next time it's released,
5018 * we will get notified again, if it still has 'notify_on_release' set.
5019 *
5020 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5021 * means only wait until the task is successfully execve()'d. The
5022 * separate release agent task is forked by call_usermodehelper(),
5023 * then control in this thread returns here, without waiting for the
5024 * release agent task. We don't bother to wait because the caller of
5025 * this routine has no use for the exit status of the release agent
5026 * task, so no sense holding our caller up for that.
5027 */
5028static void cgroup_release_agent(struct work_struct *work)
5029{
5030 BUG_ON(work != &release_agent_work);
5031 mutex_lock(&cgroup_mutex);
5032 raw_spin_lock(&release_list_lock);
5033 while (!list_empty(&release_list)) {
5034 char *argv[3], *envp[3];
5035 int i;
5036 char *pathbuf = NULL, *agentbuf = NULL;
5037 struct cgroup *cgrp = list_entry(release_list.next,
5038 struct cgroup,
5039 release_list);
5040 list_del_init(&cgrp->release_list);
5041 raw_spin_unlock(&release_list_lock);
5042 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5043 if (!pathbuf)
5044 goto continue_free;
5045 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
5046 goto continue_free;
5047 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5048 if (!agentbuf)
5049 goto continue_free;
5050
5051 i = 0;
5052 argv[i++] = agentbuf;
5053 argv[i++] = pathbuf;
5054 argv[i] = NULL;
5055
5056 i = 0;
5057 /* minimal command environment */
5058 envp[i++] = "HOME=/";
5059 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5060 envp[i] = NULL;
5061
5062 /* Drop the lock while we invoke the usermode helper,
5063 * since the exec could involve hitting disk and hence
5064 * be a slow process */
5065 mutex_unlock(&cgroup_mutex);
5066 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5067 mutex_lock(&cgroup_mutex);
5068 continue_free:
5069 kfree(pathbuf);
5070 kfree(agentbuf);
5071 raw_spin_lock(&release_list_lock);
5072 }
5073 raw_spin_unlock(&release_list_lock);
5074 mutex_unlock(&cgroup_mutex);
5075}
5076
5077static int __init cgroup_disable(char *str)
5078{
5079 int i;
5080 char *token;
5081
5082 while ((token = strsep(&str, ",")) != NULL) {
5083 if (!*token)
5084 continue;
5085 /*
5086 * cgroup_disable, being at boot time, can't know about module
5087 * subsystems, so we don't worry about them.
5088 */
5089 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
5090 struct cgroup_subsys *ss = subsys[i];
5091
5092 if (!strcmp(token, ss->name)) {
5093 ss->disabled = 1;
5094 printk(KERN_INFO "Disabling %s control group"
5095 " subsystem\n", ss->name);
5096 break;
5097 }
5098 }
5099 }
5100 return 1;
5101}
5102__setup("cgroup_disable=", cgroup_disable);
5103
5104/*
5105 * Functons for CSS ID.
5106 */
5107
5108/*
5109 *To get ID other than 0, this should be called when !cgroup_is_removed().
5110 */
5111unsigned short css_id(struct cgroup_subsys_state *css)
5112{
5113 struct css_id *cssid;
5114
5115 /*
5116 * This css_id() can return correct value when somone has refcnt
5117 * on this or this is under rcu_read_lock(). Once css->id is allocated,
5118 * it's unchanged until freed.
5119 */
5120 cssid = rcu_dereference_check(css->id, css_refcnt(css));
5121
5122 if (cssid)
5123 return cssid->id;
5124 return 0;
5125}
5126EXPORT_SYMBOL_GPL(css_id);
5127
5128unsigned short css_depth(struct cgroup_subsys_state *css)
5129{
5130 struct css_id *cssid;
5131
5132 cssid = rcu_dereference_check(css->id, css_refcnt(css));
5133
5134 if (cssid)
5135 return cssid->depth;
5136 return 0;
5137}
5138EXPORT_SYMBOL_GPL(css_depth);
5139
5140/**
5141 * css_is_ancestor - test "root" css is an ancestor of "child"
5142 * @child: the css to be tested.
5143 * @root: the css supporsed to be an ancestor of the child.
5144 *
5145 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
5146 * this function reads css->id, the caller must hold rcu_read_lock().
5147 * But, considering usual usage, the csses should be valid objects after test.
5148 * Assuming that the caller will do some action to the child if this returns
5149 * returns true, the caller must take "child";s reference count.
5150 * If "child" is valid object and this returns true, "root" is valid, too.
5151 */
5152
5153bool css_is_ancestor(struct cgroup_subsys_state *child,
5154 const struct cgroup_subsys_state *root)
5155{
5156 struct css_id *child_id;
5157 struct css_id *root_id;
5158
5159 child_id = rcu_dereference(child->id);
5160 if (!child_id)
5161 return false;
5162 root_id = rcu_dereference(root->id);
5163 if (!root_id)
5164 return false;
5165 if (child_id->depth < root_id->depth)
5166 return false;
5167 if (child_id->stack[root_id->depth] != root_id->id)
5168 return false;
5169 return true;
5170}
5171
5172void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
5173{
5174 struct css_id *id = css->id;
5175 /* When this is called before css_id initialization, id can be NULL */
5176 if (!id)
5177 return;
5178
5179 BUG_ON(!ss->use_id);
5180
5181 rcu_assign_pointer(id->css, NULL);
5182 rcu_assign_pointer(css->id, NULL);
5183 spin_lock(&ss->id_lock);
5184 idr_remove(&ss->idr, id->id);
5185 spin_unlock(&ss->id_lock);
5186 kfree_rcu(id, rcu_head);
5187}
5188EXPORT_SYMBOL_GPL(free_css_id);
5189
5190/*
5191 * This is called by init or create(). Then, calls to this function are
5192 * always serialized (By cgroup_mutex() at create()).
5193 */
5194
5195static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
5196{
5197 struct css_id *newid;
5198 int myid, error, size;
5199
5200 BUG_ON(!ss->use_id);
5201
5202 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
5203 newid = kzalloc(size, GFP_KERNEL);
5204 if (!newid)
5205 return ERR_PTR(-ENOMEM);
5206 /* get id */
5207 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
5208 error = -ENOMEM;
5209 goto err_out;
5210 }
5211 spin_lock(&ss->id_lock);
5212 /* Don't use 0. allocates an ID of 1-65535 */
5213 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
5214 spin_unlock(&ss->id_lock);
5215
5216 /* Returns error when there are no free spaces for new ID.*/
5217 if (error) {
5218 error = -ENOSPC;
5219 goto err_out;
5220 }
5221 if (myid > CSS_ID_MAX)
5222 goto remove_idr;
5223
5224 newid->id = myid;
5225 newid->depth = depth;
5226 return newid;
5227remove_idr:
5228 error = -ENOSPC;
5229 spin_lock(&ss->id_lock);
5230 idr_remove(&ss->idr, myid);
5231 spin_unlock(&ss->id_lock);
5232err_out:
5233 kfree(newid);
5234 return ERR_PTR(error);
5235
5236}
5237
5238static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
5239 struct cgroup_subsys_state *rootcss)
5240{
5241 struct css_id *newid;
5242
5243 spin_lock_init(&ss->id_lock);
5244 idr_init(&ss->idr);
5245
5246 newid = get_new_cssid(ss, 0);
5247 if (IS_ERR(newid))
5248 return PTR_ERR(newid);
5249
5250 newid->stack[0] = newid->id;
5251 newid->css = rootcss;
5252 rootcss->id = newid;
5253 return 0;
5254}
5255
5256static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
5257 struct cgroup *child)
5258{
5259 int subsys_id, i, depth = 0;
5260 struct cgroup_subsys_state *parent_css, *child_css;
5261 struct css_id *child_id, *parent_id;
5262
5263 subsys_id = ss->subsys_id;
5264 parent_css = parent->subsys[subsys_id];
5265 child_css = child->subsys[subsys_id];
5266 parent_id = parent_css->id;
5267 depth = parent_id->depth + 1;
5268
5269 child_id = get_new_cssid(ss, depth);
5270 if (IS_ERR(child_id))
5271 return PTR_ERR(child_id);
5272
5273 for (i = 0; i < depth; i++)
5274 child_id->stack[i] = parent_id->stack[i];
5275 child_id->stack[depth] = child_id->id;
5276 /*
5277 * child_id->css pointer will be set after this cgroup is available
5278 * see cgroup_populate_dir()
5279 */
5280 rcu_assign_pointer(child_css->id, child_id);
5281
5282 return 0;
5283}
5284
5285/**
5286 * css_lookup - lookup css by id
5287 * @ss: cgroup subsys to be looked into.
5288 * @id: the id
5289 *
5290 * Returns pointer to cgroup_subsys_state if there is valid one with id.
5291 * NULL if not. Should be called under rcu_read_lock()
5292 */
5293struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
5294{
5295 struct css_id *cssid = NULL;
5296
5297 BUG_ON(!ss->use_id);
5298 cssid = idr_find(&ss->idr, id);
5299
5300 if (unlikely(!cssid))
5301 return NULL;
5302
5303 return rcu_dereference(cssid->css);
5304}
5305EXPORT_SYMBOL_GPL(css_lookup);
5306
5307/**
5308 * css_get_next - lookup next cgroup under specified hierarchy.
5309 * @ss: pointer to subsystem
5310 * @id: current position of iteration.
5311 * @root: pointer to css. search tree under this.
5312 * @foundid: position of found object.
5313 *
5314 * Search next css under the specified hierarchy of rootid. Calling under
5315 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
5316 */
5317struct cgroup_subsys_state *
5318css_get_next(struct cgroup_subsys *ss, int id,
5319 struct cgroup_subsys_state *root, int *foundid)
5320{
5321 struct cgroup_subsys_state *ret = NULL;
5322 struct css_id *tmp;
5323 int tmpid;
5324 int rootid = css_id(root);
5325 int depth = css_depth(root);
5326
5327 if (!rootid)
5328 return NULL;
5329
5330 BUG_ON(!ss->use_id);
5331 WARN_ON_ONCE(!rcu_read_lock_held());
5332
5333 /* fill start point for scan */
5334 tmpid = id;
5335 while (1) {
5336 /*
5337 * scan next entry from bitmap(tree), tmpid is updated after
5338 * idr_get_next().
5339 */
5340 tmp = idr_get_next(&ss->idr, &tmpid);
5341 if (!tmp)
5342 break;
5343 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
5344 ret = rcu_dereference(tmp->css);
5345 if (ret) {
5346 *foundid = tmpid;
5347 break;
5348 }
5349 }
5350 /* continue to scan from next id */
5351 tmpid = tmpid + 1;
5352 }
5353 return ret;
5354}
5355
5356/*
5357 * get corresponding css from file open on cgroupfs directory
5358 */
5359struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
5360{
5361 struct cgroup *cgrp;
5362 struct inode *inode;
5363 struct cgroup_subsys_state *css;
5364
5365 inode = f->f_dentry->d_inode;
5366 /* check in cgroup filesystem dir */
5367 if (inode->i_op != &cgroup_dir_inode_operations)
5368 return ERR_PTR(-EBADF);
5369
5370 if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
5371 return ERR_PTR(-EINVAL);
5372
5373 /* get cgroup */
5374 cgrp = __d_cgrp(f->f_dentry);
5375 css = cgrp->subsys[id];
5376 return css ? css : ERR_PTR(-ENOENT);
5377}
5378
5379#ifdef CONFIG_CGROUP_DEBUG
5380static struct cgroup_subsys_state *debug_create(struct cgroup *cont)
5381{
5382 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5383
5384 if (!css)
5385 return ERR_PTR(-ENOMEM);
5386
5387 return css;
5388}
5389
5390static void debug_destroy(struct cgroup *cont)
5391{
5392 kfree(cont->subsys[debug_subsys_id]);
5393}
5394
5395static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
5396{
5397 return atomic_read(&cont->count);
5398}
5399
5400static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
5401{
5402 return cgroup_task_count(cont);
5403}
5404
5405static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
5406{
5407 return (u64)(unsigned long)current->cgroups;
5408}
5409
5410static u64 current_css_set_refcount_read(struct cgroup *cont,
5411 struct cftype *cft)
5412{
5413 u64 count;
5414
5415 rcu_read_lock();
5416 count = atomic_read(¤t->cgroups->refcount);
5417 rcu_read_unlock();
5418 return count;
5419}
5420
5421static int current_css_set_cg_links_read(struct cgroup *cont,
5422 struct cftype *cft,
5423 struct seq_file *seq)
5424{
5425 struct cg_cgroup_link *link;
5426 struct css_set *cg;
5427
5428 read_lock(&css_set_lock);
5429 rcu_read_lock();
5430 cg = rcu_dereference(current->cgroups);
5431 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
5432 struct cgroup *c = link->cgrp;
5433 const char *name;
5434
5435 if (c->dentry)
5436 name = c->dentry->d_name.name;
5437 else
5438 name = "?";
5439 seq_printf(seq, "Root %d group %s\n",
5440 c->root->hierarchy_id, name);
5441 }
5442 rcu_read_unlock();
5443 read_unlock(&css_set_lock);
5444 return 0;
5445}
5446
5447#define MAX_TASKS_SHOWN_PER_CSS 25
5448static int cgroup_css_links_read(struct cgroup *cont,
5449 struct cftype *cft,
5450 struct seq_file *seq)
5451{
5452 struct cg_cgroup_link *link;
5453
5454 read_lock(&css_set_lock);
5455 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
5456 struct css_set *cg = link->cg;
5457 struct task_struct *task;
5458 int count = 0;
5459 seq_printf(seq, "css_set %p\n", cg);
5460 list_for_each_entry(task, &cg->tasks, cg_list) {
5461 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5462 seq_puts(seq, " ...\n");
5463 break;
5464 } else {
5465 seq_printf(seq, " task %d\n",
5466 task_pid_vnr(task));
5467 }
5468 }
5469 }
5470 read_unlock(&css_set_lock);
5471 return 0;
5472}
5473
5474static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
5475{
5476 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
5477}
5478
5479static struct cftype debug_files[] = {
5480 {
5481 .name = "cgroup_refcount",
5482 .read_u64 = cgroup_refcount_read,
5483 },
5484 {
5485 .name = "taskcount",
5486 .read_u64 = debug_taskcount_read,
5487 },
5488
5489 {
5490 .name = "current_css_set",
5491 .read_u64 = current_css_set_read,
5492 },
5493
5494 {
5495 .name = "current_css_set_refcount",
5496 .read_u64 = current_css_set_refcount_read,
5497 },
5498
5499 {
5500 .name = "current_css_set_cg_links",
5501 .read_seq_string = current_css_set_cg_links_read,
5502 },
5503
5504 {
5505 .name = "cgroup_css_links",
5506 .read_seq_string = cgroup_css_links_read,
5507 },
5508
5509 {
5510 .name = "releasable",
5511 .read_u64 = releasable_read,
5512 },
5513
5514 { } /* terminate */
5515};
5516
5517struct cgroup_subsys debug_subsys = {
5518 .name = "debug",
5519 .create = debug_create,
5520 .destroy = debug_destroy,
5521 .subsys_id = debug_subsys_id,
5522 .base_cftypes = debug_files,
5523};
5524#endif /* CONFIG_CGROUP_DEBUG */