projects/linux·blob v4.6-rc5 · refs

6466 lines · 172.4 KiB

1/*
2 * Generic process-grouping system.
3 *
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
6 *
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
10 *
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
15 *
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
18 *
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
23 *
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
27 */
28
29#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
31#include <linux/cgroup.h>
32#include <linux/cred.h>
33#include <linux/ctype.h>
34#include <linux/errno.h>
35#include <linux/init_task.h>
36#include <linux/kernel.h>
37#include <linux/list.h>
38#include <linux/magic.h>
39#include <linux/mm.h>
40#include <linux/mutex.h>
41#include <linux/mount.h>
42#include <linux/pagemap.h>
43#include <linux/proc_fs.h>
44#include <linux/rcupdate.h>
45#include <linux/sched.h>
46#include <linux/slab.h>
47#include <linux/spinlock.h>
48#include <linux/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/proc_ns.h>
65#include <net/sock.h>
66
67/*
68 * pidlists linger the following amount before being destroyed. The goal
69 * is avoiding frequent destruction in the middle of consecutive read calls
70 * Expiring in the middle is a performance problem not a correctness one.
71 * 1 sec should be enough.
72 */
73#define CGROUP_PIDLIST_DESTROY_DELAY HZ
74
75#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 MAX_CFTYPE_NAME + 2)
77
78/*
79 * cgroup_mutex is the master lock. Any modification to cgroup or its
80 * hierarchy must be performed while holding it.
81 *
82 * css_set_lock protects task->cgroups pointer, the list of css_set
83 * objects, and the chain of tasks off each css_set.
84 *
85 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
86 * cgroup.h can use them for lockdep annotations.
87 */
88#ifdef CONFIG_PROVE_RCU
89DEFINE_MUTEX(cgroup_mutex);
90DEFINE_SPINLOCK(css_set_lock);
91EXPORT_SYMBOL_GPL(cgroup_mutex);
92EXPORT_SYMBOL_GPL(css_set_lock);
93#else
94static DEFINE_MUTEX(cgroup_mutex);
95static DEFINE_SPINLOCK(css_set_lock);
96#endif
97
98/*
99 * Protects cgroup_idr and css_idr so that IDs can be released without
100 * grabbing cgroup_mutex.
101 */
102static DEFINE_SPINLOCK(cgroup_idr_lock);
103
104/*
105 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
106 * against file removal/re-creation across css hiding.
107 */
108static DEFINE_SPINLOCK(cgroup_file_kn_lock);
109
110/*
111 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
112 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
113 */
114static DEFINE_SPINLOCK(release_agent_path_lock);
115
116struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
117
118#define cgroup_assert_mutex_or_rcu_locked() \
119 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
120 !lockdep_is_held(&cgroup_mutex), \
121 "cgroup_mutex or RCU read lock required");
122
123/*
124 * cgroup destruction makes heavy use of work items and there can be a lot
125 * of concurrent destructions. Use a separate workqueue so that cgroup
126 * destruction work items don't end up filling up max_active of system_wq
127 * which may lead to deadlock.
128 */
129static struct workqueue_struct *cgroup_destroy_wq;
130
131/*
132 * pidlist destructions need to be flushed on cgroup destruction. Use a
133 * separate workqueue as flush domain.
134 */
135static struct workqueue_struct *cgroup_pidlist_destroy_wq;
136
137/* generate an array of cgroup subsystem pointers */
138#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
139static struct cgroup_subsys *cgroup_subsys[] = {
140#include <linux/cgroup_subsys.h>
141};
142#undef SUBSYS
143
144/* array of cgroup subsystem names */
145#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
146static const char *cgroup_subsys_name[] = {
147#include <linux/cgroup_subsys.h>
148};
149#undef SUBSYS
150
151/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
152#define SUBSYS(_x) \
153 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
154 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
155 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
156 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
157#include <linux/cgroup_subsys.h>
158#undef SUBSYS
159
160#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
161static struct static_key_true *cgroup_subsys_enabled_key[] = {
162#include <linux/cgroup_subsys.h>
163};
164#undef SUBSYS
165
166#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
167static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
168#include <linux/cgroup_subsys.h>
169};
170#undef SUBSYS
171
172/*
173 * The default hierarchy, reserved for the subsystems that are otherwise
174 * unattached - it never has more than a single cgroup, and all tasks are
175 * part of that cgroup.
176 */
177struct cgroup_root cgrp_dfl_root;
178EXPORT_SYMBOL_GPL(cgrp_dfl_root);
179
180/*
181 * The default hierarchy always exists but is hidden until mounted for the
182 * first time. This is for backward compatibility.
183 */
184static bool cgrp_dfl_visible;
185
186/* Controllers blocked by the commandline in v1 */
187static u16 cgroup_no_v1_mask;
188
189/* some controllers are not supported in the default hierarchy */
190static u16 cgrp_dfl_inhibit_ss_mask;
191
192/* some controllers are implicitly enabled on the default hierarchy */
193static unsigned long cgrp_dfl_implicit_ss_mask;
194
195/* The list of hierarchy roots */
196
197static LIST_HEAD(cgroup_roots);
198static int cgroup_root_count;
199
200/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
201static DEFINE_IDR(cgroup_hierarchy_idr);
202
203/*
204 * Assign a monotonically increasing serial number to csses. It guarantees
205 * cgroups with bigger numbers are newer than those with smaller numbers.
206 * Also, as csses are always appended to the parent's ->children list, it
207 * guarantees that sibling csses are always sorted in the ascending serial
208 * number order on the list. Protected by cgroup_mutex.
209 */
210static u64 css_serial_nr_next = 1;
211
212/*
213 * These bitmask flags indicate whether tasks in the fork and exit paths have
214 * fork/exit handlers to call. This avoids us having to do extra work in the
215 * fork/exit path to check which subsystems have fork/exit callbacks.
216 */
217static u16 have_fork_callback __read_mostly;
218static u16 have_exit_callback __read_mostly;
219static u16 have_free_callback __read_mostly;
220
221/* cgroup namespace for init task */
222struct cgroup_namespace init_cgroup_ns = {
223 .count = { .counter = 2, },
224 .user_ns = &init_user_ns,
225 .ns.ops = &cgroupns_operations,
226 .ns.inum = PROC_CGROUP_INIT_INO,
227 .root_cset = &init_css_set,
228};
229
230/* Ditto for the can_fork callback. */
231static u16 have_canfork_callback __read_mostly;
232
233static struct file_system_type cgroup2_fs_type;
234static struct cftype cgroup_dfl_base_files[];
235static struct cftype cgroup_legacy_base_files[];
236
237static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
238static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
239static int cgroup_apply_control(struct cgroup *cgrp);
240static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
241static void css_task_iter_advance(struct css_task_iter *it);
242static int cgroup_destroy_locked(struct cgroup *cgrp);
243static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
244 struct cgroup_subsys *ss);
245static void css_release(struct percpu_ref *ref);
246static void kill_css(struct cgroup_subsys_state *css);
247static int cgroup_addrm_files(struct cgroup_subsys_state *css,
248 struct cgroup *cgrp, struct cftype cfts[],
249 bool is_add);
250
251/**
252 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
253 * @ssid: subsys ID of interest
254 *
255 * cgroup_subsys_enabled() can only be used with literal subsys names which
256 * is fine for individual subsystems but unsuitable for cgroup core. This
257 * is slower static_key_enabled() based test indexed by @ssid.
258 */
259static bool cgroup_ssid_enabled(int ssid)
260{
261 if (CGROUP_SUBSYS_COUNT == 0)
262 return false;
263
264 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
265}
266
267static bool cgroup_ssid_no_v1(int ssid)
268{
269 return cgroup_no_v1_mask & (1 << ssid);
270}
271
272/**
273 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
274 * @cgrp: the cgroup of interest
275 *
276 * The default hierarchy is the v2 interface of cgroup and this function
277 * can be used to test whether a cgroup is on the default hierarchy for
278 * cases where a subsystem should behave differnetly depending on the
279 * interface version.
280 *
281 * The set of behaviors which change on the default hierarchy are still
282 * being determined and the mount option is prefixed with __DEVEL__.
283 *
284 * List of changed behaviors:
285 *
286 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
287 * and "name" are disallowed.
288 *
289 * - When mounting an existing superblock, mount options should match.
290 *
291 * - Remount is disallowed.
292 *
293 * - rename(2) is disallowed.
294 *
295 * - "tasks" is removed. Everything should be at process granularity. Use
296 * "cgroup.procs" instead.
297 *
298 * - "cgroup.procs" is not sorted. pids will be unique unless they got
299 * recycled inbetween reads.
300 *
301 * - "release_agent" and "notify_on_release" are removed. Replacement
302 * notification mechanism will be implemented.
303 *
304 * - "cgroup.clone_children" is removed.
305 *
306 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
307 * and its descendants contain no task; otherwise, 1. The file also
308 * generates kernfs notification which can be monitored through poll and
309 * [di]notify when the value of the file changes.
310 *
311 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
312 * take masks of ancestors with non-empty cpus/mems, instead of being
313 * moved to an ancestor.
314 *
315 * - cpuset: a task can be moved into an empty cpuset, and again it takes
316 * masks of ancestors.
317 *
318 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
319 * is not created.
320 *
321 * - blkcg: blk-throttle becomes properly hierarchical.
322 *
323 * - debug: disallowed on the default hierarchy.
324 */
325static bool cgroup_on_dfl(const struct cgroup *cgrp)
326{
327 return cgrp->root == &cgrp_dfl_root;
328}
329
330/* IDR wrappers which synchronize using cgroup_idr_lock */
331static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
332 gfp_t gfp_mask)
333{
334 int ret;
335
336 idr_preload(gfp_mask);
337 spin_lock_bh(&cgroup_idr_lock);
338 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
339 spin_unlock_bh(&cgroup_idr_lock);
340 idr_preload_end();
341 return ret;
342}
343
344static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
345{
346 void *ret;
347
348 spin_lock_bh(&cgroup_idr_lock);
349 ret = idr_replace(idr, ptr, id);
350 spin_unlock_bh(&cgroup_idr_lock);
351 return ret;
352}
353
354static void cgroup_idr_remove(struct idr *idr, int id)
355{
356 spin_lock_bh(&cgroup_idr_lock);
357 idr_remove(idr, id);
358 spin_unlock_bh(&cgroup_idr_lock);
359}
360
361static struct cgroup *cgroup_parent(struct cgroup *cgrp)
362{
363 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
364
365 if (parent_css)
366 return container_of(parent_css, struct cgroup, self);
367 return NULL;
368}
369
370/* subsystems visibly enabled on a cgroup */
371static u16 cgroup_control(struct cgroup *cgrp)
372{
373 struct cgroup *parent = cgroup_parent(cgrp);
374 u16 root_ss_mask = cgrp->root->subsys_mask;
375
376 if (parent)
377 return parent->subtree_control;
378
379 if (cgroup_on_dfl(cgrp))
380 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
381 cgrp_dfl_implicit_ss_mask);
382 return root_ss_mask;
383}
384
385/* subsystems enabled on a cgroup */
386static u16 cgroup_ss_mask(struct cgroup *cgrp)
387{
388 struct cgroup *parent = cgroup_parent(cgrp);
389
390 if (parent)
391 return parent->subtree_ss_mask;
392
393 return cgrp->root->subsys_mask;
394}
395
396/**
397 * cgroup_css - obtain a cgroup's css for the specified subsystem
398 * @cgrp: the cgroup of interest
399 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
400 *
401 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
402 * function must be called either under cgroup_mutex or rcu_read_lock() and
403 * the caller is responsible for pinning the returned css if it wants to
404 * keep accessing it outside the said locks. This function may return
405 * %NULL if @cgrp doesn't have @subsys_id enabled.
406 */
407static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
408 struct cgroup_subsys *ss)
409{
410 if (ss)
411 return rcu_dereference_check(cgrp->subsys[ss->id],
412 lockdep_is_held(&cgroup_mutex));
413 else
414 return &cgrp->self;
415}
416
417/**
418 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
419 * @cgrp: the cgroup of interest
420 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
421 *
422 * Similar to cgroup_css() but returns the effective css, which is defined
423 * as the matching css of the nearest ancestor including self which has @ss
424 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
425 * function is guaranteed to return non-NULL css.
426 */
427static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
428 struct cgroup_subsys *ss)
429{
430 lockdep_assert_held(&cgroup_mutex);
431
432 if (!ss)
433 return &cgrp->self;
434
435 /*
436 * This function is used while updating css associations and thus
437 * can't test the csses directly. Test ss_mask.
438 */
439 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
440 cgrp = cgroup_parent(cgrp);
441 if (!cgrp)
442 return NULL;
443 }
444
445 return cgroup_css(cgrp, ss);
446}
447
448/**
449 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
450 * @cgrp: the cgroup of interest
451 * @ss: the subsystem of interest
452 *
453 * Find and get the effective css of @cgrp for @ss. The effective css is
454 * defined as the matching css of the nearest ancestor including self which
455 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
456 * the root css is returned, so this function always returns a valid css.
457 * The returned css must be put using css_put().
458 */
459struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
460 struct cgroup_subsys *ss)
461{
462 struct cgroup_subsys_state *css;
463
464 rcu_read_lock();
465
466 do {
467 css = cgroup_css(cgrp, ss);
468
469 if (css && css_tryget_online(css))
470 goto out_unlock;
471 cgrp = cgroup_parent(cgrp);
472 } while (cgrp);
473
474 css = init_css_set.subsys[ss->id];
475 css_get(css);
476out_unlock:
477 rcu_read_unlock();
478 return css;
479}
480
481/* convenient tests for these bits */
482static inline bool cgroup_is_dead(const struct cgroup *cgrp)
483{
484 return !(cgrp->self.flags & CSS_ONLINE);
485}
486
487static void cgroup_get(struct cgroup *cgrp)
488{
489 WARN_ON_ONCE(cgroup_is_dead(cgrp));
490 css_get(&cgrp->self);
491}
492
493static bool cgroup_tryget(struct cgroup *cgrp)
494{
495 return css_tryget(&cgrp->self);
496}
497
498struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
499{
500 struct cgroup *cgrp = of->kn->parent->priv;
501 struct cftype *cft = of_cft(of);
502
503 /*
504 * This is open and unprotected implementation of cgroup_css().
505 * seq_css() is only called from a kernfs file operation which has
506 * an active reference on the file. Because all the subsystem
507 * files are drained before a css is disassociated with a cgroup,
508 * the matching css from the cgroup's subsys table is guaranteed to
509 * be and stay valid until the enclosing operation is complete.
510 */
511 if (cft->ss)
512 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
513 else
514 return &cgrp->self;
515}
516EXPORT_SYMBOL_GPL(of_css);
517
518static int notify_on_release(const struct cgroup *cgrp)
519{
520 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
521}
522
523/**
524 * for_each_css - iterate all css's of a cgroup
525 * @css: the iteration cursor
526 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
527 * @cgrp: the target cgroup to iterate css's of
528 *
529 * Should be called under cgroup_[tree_]mutex.
530 */
531#define for_each_css(css, ssid, cgrp) \
532 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
533 if (!((css) = rcu_dereference_check( \
534 (cgrp)->subsys[(ssid)], \
535 lockdep_is_held(&cgroup_mutex)))) { } \
536 else
537
538/**
539 * for_each_e_css - iterate all effective css's of a cgroup
540 * @css: the iteration cursor
541 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
542 * @cgrp: the target cgroup to iterate css's of
543 *
544 * Should be called under cgroup_[tree_]mutex.
545 */
546#define for_each_e_css(css, ssid, cgrp) \
547 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
548 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
549 ; \
550 else
551
552/**
553 * for_each_subsys - iterate all enabled cgroup subsystems
554 * @ss: the iteration cursor
555 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
556 */
557#define for_each_subsys(ss, ssid) \
558 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
559 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
560
561/**
562 * do_each_subsys_mask - filter for_each_subsys with a bitmask
563 * @ss: the iteration cursor
564 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
565 * @ss_mask: the bitmask
566 *
567 * The block will only run for cases where the ssid-th bit (1 << ssid) of
568 * @ss_mask is set.
569 */
570#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
571 unsigned long __ss_mask = (ss_mask); \
572 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
573 (ssid) = 0; \
574 break; \
575 } \
576 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
577 (ss) = cgroup_subsys[ssid]; \
578 {
579
580#define while_each_subsys_mask() \
581 } \
582 } \
583} while (false)
584
585/* iterate across the hierarchies */
586#define for_each_root(root) \
587 list_for_each_entry((root), &cgroup_roots, root_list)
588
589/* iterate over child cgrps, lock should be held throughout iteration */
590#define cgroup_for_each_live_child(child, cgrp) \
591 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
592 if (({ lockdep_assert_held(&cgroup_mutex); \
593 cgroup_is_dead(child); })) \
594 ; \
595 else
596
597/* walk live descendants in preorder */
598#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
599 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
600 if (({ lockdep_assert_held(&cgroup_mutex); \
601 (dsct) = (d_css)->cgroup; \
602 cgroup_is_dead(dsct); })) \
603 ; \
604 else
605
606/* walk live descendants in postorder */
607#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
608 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
609 if (({ lockdep_assert_held(&cgroup_mutex); \
610 (dsct) = (d_css)->cgroup; \
611 cgroup_is_dead(dsct); })) \
612 ; \
613 else
614
615static void cgroup_release_agent(struct work_struct *work);
616static void check_for_release(struct cgroup *cgrp);
617
618/*
619 * A cgroup can be associated with multiple css_sets as different tasks may
620 * belong to different cgroups on different hierarchies. In the other
621 * direction, a css_set is naturally associated with multiple cgroups.
622 * This M:N relationship is represented by the following link structure
623 * which exists for each association and allows traversing the associations
624 * from both sides.
625 */
626struct cgrp_cset_link {
627 /* the cgroup and css_set this link associates */
628 struct cgroup *cgrp;
629 struct css_set *cset;
630
631 /* list of cgrp_cset_links anchored at cgrp->cset_links */
632 struct list_head cset_link;
633
634 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
635 struct list_head cgrp_link;
636};
637
638/*
639 * The default css_set - used by init and its children prior to any
640 * hierarchies being mounted. It contains a pointer to the root state
641 * for each subsystem. Also used to anchor the list of css_sets. Not
642 * reference-counted, to improve performance when child cgroups
643 * haven't been created.
644 */
645struct css_set init_css_set = {
646 .refcount = ATOMIC_INIT(1),
647 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
648 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
649 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
650 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
651 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
652 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
653};
654
655static int css_set_count = 1; /* 1 for init_css_set */
656
657/**
658 * css_set_populated - does a css_set contain any tasks?
659 * @cset: target css_set
660 */
661static bool css_set_populated(struct css_set *cset)
662{
663 lockdep_assert_held(&css_set_lock);
664
665 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
666}
667
668/**
669 * cgroup_update_populated - updated populated count of a cgroup
670 * @cgrp: the target cgroup
671 * @populated: inc or dec populated count
672 *
673 * One of the css_sets associated with @cgrp is either getting its first
674 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
675 * count is propagated towards root so that a given cgroup's populated_cnt
676 * is zero iff the cgroup and all its descendants don't contain any tasks.
677 *
678 * @cgrp's interface file "cgroup.populated" is zero if
679 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
680 * changes from or to zero, userland is notified that the content of the
681 * interface file has changed. This can be used to detect when @cgrp and
682 * its descendants become populated or empty.
683 */
684static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
685{
686 lockdep_assert_held(&css_set_lock);
687
688 do {
689 bool trigger;
690
691 if (populated)
692 trigger = !cgrp->populated_cnt++;
693 else
694 trigger = !--cgrp->populated_cnt;
695
696 if (!trigger)
697 break;
698
699 check_for_release(cgrp);
700 cgroup_file_notify(&cgrp->events_file);
701
702 cgrp = cgroup_parent(cgrp);
703 } while (cgrp);
704}
705
706/**
707 * css_set_update_populated - update populated state of a css_set
708 * @cset: target css_set
709 * @populated: whether @cset is populated or depopulated
710 *
711 * @cset is either getting the first task or losing the last. Update the
712 * ->populated_cnt of all associated cgroups accordingly.
713 */
714static void css_set_update_populated(struct css_set *cset, bool populated)
715{
716 struct cgrp_cset_link *link;
717
718 lockdep_assert_held(&css_set_lock);
719
720 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
721 cgroup_update_populated(link->cgrp, populated);
722}
723
724/**
725 * css_set_move_task - move a task from one css_set to another
726 * @task: task being moved
727 * @from_cset: css_set @task currently belongs to (may be NULL)
728 * @to_cset: new css_set @task is being moved to (may be NULL)
729 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
730 *
731 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
732 * css_set, @from_cset can be NULL. If @task is being disassociated
733 * instead of moved, @to_cset can be NULL.
734 *
735 * This function automatically handles populated_cnt updates and
736 * css_task_iter adjustments but the caller is responsible for managing
737 * @from_cset and @to_cset's reference counts.
738 */
739static void css_set_move_task(struct task_struct *task,
740 struct css_set *from_cset, struct css_set *to_cset,
741 bool use_mg_tasks)
742{
743 lockdep_assert_held(&css_set_lock);
744
745 if (to_cset && !css_set_populated(to_cset))
746 css_set_update_populated(to_cset, true);
747
748 if (from_cset) {
749 struct css_task_iter *it, *pos;
750
751 WARN_ON_ONCE(list_empty(&task->cg_list));
752
753 /*
754 * @task is leaving, advance task iterators which are
755 * pointing to it so that they can resume at the next
756 * position. Advancing an iterator might remove it from
757 * the list, use safe walk. See css_task_iter_advance*()
758 * for details.
759 */
760 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
761 iters_node)
762 if (it->task_pos == &task->cg_list)
763 css_task_iter_advance(it);
764
765 list_del_init(&task->cg_list);
766 if (!css_set_populated(from_cset))
767 css_set_update_populated(from_cset, false);
768 } else {
769 WARN_ON_ONCE(!list_empty(&task->cg_list));
770 }
771
772 if (to_cset) {
773 /*
774 * We are synchronized through cgroup_threadgroup_rwsem
775 * against PF_EXITING setting such that we can't race
776 * against cgroup_exit() changing the css_set to
777 * init_css_set and dropping the old one.
778 */
779 WARN_ON_ONCE(task->flags & PF_EXITING);
780
781 rcu_assign_pointer(task->cgroups, to_cset);
782 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
783 &to_cset->tasks);
784 }
785}
786
787/*
788 * hash table for cgroup groups. This improves the performance to find
789 * an existing css_set. This hash doesn't (currently) take into
790 * account cgroups in empty hierarchies.
791 */
792#define CSS_SET_HASH_BITS 7
793static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
794
795static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
796{
797 unsigned long key = 0UL;
798 struct cgroup_subsys *ss;
799 int i;
800
801 for_each_subsys(ss, i)
802 key += (unsigned long)css[i];
803 key = (key >> 16) ^ key;
804
805 return key;
806}
807
808static void put_css_set_locked(struct css_set *cset)
809{
810 struct cgrp_cset_link *link, *tmp_link;
811 struct cgroup_subsys *ss;
812 int ssid;
813
814 lockdep_assert_held(&css_set_lock);
815
816 if (!atomic_dec_and_test(&cset->refcount))
817 return;
818
819 /* This css_set is dead. unlink it and release cgroup and css refs */
820 for_each_subsys(ss, ssid) {
821 list_del(&cset->e_cset_node[ssid]);
822 css_put(cset->subsys[ssid]);
823 }
824 hash_del(&cset->hlist);
825 css_set_count--;
826
827 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
828 list_del(&link->cset_link);
829 list_del(&link->cgrp_link);
830 if (cgroup_parent(link->cgrp))
831 cgroup_put(link->cgrp);
832 kfree(link);
833 }
834
835 kfree_rcu(cset, rcu_head);
836}
837
838static void put_css_set(struct css_set *cset)
839{
840 /*
841 * Ensure that the refcount doesn't hit zero while any readers
842 * can see it. Similar to atomic_dec_and_lock(), but for an
843 * rwlock
844 */
845 if (atomic_add_unless(&cset->refcount, -1, 1))
846 return;
847
848 spin_lock_bh(&css_set_lock);
849 put_css_set_locked(cset);
850 spin_unlock_bh(&css_set_lock);
851}
852
853/*
854 * refcounted get/put for css_set objects
855 */
856static inline void get_css_set(struct css_set *cset)
857{
858 atomic_inc(&cset->refcount);
859}
860
861/**
862 * compare_css_sets - helper function for find_existing_css_set().
863 * @cset: candidate css_set being tested
864 * @old_cset: existing css_set for a task
865 * @new_cgrp: cgroup that's being entered by the task
866 * @template: desired set of css pointers in css_set (pre-calculated)
867 *
868 * Returns true if "cset" matches "old_cset" except for the hierarchy
869 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
870 */
871static bool compare_css_sets(struct css_set *cset,
872 struct css_set *old_cset,
873 struct cgroup *new_cgrp,
874 struct cgroup_subsys_state *template[])
875{
876 struct list_head *l1, *l2;
877
878 /*
879 * On the default hierarchy, there can be csets which are
880 * associated with the same set of cgroups but different csses.
881 * Let's first ensure that csses match.
882 */
883 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
884 return false;
885
886 /*
887 * Compare cgroup pointers in order to distinguish between
888 * different cgroups in hierarchies. As different cgroups may
889 * share the same effective css, this comparison is always
890 * necessary.
891 */
892 l1 = &cset->cgrp_links;
893 l2 = &old_cset->cgrp_links;
894 while (1) {
895 struct cgrp_cset_link *link1, *link2;
896 struct cgroup *cgrp1, *cgrp2;
897
898 l1 = l1->next;
899 l2 = l2->next;
900 /* See if we reached the end - both lists are equal length. */
901 if (l1 == &cset->cgrp_links) {
902 BUG_ON(l2 != &old_cset->cgrp_links);
903 break;
904 } else {
905 BUG_ON(l2 == &old_cset->cgrp_links);
906 }
907 /* Locate the cgroups associated with these links. */
908 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
909 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
910 cgrp1 = link1->cgrp;
911 cgrp2 = link2->cgrp;
912 /* Hierarchies should be linked in the same order. */
913 BUG_ON(cgrp1->root != cgrp2->root);
914
915 /*
916 * If this hierarchy is the hierarchy of the cgroup
917 * that's changing, then we need to check that this
918 * css_set points to the new cgroup; if it's any other
919 * hierarchy, then this css_set should point to the
920 * same cgroup as the old css_set.
921 */
922 if (cgrp1->root == new_cgrp->root) {
923 if (cgrp1 != new_cgrp)
924 return false;
925 } else {
926 if (cgrp1 != cgrp2)
927 return false;
928 }
929 }
930 return true;
931}
932
933/**
934 * find_existing_css_set - init css array and find the matching css_set
935 * @old_cset: the css_set that we're using before the cgroup transition
936 * @cgrp: the cgroup that we're moving into
937 * @template: out param for the new set of csses, should be clear on entry
938 */
939static struct css_set *find_existing_css_set(struct css_set *old_cset,
940 struct cgroup *cgrp,
941 struct cgroup_subsys_state *template[])
942{
943 struct cgroup_root *root = cgrp->root;
944 struct cgroup_subsys *ss;
945 struct css_set *cset;
946 unsigned long key;
947 int i;
948
949 /*
950 * Build the set of subsystem state objects that we want to see in the
951 * new css_set. while subsystems can change globally, the entries here
952 * won't change, so no need for locking.
953 */
954 for_each_subsys(ss, i) {
955 if (root->subsys_mask & (1UL << i)) {
956 /*
957 * @ss is in this hierarchy, so we want the
958 * effective css from @cgrp.
959 */
960 template[i] = cgroup_e_css(cgrp, ss);
961 } else {
962 /*
963 * @ss is not in this hierarchy, so we don't want
964 * to change the css.
965 */
966 template[i] = old_cset->subsys[i];
967 }
968 }
969
970 key = css_set_hash(template);
971 hash_for_each_possible(css_set_table, cset, hlist, key) {
972 if (!compare_css_sets(cset, old_cset, cgrp, template))
973 continue;
974
975 /* This css_set matches what we need */
976 return cset;
977 }
978
979 /* No existing cgroup group matched */
980 return NULL;
981}
982
983static void free_cgrp_cset_links(struct list_head *links_to_free)
984{
985 struct cgrp_cset_link *link, *tmp_link;
986
987 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
988 list_del(&link->cset_link);
989 kfree(link);
990 }
991}
992
993/**
994 * allocate_cgrp_cset_links - allocate cgrp_cset_links
995 * @count: the number of links to allocate
996 * @tmp_links: list_head the allocated links are put on
997 *
998 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
999 * through ->cset_link. Returns 0 on success or -errno.
1000 */
1001static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1002{
1003 struct cgrp_cset_link *link;
1004 int i;
1005
1006 INIT_LIST_HEAD(tmp_links);
1007
1008 for (i = 0; i < count; i++) {
1009 link = kzalloc(sizeof(*link), GFP_KERNEL);
1010 if (!link) {
1011 free_cgrp_cset_links(tmp_links);
1012 return -ENOMEM;
1013 }
1014 list_add(&link->cset_link, tmp_links);
1015 }
1016 return 0;
1017}
1018
1019/**
1020 * link_css_set - a helper function to link a css_set to a cgroup
1021 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1022 * @cset: the css_set to be linked
1023 * @cgrp: the destination cgroup
1024 */
1025static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1026 struct cgroup *cgrp)
1027{
1028 struct cgrp_cset_link *link;
1029
1030 BUG_ON(list_empty(tmp_links));
1031
1032 if (cgroup_on_dfl(cgrp))
1033 cset->dfl_cgrp = cgrp;
1034
1035 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1036 link->cset = cset;
1037 link->cgrp = cgrp;
1038
1039 /*
1040 * Always add links to the tail of the lists so that the lists are
1041 * in choronological order.
1042 */
1043 list_move_tail(&link->cset_link, &cgrp->cset_links);
1044 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1045
1046 if (cgroup_parent(cgrp))
1047 cgroup_get(cgrp);
1048}
1049
1050/**
1051 * find_css_set - return a new css_set with one cgroup updated
1052 * @old_cset: the baseline css_set
1053 * @cgrp: the cgroup to be updated
1054 *
1055 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1056 * substituted into the appropriate hierarchy.
1057 */
1058static struct css_set *find_css_set(struct css_set *old_cset,
1059 struct cgroup *cgrp)
1060{
1061 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1062 struct css_set *cset;
1063 struct list_head tmp_links;
1064 struct cgrp_cset_link *link;
1065 struct cgroup_subsys *ss;
1066 unsigned long key;
1067 int ssid;
1068
1069 lockdep_assert_held(&cgroup_mutex);
1070
1071 /* First see if we already have a cgroup group that matches
1072 * the desired set */
1073 spin_lock_bh(&css_set_lock);
1074 cset = find_existing_css_set(old_cset, cgrp, template);
1075 if (cset)
1076 get_css_set(cset);
1077 spin_unlock_bh(&css_set_lock);
1078
1079 if (cset)
1080 return cset;
1081
1082 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1083 if (!cset)
1084 return NULL;
1085
1086 /* Allocate all the cgrp_cset_link objects that we'll need */
1087 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1088 kfree(cset);
1089 return NULL;
1090 }
1091
1092 atomic_set(&cset->refcount, 1);
1093 INIT_LIST_HEAD(&cset->cgrp_links);
1094 INIT_LIST_HEAD(&cset->tasks);
1095 INIT_LIST_HEAD(&cset->mg_tasks);
1096 INIT_LIST_HEAD(&cset->mg_preload_node);
1097 INIT_LIST_HEAD(&cset->mg_node);
1098 INIT_LIST_HEAD(&cset->task_iters);
1099 INIT_HLIST_NODE(&cset->hlist);
1100
1101 /* Copy the set of subsystem state objects generated in
1102 * find_existing_css_set() */
1103 memcpy(cset->subsys, template, sizeof(cset->subsys));
1104
1105 spin_lock_bh(&css_set_lock);
1106 /* Add reference counts and links from the new css_set. */
1107 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1108 struct cgroup *c = link->cgrp;
1109
1110 if (c->root == cgrp->root)
1111 c = cgrp;
1112 link_css_set(&tmp_links, cset, c);
1113 }
1114
1115 BUG_ON(!list_empty(&tmp_links));
1116
1117 css_set_count++;
1118
1119 /* Add @cset to the hash table */
1120 key = css_set_hash(cset->subsys);
1121 hash_add(css_set_table, &cset->hlist, key);
1122
1123 for_each_subsys(ss, ssid) {
1124 struct cgroup_subsys_state *css = cset->subsys[ssid];
1125
1126 list_add_tail(&cset->e_cset_node[ssid],
1127 &css->cgroup->e_csets[ssid]);
1128 css_get(css);
1129 }
1130
1131 spin_unlock_bh(&css_set_lock);
1132
1133 return cset;
1134}
1135
1136static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1137{
1138 struct cgroup *root_cgrp = kf_root->kn->priv;
1139
1140 return root_cgrp->root;
1141}
1142
1143static int cgroup_init_root_id(struct cgroup_root *root)
1144{
1145 int id;
1146
1147 lockdep_assert_held(&cgroup_mutex);
1148
1149 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1150 if (id < 0)
1151 return id;
1152
1153 root->hierarchy_id = id;
1154 return 0;
1155}
1156
1157static void cgroup_exit_root_id(struct cgroup_root *root)
1158{
1159 lockdep_assert_held(&cgroup_mutex);
1160
1161 if (root->hierarchy_id) {
1162 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1163 root->hierarchy_id = 0;
1164 }
1165}
1166
1167static void cgroup_free_root(struct cgroup_root *root)
1168{
1169 if (root) {
1170 /* hierarchy ID should already have been released */
1171 WARN_ON_ONCE(root->hierarchy_id);
1172
1173 idr_destroy(&root->cgroup_idr);
1174 kfree(root);
1175 }
1176}
1177
1178static void cgroup_destroy_root(struct cgroup_root *root)
1179{
1180 struct cgroup *cgrp = &root->cgrp;
1181 struct cgrp_cset_link *link, *tmp_link;
1182
1183 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1184
1185 BUG_ON(atomic_read(&root->nr_cgrps));
1186 BUG_ON(!list_empty(&cgrp->self.children));
1187
1188 /* Rebind all subsystems back to the default hierarchy */
1189 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1190
1191 /*
1192 * Release all the links from cset_links to this hierarchy's
1193 * root cgroup
1194 */
1195 spin_lock_bh(&css_set_lock);
1196
1197 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1198 list_del(&link->cset_link);
1199 list_del(&link->cgrp_link);
1200 kfree(link);
1201 }
1202
1203 spin_unlock_bh(&css_set_lock);
1204
1205 if (!list_empty(&root->root_list)) {
1206 list_del(&root->root_list);
1207 cgroup_root_count--;
1208 }
1209
1210 cgroup_exit_root_id(root);
1211
1212 mutex_unlock(&cgroup_mutex);
1213
1214 kernfs_destroy_root(root->kf_root);
1215 cgroup_free_root(root);
1216}
1217
1218/* look up cgroup associated with given css_set on the specified hierarchy */
1219static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1220 struct cgroup_root *root)
1221{
1222 struct cgroup *res = NULL;
1223
1224 lockdep_assert_held(&cgroup_mutex);
1225 lockdep_assert_held(&css_set_lock);
1226
1227 if (cset == &init_css_set) {
1228 res = &root->cgrp;
1229 } else {
1230 struct cgrp_cset_link *link;
1231
1232 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1233 struct cgroup *c = link->cgrp;
1234
1235 if (c->root == root) {
1236 res = c;
1237 break;
1238 }
1239 }
1240 }
1241
1242 BUG_ON(!res);
1243 return res;
1244}
1245
1246/*
1247 * Return the cgroup for "task" from the given hierarchy. Must be
1248 * called with cgroup_mutex and css_set_lock held.
1249 */
1250static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1251 struct cgroup_root *root)
1252{
1253 /*
1254 * No need to lock the task - since we hold cgroup_mutex the
1255 * task can't change groups, so the only thing that can happen
1256 * is that it exits and its css is set back to init_css_set.
1257 */
1258 return cset_cgroup_from_root(task_css_set(task), root);
1259}
1260
1261/*
1262 * A task must hold cgroup_mutex to modify cgroups.
1263 *
1264 * Any task can increment and decrement the count field without lock.
1265 * So in general, code holding cgroup_mutex can't rely on the count
1266 * field not changing. However, if the count goes to zero, then only
1267 * cgroup_attach_task() can increment it again. Because a count of zero
1268 * means that no tasks are currently attached, therefore there is no
1269 * way a task attached to that cgroup can fork (the other way to
1270 * increment the count). So code holding cgroup_mutex can safely
1271 * assume that if the count is zero, it will stay zero. Similarly, if
1272 * a task holds cgroup_mutex on a cgroup with zero count, it
1273 * knows that the cgroup won't be removed, as cgroup_rmdir()
1274 * needs that mutex.
1275 *
1276 * A cgroup can only be deleted if both its 'count' of using tasks
1277 * is zero, and its list of 'children' cgroups is empty. Since all
1278 * tasks in the system use _some_ cgroup, and since there is always at
1279 * least one task in the system (init, pid == 1), therefore, root cgroup
1280 * always has either children cgroups and/or using tasks. So we don't
1281 * need a special hack to ensure that root cgroup cannot be deleted.
1282 *
1283 * P.S. One more locking exception. RCU is used to guard the
1284 * update of a tasks cgroup pointer by cgroup_attach_task()
1285 */
1286
1287static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1288static const struct file_operations proc_cgroupstats_operations;
1289
1290static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1291 char *buf)
1292{
1293 struct cgroup_subsys *ss = cft->ss;
1294
1295 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1296 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1297 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1298 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1299 cft->name);
1300 else
1301 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1302 return buf;
1303}
1304
1305/**
1306 * cgroup_file_mode - deduce file mode of a control file
1307 * @cft: the control file in question
1308 *
1309 * S_IRUGO for read, S_IWUSR for write.
1310 */
1311static umode_t cgroup_file_mode(const struct cftype *cft)
1312{
1313 umode_t mode = 0;
1314
1315 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1316 mode |= S_IRUGO;
1317
1318 if (cft->write_u64 || cft->write_s64 || cft->write) {
1319 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1320 mode |= S_IWUGO;
1321 else
1322 mode |= S_IWUSR;
1323 }
1324
1325 return mode;
1326}
1327
1328/**
1329 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1330 * @subtree_control: the new subtree_control mask to consider
1331 * @this_ss_mask: available subsystems
1332 *
1333 * On the default hierarchy, a subsystem may request other subsystems to be
1334 * enabled together through its ->depends_on mask. In such cases, more
1335 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1336 *
1337 * This function calculates which subsystems need to be enabled if
1338 * @subtree_control is to be applied while restricted to @this_ss_mask.
1339 */
1340static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1341{
1342 u16 cur_ss_mask = subtree_control;
1343 struct cgroup_subsys *ss;
1344 int ssid;
1345
1346 lockdep_assert_held(&cgroup_mutex);
1347
1348 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1349
1350 while (true) {
1351 u16 new_ss_mask = cur_ss_mask;
1352
1353 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1354 new_ss_mask |= ss->depends_on;
1355 } while_each_subsys_mask();
1356
1357 /*
1358 * Mask out subsystems which aren't available. This can
1359 * happen only if some depended-upon subsystems were bound
1360 * to non-default hierarchies.
1361 */
1362 new_ss_mask &= this_ss_mask;
1363
1364 if (new_ss_mask == cur_ss_mask)
1365 break;
1366 cur_ss_mask = new_ss_mask;
1367 }
1368
1369 return cur_ss_mask;
1370}
1371
1372/**
1373 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1374 * @kn: the kernfs_node being serviced
1375 *
1376 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1377 * the method finishes if locking succeeded. Note that once this function
1378 * returns the cgroup returned by cgroup_kn_lock_live() may become
1379 * inaccessible any time. If the caller intends to continue to access the
1380 * cgroup, it should pin it before invoking this function.
1381 */
1382static void cgroup_kn_unlock(struct kernfs_node *kn)
1383{
1384 struct cgroup *cgrp;
1385
1386 if (kernfs_type(kn) == KERNFS_DIR)
1387 cgrp = kn->priv;
1388 else
1389 cgrp = kn->parent->priv;
1390
1391 mutex_unlock(&cgroup_mutex);
1392
1393 kernfs_unbreak_active_protection(kn);
1394 cgroup_put(cgrp);
1395}
1396
1397/**
1398 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1399 * @kn: the kernfs_node being serviced
1400 * @drain_offline: perform offline draining on the cgroup
1401 *
1402 * This helper is to be used by a cgroup kernfs method currently servicing
1403 * @kn. It breaks the active protection, performs cgroup locking and
1404 * verifies that the associated cgroup is alive. Returns the cgroup if
1405 * alive; otherwise, %NULL. A successful return should be undone by a
1406 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1407 * cgroup is drained of offlining csses before return.
1408 *
1409 * Any cgroup kernfs method implementation which requires locking the
1410 * associated cgroup should use this helper. It avoids nesting cgroup
1411 * locking under kernfs active protection and allows all kernfs operations
1412 * including self-removal.
1413 */
1414static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1415 bool drain_offline)
1416{
1417 struct cgroup *cgrp;
1418
1419 if (kernfs_type(kn) == KERNFS_DIR)
1420 cgrp = kn->priv;
1421 else
1422 cgrp = kn->parent->priv;
1423
1424 /*
1425 * We're gonna grab cgroup_mutex which nests outside kernfs
1426 * active_ref. cgroup liveliness check alone provides enough
1427 * protection against removal. Ensure @cgrp stays accessible and
1428 * break the active_ref protection.
1429 */
1430 if (!cgroup_tryget(cgrp))
1431 return NULL;
1432 kernfs_break_active_protection(kn);
1433
1434 if (drain_offline)
1435 cgroup_lock_and_drain_offline(cgrp);
1436 else
1437 mutex_lock(&cgroup_mutex);
1438
1439 if (!cgroup_is_dead(cgrp))
1440 return cgrp;
1441
1442 cgroup_kn_unlock(kn);
1443 return NULL;
1444}
1445
1446static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1447{
1448 char name[CGROUP_FILE_NAME_MAX];
1449
1450 lockdep_assert_held(&cgroup_mutex);
1451
1452 if (cft->file_offset) {
1453 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1454 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1455
1456 spin_lock_irq(&cgroup_file_kn_lock);
1457 cfile->kn = NULL;
1458 spin_unlock_irq(&cgroup_file_kn_lock);
1459 }
1460
1461 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1462}
1463
1464/**
1465 * css_clear_dir - remove subsys files in a cgroup directory
1466 * @css: taget css
1467 */
1468static void css_clear_dir(struct cgroup_subsys_state *css)
1469{
1470 struct cgroup *cgrp = css->cgroup;
1471 struct cftype *cfts;
1472
1473 if (!(css->flags & CSS_VISIBLE))
1474 return;
1475
1476 css->flags &= ~CSS_VISIBLE;
1477
1478 list_for_each_entry(cfts, &css->ss->cfts, node)
1479 cgroup_addrm_files(css, cgrp, cfts, false);
1480}
1481
1482/**
1483 * css_populate_dir - create subsys files in a cgroup directory
1484 * @css: target css
1485 *
1486 * On failure, no file is added.
1487 */
1488static int css_populate_dir(struct cgroup_subsys_state *css)
1489{
1490 struct cgroup *cgrp = css->cgroup;
1491 struct cftype *cfts, *failed_cfts;
1492 int ret;
1493
1494 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1495 return 0;
1496
1497 if (!css->ss) {
1498 if (cgroup_on_dfl(cgrp))
1499 cfts = cgroup_dfl_base_files;
1500 else
1501 cfts = cgroup_legacy_base_files;
1502
1503 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1504 }
1505
1506 list_for_each_entry(cfts, &css->ss->cfts, node) {
1507 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1508 if (ret < 0) {
1509 failed_cfts = cfts;
1510 goto err;
1511 }
1512 }
1513
1514 css->flags |= CSS_VISIBLE;
1515
1516 return 0;
1517err:
1518 list_for_each_entry(cfts, &css->ss->cfts, node) {
1519 if (cfts == failed_cfts)
1520 break;
1521 cgroup_addrm_files(css, cgrp, cfts, false);
1522 }
1523 return ret;
1524}
1525
1526static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1527{
1528 struct cgroup *dcgrp = &dst_root->cgrp;
1529 struct cgroup_subsys *ss;
1530 int ssid, i, ret;
1531
1532 lockdep_assert_held(&cgroup_mutex);
1533
1534 do_each_subsys_mask(ss, ssid, ss_mask) {
1535 /*
1536 * If @ss has non-root csses attached to it, can't move.
1537 * If @ss is an implicit controller, it is exempt from this
1538 * rule and can be stolen.
1539 */
1540 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1541 !ss->implicit_on_dfl)
1542 return -EBUSY;
1543
1544 /* can't move between two non-dummy roots either */
1545 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1546 return -EBUSY;
1547 } while_each_subsys_mask();
1548
1549 do_each_subsys_mask(ss, ssid, ss_mask) {
1550 struct cgroup_root *src_root = ss->root;
1551 struct cgroup *scgrp = &src_root->cgrp;
1552 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1553 struct css_set *cset;
1554
1555 WARN_ON(!css || cgroup_css(dcgrp, ss));
1556
1557 /* disable from the source */
1558 src_root->subsys_mask &= ~(1 << ssid);
1559 WARN_ON(cgroup_apply_control(scgrp));
1560 cgroup_finalize_control(scgrp, 0);
1561
1562 /* rebind */
1563 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1564 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1565 ss->root = dst_root;
1566 css->cgroup = dcgrp;
1567
1568 spin_lock_bh(&css_set_lock);
1569 hash_for_each(css_set_table, i, cset, hlist)
1570 list_move_tail(&cset->e_cset_node[ss->id],
1571 &dcgrp->e_csets[ss->id]);
1572 spin_unlock_bh(&css_set_lock);
1573
1574 /* default hierarchy doesn't enable controllers by default */
1575 dst_root->subsys_mask |= 1 << ssid;
1576 if (dst_root == &cgrp_dfl_root) {
1577 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1578 } else {
1579 dcgrp->subtree_control |= 1 << ssid;
1580 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1581 }
1582
1583 ret = cgroup_apply_control(dcgrp);
1584 if (ret)
1585 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1586 ss->name, ret);
1587
1588 if (ss->bind)
1589 ss->bind(css);
1590 } while_each_subsys_mask();
1591
1592 kernfs_activate(dcgrp->kn);
1593 return 0;
1594}
1595
1596static int cgroup_show_options(struct seq_file *seq,
1597 struct kernfs_root *kf_root)
1598{
1599 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1600 struct cgroup_subsys *ss;
1601 int ssid;
1602
1603 if (root != &cgrp_dfl_root)
1604 for_each_subsys(ss, ssid)
1605 if (root->subsys_mask & (1 << ssid))
1606 seq_show_option(seq, ss->legacy_name, NULL);
1607 if (root->flags & CGRP_ROOT_NOPREFIX)
1608 seq_puts(seq, ",noprefix");
1609 if (root->flags & CGRP_ROOT_XATTR)
1610 seq_puts(seq, ",xattr");
1611
1612 spin_lock(&release_agent_path_lock);
1613 if (strlen(root->release_agent_path))
1614 seq_show_option(seq, "release_agent",
1615 root->release_agent_path);
1616 spin_unlock(&release_agent_path_lock);
1617
1618 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1619 seq_puts(seq, ",clone_children");
1620 if (strlen(root->name))
1621 seq_show_option(seq, "name", root->name);
1622 return 0;
1623}
1624
1625struct cgroup_sb_opts {
1626 u16 subsys_mask;
1627 unsigned int flags;
1628 char *release_agent;
1629 bool cpuset_clone_children;
1630 char *name;
1631 /* User explicitly requested empty subsystem */
1632 bool none;
1633};
1634
1635static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1636{
1637 char *token, *o = data;
1638 bool all_ss = false, one_ss = false;
1639 u16 mask = U16_MAX;
1640 struct cgroup_subsys *ss;
1641 int nr_opts = 0;
1642 int i;
1643
1644#ifdef CONFIG_CPUSETS
1645 mask = ~((u16)1 << cpuset_cgrp_id);
1646#endif
1647
1648 memset(opts, 0, sizeof(*opts));
1649
1650 while ((token = strsep(&o, ",")) != NULL) {
1651 nr_opts++;
1652
1653 if (!*token)
1654 return -EINVAL;
1655 if (!strcmp(token, "none")) {
1656 /* Explicitly have no subsystems */
1657 opts->none = true;
1658 continue;
1659 }
1660 if (!strcmp(token, "all")) {
1661 /* Mutually exclusive option 'all' + subsystem name */
1662 if (one_ss)
1663 return -EINVAL;
1664 all_ss = true;
1665 continue;
1666 }
1667 if (!strcmp(token, "noprefix")) {
1668 opts->flags |= CGRP_ROOT_NOPREFIX;
1669 continue;
1670 }
1671 if (!strcmp(token, "clone_children")) {
1672 opts->cpuset_clone_children = true;
1673 continue;
1674 }
1675 if (!strcmp(token, "xattr")) {
1676 opts->flags |= CGRP_ROOT_XATTR;
1677 continue;
1678 }
1679 if (!strncmp(token, "release_agent=", 14)) {
1680 /* Specifying two release agents is forbidden */
1681 if (opts->release_agent)
1682 return -EINVAL;
1683 opts->release_agent =
1684 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1685 if (!opts->release_agent)
1686 return -ENOMEM;
1687 continue;
1688 }
1689 if (!strncmp(token, "name=", 5)) {
1690 const char *name = token + 5;
1691 /* Can't specify an empty name */
1692 if (!strlen(name))
1693 return -EINVAL;
1694 /* Must match [\w.-]+ */
1695 for (i = 0; i < strlen(name); i++) {
1696 char c = name[i];
1697 if (isalnum(c))
1698 continue;
1699 if ((c == '.') || (c == '-') || (c == '_'))
1700 continue;
1701 return -EINVAL;
1702 }
1703 /* Specifying two names is forbidden */
1704 if (opts->name)
1705 return -EINVAL;
1706 opts->name = kstrndup(name,
1707 MAX_CGROUP_ROOT_NAMELEN - 1,
1708 GFP_KERNEL);
1709 if (!opts->name)
1710 return -ENOMEM;
1711
1712 continue;
1713 }
1714
1715 for_each_subsys(ss, i) {
1716 if (strcmp(token, ss->legacy_name))
1717 continue;
1718 if (!cgroup_ssid_enabled(i))
1719 continue;
1720 if (cgroup_ssid_no_v1(i))
1721 continue;
1722
1723 /* Mutually exclusive option 'all' + subsystem name */
1724 if (all_ss)
1725 return -EINVAL;
1726 opts->subsys_mask |= (1 << i);
1727 one_ss = true;
1728
1729 break;
1730 }
1731 if (i == CGROUP_SUBSYS_COUNT)
1732 return -ENOENT;
1733 }
1734
1735 /*
1736 * If the 'all' option was specified select all the subsystems,
1737 * otherwise if 'none', 'name=' and a subsystem name options were
1738 * not specified, let's default to 'all'
1739 */
1740 if (all_ss || (!one_ss && !opts->none && !opts->name))
1741 for_each_subsys(ss, i)
1742 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1743 opts->subsys_mask |= (1 << i);
1744
1745 /*
1746 * We either have to specify by name or by subsystems. (So all
1747 * empty hierarchies must have a name).
1748 */
1749 if (!opts->subsys_mask && !opts->name)
1750 return -EINVAL;
1751
1752 /*
1753 * Option noprefix was introduced just for backward compatibility
1754 * with the old cpuset, so we allow noprefix only if mounting just
1755 * the cpuset subsystem.
1756 */
1757 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1758 return -EINVAL;
1759
1760 /* Can't specify "none" and some subsystems */
1761 if (opts->subsys_mask && opts->none)
1762 return -EINVAL;
1763
1764 return 0;
1765}
1766
1767static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1768{
1769 int ret = 0;
1770 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1771 struct cgroup_sb_opts opts;
1772 u16 added_mask, removed_mask;
1773
1774 if (root == &cgrp_dfl_root) {
1775 pr_err("remount is not allowed\n");
1776 return -EINVAL;
1777 }
1778
1779 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1780
1781 /* See what subsystems are wanted */
1782 ret = parse_cgroupfs_options(data, &opts);
1783 if (ret)
1784 goto out_unlock;
1785
1786 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1787 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1788 task_tgid_nr(current), current->comm);
1789
1790 added_mask = opts.subsys_mask & ~root->subsys_mask;
1791 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1792
1793 /* Don't allow flags or name to change at remount */
1794 if ((opts.flags ^ root->flags) ||
1795 (opts.name && strcmp(opts.name, root->name))) {
1796 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1797 opts.flags, opts.name ?: "", root->flags, root->name);
1798 ret = -EINVAL;
1799 goto out_unlock;
1800 }
1801
1802 /* remounting is not allowed for populated hierarchies */
1803 if (!list_empty(&root->cgrp.self.children)) {
1804 ret = -EBUSY;
1805 goto out_unlock;
1806 }
1807
1808 ret = rebind_subsystems(root, added_mask);
1809 if (ret)
1810 goto out_unlock;
1811
1812 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1813
1814 if (opts.release_agent) {
1815 spin_lock(&release_agent_path_lock);
1816 strcpy(root->release_agent_path, opts.release_agent);
1817 spin_unlock(&release_agent_path_lock);
1818 }
1819 out_unlock:
1820 kfree(opts.release_agent);
1821 kfree(opts.name);
1822 mutex_unlock(&cgroup_mutex);
1823 return ret;
1824}
1825
1826/*
1827 * To reduce the fork() overhead for systems that are not actually using
1828 * their cgroups capability, we don't maintain the lists running through
1829 * each css_set to its tasks until we see the list actually used - in other
1830 * words after the first mount.
1831 */
1832static bool use_task_css_set_links __read_mostly;
1833
1834static void cgroup_enable_task_cg_lists(void)
1835{
1836 struct task_struct *p, *g;
1837
1838 spin_lock_bh(&css_set_lock);
1839
1840 if (use_task_css_set_links)
1841 goto out_unlock;
1842
1843 use_task_css_set_links = true;
1844
1845 /*
1846 * We need tasklist_lock because RCU is not safe against
1847 * while_each_thread(). Besides, a forking task that has passed
1848 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1849 * is not guaranteed to have its child immediately visible in the
1850 * tasklist if we walk through it with RCU.
1851 */
1852 read_lock(&tasklist_lock);
1853 do_each_thread(g, p) {
1854 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1855 task_css_set(p) != &init_css_set);
1856
1857 /*
1858 * We should check if the process is exiting, otherwise
1859 * it will race with cgroup_exit() in that the list
1860 * entry won't be deleted though the process has exited.
1861 * Do it while holding siglock so that we don't end up
1862 * racing against cgroup_exit().
1863 */
1864 spin_lock_irq(&p->sighand->siglock);
1865 if (!(p->flags & PF_EXITING)) {
1866 struct css_set *cset = task_css_set(p);
1867
1868 if (!css_set_populated(cset))
1869 css_set_update_populated(cset, true);
1870 list_add_tail(&p->cg_list, &cset->tasks);
1871 get_css_set(cset);
1872 }
1873 spin_unlock_irq(&p->sighand->siglock);
1874 } while_each_thread(g, p);
1875 read_unlock(&tasklist_lock);
1876out_unlock:
1877 spin_unlock_bh(&css_set_lock);
1878}
1879
1880static void init_cgroup_housekeeping(struct cgroup *cgrp)
1881{
1882 struct cgroup_subsys *ss;
1883 int ssid;
1884
1885 INIT_LIST_HEAD(&cgrp->self.sibling);
1886 INIT_LIST_HEAD(&cgrp->self.children);
1887 INIT_LIST_HEAD(&cgrp->cset_links);
1888 INIT_LIST_HEAD(&cgrp->pidlists);
1889 mutex_init(&cgrp->pidlist_mutex);
1890 cgrp->self.cgroup = cgrp;
1891 cgrp->self.flags |= CSS_ONLINE;
1892
1893 for_each_subsys(ss, ssid)
1894 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1895
1896 init_waitqueue_head(&cgrp->offline_waitq);
1897 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1898}
1899
1900static void init_cgroup_root(struct cgroup_root *root,
1901 struct cgroup_sb_opts *opts)
1902{
1903 struct cgroup *cgrp = &root->cgrp;
1904
1905 INIT_LIST_HEAD(&root->root_list);
1906 atomic_set(&root->nr_cgrps, 1);
1907 cgrp->root = root;
1908 init_cgroup_housekeeping(cgrp);
1909 idr_init(&root->cgroup_idr);
1910
1911 root->flags = opts->flags;
1912 if (opts->release_agent)
1913 strcpy(root->release_agent_path, opts->release_agent);
1914 if (opts->name)
1915 strcpy(root->name, opts->name);
1916 if (opts->cpuset_clone_children)
1917 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1918}
1919
1920static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1921{
1922 LIST_HEAD(tmp_links);
1923 struct cgroup *root_cgrp = &root->cgrp;
1924 struct css_set *cset;
1925 int i, ret;
1926
1927 lockdep_assert_held(&cgroup_mutex);
1928
1929 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1930 if (ret < 0)
1931 goto out;
1932 root_cgrp->id = ret;
1933 root_cgrp->ancestor_ids[0] = ret;
1934
1935 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1936 GFP_KERNEL);
1937 if (ret)
1938 goto out;
1939
1940 /*
1941 * We're accessing css_set_count without locking css_set_lock here,
1942 * but that's OK - it can only be increased by someone holding
1943 * cgroup_lock, and that's us. Later rebinding may disable
1944 * controllers on the default hierarchy and thus create new csets,
1945 * which can't be more than the existing ones. Allocate 2x.
1946 */
1947 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1948 if (ret)
1949 goto cancel_ref;
1950
1951 ret = cgroup_init_root_id(root);
1952 if (ret)
1953 goto cancel_ref;
1954
1955 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1956 KERNFS_ROOT_CREATE_DEACTIVATED,
1957 root_cgrp);
1958 if (IS_ERR(root->kf_root)) {
1959 ret = PTR_ERR(root->kf_root);
1960 goto exit_root_id;
1961 }
1962 root_cgrp->kn = root->kf_root->kn;
1963
1964 ret = css_populate_dir(&root_cgrp->self);
1965 if (ret)
1966 goto destroy_root;
1967
1968 ret = rebind_subsystems(root, ss_mask);
1969 if (ret)
1970 goto destroy_root;
1971
1972 /*
1973 * There must be no failure case after here, since rebinding takes
1974 * care of subsystems' refcounts, which are explicitly dropped in
1975 * the failure exit path.
1976 */
1977 list_add(&root->root_list, &cgroup_roots);
1978 cgroup_root_count++;
1979
1980 /*
1981 * Link the root cgroup in this hierarchy into all the css_set
1982 * objects.
1983 */
1984 spin_lock_bh(&css_set_lock);
1985 hash_for_each(css_set_table, i, cset, hlist) {
1986 link_css_set(&tmp_links, cset, root_cgrp);
1987 if (css_set_populated(cset))
1988 cgroup_update_populated(root_cgrp, true);
1989 }
1990 spin_unlock_bh(&css_set_lock);
1991
1992 BUG_ON(!list_empty(&root_cgrp->self.children));
1993 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1994
1995 kernfs_activate(root_cgrp->kn);
1996 ret = 0;
1997 goto out;
1998
1999destroy_root:
2000 kernfs_destroy_root(root->kf_root);
2001 root->kf_root = NULL;
2002exit_root_id:
2003 cgroup_exit_root_id(root);
2004cancel_ref:
2005 percpu_ref_exit(&root_cgrp->self.refcnt);
2006out:
2007 free_cgrp_cset_links(&tmp_links);
2008 return ret;
2009}
2010
2011static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2012 int flags, const char *unused_dev_name,
2013 void *data)
2014{
2015 bool is_v2 = fs_type == &cgroup2_fs_type;
2016 struct super_block *pinned_sb = NULL;
2017 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2018 struct cgroup_subsys *ss;
2019 struct cgroup_root *root;
2020 struct cgroup_sb_opts opts;
2021 struct dentry *dentry;
2022 int ret;
2023 int i;
2024 bool new_sb;
2025
2026 get_cgroup_ns(ns);
2027
2028 /* Check if the caller has permission to mount. */
2029 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2030 put_cgroup_ns(ns);
2031 return ERR_PTR(-EPERM);
2032 }
2033
2034 /*
2035 * The first time anyone tries to mount a cgroup, enable the list
2036 * linking each css_set to its tasks and fix up all existing tasks.
2037 */
2038 if (!use_task_css_set_links)
2039 cgroup_enable_task_cg_lists();
2040
2041 if (is_v2) {
2042 if (data) {
2043 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2044 put_cgroup_ns(ns);
2045 return ERR_PTR(-EINVAL);
2046 }
2047 cgrp_dfl_visible = true;
2048 root = &cgrp_dfl_root;
2049 cgroup_get(&root->cgrp);
2050 goto out_mount;
2051 }
2052
2053 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2054
2055 /* First find the desired set of subsystems */
2056 ret = parse_cgroupfs_options(data, &opts);
2057 if (ret)
2058 goto out_unlock;
2059
2060 /*
2061 * Destruction of cgroup root is asynchronous, so subsystems may
2062 * still be dying after the previous unmount. Let's drain the
2063 * dying subsystems. We just need to ensure that the ones
2064 * unmounted previously finish dying and don't care about new ones
2065 * starting. Testing ref liveliness is good enough.
2066 */
2067 for_each_subsys(ss, i) {
2068 if (!(opts.subsys_mask & (1 << i)) ||
2069 ss->root == &cgrp_dfl_root)
2070 continue;
2071
2072 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2073 mutex_unlock(&cgroup_mutex);
2074 msleep(10);
2075 ret = restart_syscall();
2076 goto out_free;
2077 }
2078 cgroup_put(&ss->root->cgrp);
2079 }
2080
2081 for_each_root(root) {
2082 bool name_match = false;
2083
2084 if (root == &cgrp_dfl_root)
2085 continue;
2086
2087 /*
2088 * If we asked for a name then it must match. Also, if
2089 * name matches but sybsys_mask doesn't, we should fail.
2090 * Remember whether name matched.
2091 */
2092 if (opts.name) {
2093 if (strcmp(opts.name, root->name))
2094 continue;
2095 name_match = true;
2096 }
2097
2098 /*
2099 * If we asked for subsystems (or explicitly for no
2100 * subsystems) then they must match.
2101 */
2102 if ((opts.subsys_mask || opts.none) &&
2103 (opts.subsys_mask != root->subsys_mask)) {
2104 if (!name_match)
2105 continue;
2106 ret = -EBUSY;
2107 goto out_unlock;
2108 }
2109
2110 if (root->flags ^ opts.flags)
2111 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2112
2113 /*
2114 * We want to reuse @root whose lifetime is governed by its
2115 * ->cgrp. Let's check whether @root is alive and keep it
2116 * that way. As cgroup_kill_sb() can happen anytime, we
2117 * want to block it by pinning the sb so that @root doesn't
2118 * get killed before mount is complete.
2119 *
2120 * With the sb pinned, tryget_live can reliably indicate
2121 * whether @root can be reused. If it's being killed,
2122 * drain it. We can use wait_queue for the wait but this
2123 * path is super cold. Let's just sleep a bit and retry.
2124 */
2125 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2126 if (IS_ERR(pinned_sb) ||
2127 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2128 mutex_unlock(&cgroup_mutex);
2129 if (!IS_ERR_OR_NULL(pinned_sb))
2130 deactivate_super(pinned_sb);
2131 msleep(10);
2132 ret = restart_syscall();
2133 goto out_free;
2134 }
2135
2136 ret = 0;
2137 goto out_unlock;
2138 }
2139
2140 /*
2141 * No such thing, create a new one. name= matching without subsys
2142 * specification is allowed for already existing hierarchies but we
2143 * can't create new one without subsys specification.
2144 */
2145 if (!opts.subsys_mask && !opts.none) {
2146 ret = -EINVAL;
2147 goto out_unlock;
2148 }
2149
2150 /*
2151 * We know this subsystem has not yet been bound. Users in a non-init
2152 * user namespace may only mount hierarchies with no bound subsystems,
2153 * i.e. 'none,name=user1'
2154 */
2155 if (!opts.none && !capable(CAP_SYS_ADMIN)) {
2156 ret = -EPERM;
2157 goto out_unlock;
2158 }
2159
2160 root = kzalloc(sizeof(*root), GFP_KERNEL);
2161 if (!root) {
2162 ret = -ENOMEM;
2163 goto out_unlock;
2164 }
2165
2166 init_cgroup_root(root, &opts);
2167
2168 ret = cgroup_setup_root(root, opts.subsys_mask);
2169 if (ret)
2170 cgroup_free_root(root);
2171
2172out_unlock:
2173 mutex_unlock(&cgroup_mutex);
2174out_free:
2175 kfree(opts.release_agent);
2176 kfree(opts.name);
2177
2178 if (ret) {
2179 put_cgroup_ns(ns);
2180 return ERR_PTR(ret);
2181 }
2182out_mount:
2183 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2184 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2185 &new_sb);
2186
2187 /*
2188 * In non-init cgroup namespace, instead of root cgroup's
2189 * dentry, we return the dentry corresponding to the
2190 * cgroupns->root_cgrp.
2191 */
2192 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2193 struct dentry *nsdentry;
2194 struct cgroup *cgrp;
2195
2196 mutex_lock(&cgroup_mutex);
2197 spin_lock_bh(&css_set_lock);
2198
2199 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2200
2201 spin_unlock_bh(&css_set_lock);
2202 mutex_unlock(&cgroup_mutex);
2203
2204 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2205 dput(dentry);
2206 dentry = nsdentry;
2207 }
2208
2209 if (IS_ERR(dentry) || !new_sb)
2210 cgroup_put(&root->cgrp);
2211
2212 /*
2213 * If @pinned_sb, we're reusing an existing root and holding an
2214 * extra ref on its sb. Mount is complete. Put the extra ref.
2215 */
2216 if (pinned_sb) {
2217 WARN_ON(new_sb);
2218 deactivate_super(pinned_sb);
2219 }
2220
2221 put_cgroup_ns(ns);
2222 return dentry;
2223}
2224
2225static void cgroup_kill_sb(struct super_block *sb)
2226{
2227 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2228 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2229
2230 /*
2231 * If @root doesn't have any mounts or children, start killing it.
2232 * This prevents new mounts by disabling percpu_ref_tryget_live().
2233 * cgroup_mount() may wait for @root's release.
2234 *
2235 * And don't kill the default root.
2236 */
2237 if (!list_empty(&root->cgrp.self.children) ||
2238 root == &cgrp_dfl_root)
2239 cgroup_put(&root->cgrp);
2240 else
2241 percpu_ref_kill(&root->cgrp.self.refcnt);
2242
2243 kernfs_kill_sb(sb);
2244}
2245
2246static struct file_system_type cgroup_fs_type = {
2247 .name = "cgroup",
2248 .mount = cgroup_mount,
2249 .kill_sb = cgroup_kill_sb,
2250 .fs_flags = FS_USERNS_MOUNT,
2251};
2252
2253static struct file_system_type cgroup2_fs_type = {
2254 .name = "cgroup2",
2255 .mount = cgroup_mount,
2256 .kill_sb = cgroup_kill_sb,
2257 .fs_flags = FS_USERNS_MOUNT,
2258};
2259
2260static char *cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2261 struct cgroup_namespace *ns)
2262{
2263 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2264 int ret;
2265
2266 ret = kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2267 if (ret < 0 || ret >= buflen)
2268 return NULL;
2269 return buf;
2270}
2271
2272char *cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2273 struct cgroup_namespace *ns)
2274{
2275 char *ret;
2276
2277 mutex_lock(&cgroup_mutex);
2278 spin_lock_bh(&css_set_lock);
2279
2280 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2281
2282 spin_unlock_bh(&css_set_lock);
2283 mutex_unlock(&cgroup_mutex);
2284
2285 return ret;
2286}
2287EXPORT_SYMBOL_GPL(cgroup_path_ns);
2288
2289/**
2290 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2291 * @task: target task
2292 * @buf: the buffer to write the path into
2293 * @buflen: the length of the buffer
2294 *
2295 * Determine @task's cgroup on the first (the one with the lowest non-zero
2296 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2297 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2298 * cgroup controller callbacks.
2299 *
2300 * Return value is the same as kernfs_path().
2301 */
2302char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2303{
2304 struct cgroup_root *root;
2305 struct cgroup *cgrp;
2306 int hierarchy_id = 1;
2307 char *path = NULL;
2308
2309 mutex_lock(&cgroup_mutex);
2310 spin_lock_bh(&css_set_lock);
2311
2312 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2313
2314 if (root) {
2315 cgrp = task_cgroup_from_root(task, root);
2316 path = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2317 } else {
2318 /* if no hierarchy exists, everyone is in "/" */
2319 if (strlcpy(buf, "/", buflen) < buflen)
2320 path = buf;
2321 }
2322
2323 spin_unlock_bh(&css_set_lock);
2324 mutex_unlock(&cgroup_mutex);
2325 return path;
2326}
2327EXPORT_SYMBOL_GPL(task_cgroup_path);
2328
2329/* used to track tasks and other necessary states during migration */
2330struct cgroup_taskset {
2331 /* the src and dst cset list running through cset->mg_node */
2332 struct list_head src_csets;
2333 struct list_head dst_csets;
2334
2335 /* the subsys currently being processed */
2336 int ssid;
2337
2338 /*
2339 * Fields for cgroup_taskset_*() iteration.
2340 *
2341 * Before migration is committed, the target migration tasks are on
2342 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2343 * the csets on ->dst_csets. ->csets point to either ->src_csets
2344 * or ->dst_csets depending on whether migration is committed.
2345 *
2346 * ->cur_csets and ->cur_task point to the current task position
2347 * during iteration.
2348 */
2349 struct list_head *csets;
2350 struct css_set *cur_cset;
2351 struct task_struct *cur_task;
2352};
2353
2354#define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2355 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2356 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2357 .csets = &tset.src_csets, \
2358}
2359
2360/**
2361 * cgroup_taskset_add - try to add a migration target task to a taskset
2362 * @task: target task
2363 * @tset: target taskset
2364 *
2365 * Add @task, which is a migration target, to @tset. This function becomes
2366 * noop if @task doesn't need to be migrated. @task's css_set should have
2367 * been added as a migration source and @task->cg_list will be moved from
2368 * the css_set's tasks list to mg_tasks one.
2369 */
2370static void cgroup_taskset_add(struct task_struct *task,
2371 struct cgroup_taskset *tset)
2372{
2373 struct css_set *cset;
2374
2375 lockdep_assert_held(&css_set_lock);
2376
2377 /* @task either already exited or can't exit until the end */
2378 if (task->flags & PF_EXITING)
2379 return;
2380
2381 /* leave @task alone if post_fork() hasn't linked it yet */
2382 if (list_empty(&task->cg_list))
2383 return;
2384
2385 cset = task_css_set(task);
2386 if (!cset->mg_src_cgrp)
2387 return;
2388
2389 list_move_tail(&task->cg_list, &cset->mg_tasks);
2390 if (list_empty(&cset->mg_node))
2391 list_add_tail(&cset->mg_node, &tset->src_csets);
2392 if (list_empty(&cset->mg_dst_cset->mg_node))
2393 list_move_tail(&cset->mg_dst_cset->mg_node,
2394 &tset->dst_csets);
2395}
2396
2397/**
2398 * cgroup_taskset_first - reset taskset and return the first task
2399 * @tset: taskset of interest
2400 * @dst_cssp: output variable for the destination css
2401 *
2402 * @tset iteration is initialized and the first task is returned.
2403 */
2404struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2405 struct cgroup_subsys_state **dst_cssp)
2406{
2407 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2408 tset->cur_task = NULL;
2409
2410 return cgroup_taskset_next(tset, dst_cssp);
2411}
2412
2413/**
2414 * cgroup_taskset_next - iterate to the next task in taskset
2415 * @tset: taskset of interest
2416 * @dst_cssp: output variable for the destination css
2417 *
2418 * Return the next task in @tset. Iteration must have been initialized
2419 * with cgroup_taskset_first().
2420 */
2421struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2422 struct cgroup_subsys_state **dst_cssp)
2423{
2424 struct css_set *cset = tset->cur_cset;
2425 struct task_struct *task = tset->cur_task;
2426
2427 while (&cset->mg_node != tset->csets) {
2428 if (!task)
2429 task = list_first_entry(&cset->mg_tasks,
2430 struct task_struct, cg_list);
2431 else
2432 task = list_next_entry(task, cg_list);
2433
2434 if (&task->cg_list != &cset->mg_tasks) {
2435 tset->cur_cset = cset;
2436 tset->cur_task = task;
2437
2438 /*
2439 * This function may be called both before and
2440 * after cgroup_taskset_migrate(). The two cases
2441 * can be distinguished by looking at whether @cset
2442 * has its ->mg_dst_cset set.
2443 */
2444 if (cset->mg_dst_cset)
2445 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2446 else
2447 *dst_cssp = cset->subsys[tset->ssid];
2448
2449 return task;
2450 }
2451
2452 cset = list_next_entry(cset, mg_node);
2453 task = NULL;
2454 }
2455
2456 return NULL;
2457}
2458
2459/**
2460 * cgroup_taskset_migrate - migrate a taskset
2461 * @tset: taget taskset
2462 * @root: cgroup root the migration is taking place on
2463 *
2464 * Migrate tasks in @tset as setup by migration preparation functions.
2465 * This function fails iff one of the ->can_attach callbacks fails and
2466 * guarantees that either all or none of the tasks in @tset are migrated.
2467 * @tset is consumed regardless of success.
2468 */
2469static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2470 struct cgroup_root *root)
2471{
2472 struct cgroup_subsys *ss;
2473 struct task_struct *task, *tmp_task;
2474 struct css_set *cset, *tmp_cset;
2475 int ssid, failed_ssid, ret;
2476
2477 /* methods shouldn't be called if no task is actually migrating */
2478 if (list_empty(&tset->src_csets))
2479 return 0;
2480
2481 /* check that we can legitimately attach to the cgroup */
2482 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2483 if (ss->can_attach) {
2484 tset->ssid = ssid;
2485 ret = ss->can_attach(tset);
2486 if (ret) {
2487 failed_ssid = ssid;
2488 goto out_cancel_attach;
2489 }
2490 }
2491 } while_each_subsys_mask();
2492
2493 /*
2494 * Now that we're guaranteed success, proceed to move all tasks to
2495 * the new cgroup. There are no failure cases after here, so this
2496 * is the commit point.
2497 */
2498 spin_lock_bh(&css_set_lock);
2499 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2500 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2501 struct css_set *from_cset = task_css_set(task);
2502 struct css_set *to_cset = cset->mg_dst_cset;
2503
2504 get_css_set(to_cset);
2505 css_set_move_task(task, from_cset, to_cset, true);
2506 put_css_set_locked(from_cset);
2507 }
2508 }
2509 spin_unlock_bh(&css_set_lock);
2510
2511 /*
2512 * Migration is committed, all target tasks are now on dst_csets.
2513 * Nothing is sensitive to fork() after this point. Notify
2514 * controllers that migration is complete.
2515 */
2516 tset->csets = &tset->dst_csets;
2517
2518 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2519 if (ss->attach) {
2520 tset->ssid = ssid;
2521 ss->attach(tset);
2522 }
2523 } while_each_subsys_mask();
2524
2525 ret = 0;
2526 goto out_release_tset;
2527
2528out_cancel_attach:
2529 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2530 if (ssid == failed_ssid)
2531 break;
2532 if (ss->cancel_attach) {
2533 tset->ssid = ssid;
2534 ss->cancel_attach(tset);
2535 }
2536 } while_each_subsys_mask();
2537out_release_tset:
2538 spin_lock_bh(&css_set_lock);
2539 list_splice_init(&tset->dst_csets, &tset->src_csets);
2540 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2541 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2542 list_del_init(&cset->mg_node);
2543 }
2544 spin_unlock_bh(&css_set_lock);
2545 return ret;
2546}
2547
2548/**
2549 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2550 * @dst_cgrp: destination cgroup to test
2551 *
2552 * On the default hierarchy, except for the root, subtree_control must be
2553 * zero for migration destination cgroups with tasks so that child cgroups
2554 * don't compete against tasks.
2555 */
2556static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2557{
2558 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2559 !dst_cgrp->subtree_control;
2560}
2561
2562/**
2563 * cgroup_migrate_finish - cleanup after attach
2564 * @preloaded_csets: list of preloaded css_sets
2565 *
2566 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2567 * those functions for details.
2568 */
2569static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2570{
2571 struct css_set *cset, *tmp_cset;
2572
2573 lockdep_assert_held(&cgroup_mutex);
2574
2575 spin_lock_bh(&css_set_lock);
2576 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2577 cset->mg_src_cgrp = NULL;
2578 cset->mg_dst_cgrp = NULL;
2579 cset->mg_dst_cset = NULL;
2580 list_del_init(&cset->mg_preload_node);
2581 put_css_set_locked(cset);
2582 }
2583 spin_unlock_bh(&css_set_lock);
2584}
2585
2586/**
2587 * cgroup_migrate_add_src - add a migration source css_set
2588 * @src_cset: the source css_set to add
2589 * @dst_cgrp: the destination cgroup
2590 * @preloaded_csets: list of preloaded css_sets
2591 *
2592 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2593 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2594 * up by cgroup_migrate_finish().
2595 *
2596 * This function may be called without holding cgroup_threadgroup_rwsem
2597 * even if the target is a process. Threads may be created and destroyed
2598 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2599 * into play and the preloaded css_sets are guaranteed to cover all
2600 * migrations.
2601 */
2602static void cgroup_migrate_add_src(struct css_set *src_cset,
2603 struct cgroup *dst_cgrp,
2604 struct list_head *preloaded_csets)
2605{
2606 struct cgroup *src_cgrp;
2607
2608 lockdep_assert_held(&cgroup_mutex);
2609 lockdep_assert_held(&css_set_lock);
2610
2611 /*
2612 * If ->dead, @src_set is associated with one or more dead cgroups
2613 * and doesn't contain any migratable tasks. Ignore it early so
2614 * that the rest of migration path doesn't get confused by it.
2615 */
2616 if (src_cset->dead)
2617 return;
2618
2619 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2620
2621 if (!list_empty(&src_cset->mg_preload_node))
2622 return;
2623
2624 WARN_ON(src_cset->mg_src_cgrp);
2625 WARN_ON(src_cset->mg_dst_cgrp);
2626 WARN_ON(!list_empty(&src_cset->mg_tasks));
2627 WARN_ON(!list_empty(&src_cset->mg_node));
2628
2629 src_cset->mg_src_cgrp = src_cgrp;
2630 src_cset->mg_dst_cgrp = dst_cgrp;
2631 get_css_set(src_cset);
2632 list_add(&src_cset->mg_preload_node, preloaded_csets);
2633}
2634
2635/**
2636 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2637 * @preloaded_csets: list of preloaded source css_sets
2638 *
2639 * Tasks are about to be moved and all the source css_sets have been
2640 * preloaded to @preloaded_csets. This function looks up and pins all
2641 * destination css_sets, links each to its source, and append them to
2642 * @preloaded_csets.
2643 *
2644 * This function must be called after cgroup_migrate_add_src() has been
2645 * called on each migration source css_set. After migration is performed
2646 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2647 * @preloaded_csets.
2648 */
2649static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2650{
2651 LIST_HEAD(csets);
2652 struct css_set *src_cset, *tmp_cset;
2653
2654 lockdep_assert_held(&cgroup_mutex);
2655
2656 /* look up the dst cset for each src cset and link it to src */
2657 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2658 struct css_set *dst_cset;
2659
2660 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2661 if (!dst_cset)
2662 goto err;
2663
2664 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2665
2666 /*
2667 * If src cset equals dst, it's noop. Drop the src.
2668 * cgroup_migrate() will skip the cset too. Note that we
2669 * can't handle src == dst as some nodes are used by both.
2670 */
2671 if (src_cset == dst_cset) {
2672 src_cset->mg_src_cgrp = NULL;
2673 src_cset->mg_dst_cgrp = NULL;
2674 list_del_init(&src_cset->mg_preload_node);
2675 put_css_set(src_cset);
2676 put_css_set(dst_cset);
2677 continue;
2678 }
2679
2680 src_cset->mg_dst_cset = dst_cset;
2681
2682 if (list_empty(&dst_cset->mg_preload_node))
2683 list_add(&dst_cset->mg_preload_node, &csets);
2684 else
2685 put_css_set(dst_cset);
2686 }
2687
2688 list_splice_tail(&csets, preloaded_csets);
2689 return 0;
2690err:
2691 cgroup_migrate_finish(&csets);
2692 return -ENOMEM;
2693}
2694
2695/**
2696 * cgroup_migrate - migrate a process or task to a cgroup
2697 * @leader: the leader of the process or the task to migrate
2698 * @threadgroup: whether @leader points to the whole process or a single task
2699 * @root: cgroup root migration is taking place on
2700 *
2701 * Migrate a process or task denoted by @leader. If migrating a process,
2702 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2703 * responsible for invoking cgroup_migrate_add_src() and
2704 * cgroup_migrate_prepare_dst() on the targets before invoking this
2705 * function and following up with cgroup_migrate_finish().
2706 *
2707 * As long as a controller's ->can_attach() doesn't fail, this function is
2708 * guaranteed to succeed. This means that, excluding ->can_attach()
2709 * failure, when migrating multiple targets, the success or failure can be
2710 * decided for all targets by invoking group_migrate_prepare_dst() before
2711 * actually starting migrating.
2712 */
2713static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2714 struct cgroup_root *root)
2715{
2716 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2717 struct task_struct *task;
2718
2719 /*
2720 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2721 * already PF_EXITING could be freed from underneath us unless we
2722 * take an rcu_read_lock.
2723 */
2724 spin_lock_bh(&css_set_lock);
2725 rcu_read_lock();
2726 task = leader;
2727 do {
2728 cgroup_taskset_add(task, &tset);
2729 if (!threadgroup)
2730 break;
2731 } while_each_thread(leader, task);
2732 rcu_read_unlock();
2733 spin_unlock_bh(&css_set_lock);
2734
2735 return cgroup_taskset_migrate(&tset, root);
2736}
2737
2738/**
2739 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2740 * @dst_cgrp: the cgroup to attach to
2741 * @leader: the task or the leader of the threadgroup to be attached
2742 * @threadgroup: attach the whole threadgroup?
2743 *
2744 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2745 */
2746static int cgroup_attach_task(struct cgroup *dst_cgrp,
2747 struct task_struct *leader, bool threadgroup)
2748{
2749 LIST_HEAD(preloaded_csets);
2750 struct task_struct *task;
2751 int ret;
2752
2753 if (!cgroup_may_migrate_to(dst_cgrp))
2754 return -EBUSY;
2755
2756 /* look up all src csets */
2757 spin_lock_bh(&css_set_lock);
2758 rcu_read_lock();
2759 task = leader;
2760 do {
2761 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2762 &preloaded_csets);
2763 if (!threadgroup)
2764 break;
2765 } while_each_thread(leader, task);
2766 rcu_read_unlock();
2767 spin_unlock_bh(&css_set_lock);
2768
2769 /* prepare dst csets and commit */
2770 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2771 if (!ret)
2772 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2773
2774 cgroup_migrate_finish(&preloaded_csets);
2775 return ret;
2776}
2777
2778static int cgroup_procs_write_permission(struct task_struct *task,
2779 struct cgroup *dst_cgrp,
2780 struct kernfs_open_file *of)
2781{
2782 const struct cred *cred = current_cred();
2783 const struct cred *tcred = get_task_cred(task);
2784 int ret = 0;
2785
2786 /*
2787 * even if we're attaching all tasks in the thread group, we only
2788 * need to check permissions on one of them.
2789 */
2790 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2791 !uid_eq(cred->euid, tcred->uid) &&
2792 !uid_eq(cred->euid, tcred->suid))
2793 ret = -EACCES;
2794
2795 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2796 struct super_block *sb = of->file->f_path.dentry->d_sb;
2797 struct cgroup *cgrp;
2798 struct inode *inode;
2799
2800 spin_lock_bh(&css_set_lock);
2801 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2802 spin_unlock_bh(&css_set_lock);
2803
2804 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2805 cgrp = cgroup_parent(cgrp);
2806
2807 ret = -ENOMEM;
2808 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2809 if (inode) {
2810 ret = inode_permission(inode, MAY_WRITE);
2811 iput(inode);
2812 }
2813 }
2814
2815 put_cred(tcred);
2816 return ret;
2817}
2818
2819/*
2820 * Find the task_struct of the task to attach by vpid and pass it along to the
2821 * function to attach either it or all tasks in its threadgroup. Will lock
2822 * cgroup_mutex and threadgroup.
2823 */
2824static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2825 size_t nbytes, loff_t off, bool threadgroup)
2826{
2827 struct task_struct *tsk;
2828 struct cgroup *cgrp;
2829 pid_t pid;
2830 int ret;
2831
2832 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2833 return -EINVAL;
2834
2835 cgrp = cgroup_kn_lock_live(of->kn, false);
2836 if (!cgrp)
2837 return -ENODEV;
2838
2839 percpu_down_write(&cgroup_threadgroup_rwsem);
2840 rcu_read_lock();
2841 if (pid) {
2842 tsk = find_task_by_vpid(pid);
2843 if (!tsk) {
2844 ret = -ESRCH;
2845 goto out_unlock_rcu;
2846 }
2847 } else {
2848 tsk = current;
2849 }
2850
2851 if (threadgroup)
2852 tsk = tsk->group_leader;
2853
2854 /*
2855 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2856 * trapped in a cpuset, or RT worker may be born in a cgroup
2857 * with no rt_runtime allocated. Just say no.
2858 */
2859 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2860 ret = -EINVAL;
2861 goto out_unlock_rcu;
2862 }
2863
2864 get_task_struct(tsk);
2865 rcu_read_unlock();
2866
2867 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2868 if (!ret)
2869 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2870
2871 put_task_struct(tsk);
2872 goto out_unlock_threadgroup;
2873
2874out_unlock_rcu:
2875 rcu_read_unlock();
2876out_unlock_threadgroup:
2877 percpu_up_write(&cgroup_threadgroup_rwsem);
2878 cgroup_kn_unlock(of->kn);
2879 cpuset_post_attach_flush();
2880 return ret ?: nbytes;
2881}
2882
2883/**
2884 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2885 * @from: attach to all cgroups of a given task
2886 * @tsk: the task to be attached
2887 */
2888int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2889{
2890 struct cgroup_root *root;
2891 int retval = 0;
2892
2893 mutex_lock(&cgroup_mutex);
2894 for_each_root(root) {
2895 struct cgroup *from_cgrp;
2896
2897 if (root == &cgrp_dfl_root)
2898 continue;
2899
2900 spin_lock_bh(&css_set_lock);
2901 from_cgrp = task_cgroup_from_root(from, root);
2902 spin_unlock_bh(&css_set_lock);
2903
2904 retval = cgroup_attach_task(from_cgrp, tsk, false);
2905 if (retval)
2906 break;
2907 }
2908 mutex_unlock(&cgroup_mutex);
2909
2910 return retval;
2911}
2912EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2913
2914static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2915 char *buf, size_t nbytes, loff_t off)
2916{
2917 return __cgroup_procs_write(of, buf, nbytes, off, false);
2918}
2919
2920static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2921 char *buf, size_t nbytes, loff_t off)
2922{
2923 return __cgroup_procs_write(of, buf, nbytes, off, true);
2924}
2925
2926static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2927 char *buf, size_t nbytes, loff_t off)
2928{
2929 struct cgroup *cgrp;
2930
2931 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2932
2933 cgrp = cgroup_kn_lock_live(of->kn, false);
2934 if (!cgrp)
2935 return -ENODEV;
2936 spin_lock(&release_agent_path_lock);
2937 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2938 sizeof(cgrp->root->release_agent_path));
2939 spin_unlock(&release_agent_path_lock);
2940 cgroup_kn_unlock(of->kn);
2941 return nbytes;
2942}
2943
2944static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2945{
2946 struct cgroup *cgrp = seq_css(seq)->cgroup;
2947
2948 spin_lock(&release_agent_path_lock);
2949 seq_puts(seq, cgrp->root->release_agent_path);
2950 spin_unlock(&release_agent_path_lock);
2951 seq_putc(seq, '\n');
2952 return 0;
2953}
2954
2955static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2956{
2957 seq_puts(seq, "0\n");
2958 return 0;
2959}
2960
2961static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2962{
2963 struct cgroup_subsys *ss;
2964 bool printed = false;
2965 int ssid;
2966
2967 do_each_subsys_mask(ss, ssid, ss_mask) {
2968 if (printed)
2969 seq_putc(seq, ' ');
2970 seq_printf(seq, "%s", ss->name);
2971 printed = true;
2972 } while_each_subsys_mask();
2973 if (printed)
2974 seq_putc(seq, '\n');
2975}
2976
2977/* show controllers which are enabled from the parent */
2978static int cgroup_controllers_show(struct seq_file *seq, void *v)
2979{
2980 struct cgroup *cgrp = seq_css(seq)->cgroup;
2981
2982 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2983 return 0;
2984}
2985
2986/* show controllers which are enabled for a given cgroup's children */
2987static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2988{
2989 struct cgroup *cgrp = seq_css(seq)->cgroup;
2990
2991 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2992 return 0;
2993}
2994
2995/**
2996 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2997 * @cgrp: root of the subtree to update csses for
2998 *
2999 * @cgrp's control masks have changed and its subtree's css associations
3000 * need to be updated accordingly. This function looks up all css_sets
3001 * which are attached to the subtree, creates the matching updated css_sets
3002 * and migrates the tasks to the new ones.
3003 */
3004static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3005{
3006 LIST_HEAD(preloaded_csets);
3007 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
3008 struct cgroup_subsys_state *d_css;
3009 struct cgroup *dsct;
3010 struct css_set *src_cset;
3011 int ret;
3012
3013 lockdep_assert_held(&cgroup_mutex);
3014
3015 percpu_down_write(&cgroup_threadgroup_rwsem);
3016
3017 /* look up all csses currently attached to @cgrp's subtree */
3018 spin_lock_bh(&css_set_lock);
3019 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3020 struct cgrp_cset_link *link;
3021
3022 list_for_each_entry(link, &dsct->cset_links, cset_link)
3023 cgroup_migrate_add_src(link->cset, dsct,
3024 &preloaded_csets);
3025 }
3026 spin_unlock_bh(&css_set_lock);
3027
3028 /* NULL dst indicates self on default hierarchy */
3029 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
3030 if (ret)
3031 goto out_finish;
3032
3033 spin_lock_bh(&css_set_lock);
3034 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3035 struct task_struct *task, *ntask;
3036
3037 /* src_csets precede dst_csets, break on the first dst_cset */
3038 if (!src_cset->mg_src_cgrp)
3039 break;
3040
3041 /* all tasks in src_csets need to be migrated */
3042 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3043 cgroup_taskset_add(task, &tset);
3044 }
3045 spin_unlock_bh(&css_set_lock);
3046
3047 ret = cgroup_taskset_migrate(&tset, cgrp->root);
3048out_finish:
3049 cgroup_migrate_finish(&preloaded_csets);
3050 percpu_up_write(&cgroup_threadgroup_rwsem);
3051 return ret;
3052}
3053
3054/**
3055 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3056 * @cgrp: root of the target subtree
3057 *
3058 * Because css offlining is asynchronous, userland may try to re-enable a
3059 * controller while the previous css is still around. This function grabs
3060 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3061 */
3062static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3063 __acquires(&cgroup_mutex)
3064{
3065 struct cgroup *dsct;
3066 struct cgroup_subsys_state *d_css;
3067 struct cgroup_subsys *ss;
3068 int ssid;
3069
3070restart:
3071 mutex_lock(&cgroup_mutex);
3072
3073 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3074 for_each_subsys(ss, ssid) {
3075 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3076 DEFINE_WAIT(wait);
3077
3078 if (!css || !percpu_ref_is_dying(&css->refcnt))
3079 continue;
3080
3081 cgroup_get(dsct);
3082 prepare_to_wait(&dsct->offline_waitq, &wait,
3083 TASK_UNINTERRUPTIBLE);
3084
3085 mutex_unlock(&cgroup_mutex);
3086 schedule();
3087 finish_wait(&dsct->offline_waitq, &wait);
3088
3089 cgroup_put(dsct);
3090 goto restart;
3091 }
3092 }
3093}
3094
3095/**
3096 * cgroup_save_control - save control masks of a subtree
3097 * @cgrp: root of the target subtree
3098 *
3099 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3100 * prefixed fields for @cgrp's subtree including @cgrp itself.
3101 */
3102static void cgroup_save_control(struct cgroup *cgrp)
3103{
3104 struct cgroup *dsct;
3105 struct cgroup_subsys_state *d_css;
3106
3107 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3108 dsct->old_subtree_control = dsct->subtree_control;
3109 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3110 }
3111}
3112
3113/**
3114 * cgroup_propagate_control - refresh control masks of a subtree
3115 * @cgrp: root of the target subtree
3116 *
3117 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3118 * ->subtree_control and propagate controller availability through the
3119 * subtree so that descendants don't have unavailable controllers enabled.
3120 */
3121static void cgroup_propagate_control(struct cgroup *cgrp)
3122{
3123 struct cgroup *dsct;
3124 struct cgroup_subsys_state *d_css;
3125
3126 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3127 dsct->subtree_control &= cgroup_control(dsct);
3128 dsct->subtree_ss_mask =
3129 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3130 cgroup_ss_mask(dsct));
3131 }
3132}
3133
3134/**
3135 * cgroup_restore_control - restore control masks of a subtree
3136 * @cgrp: root of the target subtree
3137 *
3138 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3139 * prefixed fields for @cgrp's subtree including @cgrp itself.
3140 */
3141static void cgroup_restore_control(struct cgroup *cgrp)
3142{
3143 struct cgroup *dsct;
3144 struct cgroup_subsys_state *d_css;
3145
3146 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3147 dsct->subtree_control = dsct->old_subtree_control;
3148 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3149 }
3150}
3151
3152static bool css_visible(struct cgroup_subsys_state *css)
3153{
3154 struct cgroup_subsys *ss = css->ss;
3155 struct cgroup *cgrp = css->cgroup;
3156
3157 if (cgroup_control(cgrp) & (1 << ss->id))
3158 return true;
3159 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3160 return false;
3161 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3162}
3163
3164/**
3165 * cgroup_apply_control_enable - enable or show csses according to control
3166 * @cgrp: root of the target subtree
3167 *
3168 * Walk @cgrp's subtree and create new csses or make the existing ones
3169 * visible. A css is created invisible if it's being implicitly enabled
3170 * through dependency. An invisible css is made visible when the userland
3171 * explicitly enables it.
3172 *
3173 * Returns 0 on success, -errno on failure. On failure, csses which have
3174 * been processed already aren't cleaned up. The caller is responsible for
3175 * cleaning up with cgroup_apply_control_disble().
3176 */
3177static int cgroup_apply_control_enable(struct cgroup *cgrp)
3178{
3179 struct cgroup *dsct;
3180 struct cgroup_subsys_state *d_css;
3181 struct cgroup_subsys *ss;
3182 int ssid, ret;
3183
3184 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3185 for_each_subsys(ss, ssid) {
3186 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3187
3188 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3189
3190 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3191 continue;
3192
3193 if (!css) {
3194 css = css_create(dsct, ss);
3195 if (IS_ERR(css))
3196 return PTR_ERR(css);
3197 }
3198
3199 if (css_visible(css)) {
3200 ret = css_populate_dir(css);
3201 if (ret)
3202 return ret;
3203 }
3204 }
3205 }
3206
3207 return 0;
3208}
3209
3210/**
3211 * cgroup_apply_control_disable - kill or hide csses according to control
3212 * @cgrp: root of the target subtree
3213 *
3214 * Walk @cgrp's subtree and kill and hide csses so that they match
3215 * cgroup_ss_mask() and cgroup_visible_mask().
3216 *
3217 * A css is hidden when the userland requests it to be disabled while other
3218 * subsystems are still depending on it. The css must not actively control
3219 * resources and be in the vanilla state if it's made visible again later.
3220 * Controllers which may be depended upon should provide ->css_reset() for
3221 * this purpose.
3222 */
3223static void cgroup_apply_control_disable(struct cgroup *cgrp)
3224{
3225 struct cgroup *dsct;
3226 struct cgroup_subsys_state *d_css;
3227 struct cgroup_subsys *ss;
3228 int ssid;
3229
3230 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3231 for_each_subsys(ss, ssid) {
3232 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3233
3234 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3235
3236 if (!css)
3237 continue;
3238
3239 if (css->parent &&
3240 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3241 kill_css(css);
3242 } else if (!css_visible(css)) {
3243 css_clear_dir(css);
3244 if (ss->css_reset)
3245 ss->css_reset(css);
3246 }
3247 }
3248 }
3249}
3250
3251/**
3252 * cgroup_apply_control - apply control mask updates to the subtree
3253 * @cgrp: root of the target subtree
3254 *
3255 * subsystems can be enabled and disabled in a subtree using the following
3256 * steps.
3257 *
3258 * 1. Call cgroup_save_control() to stash the current state.
3259 * 2. Update ->subtree_control masks in the subtree as desired.
3260 * 3. Call cgroup_apply_control() to apply the changes.
3261 * 4. Optionally perform other related operations.
3262 * 5. Call cgroup_finalize_control() to finish up.
3263 *
3264 * This function implements step 3 and propagates the mask changes
3265 * throughout @cgrp's subtree, updates csses accordingly and perform
3266 * process migrations.
3267 */
3268static int cgroup_apply_control(struct cgroup *cgrp)
3269{
3270 int ret;
3271
3272 cgroup_propagate_control(cgrp);
3273
3274 ret = cgroup_apply_control_enable(cgrp);
3275 if (ret)
3276 return ret;
3277
3278 /*
3279 * At this point, cgroup_e_css() results reflect the new csses
3280 * making the following cgroup_update_dfl_csses() properly update
3281 * css associations of all tasks in the subtree.
3282 */
3283 ret = cgroup_update_dfl_csses(cgrp);
3284 if (ret)
3285 return ret;
3286
3287 return 0;
3288}
3289
3290/**
3291 * cgroup_finalize_control - finalize control mask update
3292 * @cgrp: root of the target subtree
3293 * @ret: the result of the update
3294 *
3295 * Finalize control mask update. See cgroup_apply_control() for more info.
3296 */
3297static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3298{
3299 if (ret) {
3300 cgroup_restore_control(cgrp);
3301 cgroup_propagate_control(cgrp);
3302 }
3303
3304 cgroup_apply_control_disable(cgrp);
3305}
3306
3307/* change the enabled child controllers for a cgroup in the default hierarchy */
3308static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3309 char *buf, size_t nbytes,
3310 loff_t off)
3311{
3312 u16 enable = 0, disable = 0;
3313 struct cgroup *cgrp, *child;
3314 struct cgroup_subsys *ss;
3315 char *tok;
3316 int ssid, ret;
3317
3318 /*
3319 * Parse input - space separated list of subsystem names prefixed
3320 * with either + or -.
3321 */
3322 buf = strstrip(buf);
3323 while ((tok = strsep(&buf, " "))) {
3324 if (tok[0] == '\0')
3325 continue;
3326 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3327 if (!cgroup_ssid_enabled(ssid) ||
3328 strcmp(tok + 1, ss->name))
3329 continue;
3330
3331 if (*tok == '+') {
3332 enable |= 1 << ssid;
3333 disable &= ~(1 << ssid);
3334 } else if (*tok == '-') {
3335 disable |= 1 << ssid;
3336 enable &= ~(1 << ssid);
3337 } else {
3338 return -EINVAL;
3339 }
3340 break;
3341 } while_each_subsys_mask();
3342 if (ssid == CGROUP_SUBSYS_COUNT)
3343 return -EINVAL;
3344 }
3345
3346 cgrp = cgroup_kn_lock_live(of->kn, true);
3347 if (!cgrp)
3348 return -ENODEV;
3349
3350 for_each_subsys(ss, ssid) {
3351 if (enable & (1 << ssid)) {
3352 if (cgrp->subtree_control & (1 << ssid)) {
3353 enable &= ~(1 << ssid);
3354 continue;
3355 }
3356
3357 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3358 ret = -ENOENT;
3359 goto out_unlock;
3360 }
3361 } else if (disable & (1 << ssid)) {
3362 if (!(cgrp->subtree_control & (1 << ssid))) {
3363 disable &= ~(1 << ssid);
3364 continue;
3365 }
3366
3367 /* a child has it enabled? */
3368 cgroup_for_each_live_child(child, cgrp) {
3369 if (child->subtree_control & (1 << ssid)) {
3370 ret = -EBUSY;
3371 goto out_unlock;
3372 }
3373 }
3374 }
3375 }
3376
3377 if (!enable && !disable) {
3378 ret = 0;
3379 goto out_unlock;
3380 }
3381
3382 /*
3383 * Except for the root, subtree_control must be zero for a cgroup
3384 * with tasks so that child cgroups don't compete against tasks.
3385 */
3386 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3387 ret = -EBUSY;
3388 goto out_unlock;
3389 }
3390
3391 /* save and update control masks and prepare csses */
3392 cgroup_save_control(cgrp);
3393
3394 cgrp->subtree_control |= enable;
3395 cgrp->subtree_control &= ~disable;
3396
3397 ret = cgroup_apply_control(cgrp);
3398
3399 cgroup_finalize_control(cgrp, ret);
3400
3401 kernfs_activate(cgrp->kn);
3402 ret = 0;
3403out_unlock:
3404 cgroup_kn_unlock(of->kn);
3405 return ret ?: nbytes;
3406}
3407
3408static int cgroup_events_show(struct seq_file *seq, void *v)
3409{
3410 seq_printf(seq, "populated %d\n",
3411 cgroup_is_populated(seq_css(seq)->cgroup));
3412 return 0;
3413}
3414
3415static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3416 size_t nbytes, loff_t off)
3417{
3418 struct cgroup *cgrp = of->kn->parent->priv;
3419 struct cftype *cft = of->kn->priv;
3420 struct cgroup_subsys_state *css;
3421 int ret;
3422
3423 if (cft->write)
3424 return cft->write(of, buf, nbytes, off);
3425
3426 /*
3427 * kernfs guarantees that a file isn't deleted with operations in
3428 * flight, which means that the matching css is and stays alive and
3429 * doesn't need to be pinned. The RCU locking is not necessary
3430 * either. It's just for the convenience of using cgroup_css().
3431 */
3432 rcu_read_lock();
3433 css = cgroup_css(cgrp, cft->ss);
3434 rcu_read_unlock();
3435
3436 if (cft->write_u64) {
3437 unsigned long long v;
3438 ret = kstrtoull(buf, 0, &v);
3439 if (!ret)
3440 ret = cft->write_u64(css, cft, v);
3441 } else if (cft->write_s64) {
3442 long long v;
3443 ret = kstrtoll(buf, 0, &v);
3444 if (!ret)
3445 ret = cft->write_s64(css, cft, v);
3446 } else {
3447 ret = -EINVAL;
3448 }
3449
3450 return ret ?: nbytes;
3451}
3452
3453static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3454{
3455 return seq_cft(seq)->seq_start(seq, ppos);
3456}
3457
3458static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3459{
3460 return seq_cft(seq)->seq_next(seq, v, ppos);
3461}
3462
3463static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3464{
3465 seq_cft(seq)->seq_stop(seq, v);
3466}
3467
3468static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3469{
3470 struct cftype *cft = seq_cft(m);
3471 struct cgroup_subsys_state *css = seq_css(m);
3472
3473 if (cft->seq_show)
3474 return cft->seq_show(m, arg);
3475
3476 if (cft->read_u64)
3477 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3478 else if (cft->read_s64)
3479 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3480 else
3481 return -EINVAL;
3482 return 0;
3483}
3484
3485static struct kernfs_ops cgroup_kf_single_ops = {
3486 .atomic_write_len = PAGE_SIZE,
3487 .write = cgroup_file_write,
3488 .seq_show = cgroup_seqfile_show,
3489};
3490
3491static struct kernfs_ops cgroup_kf_ops = {
3492 .atomic_write_len = PAGE_SIZE,
3493 .write = cgroup_file_write,
3494 .seq_start = cgroup_seqfile_start,
3495 .seq_next = cgroup_seqfile_next,
3496 .seq_stop = cgroup_seqfile_stop,
3497 .seq_show = cgroup_seqfile_show,
3498};
3499
3500/*
3501 * cgroup_rename - Only allow simple rename of directories in place.
3502 */
3503static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3504 const char *new_name_str)
3505{
3506 struct cgroup *cgrp = kn->priv;
3507 int ret;
3508
3509 if (kernfs_type(kn) != KERNFS_DIR)
3510 return -ENOTDIR;
3511 if (kn->parent != new_parent)
3512 return -EIO;
3513
3514 /*
3515 * This isn't a proper migration and its usefulness is very
3516 * limited. Disallow on the default hierarchy.
3517 */
3518 if (cgroup_on_dfl(cgrp))
3519 return -EPERM;
3520
3521 /*
3522 * We're gonna grab cgroup_mutex which nests outside kernfs
3523 * active_ref. kernfs_rename() doesn't require active_ref
3524 * protection. Break them before grabbing cgroup_mutex.
3525 */
3526 kernfs_break_active_protection(new_parent);
3527 kernfs_break_active_protection(kn);
3528
3529 mutex_lock(&cgroup_mutex);
3530
3531 ret = kernfs_rename(kn, new_parent, new_name_str);
3532
3533 mutex_unlock(&cgroup_mutex);
3534
3535 kernfs_unbreak_active_protection(kn);
3536 kernfs_unbreak_active_protection(new_parent);
3537 return ret;
3538}
3539
3540/* set uid and gid of cgroup dirs and files to that of the creator */
3541static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3542{
3543 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3544 .ia_uid = current_fsuid(),
3545 .ia_gid = current_fsgid(), };
3546
3547 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3548 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3549 return 0;
3550
3551 return kernfs_setattr(kn, &iattr);
3552}
3553
3554static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3555 struct cftype *cft)
3556{
3557 char name[CGROUP_FILE_NAME_MAX];
3558 struct kernfs_node *kn;
3559 struct lock_class_key *key = NULL;
3560 int ret;
3561
3562#ifdef CONFIG_DEBUG_LOCK_ALLOC
3563 key = &cft->lockdep_key;
3564#endif
3565 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3566 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3567 NULL, key);
3568 if (IS_ERR(kn))
3569 return PTR_ERR(kn);
3570
3571 ret = cgroup_kn_set_ugid(kn);
3572 if (ret) {
3573 kernfs_remove(kn);
3574 return ret;
3575 }
3576
3577 if (cft->file_offset) {
3578 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3579
3580 spin_lock_irq(&cgroup_file_kn_lock);
3581 cfile->kn = kn;
3582 spin_unlock_irq(&cgroup_file_kn_lock);
3583 }
3584
3585 return 0;
3586}
3587
3588/**
3589 * cgroup_addrm_files - add or remove files to a cgroup directory
3590 * @css: the target css
3591 * @cgrp: the target cgroup (usually css->cgroup)
3592 * @cfts: array of cftypes to be added
3593 * @is_add: whether to add or remove
3594 *
3595 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3596 * For removals, this function never fails.
3597 */
3598static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3599 struct cgroup *cgrp, struct cftype cfts[],
3600 bool is_add)
3601{
3602 struct cftype *cft, *cft_end = NULL;
3603 int ret = 0;
3604
3605 lockdep_assert_held(&cgroup_mutex);
3606
3607restart:
3608 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3609 /* does cft->flags tell us to skip this file on @cgrp? */
3610 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3611 continue;
3612 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3613 continue;
3614 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3615 continue;
3616 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3617 continue;
3618
3619 if (is_add) {
3620 ret = cgroup_add_file(css, cgrp, cft);
3621 if (ret) {
3622 pr_warn("%s: failed to add %s, err=%d\n",
3623 __func__, cft->name, ret);
3624 cft_end = cft;
3625 is_add = false;
3626 goto restart;
3627 }
3628 } else {
3629 cgroup_rm_file(cgrp, cft);
3630 }
3631 }
3632 return ret;
3633}
3634
3635static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3636{
3637 LIST_HEAD(pending);
3638 struct cgroup_subsys *ss = cfts[0].ss;
3639 struct cgroup *root = &ss->root->cgrp;
3640 struct cgroup_subsys_state *css;
3641 int ret = 0;
3642
3643 lockdep_assert_held(&cgroup_mutex);
3644
3645 /* add/rm files for all cgroups created before */
3646 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3647 struct cgroup *cgrp = css->cgroup;
3648
3649 if (!(css->flags & CSS_VISIBLE))
3650 continue;
3651
3652 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3653 if (ret)
3654 break;
3655 }
3656
3657 if (is_add && !ret)
3658 kernfs_activate(root->kn);
3659 return ret;
3660}
3661
3662static void cgroup_exit_cftypes(struct cftype *cfts)
3663{
3664 struct cftype *cft;
3665
3666 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3667 /* free copy for custom atomic_write_len, see init_cftypes() */
3668 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3669 kfree(cft->kf_ops);
3670 cft->kf_ops = NULL;
3671 cft->ss = NULL;
3672
3673 /* revert flags set by cgroup core while adding @cfts */
3674 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3675 }
3676}
3677
3678static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3679{
3680 struct cftype *cft;
3681
3682 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3683 struct kernfs_ops *kf_ops;
3684
3685 WARN_ON(cft->ss || cft->kf_ops);
3686
3687 if (cft->seq_start)
3688 kf_ops = &cgroup_kf_ops;
3689 else
3690 kf_ops = &cgroup_kf_single_ops;
3691
3692 /*
3693 * Ugh... if @cft wants a custom max_write_len, we need to
3694 * make a copy of kf_ops to set its atomic_write_len.
3695 */
3696 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3697 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3698 if (!kf_ops) {
3699 cgroup_exit_cftypes(cfts);
3700 return -ENOMEM;
3701 }
3702 kf_ops->atomic_write_len = cft->max_write_len;
3703 }
3704
3705 cft->kf_ops = kf_ops;
3706 cft->ss = ss;
3707 }
3708
3709 return 0;
3710}
3711
3712static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3713{
3714 lockdep_assert_held(&cgroup_mutex);
3715
3716 if (!cfts || !cfts[0].ss)
3717 return -ENOENT;
3718
3719 list_del(&cfts->node);
3720 cgroup_apply_cftypes(cfts, false);
3721 cgroup_exit_cftypes(cfts);
3722 return 0;
3723}
3724
3725/**
3726 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3727 * @cfts: zero-length name terminated array of cftypes
3728 *
3729 * Unregister @cfts. Files described by @cfts are removed from all
3730 * existing cgroups and all future cgroups won't have them either. This
3731 * function can be called anytime whether @cfts' subsys is attached or not.
3732 *
3733 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3734 * registered.
3735 */
3736int cgroup_rm_cftypes(struct cftype *cfts)
3737{
3738 int ret;
3739
3740 mutex_lock(&cgroup_mutex);
3741 ret = cgroup_rm_cftypes_locked(cfts);
3742 mutex_unlock(&cgroup_mutex);
3743 return ret;
3744}
3745
3746/**
3747 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3748 * @ss: target cgroup subsystem
3749 * @cfts: zero-length name terminated array of cftypes
3750 *
3751 * Register @cfts to @ss. Files described by @cfts are created for all
3752 * existing cgroups to which @ss is attached and all future cgroups will
3753 * have them too. This function can be called anytime whether @ss is
3754 * attached or not.
3755 *
3756 * Returns 0 on successful registration, -errno on failure. Note that this
3757 * function currently returns 0 as long as @cfts registration is successful
3758 * even if some file creation attempts on existing cgroups fail.
3759 */
3760static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3761{
3762 int ret;
3763
3764 if (!cgroup_ssid_enabled(ss->id))
3765 return 0;
3766
3767 if (!cfts || cfts[0].name[0] == '\0')
3768 return 0;
3769
3770 ret = cgroup_init_cftypes(ss, cfts);
3771 if (ret)
3772 return ret;
3773
3774 mutex_lock(&cgroup_mutex);
3775
3776 list_add_tail(&cfts->node, &ss->cfts);
3777 ret = cgroup_apply_cftypes(cfts, true);
3778 if (ret)
3779 cgroup_rm_cftypes_locked(cfts);
3780
3781 mutex_unlock(&cgroup_mutex);
3782 return ret;
3783}
3784
3785/**
3786 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3787 * @ss: target cgroup subsystem
3788 * @cfts: zero-length name terminated array of cftypes
3789 *
3790 * Similar to cgroup_add_cftypes() but the added files are only used for
3791 * the default hierarchy.
3792 */
3793int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3794{
3795 struct cftype *cft;
3796
3797 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3798 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3799 return cgroup_add_cftypes(ss, cfts);
3800}
3801
3802/**
3803 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3804 * @ss: target cgroup subsystem
3805 * @cfts: zero-length name terminated array of cftypes
3806 *
3807 * Similar to cgroup_add_cftypes() but the added files are only used for
3808 * the legacy hierarchies.
3809 */
3810int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3811{
3812 struct cftype *cft;
3813
3814 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3815 cft->flags |= __CFTYPE_NOT_ON_DFL;
3816 return cgroup_add_cftypes(ss, cfts);
3817}
3818
3819/**
3820 * cgroup_file_notify - generate a file modified event for a cgroup_file
3821 * @cfile: target cgroup_file
3822 *
3823 * @cfile must have been obtained by setting cftype->file_offset.
3824 */
3825void cgroup_file_notify(struct cgroup_file *cfile)
3826{
3827 unsigned long flags;
3828
3829 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3830 if (cfile->kn)
3831 kernfs_notify(cfile->kn);
3832 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3833}
3834
3835/**
3836 * cgroup_task_count - count the number of tasks in a cgroup.
3837 * @cgrp: the cgroup in question
3838 *
3839 * Return the number of tasks in the cgroup.
3840 */
3841static int cgroup_task_count(const struct cgroup *cgrp)
3842{
3843 int count = 0;
3844 struct cgrp_cset_link *link;
3845
3846 spin_lock_bh(&css_set_lock);
3847 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3848 count += atomic_read(&link->cset->refcount);
3849 spin_unlock_bh(&css_set_lock);
3850 return count;
3851}
3852
3853/**
3854 * css_next_child - find the next child of a given css
3855 * @pos: the current position (%NULL to initiate traversal)
3856 * @parent: css whose children to walk
3857 *
3858 * This function returns the next child of @parent and should be called
3859 * under either cgroup_mutex or RCU read lock. The only requirement is
3860 * that @parent and @pos are accessible. The next sibling is guaranteed to
3861 * be returned regardless of their states.
3862 *
3863 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3864 * css which finished ->css_online() is guaranteed to be visible in the
3865 * future iterations and will stay visible until the last reference is put.
3866 * A css which hasn't finished ->css_online() or already finished
3867 * ->css_offline() may show up during traversal. It's each subsystem's
3868 * responsibility to synchronize against on/offlining.
3869 */
3870struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3871 struct cgroup_subsys_state *parent)
3872{
3873 struct cgroup_subsys_state *next;
3874
3875 cgroup_assert_mutex_or_rcu_locked();
3876
3877 /*
3878 * @pos could already have been unlinked from the sibling list.
3879 * Once a cgroup is removed, its ->sibling.next is no longer
3880 * updated when its next sibling changes. CSS_RELEASED is set when
3881 * @pos is taken off list, at which time its next pointer is valid,
3882 * and, as releases are serialized, the one pointed to by the next
3883 * pointer is guaranteed to not have started release yet. This
3884 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3885 * critical section, the one pointed to by its next pointer is
3886 * guaranteed to not have finished its RCU grace period even if we
3887 * have dropped rcu_read_lock() inbetween iterations.
3888 *
3889 * If @pos has CSS_RELEASED set, its next pointer can't be
3890 * dereferenced; however, as each css is given a monotonically
3891 * increasing unique serial number and always appended to the
3892 * sibling list, the next one can be found by walking the parent's
3893 * children until the first css with higher serial number than
3894 * @pos's. While this path can be slower, it happens iff iteration
3895 * races against release and the race window is very small.
3896 */
3897 if (!pos) {
3898 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3899 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3900 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3901 } else {
3902 list_for_each_entry_rcu(next, &parent->children, sibling)
3903 if (next->serial_nr > pos->serial_nr)
3904 break;
3905 }
3906
3907 /*
3908 * @next, if not pointing to the head, can be dereferenced and is
3909 * the next sibling.
3910 */
3911 if (&next->sibling != &parent->children)
3912 return next;
3913 return NULL;
3914}
3915
3916/**
3917 * css_next_descendant_pre - find the next descendant for pre-order walk
3918 * @pos: the current position (%NULL to initiate traversal)
3919 * @root: css whose descendants to walk
3920 *
3921 * To be used by css_for_each_descendant_pre(). Find the next descendant
3922 * to visit for pre-order traversal of @root's descendants. @root is
3923 * included in the iteration and the first node to be visited.
3924 *
3925 * While this function requires cgroup_mutex or RCU read locking, it
3926 * doesn't require the whole traversal to be contained in a single critical
3927 * section. This function will return the correct next descendant as long
3928 * as both @pos and @root are accessible and @pos is a descendant of @root.
3929 *
3930 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3931 * css which finished ->css_online() is guaranteed to be visible in the
3932 * future iterations and will stay visible until the last reference is put.
3933 * A css which hasn't finished ->css_online() or already finished
3934 * ->css_offline() may show up during traversal. It's each subsystem's
3935 * responsibility to synchronize against on/offlining.
3936 */
3937struct cgroup_subsys_state *
3938css_next_descendant_pre(struct cgroup_subsys_state *pos,
3939 struct cgroup_subsys_state *root)
3940{
3941 struct cgroup_subsys_state *next;
3942
3943 cgroup_assert_mutex_or_rcu_locked();
3944
3945 /* if first iteration, visit @root */
3946 if (!pos)
3947 return root;
3948
3949 /* visit the first child if exists */
3950 next = css_next_child(NULL, pos);
3951 if (next)
3952 return next;
3953
3954 /* no child, visit my or the closest ancestor's next sibling */
3955 while (pos != root) {
3956 next = css_next_child(pos, pos->parent);
3957 if (next)
3958 return next;
3959 pos = pos->parent;
3960 }
3961
3962 return NULL;
3963}
3964
3965/**
3966 * css_rightmost_descendant - return the rightmost descendant of a css
3967 * @pos: css of interest
3968 *
3969 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3970 * is returned. This can be used during pre-order traversal to skip
3971 * subtree of @pos.
3972 *
3973 * While this function requires cgroup_mutex or RCU read locking, it
3974 * doesn't require the whole traversal to be contained in a single critical
3975 * section. This function will return the correct rightmost descendant as
3976 * long as @pos is accessible.
3977 */
3978struct cgroup_subsys_state *
3979css_rightmost_descendant(struct cgroup_subsys_state *pos)
3980{
3981 struct cgroup_subsys_state *last, *tmp;
3982
3983 cgroup_assert_mutex_or_rcu_locked();
3984
3985 do {
3986 last = pos;
3987 /* ->prev isn't RCU safe, walk ->next till the end */
3988 pos = NULL;
3989 css_for_each_child(tmp, last)
3990 pos = tmp;
3991 } while (pos);
3992
3993 return last;
3994}
3995
3996static struct cgroup_subsys_state *
3997css_leftmost_descendant(struct cgroup_subsys_state *pos)
3998{
3999 struct cgroup_subsys_state *last;
4000
4001 do {
4002 last = pos;
4003 pos = css_next_child(NULL, pos);
4004 } while (pos);
4005
4006 return last;
4007}
4008
4009/**
4010 * css_next_descendant_post - find the next descendant for post-order walk
4011 * @pos: the current position (%NULL to initiate traversal)
4012 * @root: css whose descendants to walk
4013 *
4014 * To be used by css_for_each_descendant_post(). Find the next descendant
4015 * to visit for post-order traversal of @root's descendants. @root is
4016 * included in the iteration and the last node to be visited.
4017 *
4018 * While this function requires cgroup_mutex or RCU read locking, it
4019 * doesn't require the whole traversal to be contained in a single critical
4020 * section. This function will return the correct next descendant as long
4021 * as both @pos and @cgroup are accessible and @pos is a descendant of
4022 * @cgroup.
4023 *
4024 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4025 * css which finished ->css_online() is guaranteed to be visible in the
4026 * future iterations and will stay visible until the last reference is put.
4027 * A css which hasn't finished ->css_online() or already finished
4028 * ->css_offline() may show up during traversal. It's each subsystem's
4029 * responsibility to synchronize against on/offlining.
4030 */
4031struct cgroup_subsys_state *
4032css_next_descendant_post(struct cgroup_subsys_state *pos,
4033 struct cgroup_subsys_state *root)
4034{
4035 struct cgroup_subsys_state *next;
4036
4037 cgroup_assert_mutex_or_rcu_locked();
4038
4039 /* if first iteration, visit leftmost descendant which may be @root */
4040 if (!pos)
4041 return css_leftmost_descendant(root);
4042
4043 /* if we visited @root, we're done */
4044 if (pos == root)
4045 return NULL;
4046
4047 /* if there's an unvisited sibling, visit its leftmost descendant */
4048 next = css_next_child(pos, pos->parent);
4049 if (next)
4050 return css_leftmost_descendant(next);
4051
4052 /* no sibling left, visit parent */
4053 return pos->parent;
4054}
4055
4056/**
4057 * css_has_online_children - does a css have online children
4058 * @css: the target css
4059 *
4060 * Returns %true if @css has any online children; otherwise, %false. This
4061 * function can be called from any context but the caller is responsible
4062 * for synchronizing against on/offlining as necessary.
4063 */
4064bool css_has_online_children(struct cgroup_subsys_state *css)
4065{
4066 struct cgroup_subsys_state *child;
4067 bool ret = false;
4068
4069 rcu_read_lock();
4070 css_for_each_child(child, css) {
4071 if (child->flags & CSS_ONLINE) {
4072 ret = true;
4073 break;
4074 }
4075 }
4076 rcu_read_unlock();
4077 return ret;
4078}
4079
4080/**
4081 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4082 * @it: the iterator to advance
4083 *
4084 * Advance @it to the next css_set to walk.
4085 */
4086static void css_task_iter_advance_css_set(struct css_task_iter *it)
4087{
4088 struct list_head *l = it->cset_pos;
4089 struct cgrp_cset_link *link;
4090 struct css_set *cset;
4091
4092 lockdep_assert_held(&css_set_lock);
4093
4094 /* Advance to the next non-empty css_set */
4095 do {
4096 l = l->next;
4097 if (l == it->cset_head) {
4098 it->cset_pos = NULL;
4099 it->task_pos = NULL;
4100 return;
4101 }
4102
4103 if (it->ss) {
4104 cset = container_of(l, struct css_set,
4105 e_cset_node[it->ss->id]);
4106 } else {
4107 link = list_entry(l, struct cgrp_cset_link, cset_link);
4108 cset = link->cset;
4109 }
4110 } while (!css_set_populated(cset));
4111
4112 it->cset_pos = l;
4113
4114 if (!list_empty(&cset->tasks))
4115 it->task_pos = cset->tasks.next;
4116 else
4117 it->task_pos = cset->mg_tasks.next;
4118
4119 it->tasks_head = &cset->tasks;
4120 it->mg_tasks_head = &cset->mg_tasks;
4121
4122 /*
4123 * We don't keep css_sets locked across iteration steps and thus
4124 * need to take steps to ensure that iteration can be resumed after
4125 * the lock is re-acquired. Iteration is performed at two levels -
4126 * css_sets and tasks in them.
4127 *
4128 * Once created, a css_set never leaves its cgroup lists, so a
4129 * pinned css_set is guaranteed to stay put and we can resume
4130 * iteration afterwards.
4131 *
4132 * Tasks may leave @cset across iteration steps. This is resolved
4133 * by registering each iterator with the css_set currently being
4134 * walked and making css_set_move_task() advance iterators whose
4135 * next task is leaving.
4136 */
4137 if (it->cur_cset) {
4138 list_del(&it->iters_node);
4139 put_css_set_locked(it->cur_cset);
4140 }
4141 get_css_set(cset);
4142 it->cur_cset = cset;
4143 list_add(&it->iters_node, &cset->task_iters);
4144}
4145
4146static void css_task_iter_advance(struct css_task_iter *it)
4147{
4148 struct list_head *l = it->task_pos;
4149
4150 lockdep_assert_held(&css_set_lock);
4151 WARN_ON_ONCE(!l);
4152
4153 /*
4154 * Advance iterator to find next entry. cset->tasks is consumed
4155 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4156 * next cset.
4157 */
4158 l = l->next;
4159
4160 if (l == it->tasks_head)
4161 l = it->mg_tasks_head->next;
4162
4163 if (l == it->mg_tasks_head)
4164 css_task_iter_advance_css_set(it);
4165 else
4166 it->task_pos = l;
4167}
4168
4169/**
4170 * css_task_iter_start - initiate task iteration
4171 * @css: the css to walk tasks of
4172 * @it: the task iterator to use
4173 *
4174 * Initiate iteration through the tasks of @css. The caller can call
4175 * css_task_iter_next() to walk through the tasks until the function
4176 * returns NULL. On completion of iteration, css_task_iter_end() must be
4177 * called.
4178 */
4179void css_task_iter_start(struct cgroup_subsys_state *css,
4180 struct css_task_iter *it)
4181{
4182 /* no one should try to iterate before mounting cgroups */
4183 WARN_ON_ONCE(!use_task_css_set_links);
4184
4185 memset(it, 0, sizeof(*it));
4186
4187 spin_lock_bh(&css_set_lock);
4188
4189 it->ss = css->ss;
4190
4191 if (it->ss)
4192 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4193 else
4194 it->cset_pos = &css->cgroup->cset_links;
4195
4196 it->cset_head = it->cset_pos;
4197
4198 css_task_iter_advance_css_set(it);
4199
4200 spin_unlock_bh(&css_set_lock);
4201}
4202
4203/**
4204 * css_task_iter_next - return the next task for the iterator
4205 * @it: the task iterator being iterated
4206 *
4207 * The "next" function for task iteration. @it should have been
4208 * initialized via css_task_iter_start(). Returns NULL when the iteration
4209 * reaches the end.
4210 */
4211struct task_struct *css_task_iter_next(struct css_task_iter *it)
4212{
4213 if (it->cur_task) {
4214 put_task_struct(it->cur_task);
4215 it->cur_task = NULL;
4216 }
4217
4218 spin_lock_bh(&css_set_lock);
4219
4220 if (it->task_pos) {
4221 it->cur_task = list_entry(it->task_pos, struct task_struct,
4222 cg_list);
4223 get_task_struct(it->cur_task);
4224 css_task_iter_advance(it);
4225 }
4226
4227 spin_unlock_bh(&css_set_lock);
4228
4229 return it->cur_task;
4230}
4231
4232/**
4233 * css_task_iter_end - finish task iteration
4234 * @it: the task iterator to finish
4235 *
4236 * Finish task iteration started by css_task_iter_start().
4237 */
4238void css_task_iter_end(struct css_task_iter *it)
4239{
4240 if (it->cur_cset) {
4241 spin_lock_bh(&css_set_lock);
4242 list_del(&it->iters_node);
4243 put_css_set_locked(it->cur_cset);
4244 spin_unlock_bh(&css_set_lock);
4245 }
4246
4247 if (it->cur_task)
4248 put_task_struct(it->cur_task);
4249}
4250
4251/**
4252 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4253 * @to: cgroup to which the tasks will be moved
4254 * @from: cgroup in which the tasks currently reside
4255 *
4256 * Locking rules between cgroup_post_fork() and the migration path
4257 * guarantee that, if a task is forking while being migrated, the new child
4258 * is guaranteed to be either visible in the source cgroup after the
4259 * parent's migration is complete or put into the target cgroup. No task
4260 * can slip out of migration through forking.
4261 */
4262int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4263{
4264 LIST_HEAD(preloaded_csets);
4265 struct cgrp_cset_link *link;
4266 struct css_task_iter it;
4267 struct task_struct *task;
4268 int ret;
4269
4270 if (!cgroup_may_migrate_to(to))
4271 return -EBUSY;
4272
4273 mutex_lock(&cgroup_mutex);
4274
4275 /* all tasks in @from are being moved, all csets are source */
4276 spin_lock_bh(&css_set_lock);
4277 list_for_each_entry(link, &from->cset_links, cset_link)
4278 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4279 spin_unlock_bh(&css_set_lock);
4280
4281 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4282 if (ret)
4283 goto out_err;
4284
4285 /*
4286 * Migrate tasks one-by-one until @from is empty. This fails iff
4287 * ->can_attach() fails.
4288 */
4289 do {
4290 css_task_iter_start(&from->self, &it);
4291 task = css_task_iter_next(&it);
4292 if (task)
4293 get_task_struct(task);
4294 css_task_iter_end(&it);
4295
4296 if (task) {
4297 ret = cgroup_migrate(task, false, to->root);
4298 put_task_struct(task);
4299 }
4300 } while (task && !ret);
4301out_err:
4302 cgroup_migrate_finish(&preloaded_csets);
4303 mutex_unlock(&cgroup_mutex);
4304 return ret;
4305}
4306
4307/*
4308 * Stuff for reading the 'tasks'/'procs' files.
4309 *
4310 * Reading this file can return large amounts of data if a cgroup has
4311 * *lots* of attached tasks. So it may need several calls to read(),
4312 * but we cannot guarantee that the information we produce is correct
4313 * unless we produce it entirely atomically.
4314 *
4315 */
4316
4317/* which pidlist file are we talking about? */
4318enum cgroup_filetype {
4319 CGROUP_FILE_PROCS,
4320 CGROUP_FILE_TASKS,
4321};
4322
4323/*
4324 * A pidlist is a list of pids that virtually represents the contents of one
4325 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4326 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4327 * to the cgroup.
4328 */
4329struct cgroup_pidlist {
4330 /*
4331 * used to find which pidlist is wanted. doesn't change as long as
4332 * this particular list stays in the list.
4333 */
4334 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4335 /* array of xids */
4336 pid_t *list;
4337 /* how many elements the above list has */
4338 int length;
4339 /* each of these stored in a list by its cgroup */
4340 struct list_head links;
4341 /* pointer to the cgroup we belong to, for list removal purposes */
4342 struct cgroup *owner;
4343 /* for delayed destruction */
4344 struct delayed_work destroy_dwork;
4345};
4346
4347/*
4348 * The following two functions "fix" the issue where there are more pids
4349 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4350 * TODO: replace with a kernel-wide solution to this problem
4351 */
4352#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4353static void *pidlist_allocate(int count)
4354{
4355 if (PIDLIST_TOO_LARGE(count))
4356 return vmalloc(count * sizeof(pid_t));
4357 else
4358 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4359}
4360
4361static void pidlist_free(void *p)
4362{
4363 kvfree(p);
4364}
4365
4366/*
4367 * Used to destroy all pidlists lingering waiting for destroy timer. None
4368 * should be left afterwards.
4369 */
4370static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4371{
4372 struct cgroup_pidlist *l, *tmp_l;
4373
4374 mutex_lock(&cgrp->pidlist_mutex);
4375 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4376 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4377 mutex_unlock(&cgrp->pidlist_mutex);
4378
4379 flush_workqueue(cgroup_pidlist_destroy_wq);
4380 BUG_ON(!list_empty(&cgrp->pidlists));
4381}
4382
4383static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4384{
4385 struct delayed_work *dwork = to_delayed_work(work);
4386 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4387 destroy_dwork);
4388 struct cgroup_pidlist *tofree = NULL;
4389
4390 mutex_lock(&l->owner->pidlist_mutex);
4391
4392 /*
4393 * Destroy iff we didn't get queued again. The state won't change
4394 * as destroy_dwork can only be queued while locked.
4395 */
4396 if (!delayed_work_pending(dwork)) {
4397 list_del(&l->links);
4398 pidlist_free(l->list);
4399 put_pid_ns(l->key.ns);
4400 tofree = l;
4401 }
4402
4403 mutex_unlock(&l->owner->pidlist_mutex);
4404 kfree(tofree);
4405}
4406
4407/*
4408 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4409 * Returns the number of unique elements.
4410 */
4411static int pidlist_uniq(pid_t *list, int length)
4412{
4413 int src, dest = 1;
4414
4415 /*
4416 * we presume the 0th element is unique, so i starts at 1. trivial
4417 * edge cases first; no work needs to be done for either
4418 */
4419 if (length == 0 || length == 1)
4420 return length;
4421 /* src and dest walk down the list; dest counts unique elements */
4422 for (src = 1; src < length; src++) {
4423 /* find next unique element */
4424 while (list[src] == list[src-1]) {
4425 src++;
4426 if (src == length)
4427 goto after;
4428 }
4429 /* dest always points to where the next unique element goes */
4430 list[dest] = list[src];
4431 dest++;
4432 }
4433after:
4434 return dest;
4435}
4436
4437/*
4438 * The two pid files - task and cgroup.procs - guaranteed that the result
4439 * is sorted, which forced this whole pidlist fiasco. As pid order is
4440 * different per namespace, each namespace needs differently sorted list,
4441 * making it impossible to use, for example, single rbtree of member tasks
4442 * sorted by task pointer. As pidlists can be fairly large, allocating one
4443 * per open file is dangerous, so cgroup had to implement shared pool of
4444 * pidlists keyed by cgroup and namespace.
4445 *
4446 * All this extra complexity was caused by the original implementation
4447 * committing to an entirely unnecessary property. In the long term, we
4448 * want to do away with it. Explicitly scramble sort order if on the
4449 * default hierarchy so that no such expectation exists in the new
4450 * interface.
4451 *
4452 * Scrambling is done by swapping every two consecutive bits, which is
4453 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4454 */
4455static pid_t pid_fry(pid_t pid)
4456{
4457 unsigned a = pid & 0x55555555;
4458 unsigned b = pid & 0xAAAAAAAA;
4459
4460 return (a << 1) | (b >> 1);
4461}
4462
4463static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4464{
4465 if (cgroup_on_dfl(cgrp))
4466 return pid_fry(pid);
4467 else
4468 return pid;
4469}
4470
4471static int cmppid(const void *a, const void *b)
4472{
4473 return *(pid_t *)a - *(pid_t *)b;
4474}
4475
4476static int fried_cmppid(const void *a, const void *b)
4477{
4478 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4479}
4480
4481static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4482 enum cgroup_filetype type)
4483{
4484 struct cgroup_pidlist *l;
4485 /* don't need task_nsproxy() if we're looking at ourself */
4486 struct pid_namespace *ns = task_active_pid_ns(current);
4487
4488 lockdep_assert_held(&cgrp->pidlist_mutex);
4489
4490 list_for_each_entry(l, &cgrp->pidlists, links)
4491 if (l->key.type == type && l->key.ns == ns)
4492 return l;
4493 return NULL;
4494}
4495
4496/*
4497 * find the appropriate pidlist for our purpose (given procs vs tasks)
4498 * returns with the lock on that pidlist already held, and takes care
4499 * of the use count, or returns NULL with no locks held if we're out of
4500 * memory.
4501 */
4502static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4503 enum cgroup_filetype type)
4504{
4505 struct cgroup_pidlist *l;
4506
4507 lockdep_assert_held(&cgrp->pidlist_mutex);
4508
4509 l = cgroup_pidlist_find(cgrp, type);
4510 if (l)
4511 return l;
4512
4513 /* entry not found; create a new one */
4514 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4515 if (!l)
4516 return l;
4517
4518 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4519 l->key.type = type;
4520 /* don't need task_nsproxy() if we're looking at ourself */
4521 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4522 l->owner = cgrp;
4523 list_add(&l->links, &cgrp->pidlists);
4524 return l;
4525}
4526
4527/*
4528 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4529 */
4530static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4531 struct cgroup_pidlist **lp)
4532{
4533 pid_t *array;
4534 int length;
4535 int pid, n = 0; /* used for populating the array */
4536 struct css_task_iter it;
4537 struct task_struct *tsk;
4538 struct cgroup_pidlist *l;
4539
4540 lockdep_assert_held(&cgrp->pidlist_mutex);
4541
4542 /*
4543 * If cgroup gets more users after we read count, we won't have
4544 * enough space - tough. This race is indistinguishable to the
4545 * caller from the case that the additional cgroup users didn't
4546 * show up until sometime later on.
4547 */
4548 length = cgroup_task_count(cgrp);
4549 array = pidlist_allocate(length);
4550 if (!array)
4551 return -ENOMEM;
4552 /* now, populate the array */
4553 css_task_iter_start(&cgrp->self, &it);
4554 while ((tsk = css_task_iter_next(&it))) {
4555 if (unlikely(n == length))
4556 break;
4557 /* get tgid or pid for procs or tasks file respectively */
4558 if (type == CGROUP_FILE_PROCS)
4559 pid = task_tgid_vnr(tsk);
4560 else
4561 pid = task_pid_vnr(tsk);
4562 if (pid > 0) /* make sure to only use valid results */
4563 array[n++] = pid;
4564 }
4565 css_task_iter_end(&it);
4566 length = n;
4567 /* now sort & (if procs) strip out duplicates */
4568 if (cgroup_on_dfl(cgrp))
4569 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4570 else
4571 sort(array, length, sizeof(pid_t), cmppid, NULL);
4572 if (type == CGROUP_FILE_PROCS)
4573 length = pidlist_uniq(array, length);
4574
4575 l = cgroup_pidlist_find_create(cgrp, type);
4576 if (!l) {
4577 pidlist_free(array);
4578 return -ENOMEM;
4579 }
4580
4581 /* store array, freeing old if necessary */
4582 pidlist_free(l->list);
4583 l->list = array;
4584 l->length = length;
4585 *lp = l;
4586 return 0;
4587}
4588
4589/**
4590 * cgroupstats_build - build and fill cgroupstats
4591 * @stats: cgroupstats to fill information into
4592 * @dentry: A dentry entry belonging to the cgroup for which stats have
4593 * been requested.
4594 *
4595 * Build and fill cgroupstats so that taskstats can export it to user
4596 * space.
4597 */
4598int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4599{
4600 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4601 struct cgroup *cgrp;
4602 struct css_task_iter it;
4603 struct task_struct *tsk;
4604
4605 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4606 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4607 kernfs_type(kn) != KERNFS_DIR)
4608 return -EINVAL;
4609
4610 mutex_lock(&cgroup_mutex);
4611
4612 /*
4613 * We aren't being called from kernfs and there's no guarantee on
4614 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4615 * @kn->priv is RCU safe. Let's do the RCU dancing.
4616 */
4617 rcu_read_lock();
4618 cgrp = rcu_dereference(kn->priv);
4619 if (!cgrp || cgroup_is_dead(cgrp)) {
4620 rcu_read_unlock();
4621 mutex_unlock(&cgroup_mutex);
4622 return -ENOENT;
4623 }
4624 rcu_read_unlock();
4625
4626 css_task_iter_start(&cgrp->self, &it);
4627 while ((tsk = css_task_iter_next(&it))) {
4628 switch (tsk->state) {
4629 case TASK_RUNNING:
4630 stats->nr_running++;
4631 break;
4632 case TASK_INTERRUPTIBLE:
4633 stats->nr_sleeping++;
4634 break;
4635 case TASK_UNINTERRUPTIBLE:
4636 stats->nr_uninterruptible++;
4637 break;
4638 case TASK_STOPPED:
4639 stats->nr_stopped++;
4640 break;
4641 default:
4642 if (delayacct_is_task_waiting_on_io(tsk))
4643 stats->nr_io_wait++;
4644 break;
4645 }
4646 }
4647 css_task_iter_end(&it);
4648
4649 mutex_unlock(&cgroup_mutex);
4650 return 0;
4651}
4652
4653
4654/*
4655 * seq_file methods for the tasks/procs files. The seq_file position is the
4656 * next pid to display; the seq_file iterator is a pointer to the pid
4657 * in the cgroup->l->list array.
4658 */
4659
4660static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4661{
4662 /*
4663 * Initially we receive a position value that corresponds to
4664 * one more than the last pid shown (or 0 on the first call or
4665 * after a seek to the start). Use a binary-search to find the
4666 * next pid to display, if any
4667 */
4668 struct kernfs_open_file *of = s->private;
4669 struct cgroup *cgrp = seq_css(s)->cgroup;
4670 struct cgroup_pidlist *l;
4671 enum cgroup_filetype type = seq_cft(s)->private;
4672 int index = 0, pid = *pos;
4673 int *iter, ret;
4674
4675 mutex_lock(&cgrp->pidlist_mutex);
4676
4677 /*
4678 * !NULL @of->priv indicates that this isn't the first start()
4679 * after open. If the matching pidlist is around, we can use that.
4680 * Look for it. Note that @of->priv can't be used directly. It
4681 * could already have been destroyed.
4682 */
4683 if (of->priv)
4684 of->priv = cgroup_pidlist_find(cgrp, type);
4685
4686 /*
4687 * Either this is the first start() after open or the matching
4688 * pidlist has been destroyed inbetween. Create a new one.
4689 */
4690 if (!of->priv) {
4691 ret = pidlist_array_load(cgrp, type,
4692 (struct cgroup_pidlist **)&of->priv);
4693 if (ret)
4694 return ERR_PTR(ret);
4695 }
4696 l = of->priv;
4697
4698 if (pid) {
4699 int end = l->length;
4700
4701 while (index < end) {
4702 int mid = (index + end) / 2;
4703 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4704 index = mid;
4705 break;
4706 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4707 index = mid + 1;
4708 else
4709 end = mid;
4710 }
4711 }
4712 /* If we're off the end of the array, we're done */
4713 if (index >= l->length)
4714 return NULL;
4715 /* Update the abstract position to be the actual pid that we found */
4716 iter = l->list + index;
4717 *pos = cgroup_pid_fry(cgrp, *iter);
4718 return iter;
4719}
4720
4721static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4722{
4723 struct kernfs_open_file *of = s->private;
4724 struct cgroup_pidlist *l = of->priv;
4725
4726 if (l)
4727 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4728 CGROUP_PIDLIST_DESTROY_DELAY);
4729 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4730}
4731
4732static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4733{
4734 struct kernfs_open_file *of = s->private;
4735 struct cgroup_pidlist *l = of->priv;
4736 pid_t *p = v;
4737 pid_t *end = l->list + l->length;
4738 /*
4739 * Advance to the next pid in the array. If this goes off the
4740 * end, we're done
4741 */
4742 p++;
4743 if (p >= end) {
4744 return NULL;
4745 } else {
4746 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4747 return p;
4748 }
4749}
4750
4751static int cgroup_pidlist_show(struct seq_file *s, void *v)
4752{
4753 seq_printf(s, "%d\n", *(int *)v);
4754
4755 return 0;
4756}
4757
4758static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4759 struct cftype *cft)
4760{
4761 return notify_on_release(css->cgroup);
4762}
4763
4764static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4765 struct cftype *cft, u64 val)
4766{
4767 if (val)
4768 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4769 else
4770 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4771 return 0;
4772}
4773
4774static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4775 struct cftype *cft)
4776{
4777 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4778}
4779
4780static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4781 struct cftype *cft, u64 val)
4782{
4783 if (val)
4784 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4785 else
4786 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4787 return 0;
4788}
4789
4790/* cgroup core interface files for the default hierarchy */
4791static struct cftype cgroup_dfl_base_files[] = {
4792 {
4793 .name = "cgroup.procs",
4794 .file_offset = offsetof(struct cgroup, procs_file),
4795 .seq_start = cgroup_pidlist_start,
4796 .seq_next = cgroup_pidlist_next,
4797 .seq_stop = cgroup_pidlist_stop,
4798 .seq_show = cgroup_pidlist_show,
4799 .private = CGROUP_FILE_PROCS,
4800 .write = cgroup_procs_write,
4801 },
4802 {
4803 .name = "cgroup.controllers",
4804 .seq_show = cgroup_controllers_show,
4805 },
4806 {
4807 .name = "cgroup.subtree_control",
4808 .seq_show = cgroup_subtree_control_show,
4809 .write = cgroup_subtree_control_write,
4810 },
4811 {
4812 .name = "cgroup.events",
4813 .flags = CFTYPE_NOT_ON_ROOT,
4814 .file_offset = offsetof(struct cgroup, events_file),
4815 .seq_show = cgroup_events_show,
4816 },
4817 { } /* terminate */
4818};
4819
4820/* cgroup core interface files for the legacy hierarchies */
4821static struct cftype cgroup_legacy_base_files[] = {
4822 {
4823 .name = "cgroup.procs",
4824 .seq_start = cgroup_pidlist_start,
4825 .seq_next = cgroup_pidlist_next,
4826 .seq_stop = cgroup_pidlist_stop,
4827 .seq_show = cgroup_pidlist_show,
4828 .private = CGROUP_FILE_PROCS,
4829 .write = cgroup_procs_write,
4830 },
4831 {
4832 .name = "cgroup.clone_children",
4833 .read_u64 = cgroup_clone_children_read,
4834 .write_u64 = cgroup_clone_children_write,
4835 },
4836 {
4837 .name = "cgroup.sane_behavior",
4838 .flags = CFTYPE_ONLY_ON_ROOT,
4839 .seq_show = cgroup_sane_behavior_show,
4840 },
4841 {
4842 .name = "tasks",
4843 .seq_start = cgroup_pidlist_start,
4844 .seq_next = cgroup_pidlist_next,
4845 .seq_stop = cgroup_pidlist_stop,
4846 .seq_show = cgroup_pidlist_show,
4847 .private = CGROUP_FILE_TASKS,
4848 .write = cgroup_tasks_write,
4849 },
4850 {
4851 .name = "notify_on_release",
4852 .read_u64 = cgroup_read_notify_on_release,
4853 .write_u64 = cgroup_write_notify_on_release,
4854 },
4855 {
4856 .name = "release_agent",
4857 .flags = CFTYPE_ONLY_ON_ROOT,
4858 .seq_show = cgroup_release_agent_show,
4859 .write = cgroup_release_agent_write,
4860 .max_write_len = PATH_MAX - 1,
4861 },
4862 { } /* terminate */
4863};
4864
4865/*
4866 * css destruction is four-stage process.
4867 *
4868 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4869 * Implemented in kill_css().
4870 *
4871 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4872 * and thus css_tryget_online() is guaranteed to fail, the css can be
4873 * offlined by invoking offline_css(). After offlining, the base ref is
4874 * put. Implemented in css_killed_work_fn().
4875 *
4876 * 3. When the percpu_ref reaches zero, the only possible remaining
4877 * accessors are inside RCU read sections. css_release() schedules the
4878 * RCU callback.
4879 *
4880 * 4. After the grace period, the css can be freed. Implemented in
4881 * css_free_work_fn().
4882 *
4883 * It is actually hairier because both step 2 and 4 require process context
4884 * and thus involve punting to css->destroy_work adding two additional
4885 * steps to the already complex sequence.
4886 */
4887static void css_free_work_fn(struct work_struct *work)
4888{
4889 struct cgroup_subsys_state *css =
4890 container_of(work, struct cgroup_subsys_state, destroy_work);
4891 struct cgroup_subsys *ss = css->ss;
4892 struct cgroup *cgrp = css->cgroup;
4893
4894 percpu_ref_exit(&css->refcnt);
4895
4896 if (ss) {
4897 /* css free path */
4898 struct cgroup_subsys_state *parent = css->parent;
4899 int id = css->id;
4900
4901 ss->css_free(css);
4902 cgroup_idr_remove(&ss->css_idr, id);
4903 cgroup_put(cgrp);
4904
4905 if (parent)
4906 css_put(parent);
4907 } else {
4908 /* cgroup free path */
4909 atomic_dec(&cgrp->root->nr_cgrps);
4910 cgroup_pidlist_destroy_all(cgrp);
4911 cancel_work_sync(&cgrp->release_agent_work);
4912
4913 if (cgroup_parent(cgrp)) {
4914 /*
4915 * We get a ref to the parent, and put the ref when
4916 * this cgroup is being freed, so it's guaranteed
4917 * that the parent won't be destroyed before its
4918 * children.
4919 */
4920 cgroup_put(cgroup_parent(cgrp));
4921 kernfs_put(cgrp->kn);
4922 kfree(cgrp);
4923 } else {
4924 /*
4925 * This is root cgroup's refcnt reaching zero,
4926 * which indicates that the root should be
4927 * released.
4928 */
4929 cgroup_destroy_root(cgrp->root);
4930 }
4931 }
4932}
4933
4934static void css_free_rcu_fn(struct rcu_head *rcu_head)
4935{
4936 struct cgroup_subsys_state *css =
4937 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4938
4939 INIT_WORK(&css->destroy_work, css_free_work_fn);
4940 queue_work(cgroup_destroy_wq, &css->destroy_work);
4941}
4942
4943static void css_release_work_fn(struct work_struct *work)
4944{
4945 struct cgroup_subsys_state *css =
4946 container_of(work, struct cgroup_subsys_state, destroy_work);
4947 struct cgroup_subsys *ss = css->ss;
4948 struct cgroup *cgrp = css->cgroup;
4949
4950 mutex_lock(&cgroup_mutex);
4951
4952 css->flags |= CSS_RELEASED;
4953 list_del_rcu(&css->sibling);
4954
4955 if (ss) {
4956 /* css release path */
4957 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4958 if (ss->css_released)
4959 ss->css_released(css);
4960 } else {
4961 /* cgroup release path */
4962 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4963 cgrp->id = -1;
4964
4965 /*
4966 * There are two control paths which try to determine
4967 * cgroup from dentry without going through kernfs -
4968 * cgroupstats_build() and css_tryget_online_from_dir().
4969 * Those are supported by RCU protecting clearing of
4970 * cgrp->kn->priv backpointer.
4971 */
4972 if (cgrp->kn)
4973 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4974 NULL);
4975 }
4976
4977 mutex_unlock(&cgroup_mutex);
4978
4979 call_rcu(&css->rcu_head, css_free_rcu_fn);
4980}
4981
4982static void css_release(struct percpu_ref *ref)
4983{
4984 struct cgroup_subsys_state *css =
4985 container_of(ref, struct cgroup_subsys_state, refcnt);
4986
4987 INIT_WORK(&css->destroy_work, css_release_work_fn);
4988 queue_work(cgroup_destroy_wq, &css->destroy_work);
4989}
4990
4991static void init_and_link_css(struct cgroup_subsys_state *css,
4992 struct cgroup_subsys *ss, struct cgroup *cgrp)
4993{
4994 lockdep_assert_held(&cgroup_mutex);
4995
4996 cgroup_get(cgrp);
4997
4998 memset(css, 0, sizeof(*css));
4999 css->cgroup = cgrp;
5000 css->ss = ss;
5001 INIT_LIST_HEAD(&css->sibling);
5002 INIT_LIST_HEAD(&css->children);
5003 css->serial_nr = css_serial_nr_next++;
5004 atomic_set(&css->online_cnt, 0);
5005
5006 if (cgroup_parent(cgrp)) {
5007 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5008 css_get(css->parent);
5009 }
5010
5011 BUG_ON(cgroup_css(cgrp, ss));
5012}
5013
5014/* invoke ->css_online() on a new CSS and mark it online if successful */
5015static int online_css(struct cgroup_subsys_state *css)
5016{
5017 struct cgroup_subsys *ss = css->ss;
5018 int ret = 0;
5019
5020 lockdep_assert_held(&cgroup_mutex);
5021
5022 if (ss->css_online)
5023 ret = ss->css_online(css);
5024 if (!ret) {
5025 css->flags |= CSS_ONLINE;
5026 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5027
5028 atomic_inc(&css->online_cnt);
5029 if (css->parent)
5030 atomic_inc(&css->parent->online_cnt);
5031 }
5032 return ret;
5033}
5034
5035/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5036static void offline_css(struct cgroup_subsys_state *css)
5037{
5038 struct cgroup_subsys *ss = css->ss;
5039
5040 lockdep_assert_held(&cgroup_mutex);
5041
5042 if (!(css->flags & CSS_ONLINE))
5043 return;
5044
5045 if (ss->css_reset)
5046 ss->css_reset(css);
5047
5048 if (ss->css_offline)
5049 ss->css_offline(css);
5050
5051 css->flags &= ~CSS_ONLINE;
5052 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5053
5054 wake_up_all(&css->cgroup->offline_waitq);
5055}
5056
5057/**
5058 * css_create - create a cgroup_subsys_state
5059 * @cgrp: the cgroup new css will be associated with
5060 * @ss: the subsys of new css
5061 *
5062 * Create a new css associated with @cgrp - @ss pair. On success, the new
5063 * css is online and installed in @cgrp. This function doesn't create the
5064 * interface files. Returns 0 on success, -errno on failure.
5065 */
5066static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5067 struct cgroup_subsys *ss)
5068{
5069 struct cgroup *parent = cgroup_parent(cgrp);
5070 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5071 struct cgroup_subsys_state *css;
5072 int err;
5073
5074 lockdep_assert_held(&cgroup_mutex);
5075
5076 css = ss->css_alloc(parent_css);
5077 if (IS_ERR(css))
5078 return css;
5079
5080 init_and_link_css(css, ss, cgrp);
5081
5082 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5083 if (err)
5084 goto err_free_css;
5085
5086 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5087 if (err < 0)
5088 goto err_free_percpu_ref;
5089 css->id = err;
5090
5091 /* @css is ready to be brought online now, make it visible */
5092 list_add_tail_rcu(&css->sibling, &parent_css->children);
5093 cgroup_idr_replace(&ss->css_idr, css, css->id);
5094
5095 err = online_css(css);
5096 if (err)
5097 goto err_list_del;
5098
5099 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5100 cgroup_parent(parent)) {
5101 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5102 current->comm, current->pid, ss->name);
5103 if (!strcmp(ss->name, "memory"))
5104 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5105 ss->warned_broken_hierarchy = true;
5106 }
5107
5108 return css;
5109
5110err_list_del:
5111 list_del_rcu(&css->sibling);
5112 cgroup_idr_remove(&ss->css_idr, css->id);
5113err_free_percpu_ref:
5114 percpu_ref_exit(&css->refcnt);
5115err_free_css:
5116 call_rcu(&css->rcu_head, css_free_rcu_fn);
5117 return ERR_PTR(err);
5118}
5119
5120static struct cgroup *cgroup_create(struct cgroup *parent)
5121{
5122 struct cgroup_root *root = parent->root;
5123 struct cgroup *cgrp, *tcgrp;
5124 int level = parent->level + 1;
5125 int ret;
5126
5127 /* allocate the cgroup and its ID, 0 is reserved for the root */
5128 cgrp = kzalloc(sizeof(*cgrp) +
5129 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5130 if (!cgrp)
5131 return ERR_PTR(-ENOMEM);
5132
5133 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5134 if (ret)
5135 goto out_free_cgrp;
5136
5137 /*
5138 * Temporarily set the pointer to NULL, so idr_find() won't return
5139 * a half-baked cgroup.
5140 */
5141 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5142 if (cgrp->id < 0) {
5143 ret = -ENOMEM;
5144 goto out_cancel_ref;
5145 }
5146
5147 init_cgroup_housekeeping(cgrp);
5148
5149 cgrp->self.parent = &parent->self;
5150 cgrp->root = root;
5151 cgrp->level = level;
5152
5153 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5154 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5155
5156 if (notify_on_release(parent))
5157 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5158
5159 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5160 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5161
5162 cgrp->self.serial_nr = css_serial_nr_next++;
5163
5164 /* allocation complete, commit to creation */
5165 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5166 atomic_inc(&root->nr_cgrps);
5167 cgroup_get(parent);
5168
5169 /*
5170 * @cgrp is now fully operational. If something fails after this
5171 * point, it'll be released via the normal destruction path.
5172 */
5173 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5174
5175 /*
5176 * On the default hierarchy, a child doesn't automatically inherit
5177 * subtree_control from the parent. Each is configured manually.
5178 */
5179 if (!cgroup_on_dfl(cgrp))
5180 cgrp->subtree_control = cgroup_control(cgrp);
5181
5182 cgroup_propagate_control(cgrp);
5183
5184 /* @cgrp doesn't have dir yet so the following will only create csses */
5185 ret = cgroup_apply_control_enable(cgrp);
5186 if (ret)
5187 goto out_destroy;
5188
5189 return cgrp;
5190
5191out_cancel_ref:
5192 percpu_ref_exit(&cgrp->self.refcnt);
5193out_free_cgrp:
5194 kfree(cgrp);
5195 return ERR_PTR(ret);
5196out_destroy:
5197 cgroup_destroy_locked(cgrp);
5198 return ERR_PTR(ret);
5199}
5200
5201static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5202 umode_t mode)
5203{
5204 struct cgroup *parent, *cgrp;
5205 struct kernfs_node *kn;
5206 int ret;
5207
5208 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5209 if (strchr(name, '\n'))
5210 return -EINVAL;
5211
5212 parent = cgroup_kn_lock_live(parent_kn, false);
5213 if (!parent)
5214 return -ENODEV;
5215
5216 cgrp = cgroup_create(parent);
5217 if (IS_ERR(cgrp)) {
5218 ret = PTR_ERR(cgrp);
5219 goto out_unlock;
5220 }
5221
5222 /* create the directory */
5223 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5224 if (IS_ERR(kn)) {
5225 ret = PTR_ERR(kn);
5226 goto out_destroy;
5227 }
5228 cgrp->kn = kn;
5229
5230 /*
5231 * This extra ref will be put in cgroup_free_fn() and guarantees
5232 * that @cgrp->kn is always accessible.
5233 */
5234 kernfs_get(kn);
5235
5236 ret = cgroup_kn_set_ugid(kn);
5237 if (ret)
5238 goto out_destroy;
5239
5240 ret = css_populate_dir(&cgrp->self);
5241 if (ret)
5242 goto out_destroy;
5243
5244 ret = cgroup_apply_control_enable(cgrp);
5245 if (ret)
5246 goto out_destroy;
5247
5248 /* let's create and online css's */
5249 kernfs_activate(kn);
5250
5251 ret = 0;
5252 goto out_unlock;
5253
5254out_destroy:
5255 cgroup_destroy_locked(cgrp);
5256out_unlock:
5257 cgroup_kn_unlock(parent_kn);
5258 return ret;
5259}
5260
5261/*
5262 * This is called when the refcnt of a css is confirmed to be killed.
5263 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5264 * initate destruction and put the css ref from kill_css().
5265 */
5266static void css_killed_work_fn(struct work_struct *work)
5267{
5268 struct cgroup_subsys_state *css =
5269 container_of(work, struct cgroup_subsys_state, destroy_work);
5270
5271 mutex_lock(&cgroup_mutex);
5272
5273 do {
5274 offline_css(css);
5275 css_put(css);
5276 /* @css can't go away while we're holding cgroup_mutex */
5277 css = css->parent;
5278 } while (css && atomic_dec_and_test(&css->online_cnt));
5279
5280 mutex_unlock(&cgroup_mutex);
5281}
5282
5283/* css kill confirmation processing requires process context, bounce */
5284static void css_killed_ref_fn(struct percpu_ref *ref)
5285{
5286 struct cgroup_subsys_state *css =
5287 container_of(ref, struct cgroup_subsys_state, refcnt);
5288
5289 if (atomic_dec_and_test(&css->online_cnt)) {
5290 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5291 queue_work(cgroup_destroy_wq, &css->destroy_work);
5292 }
5293}
5294
5295/**
5296 * kill_css - destroy a css
5297 * @css: css to destroy
5298 *
5299 * This function initiates destruction of @css by removing cgroup interface
5300 * files and putting its base reference. ->css_offline() will be invoked
5301 * asynchronously once css_tryget_online() is guaranteed to fail and when
5302 * the reference count reaches zero, @css will be released.
5303 */
5304static void kill_css(struct cgroup_subsys_state *css)
5305{
5306 lockdep_assert_held(&cgroup_mutex);
5307
5308 /*
5309 * This must happen before css is disassociated with its cgroup.
5310 * See seq_css() for details.
5311 */
5312 css_clear_dir(css);
5313
5314 /*
5315 * Killing would put the base ref, but we need to keep it alive
5316 * until after ->css_offline().
5317 */
5318 css_get(css);
5319
5320 /*
5321 * cgroup core guarantees that, by the time ->css_offline() is
5322 * invoked, no new css reference will be given out via
5323 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5324 * proceed to offlining css's because percpu_ref_kill() doesn't
5325 * guarantee that the ref is seen as killed on all CPUs on return.
5326 *
5327 * Use percpu_ref_kill_and_confirm() to get notifications as each
5328 * css is confirmed to be seen as killed on all CPUs.
5329 */
5330 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5331}
5332
5333/**
5334 * cgroup_destroy_locked - the first stage of cgroup destruction
5335 * @cgrp: cgroup to be destroyed
5336 *
5337 * css's make use of percpu refcnts whose killing latency shouldn't be
5338 * exposed to userland and are RCU protected. Also, cgroup core needs to
5339 * guarantee that css_tryget_online() won't succeed by the time
5340 * ->css_offline() is invoked. To satisfy all the requirements,
5341 * destruction is implemented in the following two steps.
5342 *
5343 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5344 * userland visible parts and start killing the percpu refcnts of
5345 * css's. Set up so that the next stage will be kicked off once all
5346 * the percpu refcnts are confirmed to be killed.
5347 *
5348 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5349 * rest of destruction. Once all cgroup references are gone, the
5350 * cgroup is RCU-freed.
5351 *
5352 * This function implements s1. After this step, @cgrp is gone as far as
5353 * the userland is concerned and a new cgroup with the same name may be
5354 * created. As cgroup doesn't care about the names internally, this
5355 * doesn't cause any problem.
5356 */
5357static int cgroup_destroy_locked(struct cgroup *cgrp)
5358 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5359{
5360 struct cgroup_subsys_state *css;
5361 struct cgrp_cset_link *link;
5362 int ssid;
5363
5364 lockdep_assert_held(&cgroup_mutex);
5365
5366 /*
5367 * Only migration can raise populated from zero and we're already
5368 * holding cgroup_mutex.
5369 */
5370 if (cgroup_is_populated(cgrp))
5371 return -EBUSY;
5372
5373 /*
5374 * Make sure there's no live children. We can't test emptiness of
5375 * ->self.children as dead children linger on it while being
5376 * drained; otherwise, "rmdir parent/child parent" may fail.
5377 */
5378 if (css_has_online_children(&cgrp->self))
5379 return -EBUSY;
5380
5381 /*
5382 * Mark @cgrp and the associated csets dead. The former prevents
5383 * further task migration and child creation by disabling
5384 * cgroup_lock_live_group(). The latter makes the csets ignored by
5385 * the migration path.
5386 */
5387 cgrp->self.flags &= ~CSS_ONLINE;
5388
5389 spin_lock_bh(&css_set_lock);
5390 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5391 link->cset->dead = true;
5392 spin_unlock_bh(&css_set_lock);
5393
5394 /* initiate massacre of all css's */
5395 for_each_css(css, ssid, cgrp)
5396 kill_css(css);
5397
5398 /*
5399 * Remove @cgrp directory along with the base files. @cgrp has an
5400 * extra ref on its kn.
5401 */
5402 kernfs_remove(cgrp->kn);
5403
5404 check_for_release(cgroup_parent(cgrp));
5405
5406 /* put the base reference */
5407 percpu_ref_kill(&cgrp->self.refcnt);
5408
5409 return 0;
5410};
5411
5412static int cgroup_rmdir(struct kernfs_node *kn)
5413{
5414 struct cgroup *cgrp;
5415 int ret = 0;
5416
5417 cgrp = cgroup_kn_lock_live(kn, false);
5418 if (!cgrp)
5419 return 0;
5420
5421 ret = cgroup_destroy_locked(cgrp);
5422
5423 cgroup_kn_unlock(kn);
5424 return ret;
5425}
5426
5427static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5428 .remount_fs = cgroup_remount,
5429 .show_options = cgroup_show_options,
5430 .mkdir = cgroup_mkdir,
5431 .rmdir = cgroup_rmdir,
5432 .rename = cgroup_rename,
5433};
5434
5435static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5436{
5437 struct cgroup_subsys_state *css;
5438
5439 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5440
5441 mutex_lock(&cgroup_mutex);
5442
5443 idr_init(&ss->css_idr);
5444 INIT_LIST_HEAD(&ss->cfts);
5445
5446 /* Create the root cgroup state for this subsystem */
5447 ss->root = &cgrp_dfl_root;
5448 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5449 /* We don't handle early failures gracefully */
5450 BUG_ON(IS_ERR(css));
5451 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5452
5453 /*
5454 * Root csses are never destroyed and we can't initialize
5455 * percpu_ref during early init. Disable refcnting.
5456 */
5457 css->flags |= CSS_NO_REF;
5458
5459 if (early) {
5460 /* allocation can't be done safely during early init */
5461 css->id = 1;
5462 } else {
5463 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5464 BUG_ON(css->id < 0);
5465 }
5466
5467 /* Update the init_css_set to contain a subsys
5468 * pointer to this state - since the subsystem is
5469 * newly registered, all tasks and hence the
5470 * init_css_set is in the subsystem's root cgroup. */
5471 init_css_set.subsys[ss->id] = css;
5472
5473 have_fork_callback |= (bool)ss->fork << ss->id;
5474 have_exit_callback |= (bool)ss->exit << ss->id;
5475 have_free_callback |= (bool)ss->free << ss->id;
5476 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5477
5478 /* At system boot, before all subsystems have been
5479 * registered, no tasks have been forked, so we don't
5480 * need to invoke fork callbacks here. */
5481 BUG_ON(!list_empty(&init_task.tasks));
5482
5483 BUG_ON(online_css(css));
5484
5485 mutex_unlock(&cgroup_mutex);
5486}
5487
5488/**
5489 * cgroup_init_early - cgroup initialization at system boot
5490 *
5491 * Initialize cgroups at system boot, and initialize any
5492 * subsystems that request early init.
5493 */
5494int __init cgroup_init_early(void)
5495{
5496 static struct cgroup_sb_opts __initdata opts;
5497 struct cgroup_subsys *ss;
5498 int i;
5499
5500 init_cgroup_root(&cgrp_dfl_root, &opts);
5501 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5502
5503 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5504
5505 for_each_subsys(ss, i) {
5506 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5507 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5508 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5509 ss->id, ss->name);
5510 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5511 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5512
5513 ss->id = i;
5514 ss->name = cgroup_subsys_name[i];
5515 if (!ss->legacy_name)
5516 ss->legacy_name = cgroup_subsys_name[i];
5517
5518 if (ss->early_init)
5519 cgroup_init_subsys(ss, true);
5520 }
5521 return 0;
5522}
5523
5524static u16 cgroup_disable_mask __initdata;
5525
5526/**
5527 * cgroup_init - cgroup initialization
5528 *
5529 * Register cgroup filesystem and /proc file, and initialize
5530 * any subsystems that didn't request early init.
5531 */
5532int __init cgroup_init(void)
5533{
5534 struct cgroup_subsys *ss;
5535 int ssid;
5536
5537 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5538 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5539 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5540 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5541
5542 get_user_ns(init_cgroup_ns.user_ns);
5543
5544 mutex_lock(&cgroup_mutex);
5545
5546 /*
5547 * Add init_css_set to the hash table so that dfl_root can link to
5548 * it during init.
5549 */
5550 hash_add(css_set_table, &init_css_set.hlist,
5551 css_set_hash(init_css_set.subsys));
5552
5553 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5554
5555 mutex_unlock(&cgroup_mutex);
5556
5557 for_each_subsys(ss, ssid) {
5558 if (ss->early_init) {
5559 struct cgroup_subsys_state *css =
5560 init_css_set.subsys[ss->id];
5561
5562 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5563 GFP_KERNEL);
5564 BUG_ON(css->id < 0);
5565 } else {
5566 cgroup_init_subsys(ss, false);
5567 }
5568
5569 list_add_tail(&init_css_set.e_cset_node[ssid],
5570 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5571
5572 /*
5573 * Setting dfl_root subsys_mask needs to consider the
5574 * disabled flag and cftype registration needs kmalloc,
5575 * both of which aren't available during early_init.
5576 */
5577 if (cgroup_disable_mask & (1 << ssid)) {
5578 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5579 printk(KERN_INFO "Disabling %s control group subsystem\n",
5580 ss->name);
5581 continue;
5582 }
5583
5584 if (cgroup_ssid_no_v1(ssid))
5585 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5586 ss->name);
5587
5588 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5589
5590 if (ss->implicit_on_dfl)
5591 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5592 else if (!ss->dfl_cftypes)
5593 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5594
5595 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5596 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5597 } else {
5598 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5599 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5600 }
5601
5602 if (ss->bind)
5603 ss->bind(init_css_set.subsys[ssid]);
5604 }
5605
5606 /* init_css_set.subsys[] has been updated, re-hash */
5607 hash_del(&init_css_set.hlist);
5608 hash_add(css_set_table, &init_css_set.hlist,
5609 css_set_hash(init_css_set.subsys));
5610
5611 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5612 WARN_ON(register_filesystem(&cgroup_fs_type));
5613 WARN_ON(register_filesystem(&cgroup2_fs_type));
5614 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5615
5616 return 0;
5617}
5618
5619static int __init cgroup_wq_init(void)
5620{
5621 /*
5622 * There isn't much point in executing destruction path in
5623 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5624 * Use 1 for @max_active.
5625 *
5626 * We would prefer to do this in cgroup_init() above, but that
5627 * is called before init_workqueues(): so leave this until after.
5628 */
5629 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5630 BUG_ON(!cgroup_destroy_wq);
5631
5632 /*
5633 * Used to destroy pidlists and separate to serve as flush domain.
5634 * Cap @max_active to 1 too.
5635 */
5636 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5637 0, 1);
5638 BUG_ON(!cgroup_pidlist_destroy_wq);
5639
5640 return 0;
5641}
5642core_initcall(cgroup_wq_init);
5643
5644/*
5645 * proc_cgroup_show()
5646 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5647 * - Used for /proc/<pid>/cgroup.
5648 */
5649int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5650 struct pid *pid, struct task_struct *tsk)
5651{
5652 char *buf, *path;
5653 int retval;
5654 struct cgroup_root *root;
5655
5656 retval = -ENOMEM;
5657 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5658 if (!buf)
5659 goto out;
5660
5661 mutex_lock(&cgroup_mutex);
5662 spin_lock_bh(&css_set_lock);
5663
5664 for_each_root(root) {
5665 struct cgroup_subsys *ss;
5666 struct cgroup *cgrp;
5667 int ssid, count = 0;
5668
5669 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5670 continue;
5671
5672 seq_printf(m, "%d:", root->hierarchy_id);
5673 if (root != &cgrp_dfl_root)
5674 for_each_subsys(ss, ssid)
5675 if (root->subsys_mask & (1 << ssid))
5676 seq_printf(m, "%s%s", count++ ? "," : "",
5677 ss->legacy_name);
5678 if (strlen(root->name))
5679 seq_printf(m, "%sname=%s", count ? "," : "",
5680 root->name);
5681 seq_putc(m, ':');
5682
5683 cgrp = task_cgroup_from_root(tsk, root);
5684
5685 /*
5686 * On traditional hierarchies, all zombie tasks show up as
5687 * belonging to the root cgroup. On the default hierarchy,
5688 * while a zombie doesn't show up in "cgroup.procs" and
5689 * thus can't be migrated, its /proc/PID/cgroup keeps
5690 * reporting the cgroup it belonged to before exiting. If
5691 * the cgroup is removed before the zombie is reaped,
5692 * " (deleted)" is appended to the cgroup path.
5693 */
5694 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5695 path = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5696 current->nsproxy->cgroup_ns);
5697 if (!path) {
5698 retval = -ENAMETOOLONG;
5699 goto out_unlock;
5700 }
5701 } else {
5702 path = "/";
5703 }
5704
5705 seq_puts(m, path);
5706
5707 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5708 seq_puts(m, " (deleted)\n");
5709 else
5710 seq_putc(m, '\n');
5711 }
5712
5713 retval = 0;
5714out_unlock:
5715 spin_unlock_bh(&css_set_lock);
5716 mutex_unlock(&cgroup_mutex);
5717 kfree(buf);
5718out:
5719 return retval;
5720}
5721
5722/* Display information about each subsystem and each hierarchy */
5723static int proc_cgroupstats_show(struct seq_file *m, void *v)
5724{
5725 struct cgroup_subsys *ss;
5726 int i;
5727
5728 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5729 /*
5730 * ideally we don't want subsystems moving around while we do this.
5731 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5732 * subsys/hierarchy state.
5733 */
5734 mutex_lock(&cgroup_mutex);
5735
5736 for_each_subsys(ss, i)
5737 seq_printf(m, "%s\t%d\t%d\t%d\n",
5738 ss->legacy_name, ss->root->hierarchy_id,
5739 atomic_read(&ss->root->nr_cgrps),
5740 cgroup_ssid_enabled(i));
5741
5742 mutex_unlock(&cgroup_mutex);
5743 return 0;
5744}
5745
5746static int cgroupstats_open(struct inode *inode, struct file *file)
5747{
5748 return single_open(file, proc_cgroupstats_show, NULL);
5749}
5750
5751static const struct file_operations proc_cgroupstats_operations = {
5752 .open = cgroupstats_open,
5753 .read = seq_read,
5754 .llseek = seq_lseek,
5755 .release = single_release,
5756};
5757
5758/**
5759 * cgroup_fork - initialize cgroup related fields during copy_process()
5760 * @child: pointer to task_struct of forking parent process.
5761 *
5762 * A task is associated with the init_css_set until cgroup_post_fork()
5763 * attaches it to the parent's css_set. Empty cg_list indicates that
5764 * @child isn't holding reference to its css_set.
5765 */
5766void cgroup_fork(struct task_struct *child)
5767{
5768 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5769 INIT_LIST_HEAD(&child->cg_list);
5770}
5771
5772/**
5773 * cgroup_can_fork - called on a new task before the process is exposed
5774 * @child: the task in question.
5775 *
5776 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5777 * returns an error, the fork aborts with that error code. This allows for
5778 * a cgroup subsystem to conditionally allow or deny new forks.
5779 */
5780int cgroup_can_fork(struct task_struct *child)
5781{
5782 struct cgroup_subsys *ss;
5783 int i, j, ret;
5784
5785 do_each_subsys_mask(ss, i, have_canfork_callback) {
5786 ret = ss->can_fork(child);
5787 if (ret)
5788 goto out_revert;
5789 } while_each_subsys_mask();
5790
5791 return 0;
5792
5793out_revert:
5794 for_each_subsys(ss, j) {
5795 if (j >= i)
5796 break;
5797 if (ss->cancel_fork)
5798 ss->cancel_fork(child);
5799 }
5800
5801 return ret;
5802}
5803
5804/**
5805 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5806 * @child: the task in question
5807 *
5808 * This calls the cancel_fork() callbacks if a fork failed *after*
5809 * cgroup_can_fork() succeded.
5810 */
5811void cgroup_cancel_fork(struct task_struct *child)
5812{
5813 struct cgroup_subsys *ss;
5814 int i;
5815
5816 for_each_subsys(ss, i)
5817 if (ss->cancel_fork)
5818 ss->cancel_fork(child);
5819}
5820
5821/**
5822 * cgroup_post_fork - called on a new task after adding it to the task list
5823 * @child: the task in question
5824 *
5825 * Adds the task to the list running through its css_set if necessary and
5826 * call the subsystem fork() callbacks. Has to be after the task is
5827 * visible on the task list in case we race with the first call to
5828 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5829 * list.
5830 */
5831void cgroup_post_fork(struct task_struct *child)
5832{
5833 struct cgroup_subsys *ss;
5834 int i;
5835
5836 /*
5837 * This may race against cgroup_enable_task_cg_lists(). As that
5838 * function sets use_task_css_set_links before grabbing
5839 * tasklist_lock and we just went through tasklist_lock to add
5840 * @child, it's guaranteed that either we see the set
5841 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5842 * @child during its iteration.
5843 *
5844 * If we won the race, @child is associated with %current's
5845 * css_set. Grabbing css_set_lock guarantees both that the
5846 * association is stable, and, on completion of the parent's
5847 * migration, @child is visible in the source of migration or
5848 * already in the destination cgroup. This guarantee is necessary
5849 * when implementing operations which need to migrate all tasks of
5850 * a cgroup to another.
5851 *
5852 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5853 * will remain in init_css_set. This is safe because all tasks are
5854 * in the init_css_set before cg_links is enabled and there's no
5855 * operation which transfers all tasks out of init_css_set.
5856 */
5857 if (use_task_css_set_links) {
5858 struct css_set *cset;
5859
5860 spin_lock_bh(&css_set_lock);
5861 cset = task_css_set(current);
5862 if (list_empty(&child->cg_list)) {
5863 get_css_set(cset);
5864 css_set_move_task(child, NULL, cset, false);
5865 }
5866 spin_unlock_bh(&css_set_lock);
5867 }
5868
5869 /*
5870 * Call ss->fork(). This must happen after @child is linked on
5871 * css_set; otherwise, @child might change state between ->fork()
5872 * and addition to css_set.
5873 */
5874 do_each_subsys_mask(ss, i, have_fork_callback) {
5875 ss->fork(child);
5876 } while_each_subsys_mask();
5877}
5878
5879/**
5880 * cgroup_exit - detach cgroup from exiting task
5881 * @tsk: pointer to task_struct of exiting process
5882 *
5883 * Description: Detach cgroup from @tsk and release it.
5884 *
5885 * Note that cgroups marked notify_on_release force every task in
5886 * them to take the global cgroup_mutex mutex when exiting.
5887 * This could impact scaling on very large systems. Be reluctant to
5888 * use notify_on_release cgroups where very high task exit scaling
5889 * is required on large systems.
5890 *
5891 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5892 * call cgroup_exit() while the task is still competent to handle
5893 * notify_on_release(), then leave the task attached to the root cgroup in
5894 * each hierarchy for the remainder of its exit. No need to bother with
5895 * init_css_set refcnting. init_css_set never goes away and we can't race
5896 * with migration path - PF_EXITING is visible to migration path.
5897 */
5898void cgroup_exit(struct task_struct *tsk)
5899{
5900 struct cgroup_subsys *ss;
5901 struct css_set *cset;
5902 int i;
5903
5904 /*
5905 * Unlink from @tsk from its css_set. As migration path can't race
5906 * with us, we can check css_set and cg_list without synchronization.
5907 */
5908 cset = task_css_set(tsk);
5909
5910 if (!list_empty(&tsk->cg_list)) {
5911 spin_lock_bh(&css_set_lock);
5912 css_set_move_task(tsk, cset, NULL, false);
5913 spin_unlock_bh(&css_set_lock);
5914 } else {
5915 get_css_set(cset);
5916 }
5917
5918 /* see cgroup_post_fork() for details */
5919 do_each_subsys_mask(ss, i, have_exit_callback) {
5920 ss->exit(tsk);
5921 } while_each_subsys_mask();
5922}
5923
5924void cgroup_free(struct task_struct *task)
5925{
5926 struct css_set *cset = task_css_set(task);
5927 struct cgroup_subsys *ss;
5928 int ssid;
5929
5930 do_each_subsys_mask(ss, ssid, have_free_callback) {
5931 ss->free(task);
5932 } while_each_subsys_mask();
5933
5934 put_css_set(cset);
5935}
5936
5937static void check_for_release(struct cgroup *cgrp)
5938{
5939 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5940 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5941 schedule_work(&cgrp->release_agent_work);
5942}
5943
5944/*
5945 * Notify userspace when a cgroup is released, by running the
5946 * configured release agent with the name of the cgroup (path
5947 * relative to the root of cgroup file system) as the argument.
5948 *
5949 * Most likely, this user command will try to rmdir this cgroup.
5950 *
5951 * This races with the possibility that some other task will be
5952 * attached to this cgroup before it is removed, or that some other
5953 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5954 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5955 * unused, and this cgroup will be reprieved from its death sentence,
5956 * to continue to serve a useful existence. Next time it's released,
5957 * we will get notified again, if it still has 'notify_on_release' set.
5958 *
5959 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5960 * means only wait until the task is successfully execve()'d. The
5961 * separate release agent task is forked by call_usermodehelper(),
5962 * then control in this thread returns here, without waiting for the
5963 * release agent task. We don't bother to wait because the caller of
5964 * this routine has no use for the exit status of the release agent
5965 * task, so no sense holding our caller up for that.
5966 */
5967static void cgroup_release_agent(struct work_struct *work)
5968{
5969 struct cgroup *cgrp =
5970 container_of(work, struct cgroup, release_agent_work);
5971 char *pathbuf = NULL, *agentbuf = NULL, *path;
5972 char *argv[3], *envp[3];
5973
5974 mutex_lock(&cgroup_mutex);
5975
5976 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5977 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5978 if (!pathbuf || !agentbuf)
5979 goto out;
5980
5981 spin_lock_bh(&css_set_lock);
5982 path = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
5983 spin_unlock_bh(&css_set_lock);
5984 if (!path)
5985 goto out;
5986
5987 argv[0] = agentbuf;
5988 argv[1] = path;
5989 argv[2] = NULL;
5990
5991 /* minimal command environment */
5992 envp[0] = "HOME=/";
5993 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5994 envp[2] = NULL;
5995
5996 mutex_unlock(&cgroup_mutex);
5997 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5998 goto out_free;
5999out:
6000 mutex_unlock(&cgroup_mutex);
6001out_free:
6002 kfree(agentbuf);
6003 kfree(pathbuf);
6004}
6005
6006static int __init cgroup_disable(char *str)
6007{
6008 struct cgroup_subsys *ss;
6009 char *token;
6010 int i;
6011
6012 while ((token = strsep(&str, ",")) != NULL) {
6013 if (!*token)
6014 continue;
6015
6016 for_each_subsys(ss, i) {
6017 if (strcmp(token, ss->name) &&
6018 strcmp(token, ss->legacy_name))
6019 continue;
6020 cgroup_disable_mask |= 1 << i;
6021 }
6022 }
6023 return 1;
6024}
6025__setup("cgroup_disable=", cgroup_disable);
6026
6027static int __init cgroup_no_v1(char *str)
6028{
6029 struct cgroup_subsys *ss;
6030 char *token;
6031 int i;
6032
6033 while ((token = strsep(&str, ",")) != NULL) {
6034 if (!*token)
6035 continue;
6036
6037 if (!strcmp(token, "all")) {
6038 cgroup_no_v1_mask = U16_MAX;
6039 break;
6040 }
6041
6042 for_each_subsys(ss, i) {
6043 if (strcmp(token, ss->name) &&
6044 strcmp(token, ss->legacy_name))
6045 continue;
6046
6047 cgroup_no_v1_mask |= 1 << i;
6048 }
6049 }
6050 return 1;
6051}
6052__setup("cgroup_no_v1=", cgroup_no_v1);
6053
6054/**
6055 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6056 * @dentry: directory dentry of interest
6057 * @ss: subsystem of interest
6058 *
6059 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6060 * to get the corresponding css and return it. If such css doesn't exist
6061 * or can't be pinned, an ERR_PTR value is returned.
6062 */
6063struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6064 struct cgroup_subsys *ss)
6065{
6066 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6067 struct file_system_type *s_type = dentry->d_sb->s_type;
6068 struct cgroup_subsys_state *css = NULL;
6069 struct cgroup *cgrp;
6070
6071 /* is @dentry a cgroup dir? */
6072 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6073 !kn || kernfs_type(kn) != KERNFS_DIR)
6074 return ERR_PTR(-EBADF);
6075
6076 rcu_read_lock();
6077
6078 /*
6079 * This path doesn't originate from kernfs and @kn could already
6080 * have been or be removed at any point. @kn->priv is RCU
6081 * protected for this access. See css_release_work_fn() for details.
6082 */
6083 cgrp = rcu_dereference(kn->priv);
6084 if (cgrp)
6085 css = cgroup_css(cgrp, ss);
6086
6087 if (!css || !css_tryget_online(css))
6088 css = ERR_PTR(-ENOENT);
6089
6090 rcu_read_unlock();
6091 return css;
6092}
6093
6094/**
6095 * css_from_id - lookup css by id
6096 * @id: the cgroup id
6097 * @ss: cgroup subsys to be looked into
6098 *
6099 * Returns the css if there's valid one with @id, otherwise returns NULL.
6100 * Should be called under rcu_read_lock().
6101 */
6102struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6103{
6104 WARN_ON_ONCE(!rcu_read_lock_held());
6105 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
6106}
6107
6108/**
6109 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6110 * @path: path on the default hierarchy
6111 *
6112 * Find the cgroup at @path on the default hierarchy, increment its
6113 * reference count and return it. Returns pointer to the found cgroup on
6114 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6115 * if @path points to a non-directory.
6116 */
6117struct cgroup *cgroup_get_from_path(const char *path)
6118{
6119 struct kernfs_node *kn;
6120 struct cgroup *cgrp;
6121
6122 mutex_lock(&cgroup_mutex);
6123
6124 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6125 if (kn) {
6126 if (kernfs_type(kn) == KERNFS_DIR) {
6127 cgrp = kn->priv;
6128 cgroup_get(cgrp);
6129 } else {
6130 cgrp = ERR_PTR(-ENOTDIR);
6131 }
6132 kernfs_put(kn);
6133 } else {
6134 cgrp = ERR_PTR(-ENOENT);
6135 }
6136
6137 mutex_unlock(&cgroup_mutex);
6138 return cgrp;
6139}
6140EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6141
6142/*
6143 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6144 * definition in cgroup-defs.h.
6145 */
6146#ifdef CONFIG_SOCK_CGROUP_DATA
6147
6148#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6149
6150DEFINE_SPINLOCK(cgroup_sk_update_lock);
6151static bool cgroup_sk_alloc_disabled __read_mostly;
6152
6153void cgroup_sk_alloc_disable(void)
6154{
6155 if (cgroup_sk_alloc_disabled)
6156 return;
6157 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6158 cgroup_sk_alloc_disabled = true;
6159}
6160
6161#else
6162
6163#define cgroup_sk_alloc_disabled false
6164
6165#endif
6166
6167void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6168{
6169 if (cgroup_sk_alloc_disabled)
6170 return;
6171
6172 rcu_read_lock();
6173
6174 while (true) {
6175 struct css_set *cset;
6176
6177 cset = task_css_set(current);
6178 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6179 skcd->val = (unsigned long)cset->dfl_cgrp;
6180 break;
6181 }
6182 cpu_relax();
6183 }
6184
6185 rcu_read_unlock();
6186}
6187
6188void cgroup_sk_free(struct sock_cgroup_data *skcd)
6189{
6190 cgroup_put(sock_cgroup_ptr(skcd));
6191}
6192
6193#endif /* CONFIG_SOCK_CGROUP_DATA */
6194
6195/* cgroup namespaces */
6196
6197static struct cgroup_namespace *alloc_cgroup_ns(void)
6198{
6199 struct cgroup_namespace *new_ns;
6200 int ret;
6201
6202 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6203 if (!new_ns)
6204 return ERR_PTR(-ENOMEM);
6205 ret = ns_alloc_inum(&new_ns->ns);
6206 if (ret) {
6207 kfree(new_ns);
6208 return ERR_PTR(ret);
6209 }
6210 atomic_set(&new_ns->count, 1);
6211 new_ns->ns.ops = &cgroupns_operations;
6212 return new_ns;
6213}
6214
6215void free_cgroup_ns(struct cgroup_namespace *ns)
6216{
6217 put_css_set(ns->root_cset);
6218 put_user_ns(ns->user_ns);
6219 ns_free_inum(&ns->ns);
6220 kfree(ns);
6221}
6222EXPORT_SYMBOL(free_cgroup_ns);
6223
6224struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6225 struct user_namespace *user_ns,
6226 struct cgroup_namespace *old_ns)
6227{
6228 struct cgroup_namespace *new_ns;
6229 struct css_set *cset;
6230
6231 BUG_ON(!old_ns);
6232
6233 if (!(flags & CLONE_NEWCGROUP)) {
6234 get_cgroup_ns(old_ns);
6235 return old_ns;
6236 }
6237
6238 /* Allow only sysadmin to create cgroup namespace. */
6239 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6240 return ERR_PTR(-EPERM);
6241
6242 mutex_lock(&cgroup_mutex);
6243 spin_lock_bh(&css_set_lock);
6244
6245 cset = task_css_set(current);
6246 get_css_set(cset);
6247
6248 spin_unlock_bh(&css_set_lock);
6249 mutex_unlock(&cgroup_mutex);
6250
6251 new_ns = alloc_cgroup_ns();
6252 if (IS_ERR(new_ns)) {
6253 put_css_set(cset);
6254 return new_ns;
6255 }
6256
6257 new_ns->user_ns = get_user_ns(user_ns);
6258 new_ns->root_cset = cset;
6259
6260 return new_ns;
6261}
6262
6263static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6264{
6265 return container_of(ns, struct cgroup_namespace, ns);
6266}
6267
6268static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6269{
6270 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6271
6272 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6273 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6274 return -EPERM;
6275
6276 /* Don't need to do anything if we are attaching to our own cgroupns. */
6277 if (cgroup_ns == nsproxy->cgroup_ns)
6278 return 0;
6279
6280 get_cgroup_ns(cgroup_ns);
6281 put_cgroup_ns(nsproxy->cgroup_ns);
6282 nsproxy->cgroup_ns = cgroup_ns;
6283
6284 return 0;
6285}
6286
6287static struct ns_common *cgroupns_get(struct task_struct *task)
6288{
6289 struct cgroup_namespace *ns = NULL;
6290 struct nsproxy *nsproxy;
6291
6292 task_lock(task);
6293 nsproxy = task->nsproxy;
6294 if (nsproxy) {
6295 ns = nsproxy->cgroup_ns;
6296 get_cgroup_ns(ns);
6297 }
6298 task_unlock(task);
6299
6300 return ns ? &ns->ns : NULL;
6301}
6302
6303static void cgroupns_put(struct ns_common *ns)
6304{
6305 put_cgroup_ns(to_cg_ns(ns));
6306}
6307
6308const struct proc_ns_operations cgroupns_operations = {
6309 .name = "cgroup",
6310 .type = CLONE_NEWCGROUP,
6311 .get = cgroupns_get,
6312 .put = cgroupns_put,
6313 .install = cgroupns_install,
6314};
6315
6316static __init int cgroup_namespaces_init(void)
6317{
6318 return 0;
6319}
6320subsys_initcall(cgroup_namespaces_init);
6321
6322#ifdef CONFIG_CGROUP_DEBUG
6323static struct cgroup_subsys_state *
6324debug_css_alloc(struct cgroup_subsys_state *parent_css)
6325{
6326 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6327
6328 if (!css)
6329 return ERR_PTR(-ENOMEM);
6330
6331 return css;
6332}
6333
6334static void debug_css_free(struct cgroup_subsys_state *css)
6335{
6336 kfree(css);
6337}
6338
6339static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6340 struct cftype *cft)
6341{
6342 return cgroup_task_count(css->cgroup);
6343}
6344
6345static u64 current_css_set_read(struct cgroup_subsys_state *css,
6346 struct cftype *cft)
6347{
6348 return (u64)(unsigned long)current->cgroups;
6349}
6350
6351static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6352 struct cftype *cft)
6353{
6354 u64 count;
6355
6356 rcu_read_lock();
6357 count = atomic_read(&task_css_set(current)->refcount);
6358 rcu_read_unlock();
6359 return count;
6360}
6361
6362static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6363{
6364 struct cgrp_cset_link *link;
6365 struct css_set *cset;
6366 char *name_buf;
6367
6368 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6369 if (!name_buf)
6370 return -ENOMEM;
6371
6372 spin_lock_bh(&css_set_lock);
6373 rcu_read_lock();
6374 cset = rcu_dereference(current->cgroups);
6375 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6376 struct cgroup *c = link->cgrp;
6377
6378 cgroup_name(c, name_buf, NAME_MAX + 1);
6379 seq_printf(seq, "Root %d group %s\n",
6380 c->root->hierarchy_id, name_buf);
6381 }
6382 rcu_read_unlock();
6383 spin_unlock_bh(&css_set_lock);
6384 kfree(name_buf);
6385 return 0;
6386}
6387
6388#define MAX_TASKS_SHOWN_PER_CSS 25
6389static int cgroup_css_links_read(struct seq_file *seq, void *v)
6390{
6391 struct cgroup_subsys_state *css = seq_css(seq);
6392 struct cgrp_cset_link *link;
6393
6394 spin_lock_bh(&css_set_lock);
6395 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6396 struct css_set *cset = link->cset;
6397 struct task_struct *task;
6398 int count = 0;
6399
6400 seq_printf(seq, "css_set %p\n", cset);
6401
6402 list_for_each_entry(task, &cset->tasks, cg_list) {
6403 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6404 goto overflow;
6405 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6406 }
6407
6408 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6409 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6410 goto overflow;
6411 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6412 }
6413 continue;
6414 overflow:
6415 seq_puts(seq, " ...\n");
6416 }
6417 spin_unlock_bh(&css_set_lock);
6418 return 0;
6419}
6420
6421static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6422{
6423 return (!cgroup_is_populated(css->cgroup) &&
6424 !css_has_online_children(&css->cgroup->self));
6425}
6426
6427static struct cftype debug_files[] = {
6428 {
6429 .name = "taskcount",
6430 .read_u64 = debug_taskcount_read,
6431 },
6432
6433 {
6434 .name = "current_css_set",
6435 .read_u64 = current_css_set_read,
6436 },
6437
6438 {
6439 .name = "current_css_set_refcount",
6440 .read_u64 = current_css_set_refcount_read,
6441 },
6442
6443 {
6444 .name = "current_css_set_cg_links",
6445 .seq_show = current_css_set_cg_links_read,
6446 },
6447
6448 {
6449 .name = "cgroup_css_links",
6450 .seq_show = cgroup_css_links_read,
6451 },
6452
6453 {
6454 .name = "releasable",
6455 .read_u64 = releasable_read,
6456 },
6457
6458 { } /* terminate */
6459};
6460
6461struct cgroup_subsys debug_cgrp_subsys = {
6462 .css_alloc = debug_css_alloc,
6463 .css_free = debug_css_free,
6464 .legacy_cftypes = debug_files,
6465};
6466#endif /* CONFIG_CGROUP_DEBUG */