Loading...
Note: File does not exist in v3.15.
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 "cgroup-internal.h"
32
33#include <linux/cred.h>
34#include <linux/errno.h>
35#include <linux/init_task.h>
36#include <linux/kernel.h>
37#include <linux/magic.h>
38#include <linux/mutex.h>
39#include <linux/mount.h>
40#include <linux/pagemap.h>
41#include <linux/proc_fs.h>
42#include <linux/rcupdate.h>
43#include <linux/sched.h>
44#include <linux/sched/task.h>
45#include <linux/slab.h>
46#include <linux/spinlock.h>
47#include <linux/percpu-rwsem.h>
48#include <linux/string.h>
49#include <linux/hashtable.h>
50#include <linux/idr.h>
51#include <linux/kthread.h>
52#include <linux/atomic.h>
53#include <linux/cpuset.h>
54#include <linux/proc_ns.h>
55#include <linux/nsproxy.h>
56#include <linux/file.h>
57#include <net/sock.h>
58
59#define CREATE_TRACE_POINTS
60#include <trace/events/cgroup.h>
61
62#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
63 MAX_CFTYPE_NAME + 2)
64
65/*
66 * cgroup_mutex is the master lock. Any modification to cgroup or its
67 * hierarchy must be performed while holding it.
68 *
69 * css_set_lock protects task->cgroups pointer, the list of css_set
70 * objects, and the chain of tasks off each css_set.
71 *
72 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
73 * cgroup.h can use them for lockdep annotations.
74 */
75DEFINE_MUTEX(cgroup_mutex);
76DEFINE_SPINLOCK(css_set_lock);
77
78#ifdef CONFIG_PROVE_RCU
79EXPORT_SYMBOL_GPL(cgroup_mutex);
80EXPORT_SYMBOL_GPL(css_set_lock);
81#endif
82
83/*
84 * Protects cgroup_idr and css_idr so that IDs can be released without
85 * grabbing cgroup_mutex.
86 */
87static DEFINE_SPINLOCK(cgroup_idr_lock);
88
89/*
90 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
91 * against file removal/re-creation across css hiding.
92 */
93static DEFINE_SPINLOCK(cgroup_file_kn_lock);
94
95struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
96
97#define cgroup_assert_mutex_or_rcu_locked() \
98 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
99 !lockdep_is_held(&cgroup_mutex), \
100 "cgroup_mutex or RCU read lock required");
101
102/*
103 * cgroup destruction makes heavy use of work items and there can be a lot
104 * of concurrent destructions. Use a separate workqueue so that cgroup
105 * destruction work items don't end up filling up max_active of system_wq
106 * which may lead to deadlock.
107 */
108static struct workqueue_struct *cgroup_destroy_wq;
109
110/* generate an array of cgroup subsystem pointers */
111#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
112struct cgroup_subsys *cgroup_subsys[] = {
113#include <linux/cgroup_subsys.h>
114};
115#undef SUBSYS
116
117/* array of cgroup subsystem names */
118#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
119static const char *cgroup_subsys_name[] = {
120#include <linux/cgroup_subsys.h>
121};
122#undef SUBSYS
123
124/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
125#define SUBSYS(_x) \
126 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
127 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
128 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
129 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
130#include <linux/cgroup_subsys.h>
131#undef SUBSYS
132
133#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
134static struct static_key_true *cgroup_subsys_enabled_key[] = {
135#include <linux/cgroup_subsys.h>
136};
137#undef SUBSYS
138
139#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
140static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
141#include <linux/cgroup_subsys.h>
142};
143#undef SUBSYS
144
145static DEFINE_PER_CPU(struct cgroup_cpu_stat, cgrp_dfl_root_cpu_stat);
146
147/*
148 * The default hierarchy, reserved for the subsystems that are otherwise
149 * unattached - it never has more than a single cgroup, and all tasks are
150 * part of that cgroup.
151 */
152struct cgroup_root cgrp_dfl_root = { .cgrp.cpu_stat = &cgrp_dfl_root_cpu_stat };
153EXPORT_SYMBOL_GPL(cgrp_dfl_root);
154
155/*
156 * The default hierarchy always exists but is hidden until mounted for the
157 * first time. This is for backward compatibility.
158 */
159static bool cgrp_dfl_visible;
160
161/* some controllers are not supported in the default hierarchy */
162static u16 cgrp_dfl_inhibit_ss_mask;
163
164/* some controllers are implicitly enabled on the default hierarchy */
165static u16 cgrp_dfl_implicit_ss_mask;
166
167/* some controllers can be threaded on the default hierarchy */
168static u16 cgrp_dfl_threaded_ss_mask;
169
170/* The list of hierarchy roots */
171LIST_HEAD(cgroup_roots);
172static int cgroup_root_count;
173
174/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
175static DEFINE_IDR(cgroup_hierarchy_idr);
176
177/*
178 * Assign a monotonically increasing serial number to csses. It guarantees
179 * cgroups with bigger numbers are newer than those with smaller numbers.
180 * Also, as csses are always appended to the parent's ->children list, it
181 * guarantees that sibling csses are always sorted in the ascending serial
182 * number order on the list. Protected by cgroup_mutex.
183 */
184static u64 css_serial_nr_next = 1;
185
186/*
187 * These bitmasks identify subsystems with specific features to avoid
188 * having to do iterative checks repeatedly.
189 */
190static u16 have_fork_callback __read_mostly;
191static u16 have_exit_callback __read_mostly;
192static u16 have_free_callback __read_mostly;
193static u16 have_canfork_callback __read_mostly;
194
195/* cgroup namespace for init task */
196struct cgroup_namespace init_cgroup_ns = {
197 .count = REFCOUNT_INIT(2),
198 .user_ns = &init_user_ns,
199 .ns.ops = &cgroupns_operations,
200 .ns.inum = PROC_CGROUP_INIT_INO,
201 .root_cset = &init_css_set,
202};
203
204static struct file_system_type cgroup2_fs_type;
205static struct cftype cgroup_base_files[];
206
207static int cgroup_apply_control(struct cgroup *cgrp);
208static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
209static void css_task_iter_advance(struct css_task_iter *it);
210static int cgroup_destroy_locked(struct cgroup *cgrp);
211static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
212 struct cgroup_subsys *ss);
213static void css_release(struct percpu_ref *ref);
214static void kill_css(struct cgroup_subsys_state *css);
215static int cgroup_addrm_files(struct cgroup_subsys_state *css,
216 struct cgroup *cgrp, struct cftype cfts[],
217 bool is_add);
218
219/**
220 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
221 * @ssid: subsys ID of interest
222 *
223 * cgroup_subsys_enabled() can only be used with literal subsys names which
224 * is fine for individual subsystems but unsuitable for cgroup core. This
225 * is slower static_key_enabled() based test indexed by @ssid.
226 */
227bool cgroup_ssid_enabled(int ssid)
228{
229 if (CGROUP_SUBSYS_COUNT == 0)
230 return false;
231
232 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
233}
234
235/**
236 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
237 * @cgrp: the cgroup of interest
238 *
239 * The default hierarchy is the v2 interface of cgroup and this function
240 * can be used to test whether a cgroup is on the default hierarchy for
241 * cases where a subsystem should behave differnetly depending on the
242 * interface version.
243 *
244 * The set of behaviors which change on the default hierarchy are still
245 * being determined and the mount option is prefixed with __DEVEL__.
246 *
247 * List of changed behaviors:
248 *
249 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
250 * and "name" are disallowed.
251 *
252 * - When mounting an existing superblock, mount options should match.
253 *
254 * - Remount is disallowed.
255 *
256 * - rename(2) is disallowed.
257 *
258 * - "tasks" is removed. Everything should be at process granularity. Use
259 * "cgroup.procs" instead.
260 *
261 * - "cgroup.procs" is not sorted. pids will be unique unless they got
262 * recycled inbetween reads.
263 *
264 * - "release_agent" and "notify_on_release" are removed. Replacement
265 * notification mechanism will be implemented.
266 *
267 * - "cgroup.clone_children" is removed.
268 *
269 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
270 * and its descendants contain no task; otherwise, 1. The file also
271 * generates kernfs notification which can be monitored through poll and
272 * [di]notify when the value of the file changes.
273 *
274 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
275 * take masks of ancestors with non-empty cpus/mems, instead of being
276 * moved to an ancestor.
277 *
278 * - cpuset: a task can be moved into an empty cpuset, and again it takes
279 * masks of ancestors.
280 *
281 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
282 * is not created.
283 *
284 * - blkcg: blk-throttle becomes properly hierarchical.
285 *
286 * - debug: disallowed on the default hierarchy.
287 */
288bool cgroup_on_dfl(const struct cgroup *cgrp)
289{
290 return cgrp->root == &cgrp_dfl_root;
291}
292
293/* IDR wrappers which synchronize using cgroup_idr_lock */
294static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
295 gfp_t gfp_mask)
296{
297 int ret;
298
299 idr_preload(gfp_mask);
300 spin_lock_bh(&cgroup_idr_lock);
301 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
302 spin_unlock_bh(&cgroup_idr_lock);
303 idr_preload_end();
304 return ret;
305}
306
307static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
308{
309 void *ret;
310
311 spin_lock_bh(&cgroup_idr_lock);
312 ret = idr_replace(idr, ptr, id);
313 spin_unlock_bh(&cgroup_idr_lock);
314 return ret;
315}
316
317static void cgroup_idr_remove(struct idr *idr, int id)
318{
319 spin_lock_bh(&cgroup_idr_lock);
320 idr_remove(idr, id);
321 spin_unlock_bh(&cgroup_idr_lock);
322}
323
324static bool cgroup_has_tasks(struct cgroup *cgrp)
325{
326 return cgrp->nr_populated_csets;
327}
328
329bool cgroup_is_threaded(struct cgroup *cgrp)
330{
331 return cgrp->dom_cgrp != cgrp;
332}
333
334/* can @cgrp host both domain and threaded children? */
335static bool cgroup_is_mixable(struct cgroup *cgrp)
336{
337 /*
338 * Root isn't under domain level resource control exempting it from
339 * the no-internal-process constraint, so it can serve as a thread
340 * root and a parent of resource domains at the same time.
341 */
342 return !cgroup_parent(cgrp);
343}
344
345/* can @cgrp become a thread root? should always be true for a thread root */
346static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
347{
348 /* mixables don't care */
349 if (cgroup_is_mixable(cgrp))
350 return true;
351
352 /* domain roots can't be nested under threaded */
353 if (cgroup_is_threaded(cgrp))
354 return false;
355
356 /* can only have either domain or threaded children */
357 if (cgrp->nr_populated_domain_children)
358 return false;
359
360 /* and no domain controllers can be enabled */
361 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
362 return false;
363
364 return true;
365}
366
367/* is @cgrp root of a threaded subtree? */
368bool cgroup_is_thread_root(struct cgroup *cgrp)
369{
370 /* thread root should be a domain */
371 if (cgroup_is_threaded(cgrp))
372 return false;
373
374 /* a domain w/ threaded children is a thread root */
375 if (cgrp->nr_threaded_children)
376 return true;
377
378 /*
379 * A domain which has tasks and explicit threaded controllers
380 * enabled is a thread root.
381 */
382 if (cgroup_has_tasks(cgrp) &&
383 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
384 return true;
385
386 return false;
387}
388
389/* a domain which isn't connected to the root w/o brekage can't be used */
390static bool cgroup_is_valid_domain(struct cgroup *cgrp)
391{
392 /* the cgroup itself can be a thread root */
393 if (cgroup_is_threaded(cgrp))
394 return false;
395
396 /* but the ancestors can't be unless mixable */
397 while ((cgrp = cgroup_parent(cgrp))) {
398 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
399 return false;
400 if (cgroup_is_threaded(cgrp))
401 return false;
402 }
403
404 return true;
405}
406
407/* subsystems visibly enabled on a cgroup */
408static u16 cgroup_control(struct cgroup *cgrp)
409{
410 struct cgroup *parent = cgroup_parent(cgrp);
411 u16 root_ss_mask = cgrp->root->subsys_mask;
412
413 if (parent) {
414 u16 ss_mask = parent->subtree_control;
415
416 /* threaded cgroups can only have threaded controllers */
417 if (cgroup_is_threaded(cgrp))
418 ss_mask &= cgrp_dfl_threaded_ss_mask;
419 return ss_mask;
420 }
421
422 if (cgroup_on_dfl(cgrp))
423 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
424 cgrp_dfl_implicit_ss_mask);
425 return root_ss_mask;
426}
427
428/* subsystems enabled on a cgroup */
429static u16 cgroup_ss_mask(struct cgroup *cgrp)
430{
431 struct cgroup *parent = cgroup_parent(cgrp);
432
433 if (parent) {
434 u16 ss_mask = parent->subtree_ss_mask;
435
436 /* threaded cgroups can only have threaded controllers */
437 if (cgroup_is_threaded(cgrp))
438 ss_mask &= cgrp_dfl_threaded_ss_mask;
439 return ss_mask;
440 }
441
442 return cgrp->root->subsys_mask;
443}
444
445/**
446 * cgroup_css - obtain a cgroup's css for the specified subsystem
447 * @cgrp: the cgroup of interest
448 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
449 *
450 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
451 * function must be called either under cgroup_mutex or rcu_read_lock() and
452 * the caller is responsible for pinning the returned css if it wants to
453 * keep accessing it outside the said locks. This function may return
454 * %NULL if @cgrp doesn't have @subsys_id enabled.
455 */
456static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
457 struct cgroup_subsys *ss)
458{
459 if (ss)
460 return rcu_dereference_check(cgrp->subsys[ss->id],
461 lockdep_is_held(&cgroup_mutex));
462 else
463 return &cgrp->self;
464}
465
466/**
467 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
468 * @cgrp: the cgroup of interest
469 * @ss: the subsystem of interest
470 *
471 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
472 * or is offline, %NULL is returned.
473 */
474static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
475 struct cgroup_subsys *ss)
476{
477 struct cgroup_subsys_state *css;
478
479 rcu_read_lock();
480 css = cgroup_css(cgrp, ss);
481 if (!css || !css_tryget_online(css))
482 css = NULL;
483 rcu_read_unlock();
484
485 return css;
486}
487
488/**
489 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
490 * @cgrp: the cgroup of interest
491 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
492 *
493 * Similar to cgroup_css() but returns the effective css, which is defined
494 * as the matching css of the nearest ancestor including self which has @ss
495 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
496 * function is guaranteed to return non-NULL css.
497 */
498static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
499 struct cgroup_subsys *ss)
500{
501 lockdep_assert_held(&cgroup_mutex);
502
503 if (!ss)
504 return &cgrp->self;
505
506 /*
507 * This function is used while updating css associations and thus
508 * can't test the csses directly. Test ss_mask.
509 */
510 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
511 cgrp = cgroup_parent(cgrp);
512 if (!cgrp)
513 return NULL;
514 }
515
516 return cgroup_css(cgrp, ss);
517}
518
519/**
520 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
521 * @cgrp: the cgroup of interest
522 * @ss: the subsystem of interest
523 *
524 * Find and get the effective css of @cgrp for @ss. The effective css is
525 * defined as the matching css of the nearest ancestor including self which
526 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
527 * the root css is returned, so this function always returns a valid css.
528 * The returned css must be put using css_put().
529 */
530struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
531 struct cgroup_subsys *ss)
532{
533 struct cgroup_subsys_state *css;
534
535 rcu_read_lock();
536
537 do {
538 css = cgroup_css(cgrp, ss);
539
540 if (css && css_tryget_online(css))
541 goto out_unlock;
542 cgrp = cgroup_parent(cgrp);
543 } while (cgrp);
544
545 css = init_css_set.subsys[ss->id];
546 css_get(css);
547out_unlock:
548 rcu_read_unlock();
549 return css;
550}
551
552static void cgroup_get_live(struct cgroup *cgrp)
553{
554 WARN_ON_ONCE(cgroup_is_dead(cgrp));
555 css_get(&cgrp->self);
556}
557
558struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
559{
560 struct cgroup *cgrp = of->kn->parent->priv;
561 struct cftype *cft = of_cft(of);
562
563 /*
564 * This is open and unprotected implementation of cgroup_css().
565 * seq_css() is only called from a kernfs file operation which has
566 * an active reference on the file. Because all the subsystem
567 * files are drained before a css is disassociated with a cgroup,
568 * the matching css from the cgroup's subsys table is guaranteed to
569 * be and stay valid until the enclosing operation is complete.
570 */
571 if (cft->ss)
572 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
573 else
574 return &cgrp->self;
575}
576EXPORT_SYMBOL_GPL(of_css);
577
578/**
579 * for_each_css - iterate all css's of a cgroup
580 * @css: the iteration cursor
581 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
582 * @cgrp: the target cgroup to iterate css's of
583 *
584 * Should be called under cgroup_[tree_]mutex.
585 */
586#define for_each_css(css, ssid, cgrp) \
587 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
588 if (!((css) = rcu_dereference_check( \
589 (cgrp)->subsys[(ssid)], \
590 lockdep_is_held(&cgroup_mutex)))) { } \
591 else
592
593/**
594 * for_each_e_css - iterate all effective css's of a cgroup
595 * @css: the iteration cursor
596 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
597 * @cgrp: the target cgroup to iterate css's of
598 *
599 * Should be called under cgroup_[tree_]mutex.
600 */
601#define for_each_e_css(css, ssid, cgrp) \
602 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
603 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
604 ; \
605 else
606
607/**
608 * do_each_subsys_mask - filter for_each_subsys with a bitmask
609 * @ss: the iteration cursor
610 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
611 * @ss_mask: the bitmask
612 *
613 * The block will only run for cases where the ssid-th bit (1 << ssid) of
614 * @ss_mask is set.
615 */
616#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
617 unsigned long __ss_mask = (ss_mask); \
618 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
619 (ssid) = 0; \
620 break; \
621 } \
622 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
623 (ss) = cgroup_subsys[ssid]; \
624 {
625
626#define while_each_subsys_mask() \
627 } \
628 } \
629} while (false)
630
631/* iterate over child cgrps, lock should be held throughout iteration */
632#define cgroup_for_each_live_child(child, cgrp) \
633 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
634 if (({ lockdep_assert_held(&cgroup_mutex); \
635 cgroup_is_dead(child); })) \
636 ; \
637 else
638
639/* walk live descendants in preorder */
640#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
641 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
642 if (({ lockdep_assert_held(&cgroup_mutex); \
643 (dsct) = (d_css)->cgroup; \
644 cgroup_is_dead(dsct); })) \
645 ; \
646 else
647
648/* walk live descendants in postorder */
649#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
650 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
651 if (({ lockdep_assert_held(&cgroup_mutex); \
652 (dsct) = (d_css)->cgroup; \
653 cgroup_is_dead(dsct); })) \
654 ; \
655 else
656
657/*
658 * The default css_set - used by init and its children prior to any
659 * hierarchies being mounted. It contains a pointer to the root state
660 * for each subsystem. Also used to anchor the list of css_sets. Not
661 * reference-counted, to improve performance when child cgroups
662 * haven't been created.
663 */
664struct css_set init_css_set = {
665 .refcount = REFCOUNT_INIT(1),
666 .dom_cset = &init_css_set,
667 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
668 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
669 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
670 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
671 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
672 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
673 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
674
675 /*
676 * The following field is re-initialized when this cset gets linked
677 * in cgroup_init(). However, let's initialize the field
678 * statically too so that the default cgroup can be accessed safely
679 * early during boot.
680 */
681 .dfl_cgrp = &cgrp_dfl_root.cgrp,
682};
683
684static int css_set_count = 1; /* 1 for init_css_set */
685
686static bool css_set_threaded(struct css_set *cset)
687{
688 return cset->dom_cset != cset;
689}
690
691/**
692 * css_set_populated - does a css_set contain any tasks?
693 * @cset: target css_set
694 *
695 * css_set_populated() should be the same as !!cset->nr_tasks at steady
696 * state. However, css_set_populated() can be called while a task is being
697 * added to or removed from the linked list before the nr_tasks is
698 * properly updated. Hence, we can't just look at ->nr_tasks here.
699 */
700static bool css_set_populated(struct css_set *cset)
701{
702 lockdep_assert_held(&css_set_lock);
703
704 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
705}
706
707/**
708 * cgroup_update_populated - update the populated count of a cgroup
709 * @cgrp: the target cgroup
710 * @populated: inc or dec populated count
711 *
712 * One of the css_sets associated with @cgrp is either getting its first
713 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
714 * count is propagated towards root so that a given cgroup's
715 * nr_populated_children is zero iff none of its descendants contain any
716 * tasks.
717 *
718 * @cgrp's interface file "cgroup.populated" is zero if both
719 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
720 * 1 otherwise. When the sum changes from or to zero, userland is notified
721 * that the content of the interface file has changed. This can be used to
722 * detect when @cgrp and its descendants become populated or empty.
723 */
724static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
725{
726 struct cgroup *child = NULL;
727 int adj = populated ? 1 : -1;
728
729 lockdep_assert_held(&css_set_lock);
730
731 do {
732 bool was_populated = cgroup_is_populated(cgrp);
733
734 if (!child) {
735 cgrp->nr_populated_csets += adj;
736 } else {
737 if (cgroup_is_threaded(child))
738 cgrp->nr_populated_threaded_children += adj;
739 else
740 cgrp->nr_populated_domain_children += adj;
741 }
742
743 if (was_populated == cgroup_is_populated(cgrp))
744 break;
745
746 cgroup1_check_for_release(cgrp);
747 cgroup_file_notify(&cgrp->events_file);
748
749 child = cgrp;
750 cgrp = cgroup_parent(cgrp);
751 } while (cgrp);
752}
753
754/**
755 * css_set_update_populated - update populated state of a css_set
756 * @cset: target css_set
757 * @populated: whether @cset is populated or depopulated
758 *
759 * @cset is either getting the first task or losing the last. Update the
760 * populated counters of all associated cgroups accordingly.
761 */
762static void css_set_update_populated(struct css_set *cset, bool populated)
763{
764 struct cgrp_cset_link *link;
765
766 lockdep_assert_held(&css_set_lock);
767
768 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
769 cgroup_update_populated(link->cgrp, populated);
770}
771
772/**
773 * css_set_move_task - move a task from one css_set to another
774 * @task: task being moved
775 * @from_cset: css_set @task currently belongs to (may be NULL)
776 * @to_cset: new css_set @task is being moved to (may be NULL)
777 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
778 *
779 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
780 * css_set, @from_cset can be NULL. If @task is being disassociated
781 * instead of moved, @to_cset can be NULL.
782 *
783 * This function automatically handles populated counter updates and
784 * css_task_iter adjustments but the caller is responsible for managing
785 * @from_cset and @to_cset's reference counts.
786 */
787static void css_set_move_task(struct task_struct *task,
788 struct css_set *from_cset, struct css_set *to_cset,
789 bool use_mg_tasks)
790{
791 lockdep_assert_held(&css_set_lock);
792
793 if (to_cset && !css_set_populated(to_cset))
794 css_set_update_populated(to_cset, true);
795
796 if (from_cset) {
797 struct css_task_iter *it, *pos;
798
799 WARN_ON_ONCE(list_empty(&task->cg_list));
800
801 /*
802 * @task is leaving, advance task iterators which are
803 * pointing to it so that they can resume at the next
804 * position. Advancing an iterator might remove it from
805 * the list, use safe walk. See css_task_iter_advance*()
806 * for details.
807 */
808 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
809 iters_node)
810 if (it->task_pos == &task->cg_list)
811 css_task_iter_advance(it);
812
813 list_del_init(&task->cg_list);
814 if (!css_set_populated(from_cset))
815 css_set_update_populated(from_cset, false);
816 } else {
817 WARN_ON_ONCE(!list_empty(&task->cg_list));
818 }
819
820 if (to_cset) {
821 /*
822 * We are synchronized through cgroup_threadgroup_rwsem
823 * against PF_EXITING setting such that we can't race
824 * against cgroup_exit() changing the css_set to
825 * init_css_set and dropping the old one.
826 */
827 WARN_ON_ONCE(task->flags & PF_EXITING);
828
829 rcu_assign_pointer(task->cgroups, to_cset);
830 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
831 &to_cset->tasks);
832 }
833}
834
835/*
836 * hash table for cgroup groups. This improves the performance to find
837 * an existing css_set. This hash doesn't (currently) take into
838 * account cgroups in empty hierarchies.
839 */
840#define CSS_SET_HASH_BITS 7
841static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
842
843static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
844{
845 unsigned long key = 0UL;
846 struct cgroup_subsys *ss;
847 int i;
848
849 for_each_subsys(ss, i)
850 key += (unsigned long)css[i];
851 key = (key >> 16) ^ key;
852
853 return key;
854}
855
856void put_css_set_locked(struct css_set *cset)
857{
858 struct cgrp_cset_link *link, *tmp_link;
859 struct cgroup_subsys *ss;
860 int ssid;
861
862 lockdep_assert_held(&css_set_lock);
863
864 if (!refcount_dec_and_test(&cset->refcount))
865 return;
866
867 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
868
869 /* This css_set is dead. unlink it and release cgroup and css refs */
870 for_each_subsys(ss, ssid) {
871 list_del(&cset->e_cset_node[ssid]);
872 css_put(cset->subsys[ssid]);
873 }
874 hash_del(&cset->hlist);
875 css_set_count--;
876
877 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
878 list_del(&link->cset_link);
879 list_del(&link->cgrp_link);
880 if (cgroup_parent(link->cgrp))
881 cgroup_put(link->cgrp);
882 kfree(link);
883 }
884
885 if (css_set_threaded(cset)) {
886 list_del(&cset->threaded_csets_node);
887 put_css_set_locked(cset->dom_cset);
888 }
889
890 kfree_rcu(cset, rcu_head);
891}
892
893/**
894 * compare_css_sets - helper function for find_existing_css_set().
895 * @cset: candidate css_set being tested
896 * @old_cset: existing css_set for a task
897 * @new_cgrp: cgroup that's being entered by the task
898 * @template: desired set of css pointers in css_set (pre-calculated)
899 *
900 * Returns true if "cset" matches "old_cset" except for the hierarchy
901 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
902 */
903static bool compare_css_sets(struct css_set *cset,
904 struct css_set *old_cset,
905 struct cgroup *new_cgrp,
906 struct cgroup_subsys_state *template[])
907{
908 struct cgroup *new_dfl_cgrp;
909 struct list_head *l1, *l2;
910
911 /*
912 * On the default hierarchy, there can be csets which are
913 * associated with the same set of cgroups but different csses.
914 * Let's first ensure that csses match.
915 */
916 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
917 return false;
918
919
920 /* @cset's domain should match the default cgroup's */
921 if (cgroup_on_dfl(new_cgrp))
922 new_dfl_cgrp = new_cgrp;
923 else
924 new_dfl_cgrp = old_cset->dfl_cgrp;
925
926 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
927 return false;
928
929 /*
930 * Compare cgroup pointers in order to distinguish between
931 * different cgroups in hierarchies. As different cgroups may
932 * share the same effective css, this comparison is always
933 * necessary.
934 */
935 l1 = &cset->cgrp_links;
936 l2 = &old_cset->cgrp_links;
937 while (1) {
938 struct cgrp_cset_link *link1, *link2;
939 struct cgroup *cgrp1, *cgrp2;
940
941 l1 = l1->next;
942 l2 = l2->next;
943 /* See if we reached the end - both lists are equal length. */
944 if (l1 == &cset->cgrp_links) {
945 BUG_ON(l2 != &old_cset->cgrp_links);
946 break;
947 } else {
948 BUG_ON(l2 == &old_cset->cgrp_links);
949 }
950 /* Locate the cgroups associated with these links. */
951 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
952 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
953 cgrp1 = link1->cgrp;
954 cgrp2 = link2->cgrp;
955 /* Hierarchies should be linked in the same order. */
956 BUG_ON(cgrp1->root != cgrp2->root);
957
958 /*
959 * If this hierarchy is the hierarchy of the cgroup
960 * that's changing, then we need to check that this
961 * css_set points to the new cgroup; if it's any other
962 * hierarchy, then this css_set should point to the
963 * same cgroup as the old css_set.
964 */
965 if (cgrp1->root == new_cgrp->root) {
966 if (cgrp1 != new_cgrp)
967 return false;
968 } else {
969 if (cgrp1 != cgrp2)
970 return false;
971 }
972 }
973 return true;
974}
975
976/**
977 * find_existing_css_set - init css array and find the matching css_set
978 * @old_cset: the css_set that we're using before the cgroup transition
979 * @cgrp: the cgroup that we're moving into
980 * @template: out param for the new set of csses, should be clear on entry
981 */
982static struct css_set *find_existing_css_set(struct css_set *old_cset,
983 struct cgroup *cgrp,
984 struct cgroup_subsys_state *template[])
985{
986 struct cgroup_root *root = cgrp->root;
987 struct cgroup_subsys *ss;
988 struct css_set *cset;
989 unsigned long key;
990 int i;
991
992 /*
993 * Build the set of subsystem state objects that we want to see in the
994 * new css_set. while subsystems can change globally, the entries here
995 * won't change, so no need for locking.
996 */
997 for_each_subsys(ss, i) {
998 if (root->subsys_mask & (1UL << i)) {
999 /*
1000 * @ss is in this hierarchy, so we want the
1001 * effective css from @cgrp.
1002 */
1003 template[i] = cgroup_e_css(cgrp, ss);
1004 } else {
1005 /*
1006 * @ss is not in this hierarchy, so we don't want
1007 * to change the css.
1008 */
1009 template[i] = old_cset->subsys[i];
1010 }
1011 }
1012
1013 key = css_set_hash(template);
1014 hash_for_each_possible(css_set_table, cset, hlist, key) {
1015 if (!compare_css_sets(cset, old_cset, cgrp, template))
1016 continue;
1017
1018 /* This css_set matches what we need */
1019 return cset;
1020 }
1021
1022 /* No existing cgroup group matched */
1023 return NULL;
1024}
1025
1026static void free_cgrp_cset_links(struct list_head *links_to_free)
1027{
1028 struct cgrp_cset_link *link, *tmp_link;
1029
1030 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1031 list_del(&link->cset_link);
1032 kfree(link);
1033 }
1034}
1035
1036/**
1037 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1038 * @count: the number of links to allocate
1039 * @tmp_links: list_head the allocated links are put on
1040 *
1041 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1042 * through ->cset_link. Returns 0 on success or -errno.
1043 */
1044static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1045{
1046 struct cgrp_cset_link *link;
1047 int i;
1048
1049 INIT_LIST_HEAD(tmp_links);
1050
1051 for (i = 0; i < count; i++) {
1052 link = kzalloc(sizeof(*link), GFP_KERNEL);
1053 if (!link) {
1054 free_cgrp_cset_links(tmp_links);
1055 return -ENOMEM;
1056 }
1057 list_add(&link->cset_link, tmp_links);
1058 }
1059 return 0;
1060}
1061
1062/**
1063 * link_css_set - a helper function to link a css_set to a cgroup
1064 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1065 * @cset: the css_set to be linked
1066 * @cgrp: the destination cgroup
1067 */
1068static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1069 struct cgroup *cgrp)
1070{
1071 struct cgrp_cset_link *link;
1072
1073 BUG_ON(list_empty(tmp_links));
1074
1075 if (cgroup_on_dfl(cgrp))
1076 cset->dfl_cgrp = cgrp;
1077
1078 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1079 link->cset = cset;
1080 link->cgrp = cgrp;
1081
1082 /*
1083 * Always add links to the tail of the lists so that the lists are
1084 * in choronological order.
1085 */
1086 list_move_tail(&link->cset_link, &cgrp->cset_links);
1087 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1088
1089 if (cgroup_parent(cgrp))
1090 cgroup_get_live(cgrp);
1091}
1092
1093/**
1094 * find_css_set - return a new css_set with one cgroup updated
1095 * @old_cset: the baseline css_set
1096 * @cgrp: the cgroup to be updated
1097 *
1098 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1099 * substituted into the appropriate hierarchy.
1100 */
1101static struct css_set *find_css_set(struct css_set *old_cset,
1102 struct cgroup *cgrp)
1103{
1104 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1105 struct css_set *cset;
1106 struct list_head tmp_links;
1107 struct cgrp_cset_link *link;
1108 struct cgroup_subsys *ss;
1109 unsigned long key;
1110 int ssid;
1111
1112 lockdep_assert_held(&cgroup_mutex);
1113
1114 /* First see if we already have a cgroup group that matches
1115 * the desired set */
1116 spin_lock_irq(&css_set_lock);
1117 cset = find_existing_css_set(old_cset, cgrp, template);
1118 if (cset)
1119 get_css_set(cset);
1120 spin_unlock_irq(&css_set_lock);
1121
1122 if (cset)
1123 return cset;
1124
1125 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1126 if (!cset)
1127 return NULL;
1128
1129 /* Allocate all the cgrp_cset_link objects that we'll need */
1130 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1131 kfree(cset);
1132 return NULL;
1133 }
1134
1135 refcount_set(&cset->refcount, 1);
1136 cset->dom_cset = cset;
1137 INIT_LIST_HEAD(&cset->tasks);
1138 INIT_LIST_HEAD(&cset->mg_tasks);
1139 INIT_LIST_HEAD(&cset->task_iters);
1140 INIT_LIST_HEAD(&cset->threaded_csets);
1141 INIT_HLIST_NODE(&cset->hlist);
1142 INIT_LIST_HEAD(&cset->cgrp_links);
1143 INIT_LIST_HEAD(&cset->mg_preload_node);
1144 INIT_LIST_HEAD(&cset->mg_node);
1145
1146 /* Copy the set of subsystem state objects generated in
1147 * find_existing_css_set() */
1148 memcpy(cset->subsys, template, sizeof(cset->subsys));
1149
1150 spin_lock_irq(&css_set_lock);
1151 /* Add reference counts and links from the new css_set. */
1152 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1153 struct cgroup *c = link->cgrp;
1154
1155 if (c->root == cgrp->root)
1156 c = cgrp;
1157 link_css_set(&tmp_links, cset, c);
1158 }
1159
1160 BUG_ON(!list_empty(&tmp_links));
1161
1162 css_set_count++;
1163
1164 /* Add @cset to the hash table */
1165 key = css_set_hash(cset->subsys);
1166 hash_add(css_set_table, &cset->hlist, key);
1167
1168 for_each_subsys(ss, ssid) {
1169 struct cgroup_subsys_state *css = cset->subsys[ssid];
1170
1171 list_add_tail(&cset->e_cset_node[ssid],
1172 &css->cgroup->e_csets[ssid]);
1173 css_get(css);
1174 }
1175
1176 spin_unlock_irq(&css_set_lock);
1177
1178 /*
1179 * If @cset should be threaded, look up the matching dom_cset and
1180 * link them up. We first fully initialize @cset then look for the
1181 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1182 * to stay empty until we return.
1183 */
1184 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1185 struct css_set *dcset;
1186
1187 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1188 if (!dcset) {
1189 put_css_set(cset);
1190 return NULL;
1191 }
1192
1193 spin_lock_irq(&css_set_lock);
1194 cset->dom_cset = dcset;
1195 list_add_tail(&cset->threaded_csets_node,
1196 &dcset->threaded_csets);
1197 spin_unlock_irq(&css_set_lock);
1198 }
1199
1200 return cset;
1201}
1202
1203struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1204{
1205 struct cgroup *root_cgrp = kf_root->kn->priv;
1206
1207 return root_cgrp->root;
1208}
1209
1210static int cgroup_init_root_id(struct cgroup_root *root)
1211{
1212 int id;
1213
1214 lockdep_assert_held(&cgroup_mutex);
1215
1216 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1217 if (id < 0)
1218 return id;
1219
1220 root->hierarchy_id = id;
1221 return 0;
1222}
1223
1224static void cgroup_exit_root_id(struct cgroup_root *root)
1225{
1226 lockdep_assert_held(&cgroup_mutex);
1227
1228 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1229}
1230
1231void cgroup_free_root(struct cgroup_root *root)
1232{
1233 if (root) {
1234 idr_destroy(&root->cgroup_idr);
1235 kfree(root);
1236 }
1237}
1238
1239static void cgroup_destroy_root(struct cgroup_root *root)
1240{
1241 struct cgroup *cgrp = &root->cgrp;
1242 struct cgrp_cset_link *link, *tmp_link;
1243
1244 trace_cgroup_destroy_root(root);
1245
1246 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1247
1248 BUG_ON(atomic_read(&root->nr_cgrps));
1249 BUG_ON(!list_empty(&cgrp->self.children));
1250
1251 /* Rebind all subsystems back to the default hierarchy */
1252 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1253
1254 /*
1255 * Release all the links from cset_links to this hierarchy's
1256 * root cgroup
1257 */
1258 spin_lock_irq(&css_set_lock);
1259
1260 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1261 list_del(&link->cset_link);
1262 list_del(&link->cgrp_link);
1263 kfree(link);
1264 }
1265
1266 spin_unlock_irq(&css_set_lock);
1267
1268 if (!list_empty(&root->root_list)) {
1269 list_del(&root->root_list);
1270 cgroup_root_count--;
1271 }
1272
1273 cgroup_exit_root_id(root);
1274
1275 mutex_unlock(&cgroup_mutex);
1276
1277 kernfs_destroy_root(root->kf_root);
1278 cgroup_free_root(root);
1279}
1280
1281/*
1282 * look up cgroup associated with current task's cgroup namespace on the
1283 * specified hierarchy
1284 */
1285static struct cgroup *
1286current_cgns_cgroup_from_root(struct cgroup_root *root)
1287{
1288 struct cgroup *res = NULL;
1289 struct css_set *cset;
1290
1291 lockdep_assert_held(&css_set_lock);
1292
1293 rcu_read_lock();
1294
1295 cset = current->nsproxy->cgroup_ns->root_cset;
1296 if (cset == &init_css_set) {
1297 res = &root->cgrp;
1298 } else {
1299 struct cgrp_cset_link *link;
1300
1301 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1302 struct cgroup *c = link->cgrp;
1303
1304 if (c->root == root) {
1305 res = c;
1306 break;
1307 }
1308 }
1309 }
1310 rcu_read_unlock();
1311
1312 BUG_ON(!res);
1313 return res;
1314}
1315
1316/* look up cgroup associated with given css_set on the specified hierarchy */
1317static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1318 struct cgroup_root *root)
1319{
1320 struct cgroup *res = NULL;
1321
1322 lockdep_assert_held(&cgroup_mutex);
1323 lockdep_assert_held(&css_set_lock);
1324
1325 if (cset == &init_css_set) {
1326 res = &root->cgrp;
1327 } else if (root == &cgrp_dfl_root) {
1328 res = cset->dfl_cgrp;
1329 } else {
1330 struct cgrp_cset_link *link;
1331
1332 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1333 struct cgroup *c = link->cgrp;
1334
1335 if (c->root == root) {
1336 res = c;
1337 break;
1338 }
1339 }
1340 }
1341
1342 BUG_ON(!res);
1343 return res;
1344}
1345
1346/*
1347 * Return the cgroup for "task" from the given hierarchy. Must be
1348 * called with cgroup_mutex and css_set_lock held.
1349 */
1350struct cgroup *task_cgroup_from_root(struct task_struct *task,
1351 struct cgroup_root *root)
1352{
1353 /*
1354 * No need to lock the task - since we hold cgroup_mutex the
1355 * task can't change groups, so the only thing that can happen
1356 * is that it exits and its css is set back to init_css_set.
1357 */
1358 return cset_cgroup_from_root(task_css_set(task), root);
1359}
1360
1361/*
1362 * A task must hold cgroup_mutex to modify cgroups.
1363 *
1364 * Any task can increment and decrement the count field without lock.
1365 * So in general, code holding cgroup_mutex can't rely on the count
1366 * field not changing. However, if the count goes to zero, then only
1367 * cgroup_attach_task() can increment it again. Because a count of zero
1368 * means that no tasks are currently attached, therefore there is no
1369 * way a task attached to that cgroup can fork (the other way to
1370 * increment the count). So code holding cgroup_mutex can safely
1371 * assume that if the count is zero, it will stay zero. Similarly, if
1372 * a task holds cgroup_mutex on a cgroup with zero count, it
1373 * knows that the cgroup won't be removed, as cgroup_rmdir()
1374 * needs that mutex.
1375 *
1376 * A cgroup can only be deleted if both its 'count' of using tasks
1377 * is zero, and its list of 'children' cgroups is empty. Since all
1378 * tasks in the system use _some_ cgroup, and since there is always at
1379 * least one task in the system (init, pid == 1), therefore, root cgroup
1380 * always has either children cgroups and/or using tasks. So we don't
1381 * need a special hack to ensure that root cgroup cannot be deleted.
1382 *
1383 * P.S. One more locking exception. RCU is used to guard the
1384 * update of a tasks cgroup pointer by cgroup_attach_task()
1385 */
1386
1387static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1388
1389static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1390 char *buf)
1391{
1392 struct cgroup_subsys *ss = cft->ss;
1393
1394 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1395 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1396 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1397 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1398 cft->name);
1399 else
1400 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1401 return buf;
1402}
1403
1404/**
1405 * cgroup_file_mode - deduce file mode of a control file
1406 * @cft: the control file in question
1407 *
1408 * S_IRUGO for read, S_IWUSR for write.
1409 */
1410static umode_t cgroup_file_mode(const struct cftype *cft)
1411{
1412 umode_t mode = 0;
1413
1414 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1415 mode |= S_IRUGO;
1416
1417 if (cft->write_u64 || cft->write_s64 || cft->write) {
1418 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1419 mode |= S_IWUGO;
1420 else
1421 mode |= S_IWUSR;
1422 }
1423
1424 return mode;
1425}
1426
1427/**
1428 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1429 * @subtree_control: the new subtree_control mask to consider
1430 * @this_ss_mask: available subsystems
1431 *
1432 * On the default hierarchy, a subsystem may request other subsystems to be
1433 * enabled together through its ->depends_on mask. In such cases, more
1434 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1435 *
1436 * This function calculates which subsystems need to be enabled if
1437 * @subtree_control is to be applied while restricted to @this_ss_mask.
1438 */
1439static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1440{
1441 u16 cur_ss_mask = subtree_control;
1442 struct cgroup_subsys *ss;
1443 int ssid;
1444
1445 lockdep_assert_held(&cgroup_mutex);
1446
1447 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1448
1449 while (true) {
1450 u16 new_ss_mask = cur_ss_mask;
1451
1452 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1453 new_ss_mask |= ss->depends_on;
1454 } while_each_subsys_mask();
1455
1456 /*
1457 * Mask out subsystems which aren't available. This can
1458 * happen only if some depended-upon subsystems were bound
1459 * to non-default hierarchies.
1460 */
1461 new_ss_mask &= this_ss_mask;
1462
1463 if (new_ss_mask == cur_ss_mask)
1464 break;
1465 cur_ss_mask = new_ss_mask;
1466 }
1467
1468 return cur_ss_mask;
1469}
1470
1471/**
1472 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1473 * @kn: the kernfs_node being serviced
1474 *
1475 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1476 * the method finishes if locking succeeded. Note that once this function
1477 * returns the cgroup returned by cgroup_kn_lock_live() may become
1478 * inaccessible any time. If the caller intends to continue to access the
1479 * cgroup, it should pin it before invoking this function.
1480 */
1481void cgroup_kn_unlock(struct kernfs_node *kn)
1482{
1483 struct cgroup *cgrp;
1484
1485 if (kernfs_type(kn) == KERNFS_DIR)
1486 cgrp = kn->priv;
1487 else
1488 cgrp = kn->parent->priv;
1489
1490 mutex_unlock(&cgroup_mutex);
1491
1492 kernfs_unbreak_active_protection(kn);
1493 cgroup_put(cgrp);
1494}
1495
1496/**
1497 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1498 * @kn: the kernfs_node being serviced
1499 * @drain_offline: perform offline draining on the cgroup
1500 *
1501 * This helper is to be used by a cgroup kernfs method currently servicing
1502 * @kn. It breaks the active protection, performs cgroup locking and
1503 * verifies that the associated cgroup is alive. Returns the cgroup if
1504 * alive; otherwise, %NULL. A successful return should be undone by a
1505 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1506 * cgroup is drained of offlining csses before return.
1507 *
1508 * Any cgroup kernfs method implementation which requires locking the
1509 * associated cgroup should use this helper. It avoids nesting cgroup
1510 * locking under kernfs active protection and allows all kernfs operations
1511 * including self-removal.
1512 */
1513struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1514{
1515 struct cgroup *cgrp;
1516
1517 if (kernfs_type(kn) == KERNFS_DIR)
1518 cgrp = kn->priv;
1519 else
1520 cgrp = kn->parent->priv;
1521
1522 /*
1523 * We're gonna grab cgroup_mutex which nests outside kernfs
1524 * active_ref. cgroup liveliness check alone provides enough
1525 * protection against removal. Ensure @cgrp stays accessible and
1526 * break the active_ref protection.
1527 */
1528 if (!cgroup_tryget(cgrp))
1529 return NULL;
1530 kernfs_break_active_protection(kn);
1531
1532 if (drain_offline)
1533 cgroup_lock_and_drain_offline(cgrp);
1534 else
1535 mutex_lock(&cgroup_mutex);
1536
1537 if (!cgroup_is_dead(cgrp))
1538 return cgrp;
1539
1540 cgroup_kn_unlock(kn);
1541 return NULL;
1542}
1543
1544static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1545{
1546 char name[CGROUP_FILE_NAME_MAX];
1547
1548 lockdep_assert_held(&cgroup_mutex);
1549
1550 if (cft->file_offset) {
1551 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1552 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1553
1554 spin_lock_irq(&cgroup_file_kn_lock);
1555 cfile->kn = NULL;
1556 spin_unlock_irq(&cgroup_file_kn_lock);
1557 }
1558
1559 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1560}
1561
1562/**
1563 * css_clear_dir - remove subsys files in a cgroup directory
1564 * @css: taget css
1565 */
1566static void css_clear_dir(struct cgroup_subsys_state *css)
1567{
1568 struct cgroup *cgrp = css->cgroup;
1569 struct cftype *cfts;
1570
1571 if (!(css->flags & CSS_VISIBLE))
1572 return;
1573
1574 css->flags &= ~CSS_VISIBLE;
1575
1576 list_for_each_entry(cfts, &css->ss->cfts, node)
1577 cgroup_addrm_files(css, cgrp, cfts, false);
1578}
1579
1580/**
1581 * css_populate_dir - create subsys files in a cgroup directory
1582 * @css: target css
1583 *
1584 * On failure, no file is added.
1585 */
1586static int css_populate_dir(struct cgroup_subsys_state *css)
1587{
1588 struct cgroup *cgrp = css->cgroup;
1589 struct cftype *cfts, *failed_cfts;
1590 int ret;
1591
1592 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1593 return 0;
1594
1595 if (!css->ss) {
1596 if (cgroup_on_dfl(cgrp))
1597 cfts = cgroup_base_files;
1598 else
1599 cfts = cgroup1_base_files;
1600
1601 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1602 }
1603
1604 list_for_each_entry(cfts, &css->ss->cfts, node) {
1605 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1606 if (ret < 0) {
1607 failed_cfts = cfts;
1608 goto err;
1609 }
1610 }
1611
1612 css->flags |= CSS_VISIBLE;
1613
1614 return 0;
1615err:
1616 list_for_each_entry(cfts, &css->ss->cfts, node) {
1617 if (cfts == failed_cfts)
1618 break;
1619 cgroup_addrm_files(css, cgrp, cfts, false);
1620 }
1621 return ret;
1622}
1623
1624int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1625{
1626 struct cgroup *dcgrp = &dst_root->cgrp;
1627 struct cgroup_subsys *ss;
1628 int ssid, i, ret;
1629
1630 lockdep_assert_held(&cgroup_mutex);
1631
1632 do_each_subsys_mask(ss, ssid, ss_mask) {
1633 /*
1634 * If @ss has non-root csses attached to it, can't move.
1635 * If @ss is an implicit controller, it is exempt from this
1636 * rule and can be stolen.
1637 */
1638 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1639 !ss->implicit_on_dfl)
1640 return -EBUSY;
1641
1642 /* can't move between two non-dummy roots either */
1643 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1644 return -EBUSY;
1645 } while_each_subsys_mask();
1646
1647 do_each_subsys_mask(ss, ssid, ss_mask) {
1648 struct cgroup_root *src_root = ss->root;
1649 struct cgroup *scgrp = &src_root->cgrp;
1650 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1651 struct css_set *cset;
1652
1653 WARN_ON(!css || cgroup_css(dcgrp, ss));
1654
1655 /* disable from the source */
1656 src_root->subsys_mask &= ~(1 << ssid);
1657 WARN_ON(cgroup_apply_control(scgrp));
1658 cgroup_finalize_control(scgrp, 0);
1659
1660 /* rebind */
1661 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1662 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1663 ss->root = dst_root;
1664 css->cgroup = dcgrp;
1665
1666 spin_lock_irq(&css_set_lock);
1667 hash_for_each(css_set_table, i, cset, hlist)
1668 list_move_tail(&cset->e_cset_node[ss->id],
1669 &dcgrp->e_csets[ss->id]);
1670 spin_unlock_irq(&css_set_lock);
1671
1672 /* default hierarchy doesn't enable controllers by default */
1673 dst_root->subsys_mask |= 1 << ssid;
1674 if (dst_root == &cgrp_dfl_root) {
1675 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1676 } else {
1677 dcgrp->subtree_control |= 1 << ssid;
1678 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1679 }
1680
1681 ret = cgroup_apply_control(dcgrp);
1682 if (ret)
1683 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1684 ss->name, ret);
1685
1686 if (ss->bind)
1687 ss->bind(css);
1688 } while_each_subsys_mask();
1689
1690 kernfs_activate(dcgrp->kn);
1691 return 0;
1692}
1693
1694int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1695 struct kernfs_root *kf_root)
1696{
1697 int len = 0;
1698 char *buf = NULL;
1699 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1700 struct cgroup *ns_cgroup;
1701
1702 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1703 if (!buf)
1704 return -ENOMEM;
1705
1706 spin_lock_irq(&css_set_lock);
1707 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1708 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1709 spin_unlock_irq(&css_set_lock);
1710
1711 if (len >= PATH_MAX)
1712 len = -ERANGE;
1713 else if (len > 0) {
1714 seq_escape(sf, buf, " \t\n\\");
1715 len = 0;
1716 }
1717 kfree(buf);
1718 return len;
1719}
1720
1721static int parse_cgroup_root_flags(char *data, unsigned int *root_flags)
1722{
1723 char *token;
1724
1725 *root_flags = 0;
1726
1727 if (!data)
1728 return 0;
1729
1730 while ((token = strsep(&data, ",")) != NULL) {
1731 if (!strcmp(token, "nsdelegate")) {
1732 *root_flags |= CGRP_ROOT_NS_DELEGATE;
1733 continue;
1734 }
1735
1736 pr_err("cgroup2: unknown option \"%s\"\n", token);
1737 return -EINVAL;
1738 }
1739
1740 return 0;
1741}
1742
1743static void apply_cgroup_root_flags(unsigned int root_flags)
1744{
1745 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1746 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1747 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1748 else
1749 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1750 }
1751}
1752
1753static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1754{
1755 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1756 seq_puts(seq, ",nsdelegate");
1757 return 0;
1758}
1759
1760static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1761{
1762 unsigned int root_flags;
1763 int ret;
1764
1765 ret = parse_cgroup_root_flags(data, &root_flags);
1766 if (ret)
1767 return ret;
1768
1769 apply_cgroup_root_flags(root_flags);
1770 return 0;
1771}
1772
1773/*
1774 * To reduce the fork() overhead for systems that are not actually using
1775 * their cgroups capability, we don't maintain the lists running through
1776 * each css_set to its tasks until we see the list actually used - in other
1777 * words after the first mount.
1778 */
1779static bool use_task_css_set_links __read_mostly;
1780
1781static void cgroup_enable_task_cg_lists(void)
1782{
1783 struct task_struct *p, *g;
1784
1785 spin_lock_irq(&css_set_lock);
1786
1787 if (use_task_css_set_links)
1788 goto out_unlock;
1789
1790 use_task_css_set_links = true;
1791
1792 /*
1793 * We need tasklist_lock because RCU is not safe against
1794 * while_each_thread(). Besides, a forking task that has passed
1795 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1796 * is not guaranteed to have its child immediately visible in the
1797 * tasklist if we walk through it with RCU.
1798 */
1799 read_lock(&tasklist_lock);
1800 do_each_thread(g, p) {
1801 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1802 task_css_set(p) != &init_css_set);
1803
1804 /*
1805 * We should check if the process is exiting, otherwise
1806 * it will race with cgroup_exit() in that the list
1807 * entry won't be deleted though the process has exited.
1808 * Do it while holding siglock so that we don't end up
1809 * racing against cgroup_exit().
1810 *
1811 * Interrupts were already disabled while acquiring
1812 * the css_set_lock, so we do not need to disable it
1813 * again when acquiring the sighand->siglock here.
1814 */
1815 spin_lock(&p->sighand->siglock);
1816 if (!(p->flags & PF_EXITING)) {
1817 struct css_set *cset = task_css_set(p);
1818
1819 if (!css_set_populated(cset))
1820 css_set_update_populated(cset, true);
1821 list_add_tail(&p->cg_list, &cset->tasks);
1822 get_css_set(cset);
1823 cset->nr_tasks++;
1824 }
1825 spin_unlock(&p->sighand->siglock);
1826 } while_each_thread(g, p);
1827 read_unlock(&tasklist_lock);
1828out_unlock:
1829 spin_unlock_irq(&css_set_lock);
1830}
1831
1832static void init_cgroup_housekeeping(struct cgroup *cgrp)
1833{
1834 struct cgroup_subsys *ss;
1835 int ssid;
1836
1837 INIT_LIST_HEAD(&cgrp->self.sibling);
1838 INIT_LIST_HEAD(&cgrp->self.children);
1839 INIT_LIST_HEAD(&cgrp->cset_links);
1840 INIT_LIST_HEAD(&cgrp->pidlists);
1841 mutex_init(&cgrp->pidlist_mutex);
1842 cgrp->self.cgroup = cgrp;
1843 cgrp->self.flags |= CSS_ONLINE;
1844 cgrp->dom_cgrp = cgrp;
1845 cgrp->max_descendants = INT_MAX;
1846 cgrp->max_depth = INT_MAX;
1847
1848 for_each_subsys(ss, ssid)
1849 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1850
1851 init_waitqueue_head(&cgrp->offline_waitq);
1852 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1853}
1854
1855void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts)
1856{
1857 struct cgroup *cgrp = &root->cgrp;
1858
1859 INIT_LIST_HEAD(&root->root_list);
1860 atomic_set(&root->nr_cgrps, 1);
1861 cgrp->root = root;
1862 init_cgroup_housekeeping(cgrp);
1863 idr_init(&root->cgroup_idr);
1864
1865 root->flags = opts->flags;
1866 if (opts->release_agent)
1867 strscpy(root->release_agent_path, opts->release_agent, PATH_MAX);
1868 if (opts->name)
1869 strscpy(root->name, opts->name, MAX_CGROUP_ROOT_NAMELEN);
1870 if (opts->cpuset_clone_children)
1871 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1872}
1873
1874int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags)
1875{
1876 LIST_HEAD(tmp_links);
1877 struct cgroup *root_cgrp = &root->cgrp;
1878 struct kernfs_syscall_ops *kf_sops;
1879 struct css_set *cset;
1880 int i, ret;
1881
1882 lockdep_assert_held(&cgroup_mutex);
1883
1884 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1885 if (ret < 0)
1886 goto out;
1887 root_cgrp->id = ret;
1888 root_cgrp->ancestor_ids[0] = ret;
1889
1890 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1891 ref_flags, GFP_KERNEL);
1892 if (ret)
1893 goto out;
1894
1895 /*
1896 * We're accessing css_set_count without locking css_set_lock here,
1897 * but that's OK - it can only be increased by someone holding
1898 * cgroup_lock, and that's us. Later rebinding may disable
1899 * controllers on the default hierarchy and thus create new csets,
1900 * which can't be more than the existing ones. Allocate 2x.
1901 */
1902 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1903 if (ret)
1904 goto cancel_ref;
1905
1906 ret = cgroup_init_root_id(root);
1907 if (ret)
1908 goto cancel_ref;
1909
1910 kf_sops = root == &cgrp_dfl_root ?
1911 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1912
1913 root->kf_root = kernfs_create_root(kf_sops,
1914 KERNFS_ROOT_CREATE_DEACTIVATED |
1915 KERNFS_ROOT_SUPPORT_EXPORTOP,
1916 root_cgrp);
1917 if (IS_ERR(root->kf_root)) {
1918 ret = PTR_ERR(root->kf_root);
1919 goto exit_root_id;
1920 }
1921 root_cgrp->kn = root->kf_root->kn;
1922
1923 ret = css_populate_dir(&root_cgrp->self);
1924 if (ret)
1925 goto destroy_root;
1926
1927 ret = rebind_subsystems(root, ss_mask);
1928 if (ret)
1929 goto destroy_root;
1930
1931 ret = cgroup_bpf_inherit(root_cgrp);
1932 WARN_ON_ONCE(ret);
1933
1934 trace_cgroup_setup_root(root);
1935
1936 /*
1937 * There must be no failure case after here, since rebinding takes
1938 * care of subsystems' refcounts, which are explicitly dropped in
1939 * the failure exit path.
1940 */
1941 list_add(&root->root_list, &cgroup_roots);
1942 cgroup_root_count++;
1943
1944 /*
1945 * Link the root cgroup in this hierarchy into all the css_set
1946 * objects.
1947 */
1948 spin_lock_irq(&css_set_lock);
1949 hash_for_each(css_set_table, i, cset, hlist) {
1950 link_css_set(&tmp_links, cset, root_cgrp);
1951 if (css_set_populated(cset))
1952 cgroup_update_populated(root_cgrp, true);
1953 }
1954 spin_unlock_irq(&css_set_lock);
1955
1956 BUG_ON(!list_empty(&root_cgrp->self.children));
1957 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1958
1959 kernfs_activate(root_cgrp->kn);
1960 ret = 0;
1961 goto out;
1962
1963destroy_root:
1964 kernfs_destroy_root(root->kf_root);
1965 root->kf_root = NULL;
1966exit_root_id:
1967 cgroup_exit_root_id(root);
1968cancel_ref:
1969 percpu_ref_exit(&root_cgrp->self.refcnt);
1970out:
1971 free_cgrp_cset_links(&tmp_links);
1972 return ret;
1973}
1974
1975struct dentry *cgroup_do_mount(struct file_system_type *fs_type, int flags,
1976 struct cgroup_root *root, unsigned long magic,
1977 struct cgroup_namespace *ns)
1978{
1979 struct dentry *dentry;
1980 bool new_sb;
1981
1982 dentry = kernfs_mount(fs_type, flags, root->kf_root, magic, &new_sb);
1983
1984 /*
1985 * In non-init cgroup namespace, instead of root cgroup's dentry,
1986 * we return the dentry corresponding to the cgroupns->root_cgrp.
1987 */
1988 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
1989 struct dentry *nsdentry;
1990 struct cgroup *cgrp;
1991
1992 mutex_lock(&cgroup_mutex);
1993 spin_lock_irq(&css_set_lock);
1994
1995 cgrp = cset_cgroup_from_root(ns->root_cset, root);
1996
1997 spin_unlock_irq(&css_set_lock);
1998 mutex_unlock(&cgroup_mutex);
1999
2000 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2001 dput(dentry);
2002 dentry = nsdentry;
2003 }
2004
2005 if (IS_ERR(dentry) || !new_sb)
2006 cgroup_put(&root->cgrp);
2007
2008 return dentry;
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 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2016 struct dentry *dentry;
2017 int ret;
2018
2019 get_cgroup_ns(ns);
2020
2021 /* Check if the caller has permission to mount. */
2022 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2023 put_cgroup_ns(ns);
2024 return ERR_PTR(-EPERM);
2025 }
2026
2027 /*
2028 * The first time anyone tries to mount a cgroup, enable the list
2029 * linking each css_set to its tasks and fix up all existing tasks.
2030 */
2031 if (!use_task_css_set_links)
2032 cgroup_enable_task_cg_lists();
2033
2034 if (fs_type == &cgroup2_fs_type) {
2035 unsigned int root_flags;
2036
2037 ret = parse_cgroup_root_flags(data, &root_flags);
2038 if (ret) {
2039 put_cgroup_ns(ns);
2040 return ERR_PTR(ret);
2041 }
2042
2043 cgrp_dfl_visible = true;
2044 cgroup_get_live(&cgrp_dfl_root.cgrp);
2045
2046 dentry = cgroup_do_mount(&cgroup2_fs_type, flags, &cgrp_dfl_root,
2047 CGROUP2_SUPER_MAGIC, ns);
2048 if (!IS_ERR(dentry))
2049 apply_cgroup_root_flags(root_flags);
2050 } else {
2051 dentry = cgroup1_mount(&cgroup_fs_type, flags, data,
2052 CGROUP_SUPER_MAGIC, ns);
2053 }
2054
2055 put_cgroup_ns(ns);
2056 return dentry;
2057}
2058
2059static void cgroup_kill_sb(struct super_block *sb)
2060{
2061 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2062 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2063
2064 /*
2065 * If @root doesn't have any mounts or children, start killing it.
2066 * This prevents new mounts by disabling percpu_ref_tryget_live().
2067 * cgroup_mount() may wait for @root's release.
2068 *
2069 * And don't kill the default root.
2070 */
2071 if (!list_empty(&root->cgrp.self.children) ||
2072 root == &cgrp_dfl_root)
2073 cgroup_put(&root->cgrp);
2074 else
2075 percpu_ref_kill(&root->cgrp.self.refcnt);
2076
2077 kernfs_kill_sb(sb);
2078}
2079
2080struct file_system_type cgroup_fs_type = {
2081 .name = "cgroup",
2082 .mount = cgroup_mount,
2083 .kill_sb = cgroup_kill_sb,
2084 .fs_flags = FS_USERNS_MOUNT,
2085};
2086
2087static struct file_system_type cgroup2_fs_type = {
2088 .name = "cgroup2",
2089 .mount = cgroup_mount,
2090 .kill_sb = cgroup_kill_sb,
2091 .fs_flags = FS_USERNS_MOUNT,
2092};
2093
2094int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2095 struct cgroup_namespace *ns)
2096{
2097 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2098
2099 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2100}
2101
2102int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2103 struct cgroup_namespace *ns)
2104{
2105 int ret;
2106
2107 mutex_lock(&cgroup_mutex);
2108 spin_lock_irq(&css_set_lock);
2109
2110 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2111
2112 spin_unlock_irq(&css_set_lock);
2113 mutex_unlock(&cgroup_mutex);
2114
2115 return ret;
2116}
2117EXPORT_SYMBOL_GPL(cgroup_path_ns);
2118
2119/**
2120 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2121 * @task: target task
2122 * @buf: the buffer to write the path into
2123 * @buflen: the length of the buffer
2124 *
2125 * Determine @task's cgroup on the first (the one with the lowest non-zero
2126 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2127 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2128 * cgroup controller callbacks.
2129 *
2130 * Return value is the same as kernfs_path().
2131 */
2132int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2133{
2134 struct cgroup_root *root;
2135 struct cgroup *cgrp;
2136 int hierarchy_id = 1;
2137 int ret;
2138
2139 mutex_lock(&cgroup_mutex);
2140 spin_lock_irq(&css_set_lock);
2141
2142 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2143
2144 if (root) {
2145 cgrp = task_cgroup_from_root(task, root);
2146 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2147 } else {
2148 /* if no hierarchy exists, everyone is in "/" */
2149 ret = strlcpy(buf, "/", buflen);
2150 }
2151
2152 spin_unlock_irq(&css_set_lock);
2153 mutex_unlock(&cgroup_mutex);
2154 return ret;
2155}
2156EXPORT_SYMBOL_GPL(task_cgroup_path);
2157
2158/**
2159 * cgroup_migrate_add_task - add a migration target task to a migration context
2160 * @task: target task
2161 * @mgctx: target migration context
2162 *
2163 * Add @task, which is a migration target, to @mgctx->tset. This function
2164 * becomes noop if @task doesn't need to be migrated. @task's css_set
2165 * should have been added as a migration source and @task->cg_list will be
2166 * moved from the css_set's tasks list to mg_tasks one.
2167 */
2168static void cgroup_migrate_add_task(struct task_struct *task,
2169 struct cgroup_mgctx *mgctx)
2170{
2171 struct css_set *cset;
2172
2173 lockdep_assert_held(&css_set_lock);
2174
2175 /* @task either already exited or can't exit until the end */
2176 if (task->flags & PF_EXITING)
2177 return;
2178
2179 /* leave @task alone if post_fork() hasn't linked it yet */
2180 if (list_empty(&task->cg_list))
2181 return;
2182
2183 cset = task_css_set(task);
2184 if (!cset->mg_src_cgrp)
2185 return;
2186
2187 mgctx->tset.nr_tasks++;
2188
2189 list_move_tail(&task->cg_list, &cset->mg_tasks);
2190 if (list_empty(&cset->mg_node))
2191 list_add_tail(&cset->mg_node,
2192 &mgctx->tset.src_csets);
2193 if (list_empty(&cset->mg_dst_cset->mg_node))
2194 list_add_tail(&cset->mg_dst_cset->mg_node,
2195 &mgctx->tset.dst_csets);
2196}
2197
2198/**
2199 * cgroup_taskset_first - reset taskset and return the first task
2200 * @tset: taskset of interest
2201 * @dst_cssp: output variable for the destination css
2202 *
2203 * @tset iteration is initialized and the first task is returned.
2204 */
2205struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2206 struct cgroup_subsys_state **dst_cssp)
2207{
2208 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2209 tset->cur_task = NULL;
2210
2211 return cgroup_taskset_next(tset, dst_cssp);
2212}
2213
2214/**
2215 * cgroup_taskset_next - iterate to the next task in taskset
2216 * @tset: taskset of interest
2217 * @dst_cssp: output variable for the destination css
2218 *
2219 * Return the next task in @tset. Iteration must have been initialized
2220 * with cgroup_taskset_first().
2221 */
2222struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2223 struct cgroup_subsys_state **dst_cssp)
2224{
2225 struct css_set *cset = tset->cur_cset;
2226 struct task_struct *task = tset->cur_task;
2227
2228 while (&cset->mg_node != tset->csets) {
2229 if (!task)
2230 task = list_first_entry(&cset->mg_tasks,
2231 struct task_struct, cg_list);
2232 else
2233 task = list_next_entry(task, cg_list);
2234
2235 if (&task->cg_list != &cset->mg_tasks) {
2236 tset->cur_cset = cset;
2237 tset->cur_task = task;
2238
2239 /*
2240 * This function may be called both before and
2241 * after cgroup_taskset_migrate(). The two cases
2242 * can be distinguished by looking at whether @cset
2243 * has its ->mg_dst_cset set.
2244 */
2245 if (cset->mg_dst_cset)
2246 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2247 else
2248 *dst_cssp = cset->subsys[tset->ssid];
2249
2250 return task;
2251 }
2252
2253 cset = list_next_entry(cset, mg_node);
2254 task = NULL;
2255 }
2256
2257 return NULL;
2258}
2259
2260/**
2261 * cgroup_taskset_migrate - migrate a taskset
2262 * @mgctx: migration context
2263 *
2264 * Migrate tasks in @mgctx as setup by migration preparation functions.
2265 * This function fails iff one of the ->can_attach callbacks fails and
2266 * guarantees that either all or none of the tasks in @mgctx are migrated.
2267 * @mgctx is consumed regardless of success.
2268 */
2269static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2270{
2271 struct cgroup_taskset *tset = &mgctx->tset;
2272 struct cgroup_subsys *ss;
2273 struct task_struct *task, *tmp_task;
2274 struct css_set *cset, *tmp_cset;
2275 int ssid, failed_ssid, ret;
2276
2277 /* check that we can legitimately attach to the cgroup */
2278 if (tset->nr_tasks) {
2279 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2280 if (ss->can_attach) {
2281 tset->ssid = ssid;
2282 ret = ss->can_attach(tset);
2283 if (ret) {
2284 failed_ssid = ssid;
2285 goto out_cancel_attach;
2286 }
2287 }
2288 } while_each_subsys_mask();
2289 }
2290
2291 /*
2292 * Now that we're guaranteed success, proceed to move all tasks to
2293 * the new cgroup. There are no failure cases after here, so this
2294 * is the commit point.
2295 */
2296 spin_lock_irq(&css_set_lock);
2297 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2298 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2299 struct css_set *from_cset = task_css_set(task);
2300 struct css_set *to_cset = cset->mg_dst_cset;
2301
2302 get_css_set(to_cset);
2303 to_cset->nr_tasks++;
2304 css_set_move_task(task, from_cset, to_cset, true);
2305 put_css_set_locked(from_cset);
2306 from_cset->nr_tasks--;
2307 }
2308 }
2309 spin_unlock_irq(&css_set_lock);
2310
2311 /*
2312 * Migration is committed, all target tasks are now on dst_csets.
2313 * Nothing is sensitive to fork() after this point. Notify
2314 * controllers that migration is complete.
2315 */
2316 tset->csets = &tset->dst_csets;
2317
2318 if (tset->nr_tasks) {
2319 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2320 if (ss->attach) {
2321 tset->ssid = ssid;
2322 ss->attach(tset);
2323 }
2324 } while_each_subsys_mask();
2325 }
2326
2327 ret = 0;
2328 goto out_release_tset;
2329
2330out_cancel_attach:
2331 if (tset->nr_tasks) {
2332 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2333 if (ssid == failed_ssid)
2334 break;
2335 if (ss->cancel_attach) {
2336 tset->ssid = ssid;
2337 ss->cancel_attach(tset);
2338 }
2339 } while_each_subsys_mask();
2340 }
2341out_release_tset:
2342 spin_lock_irq(&css_set_lock);
2343 list_splice_init(&tset->dst_csets, &tset->src_csets);
2344 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2345 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2346 list_del_init(&cset->mg_node);
2347 }
2348 spin_unlock_irq(&css_set_lock);
2349
2350 /*
2351 * Re-initialize the cgroup_taskset structure in case it is reused
2352 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2353 * iteration.
2354 */
2355 tset->nr_tasks = 0;
2356 tset->csets = &tset->src_csets;
2357 return ret;
2358}
2359
2360/**
2361 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2362 * @dst_cgrp: destination cgroup to test
2363 *
2364 * On the default hierarchy, except for the mixable, (possible) thread root
2365 * and threaded cgroups, subtree_control must be zero for migration
2366 * destination cgroups with tasks so that child cgroups don't compete
2367 * against tasks.
2368 */
2369int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2370{
2371 /* v1 doesn't have any restriction */
2372 if (!cgroup_on_dfl(dst_cgrp))
2373 return 0;
2374
2375 /* verify @dst_cgrp can host resources */
2376 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2377 return -EOPNOTSUPP;
2378
2379 /* mixables don't care */
2380 if (cgroup_is_mixable(dst_cgrp))
2381 return 0;
2382
2383 /*
2384 * If @dst_cgrp is already or can become a thread root or is
2385 * threaded, it doesn't matter.
2386 */
2387 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2388 return 0;
2389
2390 /* apply no-internal-process constraint */
2391 if (dst_cgrp->subtree_control)
2392 return -EBUSY;
2393
2394 return 0;
2395}
2396
2397/**
2398 * cgroup_migrate_finish - cleanup after attach
2399 * @mgctx: migration context
2400 *
2401 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2402 * those functions for details.
2403 */
2404void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2405{
2406 LIST_HEAD(preloaded);
2407 struct css_set *cset, *tmp_cset;
2408
2409 lockdep_assert_held(&cgroup_mutex);
2410
2411 spin_lock_irq(&css_set_lock);
2412
2413 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2414 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2415
2416 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2417 cset->mg_src_cgrp = NULL;
2418 cset->mg_dst_cgrp = NULL;
2419 cset->mg_dst_cset = NULL;
2420 list_del_init(&cset->mg_preload_node);
2421 put_css_set_locked(cset);
2422 }
2423
2424 spin_unlock_irq(&css_set_lock);
2425}
2426
2427/**
2428 * cgroup_migrate_add_src - add a migration source css_set
2429 * @src_cset: the source css_set to add
2430 * @dst_cgrp: the destination cgroup
2431 * @mgctx: migration context
2432 *
2433 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2434 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2435 * up by cgroup_migrate_finish().
2436 *
2437 * This function may be called without holding cgroup_threadgroup_rwsem
2438 * even if the target is a process. Threads may be created and destroyed
2439 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2440 * into play and the preloaded css_sets are guaranteed to cover all
2441 * migrations.
2442 */
2443void cgroup_migrate_add_src(struct css_set *src_cset,
2444 struct cgroup *dst_cgrp,
2445 struct cgroup_mgctx *mgctx)
2446{
2447 struct cgroup *src_cgrp;
2448
2449 lockdep_assert_held(&cgroup_mutex);
2450 lockdep_assert_held(&css_set_lock);
2451
2452 /*
2453 * If ->dead, @src_set is associated with one or more dead cgroups
2454 * and doesn't contain any migratable tasks. Ignore it early so
2455 * that the rest of migration path doesn't get confused by it.
2456 */
2457 if (src_cset->dead)
2458 return;
2459
2460 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2461
2462 if (!list_empty(&src_cset->mg_preload_node))
2463 return;
2464
2465 WARN_ON(src_cset->mg_src_cgrp);
2466 WARN_ON(src_cset->mg_dst_cgrp);
2467 WARN_ON(!list_empty(&src_cset->mg_tasks));
2468 WARN_ON(!list_empty(&src_cset->mg_node));
2469
2470 src_cset->mg_src_cgrp = src_cgrp;
2471 src_cset->mg_dst_cgrp = dst_cgrp;
2472 get_css_set(src_cset);
2473 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2474}
2475
2476/**
2477 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2478 * @mgctx: migration context
2479 *
2480 * Tasks are about to be moved and all the source css_sets have been
2481 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2482 * pins all destination css_sets, links each to its source, and append them
2483 * to @mgctx->preloaded_dst_csets.
2484 *
2485 * This function must be called after cgroup_migrate_add_src() has been
2486 * called on each migration source css_set. After migration is performed
2487 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2488 * @mgctx.
2489 */
2490int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2491{
2492 struct css_set *src_cset, *tmp_cset;
2493
2494 lockdep_assert_held(&cgroup_mutex);
2495
2496 /* look up the dst cset for each src cset and link it to src */
2497 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2498 mg_preload_node) {
2499 struct css_set *dst_cset;
2500 struct cgroup_subsys *ss;
2501 int ssid;
2502
2503 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2504 if (!dst_cset)
2505 goto err;
2506
2507 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2508
2509 /*
2510 * If src cset equals dst, it's noop. Drop the src.
2511 * cgroup_migrate() will skip the cset too. Note that we
2512 * can't handle src == dst as some nodes are used by both.
2513 */
2514 if (src_cset == dst_cset) {
2515 src_cset->mg_src_cgrp = NULL;
2516 src_cset->mg_dst_cgrp = NULL;
2517 list_del_init(&src_cset->mg_preload_node);
2518 put_css_set(src_cset);
2519 put_css_set(dst_cset);
2520 continue;
2521 }
2522
2523 src_cset->mg_dst_cset = dst_cset;
2524
2525 if (list_empty(&dst_cset->mg_preload_node))
2526 list_add_tail(&dst_cset->mg_preload_node,
2527 &mgctx->preloaded_dst_csets);
2528 else
2529 put_css_set(dst_cset);
2530
2531 for_each_subsys(ss, ssid)
2532 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2533 mgctx->ss_mask |= 1 << ssid;
2534 }
2535
2536 return 0;
2537err:
2538 cgroup_migrate_finish(mgctx);
2539 return -ENOMEM;
2540}
2541
2542/**
2543 * cgroup_migrate - migrate a process or task to a cgroup
2544 * @leader: the leader of the process or the task to migrate
2545 * @threadgroup: whether @leader points to the whole process or a single task
2546 * @mgctx: migration context
2547 *
2548 * Migrate a process or task denoted by @leader. If migrating a process,
2549 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2550 * responsible for invoking cgroup_migrate_add_src() and
2551 * cgroup_migrate_prepare_dst() on the targets before invoking this
2552 * function and following up with cgroup_migrate_finish().
2553 *
2554 * As long as a controller's ->can_attach() doesn't fail, this function is
2555 * guaranteed to succeed. This means that, excluding ->can_attach()
2556 * failure, when migrating multiple targets, the success or failure can be
2557 * decided for all targets by invoking group_migrate_prepare_dst() before
2558 * actually starting migrating.
2559 */
2560int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2561 struct cgroup_mgctx *mgctx)
2562{
2563 struct task_struct *task;
2564
2565 /*
2566 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2567 * already PF_EXITING could be freed from underneath us unless we
2568 * take an rcu_read_lock.
2569 */
2570 spin_lock_irq(&css_set_lock);
2571 rcu_read_lock();
2572 task = leader;
2573 do {
2574 cgroup_migrate_add_task(task, mgctx);
2575 if (!threadgroup)
2576 break;
2577 } while_each_thread(leader, task);
2578 rcu_read_unlock();
2579 spin_unlock_irq(&css_set_lock);
2580
2581 return cgroup_migrate_execute(mgctx);
2582}
2583
2584/**
2585 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2586 * @dst_cgrp: the cgroup to attach to
2587 * @leader: the task or the leader of the threadgroup to be attached
2588 * @threadgroup: attach the whole threadgroup?
2589 *
2590 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2591 */
2592int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2593 bool threadgroup)
2594{
2595 DEFINE_CGROUP_MGCTX(mgctx);
2596 struct task_struct *task;
2597 int ret;
2598
2599 ret = cgroup_migrate_vet_dst(dst_cgrp);
2600 if (ret)
2601 return ret;
2602
2603 /* look up all src csets */
2604 spin_lock_irq(&css_set_lock);
2605 rcu_read_lock();
2606 task = leader;
2607 do {
2608 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2609 if (!threadgroup)
2610 break;
2611 } while_each_thread(leader, task);
2612 rcu_read_unlock();
2613 spin_unlock_irq(&css_set_lock);
2614
2615 /* prepare dst csets and commit */
2616 ret = cgroup_migrate_prepare_dst(&mgctx);
2617 if (!ret)
2618 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2619
2620 cgroup_migrate_finish(&mgctx);
2621
2622 if (!ret)
2623 trace_cgroup_attach_task(dst_cgrp, leader, threadgroup);
2624
2625 return ret;
2626}
2627
2628struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2629 __acquires(&cgroup_threadgroup_rwsem)
2630{
2631 struct task_struct *tsk;
2632 pid_t pid;
2633
2634 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2635 return ERR_PTR(-EINVAL);
2636
2637 percpu_down_write(&cgroup_threadgroup_rwsem);
2638
2639 rcu_read_lock();
2640 if (pid) {
2641 tsk = find_task_by_vpid(pid);
2642 if (!tsk) {
2643 tsk = ERR_PTR(-ESRCH);
2644 goto out_unlock_threadgroup;
2645 }
2646 } else {
2647 tsk = current;
2648 }
2649
2650 if (threadgroup)
2651 tsk = tsk->group_leader;
2652
2653 /*
2654 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2655 * If userland migrates such a kthread to a non-root cgroup, it can
2656 * become trapped in a cpuset, or RT kthread may be born in a
2657 * cgroup with no rt_runtime allocated. Just say no.
2658 */
2659 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2660 tsk = ERR_PTR(-EINVAL);
2661 goto out_unlock_threadgroup;
2662 }
2663
2664 get_task_struct(tsk);
2665 goto out_unlock_rcu;
2666
2667out_unlock_threadgroup:
2668 percpu_up_write(&cgroup_threadgroup_rwsem);
2669out_unlock_rcu:
2670 rcu_read_unlock();
2671 return tsk;
2672}
2673
2674void cgroup_procs_write_finish(struct task_struct *task)
2675 __releases(&cgroup_threadgroup_rwsem)
2676{
2677 struct cgroup_subsys *ss;
2678 int ssid;
2679
2680 /* release reference from cgroup_procs_write_start() */
2681 put_task_struct(task);
2682
2683 percpu_up_write(&cgroup_threadgroup_rwsem);
2684 for_each_subsys(ss, ssid)
2685 if (ss->post_attach)
2686 ss->post_attach();
2687}
2688
2689static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2690{
2691 struct cgroup_subsys *ss;
2692 bool printed = false;
2693 int ssid;
2694
2695 do_each_subsys_mask(ss, ssid, ss_mask) {
2696 if (printed)
2697 seq_putc(seq, ' ');
2698 seq_printf(seq, "%s", ss->name);
2699 printed = true;
2700 } while_each_subsys_mask();
2701 if (printed)
2702 seq_putc(seq, '\n');
2703}
2704
2705/* show controllers which are enabled from the parent */
2706static int cgroup_controllers_show(struct seq_file *seq, void *v)
2707{
2708 struct cgroup *cgrp = seq_css(seq)->cgroup;
2709
2710 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2711 return 0;
2712}
2713
2714/* show controllers which are enabled for a given cgroup's children */
2715static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2716{
2717 struct cgroup *cgrp = seq_css(seq)->cgroup;
2718
2719 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2720 return 0;
2721}
2722
2723/**
2724 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2725 * @cgrp: root of the subtree to update csses for
2726 *
2727 * @cgrp's control masks have changed and its subtree's css associations
2728 * need to be updated accordingly. This function looks up all css_sets
2729 * which are attached to the subtree, creates the matching updated css_sets
2730 * and migrates the tasks to the new ones.
2731 */
2732static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2733{
2734 DEFINE_CGROUP_MGCTX(mgctx);
2735 struct cgroup_subsys_state *d_css;
2736 struct cgroup *dsct;
2737 struct css_set *src_cset;
2738 int ret;
2739
2740 lockdep_assert_held(&cgroup_mutex);
2741
2742 percpu_down_write(&cgroup_threadgroup_rwsem);
2743
2744 /* look up all csses currently attached to @cgrp's subtree */
2745 spin_lock_irq(&css_set_lock);
2746 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2747 struct cgrp_cset_link *link;
2748
2749 list_for_each_entry(link, &dsct->cset_links, cset_link)
2750 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2751 }
2752 spin_unlock_irq(&css_set_lock);
2753
2754 /* NULL dst indicates self on default hierarchy */
2755 ret = cgroup_migrate_prepare_dst(&mgctx);
2756 if (ret)
2757 goto out_finish;
2758
2759 spin_lock_irq(&css_set_lock);
2760 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2761 struct task_struct *task, *ntask;
2762
2763 /* all tasks in src_csets need to be migrated */
2764 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2765 cgroup_migrate_add_task(task, &mgctx);
2766 }
2767 spin_unlock_irq(&css_set_lock);
2768
2769 ret = cgroup_migrate_execute(&mgctx);
2770out_finish:
2771 cgroup_migrate_finish(&mgctx);
2772 percpu_up_write(&cgroup_threadgroup_rwsem);
2773 return ret;
2774}
2775
2776/**
2777 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2778 * @cgrp: root of the target subtree
2779 *
2780 * Because css offlining is asynchronous, userland may try to re-enable a
2781 * controller while the previous css is still around. This function grabs
2782 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2783 */
2784void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2785 __acquires(&cgroup_mutex)
2786{
2787 struct cgroup *dsct;
2788 struct cgroup_subsys_state *d_css;
2789 struct cgroup_subsys *ss;
2790 int ssid;
2791
2792restart:
2793 mutex_lock(&cgroup_mutex);
2794
2795 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2796 for_each_subsys(ss, ssid) {
2797 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2798 DEFINE_WAIT(wait);
2799
2800 if (!css || !percpu_ref_is_dying(&css->refcnt))
2801 continue;
2802
2803 cgroup_get_live(dsct);
2804 prepare_to_wait(&dsct->offline_waitq, &wait,
2805 TASK_UNINTERRUPTIBLE);
2806
2807 mutex_unlock(&cgroup_mutex);
2808 schedule();
2809 finish_wait(&dsct->offline_waitq, &wait);
2810
2811 cgroup_put(dsct);
2812 goto restart;
2813 }
2814 }
2815}
2816
2817/**
2818 * cgroup_save_control - save control masks of a subtree
2819 * @cgrp: root of the target subtree
2820 *
2821 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
2822 * prefixed fields for @cgrp's subtree including @cgrp itself.
2823 */
2824static void cgroup_save_control(struct cgroup *cgrp)
2825{
2826 struct cgroup *dsct;
2827 struct cgroup_subsys_state *d_css;
2828
2829 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2830 dsct->old_subtree_control = dsct->subtree_control;
2831 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2832 }
2833}
2834
2835/**
2836 * cgroup_propagate_control - refresh control masks of a subtree
2837 * @cgrp: root of the target subtree
2838 *
2839 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2840 * ->subtree_control and propagate controller availability through the
2841 * subtree so that descendants don't have unavailable controllers enabled.
2842 */
2843static void cgroup_propagate_control(struct cgroup *cgrp)
2844{
2845 struct cgroup *dsct;
2846 struct cgroup_subsys_state *d_css;
2847
2848 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2849 dsct->subtree_control &= cgroup_control(dsct);
2850 dsct->subtree_ss_mask =
2851 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2852 cgroup_ss_mask(dsct));
2853 }
2854}
2855
2856/**
2857 * cgroup_restore_control - restore control masks of a subtree
2858 * @cgrp: root of the target subtree
2859 *
2860 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
2861 * prefixed fields for @cgrp's subtree including @cgrp itself.
2862 */
2863static void cgroup_restore_control(struct cgroup *cgrp)
2864{
2865 struct cgroup *dsct;
2866 struct cgroup_subsys_state *d_css;
2867
2868 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2869 dsct->subtree_control = dsct->old_subtree_control;
2870 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
2871 }
2872}
2873
2874static bool css_visible(struct cgroup_subsys_state *css)
2875{
2876 struct cgroup_subsys *ss = css->ss;
2877 struct cgroup *cgrp = css->cgroup;
2878
2879 if (cgroup_control(cgrp) & (1 << ss->id))
2880 return true;
2881 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
2882 return false;
2883 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
2884}
2885
2886/**
2887 * cgroup_apply_control_enable - enable or show csses according to control
2888 * @cgrp: root of the target subtree
2889 *
2890 * Walk @cgrp's subtree and create new csses or make the existing ones
2891 * visible. A css is created invisible if it's being implicitly enabled
2892 * through dependency. An invisible css is made visible when the userland
2893 * explicitly enables it.
2894 *
2895 * Returns 0 on success, -errno on failure. On failure, csses which have
2896 * been processed already aren't cleaned up. The caller is responsible for
2897 * cleaning up with cgroup_apply_control_disable().
2898 */
2899static int cgroup_apply_control_enable(struct cgroup *cgrp)
2900{
2901 struct cgroup *dsct;
2902 struct cgroup_subsys_state *d_css;
2903 struct cgroup_subsys *ss;
2904 int ssid, ret;
2905
2906 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2907 for_each_subsys(ss, ssid) {
2908 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2909
2910 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2911
2912 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
2913 continue;
2914
2915 if (!css) {
2916 css = css_create(dsct, ss);
2917 if (IS_ERR(css))
2918 return PTR_ERR(css);
2919 }
2920
2921 if (css_visible(css)) {
2922 ret = css_populate_dir(css);
2923 if (ret)
2924 return ret;
2925 }
2926 }
2927 }
2928
2929 return 0;
2930}
2931
2932/**
2933 * cgroup_apply_control_disable - kill or hide csses according to control
2934 * @cgrp: root of the target subtree
2935 *
2936 * Walk @cgrp's subtree and kill and hide csses so that they match
2937 * cgroup_ss_mask() and cgroup_visible_mask().
2938 *
2939 * A css is hidden when the userland requests it to be disabled while other
2940 * subsystems are still depending on it. The css must not actively control
2941 * resources and be in the vanilla state if it's made visible again later.
2942 * Controllers which may be depended upon should provide ->css_reset() for
2943 * this purpose.
2944 */
2945static void cgroup_apply_control_disable(struct cgroup *cgrp)
2946{
2947 struct cgroup *dsct;
2948 struct cgroup_subsys_state *d_css;
2949 struct cgroup_subsys *ss;
2950 int ssid;
2951
2952 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2953 for_each_subsys(ss, ssid) {
2954 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2955
2956 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2957
2958 if (!css)
2959 continue;
2960
2961 if (css->parent &&
2962 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
2963 kill_css(css);
2964 } else if (!css_visible(css)) {
2965 css_clear_dir(css);
2966 if (ss->css_reset)
2967 ss->css_reset(css);
2968 }
2969 }
2970 }
2971}
2972
2973/**
2974 * cgroup_apply_control - apply control mask updates to the subtree
2975 * @cgrp: root of the target subtree
2976 *
2977 * subsystems can be enabled and disabled in a subtree using the following
2978 * steps.
2979 *
2980 * 1. Call cgroup_save_control() to stash the current state.
2981 * 2. Update ->subtree_control masks in the subtree as desired.
2982 * 3. Call cgroup_apply_control() to apply the changes.
2983 * 4. Optionally perform other related operations.
2984 * 5. Call cgroup_finalize_control() to finish up.
2985 *
2986 * This function implements step 3 and propagates the mask changes
2987 * throughout @cgrp's subtree, updates csses accordingly and perform
2988 * process migrations.
2989 */
2990static int cgroup_apply_control(struct cgroup *cgrp)
2991{
2992 int ret;
2993
2994 cgroup_propagate_control(cgrp);
2995
2996 ret = cgroup_apply_control_enable(cgrp);
2997 if (ret)
2998 return ret;
2999
3000 /*
3001 * At this point, cgroup_e_css() results reflect the new csses
3002 * making the following cgroup_update_dfl_csses() properly update
3003 * css associations of all tasks in the subtree.
3004 */
3005 ret = cgroup_update_dfl_csses(cgrp);
3006 if (ret)
3007 return ret;
3008
3009 return 0;
3010}
3011
3012/**
3013 * cgroup_finalize_control - finalize control mask update
3014 * @cgrp: root of the target subtree
3015 * @ret: the result of the update
3016 *
3017 * Finalize control mask update. See cgroup_apply_control() for more info.
3018 */
3019static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3020{
3021 if (ret) {
3022 cgroup_restore_control(cgrp);
3023 cgroup_propagate_control(cgrp);
3024 }
3025
3026 cgroup_apply_control_disable(cgrp);
3027}
3028
3029static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3030{
3031 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3032
3033 /* if nothing is getting enabled, nothing to worry about */
3034 if (!enable)
3035 return 0;
3036
3037 /* can @cgrp host any resources? */
3038 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3039 return -EOPNOTSUPP;
3040
3041 /* mixables don't care */
3042 if (cgroup_is_mixable(cgrp))
3043 return 0;
3044
3045 if (domain_enable) {
3046 /* can't enable domain controllers inside a thread subtree */
3047 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3048 return -EOPNOTSUPP;
3049 } else {
3050 /*
3051 * Threaded controllers can handle internal competitions
3052 * and are always allowed inside a (prospective) thread
3053 * subtree.
3054 */
3055 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3056 return 0;
3057 }
3058
3059 /*
3060 * Controllers can't be enabled for a cgroup with tasks to avoid
3061 * child cgroups competing against tasks.
3062 */
3063 if (cgroup_has_tasks(cgrp))
3064 return -EBUSY;
3065
3066 return 0;
3067}
3068
3069/* change the enabled child controllers for a cgroup in the default hierarchy */
3070static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3071 char *buf, size_t nbytes,
3072 loff_t off)
3073{
3074 u16 enable = 0, disable = 0;
3075 struct cgroup *cgrp, *child;
3076 struct cgroup_subsys *ss;
3077 char *tok;
3078 int ssid, ret;
3079
3080 /*
3081 * Parse input - space separated list of subsystem names prefixed
3082 * with either + or -.
3083 */
3084 buf = strstrip(buf);
3085 while ((tok = strsep(&buf, " "))) {
3086 if (tok[0] == '\0')
3087 continue;
3088 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3089 if (!cgroup_ssid_enabled(ssid) ||
3090 strcmp(tok + 1, ss->name))
3091 continue;
3092
3093 if (*tok == '+') {
3094 enable |= 1 << ssid;
3095 disable &= ~(1 << ssid);
3096 } else if (*tok == '-') {
3097 disable |= 1 << ssid;
3098 enable &= ~(1 << ssid);
3099 } else {
3100 return -EINVAL;
3101 }
3102 break;
3103 } while_each_subsys_mask();
3104 if (ssid == CGROUP_SUBSYS_COUNT)
3105 return -EINVAL;
3106 }
3107
3108 cgrp = cgroup_kn_lock_live(of->kn, true);
3109 if (!cgrp)
3110 return -ENODEV;
3111
3112 for_each_subsys(ss, ssid) {
3113 if (enable & (1 << ssid)) {
3114 if (cgrp->subtree_control & (1 << ssid)) {
3115 enable &= ~(1 << ssid);
3116 continue;
3117 }
3118
3119 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3120 ret = -ENOENT;
3121 goto out_unlock;
3122 }
3123 } else if (disable & (1 << ssid)) {
3124 if (!(cgrp->subtree_control & (1 << ssid))) {
3125 disable &= ~(1 << ssid);
3126 continue;
3127 }
3128
3129 /* a child has it enabled? */
3130 cgroup_for_each_live_child(child, cgrp) {
3131 if (child->subtree_control & (1 << ssid)) {
3132 ret = -EBUSY;
3133 goto out_unlock;
3134 }
3135 }
3136 }
3137 }
3138
3139 if (!enable && !disable) {
3140 ret = 0;
3141 goto out_unlock;
3142 }
3143
3144 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3145 if (ret)
3146 goto out_unlock;
3147
3148 /* save and update control masks and prepare csses */
3149 cgroup_save_control(cgrp);
3150
3151 cgrp->subtree_control |= enable;
3152 cgrp->subtree_control &= ~disable;
3153
3154 ret = cgroup_apply_control(cgrp);
3155 cgroup_finalize_control(cgrp, ret);
3156 if (ret)
3157 goto out_unlock;
3158
3159 kernfs_activate(cgrp->kn);
3160out_unlock:
3161 cgroup_kn_unlock(of->kn);
3162 return ret ?: nbytes;
3163}
3164
3165/**
3166 * cgroup_enable_threaded - make @cgrp threaded
3167 * @cgrp: the target cgroup
3168 *
3169 * Called when "threaded" is written to the cgroup.type interface file and
3170 * tries to make @cgrp threaded and join the parent's resource domain.
3171 * This function is never called on the root cgroup as cgroup.type doesn't
3172 * exist on it.
3173 */
3174static int cgroup_enable_threaded(struct cgroup *cgrp)
3175{
3176 struct cgroup *parent = cgroup_parent(cgrp);
3177 struct cgroup *dom_cgrp = parent->dom_cgrp;
3178 int ret;
3179
3180 lockdep_assert_held(&cgroup_mutex);
3181
3182 /* noop if already threaded */
3183 if (cgroup_is_threaded(cgrp))
3184 return 0;
3185
3186 /*
3187 * If @cgroup is populated or has domain controllers enabled, it
3188 * can't be switched. While the below cgroup_can_be_thread_root()
3189 * test can catch the same conditions, that's only when @parent is
3190 * not mixable, so let's check it explicitly.
3191 */
3192 if (cgroup_is_populated(cgrp) ||
3193 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3194 return -EOPNOTSUPP;
3195
3196 /* we're joining the parent's domain, ensure its validity */
3197 if (!cgroup_is_valid_domain(dom_cgrp) ||
3198 !cgroup_can_be_thread_root(dom_cgrp))
3199 return -EOPNOTSUPP;
3200
3201 /*
3202 * The following shouldn't cause actual migrations and should
3203 * always succeed.
3204 */
3205 cgroup_save_control(cgrp);
3206
3207 cgrp->dom_cgrp = dom_cgrp;
3208 ret = cgroup_apply_control(cgrp);
3209 if (!ret)
3210 parent->nr_threaded_children++;
3211 else
3212 cgrp->dom_cgrp = cgrp;
3213
3214 cgroup_finalize_control(cgrp, ret);
3215 return ret;
3216}
3217
3218static int cgroup_type_show(struct seq_file *seq, void *v)
3219{
3220 struct cgroup *cgrp = seq_css(seq)->cgroup;
3221
3222 if (cgroup_is_threaded(cgrp))
3223 seq_puts(seq, "threaded\n");
3224 else if (!cgroup_is_valid_domain(cgrp))
3225 seq_puts(seq, "domain invalid\n");
3226 else if (cgroup_is_thread_root(cgrp))
3227 seq_puts(seq, "domain threaded\n");
3228 else
3229 seq_puts(seq, "domain\n");
3230
3231 return 0;
3232}
3233
3234static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3235 size_t nbytes, loff_t off)
3236{
3237 struct cgroup *cgrp;
3238 int ret;
3239
3240 /* only switching to threaded mode is supported */
3241 if (strcmp(strstrip(buf), "threaded"))
3242 return -EINVAL;
3243
3244 cgrp = cgroup_kn_lock_live(of->kn, false);
3245 if (!cgrp)
3246 return -ENOENT;
3247
3248 /* threaded can only be enabled */
3249 ret = cgroup_enable_threaded(cgrp);
3250
3251 cgroup_kn_unlock(of->kn);
3252 return ret ?: nbytes;
3253}
3254
3255static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3256{
3257 struct cgroup *cgrp = seq_css(seq)->cgroup;
3258 int descendants = READ_ONCE(cgrp->max_descendants);
3259
3260 if (descendants == INT_MAX)
3261 seq_puts(seq, "max\n");
3262 else
3263 seq_printf(seq, "%d\n", descendants);
3264
3265 return 0;
3266}
3267
3268static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3269 char *buf, size_t nbytes, loff_t off)
3270{
3271 struct cgroup *cgrp;
3272 int descendants;
3273 ssize_t ret;
3274
3275 buf = strstrip(buf);
3276 if (!strcmp(buf, "max")) {
3277 descendants = INT_MAX;
3278 } else {
3279 ret = kstrtoint(buf, 0, &descendants);
3280 if (ret)
3281 return ret;
3282 }
3283
3284 if (descendants < 0)
3285 return -ERANGE;
3286
3287 cgrp = cgroup_kn_lock_live(of->kn, false);
3288 if (!cgrp)
3289 return -ENOENT;
3290
3291 cgrp->max_descendants = descendants;
3292
3293 cgroup_kn_unlock(of->kn);
3294
3295 return nbytes;
3296}
3297
3298static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3299{
3300 struct cgroup *cgrp = seq_css(seq)->cgroup;
3301 int depth = READ_ONCE(cgrp->max_depth);
3302
3303 if (depth == INT_MAX)
3304 seq_puts(seq, "max\n");
3305 else
3306 seq_printf(seq, "%d\n", depth);
3307
3308 return 0;
3309}
3310
3311static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3312 char *buf, size_t nbytes, loff_t off)
3313{
3314 struct cgroup *cgrp;
3315 ssize_t ret;
3316 int depth;
3317
3318 buf = strstrip(buf);
3319 if (!strcmp(buf, "max")) {
3320 depth = INT_MAX;
3321 } else {
3322 ret = kstrtoint(buf, 0, &depth);
3323 if (ret)
3324 return ret;
3325 }
3326
3327 if (depth < 0)
3328 return -ERANGE;
3329
3330 cgrp = cgroup_kn_lock_live(of->kn, false);
3331 if (!cgrp)
3332 return -ENOENT;
3333
3334 cgrp->max_depth = depth;
3335
3336 cgroup_kn_unlock(of->kn);
3337
3338 return nbytes;
3339}
3340
3341static int cgroup_events_show(struct seq_file *seq, void *v)
3342{
3343 seq_printf(seq, "populated %d\n",
3344 cgroup_is_populated(seq_css(seq)->cgroup));
3345 return 0;
3346}
3347
3348static int cgroup_stat_show(struct seq_file *seq, void *v)
3349{
3350 struct cgroup *cgroup = seq_css(seq)->cgroup;
3351
3352 seq_printf(seq, "nr_descendants %d\n",
3353 cgroup->nr_descendants);
3354 seq_printf(seq, "nr_dying_descendants %d\n",
3355 cgroup->nr_dying_descendants);
3356
3357 return 0;
3358}
3359
3360static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3361 struct cgroup *cgrp, int ssid)
3362{
3363 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3364 struct cgroup_subsys_state *css;
3365 int ret;
3366
3367 if (!ss->css_extra_stat_show)
3368 return 0;
3369
3370 css = cgroup_tryget_css(cgrp, ss);
3371 if (!css)
3372 return 0;
3373
3374 ret = ss->css_extra_stat_show(seq, css);
3375 css_put(css);
3376 return ret;
3377}
3378
3379static int cpu_stat_show(struct seq_file *seq, void *v)
3380{
3381 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3382 int ret = 0;
3383
3384 cgroup_stat_show_cputime(seq);
3385#ifdef CONFIG_CGROUP_SCHED
3386 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3387#endif
3388 return ret;
3389}
3390
3391static int cgroup_file_open(struct kernfs_open_file *of)
3392{
3393 struct cftype *cft = of->kn->priv;
3394
3395 if (cft->open)
3396 return cft->open(of);
3397 return 0;
3398}
3399
3400static void cgroup_file_release(struct kernfs_open_file *of)
3401{
3402 struct cftype *cft = of->kn->priv;
3403
3404 if (cft->release)
3405 cft->release(of);
3406}
3407
3408static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3409 size_t nbytes, loff_t off)
3410{
3411 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3412 struct cgroup *cgrp = of->kn->parent->priv;
3413 struct cftype *cft = of->kn->priv;
3414 struct cgroup_subsys_state *css;
3415 int ret;
3416
3417 /*
3418 * If namespaces are delegation boundaries, disallow writes to
3419 * files in an non-init namespace root from inside the namespace
3420 * except for the files explicitly marked delegatable -
3421 * cgroup.procs and cgroup.subtree_control.
3422 */
3423 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3424 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3425 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3426 return -EPERM;
3427
3428 if (cft->write)
3429 return cft->write(of, buf, nbytes, off);
3430
3431 /*
3432 * kernfs guarantees that a file isn't deleted with operations in
3433 * flight, which means that the matching css is and stays alive and
3434 * doesn't need to be pinned. The RCU locking is not necessary
3435 * either. It's just for the convenience of using cgroup_css().
3436 */
3437 rcu_read_lock();
3438 css = cgroup_css(cgrp, cft->ss);
3439 rcu_read_unlock();
3440
3441 if (cft->write_u64) {
3442 unsigned long long v;
3443 ret = kstrtoull(buf, 0, &v);
3444 if (!ret)
3445 ret = cft->write_u64(css, cft, v);
3446 } else if (cft->write_s64) {
3447 long long v;
3448 ret = kstrtoll(buf, 0, &v);
3449 if (!ret)
3450 ret = cft->write_s64(css, cft, v);
3451 } else {
3452 ret = -EINVAL;
3453 }
3454
3455 return ret ?: nbytes;
3456}
3457
3458static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3459{
3460 return seq_cft(seq)->seq_start(seq, ppos);
3461}
3462
3463static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3464{
3465 return seq_cft(seq)->seq_next(seq, v, ppos);
3466}
3467
3468static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3469{
3470 if (seq_cft(seq)->seq_stop)
3471 seq_cft(seq)->seq_stop(seq, v);
3472}
3473
3474static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3475{
3476 struct cftype *cft = seq_cft(m);
3477 struct cgroup_subsys_state *css = seq_css(m);
3478
3479 if (cft->seq_show)
3480 return cft->seq_show(m, arg);
3481
3482 if (cft->read_u64)
3483 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3484 else if (cft->read_s64)
3485 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3486 else
3487 return -EINVAL;
3488 return 0;
3489}
3490
3491static struct kernfs_ops cgroup_kf_single_ops = {
3492 .atomic_write_len = PAGE_SIZE,
3493 .open = cgroup_file_open,
3494 .release = cgroup_file_release,
3495 .write = cgroup_file_write,
3496 .seq_show = cgroup_seqfile_show,
3497};
3498
3499static struct kernfs_ops cgroup_kf_ops = {
3500 .atomic_write_len = PAGE_SIZE,
3501 .open = cgroup_file_open,
3502 .release = cgroup_file_release,
3503 .write = cgroup_file_write,
3504 .seq_start = cgroup_seqfile_start,
3505 .seq_next = cgroup_seqfile_next,
3506 .seq_stop = cgroup_seqfile_stop,
3507 .seq_show = cgroup_seqfile_show,
3508};
3509
3510/* set uid and gid of cgroup dirs and files to that of the creator */
3511static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3512{
3513 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3514 .ia_uid = current_fsuid(),
3515 .ia_gid = current_fsgid(), };
3516
3517 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3518 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3519 return 0;
3520
3521 return kernfs_setattr(kn, &iattr);
3522}
3523
3524static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3525 struct cftype *cft)
3526{
3527 char name[CGROUP_FILE_NAME_MAX];
3528 struct kernfs_node *kn;
3529 struct lock_class_key *key = NULL;
3530 int ret;
3531
3532#ifdef CONFIG_DEBUG_LOCK_ALLOC
3533 key = &cft->lockdep_key;
3534#endif
3535 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3536 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3537 NULL, key);
3538 if (IS_ERR(kn))
3539 return PTR_ERR(kn);
3540
3541 ret = cgroup_kn_set_ugid(kn);
3542 if (ret) {
3543 kernfs_remove(kn);
3544 return ret;
3545 }
3546
3547 if (cft->file_offset) {
3548 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3549
3550 spin_lock_irq(&cgroup_file_kn_lock);
3551 cfile->kn = kn;
3552 spin_unlock_irq(&cgroup_file_kn_lock);
3553 }
3554
3555 return 0;
3556}
3557
3558/**
3559 * cgroup_addrm_files - add or remove files to a cgroup directory
3560 * @css: the target css
3561 * @cgrp: the target cgroup (usually css->cgroup)
3562 * @cfts: array of cftypes to be added
3563 * @is_add: whether to add or remove
3564 *
3565 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3566 * For removals, this function never fails.
3567 */
3568static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3569 struct cgroup *cgrp, struct cftype cfts[],
3570 bool is_add)
3571{
3572 struct cftype *cft, *cft_end = NULL;
3573 int ret = 0;
3574
3575 lockdep_assert_held(&cgroup_mutex);
3576
3577restart:
3578 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3579 /* does cft->flags tell us to skip this file on @cgrp? */
3580 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3581 continue;
3582 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3583 continue;
3584 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3585 continue;
3586 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3587 continue;
3588
3589 if (is_add) {
3590 ret = cgroup_add_file(css, cgrp, cft);
3591 if (ret) {
3592 pr_warn("%s: failed to add %s, err=%d\n",
3593 __func__, cft->name, ret);
3594 cft_end = cft;
3595 is_add = false;
3596 goto restart;
3597 }
3598 } else {
3599 cgroup_rm_file(cgrp, cft);
3600 }
3601 }
3602 return ret;
3603}
3604
3605static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3606{
3607 struct cgroup_subsys *ss = cfts[0].ss;
3608 struct cgroup *root = &ss->root->cgrp;
3609 struct cgroup_subsys_state *css;
3610 int ret = 0;
3611
3612 lockdep_assert_held(&cgroup_mutex);
3613
3614 /* add/rm files for all cgroups created before */
3615 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3616 struct cgroup *cgrp = css->cgroup;
3617
3618 if (!(css->flags & CSS_VISIBLE))
3619 continue;
3620
3621 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3622 if (ret)
3623 break;
3624 }
3625
3626 if (is_add && !ret)
3627 kernfs_activate(root->kn);
3628 return ret;
3629}
3630
3631static void cgroup_exit_cftypes(struct cftype *cfts)
3632{
3633 struct cftype *cft;
3634
3635 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3636 /* free copy for custom atomic_write_len, see init_cftypes() */
3637 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3638 kfree(cft->kf_ops);
3639 cft->kf_ops = NULL;
3640 cft->ss = NULL;
3641
3642 /* revert flags set by cgroup core while adding @cfts */
3643 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3644 }
3645}
3646
3647static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3648{
3649 struct cftype *cft;
3650
3651 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3652 struct kernfs_ops *kf_ops;
3653
3654 WARN_ON(cft->ss || cft->kf_ops);
3655
3656 if (cft->seq_start)
3657 kf_ops = &cgroup_kf_ops;
3658 else
3659 kf_ops = &cgroup_kf_single_ops;
3660
3661 /*
3662 * Ugh... if @cft wants a custom max_write_len, we need to
3663 * make a copy of kf_ops to set its atomic_write_len.
3664 */
3665 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3666 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3667 if (!kf_ops) {
3668 cgroup_exit_cftypes(cfts);
3669 return -ENOMEM;
3670 }
3671 kf_ops->atomic_write_len = cft->max_write_len;
3672 }
3673
3674 cft->kf_ops = kf_ops;
3675 cft->ss = ss;
3676 }
3677
3678 return 0;
3679}
3680
3681static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3682{
3683 lockdep_assert_held(&cgroup_mutex);
3684
3685 if (!cfts || !cfts[0].ss)
3686 return -ENOENT;
3687
3688 list_del(&cfts->node);
3689 cgroup_apply_cftypes(cfts, false);
3690 cgroup_exit_cftypes(cfts);
3691 return 0;
3692}
3693
3694/**
3695 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3696 * @cfts: zero-length name terminated array of cftypes
3697 *
3698 * Unregister @cfts. Files described by @cfts are removed from all
3699 * existing cgroups and all future cgroups won't have them either. This
3700 * function can be called anytime whether @cfts' subsys is attached or not.
3701 *
3702 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3703 * registered.
3704 */
3705int cgroup_rm_cftypes(struct cftype *cfts)
3706{
3707 int ret;
3708
3709 mutex_lock(&cgroup_mutex);
3710 ret = cgroup_rm_cftypes_locked(cfts);
3711 mutex_unlock(&cgroup_mutex);
3712 return ret;
3713}
3714
3715/**
3716 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3717 * @ss: target cgroup subsystem
3718 * @cfts: zero-length name terminated array of cftypes
3719 *
3720 * Register @cfts to @ss. Files described by @cfts are created for all
3721 * existing cgroups to which @ss is attached and all future cgroups will
3722 * have them too. This function can be called anytime whether @ss is
3723 * attached or not.
3724 *
3725 * Returns 0 on successful registration, -errno on failure. Note that this
3726 * function currently returns 0 as long as @cfts registration is successful
3727 * even if some file creation attempts on existing cgroups fail.
3728 */
3729static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3730{
3731 int ret;
3732
3733 if (!cgroup_ssid_enabled(ss->id))
3734 return 0;
3735
3736 if (!cfts || cfts[0].name[0] == '\0')
3737 return 0;
3738
3739 ret = cgroup_init_cftypes(ss, cfts);
3740 if (ret)
3741 return ret;
3742
3743 mutex_lock(&cgroup_mutex);
3744
3745 list_add_tail(&cfts->node, &ss->cfts);
3746 ret = cgroup_apply_cftypes(cfts, true);
3747 if (ret)
3748 cgroup_rm_cftypes_locked(cfts);
3749
3750 mutex_unlock(&cgroup_mutex);
3751 return ret;
3752}
3753
3754/**
3755 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3756 * @ss: target cgroup subsystem
3757 * @cfts: zero-length name terminated array of cftypes
3758 *
3759 * Similar to cgroup_add_cftypes() but the added files are only used for
3760 * the default hierarchy.
3761 */
3762int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3763{
3764 struct cftype *cft;
3765
3766 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3767 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3768 return cgroup_add_cftypes(ss, cfts);
3769}
3770
3771/**
3772 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3773 * @ss: target cgroup subsystem
3774 * @cfts: zero-length name terminated array of cftypes
3775 *
3776 * Similar to cgroup_add_cftypes() but the added files are only used for
3777 * the legacy hierarchies.
3778 */
3779int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3780{
3781 struct cftype *cft;
3782
3783 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3784 cft->flags |= __CFTYPE_NOT_ON_DFL;
3785 return cgroup_add_cftypes(ss, cfts);
3786}
3787
3788/**
3789 * cgroup_file_notify - generate a file modified event for a cgroup_file
3790 * @cfile: target cgroup_file
3791 *
3792 * @cfile must have been obtained by setting cftype->file_offset.
3793 */
3794void cgroup_file_notify(struct cgroup_file *cfile)
3795{
3796 unsigned long flags;
3797
3798 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3799 if (cfile->kn)
3800 kernfs_notify(cfile->kn);
3801 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3802}
3803
3804/**
3805 * css_next_child - find the next child of a given css
3806 * @pos: the current position (%NULL to initiate traversal)
3807 * @parent: css whose children to walk
3808 *
3809 * This function returns the next child of @parent and should be called
3810 * under either cgroup_mutex or RCU read lock. The only requirement is
3811 * that @parent and @pos are accessible. The next sibling is guaranteed to
3812 * be returned regardless of their states.
3813 *
3814 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3815 * css which finished ->css_online() is guaranteed to be visible in the
3816 * future iterations and will stay visible until the last reference is put.
3817 * A css which hasn't finished ->css_online() or already finished
3818 * ->css_offline() may show up during traversal. It's each subsystem's
3819 * responsibility to synchronize against on/offlining.
3820 */
3821struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3822 struct cgroup_subsys_state *parent)
3823{
3824 struct cgroup_subsys_state *next;
3825
3826 cgroup_assert_mutex_or_rcu_locked();
3827
3828 /*
3829 * @pos could already have been unlinked from the sibling list.
3830 * Once a cgroup is removed, its ->sibling.next is no longer
3831 * updated when its next sibling changes. CSS_RELEASED is set when
3832 * @pos is taken off list, at which time its next pointer is valid,
3833 * and, as releases are serialized, the one pointed to by the next
3834 * pointer is guaranteed to not have started release yet. This
3835 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3836 * critical section, the one pointed to by its next pointer is
3837 * guaranteed to not have finished its RCU grace period even if we
3838 * have dropped rcu_read_lock() inbetween iterations.
3839 *
3840 * If @pos has CSS_RELEASED set, its next pointer can't be
3841 * dereferenced; however, as each css is given a monotonically
3842 * increasing unique serial number and always appended to the
3843 * sibling list, the next one can be found by walking the parent's
3844 * children until the first css with higher serial number than
3845 * @pos's. While this path can be slower, it happens iff iteration
3846 * races against release and the race window is very small.
3847 */
3848 if (!pos) {
3849 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3850 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3851 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3852 } else {
3853 list_for_each_entry_rcu(next, &parent->children, sibling)
3854 if (next->serial_nr > pos->serial_nr)
3855 break;
3856 }
3857
3858 /*
3859 * @next, if not pointing to the head, can be dereferenced and is
3860 * the next sibling.
3861 */
3862 if (&next->sibling != &parent->children)
3863 return next;
3864 return NULL;
3865}
3866
3867/**
3868 * css_next_descendant_pre - find the next descendant for pre-order walk
3869 * @pos: the current position (%NULL to initiate traversal)
3870 * @root: css whose descendants to walk
3871 *
3872 * To be used by css_for_each_descendant_pre(). Find the next descendant
3873 * to visit for pre-order traversal of @root's descendants. @root is
3874 * included in the iteration and the first node to be visited.
3875 *
3876 * While this function requires cgroup_mutex or RCU read locking, it
3877 * doesn't require the whole traversal to be contained in a single critical
3878 * section. This function will return the correct next descendant as long
3879 * as both @pos and @root are accessible and @pos is a descendant of @root.
3880 *
3881 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3882 * css which finished ->css_online() is guaranteed to be visible in the
3883 * future iterations and will stay visible until the last reference is put.
3884 * A css which hasn't finished ->css_online() or already finished
3885 * ->css_offline() may show up during traversal. It's each subsystem's
3886 * responsibility to synchronize against on/offlining.
3887 */
3888struct cgroup_subsys_state *
3889css_next_descendant_pre(struct cgroup_subsys_state *pos,
3890 struct cgroup_subsys_state *root)
3891{
3892 struct cgroup_subsys_state *next;
3893
3894 cgroup_assert_mutex_or_rcu_locked();
3895
3896 /* if first iteration, visit @root */
3897 if (!pos)
3898 return root;
3899
3900 /* visit the first child if exists */
3901 next = css_next_child(NULL, pos);
3902 if (next)
3903 return next;
3904
3905 /* no child, visit my or the closest ancestor's next sibling */
3906 while (pos != root) {
3907 next = css_next_child(pos, pos->parent);
3908 if (next)
3909 return next;
3910 pos = pos->parent;
3911 }
3912
3913 return NULL;
3914}
3915
3916/**
3917 * css_rightmost_descendant - return the rightmost descendant of a css
3918 * @pos: css of interest
3919 *
3920 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3921 * is returned. This can be used during pre-order traversal to skip
3922 * subtree of @pos.
3923 *
3924 * While this function requires cgroup_mutex or RCU read locking, it
3925 * doesn't require the whole traversal to be contained in a single critical
3926 * section. This function will return the correct rightmost descendant as
3927 * long as @pos is accessible.
3928 */
3929struct cgroup_subsys_state *
3930css_rightmost_descendant(struct cgroup_subsys_state *pos)
3931{
3932 struct cgroup_subsys_state *last, *tmp;
3933
3934 cgroup_assert_mutex_or_rcu_locked();
3935
3936 do {
3937 last = pos;
3938 /* ->prev isn't RCU safe, walk ->next till the end */
3939 pos = NULL;
3940 css_for_each_child(tmp, last)
3941 pos = tmp;
3942 } while (pos);
3943
3944 return last;
3945}
3946
3947static struct cgroup_subsys_state *
3948css_leftmost_descendant(struct cgroup_subsys_state *pos)
3949{
3950 struct cgroup_subsys_state *last;
3951
3952 do {
3953 last = pos;
3954 pos = css_next_child(NULL, pos);
3955 } while (pos);
3956
3957 return last;
3958}
3959
3960/**
3961 * css_next_descendant_post - find the next descendant for post-order walk
3962 * @pos: the current position (%NULL to initiate traversal)
3963 * @root: css whose descendants to walk
3964 *
3965 * To be used by css_for_each_descendant_post(). Find the next descendant
3966 * to visit for post-order traversal of @root's descendants. @root is
3967 * included in the iteration and the last node to be visited.
3968 *
3969 * While this function requires cgroup_mutex or RCU read locking, it
3970 * doesn't require the whole traversal to be contained in a single critical
3971 * section. This function will return the correct next descendant as long
3972 * as both @pos and @cgroup are accessible and @pos is a descendant of
3973 * @cgroup.
3974 *
3975 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3976 * css which finished ->css_online() is guaranteed to be visible in the
3977 * future iterations and will stay visible until the last reference is put.
3978 * A css which hasn't finished ->css_online() or already finished
3979 * ->css_offline() may show up during traversal. It's each subsystem's
3980 * responsibility to synchronize against on/offlining.
3981 */
3982struct cgroup_subsys_state *
3983css_next_descendant_post(struct cgroup_subsys_state *pos,
3984 struct cgroup_subsys_state *root)
3985{
3986 struct cgroup_subsys_state *next;
3987
3988 cgroup_assert_mutex_or_rcu_locked();
3989
3990 /* if first iteration, visit leftmost descendant which may be @root */
3991 if (!pos)
3992 return css_leftmost_descendant(root);
3993
3994 /* if we visited @root, we're done */
3995 if (pos == root)
3996 return NULL;
3997
3998 /* if there's an unvisited sibling, visit its leftmost descendant */
3999 next = css_next_child(pos, pos->parent);
4000 if (next)
4001 return css_leftmost_descendant(next);
4002
4003 /* no sibling left, visit parent */
4004 return pos->parent;
4005}
4006
4007/**
4008 * css_has_online_children - does a css have online children
4009 * @css: the target css
4010 *
4011 * Returns %true if @css has any online children; otherwise, %false. This
4012 * function can be called from any context but the caller is responsible
4013 * for synchronizing against on/offlining as necessary.
4014 */
4015bool css_has_online_children(struct cgroup_subsys_state *css)
4016{
4017 struct cgroup_subsys_state *child;
4018 bool ret = false;
4019
4020 rcu_read_lock();
4021 css_for_each_child(child, css) {
4022 if (child->flags & CSS_ONLINE) {
4023 ret = true;
4024 break;
4025 }
4026 }
4027 rcu_read_unlock();
4028 return ret;
4029}
4030
4031static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4032{
4033 struct list_head *l;
4034 struct cgrp_cset_link *link;
4035 struct css_set *cset;
4036
4037 lockdep_assert_held(&css_set_lock);
4038
4039 /* find the next threaded cset */
4040 if (it->tcset_pos) {
4041 l = it->tcset_pos->next;
4042
4043 if (l != it->tcset_head) {
4044 it->tcset_pos = l;
4045 return container_of(l, struct css_set,
4046 threaded_csets_node);
4047 }
4048
4049 it->tcset_pos = NULL;
4050 }
4051
4052 /* find the next cset */
4053 l = it->cset_pos;
4054 l = l->next;
4055 if (l == it->cset_head) {
4056 it->cset_pos = NULL;
4057 return NULL;
4058 }
4059
4060 if (it->ss) {
4061 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4062 } else {
4063 link = list_entry(l, struct cgrp_cset_link, cset_link);
4064 cset = link->cset;
4065 }
4066
4067 it->cset_pos = l;
4068
4069 /* initialize threaded css_set walking */
4070 if (it->flags & CSS_TASK_ITER_THREADED) {
4071 if (it->cur_dcset)
4072 put_css_set_locked(it->cur_dcset);
4073 it->cur_dcset = cset;
4074 get_css_set(cset);
4075
4076 it->tcset_head = &cset->threaded_csets;
4077 it->tcset_pos = &cset->threaded_csets;
4078 }
4079
4080 return cset;
4081}
4082
4083/**
4084 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4085 * @it: the iterator to advance
4086 *
4087 * Advance @it to the next css_set to walk.
4088 */
4089static void css_task_iter_advance_css_set(struct css_task_iter *it)
4090{
4091 struct css_set *cset;
4092
4093 lockdep_assert_held(&css_set_lock);
4094
4095 /* Advance to the next non-empty css_set */
4096 do {
4097 cset = css_task_iter_next_css_set(it);
4098 if (!cset) {
4099 it->task_pos = NULL;
4100 return;
4101 }
4102 } while (!css_set_populated(cset));
4103
4104 if (!list_empty(&cset->tasks))
4105 it->task_pos = cset->tasks.next;
4106 else
4107 it->task_pos = cset->mg_tasks.next;
4108
4109 it->tasks_head = &cset->tasks;
4110 it->mg_tasks_head = &cset->mg_tasks;
4111
4112 /*
4113 * We don't keep css_sets locked across iteration steps and thus
4114 * need to take steps to ensure that iteration can be resumed after
4115 * the lock is re-acquired. Iteration is performed at two levels -
4116 * css_sets and tasks in them.
4117 *
4118 * Once created, a css_set never leaves its cgroup lists, so a
4119 * pinned css_set is guaranteed to stay put and we can resume
4120 * iteration afterwards.
4121 *
4122 * Tasks may leave @cset across iteration steps. This is resolved
4123 * by registering each iterator with the css_set currently being
4124 * walked and making css_set_move_task() advance iterators whose
4125 * next task is leaving.
4126 */
4127 if (it->cur_cset) {
4128 list_del(&it->iters_node);
4129 put_css_set_locked(it->cur_cset);
4130 }
4131 get_css_set(cset);
4132 it->cur_cset = cset;
4133 list_add(&it->iters_node, &cset->task_iters);
4134}
4135
4136static void css_task_iter_advance(struct css_task_iter *it)
4137{
4138 struct list_head *next;
4139
4140 lockdep_assert_held(&css_set_lock);
4141repeat:
4142 /*
4143 * Advance iterator to find next entry. cset->tasks is consumed
4144 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4145 * next cset.
4146 */
4147 next = it->task_pos->next;
4148
4149 if (next == it->tasks_head)
4150 next = it->mg_tasks_head->next;
4151
4152 if (next == it->mg_tasks_head)
4153 css_task_iter_advance_css_set(it);
4154 else
4155 it->task_pos = next;
4156
4157 /* if PROCS, skip over tasks which aren't group leaders */
4158 if ((it->flags & CSS_TASK_ITER_PROCS) && it->task_pos &&
4159 !thread_group_leader(list_entry(it->task_pos, struct task_struct,
4160 cg_list)))
4161 goto repeat;
4162}
4163
4164/**
4165 * css_task_iter_start - initiate task iteration
4166 * @css: the css to walk tasks of
4167 * @flags: CSS_TASK_ITER_* flags
4168 * @it: the task iterator to use
4169 *
4170 * Initiate iteration through the tasks of @css. The caller can call
4171 * css_task_iter_next() to walk through the tasks until the function
4172 * returns NULL. On completion of iteration, css_task_iter_end() must be
4173 * called.
4174 */
4175void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4176 struct css_task_iter *it)
4177{
4178 /* no one should try to iterate before mounting cgroups */
4179 WARN_ON_ONCE(!use_task_css_set_links);
4180
4181 memset(it, 0, sizeof(*it));
4182
4183 spin_lock_irq(&css_set_lock);
4184
4185 it->ss = css->ss;
4186 it->flags = flags;
4187
4188 if (it->ss)
4189 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4190 else
4191 it->cset_pos = &css->cgroup->cset_links;
4192
4193 it->cset_head = it->cset_pos;
4194
4195 css_task_iter_advance_css_set(it);
4196
4197 spin_unlock_irq(&css_set_lock);
4198}
4199
4200/**
4201 * css_task_iter_next - return the next task for the iterator
4202 * @it: the task iterator being iterated
4203 *
4204 * The "next" function for task iteration. @it should have been
4205 * initialized via css_task_iter_start(). Returns NULL when the iteration
4206 * reaches the end.
4207 */
4208struct task_struct *css_task_iter_next(struct css_task_iter *it)
4209{
4210 if (it->cur_task) {
4211 put_task_struct(it->cur_task);
4212 it->cur_task = NULL;
4213 }
4214
4215 spin_lock_irq(&css_set_lock);
4216
4217 if (it->task_pos) {
4218 it->cur_task = list_entry(it->task_pos, struct task_struct,
4219 cg_list);
4220 get_task_struct(it->cur_task);
4221 css_task_iter_advance(it);
4222 }
4223
4224 spin_unlock_irq(&css_set_lock);
4225
4226 return it->cur_task;
4227}
4228
4229/**
4230 * css_task_iter_end - finish task iteration
4231 * @it: the task iterator to finish
4232 *
4233 * Finish task iteration started by css_task_iter_start().
4234 */
4235void css_task_iter_end(struct css_task_iter *it)
4236{
4237 if (it->cur_cset) {
4238 spin_lock_irq(&css_set_lock);
4239 list_del(&it->iters_node);
4240 put_css_set_locked(it->cur_cset);
4241 spin_unlock_irq(&css_set_lock);
4242 }
4243
4244 if (it->cur_dcset)
4245 put_css_set(it->cur_dcset);
4246
4247 if (it->cur_task)
4248 put_task_struct(it->cur_task);
4249}
4250
4251static void cgroup_procs_release(struct kernfs_open_file *of)
4252{
4253 if (of->priv) {
4254 css_task_iter_end(of->priv);
4255 kfree(of->priv);
4256 }
4257}
4258
4259static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4260{
4261 struct kernfs_open_file *of = s->private;
4262 struct css_task_iter *it = of->priv;
4263
4264 return css_task_iter_next(it);
4265}
4266
4267static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4268 unsigned int iter_flags)
4269{
4270 struct kernfs_open_file *of = s->private;
4271 struct cgroup *cgrp = seq_css(s)->cgroup;
4272 struct css_task_iter *it = of->priv;
4273
4274 /*
4275 * When a seq_file is seeked, it's always traversed sequentially
4276 * from position 0, so we can simply keep iterating on !0 *pos.
4277 */
4278 if (!it) {
4279 if (WARN_ON_ONCE((*pos)++))
4280 return ERR_PTR(-EINVAL);
4281
4282 it = kzalloc(sizeof(*it), GFP_KERNEL);
4283 if (!it)
4284 return ERR_PTR(-ENOMEM);
4285 of->priv = it;
4286 css_task_iter_start(&cgrp->self, iter_flags, it);
4287 } else if (!(*pos)++) {
4288 css_task_iter_end(it);
4289 css_task_iter_start(&cgrp->self, iter_flags, it);
4290 }
4291
4292 return cgroup_procs_next(s, NULL, NULL);
4293}
4294
4295static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4296{
4297 struct cgroup *cgrp = seq_css(s)->cgroup;
4298
4299 /*
4300 * All processes of a threaded subtree belong to the domain cgroup
4301 * of the subtree. Only threads can be distributed across the
4302 * subtree. Reject reads on cgroup.procs in the subtree proper.
4303 * They're always empty anyway.
4304 */
4305 if (cgroup_is_threaded(cgrp))
4306 return ERR_PTR(-EOPNOTSUPP);
4307
4308 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4309 CSS_TASK_ITER_THREADED);
4310}
4311
4312static int cgroup_procs_show(struct seq_file *s, void *v)
4313{
4314 seq_printf(s, "%d\n", task_pid_vnr(v));
4315 return 0;
4316}
4317
4318static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4319 struct cgroup *dst_cgrp,
4320 struct super_block *sb)
4321{
4322 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4323 struct cgroup *com_cgrp = src_cgrp;
4324 struct inode *inode;
4325 int ret;
4326
4327 lockdep_assert_held(&cgroup_mutex);
4328
4329 /* find the common ancestor */
4330 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4331 com_cgrp = cgroup_parent(com_cgrp);
4332
4333 /* %current should be authorized to migrate to the common ancestor */
4334 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4335 if (!inode)
4336 return -ENOMEM;
4337
4338 ret = inode_permission(inode, MAY_WRITE);
4339 iput(inode);
4340 if (ret)
4341 return ret;
4342
4343 /*
4344 * If namespaces are delegation boundaries, %current must be able
4345 * to see both source and destination cgroups from its namespace.
4346 */
4347 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4348 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4349 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4350 return -ENOENT;
4351
4352 return 0;
4353}
4354
4355static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4356 char *buf, size_t nbytes, loff_t off)
4357{
4358 struct cgroup *src_cgrp, *dst_cgrp;
4359 struct task_struct *task;
4360 ssize_t ret;
4361
4362 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4363 if (!dst_cgrp)
4364 return -ENODEV;
4365
4366 task = cgroup_procs_write_start(buf, true);
4367 ret = PTR_ERR_OR_ZERO(task);
4368 if (ret)
4369 goto out_unlock;
4370
4371 /* find the source cgroup */
4372 spin_lock_irq(&css_set_lock);
4373 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4374 spin_unlock_irq(&css_set_lock);
4375
4376 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4377 of->file->f_path.dentry->d_sb);
4378 if (ret)
4379 goto out_finish;
4380
4381 ret = cgroup_attach_task(dst_cgrp, task, true);
4382
4383out_finish:
4384 cgroup_procs_write_finish(task);
4385out_unlock:
4386 cgroup_kn_unlock(of->kn);
4387
4388 return ret ?: nbytes;
4389}
4390
4391static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4392{
4393 return __cgroup_procs_start(s, pos, 0);
4394}
4395
4396static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4397 char *buf, size_t nbytes, loff_t off)
4398{
4399 struct cgroup *src_cgrp, *dst_cgrp;
4400 struct task_struct *task;
4401 ssize_t ret;
4402
4403 buf = strstrip(buf);
4404
4405 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4406 if (!dst_cgrp)
4407 return -ENODEV;
4408
4409 task = cgroup_procs_write_start(buf, false);
4410 ret = PTR_ERR_OR_ZERO(task);
4411 if (ret)
4412 goto out_unlock;
4413
4414 /* find the source cgroup */
4415 spin_lock_irq(&css_set_lock);
4416 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4417 spin_unlock_irq(&css_set_lock);
4418
4419 /* thread migrations follow the cgroup.procs delegation rule */
4420 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4421 of->file->f_path.dentry->d_sb);
4422 if (ret)
4423 goto out_finish;
4424
4425 /* and must be contained in the same domain */
4426 ret = -EOPNOTSUPP;
4427 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4428 goto out_finish;
4429
4430 ret = cgroup_attach_task(dst_cgrp, task, false);
4431
4432out_finish:
4433 cgroup_procs_write_finish(task);
4434out_unlock:
4435 cgroup_kn_unlock(of->kn);
4436
4437 return ret ?: nbytes;
4438}
4439
4440/* cgroup core interface files for the default hierarchy */
4441static struct cftype cgroup_base_files[] = {
4442 {
4443 .name = "cgroup.type",
4444 .flags = CFTYPE_NOT_ON_ROOT,
4445 .seq_show = cgroup_type_show,
4446 .write = cgroup_type_write,
4447 },
4448 {
4449 .name = "cgroup.procs",
4450 .flags = CFTYPE_NS_DELEGATABLE,
4451 .file_offset = offsetof(struct cgroup, procs_file),
4452 .release = cgroup_procs_release,
4453 .seq_start = cgroup_procs_start,
4454 .seq_next = cgroup_procs_next,
4455 .seq_show = cgroup_procs_show,
4456 .write = cgroup_procs_write,
4457 },
4458 {
4459 .name = "cgroup.threads",
4460 .flags = CFTYPE_NS_DELEGATABLE,
4461 .release = cgroup_procs_release,
4462 .seq_start = cgroup_threads_start,
4463 .seq_next = cgroup_procs_next,
4464 .seq_show = cgroup_procs_show,
4465 .write = cgroup_threads_write,
4466 },
4467 {
4468 .name = "cgroup.controllers",
4469 .seq_show = cgroup_controllers_show,
4470 },
4471 {
4472 .name = "cgroup.subtree_control",
4473 .flags = CFTYPE_NS_DELEGATABLE,
4474 .seq_show = cgroup_subtree_control_show,
4475 .write = cgroup_subtree_control_write,
4476 },
4477 {
4478 .name = "cgroup.events",
4479 .flags = CFTYPE_NOT_ON_ROOT,
4480 .file_offset = offsetof(struct cgroup, events_file),
4481 .seq_show = cgroup_events_show,
4482 },
4483 {
4484 .name = "cgroup.max.descendants",
4485 .seq_show = cgroup_max_descendants_show,
4486 .write = cgroup_max_descendants_write,
4487 },
4488 {
4489 .name = "cgroup.max.depth",
4490 .seq_show = cgroup_max_depth_show,
4491 .write = cgroup_max_depth_write,
4492 },
4493 {
4494 .name = "cgroup.stat",
4495 .seq_show = cgroup_stat_show,
4496 },
4497 {
4498 .name = "cpu.stat",
4499 .flags = CFTYPE_NOT_ON_ROOT,
4500 .seq_show = cpu_stat_show,
4501 },
4502 { } /* terminate */
4503};
4504
4505/*
4506 * css destruction is four-stage process.
4507 *
4508 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4509 * Implemented in kill_css().
4510 *
4511 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4512 * and thus css_tryget_online() is guaranteed to fail, the css can be
4513 * offlined by invoking offline_css(). After offlining, the base ref is
4514 * put. Implemented in css_killed_work_fn().
4515 *
4516 * 3. When the percpu_ref reaches zero, the only possible remaining
4517 * accessors are inside RCU read sections. css_release() schedules the
4518 * RCU callback.
4519 *
4520 * 4. After the grace period, the css can be freed. Implemented in
4521 * css_free_work_fn().
4522 *
4523 * It is actually hairier because both step 2 and 4 require process context
4524 * and thus involve punting to css->destroy_work adding two additional
4525 * steps to the already complex sequence.
4526 */
4527static void css_free_rwork_fn(struct work_struct *work)
4528{
4529 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4530 struct cgroup_subsys_state, destroy_rwork);
4531 struct cgroup_subsys *ss = css->ss;
4532 struct cgroup *cgrp = css->cgroup;
4533
4534 percpu_ref_exit(&css->refcnt);
4535
4536 if (ss) {
4537 /* css free path */
4538 struct cgroup_subsys_state *parent = css->parent;
4539 int id = css->id;
4540
4541 ss->css_free(css);
4542 cgroup_idr_remove(&ss->css_idr, id);
4543 cgroup_put(cgrp);
4544
4545 if (parent)
4546 css_put(parent);
4547 } else {
4548 /* cgroup free path */
4549 atomic_dec(&cgrp->root->nr_cgrps);
4550 cgroup1_pidlist_destroy_all(cgrp);
4551 cancel_work_sync(&cgrp->release_agent_work);
4552
4553 if (cgroup_parent(cgrp)) {
4554 /*
4555 * We get a ref to the parent, and put the ref when
4556 * this cgroup is being freed, so it's guaranteed
4557 * that the parent won't be destroyed before its
4558 * children.
4559 */
4560 cgroup_put(cgroup_parent(cgrp));
4561 kernfs_put(cgrp->kn);
4562 if (cgroup_on_dfl(cgrp))
4563 cgroup_stat_exit(cgrp);
4564 kfree(cgrp);
4565 } else {
4566 /*
4567 * This is root cgroup's refcnt reaching zero,
4568 * which indicates that the root should be
4569 * released.
4570 */
4571 cgroup_destroy_root(cgrp->root);
4572 }
4573 }
4574}
4575
4576static void css_release_work_fn(struct work_struct *work)
4577{
4578 struct cgroup_subsys_state *css =
4579 container_of(work, struct cgroup_subsys_state, destroy_work);
4580 struct cgroup_subsys *ss = css->ss;
4581 struct cgroup *cgrp = css->cgroup;
4582
4583 mutex_lock(&cgroup_mutex);
4584
4585 css->flags |= CSS_RELEASED;
4586 list_del_rcu(&css->sibling);
4587
4588 if (ss) {
4589 /* css release path */
4590 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4591 if (ss->css_released)
4592 ss->css_released(css);
4593 } else {
4594 struct cgroup *tcgrp;
4595
4596 /* cgroup release path */
4597 trace_cgroup_release(cgrp);
4598
4599 if (cgroup_on_dfl(cgrp))
4600 cgroup_stat_flush(cgrp);
4601
4602 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4603 tcgrp = cgroup_parent(tcgrp))
4604 tcgrp->nr_dying_descendants--;
4605
4606 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4607 cgrp->id = -1;
4608
4609 /*
4610 * There are two control paths which try to determine
4611 * cgroup from dentry without going through kernfs -
4612 * cgroupstats_build() and css_tryget_online_from_dir().
4613 * Those are supported by RCU protecting clearing of
4614 * cgrp->kn->priv backpointer.
4615 */
4616 if (cgrp->kn)
4617 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4618 NULL);
4619
4620 cgroup_bpf_put(cgrp);
4621 }
4622
4623 mutex_unlock(&cgroup_mutex);
4624
4625 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4626 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
4627}
4628
4629static void css_release(struct percpu_ref *ref)
4630{
4631 struct cgroup_subsys_state *css =
4632 container_of(ref, struct cgroup_subsys_state, refcnt);
4633
4634 INIT_WORK(&css->destroy_work, css_release_work_fn);
4635 queue_work(cgroup_destroy_wq, &css->destroy_work);
4636}
4637
4638static void init_and_link_css(struct cgroup_subsys_state *css,
4639 struct cgroup_subsys *ss, struct cgroup *cgrp)
4640{
4641 lockdep_assert_held(&cgroup_mutex);
4642
4643 cgroup_get_live(cgrp);
4644
4645 memset(css, 0, sizeof(*css));
4646 css->cgroup = cgrp;
4647 css->ss = ss;
4648 css->id = -1;
4649 INIT_LIST_HEAD(&css->sibling);
4650 INIT_LIST_HEAD(&css->children);
4651 css->serial_nr = css_serial_nr_next++;
4652 atomic_set(&css->online_cnt, 0);
4653
4654 if (cgroup_parent(cgrp)) {
4655 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4656 css_get(css->parent);
4657 }
4658
4659 BUG_ON(cgroup_css(cgrp, ss));
4660}
4661
4662/* invoke ->css_online() on a new CSS and mark it online if successful */
4663static int online_css(struct cgroup_subsys_state *css)
4664{
4665 struct cgroup_subsys *ss = css->ss;
4666 int ret = 0;
4667
4668 lockdep_assert_held(&cgroup_mutex);
4669
4670 if (ss->css_online)
4671 ret = ss->css_online(css);
4672 if (!ret) {
4673 css->flags |= CSS_ONLINE;
4674 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4675
4676 atomic_inc(&css->online_cnt);
4677 if (css->parent)
4678 atomic_inc(&css->parent->online_cnt);
4679 }
4680 return ret;
4681}
4682
4683/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4684static void offline_css(struct cgroup_subsys_state *css)
4685{
4686 struct cgroup_subsys *ss = css->ss;
4687
4688 lockdep_assert_held(&cgroup_mutex);
4689
4690 if (!(css->flags & CSS_ONLINE))
4691 return;
4692
4693 if (ss->css_offline)
4694 ss->css_offline(css);
4695
4696 css->flags &= ~CSS_ONLINE;
4697 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4698
4699 wake_up_all(&css->cgroup->offline_waitq);
4700}
4701
4702/**
4703 * css_create - create a cgroup_subsys_state
4704 * @cgrp: the cgroup new css will be associated with
4705 * @ss: the subsys of new css
4706 *
4707 * Create a new css associated with @cgrp - @ss pair. On success, the new
4708 * css is online and installed in @cgrp. This function doesn't create the
4709 * interface files. Returns 0 on success, -errno on failure.
4710 */
4711static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4712 struct cgroup_subsys *ss)
4713{
4714 struct cgroup *parent = cgroup_parent(cgrp);
4715 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4716 struct cgroup_subsys_state *css;
4717 int err;
4718
4719 lockdep_assert_held(&cgroup_mutex);
4720
4721 css = ss->css_alloc(parent_css);
4722 if (!css)
4723 css = ERR_PTR(-ENOMEM);
4724 if (IS_ERR(css))
4725 return css;
4726
4727 init_and_link_css(css, ss, cgrp);
4728
4729 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4730 if (err)
4731 goto err_free_css;
4732
4733 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4734 if (err < 0)
4735 goto err_free_css;
4736 css->id = err;
4737
4738 /* @css is ready to be brought online now, make it visible */
4739 list_add_tail_rcu(&css->sibling, &parent_css->children);
4740 cgroup_idr_replace(&ss->css_idr, css, css->id);
4741
4742 err = online_css(css);
4743 if (err)
4744 goto err_list_del;
4745
4746 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4747 cgroup_parent(parent)) {
4748 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4749 current->comm, current->pid, ss->name);
4750 if (!strcmp(ss->name, "memory"))
4751 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4752 ss->warned_broken_hierarchy = true;
4753 }
4754
4755 return css;
4756
4757err_list_del:
4758 list_del_rcu(&css->sibling);
4759err_free_css:
4760 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4761 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
4762 return ERR_PTR(err);
4763}
4764
4765/*
4766 * The returned cgroup is fully initialized including its control mask, but
4767 * it isn't associated with its kernfs_node and doesn't have the control
4768 * mask applied.
4769 */
4770static struct cgroup *cgroup_create(struct cgroup *parent)
4771{
4772 struct cgroup_root *root = parent->root;
4773 struct cgroup *cgrp, *tcgrp;
4774 int level = parent->level + 1;
4775 int ret;
4776
4777 /* allocate the cgroup and its ID, 0 is reserved for the root */
4778 cgrp = kzalloc(sizeof(*cgrp) +
4779 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4780 if (!cgrp)
4781 return ERR_PTR(-ENOMEM);
4782
4783 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4784 if (ret)
4785 goto out_free_cgrp;
4786
4787 if (cgroup_on_dfl(parent)) {
4788 ret = cgroup_stat_init(cgrp);
4789 if (ret)
4790 goto out_cancel_ref;
4791 }
4792
4793 /*
4794 * Temporarily set the pointer to NULL, so idr_find() won't return
4795 * a half-baked cgroup.
4796 */
4797 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4798 if (cgrp->id < 0) {
4799 ret = -ENOMEM;
4800 goto out_stat_exit;
4801 }
4802
4803 init_cgroup_housekeeping(cgrp);
4804
4805 cgrp->self.parent = &parent->self;
4806 cgrp->root = root;
4807 cgrp->level = level;
4808 ret = cgroup_bpf_inherit(cgrp);
4809 if (ret)
4810 goto out_idr_free;
4811
4812 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
4813 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4814
4815 if (tcgrp != cgrp)
4816 tcgrp->nr_descendants++;
4817 }
4818
4819 if (notify_on_release(parent))
4820 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4821
4822 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4823 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4824
4825 cgrp->self.serial_nr = css_serial_nr_next++;
4826
4827 /* allocation complete, commit to creation */
4828 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4829 atomic_inc(&root->nr_cgrps);
4830 cgroup_get_live(parent);
4831
4832 /*
4833 * @cgrp is now fully operational. If something fails after this
4834 * point, it'll be released via the normal destruction path.
4835 */
4836 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4837
4838 /*
4839 * On the default hierarchy, a child doesn't automatically inherit
4840 * subtree_control from the parent. Each is configured manually.
4841 */
4842 if (!cgroup_on_dfl(cgrp))
4843 cgrp->subtree_control = cgroup_control(cgrp);
4844
4845 cgroup_propagate_control(cgrp);
4846
4847 return cgrp;
4848
4849out_idr_free:
4850 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4851out_stat_exit:
4852 if (cgroup_on_dfl(parent))
4853 cgroup_stat_exit(cgrp);
4854out_cancel_ref:
4855 percpu_ref_exit(&cgrp->self.refcnt);
4856out_free_cgrp:
4857 kfree(cgrp);
4858 return ERR_PTR(ret);
4859}
4860
4861static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
4862{
4863 struct cgroup *cgroup;
4864 int ret = false;
4865 int level = 1;
4866
4867 lockdep_assert_held(&cgroup_mutex);
4868
4869 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
4870 if (cgroup->nr_descendants >= cgroup->max_descendants)
4871 goto fail;
4872
4873 if (level > cgroup->max_depth)
4874 goto fail;
4875
4876 level++;
4877 }
4878
4879 ret = true;
4880fail:
4881 return ret;
4882}
4883
4884int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
4885{
4886 struct cgroup *parent, *cgrp;
4887 struct kernfs_node *kn;
4888 int ret;
4889
4890 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
4891 if (strchr(name, '\n'))
4892 return -EINVAL;
4893
4894 parent = cgroup_kn_lock_live(parent_kn, false);
4895 if (!parent)
4896 return -ENODEV;
4897
4898 if (!cgroup_check_hierarchy_limits(parent)) {
4899 ret = -EAGAIN;
4900 goto out_unlock;
4901 }
4902
4903 cgrp = cgroup_create(parent);
4904 if (IS_ERR(cgrp)) {
4905 ret = PTR_ERR(cgrp);
4906 goto out_unlock;
4907 }
4908
4909 /* create the directory */
4910 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4911 if (IS_ERR(kn)) {
4912 ret = PTR_ERR(kn);
4913 goto out_destroy;
4914 }
4915 cgrp->kn = kn;
4916
4917 /*
4918 * This extra ref will be put in cgroup_free_fn() and guarantees
4919 * that @cgrp->kn is always accessible.
4920 */
4921 kernfs_get(kn);
4922
4923 ret = cgroup_kn_set_ugid(kn);
4924 if (ret)
4925 goto out_destroy;
4926
4927 ret = css_populate_dir(&cgrp->self);
4928 if (ret)
4929 goto out_destroy;
4930
4931 ret = cgroup_apply_control_enable(cgrp);
4932 if (ret)
4933 goto out_destroy;
4934
4935 trace_cgroup_mkdir(cgrp);
4936
4937 /* let's create and online css's */
4938 kernfs_activate(kn);
4939
4940 ret = 0;
4941 goto out_unlock;
4942
4943out_destroy:
4944 cgroup_destroy_locked(cgrp);
4945out_unlock:
4946 cgroup_kn_unlock(parent_kn);
4947 return ret;
4948}
4949
4950/*
4951 * This is called when the refcnt of a css is confirmed to be killed.
4952 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4953 * initate destruction and put the css ref from kill_css().
4954 */
4955static void css_killed_work_fn(struct work_struct *work)
4956{
4957 struct cgroup_subsys_state *css =
4958 container_of(work, struct cgroup_subsys_state, destroy_work);
4959
4960 mutex_lock(&cgroup_mutex);
4961
4962 do {
4963 offline_css(css);
4964 css_put(css);
4965 /* @css can't go away while we're holding cgroup_mutex */
4966 css = css->parent;
4967 } while (css && atomic_dec_and_test(&css->online_cnt));
4968
4969 mutex_unlock(&cgroup_mutex);
4970}
4971
4972/* css kill confirmation processing requires process context, bounce */
4973static void css_killed_ref_fn(struct percpu_ref *ref)
4974{
4975 struct cgroup_subsys_state *css =
4976 container_of(ref, struct cgroup_subsys_state, refcnt);
4977
4978 if (atomic_dec_and_test(&css->online_cnt)) {
4979 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4980 queue_work(cgroup_destroy_wq, &css->destroy_work);
4981 }
4982}
4983
4984/**
4985 * kill_css - destroy a css
4986 * @css: css to destroy
4987 *
4988 * This function initiates destruction of @css by removing cgroup interface
4989 * files and putting its base reference. ->css_offline() will be invoked
4990 * asynchronously once css_tryget_online() is guaranteed to fail and when
4991 * the reference count reaches zero, @css will be released.
4992 */
4993static void kill_css(struct cgroup_subsys_state *css)
4994{
4995 lockdep_assert_held(&cgroup_mutex);
4996
4997 if (css->flags & CSS_DYING)
4998 return;
4999
5000 css->flags |= CSS_DYING;
5001
5002 /*
5003 * This must happen before css is disassociated with its cgroup.
5004 * See seq_css() for details.
5005 */
5006 css_clear_dir(css);
5007
5008 /*
5009 * Killing would put the base ref, but we need to keep it alive
5010 * until after ->css_offline().
5011 */
5012 css_get(css);
5013
5014 /*
5015 * cgroup core guarantees that, by the time ->css_offline() is
5016 * invoked, no new css reference will be given out via
5017 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5018 * proceed to offlining css's because percpu_ref_kill() doesn't
5019 * guarantee that the ref is seen as killed on all CPUs on return.
5020 *
5021 * Use percpu_ref_kill_and_confirm() to get notifications as each
5022 * css is confirmed to be seen as killed on all CPUs.
5023 */
5024 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5025}
5026
5027/**
5028 * cgroup_destroy_locked - the first stage of cgroup destruction
5029 * @cgrp: cgroup to be destroyed
5030 *
5031 * css's make use of percpu refcnts whose killing latency shouldn't be
5032 * exposed to userland and are RCU protected. Also, cgroup core needs to
5033 * guarantee that css_tryget_online() won't succeed by the time
5034 * ->css_offline() is invoked. To satisfy all the requirements,
5035 * destruction is implemented in the following two steps.
5036 *
5037 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5038 * userland visible parts and start killing the percpu refcnts of
5039 * css's. Set up so that the next stage will be kicked off once all
5040 * the percpu refcnts are confirmed to be killed.
5041 *
5042 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5043 * rest of destruction. Once all cgroup references are gone, the
5044 * cgroup is RCU-freed.
5045 *
5046 * This function implements s1. After this step, @cgrp is gone as far as
5047 * the userland is concerned and a new cgroup with the same name may be
5048 * created. As cgroup doesn't care about the names internally, this
5049 * doesn't cause any problem.
5050 */
5051static int cgroup_destroy_locked(struct cgroup *cgrp)
5052 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5053{
5054 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5055 struct cgroup_subsys_state *css;
5056 struct cgrp_cset_link *link;
5057 int ssid;
5058
5059 lockdep_assert_held(&cgroup_mutex);
5060
5061 /*
5062 * Only migration can raise populated from zero and we're already
5063 * holding cgroup_mutex.
5064 */
5065 if (cgroup_is_populated(cgrp))
5066 return -EBUSY;
5067
5068 /*
5069 * Make sure there's no live children. We can't test emptiness of
5070 * ->self.children as dead children linger on it while being
5071 * drained; otherwise, "rmdir parent/child parent" may fail.
5072 */
5073 if (css_has_online_children(&cgrp->self))
5074 return -EBUSY;
5075
5076 /*
5077 * Mark @cgrp and the associated csets dead. The former prevents
5078 * further task migration and child creation by disabling
5079 * cgroup_lock_live_group(). The latter makes the csets ignored by
5080 * the migration path.
5081 */
5082 cgrp->self.flags &= ~CSS_ONLINE;
5083
5084 spin_lock_irq(&css_set_lock);
5085 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5086 link->cset->dead = true;
5087 spin_unlock_irq(&css_set_lock);
5088
5089 /* initiate massacre of all css's */
5090 for_each_css(css, ssid, cgrp)
5091 kill_css(css);
5092
5093 /*
5094 * Remove @cgrp directory along with the base files. @cgrp has an
5095 * extra ref on its kn.
5096 */
5097 kernfs_remove(cgrp->kn);
5098
5099 if (parent && cgroup_is_threaded(cgrp))
5100 parent->nr_threaded_children--;
5101
5102 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5103 tcgrp->nr_descendants--;
5104 tcgrp->nr_dying_descendants++;
5105 }
5106
5107 cgroup1_check_for_release(parent);
5108
5109 /* put the base reference */
5110 percpu_ref_kill(&cgrp->self.refcnt);
5111
5112 return 0;
5113};
5114
5115int cgroup_rmdir(struct kernfs_node *kn)
5116{
5117 struct cgroup *cgrp;
5118 int ret = 0;
5119
5120 cgrp = cgroup_kn_lock_live(kn, false);
5121 if (!cgrp)
5122 return 0;
5123
5124 ret = cgroup_destroy_locked(cgrp);
5125
5126 if (!ret)
5127 trace_cgroup_rmdir(cgrp);
5128
5129 cgroup_kn_unlock(kn);
5130 return ret;
5131}
5132
5133static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5134 .show_options = cgroup_show_options,
5135 .remount_fs = cgroup_remount,
5136 .mkdir = cgroup_mkdir,
5137 .rmdir = cgroup_rmdir,
5138 .show_path = cgroup_show_path,
5139};
5140
5141static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5142{
5143 struct cgroup_subsys_state *css;
5144
5145 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5146
5147 mutex_lock(&cgroup_mutex);
5148
5149 idr_init(&ss->css_idr);
5150 INIT_LIST_HEAD(&ss->cfts);
5151
5152 /* Create the root cgroup state for this subsystem */
5153 ss->root = &cgrp_dfl_root;
5154 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5155 /* We don't handle early failures gracefully */
5156 BUG_ON(IS_ERR(css));
5157 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5158
5159 /*
5160 * Root csses are never destroyed and we can't initialize
5161 * percpu_ref during early init. Disable refcnting.
5162 */
5163 css->flags |= CSS_NO_REF;
5164
5165 if (early) {
5166 /* allocation can't be done safely during early init */
5167 css->id = 1;
5168 } else {
5169 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5170 BUG_ON(css->id < 0);
5171 }
5172
5173 /* Update the init_css_set to contain a subsys
5174 * pointer to this state - since the subsystem is
5175 * newly registered, all tasks and hence the
5176 * init_css_set is in the subsystem's root cgroup. */
5177 init_css_set.subsys[ss->id] = css;
5178
5179 have_fork_callback |= (bool)ss->fork << ss->id;
5180 have_exit_callback |= (bool)ss->exit << ss->id;
5181 have_free_callback |= (bool)ss->free << ss->id;
5182 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5183
5184 /* At system boot, before all subsystems have been
5185 * registered, no tasks have been forked, so we don't
5186 * need to invoke fork callbacks here. */
5187 BUG_ON(!list_empty(&init_task.tasks));
5188
5189 BUG_ON(online_css(css));
5190
5191 mutex_unlock(&cgroup_mutex);
5192}
5193
5194/**
5195 * cgroup_init_early - cgroup initialization at system boot
5196 *
5197 * Initialize cgroups at system boot, and initialize any
5198 * subsystems that request early init.
5199 */
5200int __init cgroup_init_early(void)
5201{
5202 static struct cgroup_sb_opts __initdata opts;
5203 struct cgroup_subsys *ss;
5204 int i;
5205
5206 init_cgroup_root(&cgrp_dfl_root, &opts);
5207 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5208
5209 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5210
5211 for_each_subsys(ss, i) {
5212 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5213 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5214 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5215 ss->id, ss->name);
5216 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5217 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5218
5219 ss->id = i;
5220 ss->name = cgroup_subsys_name[i];
5221 if (!ss->legacy_name)
5222 ss->legacy_name = cgroup_subsys_name[i];
5223
5224 if (ss->early_init)
5225 cgroup_init_subsys(ss, true);
5226 }
5227 return 0;
5228}
5229
5230static u16 cgroup_disable_mask __initdata;
5231
5232/**
5233 * cgroup_init - cgroup initialization
5234 *
5235 * Register cgroup filesystem and /proc file, and initialize
5236 * any subsystems that didn't request early init.
5237 */
5238int __init cgroup_init(void)
5239{
5240 struct cgroup_subsys *ss;
5241 int ssid;
5242
5243 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5244 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5245 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5246 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5247
5248 cgroup_stat_boot();
5249
5250 /*
5251 * The latency of the synchronize_sched() is too high for cgroups,
5252 * avoid it at the cost of forcing all readers into the slow path.
5253 */
5254 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5255
5256 get_user_ns(init_cgroup_ns.user_ns);
5257
5258 mutex_lock(&cgroup_mutex);
5259
5260 /*
5261 * Add init_css_set to the hash table so that dfl_root can link to
5262 * it during init.
5263 */
5264 hash_add(css_set_table, &init_css_set.hlist,
5265 css_set_hash(init_css_set.subsys));
5266
5267 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0, 0));
5268
5269 mutex_unlock(&cgroup_mutex);
5270
5271 for_each_subsys(ss, ssid) {
5272 if (ss->early_init) {
5273 struct cgroup_subsys_state *css =
5274 init_css_set.subsys[ss->id];
5275
5276 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5277 GFP_KERNEL);
5278 BUG_ON(css->id < 0);
5279 } else {
5280 cgroup_init_subsys(ss, false);
5281 }
5282
5283 list_add_tail(&init_css_set.e_cset_node[ssid],
5284 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5285
5286 /*
5287 * Setting dfl_root subsys_mask needs to consider the
5288 * disabled flag and cftype registration needs kmalloc,
5289 * both of which aren't available during early_init.
5290 */
5291 if (cgroup_disable_mask & (1 << ssid)) {
5292 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5293 printk(KERN_INFO "Disabling %s control group subsystem\n",
5294 ss->name);
5295 continue;
5296 }
5297
5298 if (cgroup1_ssid_disabled(ssid))
5299 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5300 ss->name);
5301
5302 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5303
5304 /* implicit controllers must be threaded too */
5305 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5306
5307 if (ss->implicit_on_dfl)
5308 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5309 else if (!ss->dfl_cftypes)
5310 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5311
5312 if (ss->threaded)
5313 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5314
5315 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5316 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5317 } else {
5318 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5319 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5320 }
5321
5322 if (ss->bind)
5323 ss->bind(init_css_set.subsys[ssid]);
5324
5325 mutex_lock(&cgroup_mutex);
5326 css_populate_dir(init_css_set.subsys[ssid]);
5327 mutex_unlock(&cgroup_mutex);
5328 }
5329
5330 /* init_css_set.subsys[] has been updated, re-hash */
5331 hash_del(&init_css_set.hlist);
5332 hash_add(css_set_table, &init_css_set.hlist,
5333 css_set_hash(init_css_set.subsys));
5334
5335 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5336 WARN_ON(register_filesystem(&cgroup_fs_type));
5337 WARN_ON(register_filesystem(&cgroup2_fs_type));
5338 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5339
5340 return 0;
5341}
5342
5343static int __init cgroup_wq_init(void)
5344{
5345 /*
5346 * There isn't much point in executing destruction path in
5347 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5348 * Use 1 for @max_active.
5349 *
5350 * We would prefer to do this in cgroup_init() above, but that
5351 * is called before init_workqueues(): so leave this until after.
5352 */
5353 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5354 BUG_ON(!cgroup_destroy_wq);
5355 return 0;
5356}
5357core_initcall(cgroup_wq_init);
5358
5359void cgroup_path_from_kernfs_id(const union kernfs_node_id *id,
5360 char *buf, size_t buflen)
5361{
5362 struct kernfs_node *kn;
5363
5364 kn = kernfs_get_node_by_id(cgrp_dfl_root.kf_root, id);
5365 if (!kn)
5366 return;
5367 kernfs_path(kn, buf, buflen);
5368 kernfs_put(kn);
5369}
5370
5371/*
5372 * proc_cgroup_show()
5373 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5374 * - Used for /proc/<pid>/cgroup.
5375 */
5376int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5377 struct pid *pid, struct task_struct *tsk)
5378{
5379 char *buf;
5380 int retval;
5381 struct cgroup_root *root;
5382
5383 retval = -ENOMEM;
5384 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5385 if (!buf)
5386 goto out;
5387
5388 mutex_lock(&cgroup_mutex);
5389 spin_lock_irq(&css_set_lock);
5390
5391 for_each_root(root) {
5392 struct cgroup_subsys *ss;
5393 struct cgroup *cgrp;
5394 int ssid, count = 0;
5395
5396 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5397 continue;
5398
5399 seq_printf(m, "%d:", root->hierarchy_id);
5400 if (root != &cgrp_dfl_root)
5401 for_each_subsys(ss, ssid)
5402 if (root->subsys_mask & (1 << ssid))
5403 seq_printf(m, "%s%s", count++ ? "," : "",
5404 ss->legacy_name);
5405 if (strlen(root->name))
5406 seq_printf(m, "%sname=%s", count ? "," : "",
5407 root->name);
5408 seq_putc(m, ':');
5409
5410 cgrp = task_cgroup_from_root(tsk, root);
5411
5412 /*
5413 * On traditional hierarchies, all zombie tasks show up as
5414 * belonging to the root cgroup. On the default hierarchy,
5415 * while a zombie doesn't show up in "cgroup.procs" and
5416 * thus can't be migrated, its /proc/PID/cgroup keeps
5417 * reporting the cgroup it belonged to before exiting. If
5418 * the cgroup is removed before the zombie is reaped,
5419 * " (deleted)" is appended to the cgroup path.
5420 */
5421 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5422 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5423 current->nsproxy->cgroup_ns);
5424 if (retval >= PATH_MAX)
5425 retval = -ENAMETOOLONG;
5426 if (retval < 0)
5427 goto out_unlock;
5428
5429 seq_puts(m, buf);
5430 } else {
5431 seq_puts(m, "/");
5432 }
5433
5434 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5435 seq_puts(m, " (deleted)\n");
5436 else
5437 seq_putc(m, '\n');
5438 }
5439
5440 retval = 0;
5441out_unlock:
5442 spin_unlock_irq(&css_set_lock);
5443 mutex_unlock(&cgroup_mutex);
5444 kfree(buf);
5445out:
5446 return retval;
5447}
5448
5449/**
5450 * cgroup_fork - initialize cgroup related fields during copy_process()
5451 * @child: pointer to task_struct of forking parent process.
5452 *
5453 * A task is associated with the init_css_set until cgroup_post_fork()
5454 * attaches it to the parent's css_set. Empty cg_list indicates that
5455 * @child isn't holding reference to its css_set.
5456 */
5457void cgroup_fork(struct task_struct *child)
5458{
5459 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5460 INIT_LIST_HEAD(&child->cg_list);
5461}
5462
5463/**
5464 * cgroup_can_fork - called on a new task before the process is exposed
5465 * @child: the task in question.
5466 *
5467 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5468 * returns an error, the fork aborts with that error code. This allows for
5469 * a cgroup subsystem to conditionally allow or deny new forks.
5470 */
5471int cgroup_can_fork(struct task_struct *child)
5472{
5473 struct cgroup_subsys *ss;
5474 int i, j, ret;
5475
5476 do_each_subsys_mask(ss, i, have_canfork_callback) {
5477 ret = ss->can_fork(child);
5478 if (ret)
5479 goto out_revert;
5480 } while_each_subsys_mask();
5481
5482 return 0;
5483
5484out_revert:
5485 for_each_subsys(ss, j) {
5486 if (j >= i)
5487 break;
5488 if (ss->cancel_fork)
5489 ss->cancel_fork(child);
5490 }
5491
5492 return ret;
5493}
5494
5495/**
5496 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5497 * @child: the task in question
5498 *
5499 * This calls the cancel_fork() callbacks if a fork failed *after*
5500 * cgroup_can_fork() succeded.
5501 */
5502void cgroup_cancel_fork(struct task_struct *child)
5503{
5504 struct cgroup_subsys *ss;
5505 int i;
5506
5507 for_each_subsys(ss, i)
5508 if (ss->cancel_fork)
5509 ss->cancel_fork(child);
5510}
5511
5512/**
5513 * cgroup_post_fork - called on a new task after adding it to the task list
5514 * @child: the task in question
5515 *
5516 * Adds the task to the list running through its css_set if necessary and
5517 * call the subsystem fork() callbacks. Has to be after the task is
5518 * visible on the task list in case we race with the first call to
5519 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5520 * list.
5521 */
5522void cgroup_post_fork(struct task_struct *child)
5523{
5524 struct cgroup_subsys *ss;
5525 int i;
5526
5527 /*
5528 * This may race against cgroup_enable_task_cg_lists(). As that
5529 * function sets use_task_css_set_links before grabbing
5530 * tasklist_lock and we just went through tasklist_lock to add
5531 * @child, it's guaranteed that either we see the set
5532 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5533 * @child during its iteration.
5534 *
5535 * If we won the race, @child is associated with %current's
5536 * css_set. Grabbing css_set_lock guarantees both that the
5537 * association is stable, and, on completion of the parent's
5538 * migration, @child is visible in the source of migration or
5539 * already in the destination cgroup. This guarantee is necessary
5540 * when implementing operations which need to migrate all tasks of
5541 * a cgroup to another.
5542 *
5543 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5544 * will remain in init_css_set. This is safe because all tasks are
5545 * in the init_css_set before cg_links is enabled and there's no
5546 * operation which transfers all tasks out of init_css_set.
5547 */
5548 if (use_task_css_set_links) {
5549 struct css_set *cset;
5550
5551 spin_lock_irq(&css_set_lock);
5552 cset = task_css_set(current);
5553 if (list_empty(&child->cg_list)) {
5554 get_css_set(cset);
5555 cset->nr_tasks++;
5556 css_set_move_task(child, NULL, cset, false);
5557 }
5558 spin_unlock_irq(&css_set_lock);
5559 }
5560
5561 /*
5562 * Call ss->fork(). This must happen after @child is linked on
5563 * css_set; otherwise, @child might change state between ->fork()
5564 * and addition to css_set.
5565 */
5566 do_each_subsys_mask(ss, i, have_fork_callback) {
5567 ss->fork(child);
5568 } while_each_subsys_mask();
5569}
5570
5571/**
5572 * cgroup_exit - detach cgroup from exiting task
5573 * @tsk: pointer to task_struct of exiting process
5574 *
5575 * Description: Detach cgroup from @tsk and release it.
5576 *
5577 * Note that cgroups marked notify_on_release force every task in
5578 * them to take the global cgroup_mutex mutex when exiting.
5579 * This could impact scaling on very large systems. Be reluctant to
5580 * use notify_on_release cgroups where very high task exit scaling
5581 * is required on large systems.
5582 *
5583 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5584 * call cgroup_exit() while the task is still competent to handle
5585 * notify_on_release(), then leave the task attached to the root cgroup in
5586 * each hierarchy for the remainder of its exit. No need to bother with
5587 * init_css_set refcnting. init_css_set never goes away and we can't race
5588 * with migration path - PF_EXITING is visible to migration path.
5589 */
5590void cgroup_exit(struct task_struct *tsk)
5591{
5592 struct cgroup_subsys *ss;
5593 struct css_set *cset;
5594 int i;
5595
5596 /*
5597 * Unlink from @tsk from its css_set. As migration path can't race
5598 * with us, we can check css_set and cg_list without synchronization.
5599 */
5600 cset = task_css_set(tsk);
5601
5602 if (!list_empty(&tsk->cg_list)) {
5603 spin_lock_irq(&css_set_lock);
5604 css_set_move_task(tsk, cset, NULL, false);
5605 cset->nr_tasks--;
5606 spin_unlock_irq(&css_set_lock);
5607 } else {
5608 get_css_set(cset);
5609 }
5610
5611 /* see cgroup_post_fork() for details */
5612 do_each_subsys_mask(ss, i, have_exit_callback) {
5613 ss->exit(tsk);
5614 } while_each_subsys_mask();
5615}
5616
5617void cgroup_free(struct task_struct *task)
5618{
5619 struct css_set *cset = task_css_set(task);
5620 struct cgroup_subsys *ss;
5621 int ssid;
5622
5623 do_each_subsys_mask(ss, ssid, have_free_callback) {
5624 ss->free(task);
5625 } while_each_subsys_mask();
5626
5627 put_css_set(cset);
5628}
5629
5630static int __init cgroup_disable(char *str)
5631{
5632 struct cgroup_subsys *ss;
5633 char *token;
5634 int i;
5635
5636 while ((token = strsep(&str, ",")) != NULL) {
5637 if (!*token)
5638 continue;
5639
5640 for_each_subsys(ss, i) {
5641 if (strcmp(token, ss->name) &&
5642 strcmp(token, ss->legacy_name))
5643 continue;
5644 cgroup_disable_mask |= 1 << i;
5645 }
5646 }
5647 return 1;
5648}
5649__setup("cgroup_disable=", cgroup_disable);
5650
5651/**
5652 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5653 * @dentry: directory dentry of interest
5654 * @ss: subsystem of interest
5655 *
5656 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5657 * to get the corresponding css and return it. If such css doesn't exist
5658 * or can't be pinned, an ERR_PTR value is returned.
5659 */
5660struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5661 struct cgroup_subsys *ss)
5662{
5663 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5664 struct file_system_type *s_type = dentry->d_sb->s_type;
5665 struct cgroup_subsys_state *css = NULL;
5666 struct cgroup *cgrp;
5667
5668 /* is @dentry a cgroup dir? */
5669 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5670 !kn || kernfs_type(kn) != KERNFS_DIR)
5671 return ERR_PTR(-EBADF);
5672
5673 rcu_read_lock();
5674
5675 /*
5676 * This path doesn't originate from kernfs and @kn could already
5677 * have been or be removed at any point. @kn->priv is RCU
5678 * protected for this access. See css_release_work_fn() for details.
5679 */
5680 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
5681 if (cgrp)
5682 css = cgroup_css(cgrp, ss);
5683
5684 if (!css || !css_tryget_online(css))
5685 css = ERR_PTR(-ENOENT);
5686
5687 rcu_read_unlock();
5688 return css;
5689}
5690
5691/**
5692 * css_from_id - lookup css by id
5693 * @id: the cgroup id
5694 * @ss: cgroup subsys to be looked into
5695 *
5696 * Returns the css if there's valid one with @id, otherwise returns NULL.
5697 * Should be called under rcu_read_lock().
5698 */
5699struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5700{
5701 WARN_ON_ONCE(!rcu_read_lock_held());
5702 return idr_find(&ss->css_idr, id);
5703}
5704
5705/**
5706 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5707 * @path: path on the default hierarchy
5708 *
5709 * Find the cgroup at @path on the default hierarchy, increment its
5710 * reference count and return it. Returns pointer to the found cgroup on
5711 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5712 * if @path points to a non-directory.
5713 */
5714struct cgroup *cgroup_get_from_path(const char *path)
5715{
5716 struct kernfs_node *kn;
5717 struct cgroup *cgrp;
5718
5719 mutex_lock(&cgroup_mutex);
5720
5721 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5722 if (kn) {
5723 if (kernfs_type(kn) == KERNFS_DIR) {
5724 cgrp = kn->priv;
5725 cgroup_get_live(cgrp);
5726 } else {
5727 cgrp = ERR_PTR(-ENOTDIR);
5728 }
5729 kernfs_put(kn);
5730 } else {
5731 cgrp = ERR_PTR(-ENOENT);
5732 }
5733
5734 mutex_unlock(&cgroup_mutex);
5735 return cgrp;
5736}
5737EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5738
5739/**
5740 * cgroup_get_from_fd - get a cgroup pointer from a fd
5741 * @fd: fd obtained by open(cgroup2_dir)
5742 *
5743 * Find the cgroup from a fd which should be obtained
5744 * by opening a cgroup directory. Returns a pointer to the
5745 * cgroup on success. ERR_PTR is returned if the cgroup
5746 * cannot be found.
5747 */
5748struct cgroup *cgroup_get_from_fd(int fd)
5749{
5750 struct cgroup_subsys_state *css;
5751 struct cgroup *cgrp;
5752 struct file *f;
5753
5754 f = fget_raw(fd);
5755 if (!f)
5756 return ERR_PTR(-EBADF);
5757
5758 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5759 fput(f);
5760 if (IS_ERR(css))
5761 return ERR_CAST(css);
5762
5763 cgrp = css->cgroup;
5764 if (!cgroup_on_dfl(cgrp)) {
5765 cgroup_put(cgrp);
5766 return ERR_PTR(-EBADF);
5767 }
5768
5769 return cgrp;
5770}
5771EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
5772
5773/*
5774 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5775 * definition in cgroup-defs.h.
5776 */
5777#ifdef CONFIG_SOCK_CGROUP_DATA
5778
5779#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5780
5781DEFINE_SPINLOCK(cgroup_sk_update_lock);
5782static bool cgroup_sk_alloc_disabled __read_mostly;
5783
5784void cgroup_sk_alloc_disable(void)
5785{
5786 if (cgroup_sk_alloc_disabled)
5787 return;
5788 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5789 cgroup_sk_alloc_disabled = true;
5790}
5791
5792#else
5793
5794#define cgroup_sk_alloc_disabled false
5795
5796#endif
5797
5798void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5799{
5800 if (cgroup_sk_alloc_disabled)
5801 return;
5802
5803 /* Socket clone path */
5804 if (skcd->val) {
5805 /*
5806 * We might be cloning a socket which is left in an empty
5807 * cgroup and the cgroup might have already been rmdir'd.
5808 * Don't use cgroup_get_live().
5809 */
5810 cgroup_get(sock_cgroup_ptr(skcd));
5811 return;
5812 }
5813
5814 rcu_read_lock();
5815
5816 while (true) {
5817 struct css_set *cset;
5818
5819 cset = task_css_set(current);
5820 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5821 skcd->val = (unsigned long)cset->dfl_cgrp;
5822 break;
5823 }
5824 cpu_relax();
5825 }
5826
5827 rcu_read_unlock();
5828}
5829
5830void cgroup_sk_free(struct sock_cgroup_data *skcd)
5831{
5832 cgroup_put(sock_cgroup_ptr(skcd));
5833}
5834
5835#endif /* CONFIG_SOCK_CGROUP_DATA */
5836
5837#ifdef CONFIG_CGROUP_BPF
5838int cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
5839 enum bpf_attach_type type, u32 flags)
5840{
5841 int ret;
5842
5843 mutex_lock(&cgroup_mutex);
5844 ret = __cgroup_bpf_attach(cgrp, prog, type, flags);
5845 mutex_unlock(&cgroup_mutex);
5846 return ret;
5847}
5848int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
5849 enum bpf_attach_type type, u32 flags)
5850{
5851 int ret;
5852
5853 mutex_lock(&cgroup_mutex);
5854 ret = __cgroup_bpf_detach(cgrp, prog, type, flags);
5855 mutex_unlock(&cgroup_mutex);
5856 return ret;
5857}
5858int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
5859 union bpf_attr __user *uattr)
5860{
5861 int ret;
5862
5863 mutex_lock(&cgroup_mutex);
5864 ret = __cgroup_bpf_query(cgrp, attr, uattr);
5865 mutex_unlock(&cgroup_mutex);
5866 return ret;
5867}
5868#endif /* CONFIG_CGROUP_BPF */
5869
5870#ifdef CONFIG_SYSFS
5871static ssize_t show_delegatable_files(struct cftype *files, char *buf,
5872 ssize_t size, const char *prefix)
5873{
5874 struct cftype *cft;
5875 ssize_t ret = 0;
5876
5877 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
5878 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
5879 continue;
5880
5881 if (prefix)
5882 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
5883
5884 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
5885
5886 if (unlikely(ret >= size)) {
5887 WARN_ON(1);
5888 break;
5889 }
5890 }
5891
5892 return ret;
5893}
5894
5895static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
5896 char *buf)
5897{
5898 struct cgroup_subsys *ss;
5899 int ssid;
5900 ssize_t ret = 0;
5901
5902 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
5903 NULL);
5904
5905 for_each_subsys(ss, ssid)
5906 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
5907 PAGE_SIZE - ret,
5908 cgroup_subsys_name[ssid]);
5909
5910 return ret;
5911}
5912static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
5913
5914static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
5915 char *buf)
5916{
5917 return snprintf(buf, PAGE_SIZE, "nsdelegate\n");
5918}
5919static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
5920
5921static struct attribute *cgroup_sysfs_attrs[] = {
5922 &cgroup_delegate_attr.attr,
5923 &cgroup_features_attr.attr,
5924 NULL,
5925};
5926
5927static const struct attribute_group cgroup_sysfs_attr_group = {
5928 .attrs = cgroup_sysfs_attrs,
5929 .name = "cgroup",
5930};
5931
5932static int __init cgroup_sysfs_init(void)
5933{
5934 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
5935}
5936subsys_initcall(cgroup_sysfs_init);
5937#endif /* CONFIG_SYSFS */