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/bpf-cgroup.h>
34#include <linux/cred.h>
35#include <linux/errno.h>
36#include <linux/init_task.h>
37#include <linux/kernel.h>
38#include <linux/magic.h>
39#include <linux/mutex.h>
40#include <linux/mount.h>
41#include <linux/pagemap.h>
42#include <linux/proc_fs.h>
43#include <linux/rcupdate.h>
44#include <linux/sched.h>
45#include <linux/sched/task.h>
46#include <linux/slab.h>
47#include <linux/spinlock.h>
48#include <linux/percpu-rwsem.h>
49#include <linux/string.h>
50#include <linux/hashtable.h>
51#include <linux/idr.h>
52#include <linux/kthread.h>
53#include <linux/atomic.h>
54#include <linux/cpuset.h>
55#include <linux/proc_ns.h>
56#include <linux/nsproxy.h>
57#include <linux/file.h>
58#include <linux/fs_parser.h>
59#include <linux/sched/cputime.h>
60#include <linux/sched/deadline.h>
61#include <linux/psi.h>
62#include <linux/nstree.h>
63#include <net/sock.h>
64
65#define CREATE_TRACE_POINTS
66#include <trace/events/cgroup.h>
67
68#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
69 MAX_CFTYPE_NAME + 2)
70/* let's not notify more than 100 times per second */
71#define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
72
73/*
74 * To avoid confusing the compiler (and generating warnings) with code
75 * that attempts to access what would be a 0-element array (i.e. sized
76 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this
77 * constant expression can be added.
78 */
79#define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0)
80
81/*
82 * cgroup_mutex is the master lock. Any modification to cgroup or its
83 * hierarchy must be performed while holding it.
84 *
85 * css_set_lock protects task->cgroups pointer, the list of css_set
86 * objects, and the chain of tasks off each css_set.
87 *
88 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
89 * cgroup.h can use them for lockdep annotations.
90 */
91DEFINE_MUTEX(cgroup_mutex);
92DEFINE_SPINLOCK(css_set_lock);
93
94#if (defined CONFIG_PROVE_RCU || defined CONFIG_LOCKDEP)
95EXPORT_SYMBOL_GPL(cgroup_mutex);
96EXPORT_SYMBOL_GPL(css_set_lock);
97#endif
98
99struct blocking_notifier_head cgroup_lifetime_notifier =
100 BLOCKING_NOTIFIER_INIT(cgroup_lifetime_notifier);
101
102DEFINE_SPINLOCK(trace_cgroup_path_lock);
103char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
104static bool cgroup_debug __read_mostly;
105
106/*
107 * Protects cgroup_idr and css_idr so that IDs can be released without
108 * grabbing cgroup_mutex.
109 */
110static DEFINE_SPINLOCK(cgroup_idr_lock);
111
112/*
113 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
114 * against file removal/re-creation across css hiding.
115 */
116static DEFINE_SPINLOCK(cgroup_file_kn_lock);
117
118DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
119
120#define cgroup_assert_mutex_or_rcu_locked() \
121 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
122 !lockdep_is_held(&cgroup_mutex), \
123 "cgroup_mutex or RCU read lock required");
124
125/*
126 * cgroup destruction makes heavy use of work items and there can be a lot
127 * of concurrent destructions. Use a separate workqueue so that cgroup
128 * destruction work items don't end up filling up max_active of system_percpu_wq
129 * which may lead to deadlock.
130 *
131 * A cgroup destruction should enqueue work sequentially to:
132 * cgroup_offline_wq: use for css offline work
133 * cgroup_release_wq: use for css release work
134 * cgroup_free_wq: use for free work
135 *
136 * Rationale for using separate workqueues:
137 * The cgroup root free work may depend on completion of other css offline
138 * operations. If all tasks were enqueued to a single workqueue, this could
139 * create a deadlock scenario where:
140 * - Free work waits for other css offline work to complete.
141 * - But other css offline work is queued after free work in the same queue.
142 *
143 * Example deadlock scenario with single workqueue (cgroup_destroy_wq):
144 * 1. umount net_prio
145 * 2. net_prio root destruction enqueues work to cgroup_destroy_wq (CPUx)
146 * 3. perf_event CSS A offline enqueues work to same cgroup_destroy_wq (CPUx)
147 * 4. net_prio cgroup_destroy_root->cgroup_lock_and_drain_offline.
148 * 5. net_prio root destruction blocks waiting for perf_event CSS A offline,
149 * which can never complete as it's behind in the same queue and
150 * workqueue's max_active is 1.
151 */
152static struct workqueue_struct *cgroup_offline_wq;
153static struct workqueue_struct *cgroup_release_wq;
154static struct workqueue_struct *cgroup_free_wq;
155
156/* generate an array of cgroup subsystem pointers */
157#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
158struct cgroup_subsys *cgroup_subsys[] = {
159#include <linux/cgroup_subsys.h>
160};
161#undef SUBSYS
162
163/* array of cgroup subsystem names */
164#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
165static const char *cgroup_subsys_name[] = {
166#include <linux/cgroup_subsys.h>
167};
168#undef SUBSYS
169
170/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
171#define SUBSYS(_x) \
172 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
173 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
174 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
175 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
176#include <linux/cgroup_subsys.h>
177#undef SUBSYS
178
179#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
180static struct static_key_true *cgroup_subsys_enabled_key[] = {
181#include <linux/cgroup_subsys.h>
182};
183#undef SUBSYS
184
185#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
186static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
187#include <linux/cgroup_subsys.h>
188};
189#undef SUBSYS
190
191static DEFINE_PER_CPU(struct css_rstat_cpu, root_rstat_cpu);
192static DEFINE_PER_CPU(struct cgroup_rstat_base_cpu, root_rstat_base_cpu);
193
194/* the default hierarchy */
195struct cgroup_root cgrp_dfl_root = {
196 .cgrp.self.rstat_cpu = &root_rstat_cpu,
197 .cgrp.rstat_base_cpu = &root_rstat_base_cpu,
198};
199EXPORT_SYMBOL_GPL(cgrp_dfl_root);
200
201/*
202 * The default hierarchy always exists but is hidden until mounted for the
203 * first time. This is for backward compatibility.
204 */
205bool cgrp_dfl_visible;
206
207/* some controllers are not supported in the default hierarchy */
208static u16 cgrp_dfl_inhibit_ss_mask;
209
210/* some controllers are implicitly enabled on the default hierarchy */
211static u16 cgrp_dfl_implicit_ss_mask;
212
213/* some controllers can be threaded on the default hierarchy */
214static u16 cgrp_dfl_threaded_ss_mask;
215
216/* The list of hierarchy roots */
217LIST_HEAD(cgroup_roots);
218static int cgroup_root_count;
219
220/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
221static DEFINE_IDR(cgroup_hierarchy_idr);
222
223/*
224 * Assign a monotonically increasing serial number to csses. It guarantees
225 * cgroups with bigger numbers are newer than those with smaller numbers.
226 * Also, as csses are always appended to the parent's ->children list, it
227 * guarantees that sibling csses are always sorted in the ascending serial
228 * number order on the list. Protected by cgroup_mutex.
229 */
230static u64 css_serial_nr_next = 1;
231
232/*
233 * These bitmasks identify subsystems with specific features to avoid
234 * having to do iterative checks repeatedly.
235 */
236static u16 have_fork_callback __read_mostly;
237static u16 have_exit_callback __read_mostly;
238static u16 have_release_callback __read_mostly;
239static u16 have_canfork_callback __read_mostly;
240
241static bool have_favordynmods __ro_after_init = IS_ENABLED(CONFIG_CGROUP_FAVOR_DYNMODS);
242
243/*
244 * Write protected by cgroup_mutex and write-lock of cgroup_threadgroup_rwsem,
245 * read protected by either.
246 *
247 * Can only be turned on, but not turned off.
248 */
249bool cgroup_enable_per_threadgroup_rwsem __read_mostly;
250
251/* cgroup namespace for init task */
252struct cgroup_namespace init_cgroup_ns = {
253 .ns.__ns_ref = REFCOUNT_INIT(2),
254 .user_ns = &init_user_ns,
255 .ns.ops = &cgroupns_operations,
256 .ns.inum = ns_init_inum(&init_cgroup_ns),
257 .root_cset = &init_css_set,
258 .ns.ns_type = ns_common_type(&init_cgroup_ns),
259};
260
261static struct file_system_type cgroup2_fs_type;
262static struct cftype cgroup_base_files[];
263static struct cftype cgroup_psi_files[];
264
265/* cgroup optional features */
266enum cgroup_opt_features {
267#ifdef CONFIG_PSI
268 OPT_FEATURE_PRESSURE,
269#endif
270 OPT_FEATURE_COUNT
271};
272
273static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
274#ifdef CONFIG_PSI
275 "pressure",
276#endif
277};
278
279static u16 cgroup_feature_disable_mask __read_mostly;
280
281static int cgroup_apply_control(struct cgroup *cgrp);
282static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
283static void css_task_iter_skip(struct css_task_iter *it,
284 struct task_struct *task);
285static int cgroup_destroy_locked(struct cgroup *cgrp);
286static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
287 struct cgroup_subsys *ss);
288static void css_release(struct percpu_ref *ref);
289static void kill_css(struct cgroup_subsys_state *css);
290static int cgroup_addrm_files(struct cgroup_subsys_state *css,
291 struct cgroup *cgrp, struct cftype cfts[],
292 bool is_add);
293
294#ifdef CONFIG_DEBUG_CGROUP_REF
295#define CGROUP_REF_FN_ATTRS noinline
296#define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn);
297#include <linux/cgroup_refcnt.h>
298#endif
299
300/**
301 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
302 * @ssid: subsys ID of interest
303 *
304 * cgroup_subsys_enabled() can only be used with literal subsys names which
305 * is fine for individual subsystems but unsuitable for cgroup core. This
306 * is slower static_key_enabled() based test indexed by @ssid.
307 */
308bool cgroup_ssid_enabled(int ssid)
309{
310 if (!CGROUP_HAS_SUBSYS_CONFIG)
311 return false;
312
313 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
314}
315
316/**
317 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
318 * @cgrp: the cgroup of interest
319 *
320 * The default hierarchy is the v2 interface of cgroup and this function
321 * can be used to test whether a cgroup is on the default hierarchy for
322 * cases where a subsystem should behave differently depending on the
323 * interface version.
324 *
325 * List of changed behaviors:
326 *
327 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
328 * and "name" are disallowed.
329 *
330 * - When mounting an existing superblock, mount options should match.
331 *
332 * - rename(2) is disallowed.
333 *
334 * - "tasks" is removed. Everything should be at process granularity. Use
335 * "cgroup.procs" instead.
336 *
337 * - "cgroup.procs" is not sorted. pids will be unique unless they got
338 * recycled in-between reads.
339 *
340 * - "release_agent" and "notify_on_release" are removed. Replacement
341 * notification mechanism will be implemented.
342 *
343 * - "cgroup.clone_children" is removed.
344 *
345 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
346 * and its descendants contain no task; otherwise, 1. The file also
347 * generates kernfs notification which can be monitored through poll and
348 * [di]notify when the value of the file changes.
349 *
350 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
351 * take masks of ancestors with non-empty cpus/mems, instead of being
352 * moved to an ancestor.
353 *
354 * - cpuset: a task can be moved into an empty cpuset, and again it takes
355 * masks of ancestors.
356 *
357 * - blkcg: blk-throttle becomes properly hierarchical.
358 */
359bool cgroup_on_dfl(const struct cgroup *cgrp)
360{
361 return cgrp->root == &cgrp_dfl_root;
362}
363
364/* IDR wrappers which synchronize using cgroup_idr_lock */
365static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
366 gfp_t gfp_mask)
367{
368 int ret;
369
370 idr_preload(gfp_mask);
371 spin_lock_bh(lock: &cgroup_idr_lock);
372 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
373 spin_unlock_bh(lock: &cgroup_idr_lock);
374 idr_preload_end();
375 return ret;
376}
377
378static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
379{
380 void *ret;
381
382 spin_lock_bh(lock: &cgroup_idr_lock);
383 ret = idr_replace(idr, ptr, id);
384 spin_unlock_bh(lock: &cgroup_idr_lock);
385 return ret;
386}
387
388static void cgroup_idr_remove(struct idr *idr, int id)
389{
390 spin_lock_bh(lock: &cgroup_idr_lock);
391 idr_remove(idr, id);
392 spin_unlock_bh(lock: &cgroup_idr_lock);
393}
394
395static bool cgroup_has_tasks(struct cgroup *cgrp)
396{
397 return cgrp->nr_populated_csets;
398}
399
400static bool cgroup_is_threaded(struct cgroup *cgrp)
401{
402 return cgrp->dom_cgrp != cgrp;
403}
404
405/* can @cgrp host both domain and threaded children? */
406static bool cgroup_is_mixable(struct cgroup *cgrp)
407{
408 /*
409 * Root isn't under domain level resource control exempting it from
410 * the no-internal-process constraint, so it can serve as a thread
411 * root and a parent of resource domains at the same time.
412 */
413 return !cgroup_parent(cgrp);
414}
415
416/* can @cgrp become a thread root? Should always be true for a thread root */
417static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
418{
419 /* mixables don't care */
420 if (cgroup_is_mixable(cgrp))
421 return true;
422
423 /* domain roots can't be nested under threaded */
424 if (cgroup_is_threaded(cgrp))
425 return false;
426
427 /* can only have either domain or threaded children */
428 if (cgrp->nr_populated_domain_children)
429 return false;
430
431 /* and no domain controllers can be enabled */
432 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
433 return false;
434
435 return true;
436}
437
438/* is @cgrp root of a threaded subtree? */
439static bool cgroup_is_thread_root(struct cgroup *cgrp)
440{
441 /* thread root should be a domain */
442 if (cgroup_is_threaded(cgrp))
443 return false;
444
445 /* a domain w/ threaded children is a thread root */
446 if (cgrp->nr_threaded_children)
447 return true;
448
449 /*
450 * A domain which has tasks and explicit threaded controllers
451 * enabled is a thread root.
452 */
453 if (cgroup_has_tasks(cgrp) &&
454 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
455 return true;
456
457 return false;
458}
459
460/* a domain which isn't connected to the root w/o brekage can't be used */
461static bool cgroup_is_valid_domain(struct cgroup *cgrp)
462{
463 /* the cgroup itself can be a thread root */
464 if (cgroup_is_threaded(cgrp))
465 return false;
466
467 /* but the ancestors can't be unless mixable */
468 while ((cgrp = cgroup_parent(cgrp))) {
469 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
470 return false;
471 if (cgroup_is_threaded(cgrp))
472 return false;
473 }
474
475 return true;
476}
477
478/* subsystems visibly enabled on a cgroup */
479static u16 cgroup_control(struct cgroup *cgrp)
480{
481 struct cgroup *parent = cgroup_parent(cgrp);
482 u16 root_ss_mask = cgrp->root->subsys_mask;
483
484 if (parent) {
485 u16 ss_mask = parent->subtree_control;
486
487 /* threaded cgroups can only have threaded controllers */
488 if (cgroup_is_threaded(cgrp))
489 ss_mask &= cgrp_dfl_threaded_ss_mask;
490 return ss_mask;
491 }
492
493 if (cgroup_on_dfl(cgrp))
494 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
495 cgrp_dfl_implicit_ss_mask);
496 return root_ss_mask;
497}
498
499/* subsystems enabled on a cgroup */
500static u16 cgroup_ss_mask(struct cgroup *cgrp)
501{
502 struct cgroup *parent = cgroup_parent(cgrp);
503
504 if (parent) {
505 u16 ss_mask = parent->subtree_ss_mask;
506
507 /* threaded cgroups can only have threaded controllers */
508 if (cgroup_is_threaded(cgrp))
509 ss_mask &= cgrp_dfl_threaded_ss_mask;
510 return ss_mask;
511 }
512
513 return cgrp->root->subsys_mask;
514}
515
516/**
517 * cgroup_css - obtain a cgroup's css for the specified subsystem
518 * @cgrp: the cgroup of interest
519 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
520 *
521 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
522 * function must be called either under cgroup_mutex or rcu_read_lock() and
523 * the caller is responsible for pinning the returned css if it wants to
524 * keep accessing it outside the said locks. This function may return
525 * %NULL if @cgrp doesn't have @subsys_id enabled.
526 */
527static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
528 struct cgroup_subsys *ss)
529{
530 if (CGROUP_HAS_SUBSYS_CONFIG && ss)
531 return rcu_dereference_check(cgrp->subsys[ss->id],
532 lockdep_is_held(&cgroup_mutex));
533 else
534 return &cgrp->self;
535}
536
537/**
538 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
539 * @cgrp: the cgroup of interest
540 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
541 *
542 * Similar to cgroup_css() but returns the effective css, which is defined
543 * as the matching css of the nearest ancestor including self which has @ss
544 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
545 * function is guaranteed to return non-NULL css.
546 */
547static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
548 struct cgroup_subsys *ss)
549{
550 lockdep_assert_held(&cgroup_mutex);
551
552 if (!ss)
553 return &cgrp->self;
554
555 /*
556 * This function is used while updating css associations and thus
557 * can't test the csses directly. Test ss_mask.
558 */
559 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
560 cgrp = cgroup_parent(cgrp);
561 if (!cgrp)
562 return NULL;
563 }
564
565 return cgroup_css(cgrp, ss);
566}
567
568/**
569 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
570 * @cgrp: the cgroup of interest
571 * @ss: the subsystem of interest
572 *
573 * Find and get the effective css of @cgrp for @ss. The effective css is
574 * defined as the matching css of the nearest ancestor including self which
575 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
576 * the root css is returned, so this function always returns a valid css.
577 *
578 * The returned css is not guaranteed to be online, and therefore it is the
579 * callers responsibility to try get a reference for it.
580 */
581struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
582 struct cgroup_subsys *ss)
583{
584 struct cgroup_subsys_state *css;
585
586 if (!CGROUP_HAS_SUBSYS_CONFIG)
587 return NULL;
588
589 do {
590 css = cgroup_css(cgrp, ss);
591
592 if (css)
593 return css;
594 cgrp = cgroup_parent(cgrp);
595 } while (cgrp);
596
597 return init_css_set.subsys[ss->id];
598}
599
600/**
601 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
602 * @cgrp: the cgroup of interest
603 * @ss: the subsystem of interest
604 *
605 * Find and get the effective css of @cgrp for @ss. The effective css is
606 * defined as the matching css of the nearest ancestor including self which
607 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
608 * the root css is returned, so this function always returns a valid css.
609 * The returned css must be put using css_put().
610 */
611struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
612 struct cgroup_subsys *ss)
613{
614 struct cgroup_subsys_state *css;
615
616 if (!CGROUP_HAS_SUBSYS_CONFIG)
617 return NULL;
618
619 rcu_read_lock();
620
621 do {
622 css = cgroup_css(cgrp, ss);
623
624 if (css && css_tryget_online(css))
625 goto out_unlock;
626 cgrp = cgroup_parent(cgrp);
627 } while (cgrp);
628
629 css = init_css_set.subsys[ss->id];
630 css_get(css);
631out_unlock:
632 rcu_read_unlock();
633 return css;
634}
635EXPORT_SYMBOL_GPL(cgroup_get_e_css);
636
637static void cgroup_get_live(struct cgroup *cgrp)
638{
639 WARN_ON_ONCE(cgroup_is_dead(cgrp));
640 cgroup_get(cgrp);
641}
642
643/**
644 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
645 * is responsible for taking the css_set_lock.
646 * @cgrp: the cgroup in question
647 */
648int __cgroup_task_count(const struct cgroup *cgrp)
649{
650 int count = 0;
651 struct cgrp_cset_link *link;
652
653 lockdep_assert_held(&css_set_lock);
654
655 list_for_each_entry(link, &cgrp->cset_links, cset_link)
656 count += link->cset->nr_tasks;
657
658 return count;
659}
660
661/**
662 * cgroup_task_count - count the number of tasks in a cgroup.
663 * @cgrp: the cgroup in question
664 */
665int cgroup_task_count(const struct cgroup *cgrp)
666{
667 int count;
668
669 spin_lock_irq(lock: &css_set_lock);
670 count = __cgroup_task_count(cgrp);
671 spin_unlock_irq(lock: &css_set_lock);
672
673 return count;
674}
675
676static struct cgroup *kn_priv(struct kernfs_node *kn)
677{
678 struct kernfs_node *parent;
679 /*
680 * The parent can not be replaced due to KERNFS_ROOT_INVARIANT_PARENT.
681 * Therefore it is always safe to dereference this pointer outside of a
682 * RCU section.
683 */
684 parent = rcu_dereference_check(kn->__parent,
685 kernfs_root_flags(kn) & KERNFS_ROOT_INVARIANT_PARENT);
686 return parent->priv;
687}
688
689struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
690{
691 struct cgroup *cgrp = kn_priv(kn: of->kn);
692 struct cftype *cft = of_cft(of);
693
694 /*
695 * This is open and unprotected implementation of cgroup_css().
696 * seq_css() is only called from a kernfs file operation which has
697 * an active reference on the file. Because all the subsystem
698 * files are drained before a css is disassociated with a cgroup,
699 * the matching css from the cgroup's subsys table is guaranteed to
700 * be and stay valid until the enclosing operation is complete.
701 */
702 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
703 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
704 else
705 return &cgrp->self;
706}
707EXPORT_SYMBOL_GPL(of_css);
708
709/**
710 * for_each_css - iterate all css's of a cgroup
711 * @css: the iteration cursor
712 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
713 * @cgrp: the target cgroup to iterate css's of
714 *
715 * Should be called under cgroup_mutex.
716 */
717#define for_each_css(css, ssid, cgrp) \
718 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
719 if (!((css) = rcu_dereference_check( \
720 (cgrp)->subsys[(ssid)], \
721 lockdep_is_held(&cgroup_mutex)))) { } \
722 else
723
724/**
725 * do_each_subsys_mask - filter for_each_subsys with a bitmask
726 * @ss: the iteration cursor
727 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
728 * @ss_mask: the bitmask
729 *
730 * The block will only run for cases where the ssid-th bit (1 << ssid) of
731 * @ss_mask is set.
732 */
733#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
734 unsigned long __ss_mask = (ss_mask); \
735 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
736 (ssid) = 0; \
737 break; \
738 } \
739 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
740 (ss) = cgroup_subsys[ssid]; \
741 {
742
743#define while_each_subsys_mask() \
744 } \
745 } \
746} while (false)
747
748/* iterate over child cgrps, lock should be held throughout iteration */
749#define cgroup_for_each_live_child(child, cgrp) \
750 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
751 if (({ lockdep_assert_held(&cgroup_mutex); \
752 cgroup_is_dead(child); })) \
753 ; \
754 else
755
756/* walk live descendants in pre order */
757#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
758 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
759 if (({ lockdep_assert_held(&cgroup_mutex); \
760 (dsct) = (d_css)->cgroup; \
761 cgroup_is_dead(dsct); })) \
762 ; \
763 else
764
765/* walk live descendants in postorder */
766#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
767 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
768 if (({ lockdep_assert_held(&cgroup_mutex); \
769 (dsct) = (d_css)->cgroup; \
770 cgroup_is_dead(dsct); })) \
771 ; \
772 else
773
774/*
775 * The default css_set - used by init and its children prior to any
776 * hierarchies being mounted. It contains a pointer to the root state
777 * for each subsystem. Also used to anchor the list of css_sets. Not
778 * reference-counted, to improve performance when child cgroups
779 * haven't been created.
780 */
781struct css_set init_css_set = {
782 .refcount = REFCOUNT_INIT(1),
783 .dom_cset = &init_css_set,
784 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
785 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
786 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
787 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
788 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
789 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
790 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
791 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
792 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
793
794 /*
795 * The following field is re-initialized when this cset gets linked
796 * in cgroup_init(). However, let's initialize the field
797 * statically too so that the default cgroup can be accessed safely
798 * early during boot.
799 */
800 .dfl_cgrp = &cgrp_dfl_root.cgrp,
801};
802
803static int css_set_count = 1; /* 1 for init_css_set */
804
805static bool css_set_threaded(struct css_set *cset)
806{
807 return cset->dom_cset != cset;
808}
809
810/**
811 * css_set_populated - does a css_set contain any tasks?
812 * @cset: target css_set
813 *
814 * css_set_populated() should be the same as !!cset->nr_tasks at steady
815 * state. However, css_set_populated() can be called while a task is being
816 * added to or removed from the linked list before the nr_tasks is
817 * properly updated. Hence, we can't just look at ->nr_tasks here.
818 */
819static bool css_set_populated(struct css_set *cset)
820{
821 lockdep_assert_held(&css_set_lock);
822
823 return !list_empty(head: &cset->tasks) || !list_empty(head: &cset->mg_tasks);
824}
825
826/**
827 * cgroup_update_populated - update the populated count of a cgroup
828 * @cgrp: the target cgroup
829 * @populated: inc or dec populated count
830 *
831 * One of the css_sets associated with @cgrp is either getting its first
832 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
833 * count is propagated towards root so that a given cgroup's
834 * nr_populated_children is zero iff none of its descendants contain any
835 * tasks.
836 *
837 * @cgrp's interface file "cgroup.populated" is zero if both
838 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
839 * 1 otherwise. When the sum changes from or to zero, userland is notified
840 * that the content of the interface file has changed. This can be used to
841 * detect when @cgrp and its descendants become populated or empty.
842 */
843static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
844{
845 struct cgroup *child = NULL;
846 int adj = populated ? 1 : -1;
847
848 lockdep_assert_held(&css_set_lock);
849
850 do {
851 bool was_populated = cgroup_is_populated(cgrp);
852
853 if (!child) {
854 cgrp->nr_populated_csets += adj;
855 } else {
856 if (cgroup_is_threaded(cgrp: child))
857 cgrp->nr_populated_threaded_children += adj;
858 else
859 cgrp->nr_populated_domain_children += adj;
860 }
861
862 if (was_populated == cgroup_is_populated(cgrp))
863 break;
864
865 cgroup1_check_for_release(cgrp);
866 TRACE_CGROUP_PATH(notify_populated, cgrp,
867 cgroup_is_populated(cgrp));
868 cgroup_file_notify(cfile: &cgrp->events_file);
869
870 child = cgrp;
871 cgrp = cgroup_parent(cgrp);
872 } while (cgrp);
873}
874
875/**
876 * css_set_update_populated - update populated state of a css_set
877 * @cset: target css_set
878 * @populated: whether @cset is populated or depopulated
879 *
880 * @cset is either getting the first task or losing the last. Update the
881 * populated counters of all associated cgroups accordingly.
882 */
883static void css_set_update_populated(struct css_set *cset, bool populated)
884{
885 struct cgrp_cset_link *link;
886
887 lockdep_assert_held(&css_set_lock);
888
889 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
890 cgroup_update_populated(cgrp: link->cgrp, populated);
891}
892
893/*
894 * @task is leaving, advance task iterators which are pointing to it so
895 * that they can resume at the next position. Advancing an iterator might
896 * remove it from the list, use safe walk. See css_task_iter_skip() for
897 * details.
898 */
899static void css_set_skip_task_iters(struct css_set *cset,
900 struct task_struct *task)
901{
902 struct css_task_iter *it, *pos;
903
904 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
905 css_task_iter_skip(it, task);
906}
907
908/**
909 * css_set_move_task - move a task from one css_set to another
910 * @task: task being moved
911 * @from_cset: css_set @task currently belongs to (may be NULL)
912 * @to_cset: new css_set @task is being moved to (may be NULL)
913 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
914 *
915 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
916 * css_set, @from_cset can be NULL. If @task is being disassociated
917 * instead of moved, @to_cset can be NULL.
918 *
919 * This function automatically handles populated counter updates and
920 * css_task_iter adjustments but the caller is responsible for managing
921 * @from_cset and @to_cset's reference counts.
922 */
923static void css_set_move_task(struct task_struct *task,
924 struct css_set *from_cset, struct css_set *to_cset,
925 bool use_mg_tasks)
926{
927 lockdep_assert_held(&css_set_lock);
928
929 if (to_cset && !css_set_populated(cset: to_cset))
930 css_set_update_populated(cset: to_cset, populated: true);
931
932 if (from_cset) {
933 WARN_ON_ONCE(list_empty(&task->cg_list));
934
935 css_set_skip_task_iters(cset: from_cset, task);
936 list_del_init(entry: &task->cg_list);
937 if (!css_set_populated(cset: from_cset))
938 css_set_update_populated(cset: from_cset, populated: false);
939 } else {
940 WARN_ON_ONCE(!list_empty(&task->cg_list));
941 }
942
943 if (to_cset) {
944 /*
945 * We are synchronized through cgroup_threadgroup_rwsem
946 * against PF_EXITING setting such that we can't race
947 * against cgroup_exit()/cgroup_free() dropping the css_set.
948 */
949 WARN_ON_ONCE(task->flags & PF_EXITING);
950
951 cgroup_move_task(p: task, to: to_cset);
952 list_add_tail(new: &task->cg_list, head: use_mg_tasks ? &to_cset->mg_tasks :
953 &to_cset->tasks);
954 }
955}
956
957/*
958 * hash table for cgroup groups. This improves the performance to find
959 * an existing css_set. This hash doesn't (currently) take into
960 * account cgroups in empty hierarchies.
961 */
962#define CSS_SET_HASH_BITS 7
963static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
964
965static unsigned long css_set_hash(struct cgroup_subsys_state **css)
966{
967 unsigned long key = 0UL;
968 struct cgroup_subsys *ss;
969 int i;
970
971 for_each_subsys(ss, i)
972 key += (unsigned long)css[i];
973 key = (key >> 16) ^ key;
974
975 return key;
976}
977
978void put_css_set_locked(struct css_set *cset)
979{
980 struct cgrp_cset_link *link, *tmp_link;
981 struct cgroup_subsys *ss;
982 int ssid;
983
984 lockdep_assert_held(&css_set_lock);
985
986 if (!refcount_dec_and_test(r: &cset->refcount))
987 return;
988
989 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
990
991 /* This css_set is dead. Unlink it and release cgroup and css refs */
992 for_each_subsys(ss, ssid) {
993 list_del(entry: &cset->e_cset_node[ssid]);
994 css_put(css: cset->subsys[ssid]);
995 }
996 hash_del(node: &cset->hlist);
997 css_set_count--;
998
999 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
1000 list_del(entry: &link->cset_link);
1001 list_del(entry: &link->cgrp_link);
1002 if (cgroup_parent(cgrp: link->cgrp))
1003 cgroup_put(cgrp: link->cgrp);
1004 kfree(objp: link);
1005 }
1006
1007 if (css_set_threaded(cset)) {
1008 list_del(entry: &cset->threaded_csets_node);
1009 put_css_set_locked(cset: cset->dom_cset);
1010 }
1011
1012 kfree_rcu(cset, rcu_head);
1013}
1014
1015/**
1016 * compare_css_sets - helper function for find_existing_css_set().
1017 * @cset: candidate css_set being tested
1018 * @old_cset: existing css_set for a task
1019 * @new_cgrp: cgroup that's being entered by the task
1020 * @template: desired set of css pointers in css_set (pre-calculated)
1021 *
1022 * Returns true if "cset" matches "old_cset" except for the hierarchy
1023 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
1024 */
1025static bool compare_css_sets(struct css_set *cset,
1026 struct css_set *old_cset,
1027 struct cgroup *new_cgrp,
1028 struct cgroup_subsys_state *template[])
1029{
1030 struct cgroup *new_dfl_cgrp;
1031 struct list_head *l1, *l2;
1032
1033 /*
1034 * On the default hierarchy, there can be csets which are
1035 * associated with the same set of cgroups but different csses.
1036 * Let's first ensure that csses match.
1037 */
1038 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1039 return false;
1040
1041
1042 /* @cset's domain should match the default cgroup's */
1043 if (cgroup_on_dfl(cgrp: new_cgrp))
1044 new_dfl_cgrp = new_cgrp;
1045 else
1046 new_dfl_cgrp = old_cset->dfl_cgrp;
1047
1048 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1049 return false;
1050
1051 /*
1052 * Compare cgroup pointers in order to distinguish between
1053 * different cgroups in hierarchies. As different cgroups may
1054 * share the same effective css, this comparison is always
1055 * necessary.
1056 */
1057 l1 = &cset->cgrp_links;
1058 l2 = &old_cset->cgrp_links;
1059 while (1) {
1060 struct cgrp_cset_link *link1, *link2;
1061 struct cgroup *cgrp1, *cgrp2;
1062
1063 l1 = l1->next;
1064 l2 = l2->next;
1065 /* See if we reached the end - both lists are equal length. */
1066 if (l1 == &cset->cgrp_links) {
1067 BUG_ON(l2 != &old_cset->cgrp_links);
1068 break;
1069 } else {
1070 BUG_ON(l2 == &old_cset->cgrp_links);
1071 }
1072 /* Locate the cgroups associated with these links. */
1073 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1074 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1075 cgrp1 = link1->cgrp;
1076 cgrp2 = link2->cgrp;
1077 /* Hierarchies should be linked in the same order. */
1078 BUG_ON(cgrp1->root != cgrp2->root);
1079
1080 /*
1081 * If this hierarchy is the hierarchy of the cgroup
1082 * that's changing, then we need to check that this
1083 * css_set points to the new cgroup; if it's any other
1084 * hierarchy, then this css_set should point to the
1085 * same cgroup as the old css_set.
1086 */
1087 if (cgrp1->root == new_cgrp->root) {
1088 if (cgrp1 != new_cgrp)
1089 return false;
1090 } else {
1091 if (cgrp1 != cgrp2)
1092 return false;
1093 }
1094 }
1095 return true;
1096}
1097
1098/**
1099 * find_existing_css_set - init css array and find the matching css_set
1100 * @old_cset: the css_set that we're using before the cgroup transition
1101 * @cgrp: the cgroup that we're moving into
1102 * @template: out param for the new set of csses, should be clear on entry
1103 */
1104static struct css_set *find_existing_css_set(struct css_set *old_cset,
1105 struct cgroup *cgrp,
1106 struct cgroup_subsys_state **template)
1107{
1108 struct cgroup_root *root = cgrp->root;
1109 struct cgroup_subsys *ss;
1110 struct css_set *cset;
1111 unsigned long key;
1112 int i;
1113
1114 /*
1115 * Build the set of subsystem state objects that we want to see in the
1116 * new css_set. While subsystems can change globally, the entries here
1117 * won't change, so no need for locking.
1118 */
1119 for_each_subsys(ss, i) {
1120 if (root->subsys_mask & (1UL << i)) {
1121 /*
1122 * @ss is in this hierarchy, so we want the
1123 * effective css from @cgrp.
1124 */
1125 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1126 } else {
1127 /*
1128 * @ss is not in this hierarchy, so we don't want
1129 * to change the css.
1130 */
1131 template[i] = old_cset->subsys[i];
1132 }
1133 }
1134
1135 key = css_set_hash(css: template);
1136 hash_for_each_possible(css_set_table, cset, hlist, key) {
1137 if (!compare_css_sets(cset, old_cset, new_cgrp: cgrp, template))
1138 continue;
1139
1140 /* This css_set matches what we need */
1141 return cset;
1142 }
1143
1144 /* No existing cgroup group matched */
1145 return NULL;
1146}
1147
1148static void free_cgrp_cset_links(struct list_head *links_to_free)
1149{
1150 struct cgrp_cset_link *link, *tmp_link;
1151
1152 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1153 list_del(entry: &link->cset_link);
1154 kfree(objp: link);
1155 }
1156}
1157
1158/**
1159 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1160 * @count: the number of links to allocate
1161 * @tmp_links: list_head the allocated links are put on
1162 *
1163 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1164 * through ->cset_link. Returns 0 on success or -errno.
1165 */
1166static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1167{
1168 struct cgrp_cset_link *link;
1169 int i;
1170
1171 INIT_LIST_HEAD(list: tmp_links);
1172
1173 for (i = 0; i < count; i++) {
1174 link = kzalloc(sizeof(*link), GFP_KERNEL);
1175 if (!link) {
1176 free_cgrp_cset_links(links_to_free: tmp_links);
1177 return -ENOMEM;
1178 }
1179 list_add(new: &link->cset_link, head: tmp_links);
1180 }
1181 return 0;
1182}
1183
1184/**
1185 * link_css_set - a helper function to link a css_set to a cgroup
1186 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1187 * @cset: the css_set to be linked
1188 * @cgrp: the destination cgroup
1189 */
1190static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1191 struct cgroup *cgrp)
1192{
1193 struct cgrp_cset_link *link;
1194
1195 BUG_ON(list_empty(tmp_links));
1196
1197 if (cgroup_on_dfl(cgrp))
1198 cset->dfl_cgrp = cgrp;
1199
1200 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1201 link->cset = cset;
1202 link->cgrp = cgrp;
1203
1204 /*
1205 * Always add links to the tail of the lists so that the lists are
1206 * in chronological order.
1207 */
1208 list_move_tail(list: &link->cset_link, head: &cgrp->cset_links);
1209 list_add_tail(new: &link->cgrp_link, head: &cset->cgrp_links);
1210
1211 if (cgroup_parent(cgrp))
1212 cgroup_get_live(cgrp);
1213}
1214
1215/**
1216 * find_css_set - return a new css_set with one cgroup updated
1217 * @old_cset: the baseline css_set
1218 * @cgrp: the cgroup to be updated
1219 *
1220 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1221 * substituted into the appropriate hierarchy.
1222 */
1223static struct css_set *find_css_set(struct css_set *old_cset,
1224 struct cgroup *cgrp)
1225{
1226 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1227 struct css_set *cset;
1228 struct list_head tmp_links;
1229 struct cgrp_cset_link *link;
1230 struct cgroup_subsys *ss;
1231 unsigned long key;
1232 int ssid;
1233
1234 lockdep_assert_held(&cgroup_mutex);
1235
1236 /* First see if we already have a cgroup group that matches
1237 * the desired set */
1238 spin_lock_irq(lock: &css_set_lock);
1239 cset = find_existing_css_set(old_cset, cgrp, template);
1240 if (cset)
1241 get_css_set(cset);
1242 spin_unlock_irq(lock: &css_set_lock);
1243
1244 if (cset)
1245 return cset;
1246
1247 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1248 if (!cset)
1249 return NULL;
1250
1251 /* Allocate all the cgrp_cset_link objects that we'll need */
1252 if (allocate_cgrp_cset_links(count: cgroup_root_count, tmp_links: &tmp_links) < 0) {
1253 kfree(objp: cset);
1254 return NULL;
1255 }
1256
1257 refcount_set(r: &cset->refcount, n: 1);
1258 cset->dom_cset = cset;
1259 INIT_LIST_HEAD(list: &cset->tasks);
1260 INIT_LIST_HEAD(list: &cset->mg_tasks);
1261 INIT_LIST_HEAD(list: &cset->dying_tasks);
1262 INIT_LIST_HEAD(list: &cset->task_iters);
1263 INIT_LIST_HEAD(list: &cset->threaded_csets);
1264 INIT_HLIST_NODE(h: &cset->hlist);
1265 INIT_LIST_HEAD(list: &cset->cgrp_links);
1266 INIT_LIST_HEAD(list: &cset->mg_src_preload_node);
1267 INIT_LIST_HEAD(list: &cset->mg_dst_preload_node);
1268 INIT_LIST_HEAD(list: &cset->mg_node);
1269
1270 /* Copy the set of subsystem state objects generated in
1271 * find_existing_css_set() */
1272 memcpy(to: cset->subsys, from: template, len: sizeof(cset->subsys));
1273
1274 spin_lock_irq(lock: &css_set_lock);
1275 /* Add reference counts and links from the new css_set. */
1276 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1277 struct cgroup *c = link->cgrp;
1278
1279 if (c->root == cgrp->root)
1280 c = cgrp;
1281 link_css_set(tmp_links: &tmp_links, cset, cgrp: c);
1282 }
1283
1284 BUG_ON(!list_empty(&tmp_links));
1285
1286 css_set_count++;
1287
1288 /* Add @cset to the hash table */
1289 key = css_set_hash(css: cset->subsys);
1290 hash_add(css_set_table, &cset->hlist, key);
1291
1292 for_each_subsys(ss, ssid) {
1293 struct cgroup_subsys_state *css = cset->subsys[ssid];
1294
1295 list_add_tail(new: &cset->e_cset_node[ssid],
1296 head: &css->cgroup->e_csets[ssid]);
1297 css_get(css);
1298 }
1299
1300 spin_unlock_irq(lock: &css_set_lock);
1301
1302 /*
1303 * If @cset should be threaded, look up the matching dom_cset and
1304 * link them up. We first fully initialize @cset then look for the
1305 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1306 * to stay empty until we return.
1307 */
1308 if (cgroup_is_threaded(cgrp: cset->dfl_cgrp)) {
1309 struct css_set *dcset;
1310
1311 dcset = find_css_set(old_cset: cset, cgrp: cset->dfl_cgrp->dom_cgrp);
1312 if (!dcset) {
1313 put_css_set(cset);
1314 return NULL;
1315 }
1316
1317 spin_lock_irq(lock: &css_set_lock);
1318 cset->dom_cset = dcset;
1319 list_add_tail(new: &cset->threaded_csets_node,
1320 head: &dcset->threaded_csets);
1321 spin_unlock_irq(lock: &css_set_lock);
1322 }
1323
1324 return cset;
1325}
1326
1327struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1328{
1329 struct cgroup *root_cgrp = kernfs_root_to_node(root: kf_root)->priv;
1330
1331 return root_cgrp->root;
1332}
1333
1334void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1335{
1336 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1337
1338 /*
1339 * see the comment above CGRP_ROOT_FAVOR_DYNMODS definition.
1340 * favordynmods can flip while task is between
1341 * cgroup_threadgroup_change_begin() and end(), so down_write global
1342 * cgroup_threadgroup_rwsem to synchronize them.
1343 *
1344 * Once cgroup_enable_per_threadgroup_rwsem is enabled, holding
1345 * cgroup_threadgroup_rwsem doesn't exlude tasks between
1346 * cgroup_thread_group_change_begin() and end() and thus it's unsafe to
1347 * turn off. As the scenario is unlikely, simply disallow disabling once
1348 * enabled and print out a warning.
1349 */
1350 percpu_down_write(&cgroup_threadgroup_rwsem);
1351 if (favor && !favoring) {
1352 cgroup_enable_per_threadgroup_rwsem = true;
1353 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1354 root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1355 } else if (!favor && favoring) {
1356 if (cgroup_enable_per_threadgroup_rwsem)
1357 pr_warn_once("cgroup favordynmods: per threadgroup rwsem mechanism can't be disabled\n");
1358 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1359 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1360 }
1361 percpu_up_write(&cgroup_threadgroup_rwsem);
1362}
1363
1364static int cgroup_init_root_id(struct cgroup_root *root)
1365{
1366 int id;
1367
1368 lockdep_assert_held(&cgroup_mutex);
1369
1370 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, ptr: root, start: 0, end: 0, GFP_KERNEL);
1371 if (id < 0)
1372 return id;
1373
1374 root->hierarchy_id = id;
1375 return 0;
1376}
1377
1378static void cgroup_exit_root_id(struct cgroup_root *root)
1379{
1380 lockdep_assert_held(&cgroup_mutex);
1381
1382 idr_remove(&cgroup_hierarchy_idr, id: root->hierarchy_id);
1383}
1384
1385void cgroup_free_root(struct cgroup_root *root)
1386{
1387 kfree_rcu(root, rcu);
1388}
1389
1390static void cgroup_destroy_root(struct cgroup_root *root)
1391{
1392 struct cgroup *cgrp = &root->cgrp;
1393 struct cgrp_cset_link *link, *tmp_link;
1394 int ret;
1395
1396 trace_cgroup_destroy_root(root);
1397
1398 cgroup_lock_and_drain_offline(cgrp: &cgrp_dfl_root.cgrp);
1399
1400 BUG_ON(atomic_read(&root->nr_cgrps));
1401 BUG_ON(!list_empty(&cgrp->self.children));
1402
1403 ret = blocking_notifier_call_chain(nh: &cgroup_lifetime_notifier,
1404 val: CGROUP_LIFETIME_OFFLINE, v: cgrp);
1405 WARN_ON_ONCE(notifier_to_errno(ret));
1406
1407 /* Rebind all subsystems back to the default hierarchy */
1408 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1409
1410 /*
1411 * Release all the links from cset_links to this hierarchy's
1412 * root cgroup
1413 */
1414 spin_lock_irq(lock: &css_set_lock);
1415
1416 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1417 list_del(entry: &link->cset_link);
1418 list_del(entry: &link->cgrp_link);
1419 kfree(objp: link);
1420 }
1421
1422 spin_unlock_irq(lock: &css_set_lock);
1423
1424 WARN_ON_ONCE(list_empty(&root->root_list));
1425 list_del_rcu(entry: &root->root_list);
1426 cgroup_root_count--;
1427
1428 if (!have_favordynmods)
1429 cgroup_favor_dynmods(root, favor: false);
1430
1431 cgroup_exit_root_id(root);
1432
1433 cgroup_unlock();
1434
1435 kernfs_destroy_root(root: root->kf_root);
1436 cgroup_free_root(root);
1437}
1438
1439/*
1440 * Returned cgroup is without refcount but it's valid as long as cset pins it.
1441 */
1442static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1443 struct cgroup_root *root)
1444{
1445 struct cgroup *res_cgroup = NULL;
1446
1447 if (cset == &init_css_set) {
1448 res_cgroup = &root->cgrp;
1449 } else if (root == &cgrp_dfl_root) {
1450 res_cgroup = cset->dfl_cgrp;
1451 } else {
1452 struct cgrp_cset_link *link;
1453 lockdep_assert_held(&css_set_lock);
1454
1455 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1456 struct cgroup *c = link->cgrp;
1457
1458 if (c->root == root) {
1459 res_cgroup = c;
1460 break;
1461 }
1462 }
1463 }
1464
1465 /*
1466 * If cgroup_mutex is not held, the cgrp_cset_link will be freed
1467 * before we remove the cgroup root from the root_list. Consequently,
1468 * when accessing a cgroup root, the cset_link may have already been
1469 * freed, resulting in a NULL res_cgroup. However, by holding the
1470 * cgroup_mutex, we ensure that res_cgroup can't be NULL.
1471 * If we don't hold cgroup_mutex in the caller, we must do the NULL
1472 * check.
1473 */
1474 return res_cgroup;
1475}
1476
1477/*
1478 * look up cgroup associated with current task's cgroup namespace on the
1479 * specified hierarchy
1480 */
1481static struct cgroup *
1482current_cgns_cgroup_from_root(struct cgroup_root *root)
1483{
1484 struct cgroup *res = NULL;
1485 struct css_set *cset;
1486
1487 lockdep_assert_held(&css_set_lock);
1488
1489 rcu_read_lock();
1490
1491 cset = current->nsproxy->cgroup_ns->root_cset;
1492 res = __cset_cgroup_from_root(cset, root);
1493
1494 rcu_read_unlock();
1495
1496 /*
1497 * The namespace_sem is held by current, so the root cgroup can't
1498 * be umounted. Therefore, we can ensure that the res is non-NULL.
1499 */
1500 WARN_ON_ONCE(!res);
1501 return res;
1502}
1503
1504/*
1505 * Look up cgroup associated with current task's cgroup namespace on the default
1506 * hierarchy.
1507 *
1508 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1509 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1510 * pointers.
1511 * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1512 * - As a bonus returned cgrp is pinned with the current because it cannot
1513 * switch cgroup_ns asynchronously.
1514 */
1515static struct cgroup *current_cgns_cgroup_dfl(void)
1516{
1517 struct css_set *cset;
1518
1519 if (current->nsproxy) {
1520 cset = current->nsproxy->cgroup_ns->root_cset;
1521 return __cset_cgroup_from_root(cset, root: &cgrp_dfl_root);
1522 } else {
1523 /*
1524 * NOTE: This function may be called from bpf_cgroup_from_id()
1525 * on a task which has already passed exit_task_namespaces() and
1526 * nsproxy == NULL. Fall back to cgrp_dfl_root which will make all
1527 * cgroups visible for lookups.
1528 */
1529 return &cgrp_dfl_root.cgrp;
1530 }
1531}
1532
1533/* look up cgroup associated with given css_set on the specified hierarchy */
1534static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1535 struct cgroup_root *root)
1536{
1537 lockdep_assert_held(&css_set_lock);
1538
1539 return __cset_cgroup_from_root(cset, root);
1540}
1541
1542/*
1543 * Return the cgroup for "task" from the given hierarchy. Must be
1544 * called with css_set_lock held to prevent task's groups from being modified.
1545 * Must be called with either cgroup_mutex or rcu read lock to prevent the
1546 * cgroup root from being destroyed.
1547 */
1548struct cgroup *task_cgroup_from_root(struct task_struct *task,
1549 struct cgroup_root *root)
1550{
1551 /*
1552 * No need to lock the task - since we hold css_set_lock the
1553 * task can't change groups.
1554 */
1555 return cset_cgroup_from_root(cset: task_css_set(task), root);
1556}
1557
1558/*
1559 * A task must hold cgroup_mutex to modify cgroups.
1560 *
1561 * Any task can increment and decrement the count field without lock.
1562 * So in general, code holding cgroup_mutex can't rely on the count
1563 * field not changing. However, if the count goes to zero, then only
1564 * cgroup_attach_task() can increment it again. Because a count of zero
1565 * means that no tasks are currently attached, therefore there is no
1566 * way a task attached to that cgroup can fork (the other way to
1567 * increment the count). So code holding cgroup_mutex can safely
1568 * assume that if the count is zero, it will stay zero. Similarly, if
1569 * a task holds cgroup_mutex on a cgroup with zero count, it
1570 * knows that the cgroup won't be removed, as cgroup_rmdir()
1571 * needs that mutex.
1572 *
1573 * A cgroup can only be deleted if both its 'count' of using tasks
1574 * is zero, and its list of 'children' cgroups is empty. Since all
1575 * tasks in the system use _some_ cgroup, and since there is always at
1576 * least one task in the system (init, pid == 1), therefore, root cgroup
1577 * always has either children cgroups and/or using tasks. So we don't
1578 * need a special hack to ensure that root cgroup cannot be deleted.
1579 *
1580 * P.S. One more locking exception. RCU is used to guard the
1581 * update of a tasks cgroup pointer by cgroup_attach_task()
1582 */
1583
1584static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1585
1586static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1587 char *buf)
1588{
1589 struct cgroup_subsys *ss = cft->ss;
1590
1591 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1592 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1593 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1594
1595 snprintf(buf, CGROUP_FILE_NAME_MAX, fmt: "%s%s.%s",
1596 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1597 cft->name);
1598 } else {
1599 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1600 }
1601 return buf;
1602}
1603
1604/**
1605 * cgroup_file_mode - deduce file mode of a control file
1606 * @cft: the control file in question
1607 *
1608 * S_IRUGO for read, S_IWUSR for write.
1609 */
1610static umode_t cgroup_file_mode(const struct cftype *cft)
1611{
1612 umode_t mode = 0;
1613
1614 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1615 mode |= S_IRUGO;
1616
1617 if (cft->write_u64 || cft->write_s64 || cft->write) {
1618 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1619 mode |= S_IWUGO;
1620 else
1621 mode |= S_IWUSR;
1622 }
1623
1624 return mode;
1625}
1626
1627/**
1628 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1629 * @subtree_control: the new subtree_control mask to consider
1630 * @this_ss_mask: available subsystems
1631 *
1632 * On the default hierarchy, a subsystem may request other subsystems to be
1633 * enabled together through its ->depends_on mask. In such cases, more
1634 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1635 *
1636 * This function calculates which subsystems need to be enabled if
1637 * @subtree_control is to be applied while restricted to @this_ss_mask.
1638 */
1639static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1640{
1641 u16 cur_ss_mask = subtree_control;
1642 struct cgroup_subsys *ss;
1643 int ssid;
1644
1645 lockdep_assert_held(&cgroup_mutex);
1646
1647 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1648
1649 while (true) {
1650 u16 new_ss_mask = cur_ss_mask;
1651
1652 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1653 new_ss_mask |= ss->depends_on;
1654 } while_each_subsys_mask();
1655
1656 /*
1657 * Mask out subsystems which aren't available. This can
1658 * happen only if some depended-upon subsystems were bound
1659 * to non-default hierarchies.
1660 */
1661 new_ss_mask &= this_ss_mask;
1662
1663 if (new_ss_mask == cur_ss_mask)
1664 break;
1665 cur_ss_mask = new_ss_mask;
1666 }
1667
1668 return cur_ss_mask;
1669}
1670
1671/**
1672 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1673 * @kn: the kernfs_node being serviced
1674 *
1675 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1676 * the method finishes if locking succeeded. Note that once this function
1677 * returns the cgroup returned by cgroup_kn_lock_live() may become
1678 * inaccessible any time. If the caller intends to continue to access the
1679 * cgroup, it should pin it before invoking this function.
1680 */
1681void cgroup_kn_unlock(struct kernfs_node *kn)
1682{
1683 struct cgroup *cgrp;
1684
1685 if (kernfs_type(kn) == KERNFS_DIR)
1686 cgrp = kn->priv;
1687 else
1688 cgrp = kn_priv(kn);
1689
1690 cgroup_unlock();
1691
1692 kernfs_unbreak_active_protection(kn);
1693 cgroup_put(cgrp);
1694}
1695
1696/**
1697 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1698 * @kn: the kernfs_node being serviced
1699 * @drain_offline: perform offline draining on the cgroup
1700 *
1701 * This helper is to be used by a cgroup kernfs method currently servicing
1702 * @kn. It breaks the active protection, performs cgroup locking and
1703 * verifies that the associated cgroup is alive. Returns the cgroup if
1704 * alive; otherwise, %NULL. A successful return should be undone by a
1705 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1706 * cgroup is drained of offlining csses before return.
1707 *
1708 * Any cgroup kernfs method implementation which requires locking the
1709 * associated cgroup should use this helper. It avoids nesting cgroup
1710 * locking under kernfs active protection and allows all kernfs operations
1711 * including self-removal.
1712 */
1713struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1714{
1715 struct cgroup *cgrp;
1716
1717 if (kernfs_type(kn) == KERNFS_DIR)
1718 cgrp = kn->priv;
1719 else
1720 cgrp = kn_priv(kn);
1721
1722 /*
1723 * We're gonna grab cgroup_mutex which nests outside kernfs
1724 * active_ref. cgroup liveliness check alone provides enough
1725 * protection against removal. Ensure @cgrp stays accessible and
1726 * break the active_ref protection.
1727 */
1728 if (!cgroup_tryget(cgrp))
1729 return NULL;
1730 kernfs_break_active_protection(kn);
1731
1732 if (drain_offline)
1733 cgroup_lock_and_drain_offline(cgrp);
1734 else
1735 cgroup_lock();
1736
1737 if (!cgroup_is_dead(cgrp))
1738 return cgrp;
1739
1740 cgroup_kn_unlock(kn);
1741 return NULL;
1742}
1743
1744static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1745{
1746 char name[CGROUP_FILE_NAME_MAX];
1747
1748 lockdep_assert_held(&cgroup_mutex);
1749
1750 if (cft->file_offset) {
1751 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss: cft->ss);
1752 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1753
1754 spin_lock_irq(lock: &cgroup_file_kn_lock);
1755 cfile->kn = NULL;
1756 spin_unlock_irq(lock: &cgroup_file_kn_lock);
1757
1758 timer_delete_sync(timer: &cfile->notify_timer);
1759 }
1760
1761 kernfs_remove_by_name(parent: cgrp->kn, name: cgroup_file_name(cgrp, cft, buf: name));
1762}
1763
1764/**
1765 * css_clear_dir - remove subsys files in a cgroup directory
1766 * @css: target css
1767 */
1768static void css_clear_dir(struct cgroup_subsys_state *css)
1769{
1770 struct cgroup *cgrp = css->cgroup;
1771 struct cftype *cfts;
1772
1773 if (!(css->flags & CSS_VISIBLE))
1774 return;
1775
1776 css->flags &= ~CSS_VISIBLE;
1777
1778 if (css_is_self(css)) {
1779 if (cgroup_on_dfl(cgrp)) {
1780 cgroup_addrm_files(css, cgrp,
1781 cfts: cgroup_base_files, is_add: false);
1782 if (cgroup_psi_enabled())
1783 cgroup_addrm_files(css, cgrp,
1784 cfts: cgroup_psi_files, is_add: false);
1785 } else {
1786 cgroup_addrm_files(css, cgrp,
1787 cfts: cgroup1_base_files, is_add: false);
1788 }
1789 } else {
1790 list_for_each_entry(cfts, &css->ss->cfts, node)
1791 cgroup_addrm_files(css, cgrp, cfts, is_add: false);
1792 }
1793}
1794
1795/**
1796 * css_populate_dir - create subsys files in a cgroup directory
1797 * @css: target css
1798 *
1799 * On failure, no file is added.
1800 */
1801static int css_populate_dir(struct cgroup_subsys_state *css)
1802{
1803 struct cgroup *cgrp = css->cgroup;
1804 struct cftype *cfts, *failed_cfts;
1805 int ret;
1806
1807 if (css->flags & CSS_VISIBLE)
1808 return 0;
1809
1810 if (css_is_self(css)) {
1811 if (cgroup_on_dfl(cgrp)) {
1812 ret = cgroup_addrm_files(css, cgrp,
1813 cfts: cgroup_base_files, is_add: true);
1814 if (ret < 0)
1815 return ret;
1816
1817 if (cgroup_psi_enabled()) {
1818 ret = cgroup_addrm_files(css, cgrp,
1819 cfts: cgroup_psi_files, is_add: true);
1820 if (ret < 0) {
1821 cgroup_addrm_files(css, cgrp,
1822 cfts: cgroup_base_files, is_add: false);
1823 return ret;
1824 }
1825 }
1826 } else {
1827 ret = cgroup_addrm_files(css, cgrp,
1828 cfts: cgroup1_base_files, is_add: true);
1829 if (ret < 0)
1830 return ret;
1831 }
1832 } else {
1833 list_for_each_entry(cfts, &css->ss->cfts, node) {
1834 ret = cgroup_addrm_files(css, cgrp, cfts, is_add: true);
1835 if (ret < 0) {
1836 failed_cfts = cfts;
1837 goto err;
1838 }
1839 }
1840 }
1841
1842 css->flags |= CSS_VISIBLE;
1843
1844 return 0;
1845err:
1846 list_for_each_entry(cfts, &css->ss->cfts, node) {
1847 if (cfts == failed_cfts)
1848 break;
1849 cgroup_addrm_files(css, cgrp, cfts, is_add: false);
1850 }
1851 return ret;
1852}
1853
1854int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1855{
1856 struct cgroup *dcgrp = &dst_root->cgrp;
1857 struct cgroup_subsys *ss;
1858 int ssid, ret;
1859 u16 dfl_disable_ss_mask = 0;
1860
1861 lockdep_assert_held(&cgroup_mutex);
1862
1863 do_each_subsys_mask(ss, ssid, ss_mask) {
1864 /*
1865 * If @ss has non-root csses attached to it, can't move.
1866 * If @ss is an implicit controller, it is exempt from this
1867 * rule and can be stolen.
1868 */
1869 if (css_next_child(NULL, parent: cgroup_css(cgrp: &ss->root->cgrp, ss)) &&
1870 !ss->implicit_on_dfl)
1871 return -EBUSY;
1872
1873 /* can't move between two non-dummy roots either */
1874 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1875 return -EBUSY;
1876
1877 /*
1878 * Collect ssid's that need to be disabled from default
1879 * hierarchy.
1880 */
1881 if (ss->root == &cgrp_dfl_root)
1882 dfl_disable_ss_mask |= 1 << ssid;
1883
1884 } while_each_subsys_mask();
1885
1886 if (dfl_disable_ss_mask) {
1887 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1888
1889 /*
1890 * Controllers from default hierarchy that need to be rebound
1891 * are all disabled together in one go.
1892 */
1893 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1894 WARN_ON(cgroup_apply_control(scgrp));
1895 cgroup_finalize_control(cgrp: scgrp, ret: 0);
1896 }
1897
1898 do_each_subsys_mask(ss, ssid, ss_mask) {
1899 struct cgroup_root *src_root = ss->root;
1900 struct cgroup *scgrp = &src_root->cgrp;
1901 struct cgroup_subsys_state *css = cgroup_css(cgrp: scgrp, ss);
1902 struct css_set *cset, *cset_pos;
1903 struct css_task_iter *it;
1904
1905 WARN_ON(!css || cgroup_css(dcgrp, ss));
1906
1907 if (src_root != &cgrp_dfl_root) {
1908 /* disable from the source */
1909 src_root->subsys_mask &= ~(1 << ssid);
1910 WARN_ON(cgroup_apply_control(scgrp));
1911 cgroup_finalize_control(cgrp: scgrp, ret: 0);
1912 }
1913
1914 /* rebind */
1915 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1916 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1917 ss->root = dst_root;
1918
1919 spin_lock_irq(lock: &css_set_lock);
1920 css->cgroup = dcgrp;
1921 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id]));
1922 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id],
1923 e_cset_node[ss->id]) {
1924 list_move_tail(list: &cset->e_cset_node[ss->id],
1925 head: &dcgrp->e_csets[ss->id]);
1926 /*
1927 * all css_sets of scgrp together in same order to dcgrp,
1928 * patch in-flight iterators to preserve correct iteration.
1929 * since the iterator is always advanced right away and
1930 * finished when it->cset_pos meets it->cset_head, so only
1931 * update it->cset_head is enough here.
1932 */
1933 list_for_each_entry(it, &cset->task_iters, iters_node)
1934 if (it->cset_head == &scgrp->e_csets[ss->id])
1935 it->cset_head = &dcgrp->e_csets[ss->id];
1936 }
1937 spin_unlock_irq(lock: &css_set_lock);
1938
1939 /* default hierarchy doesn't enable controllers by default */
1940 dst_root->subsys_mask |= 1 << ssid;
1941 if (dst_root == &cgrp_dfl_root) {
1942 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1943 } else {
1944 dcgrp->subtree_control |= 1 << ssid;
1945 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1946 }
1947
1948 ret = cgroup_apply_control(cgrp: dcgrp);
1949 if (ret)
1950 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1951 ss->name, ret);
1952
1953 if (ss->bind)
1954 ss->bind(css);
1955 } while_each_subsys_mask();
1956
1957 kernfs_activate(kn: dcgrp->kn);
1958 return 0;
1959}
1960
1961int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1962 struct kernfs_root *kf_root)
1963{
1964 int len = 0;
1965 char *buf = NULL;
1966 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1967 struct cgroup *ns_cgroup;
1968
1969 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1970 if (!buf)
1971 return -ENOMEM;
1972
1973 spin_lock_irq(lock: &css_set_lock);
1974 ns_cgroup = current_cgns_cgroup_from_root(root: kf_cgroot);
1975 len = kernfs_path_from_node(kn_to: kf_node, kn_from: ns_cgroup->kn, buf, PATH_MAX);
1976 spin_unlock_irq(lock: &css_set_lock);
1977
1978 if (len == -E2BIG)
1979 len = -ERANGE;
1980 else if (len > 0) {
1981 seq_escape(m: sf, s: buf, esc: " \t\n\\");
1982 len = 0;
1983 }
1984 kfree(objp: buf);
1985 return len;
1986}
1987
1988enum cgroup2_param {
1989 Opt_nsdelegate,
1990 Opt_favordynmods,
1991 Opt_memory_localevents,
1992 Opt_memory_recursiveprot,
1993 Opt_memory_hugetlb_accounting,
1994 Opt_pids_localevents,
1995 nr__cgroup2_params
1996};
1997
1998static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1999 fsparam_flag("nsdelegate", Opt_nsdelegate),
2000 fsparam_flag("favordynmods", Opt_favordynmods),
2001 fsparam_flag("memory_localevents", Opt_memory_localevents),
2002 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
2003 fsparam_flag("memory_hugetlb_accounting", Opt_memory_hugetlb_accounting),
2004 fsparam_flag("pids_localevents", Opt_pids_localevents),
2005 {}
2006};
2007
2008static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
2009{
2010 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2011 struct fs_parse_result result;
2012 int opt;
2013
2014 opt = fs_parse(fc, desc: cgroup2_fs_parameters, param, result: &result);
2015 if (opt < 0)
2016 return opt;
2017
2018 switch (opt) {
2019 case Opt_nsdelegate:
2020 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
2021 return 0;
2022 case Opt_favordynmods:
2023 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2024 return 0;
2025 case Opt_memory_localevents:
2026 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
2027 return 0;
2028 case Opt_memory_recursiveprot:
2029 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2030 return 0;
2031 case Opt_memory_hugetlb_accounting:
2032 ctx->flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2033 return 0;
2034 case Opt_pids_localevents:
2035 ctx->flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS;
2036 return 0;
2037 }
2038 return -EINVAL;
2039}
2040
2041struct cgroup_of_peak *of_peak(struct kernfs_open_file *of)
2042{
2043 struct cgroup_file_ctx *ctx = of->priv;
2044
2045 return &ctx->peak;
2046}
2047
2048static void apply_cgroup_root_flags(unsigned int root_flags)
2049{
2050 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
2051 if (root_flags & CGRP_ROOT_NS_DELEGATE)
2052 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
2053 else
2054 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
2055
2056 cgroup_favor_dynmods(root: &cgrp_dfl_root,
2057 favor: root_flags & CGRP_ROOT_FAVOR_DYNMODS);
2058
2059 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
2060 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
2061 else
2062 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
2063
2064 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
2065 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2066 else
2067 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2068
2069 if (root_flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2070 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2071 else
2072 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2073
2074 if (root_flags & CGRP_ROOT_PIDS_LOCAL_EVENTS)
2075 cgrp_dfl_root.flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS;
2076 else
2077 cgrp_dfl_root.flags &= ~CGRP_ROOT_PIDS_LOCAL_EVENTS;
2078 }
2079}
2080
2081static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
2082{
2083 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
2084 seq_puts(m: seq, s: ",nsdelegate");
2085 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
2086 seq_puts(m: seq, s: ",favordynmods");
2087 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
2088 seq_puts(m: seq, s: ",memory_localevents");
2089 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
2090 seq_puts(m: seq, s: ",memory_recursiveprot");
2091 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2092 seq_puts(m: seq, s: ",memory_hugetlb_accounting");
2093 if (cgrp_dfl_root.flags & CGRP_ROOT_PIDS_LOCAL_EVENTS)
2094 seq_puts(m: seq, s: ",pids_localevents");
2095 return 0;
2096}
2097
2098static int cgroup_reconfigure(struct fs_context *fc)
2099{
2100 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2101
2102 apply_cgroup_root_flags(root_flags: ctx->flags);
2103 return 0;
2104}
2105
2106static void init_cgroup_housekeeping(struct cgroup *cgrp)
2107{
2108 struct cgroup_subsys *ss;
2109 int ssid;
2110
2111 INIT_LIST_HEAD(list: &cgrp->self.sibling);
2112 INIT_LIST_HEAD(list: &cgrp->self.children);
2113 INIT_LIST_HEAD(list: &cgrp->cset_links);
2114 INIT_LIST_HEAD(list: &cgrp->pidlists);
2115 mutex_init(&cgrp->pidlist_mutex);
2116 cgrp->self.cgroup = cgrp;
2117 cgrp->self.flags |= CSS_ONLINE;
2118 cgrp->dom_cgrp = cgrp;
2119 cgrp->max_descendants = INT_MAX;
2120 cgrp->max_depth = INT_MAX;
2121 prev_cputime_init(prev: &cgrp->prev_cputime);
2122
2123 for_each_subsys(ss, ssid)
2124 INIT_LIST_HEAD(list: &cgrp->e_csets[ssid]);
2125
2126#ifdef CONFIG_CGROUP_BPF
2127 for (int i = 0; i < ARRAY_SIZE(cgrp->bpf.revisions); i++)
2128 cgrp->bpf.revisions[i] = 1;
2129#endif
2130
2131 init_waitqueue_head(&cgrp->offline_waitq);
2132 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2133}
2134
2135void init_cgroup_root(struct cgroup_fs_context *ctx)
2136{
2137 struct cgroup_root *root = ctx->root;
2138 struct cgroup *cgrp = &root->cgrp;
2139
2140 INIT_LIST_HEAD_RCU(list: &root->root_list);
2141 atomic_set(v: &root->nr_cgrps, i: 1);
2142 cgrp->root = root;
2143 init_cgroup_housekeeping(cgrp);
2144
2145 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2146 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2147 if (ctx->release_agent)
2148 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2149 if (ctx->name)
2150 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2151 if (ctx->cpuset_clone_children)
2152 set_bit(nr: CGRP_CPUSET_CLONE_CHILDREN, addr: &root->cgrp.flags);
2153}
2154
2155int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2156{
2157 LIST_HEAD(tmp_links);
2158 struct cgroup *root_cgrp = &root->cgrp;
2159 struct kernfs_syscall_ops *kf_sops;
2160 struct css_set *cset;
2161 int i, ret;
2162
2163 lockdep_assert_held(&cgroup_mutex);
2164
2165 ret = percpu_ref_init(ref: &root_cgrp->self.refcnt, release: css_release,
2166 flags: 0, GFP_KERNEL);
2167 if (ret)
2168 goto out;
2169
2170 /*
2171 * We're accessing css_set_count without locking css_set_lock here,
2172 * but that's OK - it can only be increased by someone holding
2173 * cgroup_lock, and that's us. Later rebinding may disable
2174 * controllers on the default hierarchy and thus create new csets,
2175 * which can't be more than the existing ones. Allocate 2x.
2176 */
2177 ret = allocate_cgrp_cset_links(count: 2 * css_set_count, tmp_links: &tmp_links);
2178 if (ret)
2179 goto cancel_ref;
2180
2181 ret = cgroup_init_root_id(root);
2182 if (ret)
2183 goto cancel_ref;
2184
2185 kf_sops = root == &cgrp_dfl_root ?
2186 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2187
2188 root->kf_root = kernfs_create_root(scops: kf_sops,
2189 flags: KERNFS_ROOT_CREATE_DEACTIVATED |
2190 KERNFS_ROOT_SUPPORT_EXPORTOP |
2191 KERNFS_ROOT_SUPPORT_USER_XATTR |
2192 KERNFS_ROOT_INVARIANT_PARENT,
2193 priv: root_cgrp);
2194 if (IS_ERR(ptr: root->kf_root)) {
2195 ret = PTR_ERR(ptr: root->kf_root);
2196 goto exit_root_id;
2197 }
2198 root_cgrp->kn = kernfs_root_to_node(root: root->kf_root);
2199 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2200 root_cgrp->ancestors[0] = root_cgrp;
2201
2202 ret = css_populate_dir(css: &root_cgrp->self);
2203 if (ret)
2204 goto destroy_root;
2205
2206 ret = css_rstat_init(css: &root_cgrp->self);
2207 if (ret)
2208 goto destroy_root;
2209
2210 ret = rebind_subsystems(dst_root: root, ss_mask);
2211 if (ret)
2212 goto exit_stats;
2213
2214 ret = blocking_notifier_call_chain(nh: &cgroup_lifetime_notifier,
2215 val: CGROUP_LIFETIME_ONLINE, v: root_cgrp);
2216 WARN_ON_ONCE(notifier_to_errno(ret));
2217
2218 trace_cgroup_setup_root(root);
2219
2220 /*
2221 * There must be no failure case after here, since rebinding takes
2222 * care of subsystems' refcounts, which are explicitly dropped in
2223 * the failure exit path.
2224 */
2225 list_add_rcu(new: &root->root_list, head: &cgroup_roots);
2226 cgroup_root_count++;
2227
2228 /*
2229 * Link the root cgroup in this hierarchy into all the css_set
2230 * objects.
2231 */
2232 spin_lock_irq(lock: &css_set_lock);
2233 hash_for_each(css_set_table, i, cset, hlist) {
2234 link_css_set(tmp_links: &tmp_links, cset, cgrp: root_cgrp);
2235 if (css_set_populated(cset))
2236 cgroup_update_populated(cgrp: root_cgrp, populated: true);
2237 }
2238 spin_unlock_irq(lock: &css_set_lock);
2239
2240 BUG_ON(!list_empty(&root_cgrp->self.children));
2241 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2242
2243 ret = 0;
2244 goto out;
2245
2246exit_stats:
2247 css_rstat_exit(css: &root_cgrp->self);
2248destroy_root:
2249 kernfs_destroy_root(root: root->kf_root);
2250 root->kf_root = NULL;
2251exit_root_id:
2252 cgroup_exit_root_id(root);
2253cancel_ref:
2254 percpu_ref_exit(ref: &root_cgrp->self.refcnt);
2255out:
2256 free_cgrp_cset_links(links_to_free: &tmp_links);
2257 return ret;
2258}
2259
2260int cgroup_do_get_tree(struct fs_context *fc)
2261{
2262 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2263 int ret;
2264
2265 ctx->kfc.root = ctx->root->kf_root;
2266 if (fc->fs_type == &cgroup2_fs_type)
2267 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2268 else
2269 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2270 ret = kernfs_get_tree(fc);
2271
2272 /*
2273 * In non-init cgroup namespace, instead of root cgroup's dentry,
2274 * we return the dentry corresponding to the cgroupns->root_cgrp.
2275 */
2276 if (!ret && ctx->ns != &init_cgroup_ns) {
2277 struct dentry *nsdentry;
2278 struct super_block *sb = fc->root->d_sb;
2279 struct cgroup *cgrp;
2280
2281 cgroup_lock();
2282 spin_lock_irq(lock: &css_set_lock);
2283
2284 cgrp = cset_cgroup_from_root(cset: ctx->ns->root_cset, root: ctx->root);
2285
2286 spin_unlock_irq(lock: &css_set_lock);
2287 cgroup_unlock();
2288
2289 nsdentry = kernfs_node_dentry(kn: cgrp->kn, sb);
2290 dput(fc->root);
2291 if (IS_ERR(ptr: nsdentry)) {
2292 deactivate_locked_super(sb);
2293 ret = PTR_ERR(ptr: nsdentry);
2294 nsdentry = NULL;
2295 }
2296 fc->root = nsdentry;
2297 }
2298
2299 if (!ctx->kfc.new_sb_created)
2300 cgroup_put(cgrp: &ctx->root->cgrp);
2301
2302 return ret;
2303}
2304
2305/*
2306 * Destroy a cgroup filesystem context.
2307 */
2308static void cgroup_fs_context_free(struct fs_context *fc)
2309{
2310 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2311
2312 kfree(objp: ctx->name);
2313 kfree(objp: ctx->release_agent);
2314 put_cgroup_ns(ns: ctx->ns);
2315 kernfs_free_fs_context(fc);
2316 kfree(objp: ctx);
2317}
2318
2319static int cgroup_get_tree(struct fs_context *fc)
2320{
2321 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2322 int ret;
2323
2324 WRITE_ONCE(cgrp_dfl_visible, true);
2325 cgroup_get_live(cgrp: &cgrp_dfl_root.cgrp);
2326 ctx->root = &cgrp_dfl_root;
2327
2328 ret = cgroup_do_get_tree(fc);
2329 if (!ret)
2330 apply_cgroup_root_flags(root_flags: ctx->flags);
2331 return ret;
2332}
2333
2334static const struct fs_context_operations cgroup_fs_context_ops = {
2335 .free = cgroup_fs_context_free,
2336 .parse_param = cgroup2_parse_param,
2337 .get_tree = cgroup_get_tree,
2338 .reconfigure = cgroup_reconfigure,
2339};
2340
2341static const struct fs_context_operations cgroup1_fs_context_ops = {
2342 .free = cgroup_fs_context_free,
2343 .parse_param = cgroup1_parse_param,
2344 .get_tree = cgroup1_get_tree,
2345 .reconfigure = cgroup1_reconfigure,
2346};
2347
2348/*
2349 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2350 * we select the namespace we're going to use.
2351 */
2352static int cgroup_init_fs_context(struct fs_context *fc)
2353{
2354 struct cgroup_fs_context *ctx;
2355
2356 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2357 if (!ctx)
2358 return -ENOMEM;
2359
2360 ctx->ns = current->nsproxy->cgroup_ns;
2361 get_cgroup_ns(ns: ctx->ns);
2362 fc->fs_private = &ctx->kfc;
2363 if (fc->fs_type == &cgroup2_fs_type)
2364 fc->ops = &cgroup_fs_context_ops;
2365 else
2366 fc->ops = &cgroup1_fs_context_ops;
2367 put_user_ns(ns: fc->user_ns);
2368 fc->user_ns = get_user_ns(ns: ctx->ns->user_ns);
2369 fc->global = true;
2370
2371 if (have_favordynmods)
2372 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2373
2374 return 0;
2375}
2376
2377static void cgroup_kill_sb(struct super_block *sb)
2378{
2379 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2380 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2381
2382 /*
2383 * If @root doesn't have any children, start killing it.
2384 * This prevents new mounts by disabling percpu_ref_tryget_live().
2385 *
2386 * And don't kill the default root.
2387 */
2388 if (list_empty(head: &root->cgrp.self.children) && root != &cgrp_dfl_root &&
2389 !percpu_ref_is_dying(ref: &root->cgrp.self.refcnt))
2390 percpu_ref_kill(ref: &root->cgrp.self.refcnt);
2391 cgroup_put(cgrp: &root->cgrp);
2392 kernfs_kill_sb(sb);
2393}
2394
2395struct file_system_type cgroup_fs_type = {
2396 .name = "cgroup",
2397 .init_fs_context = cgroup_init_fs_context,
2398 .parameters = cgroup1_fs_parameters,
2399 .kill_sb = cgroup_kill_sb,
2400 .fs_flags = FS_USERNS_MOUNT,
2401};
2402
2403static struct file_system_type cgroup2_fs_type = {
2404 .name = "cgroup2",
2405 .init_fs_context = cgroup_init_fs_context,
2406 .parameters = cgroup2_fs_parameters,
2407 .kill_sb = cgroup_kill_sb,
2408 .fs_flags = FS_USERNS_MOUNT,
2409};
2410
2411#ifdef CONFIG_CPUSETS_V1
2412enum cpuset_param {
2413 Opt_cpuset_v2_mode,
2414};
2415
2416static const struct fs_parameter_spec cpuset_fs_parameters[] = {
2417 fsparam_flag ("cpuset_v2_mode", Opt_cpuset_v2_mode),
2418 {}
2419};
2420
2421static int cpuset_parse_param(struct fs_context *fc, struct fs_parameter *param)
2422{
2423 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2424 struct fs_parse_result result;
2425 int opt;
2426
2427 opt = fs_parse(fc, cpuset_fs_parameters, param, &result);
2428 if (opt < 0)
2429 return opt;
2430
2431 switch (opt) {
2432 case Opt_cpuset_v2_mode:
2433 ctx->flags |= CGRP_ROOT_CPUSET_V2_MODE;
2434 return 0;
2435 }
2436 return -EINVAL;
2437}
2438
2439static const struct fs_context_operations cpuset_fs_context_ops = {
2440 .get_tree = cgroup1_get_tree,
2441 .free = cgroup_fs_context_free,
2442 .parse_param = cpuset_parse_param,
2443};
2444
2445/*
2446 * This is ugly, but preserves the userspace API for existing cpuset
2447 * users. If someone tries to mount the "cpuset" filesystem, we
2448 * silently switch it to mount "cgroup" instead
2449 */
2450static int cpuset_init_fs_context(struct fs_context *fc)
2451{
2452 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2453 struct cgroup_fs_context *ctx;
2454 int err;
2455
2456 err = cgroup_init_fs_context(fc);
2457 if (err) {
2458 kfree(agent);
2459 return err;
2460 }
2461
2462 fc->ops = &cpuset_fs_context_ops;
2463
2464 ctx = cgroup_fc2context(fc);
2465 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2466 ctx->flags |= CGRP_ROOT_NOPREFIX;
2467 ctx->release_agent = agent;
2468
2469 get_filesystem(&cgroup_fs_type);
2470 put_filesystem(fc->fs_type);
2471 fc->fs_type = &cgroup_fs_type;
2472
2473 return 0;
2474}
2475
2476static struct file_system_type cpuset_fs_type = {
2477 .name = "cpuset",
2478 .init_fs_context = cpuset_init_fs_context,
2479 .parameters = cpuset_fs_parameters,
2480 .fs_flags = FS_USERNS_MOUNT,
2481};
2482#endif
2483
2484int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2485 struct cgroup_namespace *ns)
2486{
2487 struct cgroup *root = cset_cgroup_from_root(cset: ns->root_cset, root: cgrp->root);
2488
2489 return kernfs_path_from_node(kn_to: cgrp->kn, kn_from: root->kn, buf, buflen);
2490}
2491
2492int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2493 struct cgroup_namespace *ns)
2494{
2495 int ret;
2496
2497 cgroup_lock();
2498 spin_lock_irq(lock: &css_set_lock);
2499
2500 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2501
2502 spin_unlock_irq(lock: &css_set_lock);
2503 cgroup_unlock();
2504
2505 return ret;
2506}
2507EXPORT_SYMBOL_GPL(cgroup_path_ns);
2508
2509/**
2510 * cgroup_attach_lock - Lock for ->attach()
2511 * @lock_mode: whether acquire and acquire which rwsem
2512 * @tsk: thread group to lock
2513 *
2514 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2515 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2516 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2517 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2518 * lead to deadlocks.
2519 *
2520 * Bringing up a CPU may involve creating and destroying tasks which requires
2521 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2522 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2523 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2524 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2525 * the threadgroup_rwsem to be released to create new tasks. For more details:
2526 *
2527 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2528 *
2529 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2530 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2531 * CPU hotplug is disabled on entry.
2532 *
2533 * When favordynmods is enabled, take per threadgroup rwsem to reduce overhead
2534 * on dynamic cgroup modifications. see the comment above
2535 * CGRP_ROOT_FAVOR_DYNMODS definition.
2536 *
2537 * tsk is not NULL only when writing to cgroup.procs.
2538 */
2539void cgroup_attach_lock(enum cgroup_attach_lock_mode lock_mode,
2540 struct task_struct *tsk)
2541{
2542 cpus_read_lock();
2543
2544 switch (lock_mode) {
2545 case CGRP_ATTACH_LOCK_NONE:
2546 break;
2547 case CGRP_ATTACH_LOCK_GLOBAL:
2548 percpu_down_write(&cgroup_threadgroup_rwsem);
2549 break;
2550 case CGRP_ATTACH_LOCK_PER_THREADGROUP:
2551 down_write(sem: &tsk->signal->cgroup_threadgroup_rwsem);
2552 break;
2553 default:
2554 pr_warn("cgroup: Unexpected attach lock mode.");
2555 break;
2556 }
2557}
2558
2559/**
2560 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2561 * @lock_mode: whether release and release which rwsem
2562 * @tsk: thread group to lock
2563 */
2564void cgroup_attach_unlock(enum cgroup_attach_lock_mode lock_mode,
2565 struct task_struct *tsk)
2566{
2567 switch (lock_mode) {
2568 case CGRP_ATTACH_LOCK_NONE:
2569 break;
2570 case CGRP_ATTACH_LOCK_GLOBAL:
2571 percpu_up_write(&cgroup_threadgroup_rwsem);
2572 break;
2573 case CGRP_ATTACH_LOCK_PER_THREADGROUP:
2574 up_write(sem: &tsk->signal->cgroup_threadgroup_rwsem);
2575 break;
2576 default:
2577 pr_warn("cgroup: Unexpected attach lock mode.");
2578 break;
2579 }
2580
2581 cpus_read_unlock();
2582}
2583
2584/**
2585 * cgroup_migrate_add_task - add a migration target task to a migration context
2586 * @task: target task
2587 * @mgctx: target migration context
2588 *
2589 * Add @task, which is a migration target, to @mgctx->tset. This function
2590 * becomes noop if @task doesn't need to be migrated. @task's css_set
2591 * should have been added as a migration source and @task->cg_list will be
2592 * moved from the css_set's tasks list to mg_tasks one.
2593 */
2594static void cgroup_migrate_add_task(struct task_struct *task,
2595 struct cgroup_mgctx *mgctx)
2596{
2597 struct css_set *cset;
2598
2599 lockdep_assert_held(&css_set_lock);
2600
2601 /* @task either already exited or can't exit until the end */
2602 if (task->flags & PF_EXITING)
2603 return;
2604
2605 /* cgroup_threadgroup_rwsem protects racing against forks */
2606 WARN_ON_ONCE(list_empty(&task->cg_list));
2607
2608 cset = task_css_set(task);
2609 if (!cset->mg_src_cgrp)
2610 return;
2611
2612 mgctx->tset.nr_tasks++;
2613
2614 list_move_tail(list: &task->cg_list, head: &cset->mg_tasks);
2615 if (list_empty(head: &cset->mg_node))
2616 list_add_tail(new: &cset->mg_node,
2617 head: &mgctx->tset.src_csets);
2618 if (list_empty(head: &cset->mg_dst_cset->mg_node))
2619 list_add_tail(new: &cset->mg_dst_cset->mg_node,
2620 head: &mgctx->tset.dst_csets);
2621}
2622
2623/**
2624 * cgroup_taskset_first - reset taskset and return the first task
2625 * @tset: taskset of interest
2626 * @dst_cssp: output variable for the destination css
2627 *
2628 * @tset iteration is initialized and the first task is returned.
2629 */
2630struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2631 struct cgroup_subsys_state **dst_cssp)
2632{
2633 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2634 tset->cur_task = NULL;
2635
2636 return cgroup_taskset_next(tset, dst_cssp);
2637}
2638
2639/**
2640 * cgroup_taskset_next - iterate to the next task in taskset
2641 * @tset: taskset of interest
2642 * @dst_cssp: output variable for the destination css
2643 *
2644 * Return the next task in @tset. Iteration must have been initialized
2645 * with cgroup_taskset_first().
2646 */
2647struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2648 struct cgroup_subsys_state **dst_cssp)
2649{
2650 struct css_set *cset = tset->cur_cset;
2651 struct task_struct *task = tset->cur_task;
2652
2653 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2654 if (!task)
2655 task = list_first_entry(&cset->mg_tasks,
2656 struct task_struct, cg_list);
2657 else
2658 task = list_next_entry(task, cg_list);
2659
2660 if (&task->cg_list != &cset->mg_tasks) {
2661 tset->cur_cset = cset;
2662 tset->cur_task = task;
2663
2664 /*
2665 * This function may be called both before and
2666 * after cgroup_migrate_execute(). The two cases
2667 * can be distinguished by looking at whether @cset
2668 * has its ->mg_dst_cset set.
2669 */
2670 if (cset->mg_dst_cset)
2671 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2672 else
2673 *dst_cssp = cset->subsys[tset->ssid];
2674
2675 return task;
2676 }
2677
2678 cset = list_next_entry(cset, mg_node);
2679 task = NULL;
2680 }
2681
2682 return NULL;
2683}
2684
2685/**
2686 * cgroup_migrate_execute - migrate a taskset
2687 * @mgctx: migration context
2688 *
2689 * Migrate tasks in @mgctx as setup by migration preparation functions.
2690 * This function fails iff one of the ->can_attach callbacks fails and
2691 * guarantees that either all or none of the tasks in @mgctx are migrated.
2692 * @mgctx is consumed regardless of success.
2693 */
2694static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2695{
2696 struct cgroup_taskset *tset = &mgctx->tset;
2697 struct cgroup_subsys *ss;
2698 struct task_struct *task, *tmp_task;
2699 struct css_set *cset, *tmp_cset;
2700 int ssid, failed_ssid, ret;
2701
2702 /* check that we can legitimately attach to the cgroup */
2703 if (tset->nr_tasks) {
2704 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2705 if (ss->can_attach) {
2706 tset->ssid = ssid;
2707 ret = ss->can_attach(tset);
2708 if (ret) {
2709 failed_ssid = ssid;
2710 goto out_cancel_attach;
2711 }
2712 }
2713 } while_each_subsys_mask();
2714 }
2715
2716 /*
2717 * Now that we're guaranteed success, proceed to move all tasks to
2718 * the new cgroup. There are no failure cases after here, so this
2719 * is the commit point.
2720 */
2721 spin_lock_irq(lock: &css_set_lock);
2722 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2723 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2724 struct css_set *from_cset = task_css_set(task);
2725 struct css_set *to_cset = cset->mg_dst_cset;
2726
2727 get_css_set(cset: to_cset);
2728 to_cset->nr_tasks++;
2729 css_set_move_task(task, from_cset, to_cset, use_mg_tasks: true);
2730 from_cset->nr_tasks--;
2731 /*
2732 * If the source or destination cgroup is frozen,
2733 * the task might require to change its state.
2734 */
2735 cgroup_freezer_migrate_task(task, src: from_cset->dfl_cgrp,
2736 dst: to_cset->dfl_cgrp);
2737 put_css_set_locked(cset: from_cset);
2738
2739 }
2740 }
2741 spin_unlock_irq(lock: &css_set_lock);
2742
2743 /*
2744 * Migration is committed, all target tasks are now on dst_csets.
2745 * Nothing is sensitive to fork() after this point. Notify
2746 * controllers that migration is complete.
2747 */
2748 tset->csets = &tset->dst_csets;
2749
2750 if (tset->nr_tasks) {
2751 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2752 if (ss->attach) {
2753 tset->ssid = ssid;
2754 ss->attach(tset);
2755 }
2756 } while_each_subsys_mask();
2757 }
2758
2759 ret = 0;
2760 goto out_release_tset;
2761
2762out_cancel_attach:
2763 if (tset->nr_tasks) {
2764 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2765 if (ssid == failed_ssid)
2766 break;
2767 if (ss->cancel_attach) {
2768 tset->ssid = ssid;
2769 ss->cancel_attach(tset);
2770 }
2771 } while_each_subsys_mask();
2772 }
2773out_release_tset:
2774 spin_lock_irq(lock: &css_set_lock);
2775 list_splice_init(list: &tset->dst_csets, head: &tset->src_csets);
2776 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2777 list_splice_tail_init(list: &cset->mg_tasks, head: &cset->tasks);
2778 list_del_init(entry: &cset->mg_node);
2779 }
2780 spin_unlock_irq(lock: &css_set_lock);
2781
2782 /*
2783 * Re-initialize the cgroup_taskset structure in case it is reused
2784 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2785 * iteration.
2786 */
2787 tset->nr_tasks = 0;
2788 tset->csets = &tset->src_csets;
2789 return ret;
2790}
2791
2792/**
2793 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2794 * @dst_cgrp: destination cgroup to test
2795 *
2796 * On the default hierarchy, except for the mixable, (possible) thread root
2797 * and threaded cgroups, subtree_control must be zero for migration
2798 * destination cgroups with tasks so that child cgroups don't compete
2799 * against tasks.
2800 */
2801int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2802{
2803 /* v1 doesn't have any restriction */
2804 if (!cgroup_on_dfl(cgrp: dst_cgrp))
2805 return 0;
2806
2807 /* verify @dst_cgrp can host resources */
2808 if (!cgroup_is_valid_domain(cgrp: dst_cgrp->dom_cgrp))
2809 return -EOPNOTSUPP;
2810
2811 /*
2812 * If @dst_cgrp is already or can become a thread root or is
2813 * threaded, it doesn't matter.
2814 */
2815 if (cgroup_can_be_thread_root(cgrp: dst_cgrp) || cgroup_is_threaded(cgrp: dst_cgrp))
2816 return 0;
2817
2818 /* apply no-internal-process constraint */
2819 if (dst_cgrp->subtree_control)
2820 return -EBUSY;
2821
2822 return 0;
2823}
2824
2825/**
2826 * cgroup_migrate_finish - cleanup after attach
2827 * @mgctx: migration context
2828 *
2829 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2830 * those functions for details.
2831 */
2832void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2833{
2834 struct css_set *cset, *tmp_cset;
2835
2836 lockdep_assert_held(&cgroup_mutex);
2837
2838 spin_lock_irq(lock: &css_set_lock);
2839
2840 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2841 mg_src_preload_node) {
2842 cset->mg_src_cgrp = NULL;
2843 cset->mg_dst_cgrp = NULL;
2844 cset->mg_dst_cset = NULL;
2845 list_del_init(entry: &cset->mg_src_preload_node);
2846 put_css_set_locked(cset);
2847 }
2848
2849 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2850 mg_dst_preload_node) {
2851 cset->mg_src_cgrp = NULL;
2852 cset->mg_dst_cgrp = NULL;
2853 cset->mg_dst_cset = NULL;
2854 list_del_init(entry: &cset->mg_dst_preload_node);
2855 put_css_set_locked(cset);
2856 }
2857
2858 spin_unlock_irq(lock: &css_set_lock);
2859}
2860
2861/**
2862 * cgroup_migrate_add_src - add a migration source css_set
2863 * @src_cset: the source css_set to add
2864 * @dst_cgrp: the destination cgroup
2865 * @mgctx: migration context
2866 *
2867 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2868 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2869 * up by cgroup_migrate_finish().
2870 *
2871 * This function may be called without holding cgroup_threadgroup_rwsem
2872 * even if the target is a process. Threads may be created and destroyed
2873 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2874 * into play and the preloaded css_sets are guaranteed to cover all
2875 * migrations.
2876 */
2877void cgroup_migrate_add_src(struct css_set *src_cset,
2878 struct cgroup *dst_cgrp,
2879 struct cgroup_mgctx *mgctx)
2880{
2881 struct cgroup *src_cgrp;
2882
2883 lockdep_assert_held(&cgroup_mutex);
2884 lockdep_assert_held(&css_set_lock);
2885
2886 /*
2887 * If ->dead, @src_set is associated with one or more dead cgroups
2888 * and doesn't contain any migratable tasks. Ignore it early so
2889 * that the rest of migration path doesn't get confused by it.
2890 */
2891 if (src_cset->dead)
2892 return;
2893
2894 if (!list_empty(head: &src_cset->mg_src_preload_node))
2895 return;
2896
2897 src_cgrp = cset_cgroup_from_root(cset: src_cset, root: dst_cgrp->root);
2898
2899 WARN_ON(src_cset->mg_src_cgrp);
2900 WARN_ON(src_cset->mg_dst_cgrp);
2901 WARN_ON(!list_empty(&src_cset->mg_tasks));
2902 WARN_ON(!list_empty(&src_cset->mg_node));
2903
2904 src_cset->mg_src_cgrp = src_cgrp;
2905 src_cset->mg_dst_cgrp = dst_cgrp;
2906 get_css_set(cset: src_cset);
2907 list_add_tail(new: &src_cset->mg_src_preload_node, head: &mgctx->preloaded_src_csets);
2908}
2909
2910/**
2911 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2912 * @mgctx: migration context
2913 *
2914 * Tasks are about to be moved and all the source css_sets have been
2915 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2916 * pins all destination css_sets, links each to its source, and append them
2917 * to @mgctx->preloaded_dst_csets.
2918 *
2919 * This function must be called after cgroup_migrate_add_src() has been
2920 * called on each migration source css_set. After migration is performed
2921 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2922 * @mgctx.
2923 */
2924int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2925{
2926 struct css_set *src_cset, *tmp_cset;
2927
2928 lockdep_assert_held(&cgroup_mutex);
2929
2930 /* look up the dst cset for each src cset and link it to src */
2931 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2932 mg_src_preload_node) {
2933 struct css_set *dst_cset;
2934 struct cgroup_subsys *ss;
2935 int ssid;
2936
2937 dst_cset = find_css_set(old_cset: src_cset, cgrp: src_cset->mg_dst_cgrp);
2938 if (!dst_cset)
2939 return -ENOMEM;
2940
2941 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2942
2943 /*
2944 * If src cset equals dst, it's noop. Drop the src.
2945 * cgroup_migrate() will skip the cset too. Note that we
2946 * can't handle src == dst as some nodes are used by both.
2947 */
2948 if (src_cset == dst_cset) {
2949 src_cset->mg_src_cgrp = NULL;
2950 src_cset->mg_dst_cgrp = NULL;
2951 list_del_init(entry: &src_cset->mg_src_preload_node);
2952 put_css_set(cset: src_cset);
2953 put_css_set(cset: dst_cset);
2954 continue;
2955 }
2956
2957 src_cset->mg_dst_cset = dst_cset;
2958
2959 if (list_empty(head: &dst_cset->mg_dst_preload_node))
2960 list_add_tail(new: &dst_cset->mg_dst_preload_node,
2961 head: &mgctx->preloaded_dst_csets);
2962 else
2963 put_css_set(cset: dst_cset);
2964
2965 for_each_subsys(ss, ssid)
2966 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2967 mgctx->ss_mask |= 1 << ssid;
2968 }
2969
2970 return 0;
2971}
2972
2973/**
2974 * cgroup_migrate - migrate a process or task to a cgroup
2975 * @leader: the leader of the process or the task to migrate
2976 * @threadgroup: whether @leader points to the whole process or a single task
2977 * @mgctx: migration context
2978 *
2979 * Migrate a process or task denoted by @leader. If migrating a process,
2980 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2981 * responsible for invoking cgroup_migrate_add_src() and
2982 * cgroup_migrate_prepare_dst() on the targets before invoking this
2983 * function and following up with cgroup_migrate_finish().
2984 *
2985 * As long as a controller's ->can_attach() doesn't fail, this function is
2986 * guaranteed to succeed. This means that, excluding ->can_attach()
2987 * failure, when migrating multiple targets, the success or failure can be
2988 * decided for all targets by invoking group_migrate_prepare_dst() before
2989 * actually starting migrating.
2990 */
2991int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2992 struct cgroup_mgctx *mgctx)
2993{
2994 struct task_struct *task;
2995
2996 /*
2997 * The following thread iteration should be inside an RCU critical
2998 * section to prevent tasks from being freed while taking the snapshot.
2999 * spin_lock_irq() implies RCU critical section here.
3000 */
3001 spin_lock_irq(lock: &css_set_lock);
3002 task = leader;
3003 do {
3004 cgroup_migrate_add_task(task, mgctx);
3005 if (!threadgroup)
3006 break;
3007 } while_each_thread(leader, task);
3008 spin_unlock_irq(lock: &css_set_lock);
3009
3010 return cgroup_migrate_execute(mgctx);
3011}
3012
3013/**
3014 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
3015 * @dst_cgrp: the cgroup to attach to
3016 * @leader: the task or the leader of the threadgroup to be attached
3017 * @threadgroup: attach the whole threadgroup?
3018 *
3019 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
3020 */
3021int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
3022 bool threadgroup)
3023{
3024 DEFINE_CGROUP_MGCTX(mgctx);
3025 struct task_struct *task;
3026 int ret = 0;
3027
3028 /* look up all src csets */
3029 spin_lock_irq(lock: &css_set_lock);
3030 task = leader;
3031 do {
3032 cgroup_migrate_add_src(src_cset: task_css_set(task), dst_cgrp, mgctx: &mgctx);
3033 if (!threadgroup)
3034 break;
3035 } while_each_thread(leader, task);
3036 spin_unlock_irq(lock: &css_set_lock);
3037
3038 /* prepare dst csets and commit */
3039 ret = cgroup_migrate_prepare_dst(mgctx: &mgctx);
3040 if (!ret)
3041 ret = cgroup_migrate(leader, threadgroup, mgctx: &mgctx);
3042
3043 cgroup_migrate_finish(mgctx: &mgctx);
3044
3045 if (!ret)
3046 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
3047
3048 return ret;
3049}
3050
3051struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
3052 enum cgroup_attach_lock_mode *lock_mode)
3053{
3054 struct task_struct *tsk;
3055 pid_t pid;
3056
3057 if (kstrtoint(s: strstrip(str: buf), base: 0, res: &pid) || pid < 0)
3058 return ERR_PTR(error: -EINVAL);
3059
3060retry_find_task:
3061 rcu_read_lock();
3062 if (pid) {
3063 tsk = find_task_by_vpid(nr: pid);
3064 if (!tsk) {
3065 tsk = ERR_PTR(error: -ESRCH);
3066 goto out_unlock_rcu;
3067 }
3068 } else {
3069 tsk = current;
3070 }
3071
3072 if (threadgroup)
3073 tsk = tsk->group_leader;
3074
3075 /*
3076 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
3077 * If userland migrates such a kthread to a non-root cgroup, it can
3078 * become trapped in a cpuset, or RT kthread may be born in a
3079 * cgroup with no rt_runtime allocated. Just say no.
3080 */
3081 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
3082 tsk = ERR_PTR(error: -EINVAL);
3083 goto out_unlock_rcu;
3084 }
3085 get_task_struct(t: tsk);
3086 rcu_read_unlock();
3087
3088 /*
3089 * If we migrate a single thread, we don't care about threadgroup
3090 * stability. If the thread is `current`, it won't exit(2) under our
3091 * hands or change PID through exec(2). We exclude
3092 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write callers
3093 * by cgroup_mutex. Therefore, we can skip the global lock.
3094 */
3095 lockdep_assert_held(&cgroup_mutex);
3096
3097 if (pid || threadgroup) {
3098 if (cgroup_enable_per_threadgroup_rwsem)
3099 *lock_mode = CGRP_ATTACH_LOCK_PER_THREADGROUP;
3100 else
3101 *lock_mode = CGRP_ATTACH_LOCK_GLOBAL;
3102 } else {
3103 *lock_mode = CGRP_ATTACH_LOCK_NONE;
3104 }
3105
3106 cgroup_attach_lock(lock_mode: *lock_mode, tsk);
3107
3108 if (threadgroup) {
3109 if (!thread_group_leader(p: tsk)) {
3110 /*
3111 * A race with de_thread from another thread's exec()
3112 * may strip us of our leadership. If this happens,
3113 * throw this task away and try again.
3114 */
3115 cgroup_attach_unlock(lock_mode: *lock_mode, tsk);
3116 put_task_struct(t: tsk);
3117 goto retry_find_task;
3118 }
3119 }
3120
3121 return tsk;
3122
3123out_unlock_rcu:
3124 rcu_read_unlock();
3125 return tsk;
3126}
3127
3128void cgroup_procs_write_finish(struct task_struct *task,
3129 enum cgroup_attach_lock_mode lock_mode)
3130{
3131 cgroup_attach_unlock(lock_mode, tsk: task);
3132
3133 /* release reference from cgroup_procs_write_start() */
3134 put_task_struct(t: task);
3135}
3136
3137static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
3138{
3139 struct cgroup_subsys *ss;
3140 bool printed = false;
3141 int ssid;
3142
3143 do_each_subsys_mask(ss, ssid, ss_mask) {
3144 if (printed)
3145 seq_putc(m: seq, c: ' ');
3146 seq_puts(m: seq, s: ss->name);
3147 printed = true;
3148 } while_each_subsys_mask();
3149 if (printed)
3150 seq_putc(m: seq, c: '\n');
3151}
3152
3153/* show controllers which are enabled from the parent */
3154static int cgroup_controllers_show(struct seq_file *seq, void *v)
3155{
3156 struct cgroup *cgrp = seq_css(seq)->cgroup;
3157
3158 cgroup_print_ss_mask(seq, ss_mask: cgroup_control(cgrp));
3159 return 0;
3160}
3161
3162/* show controllers which are enabled for a given cgroup's children */
3163static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3164{
3165 struct cgroup *cgrp = seq_css(seq)->cgroup;
3166
3167 cgroup_print_ss_mask(seq, ss_mask: cgrp->subtree_control);
3168 return 0;
3169}
3170
3171/**
3172 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3173 * @cgrp: root of the subtree to update csses for
3174 *
3175 * @cgrp's control masks have changed and its subtree's css associations
3176 * need to be updated accordingly. This function looks up all css_sets
3177 * which are attached to the subtree, creates the matching updated css_sets
3178 * and migrates the tasks to the new ones.
3179 */
3180static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3181{
3182 DEFINE_CGROUP_MGCTX(mgctx);
3183 struct cgroup_subsys_state *d_css;
3184 struct cgroup *dsct;
3185 struct css_set *src_cset;
3186 enum cgroup_attach_lock_mode lock_mode;
3187 bool has_tasks;
3188 int ret;
3189
3190 lockdep_assert_held(&cgroup_mutex);
3191
3192 /* look up all csses currently attached to @cgrp's subtree */
3193 spin_lock_irq(lock: &css_set_lock);
3194 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3195 struct cgrp_cset_link *link;
3196
3197 /*
3198 * As cgroup_update_dfl_csses() is only called by
3199 * cgroup_apply_control(). The csses associated with the
3200 * given cgrp will not be affected by changes made to
3201 * its subtree_control file. We can skip them.
3202 */
3203 if (dsct == cgrp)
3204 continue;
3205
3206 list_for_each_entry(link, &dsct->cset_links, cset_link)
3207 cgroup_migrate_add_src(src_cset: link->cset, dst_cgrp: dsct, mgctx: &mgctx);
3208 }
3209 spin_unlock_irq(lock: &css_set_lock);
3210
3211 /*
3212 * We need to write-lock threadgroup_rwsem while migrating tasks.
3213 * However, if there are no source csets for @cgrp, changing its
3214 * controllers isn't gonna produce any task migrations and the
3215 * write-locking can be skipped safely.
3216 */
3217 has_tasks = !list_empty(head: &mgctx.preloaded_src_csets);
3218
3219 if (has_tasks)
3220 lock_mode = CGRP_ATTACH_LOCK_GLOBAL;
3221 else
3222 lock_mode = CGRP_ATTACH_LOCK_NONE;
3223
3224 cgroup_attach_lock(lock_mode, NULL);
3225
3226 /* NULL dst indicates self on default hierarchy */
3227 ret = cgroup_migrate_prepare_dst(mgctx: &mgctx);
3228 if (ret)
3229 goto out_finish;
3230
3231 spin_lock_irq(lock: &css_set_lock);
3232 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3233 mg_src_preload_node) {
3234 struct task_struct *task, *ntask;
3235
3236 /* all tasks in src_csets need to be migrated */
3237 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3238 cgroup_migrate_add_task(task, mgctx: &mgctx);
3239 }
3240 spin_unlock_irq(lock: &css_set_lock);
3241
3242 ret = cgroup_migrate_execute(mgctx: &mgctx);
3243out_finish:
3244 cgroup_migrate_finish(mgctx: &mgctx);
3245 cgroup_attach_unlock(lock_mode, NULL);
3246 return ret;
3247}
3248
3249/**
3250 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3251 * @cgrp: root of the target subtree
3252 *
3253 * Because css offlining is asynchronous, userland may try to re-enable a
3254 * controller while the previous css is still around. This function grabs
3255 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3256 */
3257void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3258 __acquires(&cgroup_mutex)
3259{
3260 struct cgroup *dsct;
3261 struct cgroup_subsys_state *d_css;
3262 struct cgroup_subsys *ss;
3263 int ssid;
3264
3265restart:
3266 cgroup_lock();
3267
3268 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3269 for_each_subsys(ss, ssid) {
3270 struct cgroup_subsys_state *css = cgroup_css(cgrp: dsct, ss);
3271 DEFINE_WAIT(wait);
3272
3273 if (!css || !percpu_ref_is_dying(ref: &css->refcnt))
3274 continue;
3275
3276 cgroup_get_live(cgrp: dsct);
3277 prepare_to_wait(wq_head: &dsct->offline_waitq, wq_entry: &wait,
3278 TASK_UNINTERRUPTIBLE);
3279
3280 cgroup_unlock();
3281 schedule();
3282 finish_wait(wq_head: &dsct->offline_waitq, wq_entry: &wait);
3283
3284 cgroup_put(cgrp: dsct);
3285 goto restart;
3286 }
3287 }
3288}
3289
3290/**
3291 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3292 * @cgrp: root of the target subtree
3293 *
3294 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3295 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3296 * itself.
3297 */
3298static void cgroup_save_control(struct cgroup *cgrp)
3299{
3300 struct cgroup *dsct;
3301 struct cgroup_subsys_state *d_css;
3302
3303 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3304 dsct->old_subtree_control = dsct->subtree_control;
3305 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3306 dsct->old_dom_cgrp = dsct->dom_cgrp;
3307 }
3308}
3309
3310/**
3311 * cgroup_propagate_control - refresh control masks of a subtree
3312 * @cgrp: root of the target subtree
3313 *
3314 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3315 * ->subtree_control and propagate controller availability through the
3316 * subtree so that descendants don't have unavailable controllers enabled.
3317 */
3318static void cgroup_propagate_control(struct cgroup *cgrp)
3319{
3320 struct cgroup *dsct;
3321 struct cgroup_subsys_state *d_css;
3322
3323 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3324 dsct->subtree_control &= cgroup_control(cgrp: dsct);
3325 dsct->subtree_ss_mask =
3326 cgroup_calc_subtree_ss_mask(subtree_control: dsct->subtree_control,
3327 this_ss_mask: cgroup_ss_mask(cgrp: dsct));
3328 }
3329}
3330
3331/**
3332 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3333 * @cgrp: root of the target subtree
3334 *
3335 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3336 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3337 * itself.
3338 */
3339static void cgroup_restore_control(struct cgroup *cgrp)
3340{
3341 struct cgroup *dsct;
3342 struct cgroup_subsys_state *d_css;
3343
3344 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3345 dsct->subtree_control = dsct->old_subtree_control;
3346 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3347 dsct->dom_cgrp = dsct->old_dom_cgrp;
3348 }
3349}
3350
3351static bool css_visible(struct cgroup_subsys_state *css)
3352{
3353 struct cgroup_subsys *ss = css->ss;
3354 struct cgroup *cgrp = css->cgroup;
3355
3356 if (cgroup_control(cgrp) & (1 << ss->id))
3357 return true;
3358 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3359 return false;
3360 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3361}
3362
3363/**
3364 * cgroup_apply_control_enable - enable or show csses according to control
3365 * @cgrp: root of the target subtree
3366 *
3367 * Walk @cgrp's subtree and create new csses or make the existing ones
3368 * visible. A css is created invisible if it's being implicitly enabled
3369 * through dependency. An invisible css is made visible when the userland
3370 * explicitly enables it.
3371 *
3372 * Returns 0 on success, -errno on failure. On failure, csses which have
3373 * been processed already aren't cleaned up. The caller is responsible for
3374 * cleaning up with cgroup_apply_control_disable().
3375 */
3376static int cgroup_apply_control_enable(struct cgroup *cgrp)
3377{
3378 struct cgroup *dsct;
3379 struct cgroup_subsys_state *d_css;
3380 struct cgroup_subsys *ss;
3381 int ssid, ret;
3382
3383 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3384 for_each_subsys(ss, ssid) {
3385 struct cgroup_subsys_state *css = cgroup_css(cgrp: dsct, ss);
3386
3387 if (!(cgroup_ss_mask(cgrp: dsct) & (1 << ss->id)))
3388 continue;
3389
3390 if (!css) {
3391 css = css_create(cgrp: dsct, ss);
3392 if (IS_ERR(ptr: css))
3393 return PTR_ERR(ptr: css);
3394 }
3395
3396 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3397
3398 if (css_visible(css)) {
3399 ret = css_populate_dir(css);
3400 if (ret)
3401 return ret;
3402 }
3403 }
3404 }
3405
3406 return 0;
3407}
3408
3409/**
3410 * cgroup_apply_control_disable - kill or hide csses according to control
3411 * @cgrp: root of the target subtree
3412 *
3413 * Walk @cgrp's subtree and kill and hide csses so that they match
3414 * cgroup_ss_mask() and cgroup_visible_mask().
3415 *
3416 * A css is hidden when the userland requests it to be disabled while other
3417 * subsystems are still depending on it. The css must not actively control
3418 * resources and be in the vanilla state if it's made visible again later.
3419 * Controllers which may be depended upon should provide ->css_reset() for
3420 * this purpose.
3421 */
3422static void cgroup_apply_control_disable(struct cgroup *cgrp)
3423{
3424 struct cgroup *dsct;
3425 struct cgroup_subsys_state *d_css;
3426 struct cgroup_subsys *ss;
3427 int ssid;
3428
3429 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3430 for_each_subsys(ss, ssid) {
3431 struct cgroup_subsys_state *css = cgroup_css(cgrp: dsct, ss);
3432
3433 if (!css)
3434 continue;
3435
3436 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3437
3438 if (css->parent &&
3439 !(cgroup_ss_mask(cgrp: dsct) & (1 << ss->id))) {
3440 kill_css(css);
3441 } else if (!css_visible(css)) {
3442 css_clear_dir(css);
3443 if (ss->css_reset)
3444 ss->css_reset(css);
3445 }
3446 }
3447 }
3448}
3449
3450/**
3451 * cgroup_apply_control - apply control mask updates to the subtree
3452 * @cgrp: root of the target subtree
3453 *
3454 * subsystems can be enabled and disabled in a subtree using the following
3455 * steps.
3456 *
3457 * 1. Call cgroup_save_control() to stash the current state.
3458 * 2. Update ->subtree_control masks in the subtree as desired.
3459 * 3. Call cgroup_apply_control() to apply the changes.
3460 * 4. Optionally perform other related operations.
3461 * 5. Call cgroup_finalize_control() to finish up.
3462 *
3463 * This function implements step 3 and propagates the mask changes
3464 * throughout @cgrp's subtree, updates csses accordingly and perform
3465 * process migrations.
3466 */
3467static int cgroup_apply_control(struct cgroup *cgrp)
3468{
3469 int ret;
3470
3471 cgroup_propagate_control(cgrp);
3472
3473 ret = cgroup_apply_control_enable(cgrp);
3474 if (ret)
3475 return ret;
3476
3477 /*
3478 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3479 * making the following cgroup_update_dfl_csses() properly update
3480 * css associations of all tasks in the subtree.
3481 */
3482 return cgroup_update_dfl_csses(cgrp);
3483}
3484
3485/**
3486 * cgroup_finalize_control - finalize control mask update
3487 * @cgrp: root of the target subtree
3488 * @ret: the result of the update
3489 *
3490 * Finalize control mask update. See cgroup_apply_control() for more info.
3491 */
3492static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3493{
3494 if (ret) {
3495 cgroup_restore_control(cgrp);
3496 cgroup_propagate_control(cgrp);
3497 }
3498
3499 cgroup_apply_control_disable(cgrp);
3500}
3501
3502static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3503{
3504 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3505
3506 /* if nothing is getting enabled, nothing to worry about */
3507 if (!enable)
3508 return 0;
3509
3510 /* can @cgrp host any resources? */
3511 if (!cgroup_is_valid_domain(cgrp: cgrp->dom_cgrp))
3512 return -EOPNOTSUPP;
3513
3514 /* mixables don't care */
3515 if (cgroup_is_mixable(cgrp))
3516 return 0;
3517
3518 if (domain_enable) {
3519 /* can't enable domain controllers inside a thread subtree */
3520 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3521 return -EOPNOTSUPP;
3522 } else {
3523 /*
3524 * Threaded controllers can handle internal competitions
3525 * and are always allowed inside a (prospective) thread
3526 * subtree.
3527 */
3528 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3529 return 0;
3530 }
3531
3532 /*
3533 * Controllers can't be enabled for a cgroup with tasks to avoid
3534 * child cgroups competing against tasks.
3535 */
3536 if (cgroup_has_tasks(cgrp))
3537 return -EBUSY;
3538
3539 return 0;
3540}
3541
3542/* change the enabled child controllers for a cgroup in the default hierarchy */
3543static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3544 char *buf, size_t nbytes,
3545 loff_t off)
3546{
3547 u16 enable = 0, disable = 0;
3548 struct cgroup *cgrp, *child;
3549 struct cgroup_subsys *ss;
3550 char *tok;
3551 int ssid, ret;
3552
3553 /*
3554 * Parse input - space separated list of subsystem names prefixed
3555 * with either + or -.
3556 */
3557 buf = strstrip(str: buf);
3558 while ((tok = strsep(&buf, " "))) {
3559 if (tok[0] == '\0')
3560 continue;
3561 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3562 if (!cgroup_ssid_enabled(ssid) ||
3563 strcmp(tok + 1, ss->name))
3564 continue;
3565
3566 if (*tok == '+') {
3567 enable |= 1 << ssid;
3568 disable &= ~(1 << ssid);
3569 } else if (*tok == '-') {
3570 disable |= 1 << ssid;
3571 enable &= ~(1 << ssid);
3572 } else {
3573 return -EINVAL;
3574 }
3575 break;
3576 } while_each_subsys_mask();
3577 if (ssid == CGROUP_SUBSYS_COUNT)
3578 return -EINVAL;
3579 }
3580
3581 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: true);
3582 if (!cgrp)
3583 return -ENODEV;
3584
3585 for_each_subsys(ss, ssid) {
3586 if (enable & (1 << ssid)) {
3587 if (cgrp->subtree_control & (1 << ssid)) {
3588 enable &= ~(1 << ssid);
3589 continue;
3590 }
3591
3592 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3593 ret = -ENOENT;
3594 goto out_unlock;
3595 }
3596 } else if (disable & (1 << ssid)) {
3597 if (!(cgrp->subtree_control & (1 << ssid))) {
3598 disable &= ~(1 << ssid);
3599 continue;
3600 }
3601
3602 /* a child has it enabled? */
3603 cgroup_for_each_live_child(child, cgrp) {
3604 if (child->subtree_control & (1 << ssid)) {
3605 ret = -EBUSY;
3606 goto out_unlock;
3607 }
3608 }
3609 }
3610 }
3611
3612 if (!enable && !disable) {
3613 ret = 0;
3614 goto out_unlock;
3615 }
3616
3617 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3618 if (ret)
3619 goto out_unlock;
3620
3621 /* save and update control masks and prepare csses */
3622 cgroup_save_control(cgrp);
3623
3624 cgrp->subtree_control |= enable;
3625 cgrp->subtree_control &= ~disable;
3626
3627 ret = cgroup_apply_control(cgrp);
3628 cgroup_finalize_control(cgrp, ret);
3629 if (ret)
3630 goto out_unlock;
3631
3632 kernfs_activate(kn: cgrp->kn);
3633out_unlock:
3634 cgroup_kn_unlock(kn: of->kn);
3635 return ret ?: nbytes;
3636}
3637
3638/**
3639 * cgroup_enable_threaded - make @cgrp threaded
3640 * @cgrp: the target cgroup
3641 *
3642 * Called when "threaded" is written to the cgroup.type interface file and
3643 * tries to make @cgrp threaded and join the parent's resource domain.
3644 * This function is never called on the root cgroup as cgroup.type doesn't
3645 * exist on it.
3646 */
3647static int cgroup_enable_threaded(struct cgroup *cgrp)
3648{
3649 struct cgroup *parent = cgroup_parent(cgrp);
3650 struct cgroup *dom_cgrp = parent->dom_cgrp;
3651 struct cgroup *dsct;
3652 struct cgroup_subsys_state *d_css;
3653 int ret;
3654
3655 lockdep_assert_held(&cgroup_mutex);
3656
3657 /* noop if already threaded */
3658 if (cgroup_is_threaded(cgrp))
3659 return 0;
3660
3661 /*
3662 * If @cgroup is populated or has domain controllers enabled, it
3663 * can't be switched. While the below cgroup_can_be_thread_root()
3664 * test can catch the same conditions, that's only when @parent is
3665 * not mixable, so let's check it explicitly.
3666 */
3667 if (cgroup_is_populated(cgrp) ||
3668 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3669 return -EOPNOTSUPP;
3670
3671 /* we're joining the parent's domain, ensure its validity */
3672 if (!cgroup_is_valid_domain(cgrp: dom_cgrp) ||
3673 !cgroup_can_be_thread_root(cgrp: dom_cgrp))
3674 return -EOPNOTSUPP;
3675
3676 /*
3677 * The following shouldn't cause actual migrations and should
3678 * always succeed.
3679 */
3680 cgroup_save_control(cgrp);
3681
3682 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3683 if (dsct == cgrp || cgroup_is_threaded(cgrp: dsct))
3684 dsct->dom_cgrp = dom_cgrp;
3685
3686 ret = cgroup_apply_control(cgrp);
3687 if (!ret)
3688 parent->nr_threaded_children++;
3689
3690 cgroup_finalize_control(cgrp, ret);
3691 return ret;
3692}
3693
3694static int cgroup_type_show(struct seq_file *seq, void *v)
3695{
3696 struct cgroup *cgrp = seq_css(seq)->cgroup;
3697
3698 if (cgroup_is_threaded(cgrp))
3699 seq_puts(m: seq, s: "threaded\n");
3700 else if (!cgroup_is_valid_domain(cgrp))
3701 seq_puts(m: seq, s: "domain invalid\n");
3702 else if (cgroup_is_thread_root(cgrp))
3703 seq_puts(m: seq, s: "domain threaded\n");
3704 else
3705 seq_puts(m: seq, s: "domain\n");
3706
3707 return 0;
3708}
3709
3710static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3711 size_t nbytes, loff_t off)
3712{
3713 struct cgroup *cgrp;
3714 int ret;
3715
3716 /* only switching to threaded mode is supported */
3717 if (strcmp(strstrip(str: buf), "threaded"))
3718 return -EINVAL;
3719
3720 /* drain dying csses before we re-apply (threaded) subtree control */
3721 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: true);
3722 if (!cgrp)
3723 return -ENOENT;
3724
3725 /* threaded can only be enabled */
3726 ret = cgroup_enable_threaded(cgrp);
3727
3728 cgroup_kn_unlock(kn: of->kn);
3729 return ret ?: nbytes;
3730}
3731
3732static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3733{
3734 struct cgroup *cgrp = seq_css(seq)->cgroup;
3735 int descendants = READ_ONCE(cgrp->max_descendants);
3736
3737 if (descendants == INT_MAX)
3738 seq_puts(m: seq, s: "max\n");
3739 else
3740 seq_printf(m: seq, fmt: "%d\n", descendants);
3741
3742 return 0;
3743}
3744
3745static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3746 char *buf, size_t nbytes, loff_t off)
3747{
3748 struct cgroup *cgrp;
3749 int descendants;
3750 ssize_t ret;
3751
3752 buf = strstrip(str: buf);
3753 if (!strcmp(buf, "max")) {
3754 descendants = INT_MAX;
3755 } else {
3756 ret = kstrtoint(s: buf, base: 0, res: &descendants);
3757 if (ret)
3758 return ret;
3759 }
3760
3761 if (descendants < 0)
3762 return -ERANGE;
3763
3764 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
3765 if (!cgrp)
3766 return -ENOENT;
3767
3768 cgrp->max_descendants = descendants;
3769
3770 cgroup_kn_unlock(kn: of->kn);
3771
3772 return nbytes;
3773}
3774
3775static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3776{
3777 struct cgroup *cgrp = seq_css(seq)->cgroup;
3778 int depth = READ_ONCE(cgrp->max_depth);
3779
3780 if (depth == INT_MAX)
3781 seq_puts(m: seq, s: "max\n");
3782 else
3783 seq_printf(m: seq, fmt: "%d\n", depth);
3784
3785 return 0;
3786}
3787
3788static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3789 char *buf, size_t nbytes, loff_t off)
3790{
3791 struct cgroup *cgrp;
3792 ssize_t ret;
3793 int depth;
3794
3795 buf = strstrip(str: buf);
3796 if (!strcmp(buf, "max")) {
3797 depth = INT_MAX;
3798 } else {
3799 ret = kstrtoint(s: buf, base: 0, res: &depth);
3800 if (ret)
3801 return ret;
3802 }
3803
3804 if (depth < 0)
3805 return -ERANGE;
3806
3807 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
3808 if (!cgrp)
3809 return -ENOENT;
3810
3811 cgrp->max_depth = depth;
3812
3813 cgroup_kn_unlock(kn: of->kn);
3814
3815 return nbytes;
3816}
3817
3818static int cgroup_events_show(struct seq_file *seq, void *v)
3819{
3820 struct cgroup *cgrp = seq_css(seq)->cgroup;
3821
3822 seq_printf(m: seq, fmt: "populated %d\n", cgroup_is_populated(cgrp));
3823 seq_printf(m: seq, fmt: "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3824
3825 return 0;
3826}
3827
3828static int cgroup_stat_show(struct seq_file *seq, void *v)
3829{
3830 struct cgroup *cgroup = seq_css(seq)->cgroup;
3831 struct cgroup_subsys_state *css;
3832 int dying_cnt[CGROUP_SUBSYS_COUNT];
3833 int ssid;
3834
3835 seq_printf(m: seq, fmt: "nr_descendants %d\n",
3836 cgroup->nr_descendants);
3837
3838 /*
3839 * Show the number of live and dying csses associated with each of
3840 * non-inhibited cgroup subsystems that is bound to cgroup v2.
3841 *
3842 * Without proper lock protection, racing is possible. So the
3843 * numbers may not be consistent when that happens.
3844 */
3845 rcu_read_lock();
3846 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) {
3847 dying_cnt[ssid] = -1;
3848 if ((BIT(ssid) & cgrp_dfl_inhibit_ss_mask) ||
3849 (cgroup_subsys[ssid]->root != &cgrp_dfl_root))
3850 continue;
3851 css = rcu_dereference_raw(cgroup->subsys[ssid]);
3852 dying_cnt[ssid] = cgroup->nr_dying_subsys[ssid];
3853 seq_printf(m: seq, fmt: "nr_subsys_%s %d\n", cgroup_subsys[ssid]->name,
3854 css ? (css->nr_descendants + 1) : 0);
3855 }
3856
3857 seq_printf(m: seq, fmt: "nr_dying_descendants %d\n",
3858 cgroup->nr_dying_descendants);
3859 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) {
3860 if (dying_cnt[ssid] >= 0)
3861 seq_printf(m: seq, fmt: "nr_dying_subsys_%s %d\n",
3862 cgroup_subsys[ssid]->name, dying_cnt[ssid]);
3863 }
3864 rcu_read_unlock();
3865 return 0;
3866}
3867
3868static int cgroup_core_local_stat_show(struct seq_file *seq, void *v)
3869{
3870 struct cgroup *cgrp = seq_css(seq)->cgroup;
3871 unsigned int sequence;
3872 u64 freeze_time;
3873
3874 do {
3875 sequence = read_seqcount_begin(&cgrp->freezer.freeze_seq);
3876 freeze_time = cgrp->freezer.frozen_nsec;
3877 /* Add in current freezer interval if the cgroup is freezing. */
3878 if (test_bit(CGRP_FREEZE, &cgrp->flags))
3879 freeze_time += (ktime_get_ns() -
3880 cgrp->freezer.freeze_start_nsec);
3881 } while (read_seqcount_retry(&cgrp->freezer.freeze_seq, sequence));
3882
3883 do_div(freeze_time, NSEC_PER_USEC);
3884 seq_printf(m: seq, fmt: "frozen_usec %llu\n", freeze_time);
3885
3886 return 0;
3887}
3888
3889#ifdef CONFIG_CGROUP_SCHED
3890/**
3891 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
3892 * @cgrp: the cgroup of interest
3893 * @ss: the subsystem of interest
3894 *
3895 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
3896 * or is offline, %NULL is returned.
3897 */
3898static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
3899 struct cgroup_subsys *ss)
3900{
3901 struct cgroup_subsys_state *css;
3902
3903 rcu_read_lock();
3904 css = cgroup_css(cgrp, ss);
3905 if (css && !css_tryget_online(css))
3906 css = NULL;
3907 rcu_read_unlock();
3908
3909 return css;
3910}
3911
3912static int cgroup_extra_stat_show(struct seq_file *seq, int ssid)
3913{
3914 struct cgroup *cgrp = seq_css(seq)->cgroup;
3915 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3916 struct cgroup_subsys_state *css;
3917 int ret;
3918
3919 if (!ss->css_extra_stat_show)
3920 return 0;
3921
3922 css = cgroup_tryget_css(cgrp, ss);
3923 if (!css)
3924 return 0;
3925
3926 ret = ss->css_extra_stat_show(seq, css);
3927 css_put(css);
3928 return ret;
3929}
3930
3931static int cgroup_local_stat_show(struct seq_file *seq,
3932 struct cgroup *cgrp, int ssid)
3933{
3934 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3935 struct cgroup_subsys_state *css;
3936 int ret;
3937
3938 if (!ss->css_local_stat_show)
3939 return 0;
3940
3941 css = cgroup_tryget_css(cgrp, ss);
3942 if (!css)
3943 return 0;
3944
3945 ret = ss->css_local_stat_show(seq, css);
3946 css_put(css);
3947 return ret;
3948}
3949#endif
3950
3951static int cpu_stat_show(struct seq_file *seq, void *v)
3952{
3953 int ret = 0;
3954
3955 cgroup_base_stat_cputime_show(seq);
3956#ifdef CONFIG_CGROUP_SCHED
3957 ret = cgroup_extra_stat_show(seq, ssid: cpu_cgrp_id);
3958#endif
3959 return ret;
3960}
3961
3962static int cpu_local_stat_show(struct seq_file *seq, void *v)
3963{
3964 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3965 int ret = 0;
3966
3967#ifdef CONFIG_CGROUP_SCHED
3968 ret = cgroup_local_stat_show(seq, cgrp, ssid: cpu_cgrp_id);
3969#endif
3970 return ret;
3971}
3972
3973#ifdef CONFIG_PSI
3974static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3975{
3976 struct cgroup *cgrp = seq_css(seq)->cgroup;
3977 struct psi_group *psi = cgroup_psi(cgrp);
3978
3979 return psi_show(seq, psi, PSI_IO);
3980}
3981static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3982{
3983 struct cgroup *cgrp = seq_css(seq)->cgroup;
3984 struct psi_group *psi = cgroup_psi(cgrp);
3985
3986 return psi_show(seq, psi, PSI_MEM);
3987}
3988static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3989{
3990 struct cgroup *cgrp = seq_css(seq)->cgroup;
3991 struct psi_group *psi = cgroup_psi(cgrp);
3992
3993 return psi_show(seq, psi, PSI_CPU);
3994}
3995
3996static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3997 size_t nbytes, enum psi_res res)
3998{
3999 struct cgroup_file_ctx *ctx = of->priv;
4000 struct psi_trigger *new;
4001 struct cgroup *cgrp;
4002 struct psi_group *psi;
4003
4004 cgrp = cgroup_kn_lock_live(of->kn, false);
4005 if (!cgrp)
4006 return -ENODEV;
4007
4008 cgroup_get(cgrp);
4009 cgroup_kn_unlock(of->kn);
4010
4011 /* Allow only one trigger per file descriptor */
4012 if (ctx->psi.trigger) {
4013 cgroup_put(cgrp);
4014 return -EBUSY;
4015 }
4016
4017 psi = cgroup_psi(cgrp);
4018 new = psi_trigger_create(psi, buf, res, of->file, of);
4019 if (IS_ERR(new)) {
4020 cgroup_put(cgrp);
4021 return PTR_ERR(new);
4022 }
4023
4024 smp_store_release(&ctx->psi.trigger, new);
4025 cgroup_put(cgrp);
4026
4027 return nbytes;
4028}
4029
4030static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
4031 char *buf, size_t nbytes,
4032 loff_t off)
4033{
4034 return pressure_write(of, buf, nbytes, PSI_IO);
4035}
4036
4037static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
4038 char *buf, size_t nbytes,
4039 loff_t off)
4040{
4041 return pressure_write(of, buf, nbytes, PSI_MEM);
4042}
4043
4044static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
4045 char *buf, size_t nbytes,
4046 loff_t off)
4047{
4048 return pressure_write(of, buf, nbytes, PSI_CPU);
4049}
4050
4051#ifdef CONFIG_IRQ_TIME_ACCOUNTING
4052static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
4053{
4054 struct cgroup *cgrp = seq_css(seq)->cgroup;
4055 struct psi_group *psi = cgroup_psi(cgrp);
4056
4057 return psi_show(seq, psi, PSI_IRQ);
4058}
4059
4060static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
4061 char *buf, size_t nbytes,
4062 loff_t off)
4063{
4064 return pressure_write(of, buf, nbytes, PSI_IRQ);
4065}
4066#endif
4067
4068static int cgroup_pressure_show(struct seq_file *seq, void *v)
4069{
4070 struct cgroup *cgrp = seq_css(seq)->cgroup;
4071 struct psi_group *psi = cgroup_psi(cgrp);
4072
4073 seq_printf(seq, "%d\n", psi->enabled);
4074
4075 return 0;
4076}
4077
4078static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
4079 char *buf, size_t nbytes,
4080 loff_t off)
4081{
4082 ssize_t ret;
4083 int enable;
4084 struct cgroup *cgrp;
4085 struct psi_group *psi;
4086
4087 ret = kstrtoint(strstrip(buf), 0, &enable);
4088 if (ret)
4089 return ret;
4090
4091 if (enable < 0 || enable > 1)
4092 return -ERANGE;
4093
4094 cgrp = cgroup_kn_lock_live(of->kn, false);
4095 if (!cgrp)
4096 return -ENOENT;
4097
4098 psi = cgroup_psi(cgrp);
4099 if (psi->enabled != enable) {
4100 int i;
4101
4102 /* show or hide {cpu,memory,io,irq}.pressure files */
4103 for (i = 0; i < NR_PSI_RESOURCES; i++)
4104 cgroup_file_show(&cgrp->psi_files[i], enable);
4105
4106 psi->enabled = enable;
4107 if (enable)
4108 psi_cgroup_restart(psi);
4109 }
4110
4111 cgroup_kn_unlock(of->kn);
4112
4113 return nbytes;
4114}
4115
4116static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
4117 poll_table *pt)
4118{
4119 struct cgroup_file_ctx *ctx = of->priv;
4120
4121 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
4122}
4123
4124static void cgroup_pressure_release(struct kernfs_open_file *of)
4125{
4126 struct cgroup_file_ctx *ctx = of->priv;
4127
4128 psi_trigger_destroy(ctx->psi.trigger);
4129}
4130
4131bool cgroup_psi_enabled(void)
4132{
4133 if (static_branch_likely(&psi_disabled))
4134 return false;
4135
4136 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
4137}
4138
4139#else /* CONFIG_PSI */
4140bool cgroup_psi_enabled(void)
4141{
4142 return false;
4143}
4144
4145#endif /* CONFIG_PSI */
4146
4147static int cgroup_freeze_show(struct seq_file *seq, void *v)
4148{
4149 struct cgroup *cgrp = seq_css(seq)->cgroup;
4150
4151 seq_printf(m: seq, fmt: "%d\n", cgrp->freezer.freeze);
4152
4153 return 0;
4154}
4155
4156static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
4157 char *buf, size_t nbytes, loff_t off)
4158{
4159 struct cgroup *cgrp;
4160 ssize_t ret;
4161 int freeze;
4162
4163 ret = kstrtoint(s: strstrip(str: buf), base: 0, res: &freeze);
4164 if (ret)
4165 return ret;
4166
4167 if (freeze < 0 || freeze > 1)
4168 return -ERANGE;
4169
4170 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
4171 if (!cgrp)
4172 return -ENOENT;
4173
4174 cgroup_freeze(cgrp, freeze);
4175
4176 cgroup_kn_unlock(kn: of->kn);
4177
4178 return nbytes;
4179}
4180
4181static void __cgroup_kill(struct cgroup *cgrp)
4182{
4183 struct css_task_iter it;
4184 struct task_struct *task;
4185
4186 lockdep_assert_held(&cgroup_mutex);
4187
4188 spin_lock_irq(lock: &css_set_lock);
4189 cgrp->kill_seq++;
4190 spin_unlock_irq(lock: &css_set_lock);
4191
4192 css_task_iter_start(css: &cgrp->self, flags: CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, it: &it);
4193 while ((task = css_task_iter_next(it: &it))) {
4194 /* Ignore kernel threads here. */
4195 if (task->flags & PF_KTHREAD)
4196 continue;
4197
4198 /* Skip tasks that are already dying. */
4199 if (__fatal_signal_pending(p: task))
4200 continue;
4201
4202 send_sig(SIGKILL, task, 0);
4203 }
4204 css_task_iter_end(it: &it);
4205}
4206
4207static void cgroup_kill(struct cgroup *cgrp)
4208{
4209 struct cgroup_subsys_state *css;
4210 struct cgroup *dsct;
4211
4212 lockdep_assert_held(&cgroup_mutex);
4213
4214 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
4215 __cgroup_kill(cgrp: dsct);
4216}
4217
4218static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
4219 size_t nbytes, loff_t off)
4220{
4221 ssize_t ret = 0;
4222 int kill;
4223 struct cgroup *cgrp;
4224
4225 ret = kstrtoint(s: strstrip(str: buf), base: 0, res: &kill);
4226 if (ret)
4227 return ret;
4228
4229 if (kill != 1)
4230 return -ERANGE;
4231
4232 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
4233 if (!cgrp)
4234 return -ENOENT;
4235
4236 /*
4237 * Killing is a process directed operation, i.e. the whole thread-group
4238 * is taken down so act like we do for cgroup.procs and only make this
4239 * writable in non-threaded cgroups.
4240 */
4241 if (cgroup_is_threaded(cgrp))
4242 ret = -EOPNOTSUPP;
4243 else
4244 cgroup_kill(cgrp);
4245
4246 cgroup_kn_unlock(kn: of->kn);
4247
4248 return ret ?: nbytes;
4249}
4250
4251static int cgroup_file_open(struct kernfs_open_file *of)
4252{
4253 struct cftype *cft = of_cft(of);
4254 struct cgroup_file_ctx *ctx;
4255 int ret;
4256
4257 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
4258 if (!ctx)
4259 return -ENOMEM;
4260
4261 ctx->ns = current->nsproxy->cgroup_ns;
4262 get_cgroup_ns(ns: ctx->ns);
4263 of->priv = ctx;
4264
4265 if (!cft->open)
4266 return 0;
4267
4268 ret = cft->open(of);
4269 if (ret) {
4270 put_cgroup_ns(ns: ctx->ns);
4271 kfree(objp: ctx);
4272 }
4273 return ret;
4274}
4275
4276static void cgroup_file_release(struct kernfs_open_file *of)
4277{
4278 struct cftype *cft = of_cft(of);
4279 struct cgroup_file_ctx *ctx = of->priv;
4280
4281 if (cft->release)
4282 cft->release(of);
4283 put_cgroup_ns(ns: ctx->ns);
4284 kfree(objp: ctx);
4285 of->priv = NULL;
4286}
4287
4288static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4289 size_t nbytes, loff_t off)
4290{
4291 struct cgroup_file_ctx *ctx = of->priv;
4292 struct cgroup *cgrp = kn_priv(kn: of->kn);
4293 struct cftype *cft = of_cft(of);
4294 struct cgroup_subsys_state *css;
4295 int ret;
4296
4297 if (!nbytes)
4298 return 0;
4299
4300 /*
4301 * If namespaces are delegation boundaries, disallow writes to
4302 * files in an non-init namespace root from inside the namespace
4303 * except for the files explicitly marked delegatable -
4304 * eg. cgroup.procs, cgroup.threads and cgroup.subtree_control.
4305 */
4306 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4307 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4308 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4309 return -EPERM;
4310
4311 if (cft->write)
4312 return cft->write(of, buf, nbytes, off);
4313
4314 /*
4315 * kernfs guarantees that a file isn't deleted with operations in
4316 * flight, which means that the matching css is and stays alive and
4317 * doesn't need to be pinned. The RCU locking is not necessary
4318 * either. It's just for the convenience of using cgroup_css().
4319 */
4320 rcu_read_lock();
4321 css = cgroup_css(cgrp, ss: cft->ss);
4322 rcu_read_unlock();
4323
4324 if (cft->write_u64) {
4325 unsigned long long v;
4326 ret = kstrtoull(s: buf, base: 0, res: &v);
4327 if (!ret)
4328 ret = cft->write_u64(css, cft, v);
4329 } else if (cft->write_s64) {
4330 long long v;
4331 ret = kstrtoll(s: buf, base: 0, res: &v);
4332 if (!ret)
4333 ret = cft->write_s64(css, cft, v);
4334 } else {
4335 ret = -EINVAL;
4336 }
4337
4338 return ret ?: nbytes;
4339}
4340
4341static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4342{
4343 struct cftype *cft = of_cft(of);
4344
4345 if (cft->poll)
4346 return cft->poll(of, pt);
4347
4348 return kernfs_generic_poll(of, pt);
4349}
4350
4351static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4352{
4353 return seq_cft(seq)->seq_start(seq, ppos);
4354}
4355
4356static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4357{
4358 return seq_cft(seq)->seq_next(seq, v, ppos);
4359}
4360
4361static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4362{
4363 if (seq_cft(seq)->seq_stop)
4364 seq_cft(seq)->seq_stop(seq, v);
4365}
4366
4367static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4368{
4369 struct cftype *cft = seq_cft(seq: m);
4370 struct cgroup_subsys_state *css = seq_css(seq: m);
4371
4372 if (cft->seq_show)
4373 return cft->seq_show(m, arg);
4374
4375 if (cft->read_u64)
4376 seq_printf(m, fmt: "%llu\n", cft->read_u64(css, cft));
4377 else if (cft->read_s64)
4378 seq_printf(m, fmt: "%lld\n", cft->read_s64(css, cft));
4379 else
4380 return -EINVAL;
4381 return 0;
4382}
4383
4384static struct kernfs_ops cgroup_kf_single_ops = {
4385 .atomic_write_len = PAGE_SIZE,
4386 .open = cgroup_file_open,
4387 .release = cgroup_file_release,
4388 .write = cgroup_file_write,
4389 .poll = cgroup_file_poll,
4390 .seq_show = cgroup_seqfile_show,
4391};
4392
4393static struct kernfs_ops cgroup_kf_ops = {
4394 .atomic_write_len = PAGE_SIZE,
4395 .open = cgroup_file_open,
4396 .release = cgroup_file_release,
4397 .write = cgroup_file_write,
4398 .poll = cgroup_file_poll,
4399 .seq_start = cgroup_seqfile_start,
4400 .seq_next = cgroup_seqfile_next,
4401 .seq_stop = cgroup_seqfile_stop,
4402 .seq_show = cgroup_seqfile_show,
4403};
4404
4405static void cgroup_file_notify_timer(struct timer_list *timer)
4406{
4407 cgroup_file_notify(container_of(timer, struct cgroup_file,
4408 notify_timer));
4409}
4410
4411static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4412 struct cftype *cft)
4413{
4414 char name[CGROUP_FILE_NAME_MAX];
4415 struct kernfs_node *kn;
4416 struct lock_class_key *key = NULL;
4417
4418#ifdef CONFIG_DEBUG_LOCK_ALLOC
4419 key = &cft->lockdep_key;
4420#endif
4421 kn = __kernfs_create_file(parent: cgrp->kn, name: cgroup_file_name(cgrp, cft, buf: name),
4422 mode: cgroup_file_mode(cft),
4423 current_fsuid(), current_fsgid(),
4424 size: 0, ops: cft->kf_ops, priv: cft,
4425 NULL, key);
4426 if (IS_ERR(ptr: kn))
4427 return PTR_ERR(ptr: kn);
4428
4429 if (cft->file_offset) {
4430 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4431
4432 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4433
4434 spin_lock_irq(lock: &cgroup_file_kn_lock);
4435 cfile->kn = kn;
4436 spin_unlock_irq(lock: &cgroup_file_kn_lock);
4437 }
4438
4439 return 0;
4440}
4441
4442/**
4443 * cgroup_addrm_files - add or remove files to a cgroup directory
4444 * @css: the target css
4445 * @cgrp: the target cgroup (usually css->cgroup)
4446 * @cfts: array of cftypes to be added
4447 * @is_add: whether to add or remove
4448 *
4449 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4450 * For removals, this function never fails.
4451 */
4452static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4453 struct cgroup *cgrp, struct cftype cfts[],
4454 bool is_add)
4455{
4456 struct cftype *cft, *cft_end = NULL;
4457 int ret = 0;
4458
4459 lockdep_assert_held(&cgroup_mutex);
4460
4461restart:
4462 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4463 /* does cft->flags tell us to skip this file on @cgrp? */
4464 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4465 continue;
4466 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4467 continue;
4468 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4469 continue;
4470 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4471 continue;
4472 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4473 continue;
4474 if (is_add) {
4475 ret = cgroup_add_file(css, cgrp, cft);
4476 if (ret) {
4477 pr_warn("%s: failed to add %s, err=%d\n",
4478 __func__, cft->name, ret);
4479 cft_end = cft;
4480 is_add = false;
4481 goto restart;
4482 }
4483 } else {
4484 cgroup_rm_file(cgrp, cft);
4485 }
4486 }
4487 return ret;
4488}
4489
4490static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4491{
4492 struct cgroup_subsys *ss = cfts[0].ss;
4493 struct cgroup *root = &ss->root->cgrp;
4494 struct cgroup_subsys_state *css;
4495 int ret = 0;
4496
4497 lockdep_assert_held(&cgroup_mutex);
4498
4499 /* add/rm files for all cgroups created before */
4500 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4501 struct cgroup *cgrp = css->cgroup;
4502
4503 if (!(css->flags & CSS_VISIBLE))
4504 continue;
4505
4506 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4507 if (ret)
4508 break;
4509 }
4510
4511 if (is_add && !ret)
4512 kernfs_activate(kn: root->kn);
4513 return ret;
4514}
4515
4516static void cgroup_exit_cftypes(struct cftype *cfts)
4517{
4518 struct cftype *cft;
4519
4520 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4521 /* free copy for custom atomic_write_len, see init_cftypes() */
4522 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4523 kfree(objp: cft->kf_ops);
4524 cft->kf_ops = NULL;
4525 cft->ss = NULL;
4526
4527 /* revert flags set by cgroup core while adding @cfts */
4528 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4529 __CFTYPE_ADDED);
4530 }
4531}
4532
4533static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4534{
4535 struct cftype *cft;
4536 int ret = 0;
4537
4538 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4539 struct kernfs_ops *kf_ops;
4540
4541 WARN_ON(cft->ss || cft->kf_ops);
4542
4543 if (cft->flags & __CFTYPE_ADDED) {
4544 ret = -EBUSY;
4545 break;
4546 }
4547
4548 if (cft->seq_start)
4549 kf_ops = &cgroup_kf_ops;
4550 else
4551 kf_ops = &cgroup_kf_single_ops;
4552
4553 /*
4554 * Ugh... if @cft wants a custom max_write_len, we need to
4555 * make a copy of kf_ops to set its atomic_write_len.
4556 */
4557 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4558 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4559 if (!kf_ops) {
4560 ret = -ENOMEM;
4561 break;
4562 }
4563 kf_ops->atomic_write_len = cft->max_write_len;
4564 }
4565
4566 cft->kf_ops = kf_ops;
4567 cft->ss = ss;
4568 cft->flags |= __CFTYPE_ADDED;
4569 }
4570
4571 if (ret)
4572 cgroup_exit_cftypes(cfts);
4573 return ret;
4574}
4575
4576static void cgroup_rm_cftypes_locked(struct cftype *cfts)
4577{
4578 lockdep_assert_held(&cgroup_mutex);
4579
4580 list_del(entry: &cfts->node);
4581 cgroup_apply_cftypes(cfts, is_add: false);
4582 cgroup_exit_cftypes(cfts);
4583}
4584
4585/**
4586 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4587 * @cfts: zero-length name terminated array of cftypes
4588 *
4589 * Unregister @cfts. Files described by @cfts are removed from all
4590 * existing cgroups and all future cgroups won't have them either. This
4591 * function can be called anytime whether @cfts' subsys is attached or not.
4592 *
4593 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4594 * registered.
4595 */
4596int cgroup_rm_cftypes(struct cftype *cfts)
4597{
4598 if (!cfts || cfts[0].name[0] == '\0')
4599 return 0;
4600
4601 if (!(cfts[0].flags & __CFTYPE_ADDED))
4602 return -ENOENT;
4603
4604 cgroup_lock();
4605 cgroup_rm_cftypes_locked(cfts);
4606 cgroup_unlock();
4607 return 0;
4608}
4609
4610/**
4611 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4612 * @ss: target cgroup subsystem
4613 * @cfts: zero-length name terminated array of cftypes
4614 *
4615 * Register @cfts to @ss. Files described by @cfts are created for all
4616 * existing cgroups to which @ss is attached and all future cgroups will
4617 * have them too. This function can be called anytime whether @ss is
4618 * attached or not.
4619 *
4620 * Returns 0 on successful registration, -errno on failure. Note that this
4621 * function currently returns 0 as long as @cfts registration is successful
4622 * even if some file creation attempts on existing cgroups fail.
4623 */
4624int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4625{
4626 int ret;
4627
4628 if (!cgroup_ssid_enabled(ssid: ss->id))
4629 return 0;
4630
4631 if (!cfts || cfts[0].name[0] == '\0')
4632 return 0;
4633
4634 ret = cgroup_init_cftypes(ss, cfts);
4635 if (ret)
4636 return ret;
4637
4638 cgroup_lock();
4639
4640 list_add_tail(new: &cfts->node, head: &ss->cfts);
4641 ret = cgroup_apply_cftypes(cfts, is_add: true);
4642 if (ret)
4643 cgroup_rm_cftypes_locked(cfts);
4644
4645 cgroup_unlock();
4646 return ret;
4647}
4648
4649/**
4650 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4651 * @ss: target cgroup subsystem
4652 * @cfts: zero-length name terminated array of cftypes
4653 *
4654 * Similar to cgroup_add_cftypes() but the added files are only used for
4655 * the default hierarchy.
4656 */
4657int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4658{
4659 struct cftype *cft;
4660
4661 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4662 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4663 return cgroup_add_cftypes(ss, cfts);
4664}
4665
4666/**
4667 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4668 * @ss: target cgroup subsystem
4669 * @cfts: zero-length name terminated array of cftypes
4670 *
4671 * Similar to cgroup_add_cftypes() but the added files are only used for
4672 * the legacy hierarchies.
4673 */
4674int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4675{
4676 struct cftype *cft;
4677
4678 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4679 cft->flags |= __CFTYPE_NOT_ON_DFL;
4680 return cgroup_add_cftypes(ss, cfts);
4681}
4682
4683/**
4684 * cgroup_file_notify - generate a file modified event for a cgroup_file
4685 * @cfile: target cgroup_file
4686 *
4687 * @cfile must have been obtained by setting cftype->file_offset.
4688 */
4689void cgroup_file_notify(struct cgroup_file *cfile)
4690{
4691 unsigned long flags;
4692
4693 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4694 if (cfile->kn) {
4695 unsigned long last = cfile->notified_at;
4696 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4697
4698 if (time_in_range(jiffies, last, next)) {
4699 timer_reduce(timer: &cfile->notify_timer, expires: next);
4700 } else {
4701 kernfs_notify(kn: cfile->kn);
4702 cfile->notified_at = jiffies;
4703 }
4704 }
4705 spin_unlock_irqrestore(lock: &cgroup_file_kn_lock, flags);
4706}
4707
4708/**
4709 * cgroup_file_show - show or hide a hidden cgroup file
4710 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4711 * @show: whether to show or hide
4712 */
4713void cgroup_file_show(struct cgroup_file *cfile, bool show)
4714{
4715 struct kernfs_node *kn;
4716
4717 spin_lock_irq(lock: &cgroup_file_kn_lock);
4718 kn = cfile->kn;
4719 kernfs_get(kn);
4720 spin_unlock_irq(lock: &cgroup_file_kn_lock);
4721
4722 if (kn)
4723 kernfs_show(kn, show);
4724
4725 kernfs_put(kn);
4726}
4727
4728/**
4729 * css_next_child - find the next child of a given css
4730 * @pos: the current position (%NULL to initiate traversal)
4731 * @parent: css whose children to walk
4732 *
4733 * This function returns the next child of @parent and should be called
4734 * under either cgroup_mutex or RCU read lock. The only requirement is
4735 * that @parent and @pos are accessible. The next sibling is guaranteed to
4736 * be returned regardless of their states.
4737 *
4738 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4739 * css which finished ->css_online() is guaranteed to be visible in the
4740 * future iterations and will stay visible until the last reference is put.
4741 * A css which hasn't finished ->css_online() or already finished
4742 * ->css_offline() may show up during traversal. It's each subsystem's
4743 * responsibility to synchronize against on/offlining.
4744 */
4745struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4746 struct cgroup_subsys_state *parent)
4747{
4748 struct cgroup_subsys_state *next;
4749
4750 cgroup_assert_mutex_or_rcu_locked();
4751
4752 /*
4753 * @pos could already have been unlinked from the sibling list.
4754 * Once a cgroup is removed, its ->sibling.next is no longer
4755 * updated when its next sibling changes. CSS_RELEASED is set when
4756 * @pos is taken off list, at which time its next pointer is valid,
4757 * and, as releases are serialized, the one pointed to by the next
4758 * pointer is guaranteed to not have started release yet. This
4759 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4760 * critical section, the one pointed to by its next pointer is
4761 * guaranteed to not have finished its RCU grace period even if we
4762 * have dropped rcu_read_lock() in-between iterations.
4763 *
4764 * If @pos has CSS_RELEASED set, its next pointer can't be
4765 * dereferenced; however, as each css is given a monotonically
4766 * increasing unique serial number and always appended to the
4767 * sibling list, the next one can be found by walking the parent's
4768 * children until the first css with higher serial number than
4769 * @pos's. While this path can be slower, it happens iff iteration
4770 * races against release and the race window is very small.
4771 */
4772 if (!pos) {
4773 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4774 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4775 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4776 } else {
4777 list_for_each_entry_rcu(next, &parent->children, sibling,
4778 lockdep_is_held(&cgroup_mutex))
4779 if (next->serial_nr > pos->serial_nr)
4780 break;
4781 }
4782
4783 /*
4784 * @next, if not pointing to the head, can be dereferenced and is
4785 * the next sibling.
4786 */
4787 if (&next->sibling != &parent->children)
4788 return next;
4789 return NULL;
4790}
4791
4792/**
4793 * css_next_descendant_pre - find the next descendant for pre-order walk
4794 * @pos: the current position (%NULL to initiate traversal)
4795 * @root: css whose descendants to walk
4796 *
4797 * To be used by css_for_each_descendant_pre(). Find the next descendant
4798 * to visit for pre-order traversal of @root's descendants. @root is
4799 * included in the iteration and the first node to be visited.
4800 *
4801 * While this function requires cgroup_mutex or RCU read locking, it
4802 * doesn't require the whole traversal to be contained in a single critical
4803 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4804 * This function will return the correct next descendant as long as both @pos
4805 * and @root are accessible and @pos is a descendant of @root.
4806 *
4807 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4808 * css which finished ->css_online() is guaranteed to be visible in the
4809 * future iterations and will stay visible until the last reference is put.
4810 * A css which hasn't finished ->css_online() or already finished
4811 * ->css_offline() may show up during traversal. It's each subsystem's
4812 * responsibility to synchronize against on/offlining.
4813 */
4814struct cgroup_subsys_state *
4815css_next_descendant_pre(struct cgroup_subsys_state *pos,
4816 struct cgroup_subsys_state *root)
4817{
4818 struct cgroup_subsys_state *next;
4819
4820 cgroup_assert_mutex_or_rcu_locked();
4821
4822 /* if first iteration, visit @root */
4823 if (!pos)
4824 return root;
4825
4826 /* visit the first child if exists */
4827 next = css_next_child(NULL, parent: pos);
4828 if (next)
4829 return next;
4830
4831 /* no child, visit my or the closest ancestor's next sibling */
4832 while (pos != root) {
4833 next = css_next_child(pos, parent: pos->parent);
4834 if (next)
4835 return next;
4836 pos = pos->parent;
4837 }
4838
4839 return NULL;
4840}
4841EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4842
4843/**
4844 * css_rightmost_descendant - return the rightmost descendant of a css
4845 * @pos: css of interest
4846 *
4847 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4848 * is returned. This can be used during pre-order traversal to skip
4849 * subtree of @pos.
4850 *
4851 * While this function requires cgroup_mutex or RCU read locking, it
4852 * doesn't require the whole traversal to be contained in a single critical
4853 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4854 * This function will return the correct rightmost descendant as long as @pos
4855 * is accessible.
4856 */
4857struct cgroup_subsys_state *
4858css_rightmost_descendant(struct cgroup_subsys_state *pos)
4859{
4860 struct cgroup_subsys_state *last, *tmp;
4861
4862 cgroup_assert_mutex_or_rcu_locked();
4863
4864 do {
4865 last = pos;
4866 /* ->prev isn't RCU safe, walk ->next till the end */
4867 pos = NULL;
4868 css_for_each_child(tmp, last)
4869 pos = tmp;
4870 } while (pos);
4871
4872 return last;
4873}
4874
4875static struct cgroup_subsys_state *
4876css_leftmost_descendant(struct cgroup_subsys_state *pos)
4877{
4878 struct cgroup_subsys_state *last;
4879
4880 do {
4881 last = pos;
4882 pos = css_next_child(NULL, parent: pos);
4883 } while (pos);
4884
4885 return last;
4886}
4887
4888/**
4889 * css_next_descendant_post - find the next descendant for post-order walk
4890 * @pos: the current position (%NULL to initiate traversal)
4891 * @root: css whose descendants to walk
4892 *
4893 * To be used by css_for_each_descendant_post(). Find the next descendant
4894 * to visit for post-order traversal of @root's descendants. @root is
4895 * included in the iteration and the last node to be visited.
4896 *
4897 * While this function requires cgroup_mutex or RCU read locking, it
4898 * doesn't require the whole traversal to be contained in a single critical
4899 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4900 * This function will return the correct next descendant as long as both @pos
4901 * and @cgroup are accessible and @pos is a descendant of @cgroup.
4902 *
4903 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4904 * css which finished ->css_online() is guaranteed to be visible in the
4905 * future iterations and will stay visible until the last reference is put.
4906 * A css which hasn't finished ->css_online() or already finished
4907 * ->css_offline() may show up during traversal. It's each subsystem's
4908 * responsibility to synchronize against on/offlining.
4909 */
4910struct cgroup_subsys_state *
4911css_next_descendant_post(struct cgroup_subsys_state *pos,
4912 struct cgroup_subsys_state *root)
4913{
4914 struct cgroup_subsys_state *next;
4915
4916 cgroup_assert_mutex_or_rcu_locked();
4917
4918 /* if first iteration, visit leftmost descendant which may be @root */
4919 if (!pos)
4920 return css_leftmost_descendant(pos: root);
4921
4922 /* if we visited @root, we're done */
4923 if (pos == root)
4924 return NULL;
4925
4926 /* if there's an unvisited sibling, visit its leftmost descendant */
4927 next = css_next_child(pos, parent: pos->parent);
4928 if (next)
4929 return css_leftmost_descendant(pos: next);
4930
4931 /* no sibling left, visit parent */
4932 return pos->parent;
4933}
4934
4935/**
4936 * css_has_online_children - does a css have online children
4937 * @css: the target css
4938 *
4939 * Returns %true if @css has any online children; otherwise, %false. This
4940 * function can be called from any context but the caller is responsible
4941 * for synchronizing against on/offlining as necessary.
4942 */
4943bool css_has_online_children(struct cgroup_subsys_state *css)
4944{
4945 struct cgroup_subsys_state *child;
4946 bool ret = false;
4947
4948 rcu_read_lock();
4949 css_for_each_child(child, css) {
4950 if (child->flags & CSS_ONLINE) {
4951 ret = true;
4952 break;
4953 }
4954 }
4955 rcu_read_unlock();
4956 return ret;
4957}
4958
4959static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4960{
4961 struct list_head *l;
4962 struct cgrp_cset_link *link;
4963 struct css_set *cset;
4964
4965 lockdep_assert_held(&css_set_lock);
4966
4967 /* find the next threaded cset */
4968 if (it->tcset_pos) {
4969 l = it->tcset_pos->next;
4970
4971 if (l != it->tcset_head) {
4972 it->tcset_pos = l;
4973 return container_of(l, struct css_set,
4974 threaded_csets_node);
4975 }
4976
4977 it->tcset_pos = NULL;
4978 }
4979
4980 /* find the next cset */
4981 l = it->cset_pos;
4982 l = l->next;
4983 if (l == it->cset_head) {
4984 it->cset_pos = NULL;
4985 return NULL;
4986 }
4987
4988 if (it->ss) {
4989 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4990 } else {
4991 link = list_entry(l, struct cgrp_cset_link, cset_link);
4992 cset = link->cset;
4993 }
4994
4995 it->cset_pos = l;
4996
4997 /* initialize threaded css_set walking */
4998 if (it->flags & CSS_TASK_ITER_THREADED) {
4999 if (it->cur_dcset)
5000 put_css_set_locked(cset: it->cur_dcset);
5001 it->cur_dcset = cset;
5002 get_css_set(cset);
5003
5004 it->tcset_head = &cset->threaded_csets;
5005 it->tcset_pos = &cset->threaded_csets;
5006 }
5007
5008 return cset;
5009}
5010
5011/**
5012 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
5013 * @it: the iterator to advance
5014 *
5015 * Advance @it to the next css_set to walk.
5016 */
5017static void css_task_iter_advance_css_set(struct css_task_iter *it)
5018{
5019 struct css_set *cset;
5020
5021 lockdep_assert_held(&css_set_lock);
5022
5023 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
5024 while ((cset = css_task_iter_next_css_set(it))) {
5025 if (!list_empty(head: &cset->tasks)) {
5026 it->cur_tasks_head = &cset->tasks;
5027 break;
5028 } else if (!list_empty(head: &cset->mg_tasks)) {
5029 it->cur_tasks_head = &cset->mg_tasks;
5030 break;
5031 } else if (!list_empty(head: &cset->dying_tasks)) {
5032 it->cur_tasks_head = &cset->dying_tasks;
5033 break;
5034 }
5035 }
5036 if (!cset) {
5037 it->task_pos = NULL;
5038 return;
5039 }
5040 it->task_pos = it->cur_tasks_head->next;
5041
5042 /*
5043 * We don't keep css_sets locked across iteration steps and thus
5044 * need to take steps to ensure that iteration can be resumed after
5045 * the lock is re-acquired. Iteration is performed at two levels -
5046 * css_sets and tasks in them.
5047 *
5048 * Once created, a css_set never leaves its cgroup lists, so a
5049 * pinned css_set is guaranteed to stay put and we can resume
5050 * iteration afterwards.
5051 *
5052 * Tasks may leave @cset across iteration steps. This is resolved
5053 * by registering each iterator with the css_set currently being
5054 * walked and making css_set_move_task() advance iterators whose
5055 * next task is leaving.
5056 */
5057 if (it->cur_cset) {
5058 list_del(entry: &it->iters_node);
5059 put_css_set_locked(cset: it->cur_cset);
5060 }
5061 get_css_set(cset);
5062 it->cur_cset = cset;
5063 list_add(new: &it->iters_node, head: &cset->task_iters);
5064}
5065
5066static void css_task_iter_skip(struct css_task_iter *it,
5067 struct task_struct *task)
5068{
5069 lockdep_assert_held(&css_set_lock);
5070
5071 if (it->task_pos == &task->cg_list) {
5072 it->task_pos = it->task_pos->next;
5073 it->flags |= CSS_TASK_ITER_SKIPPED;
5074 }
5075}
5076
5077static void css_task_iter_advance(struct css_task_iter *it)
5078{
5079 struct task_struct *task;
5080
5081 lockdep_assert_held(&css_set_lock);
5082repeat:
5083 if (it->task_pos) {
5084 /*
5085 * Advance iterator to find next entry. We go through cset
5086 * tasks, mg_tasks and dying_tasks, when consumed we move onto
5087 * the next cset.
5088 */
5089 if (it->flags & CSS_TASK_ITER_SKIPPED)
5090 it->flags &= ~CSS_TASK_ITER_SKIPPED;
5091 else
5092 it->task_pos = it->task_pos->next;
5093
5094 if (it->task_pos == &it->cur_cset->tasks) {
5095 it->cur_tasks_head = &it->cur_cset->mg_tasks;
5096 it->task_pos = it->cur_tasks_head->next;
5097 }
5098 if (it->task_pos == &it->cur_cset->mg_tasks) {
5099 it->cur_tasks_head = &it->cur_cset->dying_tasks;
5100 it->task_pos = it->cur_tasks_head->next;
5101 }
5102 if (it->task_pos == &it->cur_cset->dying_tasks)
5103 css_task_iter_advance_css_set(it);
5104 } else {
5105 /* called from start, proceed to the first cset */
5106 css_task_iter_advance_css_set(it);
5107 }
5108
5109 if (!it->task_pos)
5110 return;
5111
5112 task = list_entry(it->task_pos, struct task_struct, cg_list);
5113
5114 if (it->flags & CSS_TASK_ITER_PROCS) {
5115 /* if PROCS, skip over tasks which aren't group leaders */
5116 if (!thread_group_leader(p: task))
5117 goto repeat;
5118
5119 /* and dying leaders w/o live member threads */
5120 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
5121 !atomic_read(v: &task->signal->live))
5122 goto repeat;
5123 } else {
5124 /* skip all dying ones */
5125 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
5126 goto repeat;
5127 }
5128}
5129
5130/**
5131 * css_task_iter_start - initiate task iteration
5132 * @css: the css to walk tasks of
5133 * @flags: CSS_TASK_ITER_* flags
5134 * @it: the task iterator to use
5135 *
5136 * Initiate iteration through the tasks of @css. The caller can call
5137 * css_task_iter_next() to walk through the tasks until the function
5138 * returns NULL. On completion of iteration, css_task_iter_end() must be
5139 * called.
5140 */
5141void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
5142 struct css_task_iter *it)
5143{
5144 unsigned long irqflags;
5145
5146 memset(s: it, c: 0, n: sizeof(*it));
5147
5148 spin_lock_irqsave(&css_set_lock, irqflags);
5149
5150 it->ss = css->ss;
5151 it->flags = flags;
5152
5153 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
5154 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
5155 else
5156 it->cset_pos = &css->cgroup->cset_links;
5157
5158 it->cset_head = it->cset_pos;
5159
5160 css_task_iter_advance(it);
5161
5162 spin_unlock_irqrestore(lock: &css_set_lock, flags: irqflags);
5163}
5164
5165/**
5166 * css_task_iter_next - return the next task for the iterator
5167 * @it: the task iterator being iterated
5168 *
5169 * The "next" function for task iteration. @it should have been
5170 * initialized via css_task_iter_start(). Returns NULL when the iteration
5171 * reaches the end.
5172 */
5173struct task_struct *css_task_iter_next(struct css_task_iter *it)
5174{
5175 unsigned long irqflags;
5176
5177 if (it->cur_task) {
5178 put_task_struct(t: it->cur_task);
5179 it->cur_task = NULL;
5180 }
5181
5182 spin_lock_irqsave(&css_set_lock, irqflags);
5183
5184 /* @it may be half-advanced by skips, finish advancing */
5185 if (it->flags & CSS_TASK_ITER_SKIPPED)
5186 css_task_iter_advance(it);
5187
5188 if (it->task_pos) {
5189 it->cur_task = list_entry(it->task_pos, struct task_struct,
5190 cg_list);
5191 get_task_struct(t: it->cur_task);
5192 css_task_iter_advance(it);
5193 }
5194
5195 spin_unlock_irqrestore(lock: &css_set_lock, flags: irqflags);
5196
5197 return it->cur_task;
5198}
5199
5200/**
5201 * css_task_iter_end - finish task iteration
5202 * @it: the task iterator to finish
5203 *
5204 * Finish task iteration started by css_task_iter_start().
5205 */
5206void css_task_iter_end(struct css_task_iter *it)
5207{
5208 unsigned long irqflags;
5209
5210 if (it->cur_cset) {
5211 spin_lock_irqsave(&css_set_lock, irqflags);
5212 list_del(entry: &it->iters_node);
5213 put_css_set_locked(cset: it->cur_cset);
5214 spin_unlock_irqrestore(lock: &css_set_lock, flags: irqflags);
5215 }
5216
5217 if (it->cur_dcset)
5218 put_css_set(cset: it->cur_dcset);
5219
5220 if (it->cur_task)
5221 put_task_struct(t: it->cur_task);
5222}
5223
5224static void cgroup_procs_release(struct kernfs_open_file *of)
5225{
5226 struct cgroup_file_ctx *ctx = of->priv;
5227
5228 if (ctx->procs.started)
5229 css_task_iter_end(it: &ctx->procs.iter);
5230}
5231
5232static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
5233{
5234 struct kernfs_open_file *of = s->private;
5235 struct cgroup_file_ctx *ctx = of->priv;
5236
5237 if (pos)
5238 (*pos)++;
5239
5240 return css_task_iter_next(it: &ctx->procs.iter);
5241}
5242
5243static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5244 unsigned int iter_flags)
5245{
5246 struct kernfs_open_file *of = s->private;
5247 struct cgroup *cgrp = seq_css(seq: s)->cgroup;
5248 struct cgroup_file_ctx *ctx = of->priv;
5249 struct css_task_iter *it = &ctx->procs.iter;
5250
5251 /*
5252 * When a seq_file is seeked, it's always traversed sequentially
5253 * from position 0, so we can simply keep iterating on !0 *pos.
5254 */
5255 if (!ctx->procs.started) {
5256 if (WARN_ON_ONCE((*pos)))
5257 return ERR_PTR(error: -EINVAL);
5258 css_task_iter_start(css: &cgrp->self, flags: iter_flags, it);
5259 ctx->procs.started = true;
5260 } else if (!(*pos)) {
5261 css_task_iter_end(it);
5262 css_task_iter_start(css: &cgrp->self, flags: iter_flags, it);
5263 } else
5264 return it->cur_task;
5265
5266 return cgroup_procs_next(s, NULL, NULL);
5267}
5268
5269static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5270{
5271 struct cgroup *cgrp = seq_css(seq: s)->cgroup;
5272
5273 /*
5274 * All processes of a threaded subtree belong to the domain cgroup
5275 * of the subtree. Only threads can be distributed across the
5276 * subtree. Reject reads on cgroup.procs in the subtree proper.
5277 * They're always empty anyway.
5278 */
5279 if (cgroup_is_threaded(cgrp))
5280 return ERR_PTR(error: -EOPNOTSUPP);
5281
5282 return __cgroup_procs_start(s, pos, iter_flags: CSS_TASK_ITER_PROCS |
5283 CSS_TASK_ITER_THREADED);
5284}
5285
5286static int cgroup_procs_show(struct seq_file *s, void *v)
5287{
5288 seq_printf(m: s, fmt: "%d\n", task_pid_vnr(tsk: v));
5289 return 0;
5290}
5291
5292static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5293{
5294 int ret;
5295 struct inode *inode;
5296
5297 lockdep_assert_held(&cgroup_mutex);
5298
5299 inode = kernfs_get_inode(sb, kn: cgrp->procs_file.kn);
5300 if (!inode)
5301 return -ENOMEM;
5302
5303 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5304 iput(inode);
5305 return ret;
5306}
5307
5308static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5309 struct cgroup *dst_cgrp,
5310 struct super_block *sb,
5311 struct cgroup_namespace *ns)
5312{
5313 struct cgroup *com_cgrp = src_cgrp;
5314 int ret;
5315
5316 lockdep_assert_held(&cgroup_mutex);
5317
5318 /* find the common ancestor */
5319 while (!cgroup_is_descendant(cgrp: dst_cgrp, ancestor: com_cgrp))
5320 com_cgrp = cgroup_parent(cgrp: com_cgrp);
5321
5322 /* %current should be authorized to migrate to the common ancestor */
5323 ret = cgroup_may_write(cgrp: com_cgrp, sb);
5324 if (ret)
5325 return ret;
5326
5327 /*
5328 * If namespaces are delegation boundaries, %current must be able
5329 * to see both source and destination cgroups from its namespace.
5330 */
5331 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5332 (!cgroup_is_descendant(cgrp: src_cgrp, ancestor: ns->root_cset->dfl_cgrp) ||
5333 !cgroup_is_descendant(cgrp: dst_cgrp, ancestor: ns->root_cset->dfl_cgrp)))
5334 return -ENOENT;
5335
5336 return 0;
5337}
5338
5339static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5340 struct cgroup *dst_cgrp,
5341 struct super_block *sb, bool threadgroup,
5342 struct cgroup_namespace *ns)
5343{
5344 int ret = 0;
5345
5346 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5347 if (ret)
5348 return ret;
5349
5350 ret = cgroup_migrate_vet_dst(dst_cgrp);
5351 if (ret)
5352 return ret;
5353
5354 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5355 ret = -EOPNOTSUPP;
5356
5357 return ret;
5358}
5359
5360static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5361 bool threadgroup)
5362{
5363 struct cgroup_file_ctx *ctx = of->priv;
5364 struct cgroup *src_cgrp, *dst_cgrp;
5365 struct task_struct *task;
5366 const struct cred *saved_cred;
5367 ssize_t ret;
5368 enum cgroup_attach_lock_mode lock_mode;
5369
5370 dst_cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
5371 if (!dst_cgrp)
5372 return -ENODEV;
5373
5374 task = cgroup_procs_write_start(buf, threadgroup, lock_mode: &lock_mode);
5375 ret = PTR_ERR_OR_ZERO(ptr: task);
5376 if (ret)
5377 goto out_unlock;
5378
5379 /* find the source cgroup */
5380 spin_lock_irq(lock: &css_set_lock);
5381 src_cgrp = task_cgroup_from_root(task, root: &cgrp_dfl_root);
5382 spin_unlock_irq(lock: &css_set_lock);
5383
5384 /*
5385 * Process and thread migrations follow same delegation rule. Check
5386 * permissions using the credentials from file open to protect against
5387 * inherited fd attacks.
5388 */
5389 saved_cred = override_creds(override_cred: of->file->f_cred);
5390 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5391 sb: of->file->f_path.dentry->d_sb,
5392 threadgroup, ns: ctx->ns);
5393 revert_creds(revert_cred: saved_cred);
5394 if (ret)
5395 goto out_finish;
5396
5397 ret = cgroup_attach_task(dst_cgrp, leader: task, threadgroup);
5398
5399out_finish:
5400 cgroup_procs_write_finish(task, lock_mode);
5401out_unlock:
5402 cgroup_kn_unlock(kn: of->kn);
5403
5404 return ret;
5405}
5406
5407static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5408 char *buf, size_t nbytes, loff_t off)
5409{
5410 return __cgroup_procs_write(of, buf, threadgroup: true) ?: nbytes;
5411}
5412
5413static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5414{
5415 return __cgroup_procs_start(s, pos, iter_flags: 0);
5416}
5417
5418static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5419 char *buf, size_t nbytes, loff_t off)
5420{
5421 return __cgroup_procs_write(of, buf, threadgroup: false) ?: nbytes;
5422}
5423
5424/* cgroup core interface files for the default hierarchy */
5425static struct cftype cgroup_base_files[] = {
5426 {
5427 .name = "cgroup.type",
5428 .flags = CFTYPE_NOT_ON_ROOT,
5429 .seq_show = cgroup_type_show,
5430 .write = cgroup_type_write,
5431 },
5432 {
5433 .name = "cgroup.procs",
5434 .flags = CFTYPE_NS_DELEGATABLE,
5435 .file_offset = offsetof(struct cgroup, procs_file),
5436 .release = cgroup_procs_release,
5437 .seq_start = cgroup_procs_start,
5438 .seq_next = cgroup_procs_next,
5439 .seq_show = cgroup_procs_show,
5440 .write = cgroup_procs_write,
5441 },
5442 {
5443 .name = "cgroup.threads",
5444 .flags = CFTYPE_NS_DELEGATABLE,
5445 .release = cgroup_procs_release,
5446 .seq_start = cgroup_threads_start,
5447 .seq_next = cgroup_procs_next,
5448 .seq_show = cgroup_procs_show,
5449 .write = cgroup_threads_write,
5450 },
5451 {
5452 .name = "cgroup.controllers",
5453 .seq_show = cgroup_controllers_show,
5454 },
5455 {
5456 .name = "cgroup.subtree_control",
5457 .flags = CFTYPE_NS_DELEGATABLE,
5458 .seq_show = cgroup_subtree_control_show,
5459 .write = cgroup_subtree_control_write,
5460 },
5461 {
5462 .name = "cgroup.events",
5463 .flags = CFTYPE_NOT_ON_ROOT,
5464 .file_offset = offsetof(struct cgroup, events_file),
5465 .seq_show = cgroup_events_show,
5466 },
5467 {
5468 .name = "cgroup.max.descendants",
5469 .seq_show = cgroup_max_descendants_show,
5470 .write = cgroup_max_descendants_write,
5471 },
5472 {
5473 .name = "cgroup.max.depth",
5474 .seq_show = cgroup_max_depth_show,
5475 .write = cgroup_max_depth_write,
5476 },
5477 {
5478 .name = "cgroup.stat",
5479 .seq_show = cgroup_stat_show,
5480 },
5481 {
5482 .name = "cgroup.stat.local",
5483 .flags = CFTYPE_NOT_ON_ROOT,
5484 .seq_show = cgroup_core_local_stat_show,
5485 },
5486 {
5487 .name = "cgroup.freeze",
5488 .flags = CFTYPE_NOT_ON_ROOT,
5489 .seq_show = cgroup_freeze_show,
5490 .write = cgroup_freeze_write,
5491 },
5492 {
5493 .name = "cgroup.kill",
5494 .flags = CFTYPE_NOT_ON_ROOT,
5495 .write = cgroup_kill_write,
5496 },
5497 {
5498 .name = "cpu.stat",
5499 .seq_show = cpu_stat_show,
5500 },
5501 {
5502 .name = "cpu.stat.local",
5503 .seq_show = cpu_local_stat_show,
5504 },
5505 { } /* terminate */
5506};
5507
5508static struct cftype cgroup_psi_files[] = {
5509#ifdef CONFIG_PSI
5510 {
5511 .name = "io.pressure",
5512 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5513 .seq_show = cgroup_io_pressure_show,
5514 .write = cgroup_io_pressure_write,
5515 .poll = cgroup_pressure_poll,
5516 .release = cgroup_pressure_release,
5517 },
5518 {
5519 .name = "memory.pressure",
5520 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5521 .seq_show = cgroup_memory_pressure_show,
5522 .write = cgroup_memory_pressure_write,
5523 .poll = cgroup_pressure_poll,
5524 .release = cgroup_pressure_release,
5525 },
5526 {
5527 .name = "cpu.pressure",
5528 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5529 .seq_show = cgroup_cpu_pressure_show,
5530 .write = cgroup_cpu_pressure_write,
5531 .poll = cgroup_pressure_poll,
5532 .release = cgroup_pressure_release,
5533 },
5534#ifdef CONFIG_IRQ_TIME_ACCOUNTING
5535 {
5536 .name = "irq.pressure",
5537 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5538 .seq_show = cgroup_irq_pressure_show,
5539 .write = cgroup_irq_pressure_write,
5540 .poll = cgroup_pressure_poll,
5541 .release = cgroup_pressure_release,
5542 },
5543#endif
5544 {
5545 .name = "cgroup.pressure",
5546 .seq_show = cgroup_pressure_show,
5547 .write = cgroup_pressure_write,
5548 },
5549#endif /* CONFIG_PSI */
5550 { } /* terminate */
5551};
5552
5553/*
5554 * css destruction is four-stage process.
5555 *
5556 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5557 * Implemented in kill_css().
5558 *
5559 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5560 * and thus css_tryget_online() is guaranteed to fail, the css can be
5561 * offlined by invoking offline_css(). After offlining, the base ref is
5562 * put. Implemented in css_killed_work_fn().
5563 *
5564 * 3. When the percpu_ref reaches zero, the only possible remaining
5565 * accessors are inside RCU read sections. css_release() schedules the
5566 * RCU callback.
5567 *
5568 * 4. After the grace period, the css can be freed. Implemented in
5569 * css_free_rwork_fn().
5570 *
5571 * It is actually hairier because both step 2 and 4 require process context
5572 * and thus involve punting to css->destroy_work adding two additional
5573 * steps to the already complex sequence.
5574 */
5575static void css_free_rwork_fn(struct work_struct *work)
5576{
5577 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5578 struct cgroup_subsys_state, destroy_rwork);
5579 struct cgroup_subsys *ss = css->ss;
5580 struct cgroup *cgrp = css->cgroup;
5581
5582 percpu_ref_exit(ref: &css->refcnt);
5583 css_rstat_exit(css);
5584
5585 if (!css_is_self(css)) {
5586 /* css free path */
5587 struct cgroup_subsys_state *parent = css->parent;
5588 int id = css->id;
5589
5590 ss->css_free(css);
5591 cgroup_idr_remove(idr: &ss->css_idr, id);
5592 cgroup_put(cgrp);
5593
5594 if (parent)
5595 css_put(css: parent);
5596 } else {
5597 /* cgroup free path */
5598 atomic_dec(v: &cgrp->root->nr_cgrps);
5599 if (!cgroup_on_dfl(cgrp))
5600 cgroup1_pidlist_destroy_all(cgrp);
5601 cancel_work_sync(work: &cgrp->release_agent_work);
5602 bpf_cgrp_storage_free(cgroup: cgrp);
5603
5604 if (cgroup_parent(cgrp)) {
5605 /*
5606 * We get a ref to the parent, and put the ref when
5607 * this cgroup is being freed, so it's guaranteed
5608 * that the parent won't be destroyed before its
5609 * children.
5610 */
5611 cgroup_put(cgrp: cgroup_parent(cgrp));
5612 kernfs_put(kn: cgrp->kn);
5613 psi_cgroup_free(cgrp);
5614 kfree(objp: cgrp);
5615 } else {
5616 /*
5617 * This is root cgroup's refcnt reaching zero,
5618 * which indicates that the root should be
5619 * released.
5620 */
5621 cgroup_destroy_root(root: cgrp->root);
5622 }
5623 }
5624}
5625
5626static void css_release_work_fn(struct work_struct *work)
5627{
5628 struct cgroup_subsys_state *css =
5629 container_of(work, struct cgroup_subsys_state, destroy_work);
5630 struct cgroup_subsys *ss = css->ss;
5631 struct cgroup *cgrp = css->cgroup;
5632
5633 cgroup_lock();
5634
5635 css->flags |= CSS_RELEASED;
5636 list_del_rcu(entry: &css->sibling);
5637
5638 if (!css_is_self(css)) {
5639 struct cgroup *parent_cgrp;
5640
5641 css_rstat_flush(css);
5642
5643 cgroup_idr_replace(idr: &ss->css_idr, NULL, id: css->id);
5644 if (ss->css_released)
5645 ss->css_released(css);
5646
5647 cgrp->nr_dying_subsys[ss->id]--;
5648 /*
5649 * When a css is released and ready to be freed, its
5650 * nr_descendants must be zero. However, the corresponding
5651 * cgrp->nr_dying_subsys[ss->id] may not be 0 if a subsystem
5652 * is activated and deactivated multiple times with one or
5653 * more of its previous activation leaving behind dying csses.
5654 */
5655 WARN_ON_ONCE(css->nr_descendants);
5656 parent_cgrp = cgroup_parent(cgrp);
5657 while (parent_cgrp) {
5658 parent_cgrp->nr_dying_subsys[ss->id]--;
5659 parent_cgrp = cgroup_parent(cgrp: parent_cgrp);
5660 }
5661 } else {
5662 struct cgroup *tcgrp;
5663
5664 /* cgroup release path */
5665 TRACE_CGROUP_PATH(release, cgrp);
5666
5667 css_rstat_flush(css: &cgrp->self);
5668
5669 spin_lock_irq(lock: &css_set_lock);
5670 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5671 tcgrp = cgroup_parent(cgrp: tcgrp))
5672 tcgrp->nr_dying_descendants--;
5673 spin_unlock_irq(lock: &css_set_lock);
5674
5675 /*
5676 * There are two control paths which try to determine
5677 * cgroup from dentry without going through kernfs -
5678 * cgroupstats_build() and css_tryget_online_from_dir().
5679 * Those are supported by RCU protecting clearing of
5680 * cgrp->kn->priv backpointer.
5681 */
5682 if (cgrp->kn)
5683 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5684 NULL);
5685 }
5686
5687 cgroup_unlock();
5688
5689 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5690 queue_rcu_work(wq: cgroup_free_wq, rwork: &css->destroy_rwork);
5691}
5692
5693static void css_release(struct percpu_ref *ref)
5694{
5695 struct cgroup_subsys_state *css =
5696 container_of(ref, struct cgroup_subsys_state, refcnt);
5697
5698 INIT_WORK(&css->destroy_work, css_release_work_fn);
5699 queue_work(wq: cgroup_release_wq, work: &css->destroy_work);
5700}
5701
5702static void init_and_link_css(struct cgroup_subsys_state *css,
5703 struct cgroup_subsys *ss, struct cgroup *cgrp)
5704{
5705 lockdep_assert_held(&cgroup_mutex);
5706
5707 cgroup_get_live(cgrp);
5708
5709 memset(s: css, c: 0, n: sizeof(*css));
5710 css->cgroup = cgrp;
5711 css->ss = ss;
5712 css->id = -1;
5713 INIT_LIST_HEAD(list: &css->sibling);
5714 INIT_LIST_HEAD(list: &css->children);
5715 css->serial_nr = css_serial_nr_next++;
5716 atomic_set(v: &css->online_cnt, i: 0);
5717
5718 if (cgroup_parent(cgrp)) {
5719 css->parent = cgroup_css(cgrp: cgroup_parent(cgrp), ss);
5720 css_get(css: css->parent);
5721 }
5722
5723 BUG_ON(cgroup_css(cgrp, ss));
5724}
5725
5726/* invoke ->css_online() on a new CSS and mark it online if successful */
5727static int online_css(struct cgroup_subsys_state *css)
5728{
5729 struct cgroup_subsys *ss = css->ss;
5730 int ret = 0;
5731
5732 lockdep_assert_held(&cgroup_mutex);
5733
5734 if (ss->css_online)
5735 ret = ss->css_online(css);
5736 if (!ret) {
5737 css->flags |= CSS_ONLINE;
5738 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5739
5740 atomic_inc(v: &css->online_cnt);
5741 if (css->parent) {
5742 atomic_inc(v: &css->parent->online_cnt);
5743 while ((css = css->parent))
5744 css->nr_descendants++;
5745 }
5746 }
5747 return ret;
5748}
5749
5750/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5751static void offline_css(struct cgroup_subsys_state *css)
5752{
5753 struct cgroup_subsys *ss = css->ss;
5754
5755 lockdep_assert_held(&cgroup_mutex);
5756
5757 if (!(css->flags & CSS_ONLINE))
5758 return;
5759
5760 if (ss->css_offline)
5761 ss->css_offline(css);
5762
5763 css->flags &= ~CSS_ONLINE;
5764 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5765
5766 wake_up_all(&css->cgroup->offline_waitq);
5767
5768 css->cgroup->nr_dying_subsys[ss->id]++;
5769 /*
5770 * Parent css and cgroup cannot be freed until after the freeing
5771 * of child css, see css_free_rwork_fn().
5772 */
5773 while ((css = css->parent)) {
5774 css->nr_descendants--;
5775 css->cgroup->nr_dying_subsys[ss->id]++;
5776 }
5777}
5778
5779/**
5780 * css_create - create a cgroup_subsys_state
5781 * @cgrp: the cgroup new css will be associated with
5782 * @ss: the subsys of new css
5783 *
5784 * Create a new css associated with @cgrp - @ss pair. On success, the new
5785 * css is online and installed in @cgrp. This function doesn't create the
5786 * interface files. Returns 0 on success, -errno on failure.
5787 */
5788static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5789 struct cgroup_subsys *ss)
5790{
5791 struct cgroup *parent = cgroup_parent(cgrp);
5792 struct cgroup_subsys_state *parent_css = cgroup_css(cgrp: parent, ss);
5793 struct cgroup_subsys_state *css;
5794 int err;
5795
5796 lockdep_assert_held(&cgroup_mutex);
5797
5798 css = ss->css_alloc(parent_css);
5799 if (!css)
5800 css = ERR_PTR(error: -ENOMEM);
5801 if (IS_ERR(ptr: css))
5802 return css;
5803
5804 init_and_link_css(css, ss, cgrp);
5805
5806 err = percpu_ref_init(ref: &css->refcnt, release: css_release, flags: 0, GFP_KERNEL);
5807 if (err)
5808 goto err_free_css;
5809
5810 err = cgroup_idr_alloc(idr: &ss->css_idr, NULL, start: 2, end: 0, GFP_KERNEL);
5811 if (err < 0)
5812 goto err_free_css;
5813 css->id = err;
5814
5815 err = css_rstat_init(css);
5816 if (err)
5817 goto err_free_css;
5818
5819 /* @css is ready to be brought online now, make it visible */
5820 list_add_tail_rcu(new: &css->sibling, head: &parent_css->children);
5821 cgroup_idr_replace(idr: &ss->css_idr, ptr: css, id: css->id);
5822
5823 err = online_css(css);
5824 if (err)
5825 goto err_list_del;
5826
5827 return css;
5828
5829err_list_del:
5830 list_del_rcu(entry: &css->sibling);
5831err_free_css:
5832 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5833 queue_rcu_work(wq: cgroup_free_wq, rwork: &css->destroy_rwork);
5834 return ERR_PTR(error: err);
5835}
5836
5837/*
5838 * The returned cgroup is fully initialized including its control mask, but
5839 * it doesn't have the control mask applied.
5840 */
5841static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5842 umode_t mode)
5843{
5844 struct cgroup_root *root = parent->root;
5845 struct cgroup *cgrp, *tcgrp;
5846 struct kernfs_node *kn;
5847 int i, level = parent->level + 1;
5848 int ret;
5849
5850 /* allocate the cgroup and its ID, 0 is reserved for the root */
5851 cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5852 if (!cgrp)
5853 return ERR_PTR(error: -ENOMEM);
5854
5855 ret = percpu_ref_init(ref: &cgrp->self.refcnt, release: css_release, flags: 0, GFP_KERNEL);
5856 if (ret)
5857 goto out_free_cgrp;
5858
5859 /* create the directory */
5860 kn = kernfs_create_dir_ns(parent: parent->kn, name, mode,
5861 current_fsuid(), current_fsgid(),
5862 priv: cgrp, NULL);
5863 if (IS_ERR(ptr: kn)) {
5864 ret = PTR_ERR(ptr: kn);
5865 goto out_cancel_ref;
5866 }
5867 cgrp->kn = kn;
5868
5869 init_cgroup_housekeeping(cgrp);
5870
5871 cgrp->self.parent = &parent->self;
5872 cgrp->root = root;
5873 cgrp->level = level;
5874
5875 /*
5876 * Now that init_cgroup_housekeeping() has been called and cgrp->self
5877 * is setup, it is safe to perform rstat initialization on it.
5878 */
5879 ret = css_rstat_init(css: &cgrp->self);
5880 if (ret)
5881 goto out_kernfs_remove;
5882
5883 ret = psi_cgroup_alloc(cgrp);
5884 if (ret)
5885 goto out_stat_exit;
5886
5887 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(cgrp: tcgrp))
5888 cgrp->ancestors[tcgrp->level] = tcgrp;
5889
5890 /*
5891 * New cgroup inherits effective freeze counter, and
5892 * if the parent has to be frozen, the child has too.
5893 */
5894 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5895 seqcount_init(&cgrp->freezer.freeze_seq);
5896 if (cgrp->freezer.e_freeze) {
5897 /*
5898 * Set the CGRP_FREEZE flag, so when a process will be
5899 * attached to the child cgroup, it will become frozen.
5900 * At this point the new cgroup is unpopulated, so we can
5901 * consider it frozen immediately.
5902 */
5903 set_bit(nr: CGRP_FREEZE, addr: &cgrp->flags);
5904 cgrp->freezer.freeze_start_nsec = ktime_get_ns();
5905 set_bit(nr: CGRP_FROZEN, addr: &cgrp->flags);
5906 }
5907
5908 if (notify_on_release(cgrp: parent))
5909 set_bit(nr: CGRP_NOTIFY_ON_RELEASE, addr: &cgrp->flags);
5910
5911 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5912 set_bit(nr: CGRP_CPUSET_CLONE_CHILDREN, addr: &cgrp->flags);
5913
5914 cgrp->self.serial_nr = css_serial_nr_next++;
5915
5916 ret = blocking_notifier_call_chain_robust(nh: &cgroup_lifetime_notifier,
5917 val_up: CGROUP_LIFETIME_ONLINE,
5918 val_down: CGROUP_LIFETIME_OFFLINE, v: cgrp);
5919 ret = notifier_to_errno(ret);
5920 if (ret)
5921 goto out_psi_free;
5922
5923 /* allocation complete, commit to creation */
5924 spin_lock_irq(lock: &css_set_lock);
5925 for (i = 0; i < level; i++) {
5926 tcgrp = cgrp->ancestors[i];
5927 tcgrp->nr_descendants++;
5928
5929 /*
5930 * If the new cgroup is frozen, all ancestor cgroups get a new
5931 * frozen descendant, but their state can't change because of
5932 * this.
5933 */
5934 if (cgrp->freezer.e_freeze)
5935 tcgrp->freezer.nr_frozen_descendants++;
5936 }
5937 spin_unlock_irq(lock: &css_set_lock);
5938
5939 list_add_tail_rcu(new: &cgrp->self.sibling, head: &cgroup_parent(cgrp)->self.children);
5940 atomic_inc(v: &root->nr_cgrps);
5941 cgroup_get_live(cgrp: parent);
5942
5943 /*
5944 * On the default hierarchy, a child doesn't automatically inherit
5945 * subtree_control from the parent. Each is configured manually.
5946 */
5947 if (!cgroup_on_dfl(cgrp))
5948 cgrp->subtree_control = cgroup_control(cgrp);
5949
5950 cgroup_propagate_control(cgrp);
5951
5952 return cgrp;
5953
5954out_psi_free:
5955 psi_cgroup_free(cgrp);
5956out_stat_exit:
5957 css_rstat_exit(css: &cgrp->self);
5958out_kernfs_remove:
5959 kernfs_remove(kn: cgrp->kn);
5960out_cancel_ref:
5961 percpu_ref_exit(ref: &cgrp->self.refcnt);
5962out_free_cgrp:
5963 kfree(objp: cgrp);
5964 return ERR_PTR(error: ret);
5965}
5966
5967static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5968{
5969 struct cgroup *cgroup;
5970 int ret = false;
5971 int level = 0;
5972
5973 lockdep_assert_held(&cgroup_mutex);
5974
5975 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgrp: cgroup)) {
5976 if (cgroup->nr_descendants >= cgroup->max_descendants)
5977 goto fail;
5978
5979 if (level >= cgroup->max_depth)
5980 goto fail;
5981
5982 level++;
5983 }
5984
5985 ret = true;
5986fail:
5987 return ret;
5988}
5989
5990int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5991{
5992 struct cgroup *parent, *cgrp;
5993 int ret;
5994
5995 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5996 if (strchr(name, '\n'))
5997 return -EINVAL;
5998
5999 parent = cgroup_kn_lock_live(kn: parent_kn, drain_offline: false);
6000 if (!parent)
6001 return -ENODEV;
6002
6003 if (!cgroup_check_hierarchy_limits(parent)) {
6004 ret = -EAGAIN;
6005 goto out_unlock;
6006 }
6007
6008 cgrp = cgroup_create(parent, name, mode);
6009 if (IS_ERR(ptr: cgrp)) {
6010 ret = PTR_ERR(ptr: cgrp);
6011 goto out_unlock;
6012 }
6013
6014 /*
6015 * This extra ref will be put in css_free_rwork_fn() and guarantees
6016 * that @cgrp->kn is always accessible.
6017 */
6018 kernfs_get(kn: cgrp->kn);
6019
6020 ret = css_populate_dir(css: &cgrp->self);
6021 if (ret)
6022 goto out_destroy;
6023
6024 ret = cgroup_apply_control_enable(cgrp);
6025 if (ret)
6026 goto out_destroy;
6027
6028 TRACE_CGROUP_PATH(mkdir, cgrp);
6029
6030 /* let's create and online css's */
6031 kernfs_activate(kn: cgrp->kn);
6032
6033 ret = 0;
6034 goto out_unlock;
6035
6036out_destroy:
6037 cgroup_destroy_locked(cgrp);
6038out_unlock:
6039 cgroup_kn_unlock(kn: parent_kn);
6040 return ret;
6041}
6042
6043/*
6044 * This is called when the refcnt of a css is confirmed to be killed.
6045 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
6046 * initiate destruction and put the css ref from kill_css().
6047 */
6048static void css_killed_work_fn(struct work_struct *work)
6049{
6050 struct cgroup_subsys_state *css =
6051 container_of(work, struct cgroup_subsys_state, destroy_work);
6052
6053 cgroup_lock();
6054
6055 do {
6056 offline_css(css);
6057 css_put(css);
6058 /* @css can't go away while we're holding cgroup_mutex */
6059 css = css->parent;
6060 } while (css && atomic_dec_and_test(v: &css->online_cnt));
6061
6062 cgroup_unlock();
6063}
6064
6065/* css kill confirmation processing requires process context, bounce */
6066static void css_killed_ref_fn(struct percpu_ref *ref)
6067{
6068 struct cgroup_subsys_state *css =
6069 container_of(ref, struct cgroup_subsys_state, refcnt);
6070
6071 if (atomic_dec_and_test(v: &css->online_cnt)) {
6072 INIT_WORK(&css->destroy_work, css_killed_work_fn);
6073 queue_work(wq: cgroup_offline_wq, work: &css->destroy_work);
6074 }
6075}
6076
6077/**
6078 * kill_css - destroy a css
6079 * @css: css to destroy
6080 *
6081 * This function initiates destruction of @css by removing cgroup interface
6082 * files and putting its base reference. ->css_offline() will be invoked
6083 * asynchronously once css_tryget_online() is guaranteed to fail and when
6084 * the reference count reaches zero, @css will be released.
6085 */
6086static void kill_css(struct cgroup_subsys_state *css)
6087{
6088 lockdep_assert_held(&cgroup_mutex);
6089
6090 if (css->flags & CSS_DYING)
6091 return;
6092
6093 /*
6094 * Call css_killed(), if defined, before setting the CSS_DYING flag
6095 */
6096 if (css->ss->css_killed)
6097 css->ss->css_killed(css);
6098
6099 css->flags |= CSS_DYING;
6100
6101 /*
6102 * This must happen before css is disassociated with its cgroup.
6103 * See seq_css() for details.
6104 */
6105 css_clear_dir(css);
6106
6107 /*
6108 * Killing would put the base ref, but we need to keep it alive
6109 * until after ->css_offline().
6110 */
6111 css_get(css);
6112
6113 /*
6114 * cgroup core guarantees that, by the time ->css_offline() is
6115 * invoked, no new css reference will be given out via
6116 * css_tryget_online(). We can't simply call percpu_ref_kill() and
6117 * proceed to offlining css's because percpu_ref_kill() doesn't
6118 * guarantee that the ref is seen as killed on all CPUs on return.
6119 *
6120 * Use percpu_ref_kill_and_confirm() to get notifications as each
6121 * css is confirmed to be seen as killed on all CPUs.
6122 */
6123 percpu_ref_kill_and_confirm(ref: &css->refcnt, confirm_kill: css_killed_ref_fn);
6124}
6125
6126/**
6127 * cgroup_destroy_locked - the first stage of cgroup destruction
6128 * @cgrp: cgroup to be destroyed
6129 *
6130 * css's make use of percpu refcnts whose killing latency shouldn't be
6131 * exposed to userland and are RCU protected. Also, cgroup core needs to
6132 * guarantee that css_tryget_online() won't succeed by the time
6133 * ->css_offline() is invoked. To satisfy all the requirements,
6134 * destruction is implemented in the following two steps.
6135 *
6136 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
6137 * userland visible parts and start killing the percpu refcnts of
6138 * css's. Set up so that the next stage will be kicked off once all
6139 * the percpu refcnts are confirmed to be killed.
6140 *
6141 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
6142 * rest of destruction. Once all cgroup references are gone, the
6143 * cgroup is RCU-freed.
6144 *
6145 * This function implements s1. After this step, @cgrp is gone as far as
6146 * the userland is concerned and a new cgroup with the same name may be
6147 * created. As cgroup doesn't care about the names internally, this
6148 * doesn't cause any problem.
6149 */
6150static int cgroup_destroy_locked(struct cgroup *cgrp)
6151 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
6152{
6153 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
6154 struct cgroup_subsys_state *css;
6155 struct cgrp_cset_link *link;
6156 int ssid, ret;
6157
6158 lockdep_assert_held(&cgroup_mutex);
6159
6160 /*
6161 * Only migration can raise populated from zero and we're already
6162 * holding cgroup_mutex.
6163 */
6164 if (cgroup_is_populated(cgrp))
6165 return -EBUSY;
6166
6167 /*
6168 * Make sure there's no live children. We can't test emptiness of
6169 * ->self.children as dead children linger on it while being
6170 * drained; otherwise, "rmdir parent/child parent" may fail.
6171 */
6172 if (css_has_online_children(css: &cgrp->self))
6173 return -EBUSY;
6174
6175 /*
6176 * Mark @cgrp and the associated csets dead. The former prevents
6177 * further task migration and child creation by disabling
6178 * cgroup_kn_lock_live(). The latter makes the csets ignored by
6179 * the migration path.
6180 */
6181 cgrp->self.flags &= ~CSS_ONLINE;
6182
6183 spin_lock_irq(lock: &css_set_lock);
6184 list_for_each_entry(link, &cgrp->cset_links, cset_link)
6185 link->cset->dead = true;
6186 spin_unlock_irq(lock: &css_set_lock);
6187
6188 /* initiate massacre of all css's */
6189 for_each_css(css, ssid, cgrp)
6190 kill_css(css);
6191
6192 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
6193 css_clear_dir(css: &cgrp->self);
6194 kernfs_remove(kn: cgrp->kn);
6195
6196 if (cgroup_is_threaded(cgrp))
6197 parent->nr_threaded_children--;
6198
6199 spin_lock_irq(lock: &css_set_lock);
6200 for (tcgrp = parent; tcgrp; tcgrp = cgroup_parent(cgrp: tcgrp)) {
6201 tcgrp->nr_descendants--;
6202 tcgrp->nr_dying_descendants++;
6203 /*
6204 * If the dying cgroup is frozen, decrease frozen descendants
6205 * counters of ancestor cgroups.
6206 */
6207 if (test_bit(CGRP_FROZEN, &cgrp->flags))
6208 tcgrp->freezer.nr_frozen_descendants--;
6209 }
6210 spin_unlock_irq(lock: &css_set_lock);
6211
6212 cgroup1_check_for_release(cgrp: parent);
6213
6214 ret = blocking_notifier_call_chain(nh: &cgroup_lifetime_notifier,
6215 val: CGROUP_LIFETIME_OFFLINE, v: cgrp);
6216 WARN_ON_ONCE(notifier_to_errno(ret));
6217
6218 /* put the base reference */
6219 percpu_ref_kill(ref: &cgrp->self.refcnt);
6220
6221 return 0;
6222};
6223
6224int cgroup_rmdir(struct kernfs_node *kn)
6225{
6226 struct cgroup *cgrp;
6227 int ret = 0;
6228
6229 cgrp = cgroup_kn_lock_live(kn, drain_offline: false);
6230 if (!cgrp)
6231 return 0;
6232
6233 ret = cgroup_destroy_locked(cgrp);
6234 if (!ret)
6235 TRACE_CGROUP_PATH(rmdir, cgrp);
6236
6237 cgroup_kn_unlock(kn);
6238 return ret;
6239}
6240
6241static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
6242 .show_options = cgroup_show_options,
6243 .mkdir = cgroup_mkdir,
6244 .rmdir = cgroup_rmdir,
6245 .show_path = cgroup_show_path,
6246};
6247
6248static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
6249{
6250 struct cgroup_subsys_state *css;
6251
6252 pr_debug("Initializing cgroup subsys %s\n", ss->name);
6253
6254 cgroup_lock();
6255
6256 idr_init(idr: &ss->css_idr);
6257 INIT_LIST_HEAD(list: &ss->cfts);
6258
6259 /* Create the root cgroup state for this subsystem */
6260 ss->root = &cgrp_dfl_root;
6261 css = ss->css_alloc(NULL);
6262 /* We don't handle early failures gracefully */
6263 BUG_ON(IS_ERR(css));
6264 init_and_link_css(css, ss, cgrp: &cgrp_dfl_root.cgrp);
6265
6266 /*
6267 * Root csses are never destroyed and we can't initialize
6268 * percpu_ref during early init. Disable refcnting.
6269 */
6270 css->flags |= CSS_NO_REF;
6271
6272 if (early) {
6273 /* allocation can't be done safely during early init */
6274 css->id = 1;
6275 } else {
6276 css->id = cgroup_idr_alloc(idr: &ss->css_idr, ptr: css, start: 1, end: 2, GFP_KERNEL);
6277 BUG_ON(css->id < 0);
6278
6279 BUG_ON(ss_rstat_init(ss));
6280 BUG_ON(css_rstat_init(css));
6281 }
6282
6283 /* Update the init_css_set to contain a subsys
6284 * pointer to this state - since the subsystem is
6285 * newly registered, all tasks and hence the
6286 * init_css_set is in the subsystem's root cgroup. */
6287 init_css_set.subsys[ss->id] = css;
6288
6289 have_fork_callback |= (bool)ss->fork << ss->id;
6290 have_exit_callback |= (bool)ss->exit << ss->id;
6291 have_release_callback |= (bool)ss->release << ss->id;
6292 have_canfork_callback |= (bool)ss->can_fork << ss->id;
6293
6294 /* At system boot, before all subsystems have been
6295 * registered, no tasks have been forked, so we don't
6296 * need to invoke fork callbacks here. */
6297 BUG_ON(!list_empty(&init_task.tasks));
6298
6299 BUG_ON(online_css(css));
6300
6301 cgroup_unlock();
6302}
6303
6304/**
6305 * cgroup_init_early - cgroup initialization at system boot
6306 *
6307 * Initialize cgroups at system boot, and initialize any
6308 * subsystems that request early init.
6309 */
6310int __init cgroup_init_early(void)
6311{
6312 static struct cgroup_fs_context __initdata ctx;
6313 struct cgroup_subsys *ss;
6314 int i;
6315
6316 ctx.root = &cgrp_dfl_root;
6317 init_cgroup_root(ctx: &ctx);
6318 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6319
6320 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6321
6322 for_each_subsys(ss, i) {
6323 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6324 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6325 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6326 ss->id, ss->name);
6327 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6328 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6329 WARN(ss->early_init && ss->css_rstat_flush,
6330 "cgroup rstat cannot be used with early init subsystem\n");
6331
6332 ss->id = i;
6333 ss->name = cgroup_subsys_name[i];
6334 if (!ss->legacy_name)
6335 ss->legacy_name = cgroup_subsys_name[i];
6336
6337 if (ss->early_init)
6338 cgroup_init_subsys(ss, early: true);
6339 }
6340 return 0;
6341}
6342
6343/**
6344 * cgroup_init - cgroup initialization
6345 *
6346 * Register cgroup filesystem and /proc file, and initialize
6347 * any subsystems that didn't request early init.
6348 */
6349int __init cgroup_init(void)
6350{
6351 struct cgroup_subsys *ss;
6352 int ssid;
6353
6354 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6355 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6356 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6357 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6358
6359 BUG_ON(ss_rstat_init(NULL));
6360
6361 get_user_ns(ns: init_cgroup_ns.user_ns);
6362
6363 cgroup_lock();
6364
6365 /*
6366 * Add init_css_set to the hash table so that dfl_root can link to
6367 * it during init.
6368 */
6369 hash_add(css_set_table, &init_css_set.hlist,
6370 css_set_hash(init_css_set.subsys));
6371
6372 cgroup_bpf_lifetime_notifier_init();
6373
6374 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6375
6376 cgroup_unlock();
6377
6378 for_each_subsys(ss, ssid) {
6379 if (ss->early_init) {
6380 struct cgroup_subsys_state *css =
6381 init_css_set.subsys[ss->id];
6382
6383 css->id = cgroup_idr_alloc(idr: &ss->css_idr, ptr: css, start: 1, end: 2,
6384 GFP_KERNEL);
6385 BUG_ON(css->id < 0);
6386 } else {
6387 cgroup_init_subsys(ss, early: false);
6388 }
6389
6390 list_add_tail(new: &init_css_set.e_cset_node[ssid],
6391 head: &cgrp_dfl_root.cgrp.e_csets[ssid]);
6392
6393 /*
6394 * Setting dfl_root subsys_mask needs to consider the
6395 * disabled flag and cftype registration needs kmalloc,
6396 * both of which aren't available during early_init.
6397 */
6398 if (!cgroup_ssid_enabled(ssid))
6399 continue;
6400
6401 if (cgroup1_ssid_disabled(ssid))
6402 pr_info("Disabling %s control group subsystem in v1 mounts\n",
6403 ss->legacy_name);
6404
6405 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6406
6407 /* implicit controllers must be threaded too */
6408 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6409
6410 if (ss->implicit_on_dfl)
6411 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6412 else if (!ss->dfl_cftypes)
6413 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6414
6415 if (ss->threaded)
6416 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6417
6418 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6419 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6420 } else {
6421 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6422 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6423 }
6424
6425 if (ss->bind)
6426 ss->bind(init_css_set.subsys[ssid]);
6427
6428 cgroup_lock();
6429 css_populate_dir(css: init_css_set.subsys[ssid]);
6430 cgroup_unlock();
6431 }
6432
6433 /* init_css_set.subsys[] has been updated, re-hash */
6434 hash_del(node: &init_css_set.hlist);
6435 hash_add(css_set_table, &init_css_set.hlist,
6436 css_set_hash(init_css_set.subsys));
6437
6438 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6439 WARN_ON(register_filesystem(&cgroup_fs_type));
6440 WARN_ON(register_filesystem(&cgroup2_fs_type));
6441 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6442#ifdef CONFIG_CPUSETS_V1
6443 WARN_ON(register_filesystem(&cpuset_fs_type));
6444#endif
6445
6446 ns_tree_add(&init_cgroup_ns);
6447 return 0;
6448}
6449
6450static int __init cgroup_wq_init(void)
6451{
6452 /*
6453 * There isn't much point in executing destruction path in
6454 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6455 * Use 1 for @max_active.
6456 *
6457 * We would prefer to do this in cgroup_init() above, but that
6458 * is called before init_workqueues(): so leave this until after.
6459 */
6460 cgroup_offline_wq = alloc_workqueue("cgroup_offline", WQ_PERCPU, 1);
6461 BUG_ON(!cgroup_offline_wq);
6462
6463 cgroup_release_wq = alloc_workqueue("cgroup_release", WQ_PERCPU, 1);
6464 BUG_ON(!cgroup_release_wq);
6465
6466 cgroup_free_wq = alloc_workqueue("cgroup_free", WQ_PERCPU, 1);
6467 BUG_ON(!cgroup_free_wq);
6468 return 0;
6469}
6470core_initcall(cgroup_wq_init);
6471
6472void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6473{
6474 struct kernfs_node *kn;
6475
6476 kn = kernfs_find_and_get_node_by_id(root: cgrp_dfl_root.kf_root, id);
6477 if (!kn)
6478 return;
6479 kernfs_path(kn, buf, buflen);
6480 kernfs_put(kn);
6481}
6482
6483/*
6484 * __cgroup_get_from_id : get the cgroup associated with cgroup id
6485 * @id: cgroup id
6486 * On success return the cgrp or ERR_PTR on failure
6487 * There are no cgroup NS restrictions.
6488 */
6489struct cgroup *__cgroup_get_from_id(u64 id)
6490{
6491 struct kernfs_node *kn;
6492 struct cgroup *cgrp;
6493
6494 kn = kernfs_find_and_get_node_by_id(root: cgrp_dfl_root.kf_root, id);
6495 if (!kn)
6496 return ERR_PTR(error: -ENOENT);
6497
6498 if (kernfs_type(kn) != KERNFS_DIR) {
6499 kernfs_put(kn);
6500 return ERR_PTR(error: -ENOENT);
6501 }
6502
6503 rcu_read_lock();
6504
6505 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6506 if (cgrp && !cgroup_tryget(cgrp))
6507 cgrp = NULL;
6508
6509 rcu_read_unlock();
6510 kernfs_put(kn);
6511
6512 if (!cgrp)
6513 return ERR_PTR(error: -ENOENT);
6514 return cgrp;
6515}
6516
6517/*
6518 * cgroup_get_from_id : get the cgroup associated with cgroup id
6519 * @id: cgroup id
6520 * On success return the cgrp or ERR_PTR on failure
6521 * Only cgroups within current task's cgroup NS are valid.
6522 */
6523struct cgroup *cgroup_get_from_id(u64 id)
6524{
6525 struct cgroup *cgrp, *root_cgrp;
6526
6527 cgrp = __cgroup_get_from_id(id);
6528 if (IS_ERR(ptr: cgrp))
6529 return cgrp;
6530
6531 root_cgrp = current_cgns_cgroup_dfl();
6532 if (!cgroup_is_descendant(cgrp, ancestor: root_cgrp)) {
6533 cgroup_put(cgrp);
6534 return ERR_PTR(error: -ENOENT);
6535 }
6536
6537 return cgrp;
6538}
6539EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6540
6541/*
6542 * proc_cgroup_show()
6543 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6544 * - Used for /proc/<pid>/cgroup.
6545 */
6546int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6547 struct pid *pid, struct task_struct *tsk)
6548{
6549 char *buf;
6550 int retval;
6551 struct cgroup_root *root;
6552
6553 retval = -ENOMEM;
6554 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6555 if (!buf)
6556 goto out;
6557
6558 rcu_read_lock();
6559 spin_lock_irq(lock: &css_set_lock);
6560
6561 for_each_root(root) {
6562 struct cgroup_subsys *ss;
6563 struct cgroup *cgrp;
6564 int ssid, count = 0;
6565
6566 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6567 continue;
6568
6569 cgrp = task_cgroup_from_root(task: tsk, root);
6570 /* The root has already been unmounted. */
6571 if (!cgrp)
6572 continue;
6573
6574 seq_printf(m, fmt: "%d:", root->hierarchy_id);
6575 if (root != &cgrp_dfl_root)
6576 for_each_subsys(ss, ssid)
6577 if (root->subsys_mask & (1 << ssid))
6578 seq_printf(m, fmt: "%s%s", count++ ? "," : "",
6579 ss->legacy_name);
6580 if (strlen(root->name))
6581 seq_printf(m, fmt: "%sname=%s", count ? "," : "",
6582 root->name);
6583 seq_putc(m, c: ':');
6584 /*
6585 * On traditional hierarchies, all zombie tasks show up as
6586 * belonging to the root cgroup. On the default hierarchy,
6587 * while a zombie doesn't show up in "cgroup.procs" and
6588 * thus can't be migrated, its /proc/PID/cgroup keeps
6589 * reporting the cgroup it belonged to before exiting. If
6590 * the cgroup is removed before the zombie is reaped,
6591 * " (deleted)" is appended to the cgroup path.
6592 */
6593 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6594 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6595 current->nsproxy->cgroup_ns);
6596 if (retval == -E2BIG)
6597 retval = -ENAMETOOLONG;
6598 if (retval < 0)
6599 goto out_unlock;
6600
6601 seq_puts(m, s: buf);
6602 } else {
6603 seq_puts(m, s: "/");
6604 }
6605
6606 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6607 seq_puts(m, s: " (deleted)\n");
6608 else
6609 seq_putc(m, c: '\n');
6610 }
6611
6612 retval = 0;
6613out_unlock:
6614 spin_unlock_irq(lock: &css_set_lock);
6615 rcu_read_unlock();
6616 kfree(objp: buf);
6617out:
6618 return retval;
6619}
6620
6621/**
6622 * cgroup_fork - initialize cgroup related fields during copy_process()
6623 * @child: pointer to task_struct of forking parent process.
6624 *
6625 * A task is associated with the init_css_set until cgroup_post_fork()
6626 * attaches it to the target css_set.
6627 */
6628void cgroup_fork(struct task_struct *child)
6629{
6630 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6631 INIT_LIST_HEAD(list: &child->cg_list);
6632}
6633
6634/**
6635 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6636 * @f: file corresponding to cgroup_dir
6637 *
6638 * Find the cgroup from a file pointer associated with a cgroup directory.
6639 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6640 * cgroup cannot be found.
6641 */
6642static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6643{
6644 struct cgroup_subsys_state *css;
6645
6646 css = css_tryget_online_from_dir(dentry: f->f_path.dentry, NULL);
6647 if (IS_ERR(ptr: css))
6648 return ERR_CAST(ptr: css);
6649
6650 return css->cgroup;
6651}
6652
6653/**
6654 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6655 * cgroup2.
6656 * @f: file corresponding to cgroup2_dir
6657 */
6658static struct cgroup *cgroup_get_from_file(struct file *f)
6659{
6660 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6661
6662 if (IS_ERR(ptr: cgrp))
6663 return ERR_CAST(ptr: cgrp);
6664
6665 if (!cgroup_on_dfl(cgrp)) {
6666 cgroup_put(cgrp);
6667 return ERR_PTR(error: -EBADF);
6668 }
6669
6670 return cgrp;
6671}
6672
6673/**
6674 * cgroup_css_set_fork - find or create a css_set for a child process
6675 * @kargs: the arguments passed to create the child process
6676 *
6677 * This functions finds or creates a new css_set which the child
6678 * process will be attached to in cgroup_post_fork(). By default,
6679 * the child process will be given the same css_set as its parent.
6680 *
6681 * If CLONE_INTO_CGROUP is specified this function will try to find an
6682 * existing css_set which includes the requested cgroup and if not create
6683 * a new css_set that the child will be attached to later. If this function
6684 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6685 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6686 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6687 * to the target cgroup.
6688 */
6689static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6690 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6691{
6692 int ret;
6693 struct cgroup *dst_cgrp = NULL;
6694 struct css_set *cset;
6695 struct super_block *sb;
6696
6697 if (kargs->flags & CLONE_INTO_CGROUP)
6698 cgroup_lock();
6699
6700 cgroup_threadgroup_change_begin(current);
6701
6702 spin_lock_irq(lock: &css_set_lock);
6703 cset = task_css_set(current);
6704 get_css_set(cset);
6705 if (kargs->cgrp)
6706 kargs->kill_seq = kargs->cgrp->kill_seq;
6707 else
6708 kargs->kill_seq = cset->dfl_cgrp->kill_seq;
6709 spin_unlock_irq(lock: &css_set_lock);
6710
6711 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6712 kargs->cset = cset;
6713 return 0;
6714 }
6715
6716 CLASS(fd_raw, f)(fd: kargs->cgroup);
6717 if (fd_empty(f)) {
6718 ret = -EBADF;
6719 goto err;
6720 }
6721 sb = fd_file(f)->f_path.dentry->d_sb;
6722
6723 dst_cgrp = cgroup_get_from_file(fd_file(f));
6724 if (IS_ERR(ptr: dst_cgrp)) {
6725 ret = PTR_ERR(ptr: dst_cgrp);
6726 dst_cgrp = NULL;
6727 goto err;
6728 }
6729
6730 if (cgroup_is_dead(cgrp: dst_cgrp)) {
6731 ret = -ENODEV;
6732 goto err;
6733 }
6734
6735 /*
6736 * Verify that we the target cgroup is writable for us. This is
6737 * usually done by the vfs layer but since we're not going through
6738 * the vfs layer here we need to do it "manually".
6739 */
6740 ret = cgroup_may_write(cgrp: dst_cgrp, sb);
6741 if (ret)
6742 goto err;
6743
6744 /*
6745 * Spawning a task directly into a cgroup works by passing a file
6746 * descriptor to the target cgroup directory. This can even be an O_PATH
6747 * file descriptor. But it can never be a cgroup.procs file descriptor.
6748 * This was done on purpose so spawning into a cgroup could be
6749 * conceptualized as an atomic
6750 *
6751 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6752 * write(fd, <child-pid>, ...);
6753 *
6754 * sequence, i.e. it's a shorthand for the caller opening and writing
6755 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6756 * to always use the caller's credentials.
6757 */
6758 ret = cgroup_attach_permissions(src_cgrp: cset->dfl_cgrp, dst_cgrp, sb,
6759 threadgroup: !(kargs->flags & CLONE_THREAD),
6760 current->nsproxy->cgroup_ns);
6761 if (ret)
6762 goto err;
6763
6764 kargs->cset = find_css_set(old_cset: cset, cgrp: dst_cgrp);
6765 if (!kargs->cset) {
6766 ret = -ENOMEM;
6767 goto err;
6768 }
6769
6770 put_css_set(cset);
6771 kargs->cgrp = dst_cgrp;
6772 return ret;
6773
6774err:
6775 cgroup_threadgroup_change_end(current);
6776 cgroup_unlock();
6777 if (dst_cgrp)
6778 cgroup_put(cgrp: dst_cgrp);
6779 put_css_set(cset);
6780 if (kargs->cset)
6781 put_css_set(cset: kargs->cset);
6782 return ret;
6783}
6784
6785/**
6786 * cgroup_css_set_put_fork - drop references we took during fork
6787 * @kargs: the arguments passed to create the child process
6788 *
6789 * Drop references to the prepared css_set and target cgroup if
6790 * CLONE_INTO_CGROUP was requested.
6791 */
6792static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6793 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6794{
6795 struct cgroup *cgrp = kargs->cgrp;
6796 struct css_set *cset = kargs->cset;
6797
6798 cgroup_threadgroup_change_end(current);
6799
6800 if (cset) {
6801 put_css_set(cset);
6802 kargs->cset = NULL;
6803 }
6804
6805 if (kargs->flags & CLONE_INTO_CGROUP) {
6806 cgroup_unlock();
6807 if (cgrp) {
6808 cgroup_put(cgrp);
6809 kargs->cgrp = NULL;
6810 }
6811 }
6812}
6813
6814/**
6815 * cgroup_can_fork - called on a new task before the process is exposed
6816 * @child: the child process
6817 * @kargs: the arguments passed to create the child process
6818 *
6819 * This prepares a new css_set for the child process which the child will
6820 * be attached to in cgroup_post_fork().
6821 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6822 * callback returns an error, the fork aborts with that error code. This
6823 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6824 */
6825int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6826{
6827 struct cgroup_subsys *ss;
6828 int i, j, ret;
6829
6830 ret = cgroup_css_set_fork(kargs);
6831 if (ret)
6832 return ret;
6833
6834 do_each_subsys_mask(ss, i, have_canfork_callback) {
6835 ret = ss->can_fork(child, kargs->cset);
6836 if (ret)
6837 goto out_revert;
6838 } while_each_subsys_mask();
6839
6840 return 0;
6841
6842out_revert:
6843 for_each_subsys(ss, j) {
6844 if (j >= i)
6845 break;
6846 if (ss->cancel_fork)
6847 ss->cancel_fork(child, kargs->cset);
6848 }
6849
6850 cgroup_css_set_put_fork(kargs);
6851
6852 return ret;
6853}
6854
6855/**
6856 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6857 * @child: the child process
6858 * @kargs: the arguments passed to create the child process
6859 *
6860 * This calls the cancel_fork() callbacks if a fork failed *after*
6861 * cgroup_can_fork() succeeded and cleans up references we took to
6862 * prepare a new css_set for the child process in cgroup_can_fork().
6863 */
6864void cgroup_cancel_fork(struct task_struct *child,
6865 struct kernel_clone_args *kargs)
6866{
6867 struct cgroup_subsys *ss;
6868 int i;
6869
6870 for_each_subsys(ss, i)
6871 if (ss->cancel_fork)
6872 ss->cancel_fork(child, kargs->cset);
6873
6874 cgroup_css_set_put_fork(kargs);
6875}
6876
6877/**
6878 * cgroup_post_fork - finalize cgroup setup for the child process
6879 * @child: the child process
6880 * @kargs: the arguments passed to create the child process
6881 *
6882 * Attach the child process to its css_set calling the subsystem fork()
6883 * callbacks.
6884 */
6885void cgroup_post_fork(struct task_struct *child,
6886 struct kernel_clone_args *kargs)
6887 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6888{
6889 unsigned int cgrp_kill_seq = 0;
6890 unsigned long cgrp_flags = 0;
6891 bool kill = false;
6892 struct cgroup_subsys *ss;
6893 struct css_set *cset;
6894 int i;
6895
6896 cset = kargs->cset;
6897 kargs->cset = NULL;
6898
6899 spin_lock_irq(lock: &css_set_lock);
6900
6901 /* init tasks are special, only link regular threads */
6902 if (likely(child->pid)) {
6903 if (kargs->cgrp) {
6904 cgrp_flags = kargs->cgrp->flags;
6905 cgrp_kill_seq = kargs->cgrp->kill_seq;
6906 } else {
6907 cgrp_flags = cset->dfl_cgrp->flags;
6908 cgrp_kill_seq = cset->dfl_cgrp->kill_seq;
6909 }
6910
6911 WARN_ON_ONCE(!list_empty(&child->cg_list));
6912 cset->nr_tasks++;
6913 css_set_move_task(task: child, NULL, to_cset: cset, use_mg_tasks: false);
6914 } else {
6915 put_css_set(cset);
6916 cset = NULL;
6917 }
6918
6919 if (!(child->flags & PF_KTHREAD)) {
6920 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6921 /*
6922 * If the cgroup has to be frozen, the new task has
6923 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6924 * get the task into the frozen state.
6925 */
6926 spin_lock(lock: &child->sighand->siglock);
6927 WARN_ON_ONCE(child->frozen);
6928 child->jobctl |= JOBCTL_TRAP_FREEZE;
6929 spin_unlock(lock: &child->sighand->siglock);
6930
6931 /*
6932 * Calling cgroup_update_frozen() isn't required here,
6933 * because it will be called anyway a bit later from
6934 * do_freezer_trap(). So we avoid cgroup's transient
6935 * switch from the frozen state and back.
6936 */
6937 }
6938
6939 /*
6940 * If the cgroup is to be killed notice it now and take the
6941 * child down right after we finished preparing it for
6942 * userspace.
6943 */
6944 kill = kargs->kill_seq != cgrp_kill_seq;
6945 }
6946
6947 spin_unlock_irq(lock: &css_set_lock);
6948
6949 /*
6950 * Call ss->fork(). This must happen after @child is linked on
6951 * css_set; otherwise, @child might change state between ->fork()
6952 * and addition to css_set.
6953 */
6954 do_each_subsys_mask(ss, i, have_fork_callback) {
6955 ss->fork(child);
6956 } while_each_subsys_mask();
6957
6958 /* Make the new cset the root_cset of the new cgroup namespace. */
6959 if (kargs->flags & CLONE_NEWCGROUP) {
6960 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6961
6962 get_css_set(cset);
6963 child->nsproxy->cgroup_ns->root_cset = cset;
6964 put_css_set(cset: rcset);
6965 }
6966
6967 /* Cgroup has to be killed so take down child immediately. */
6968 if (unlikely(kill))
6969 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, p: child, type: PIDTYPE_TGID);
6970
6971 cgroup_css_set_put_fork(kargs);
6972}
6973
6974/**
6975 * cgroup_exit - detach cgroup from exiting task
6976 * @tsk: pointer to task_struct of exiting process
6977 *
6978 * Description: Detach cgroup from @tsk.
6979 *
6980 */
6981void cgroup_exit(struct task_struct *tsk)
6982{
6983 struct cgroup_subsys *ss;
6984 struct css_set *cset;
6985 int i;
6986
6987 spin_lock_irq(lock: &css_set_lock);
6988
6989 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6990 cset = task_css_set(task: tsk);
6991 css_set_move_task(task: tsk, from_cset: cset, NULL, use_mg_tasks: false);
6992 cset->nr_tasks--;
6993 /* matches the signal->live check in css_task_iter_advance() */
6994 if (thread_group_leader(p: tsk) && atomic_read(v: &tsk->signal->live))
6995 list_add_tail(new: &tsk->cg_list, head: &cset->dying_tasks);
6996
6997 if (dl_task(p: tsk))
6998 dec_dl_tasks_cs(task: tsk);
6999
7000 WARN_ON_ONCE(cgroup_task_frozen(tsk));
7001 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
7002 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
7003 cgroup_update_frozen(cgrp: task_dfl_cgroup(task: tsk));
7004
7005 spin_unlock_irq(lock: &css_set_lock);
7006
7007 /* see cgroup_post_fork() for details */
7008 do_each_subsys_mask(ss, i, have_exit_callback) {
7009 ss->exit(tsk);
7010 } while_each_subsys_mask();
7011}
7012
7013void cgroup_release(struct task_struct *task)
7014{
7015 struct cgroup_subsys *ss;
7016 int ssid;
7017
7018 do_each_subsys_mask(ss, ssid, have_release_callback) {
7019 ss->release(task);
7020 } while_each_subsys_mask();
7021
7022 if (!list_empty(head: &task->cg_list)) {
7023 spin_lock_irq(lock: &css_set_lock);
7024 css_set_skip_task_iters(cset: task_css_set(task), task);
7025 list_del_init(entry: &task->cg_list);
7026 spin_unlock_irq(lock: &css_set_lock);
7027 }
7028}
7029
7030void cgroup_free(struct task_struct *task)
7031{
7032 struct css_set *cset = task_css_set(task);
7033 put_css_set(cset);
7034}
7035
7036static int __init cgroup_disable(char *str)
7037{
7038 struct cgroup_subsys *ss;
7039 char *token;
7040 int i;
7041
7042 while ((token = strsep(&str, ",")) != NULL) {
7043 if (!*token)
7044 continue;
7045
7046 for_each_subsys(ss, i) {
7047 if (strcmp(token, ss->name) &&
7048 strcmp(token, ss->legacy_name))
7049 continue;
7050
7051 static_branch_disable(cgroup_subsys_enabled_key[i]);
7052 pr_info("Disabling %s control group subsystem\n",
7053 ss->name);
7054 }
7055
7056 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
7057 if (strcmp(token, cgroup_opt_feature_names[i]))
7058 continue;
7059 cgroup_feature_disable_mask |= 1 << i;
7060 pr_info("Disabling %s control group feature\n",
7061 cgroup_opt_feature_names[i]);
7062 break;
7063 }
7064 }
7065 return 1;
7066}
7067__setup("cgroup_disable=", cgroup_disable);
7068
7069void __init __weak enable_debug_cgroup(void) { }
7070
7071static int __init enable_cgroup_debug(char *str)
7072{
7073 cgroup_debug = true;
7074 enable_debug_cgroup();
7075 return 1;
7076}
7077__setup("cgroup_debug", enable_cgroup_debug);
7078
7079static int __init cgroup_favordynmods_setup(char *str)
7080{
7081 return (kstrtobool(s: str, res: &have_favordynmods) == 0);
7082}
7083__setup("cgroup_favordynmods=", cgroup_favordynmods_setup);
7084
7085/**
7086 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
7087 * @dentry: directory dentry of interest
7088 * @ss: subsystem of interest
7089 *
7090 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
7091 * to get the corresponding css and return it. If such css doesn't exist
7092 * or can't be pinned, an ERR_PTR value is returned.
7093 */
7094struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
7095 struct cgroup_subsys *ss)
7096{
7097 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
7098 struct file_system_type *s_type = dentry->d_sb->s_type;
7099 struct cgroup_subsys_state *css = NULL;
7100 struct cgroup *cgrp;
7101
7102 /* is @dentry a cgroup dir? */
7103 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
7104 !kn || kernfs_type(kn) != KERNFS_DIR)
7105 return ERR_PTR(error: -EBADF);
7106
7107 rcu_read_lock();
7108
7109 /*
7110 * This path doesn't originate from kernfs and @kn could already
7111 * have been or be removed at any point. @kn->priv is RCU
7112 * protected for this access. See css_release_work_fn() for details.
7113 */
7114 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
7115 if (cgrp)
7116 css = cgroup_css(cgrp, ss);
7117
7118 if (!css || !css_tryget_online(css))
7119 css = ERR_PTR(error: -ENOENT);
7120
7121 rcu_read_unlock();
7122 return css;
7123}
7124
7125/**
7126 * css_from_id - lookup css by id
7127 * @id: the cgroup id
7128 * @ss: cgroup subsys to be looked into
7129 *
7130 * Returns the css if there's valid one with @id, otherwise returns NULL.
7131 * Should be called under rcu_read_lock().
7132 */
7133struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
7134{
7135 WARN_ON_ONCE(!rcu_read_lock_held());
7136 return idr_find(&ss->css_idr, id);
7137}
7138
7139/**
7140 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
7141 * @path: path on the default hierarchy
7142 *
7143 * Find the cgroup at @path on the default hierarchy, increment its
7144 * reference count and return it. Returns pointer to the found cgroup on
7145 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
7146 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
7147 */
7148struct cgroup *cgroup_get_from_path(const char *path)
7149{
7150 struct kernfs_node *kn;
7151 struct cgroup *cgrp = ERR_PTR(error: -ENOENT);
7152 struct cgroup *root_cgrp;
7153
7154 root_cgrp = current_cgns_cgroup_dfl();
7155 kn = kernfs_walk_and_get(kn: root_cgrp->kn, path);
7156 if (!kn)
7157 goto out;
7158
7159 if (kernfs_type(kn) != KERNFS_DIR) {
7160 cgrp = ERR_PTR(error: -ENOTDIR);
7161 goto out_kernfs;
7162 }
7163
7164 rcu_read_lock();
7165
7166 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
7167 if (!cgrp || !cgroup_tryget(cgrp))
7168 cgrp = ERR_PTR(error: -ENOENT);
7169
7170 rcu_read_unlock();
7171
7172out_kernfs:
7173 kernfs_put(kn);
7174out:
7175 return cgrp;
7176}
7177EXPORT_SYMBOL_GPL(cgroup_get_from_path);
7178
7179/**
7180 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
7181 * @fd: fd obtained by open(cgroup_dir)
7182 *
7183 * Find the cgroup from a fd which should be obtained
7184 * by opening a cgroup directory. Returns a pointer to the
7185 * cgroup on success. ERR_PTR is returned if the cgroup
7186 * cannot be found.
7187 */
7188struct cgroup *cgroup_v1v2_get_from_fd(int fd)
7189{
7190 CLASS(fd_raw, f)(fd);
7191 if (fd_empty(f))
7192 return ERR_PTR(error: -EBADF);
7193
7194 return cgroup_v1v2_get_from_file(fd_file(f));
7195}
7196
7197/**
7198 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
7199 * cgroup2.
7200 * @fd: fd obtained by open(cgroup2_dir)
7201 */
7202struct cgroup *cgroup_get_from_fd(int fd)
7203{
7204 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
7205
7206 if (IS_ERR(ptr: cgrp))
7207 return ERR_CAST(ptr: cgrp);
7208
7209 if (!cgroup_on_dfl(cgrp)) {
7210 cgroup_put(cgrp);
7211 return ERR_PTR(error: -EBADF);
7212 }
7213 return cgrp;
7214}
7215EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
7216
7217static u64 power_of_ten(int power)
7218{
7219 u64 v = 1;
7220 while (power--)
7221 v *= 10;
7222 return v;
7223}
7224
7225/**
7226 * cgroup_parse_float - parse a floating number
7227 * @input: input string
7228 * @dec_shift: number of decimal digits to shift
7229 * @v: output
7230 *
7231 * Parse a decimal floating point number in @input and store the result in
7232 * @v with decimal point right shifted @dec_shift times. For example, if
7233 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
7234 * Returns 0 on success, -errno otherwise.
7235 *
7236 * There's nothing cgroup specific about this function except that it's
7237 * currently the only user.
7238 */
7239int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
7240{
7241 s64 whole, frac = 0;
7242 int fstart = 0, fend = 0, flen;
7243
7244 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
7245 return -EINVAL;
7246 if (frac < 0)
7247 return -EINVAL;
7248
7249 flen = fend > fstart ? fend - fstart : 0;
7250 if (flen < dec_shift)
7251 frac *= power_of_ten(power: dec_shift - flen);
7252 else
7253 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
7254
7255 *v = whole * power_of_ten(power: dec_shift) + frac;
7256 return 0;
7257}
7258
7259/*
7260 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
7261 * definition in cgroup-defs.h.
7262 */
7263#ifdef CONFIG_SOCK_CGROUP_DATA
7264
7265void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
7266{
7267 struct cgroup *cgroup;
7268
7269 rcu_read_lock();
7270 /* Don't associate the sock with unrelated interrupted task's cgroup. */
7271 if (in_interrupt()) {
7272 cgroup = &cgrp_dfl_root.cgrp;
7273 cgroup_get(cgrp: cgroup);
7274 goto out;
7275 }
7276
7277 while (true) {
7278 struct css_set *cset;
7279
7280 cset = task_css_set(current);
7281 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
7282 cgroup = cset->dfl_cgrp;
7283 break;
7284 }
7285 cpu_relax();
7286 }
7287out:
7288 skcd->cgroup = cgroup;
7289 cgroup_bpf_get(cgrp: cgroup);
7290 rcu_read_unlock();
7291}
7292
7293void cgroup_sk_clone(struct sock_cgroup_data *skcd)
7294{
7295 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7296
7297 /*
7298 * We might be cloning a socket which is left in an empty
7299 * cgroup and the cgroup might have already been rmdir'd.
7300 * Don't use cgroup_get_live().
7301 */
7302 cgroup_get(cgrp);
7303 cgroup_bpf_get(cgrp);
7304}
7305
7306void cgroup_sk_free(struct sock_cgroup_data *skcd)
7307{
7308 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7309
7310 cgroup_bpf_put(cgrp);
7311 cgroup_put(cgrp);
7312}
7313
7314#endif /* CONFIG_SOCK_CGROUP_DATA */
7315
7316#ifdef CONFIG_SYSFS
7317static ssize_t show_delegatable_files(struct cftype *files, char *buf,
7318 ssize_t size, const char *prefix)
7319{
7320 struct cftype *cft;
7321 ssize_t ret = 0;
7322
7323 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
7324 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
7325 continue;
7326
7327 if (prefix)
7328 ret += snprintf(buf: buf + ret, size: size - ret, fmt: "%s.", prefix);
7329
7330 ret += snprintf(buf: buf + ret, size: size - ret, fmt: "%s\n", cft->name);
7331
7332 if (WARN_ON(ret >= size))
7333 break;
7334 }
7335
7336 return ret;
7337}
7338
7339static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7340 char *buf)
7341{
7342 struct cgroup_subsys *ss;
7343 int ssid;
7344 ssize_t ret = 0;
7345
7346 ret = show_delegatable_files(files: cgroup_base_files, buf: buf + ret,
7347 PAGE_SIZE - ret, NULL);
7348 if (cgroup_psi_enabled())
7349 ret += show_delegatable_files(files: cgroup_psi_files, buf: buf + ret,
7350 PAGE_SIZE - ret, NULL);
7351
7352 for_each_subsys(ss, ssid)
7353 ret += show_delegatable_files(files: ss->dfl_cftypes, buf: buf + ret,
7354 PAGE_SIZE - ret,
7355 prefix: cgroup_subsys_name[ssid]);
7356
7357 return ret;
7358}
7359static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7360
7361static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7362 char *buf)
7363{
7364 return snprintf(buf, PAGE_SIZE,
7365 fmt: "nsdelegate\n"
7366 "favordynmods\n"
7367 "memory_localevents\n"
7368 "memory_recursiveprot\n"
7369 "memory_hugetlb_accounting\n"
7370 "pids_localevents\n");
7371}
7372static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7373
7374static struct attribute *cgroup_sysfs_attrs[] = {
7375 &cgroup_delegate_attr.attr,
7376 &cgroup_features_attr.attr,
7377 NULL,
7378};
7379
7380static const struct attribute_group cgroup_sysfs_attr_group = {
7381 .attrs = cgroup_sysfs_attrs,
7382 .name = "cgroup",
7383};
7384
7385static int __init cgroup_sysfs_init(void)
7386{
7387 return sysfs_create_group(kobj: kernel_kobj, grp: &cgroup_sysfs_attr_group);
7388}
7389subsys_initcall(cgroup_sysfs_init);
7390
7391#endif /* CONFIG_SYSFS */
7392