projects/linux·blob master · refs

7620 lines · 200.7 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic process-grouping system.
4 *
5 * Based originally on the cpuset system, extracted by Paul Menage
6 * Copyright (C) 2006 Google, Inc
7 *
8 * Notifications support
9 * Copyright (C) 2009 Nokia Corporation
10 * Author: Kirill A. Shutemov
11 *
12 * Copyright notices from the original cpuset code:
13 * --------------------------------------------------
14 * Copyright (C) 2003 BULL SA.
15 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 *
17 * Portions derived from Patrick Mochel's sysfs code.
18 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 *
20 * 2003-10-10 Written by Simon Derr.
21 * 2003-10-22 Updates by Stephen Hemminger.
22 * 2004 May-July Rework by Paul Jackson.
23 * ---------------------------------------------------
24 */
25
26#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
27
28#include "cgroup-internal.h"
29
30#include <linux/bpf-cgroup.h>
31#include <linux/cred.h>
32#include <linux/errno.h>
33#include <linux/init_task.h>
34#include <linux/kernel.h>
35#include <linux/magic.h>
36#include <linux/mutex.h>
37#include <linux/mount.h>
38#include <linux/pagemap.h>
39#include <linux/proc_fs.h>
40#include <linux/rcupdate.h>
41#include <linux/sched.h>
42#include <linux/sched/task.h>
43#include <linux/slab.h>
44#include <linux/spinlock.h>
45#include <linux/percpu-rwsem.h>
46#include <linux/string.h>
47#include <linux/hashtable.h>
48#include <linux/idr.h>
49#include <linux/kthread.h>
50#include <linux/atomic.h>
51#include <linux/cpuset.h>
52#include <linux/proc_ns.h>
53#include <linux/nsproxy.h>
54#include <linux/file.h>
55#include <linux/fs_parser.h>
56#include <linux/sched/cputime.h>
57#include <linux/sched/deadline.h>
58#include <linux/psi.h>
59#include <linux/nstree.h>
60#include <linux/irq_work.h>
61#include <net/sock.h>
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/cgroup.h>
65
66#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
67 MAX_CFTYPE_NAME + 2)
68/* let's not notify more than 100 times per second */
69#define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
70
71/*
72 * cgroup_mutex is the master lock. Any modification to cgroup or its
73 * hierarchy must be performed while holding it.
74 *
75 * css_set_lock protects task->cgroups pointer, the list of css_set
76 * objects, and the chain of tasks off each css_set.
77 *
78 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
79 * cgroup.h can use them for lockdep annotations.
80 */
81DEFINE_MUTEX(cgroup_mutex);
82DEFINE_SPINLOCK(css_set_lock);
83
84#if (defined CONFIG_PROVE_RCU || defined CONFIG_LOCKDEP)
85EXPORT_SYMBOL_GPL(cgroup_mutex);
86EXPORT_SYMBOL_GPL(css_set_lock);
87#endif
88
89struct blocking_notifier_head cgroup_lifetime_notifier =
90 BLOCKING_NOTIFIER_INIT(cgroup_lifetime_notifier);
91struct blocking_notifier_head cgroup_task_notifier =
92 BLOCKING_NOTIFIER_INIT(cgroup_task_notifier);
93
94DEFINE_SPINLOCK(trace_cgroup_path_lock);
95char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
96static bool cgroup_debug __read_mostly;
97
98/*
99 * Protects cgroup_idr and css_idr so that IDs can be released without
100 * grabbing cgroup_mutex.
101 */
102static DEFINE_SPINLOCK(cgroup_idr_lock);
103
104DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
105
106#define cgroup_assert_mutex_or_rcu_locked() \
107 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
108 !lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required")
110
111/*
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_percpu_wq
115 * which may lead to deadlock.
116 *
117 * A cgroup destruction should enqueue work sequentially to:
118 * cgroup_offline_wq: use for css offline work
119 * cgroup_release_wq: use for css release work
120 * cgroup_free_wq: use for free work
121 *
122 * Rationale for using separate workqueues:
123 * The cgroup root free work may depend on completion of other css offline
124 * operations. If all tasks were enqueued to a single workqueue, this could
125 * create a deadlock scenario where:
126 * - Free work waits for other css offline work to complete.
127 * - But other css offline work is queued after free work in the same queue.
128 *
129 * Example deadlock scenario with single workqueue (cgroup_destroy_wq):
130 * 1. umount net_prio
131 * 2. net_prio root destruction enqueues work to cgroup_destroy_wq (CPUx)
132 * 3. perf_event CSS A offline enqueues work to same cgroup_destroy_wq (CPUx)
133 * 4. net_prio cgroup_destroy_root->cgroup_lock_and_drain_offline.
134 * 5. net_prio root destruction blocks waiting for perf_event CSS A offline,
135 * which can never complete as it's behind in the same queue and
136 * workqueue's max_active is 1.
137 */
138static struct workqueue_struct *cgroup_offline_wq;
139static struct workqueue_struct *cgroup_release_wq;
140static struct workqueue_struct *cgroup_free_wq;
141
142/* generate an array of cgroup subsystem pointers */
143#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
144struct cgroup_subsys *cgroup_subsys[] = {
145#include <linux/cgroup_subsys.h>
146};
147#undef SUBSYS
148
149/* array of cgroup subsystem names */
150#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
151static const char *cgroup_subsys_name[] = {
152#include <linux/cgroup_subsys.h>
153};
154#undef SUBSYS
155
156/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
157#define SUBSYS(_x) \
158 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
159 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
160 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
161 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
162#include <linux/cgroup_subsys.h>
163#undef SUBSYS
164
165#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
166static struct static_key_true *cgroup_subsys_enabled_key[] = {
167#include <linux/cgroup_subsys.h>
168};
169#undef SUBSYS
170
171#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
172static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
173#include <linux/cgroup_subsys.h>
174};
175#undef SUBSYS
176
177static DEFINE_PER_CPU(struct css_rstat_cpu, root_rstat_cpu);
178static DEFINE_PER_CPU(struct cgroup_rstat_base_cpu, root_rstat_base_cpu);
179
180/* the default hierarchy */
181struct cgroup_root cgrp_dfl_root = {
182 .cgrp.self.rstat_cpu = &root_rstat_cpu,
183 .cgrp.rstat_base_cpu = &root_rstat_base_cpu,
184};
185EXPORT_SYMBOL_GPL(cgrp_dfl_root);
186
187/*
188 * The default hierarchy always exists but is hidden until mounted for the
189 * first time. This is for backward compatibility.
190 */
191bool cgrp_dfl_visible;
192
193/* some controllers are not supported in the default hierarchy */
194static u32 cgrp_dfl_inhibit_ss_mask;
195
196/* some controllers are implicitly enabled on the default hierarchy */
197static u32 cgrp_dfl_implicit_ss_mask;
198
199/* some controllers can be threaded on the default hierarchy */
200static u32 cgrp_dfl_threaded_ss_mask;
201
202/*
203 * Set across rebind_subsystems() to the controllers leaving a hierarchy.
204 * Guarded by cgroup_mutex. Makes find_existing_css_set() resolve them to the
205 * root css so the affected tasks are migrated there before
206 * cgroup_apply_control_disable() kills the per-cgroup csses.
207 */
208static u32 cgroup_rebind_ss_mask;
209
210/* The list of hierarchy roots */
211LIST_HEAD(cgroup_roots);
212static int cgroup_root_count;
213
214/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
215static DEFINE_IDR(cgroup_hierarchy_idr);
216
217/*
218 * Assign a monotonically increasing serial number to csses. It guarantees
219 * cgroups with bigger numbers are newer than those with smaller numbers.
220 * Also, as csses are always appended to the parent's ->children list, it
221 * guarantees that sibling csses are always sorted in the ascending serial
222 * number order on the list. Protected by cgroup_mutex.
223 */
224static u64 css_serial_nr_next = 1;
225
226/*
227 * These bitmasks identify subsystems with specific features to avoid
228 * having to do iterative checks repeatedly.
229 */
230static u32 have_fork_callback __read_mostly;
231static u32 have_exit_callback __read_mostly;
232static u32 have_release_callback __read_mostly;
233static u32 have_canfork_callback __read_mostly;
234
235static bool have_favordynmods __ro_after_init = IS_ENABLED(CONFIG_CGROUP_FAVOR_DYNMODS);
236
237/*
238 * Write protected by cgroup_mutex and write-lock of cgroup_threadgroup_rwsem,
239 * read protected by either.
240 *
241 * Can only be turned on, but not turned off.
242 */
243bool cgroup_enable_per_threadgroup_rwsem __read_mostly;
244
245/* cgroup namespace for init task */
246struct cgroup_namespace init_cgroup_ns = {
247 .ns = NS_COMMON_INIT(init_cgroup_ns),
248 .user_ns = &init_user_ns,
249 .root_cset = &init_css_set,
250};
251
252static struct file_system_type cgroup2_fs_type;
253static struct cftype cgroup_base_files[];
254static struct cftype cgroup_psi_files[];
255
256/* cgroup optional features */
257enum cgroup_opt_features {
258#ifdef CONFIG_PSI
259 OPT_FEATURE_PRESSURE,
260#endif
261 OPT_FEATURE_COUNT
262};
263
264static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
265#ifdef CONFIG_PSI
266 "pressure",
267#endif
268};
269
270static u16 cgroup_feature_disable_mask __read_mostly;
271
272static int cgroup_apply_control(struct cgroup *cgrp);
273static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
274static void css_task_iter_skip(struct css_task_iter *it,
275 struct task_struct *task);
276static int cgroup_destroy_locked(struct cgroup *cgrp);
277static void kill_css_sync(struct cgroup_subsys_state *css);
278static void kill_css_finish(struct cgroup_subsys_state *css);
279static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
280 struct cgroup_subsys *ss);
281static void css_release(struct percpu_ref *ref);
282static int cgroup_addrm_files(struct cgroup_subsys_state *css,
283 struct cgroup *cgrp, struct cftype cfts[],
284 bool is_add);
285static void cgroup_rt_init(void);
286
287#ifdef CONFIG_DEBUG_CGROUP_REF
288#define CGROUP_REF_FN_ATTRS noinline
289#define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn);
290#include <linux/cgroup_refcnt.h>
291#endif
292
293/**
294 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
295 * @ssid: subsys ID of interest
296 *
297 * cgroup_subsys_enabled() can only be used with literal subsys names which
298 * is fine for individual subsystems but unsuitable for cgroup core. This
299 * is slower static_key_enabled() based test indexed by @ssid.
300 */
301bool cgroup_ssid_enabled(int ssid)
302{
303 if (!CGROUP_HAS_SUBSYS_CONFIG)
304 return false;
305
306 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
307}
308
309/**
310 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
311 * @cgrp: the cgroup of interest
312 *
313 * The default hierarchy is the v2 interface of cgroup and this function
314 * can be used to test whether a cgroup is on the default hierarchy for
315 * cases where a subsystem should behave differently depending on the
316 * interface version.
317 *
318 * List of changed behaviors:
319 *
320 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
321 * and "name" are disallowed.
322 *
323 * - When mounting an existing superblock, mount options should match.
324 *
325 * - rename(2) is disallowed.
326 *
327 * - "tasks" is removed. Everything should be at process granularity. Use
328 * "cgroup.procs" instead.
329 *
330 * - "cgroup.procs" is not sorted. pids will be unique unless they got
331 * recycled in-between reads.
332 *
333 * - "release_agent" and "notify_on_release" are removed. Replacement
334 * notification mechanism will be implemented.
335 *
336 * - "cgroup.clone_children" is removed.
337 *
338 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
339 * and its descendants contain no task; otherwise, 1. The file also
340 * generates kernfs notification which can be monitored through poll and
341 * [di]notify when the value of the file changes.
342 *
343 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
344 * take masks of ancestors with non-empty cpus/mems, instead of being
345 * moved to an ancestor.
346 *
347 * - cpuset: a task can be moved into an empty cpuset, and again it takes
348 * masks of ancestors.
349 *
350 * - blkcg: blk-throttle becomes properly hierarchical.
351 */
352bool cgroup_on_dfl(const struct cgroup *cgrp)
353{
354 return cgrp->root == &cgrp_dfl_root;
355}
356
357/* IDR wrappers which synchronize using cgroup_idr_lock */
358static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
359 gfp_t gfp_mask)
360{
361 int ret;
362
363 idr_preload(gfp_mask);
364 spin_lock_bh(&cgroup_idr_lock);
365 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
366 spin_unlock_bh(&cgroup_idr_lock);
367 idr_preload_end();
368 return ret;
369}
370
371static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
372{
373 void *ret;
374
375 spin_lock_bh(&cgroup_idr_lock);
376 ret = idr_replace(idr, ptr, id);
377 spin_unlock_bh(&cgroup_idr_lock);
378 return ret;
379}
380
381static void cgroup_idr_remove(struct idr *idr, int id)
382{
383 spin_lock_bh(&cgroup_idr_lock);
384 idr_remove(idr, id);
385 spin_unlock_bh(&cgroup_idr_lock);
386}
387
388static bool cgroup_is_threaded(struct cgroup *cgrp)
389{
390 return cgrp->dom_cgrp != cgrp;
391}
392
393/* can @cgrp host both domain and threaded children? */
394static bool cgroup_is_mixable(struct cgroup *cgrp)
395{
396 /*
397 * Root isn't under domain level resource control exempting it from
398 * the no-internal-process constraint, so it can serve as a thread
399 * root and a parent of resource domains at the same time.
400 */
401 return !cgroup_parent(cgrp);
402}
403
404/* can @cgrp become a thread root? Should always be true for a thread root */
405static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
406{
407 /* mixables don't care */
408 if (cgroup_is_mixable(cgrp))
409 return true;
410
411 /* domain roots can't be nested under threaded */
412 if (cgroup_is_threaded(cgrp))
413 return false;
414
415 /* can only have either domain or threaded children */
416 if (READ_ONCE(cgrp->nr_populated_domain_children))
417 return false;
418
419 /* and no domain controllers can be enabled */
420 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
421 return false;
422
423 return true;
424}
425
426/* is @cgrp root of a threaded subtree? */
427static bool cgroup_is_thread_root(struct cgroup *cgrp)
428{
429 /* thread root should be a domain */
430 if (cgroup_is_threaded(cgrp))
431 return false;
432
433 /* a domain w/ threaded children is a thread root */
434 if (cgrp->nr_threaded_children)
435 return true;
436
437 /*
438 * A domain which has tasks and explicit threaded controllers
439 * enabled is a thread root.
440 */
441 if (cgroup_has_tasks(cgrp) &&
442 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
443 return true;
444
445 return false;
446}
447
448/* a domain which isn't connected to the root w/o brekage can't be used */
449static bool cgroup_is_valid_domain(struct cgroup *cgrp)
450{
451 /* the cgroup itself can be a thread root */
452 if (cgroup_is_threaded(cgrp))
453 return false;
454
455 /* but the ancestors can't be unless mixable */
456 while ((cgrp = cgroup_parent(cgrp))) {
457 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
458 return false;
459 if (cgroup_is_threaded(cgrp))
460 return false;
461 }
462
463 return true;
464}
465
466/* subsystems visibly enabled on a cgroup */
467static u32 cgroup_control(struct cgroup *cgrp)
468{
469 struct cgroup *parent = cgroup_parent(cgrp);
470 u32 root_ss_mask = cgrp->root->subsys_mask;
471
472 if (parent) {
473 u32 ss_mask = parent->subtree_control;
474
475 /* threaded cgroups can only have threaded controllers */
476 if (cgroup_is_threaded(cgrp))
477 ss_mask &= cgrp_dfl_threaded_ss_mask;
478 return ss_mask;
479 }
480
481 if (cgroup_on_dfl(cgrp))
482 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
483 cgrp_dfl_implicit_ss_mask);
484 return root_ss_mask;
485}
486
487/* subsystems enabled on a cgroup */
488static u32 cgroup_ss_mask(struct cgroup *cgrp)
489{
490 struct cgroup *parent = cgroup_parent(cgrp);
491
492 if (parent) {
493 u32 ss_mask = parent->subtree_ss_mask;
494
495 /* threaded cgroups can only have threaded controllers */
496 if (cgroup_is_threaded(cgrp))
497 ss_mask &= cgrp_dfl_threaded_ss_mask;
498 return ss_mask;
499 }
500
501 return cgrp->root->subsys_mask;
502}
503
504/**
505 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
506 * @cgrp: the cgroup of interest
507 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
508 *
509 * Similar to cgroup_css() but returns the effective css, which is defined
510 * as the matching css of the nearest ancestor including self which has @ss
511 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
512 * function is guaranteed to return non-NULL css.
513 */
514static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
515 struct cgroup_subsys *ss)
516{
517 lockdep_assert_held(&cgroup_mutex);
518
519 if (!ss)
520 return &cgrp->self;
521
522 /*
523 * This function is used while updating css associations and thus
524 * can't test the csses directly. Test ss_mask.
525 */
526 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
527 cgrp = cgroup_parent(cgrp);
528 if (!cgrp)
529 return NULL;
530 }
531
532 return cgroup_css(cgrp, ss);
533}
534
535/**
536 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
537 * @cgrp: the cgroup of interest
538 * @ss: the subsystem of interest
539 *
540 * Find and get the effective css of @cgrp for @ss. The effective css is
541 * defined as the matching css of the nearest ancestor including self which
542 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
543 * the root css is returned, so this function always returns a valid css.
544 *
545 * The returned css is not guaranteed to be online, and therefore it is the
546 * callers responsibility to try get a reference for it.
547 */
548struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
549 struct cgroup_subsys *ss)
550{
551 struct cgroup_subsys_state *css;
552
553 if (!CGROUP_HAS_SUBSYS_CONFIG)
554 return NULL;
555
556 do {
557 css = cgroup_css(cgrp, ss);
558
559 if (css)
560 return css;
561 cgrp = cgroup_parent(cgrp);
562 } while (cgrp);
563
564 return init_css_set.subsys[ss->id];
565}
566
567/**
568 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
569 * @cgrp: the cgroup of interest
570 * @ss: the subsystem of interest
571 *
572 * Find and get the effective css of @cgrp for @ss. The effective css is
573 * defined as the matching css of the nearest ancestor including self which
574 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
575 * the root css is returned, so this function always returns a valid css.
576 * The returned css must be put using css_put().
577 */
578struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
579 struct cgroup_subsys *ss)
580{
581 struct cgroup_subsys_state *css;
582
583 if (!CGROUP_HAS_SUBSYS_CONFIG)
584 return NULL;
585
586 rcu_read_lock();
587
588 do {
589 css = cgroup_css(cgrp, ss);
590
591 if (css && css_tryget_online(css))
592 goto out_unlock;
593 cgrp = cgroup_parent(cgrp);
594 } while (cgrp);
595
596 css = init_css_set.subsys[ss->id];
597 css_get(css);
598out_unlock:
599 rcu_read_unlock();
600 return css;
601}
602EXPORT_SYMBOL_GPL(cgroup_get_e_css);
603
604static void cgroup_get_live(struct cgroup *cgrp)
605{
606 WARN_ON_ONCE(cgroup_is_dead(cgrp));
607 cgroup_get(cgrp);
608}
609
610/**
611 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
612 * is responsible for taking the css_set_lock.
613 * @cgrp: the cgroup in question
614 */
615int __cgroup_task_count(const struct cgroup *cgrp)
616{
617 int count = 0;
618 struct cgrp_cset_link *link;
619
620 lockdep_assert_held(&css_set_lock);
621
622 list_for_each_entry(link, &cgrp->cset_links, cset_link)
623 count += link->cset->nr_tasks;
624
625 return count;
626}
627
628/**
629 * cgroup_task_count - count the number of tasks in a cgroup.
630 * @cgrp: the cgroup in question
631 */
632int cgroup_task_count(const struct cgroup *cgrp)
633{
634 int count;
635
636 spin_lock_irq(&css_set_lock);
637 count = __cgroup_task_count(cgrp);
638 spin_unlock_irq(&css_set_lock);
639
640 return count;
641}
642
643static struct cgroup *kn_priv(struct kernfs_node *kn)
644{
645 struct kernfs_node *parent;
646 /*
647 * The parent can not be replaced due to KERNFS_ROOT_INVARIANT_PARENT.
648 * Therefore it is always safe to dereference this pointer outside of a
649 * RCU section.
650 */
651 parent = rcu_dereference_check(kn->__parent,
652 kernfs_root_flags(kn) & KERNFS_ROOT_INVARIANT_PARENT);
653 return parent->priv;
654}
655
656struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
657{
658 struct cgroup *cgrp = kn_priv(of->kn);
659 struct cftype *cft = of_cft(of);
660
661 /*
662 * This is open and unprotected implementation of cgroup_css().
663 * seq_css() is only called from a kernfs file operation which has
664 * an active reference on the file. Because all the subsystem
665 * files are drained before a css is disassociated with a cgroup,
666 * the matching css from the cgroup's subsys table is guaranteed to
667 * be and stay valid until the enclosing operation is complete.
668 */
669 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
670 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
671 else
672 return &cgrp->self;
673}
674EXPORT_SYMBOL_GPL(of_css);
675
676/**
677 * for_each_css - iterate all css's of a cgroup
678 * @css: the iteration cursor
679 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
680 * @cgrp: the target cgroup to iterate css's of
681 *
682 * Should be called under cgroup_mutex.
683 */
684#define for_each_css(css, ssid, cgrp) \
685 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
686 if (!((css) = rcu_dereference_check( \
687 (cgrp)->subsys[(ssid)], \
688 lockdep_is_held(&cgroup_mutex)))) { } \
689 else
690
691/**
692 * do_each_subsys_mask - filter for_each_subsys with a bitmask
693 * @ss: the iteration cursor
694 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
695 * @ss_mask: the bitmask
696 *
697 * The block will only run for cases where the ssid-th bit (1 << ssid) of
698 * @ss_mask is set.
699 */
700#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
701 unsigned long __ss_mask = (ss_mask); \
702 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
703 (ssid) = 0; \
704 break; \
705 } \
706 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
707 (ss) = cgroup_subsys[ssid]; \
708 {
709
710#define while_each_subsys_mask() \
711 } \
712 } \
713} while (false)
714
715/*
716 * The default css_set - used by init and its children prior to any
717 * hierarchies being mounted. It contains a pointer to the root state
718 * for each subsystem. Also used to anchor the list of css_sets. Not
719 * reference-counted, to improve performance when child cgroups
720 * haven't been created.
721 */
722struct css_set init_css_set = {
723 .refcount = REFCOUNT_INIT(1),
724 .dom_cset = &init_css_set,
725 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
726 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
727 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
728 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
729 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
730 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
731 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
732 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
733 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
734
735 /*
736 * The following field is re-initialized when this cset gets linked
737 * in cgroup_init(). However, let's initialize the field
738 * statically too so that the default cgroup can be accessed safely
739 * early during boot.
740 */
741 .dfl_cgrp = &cgrp_dfl_root.cgrp,
742};
743
744static int css_set_count = 1; /* 1 for init_css_set */
745
746static bool css_set_threaded(struct css_set *cset)
747{
748 return cset->dom_cset != cset;
749}
750
751/**
752 * css_set_populated - does a css_set contain any tasks?
753 * @cset: target css_set
754 *
755 * css_set_populated() should be the same as !!cset->nr_tasks at steady
756 * state. However, css_set_populated() can be called while a task is being
757 * added to or removed from the linked list before the nr_tasks is
758 * properly updated. Hence, we can't just look at ->nr_tasks here.
759 */
760static bool css_set_populated(struct css_set *cset)
761{
762 lockdep_assert_held(&css_set_lock);
763
764 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
765}
766
767/**
768 * css_update_populated - update the populated state of a css and ancestors
769 * @css: leaf css whose own populated count is changing
770 * @populated: inc or dec
771 *
772 * One of the css_sets pinned by @css is getting its first task or losing the
773 * last. Propagate the transition up the parent chain so that a css's
774 * nr_populated_children is zero iff none of its descendants contain any tasks.
775 *
776 * For a cgroup->self walk, also runs cgroup-side bookkeeping at each level:
777 * domain/threaded child split, deferred-destroy trigger, and notification via
778 * "cgroup.populated" (zero iff cgrp->self has neither populated csets nor
779 * populated children; userland is notified on transitions).
780 */
781static void css_update_populated(struct cgroup_subsys_state *css, bool populated)
782{
783 struct cgroup_subsys_state *child = NULL;
784 int adj = populated ? 1 : -1;
785
786 lockdep_assert_held(&css_set_lock);
787
788 do {
789 /* non-NULL only on the cgroup->self walk */
790 struct cgroup *cgrp = css_is_self(css) ? css->cgroup : NULL;
791 bool was_populated = css_is_populated(css);
792
793 if (!child) {
794 WRITE_ONCE(css->nr_populated_csets,
795 css->nr_populated_csets + adj);
796 } else {
797 WRITE_ONCE(css->nr_populated_children,
798 css->nr_populated_children + adj);
799 if (cgrp) {
800 if (cgroup_is_threaded(child->cgroup))
801 WRITE_ONCE(cgrp->nr_populated_threaded_children,
802 cgrp->nr_populated_threaded_children + adj);
803 else
804 WRITE_ONCE(cgrp->nr_populated_domain_children,
805 cgrp->nr_populated_domain_children + adj);
806 }
807 }
808
809 if (was_populated == css_is_populated(css))
810 break;
811
812 /*
813 * Pair with smp_mb() in kill_css_sync(). Either we observe
814 * CSS_DYING and queue, or the caller observes our decrement
815 * and fires synchronously.
816 */
817 smp_mb();
818
819 /*
820 * Subtree just emptied below a dying css. Fire deferred kill.
821 * The transition is one-shot for a dying css.
822 */
823 if (was_populated && css_is_dying(css)) {
824 css_get(css);
825 WARN_ON_ONCE(!queue_work(cgroup_offline_wq, &css->kill_finish_work));
826 }
827
828 if (cgrp) {
829 cgroup1_check_for_release(cgrp);
830 TRACE_CGROUP_PATH(notify_populated, cgrp,
831 cgroup_is_populated(cgrp));
832 cgroup_file_notify(&cgrp->events_file);
833 }
834
835 child = css;
836 css = css->parent;
837 } while (css);
838}
839
840/**
841 * css_set_update_populated - update populated state of a css_set
842 * @cset: target css_set
843 * @populated: whether @cset is populated or depopulated
844 *
845 * @cset is either getting the first task or losing the last. Update the
846 * populated counters along each linked cgroup's self chain and each
847 * subsystem css that @cset pins.
848 */
849static void css_set_update_populated(struct css_set *cset, bool populated)
850{
851 struct cgrp_cset_link *link;
852 struct cgroup_subsys *ss;
853 int ssid;
854
855 lockdep_assert_held(&css_set_lock);
856
857 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
858 css_update_populated(&link->cgrp->self, populated);
859
860 for_each_subsys(ss, ssid) {
861 struct cgroup_subsys_state *css = cset->subsys[ssid];
862
863 if (css)
864 css_update_populated(css, populated);
865 }
866}
867
868/*
869 * @task is leaving, advance task iterators which are pointing to it so
870 * that they can resume at the next position. Advancing an iterator might
871 * remove it from the list, use safe walk. See css_task_iter_skip() for
872 * details.
873 */
874static void css_set_skip_task_iters(struct css_set *cset,
875 struct task_struct *task)
876{
877 struct css_task_iter *it, *pos;
878
879 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
880 css_task_iter_skip(it, task);
881}
882
883/**
884 * css_set_move_task - move a task from one css_set to another
885 * @task: task being moved
886 * @from_cset: css_set @task currently belongs to (may be NULL)
887 * @to_cset: new css_set @task is being moved to (may be NULL)
888 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
889 *
890 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
891 * css_set, @from_cset can be NULL. If @task is being disassociated
892 * instead of moved, @to_cset can be NULL.
893 *
894 * This function automatically handles populated counter updates and
895 * css_task_iter adjustments but the caller is responsible for managing
896 * @from_cset and @to_cset's reference counts.
897 */
898static void css_set_move_task(struct task_struct *task,
899 struct css_set *from_cset, struct css_set *to_cset,
900 bool use_mg_tasks)
901{
902 lockdep_assert_held(&css_set_lock);
903
904 if (to_cset && !css_set_populated(to_cset))
905 css_set_update_populated(to_cset, true);
906
907 if (from_cset) {
908 WARN_ON_ONCE(list_empty(&task->cg_list));
909
910 css_set_skip_task_iters(from_cset, task);
911 list_del_init(&task->cg_list);
912 if (!css_set_populated(from_cset))
913 css_set_update_populated(from_cset, false);
914 } else {
915 WARN_ON_ONCE(!list_empty(&task->cg_list));
916 }
917
918 if (to_cset) {
919 /*
920 * We are synchronized through cgroup_threadgroup_rwsem
921 * against PF_EXITING setting such that we can't race
922 * against cgroup_task_dead()/cgroup_task_free() dropping
923 * the css_set.
924 */
925 WARN_ON_ONCE(task->flags & PF_EXITING);
926
927 cgroup_move_task(task, to_cset);
928 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
929 &to_cset->tasks);
930 }
931}
932
933/*
934 * hash table for cgroup groups. This improves the performance to find
935 * an existing css_set. This hash doesn't (currently) take into
936 * account cgroups in empty hierarchies.
937 */
938#define CSS_SET_HASH_BITS 7
939static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
940
941static unsigned long css_set_hash(struct cgroup_subsys_state **css)
942{
943 unsigned long key = 0UL;
944 struct cgroup_subsys *ss;
945 int i;
946
947 for_each_subsys(ss, i)
948 key += (unsigned long)css[i];
949 key = (key >> 16) ^ key;
950
951 return key;
952}
953
954void put_css_set_locked(struct css_set *cset)
955{
956 struct cgrp_cset_link *link, *tmp_link;
957 struct cgroup_subsys *ss;
958 int ssid;
959
960 lockdep_assert_held(&css_set_lock);
961
962 if (!refcount_dec_and_test(&cset->refcount))
963 return;
964
965 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
966
967 /* This css_set is dead. Unlink it and release cgroup and css refs */
968 for_each_subsys(ss, ssid) {
969 list_del(&cset->e_cset_node[ssid]);
970 css_put(cset->subsys[ssid]);
971 }
972 hash_del(&cset->hlist);
973 css_set_count--;
974
975 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
976 list_del(&link->cset_link);
977 list_del(&link->cgrp_link);
978 if (cgroup_parent(link->cgrp))
979 cgroup_put(link->cgrp);
980 kfree(link);
981 }
982
983 if (css_set_threaded(cset)) {
984 list_del(&cset->threaded_csets_node);
985 put_css_set_locked(cset->dom_cset);
986 }
987
988 kfree_rcu(cset, rcu_head);
989}
990
991/**
992 * compare_css_sets - helper function for find_existing_css_set().
993 * @cset: candidate css_set being tested
994 * @old_cset: existing css_set for a task
995 * @new_cgrp: cgroup that's being entered by the task
996 * @template: desired set of css pointers in css_set (pre-calculated)
997 *
998 * Returns true if "cset" matches "old_cset" except for the hierarchy
999 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
1000 */
1001static bool compare_css_sets(struct css_set *cset,
1002 struct css_set *old_cset,
1003 struct cgroup *new_cgrp,
1004 struct cgroup_subsys_state *template[])
1005{
1006 struct cgroup *new_dfl_cgrp;
1007 struct list_head *l1, *l2;
1008
1009 /*
1010 * On the default hierarchy, there can be csets which are
1011 * associated with the same set of cgroups but different csses.
1012 * Let's first ensure that csses match.
1013 */
1014 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1015 return false;
1016
1017
1018 /* @cset's domain should match the default cgroup's */
1019 if (cgroup_on_dfl(new_cgrp))
1020 new_dfl_cgrp = new_cgrp;
1021 else
1022 new_dfl_cgrp = old_cset->dfl_cgrp;
1023
1024 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1025 return false;
1026
1027 /*
1028 * Compare cgroup pointers in order to distinguish between
1029 * different cgroups in hierarchies. As different cgroups may
1030 * share the same effective css, this comparison is always
1031 * necessary.
1032 */
1033 l1 = &cset->cgrp_links;
1034 l2 = &old_cset->cgrp_links;
1035 while (1) {
1036 struct cgrp_cset_link *link1, *link2;
1037 struct cgroup *cgrp1, *cgrp2;
1038
1039 l1 = l1->next;
1040 l2 = l2->next;
1041 /* See if we reached the end - both lists are equal length. */
1042 if (l1 == &cset->cgrp_links) {
1043 BUG_ON(l2 != &old_cset->cgrp_links);
1044 break;
1045 } else {
1046 BUG_ON(l2 == &old_cset->cgrp_links);
1047 }
1048 /* Locate the cgroups associated with these links. */
1049 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1050 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1051 cgrp1 = link1->cgrp;
1052 cgrp2 = link2->cgrp;
1053 /* Hierarchies should be linked in the same order. */
1054 BUG_ON(cgrp1->root != cgrp2->root);
1055
1056 /*
1057 * If this hierarchy is the hierarchy of the cgroup
1058 * that's changing, then we need to check that this
1059 * css_set points to the new cgroup; if it's any other
1060 * hierarchy, then this css_set should point to the
1061 * same cgroup as the old css_set.
1062 */
1063 if (cgrp1->root == new_cgrp->root) {
1064 if (cgrp1 != new_cgrp)
1065 return false;
1066 } else {
1067 if (cgrp1 != cgrp2)
1068 return false;
1069 }
1070 }
1071 return true;
1072}
1073
1074/**
1075 * find_existing_css_set - init css array and find the matching css_set
1076 * @old_cset: the css_set that we're using before the cgroup transition
1077 * @cgrp: the cgroup that we're moving into
1078 * @template: out param for the new set of csses, should be clear on entry
1079 */
1080static struct css_set *find_existing_css_set(struct css_set *old_cset,
1081 struct cgroup *cgrp,
1082 struct cgroup_subsys_state **template)
1083{
1084 struct cgroup_root *root = cgrp->root;
1085 struct cgroup_subsys *ss;
1086 struct css_set *cset;
1087 unsigned long key;
1088 int i;
1089
1090 /*
1091 * Build the set of subsystem state objects that we want to see in the
1092 * new css_set. While subsystems can change globally, the entries here
1093 * won't change, so no need for locking.
1094 */
1095 for_each_subsys(ss, i) {
1096 if (unlikely(cgroup_rebind_ss_mask & (1UL << i))) {
1097 /*
1098 * @ss is leaving this hierarchy and its per-cgroup
1099 * csses are about to be killed. Resolve to the
1100 * surviving root css so the tasks are migrated there.
1101 */
1102 template[i] = cgroup_css(&root->cgrp, ss);
1103 WARN_ON_ONCE(!template[i]);
1104 } else if (root->subsys_mask & (1UL << i)) {
1105 /*
1106 * @ss is in this hierarchy, so we want the
1107 * effective css from @cgrp.
1108 */
1109 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1110 } else {
1111 /*
1112 * @ss is not in this hierarchy, so we don't want
1113 * to change the css.
1114 */
1115 template[i] = old_cset->subsys[i];
1116 }
1117 }
1118
1119 key = css_set_hash(template);
1120 hash_for_each_possible(css_set_table, cset, hlist, key) {
1121 if (!compare_css_sets(cset, old_cset, cgrp, template))
1122 continue;
1123
1124 /* This css_set matches what we need */
1125 return cset;
1126 }
1127
1128 /* No existing cgroup group matched */
1129 return NULL;
1130}
1131
1132static void free_cgrp_cset_links(struct list_head *links_to_free)
1133{
1134 struct cgrp_cset_link *link, *tmp_link;
1135
1136 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1137 list_del(&link->cset_link);
1138 kfree(link);
1139 }
1140}
1141
1142/**
1143 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1144 * @count: the number of links to allocate
1145 * @tmp_links: list_head the allocated links are put on
1146 *
1147 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1148 * through ->cset_link. Returns 0 on success or -errno.
1149 */
1150static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1151{
1152 struct cgrp_cset_link *link;
1153 int i;
1154
1155 INIT_LIST_HEAD(tmp_links);
1156
1157 for (i = 0; i < count; i++) {
1158 link = kzalloc_obj(*link);
1159 if (!link) {
1160 free_cgrp_cset_links(tmp_links);
1161 return -ENOMEM;
1162 }
1163 list_add(&link->cset_link, tmp_links);
1164 }
1165 return 0;
1166}
1167
1168/**
1169 * link_css_set - a helper function to link a css_set to a cgroup
1170 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1171 * @cset: the css_set to be linked
1172 * @cgrp: the destination cgroup
1173 */
1174static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1175 struct cgroup *cgrp)
1176{
1177 struct cgrp_cset_link *link;
1178
1179 BUG_ON(list_empty(tmp_links));
1180
1181 if (cgroup_on_dfl(cgrp))
1182 cset->dfl_cgrp = cgrp;
1183
1184 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1185 link->cset = cset;
1186 link->cgrp = cgrp;
1187
1188 /*
1189 * Always add links to the tail of the lists so that the lists are
1190 * in chronological order.
1191 */
1192 list_move_tail(&link->cset_link, &cgrp->cset_links);
1193 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1194
1195 if (cgroup_parent(cgrp))
1196 cgroup_get_live(cgrp);
1197}
1198
1199/**
1200 * find_css_set - return a new css_set with one cgroup updated
1201 * @old_cset: the baseline css_set
1202 * @cgrp: the cgroup to be updated
1203 *
1204 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1205 * substituted into the appropriate hierarchy.
1206 */
1207static struct css_set *find_css_set(struct css_set *old_cset,
1208 struct cgroup *cgrp)
1209{
1210 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1211 struct css_set *cset;
1212 struct list_head tmp_links;
1213 struct cgrp_cset_link *link;
1214 struct cgroup_subsys *ss;
1215 unsigned long key;
1216 int ssid;
1217
1218 lockdep_assert_held(&cgroup_mutex);
1219
1220 /* First see if we already have a cgroup group that matches
1221 * the desired set */
1222 spin_lock_irq(&css_set_lock);
1223 cset = find_existing_css_set(old_cset, cgrp, template);
1224 if (cset)
1225 get_css_set(cset);
1226 spin_unlock_irq(&css_set_lock);
1227
1228 if (cset)
1229 return cset;
1230
1231 cset = kzalloc_obj(*cset);
1232 if (!cset)
1233 return NULL;
1234
1235 /* Allocate all the cgrp_cset_link objects that we'll need */
1236 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1237 kfree(cset);
1238 return NULL;
1239 }
1240
1241 refcount_set(&cset->refcount, 1);
1242 cset->dom_cset = cset;
1243 INIT_LIST_HEAD(&cset->tasks);
1244 INIT_LIST_HEAD(&cset->mg_tasks);
1245 INIT_LIST_HEAD(&cset->dying_tasks);
1246 INIT_LIST_HEAD(&cset->task_iters);
1247 INIT_LIST_HEAD(&cset->threaded_csets);
1248 INIT_HLIST_NODE(&cset->hlist);
1249 INIT_LIST_HEAD(&cset->cgrp_links);
1250 INIT_LIST_HEAD(&cset->mg_src_preload_node);
1251 INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1252 INIT_LIST_HEAD(&cset->mg_node);
1253
1254 /* Copy the set of subsystem state objects generated in
1255 * find_existing_css_set() */
1256 memcpy(cset->subsys, template, sizeof(cset->subsys));
1257
1258 spin_lock_irq(&css_set_lock);
1259 /* Add reference counts and links from the new css_set. */
1260 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1261 struct cgroup *c = link->cgrp;
1262
1263 if (c->root == cgrp->root)
1264 c = cgrp;
1265 link_css_set(&tmp_links, cset, c);
1266 }
1267
1268 BUG_ON(!list_empty(&tmp_links));
1269
1270 css_set_count++;
1271
1272 /* Add @cset to the hash table */
1273 key = css_set_hash(cset->subsys);
1274 hash_add(css_set_table, &cset->hlist, key);
1275
1276 for_each_subsys(ss, ssid) {
1277 struct cgroup_subsys_state *css = cset->subsys[ssid];
1278
1279 list_add_tail(&cset->e_cset_node[ssid],
1280 &css->cgroup->e_csets[ssid]);
1281 css_get(css);
1282 }
1283
1284 spin_unlock_irq(&css_set_lock);
1285
1286 /*
1287 * If @cset should be threaded, look up the matching dom_cset and
1288 * link them up. We first fully initialize @cset then look for the
1289 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1290 * to stay empty until we return.
1291 */
1292 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1293 struct css_set *dcset;
1294
1295 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1296 if (!dcset) {
1297 put_css_set(cset);
1298 return NULL;
1299 }
1300
1301 spin_lock_irq(&css_set_lock);
1302 cset->dom_cset = dcset;
1303 list_add_tail(&cset->threaded_csets_node,
1304 &dcset->threaded_csets);
1305 spin_unlock_irq(&css_set_lock);
1306 }
1307
1308 return cset;
1309}
1310
1311struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1312{
1313 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1314
1315 return root_cgrp->root;
1316}
1317
1318void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1319{
1320 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1321
1322 /*
1323 * see the comment above CGRP_ROOT_FAVOR_DYNMODS definition.
1324 * favordynmods can flip while task is between
1325 * cgroup_threadgroup_change_begin() and end(), so down_write global
1326 * cgroup_threadgroup_rwsem to synchronize them.
1327 *
1328 * Once cgroup_enable_per_threadgroup_rwsem is enabled, holding
1329 * cgroup_threadgroup_rwsem doesn't exlude tasks between
1330 * cgroup_thread_group_change_begin() and end() and thus it's unsafe to
1331 * turn off. As the scenario is unlikely, simply disallow disabling once
1332 * enabled and print out a warning.
1333 */
1334 percpu_down_write(&cgroup_threadgroup_rwsem);
1335 if (favor && !favoring) {
1336 cgroup_enable_per_threadgroup_rwsem = true;
1337 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1338 root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1339 } else if (!favor && favoring) {
1340 if (cgroup_enable_per_threadgroup_rwsem)
1341 pr_warn_once("cgroup favordynmods: per threadgroup rwsem mechanism can't be disabled\n");
1342 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1343 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1344 }
1345 percpu_up_write(&cgroup_threadgroup_rwsem);
1346}
1347
1348static int cgroup_init_root_id(struct cgroup_root *root)
1349{
1350 int id;
1351
1352 lockdep_assert_held(&cgroup_mutex);
1353
1354 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1355 if (id < 0)
1356 return id;
1357
1358 root->hierarchy_id = id;
1359 return 0;
1360}
1361
1362static void cgroup_exit_root_id(struct cgroup_root *root)
1363{
1364 lockdep_assert_held(&cgroup_mutex);
1365
1366 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1367}
1368
1369void cgroup_free_root(struct cgroup_root *root)
1370{
1371 kfree_rcu(root, rcu);
1372}
1373
1374static void cgroup_destroy_root(struct cgroup_root *root)
1375{
1376 struct cgroup *cgrp = &root->cgrp;
1377 struct cgrp_cset_link *link, *tmp_link;
1378 int ret;
1379
1380 trace_cgroup_destroy_root(root);
1381
1382 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1383
1384 BUG_ON(atomic_read(&root->nr_cgrps));
1385 BUG_ON(!list_empty(&cgrp->self.children));
1386
1387 ret = blocking_notifier_call_chain(&cgroup_lifetime_notifier,
1388 CGROUP_LIFETIME_OFFLINE, cgrp);
1389 WARN_ON_ONCE(notifier_to_errno(ret));
1390
1391 /* Rebind all subsystems back to the default hierarchy */
1392 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1393
1394 /*
1395 * Release all the links from cset_links to this hierarchy's
1396 * root cgroup
1397 */
1398 spin_lock_irq(&css_set_lock);
1399
1400 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1401 list_del(&link->cset_link);
1402 list_del(&link->cgrp_link);
1403 kfree(link);
1404 }
1405
1406 spin_unlock_irq(&css_set_lock);
1407
1408 WARN_ON_ONCE(list_empty(&root->root_list));
1409 list_del_rcu(&root->root_list);
1410 cgroup_root_count--;
1411
1412 if (!have_favordynmods)
1413 cgroup_favor_dynmods(root, false);
1414
1415 cgroup_exit_root_id(root);
1416
1417 cgroup_unlock();
1418
1419 kernfs_destroy_root(root->kf_root);
1420 cgroup_free_root(root);
1421}
1422
1423/*
1424 * Returned cgroup is without refcount but it's valid as long as cset pins it.
1425 */
1426static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1427 struct cgroup_root *root)
1428{
1429 struct cgroup *res_cgroup = NULL;
1430
1431 if (cset == &init_css_set) {
1432 res_cgroup = &root->cgrp;
1433 } else if (root == &cgrp_dfl_root) {
1434 res_cgroup = cset->dfl_cgrp;
1435 } else {
1436 struct cgrp_cset_link *link;
1437 lockdep_assert_held(&css_set_lock);
1438
1439 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1440 struct cgroup *c = link->cgrp;
1441
1442 if (c->root == root) {
1443 res_cgroup = c;
1444 break;
1445 }
1446 }
1447 }
1448
1449 /*
1450 * If cgroup_mutex is not held, the cgrp_cset_link will be freed
1451 * before we remove the cgroup root from the root_list. Consequently,
1452 * when accessing a cgroup root, the cset_link may have already been
1453 * freed, resulting in a NULL res_cgroup. However, by holding the
1454 * cgroup_mutex, we ensure that res_cgroup can't be NULL.
1455 * If we don't hold cgroup_mutex in the caller, we must do the NULL
1456 * check.
1457 */
1458 return res_cgroup;
1459}
1460
1461/*
1462 * look up cgroup associated with current task's cgroup namespace on the
1463 * specified hierarchy
1464 */
1465static struct cgroup *
1466current_cgns_cgroup_from_root(struct cgroup_root *root)
1467{
1468 struct cgroup *res = NULL;
1469 struct css_set *cset;
1470
1471 lockdep_assert_held(&css_set_lock);
1472
1473 rcu_read_lock();
1474
1475 cset = current->nsproxy->cgroup_ns->root_cset;
1476 res = __cset_cgroup_from_root(cset, root);
1477
1478 rcu_read_unlock();
1479
1480 /*
1481 * The namespace_sem is held by current, so the root cgroup can't
1482 * be umounted. Therefore, we can ensure that the res is non-NULL.
1483 */
1484 WARN_ON_ONCE(!res);
1485 return res;
1486}
1487
1488/*
1489 * Look up cgroup associated with current task's cgroup namespace on the default
1490 * hierarchy.
1491 *
1492 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1493 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1494 * pointers.
1495 * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1496 * - As a bonus returned cgrp is pinned with the current because it cannot
1497 * switch cgroup_ns asynchronously.
1498 */
1499static struct cgroup *current_cgns_cgroup_dfl(void)
1500{
1501 struct css_set *cset;
1502
1503 if (current->nsproxy) {
1504 cset = current->nsproxy->cgroup_ns->root_cset;
1505 return __cset_cgroup_from_root(cset, &cgrp_dfl_root);
1506 } else {
1507 /*
1508 * NOTE: This function may be called from bpf_cgroup_from_id()
1509 * on a task which has already passed exit_nsproxy_namespaces()
1510 * and nsproxy == NULL. Fall back to cgrp_dfl_root which will
1511 * make all cgroups visible for lookups.
1512 */
1513 return &cgrp_dfl_root.cgrp;
1514 }
1515}
1516
1517/* look up cgroup associated with given css_set on the specified hierarchy */
1518static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1519 struct cgroup_root *root)
1520{
1521 lockdep_assert_held(&css_set_lock);
1522
1523 return __cset_cgroup_from_root(cset, root);
1524}
1525
1526/*
1527 * Return the cgroup for "task" from the given hierarchy. Must be
1528 * called with css_set_lock held to prevent task's groups from being modified.
1529 * Must be called with either cgroup_mutex or rcu read lock to prevent the
1530 * cgroup root from being destroyed.
1531 */
1532struct cgroup *task_cgroup_from_root(struct task_struct *task,
1533 struct cgroup_root *root)
1534{
1535 /*
1536 * No need to lock the task - since we hold css_set_lock the
1537 * task can't change groups.
1538 */
1539 return cset_cgroup_from_root(task_css_set(task), root);
1540}
1541
1542/*
1543 * A task must hold cgroup_mutex to modify cgroups.
1544 *
1545 * Any task can increment and decrement the count field without lock.
1546 * So in general, code holding cgroup_mutex can't rely on the count
1547 * field not changing. However, if the count goes to zero, then only
1548 * cgroup_attach_task() can increment it again. Because a count of zero
1549 * means that no tasks are currently attached, therefore there is no
1550 * way a task attached to that cgroup can fork (the other way to
1551 * increment the count). So code holding cgroup_mutex can safely
1552 * assume that if the count is zero, it will stay zero. Similarly, if
1553 * a task holds cgroup_mutex on a cgroup with zero count, it
1554 * knows that the cgroup won't be removed, as cgroup_rmdir()
1555 * needs that mutex.
1556 *
1557 * A cgroup can only be deleted if both its 'count' of using tasks
1558 * is zero, and its list of 'children' cgroups is empty. Since all
1559 * tasks in the system use _some_ cgroup, and since there is always at
1560 * least one task in the system (init, pid == 1), therefore, root cgroup
1561 * always has either children cgroups and/or using tasks. So we don't
1562 * need a special hack to ensure that root cgroup cannot be deleted.
1563 *
1564 * P.S. One more locking exception. RCU is used to guard the
1565 * update of a tasks cgroup pointer by cgroup_attach_task()
1566 */
1567
1568static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1569
1570static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1571 char *buf)
1572{
1573 struct cgroup_subsys *ss = cft->ss;
1574
1575 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1576 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1577 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1578
1579 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1580 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1581 cft->name);
1582 } else {
1583 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1584 }
1585 return buf;
1586}
1587
1588/**
1589 * cgroup_file_mode - deduce file mode of a control file
1590 * @cft: the control file in question
1591 *
1592 * S_IRUGO for read, S_IWUSR for write.
1593 */
1594static umode_t cgroup_file_mode(const struct cftype *cft)
1595{
1596 umode_t mode = 0;
1597
1598 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1599 mode |= S_IRUGO;
1600
1601 if (cft->write_u64 || cft->write_s64 || cft->write) {
1602 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1603 mode |= S_IWUGO;
1604 else
1605 mode |= S_IWUSR;
1606 }
1607
1608 return mode;
1609}
1610
1611/**
1612 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1613 * @subtree_control: the new subtree_control mask to consider
1614 * @this_ss_mask: available subsystems
1615 *
1616 * On the default hierarchy, a subsystem may request other subsystems to be
1617 * enabled together through its ->depends_on mask. In such cases, more
1618 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1619 *
1620 * This function calculates which subsystems need to be enabled if
1621 * @subtree_control is to be applied while restricted to @this_ss_mask.
1622 */
1623static u32 cgroup_calc_subtree_ss_mask(u32 subtree_control, u32 this_ss_mask)
1624{
1625 u32 cur_ss_mask = subtree_control;
1626 struct cgroup_subsys *ss;
1627 int ssid;
1628
1629 lockdep_assert_held(&cgroup_mutex);
1630
1631 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1632
1633 while (true) {
1634 u32 new_ss_mask = cur_ss_mask;
1635
1636 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1637 new_ss_mask |= ss->depends_on;
1638 } while_each_subsys_mask();
1639
1640 /*
1641 * Mask out subsystems which aren't available. This can
1642 * happen only if some depended-upon subsystems were bound
1643 * to non-default hierarchies.
1644 */
1645 new_ss_mask &= this_ss_mask;
1646
1647 if (new_ss_mask == cur_ss_mask)
1648 break;
1649 cur_ss_mask = new_ss_mask;
1650 }
1651
1652 return cur_ss_mask;
1653}
1654
1655/**
1656 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1657 * @kn: the kernfs_node being serviced
1658 *
1659 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1660 * the method finishes if locking succeeded. Note that once this function
1661 * returns the cgroup returned by cgroup_kn_lock_live() may become
1662 * inaccessible any time. If the caller intends to continue to access the
1663 * cgroup, it should pin it before invoking this function.
1664 */
1665void cgroup_kn_unlock(struct kernfs_node *kn)
1666{
1667 struct cgroup *cgrp;
1668
1669 if (kernfs_type(kn) == KERNFS_DIR)
1670 cgrp = kn->priv;
1671 else
1672 cgrp = kn_priv(kn);
1673
1674 cgroup_unlock();
1675
1676 kernfs_unbreak_active_protection(kn);
1677 cgroup_put(cgrp);
1678}
1679
1680/**
1681 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1682 * @kn: the kernfs_node being serviced
1683 * @drain_offline: perform offline draining on the cgroup
1684 *
1685 * This helper is to be used by a cgroup kernfs method currently servicing
1686 * @kn. It breaks the active protection, performs cgroup locking and
1687 * verifies that the associated cgroup is alive. Returns the cgroup if
1688 * alive; otherwise, %NULL. A successful return should be undone by a
1689 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1690 * cgroup is drained of offlining csses before return.
1691 *
1692 * Any cgroup kernfs method implementation which requires locking the
1693 * associated cgroup should use this helper. It avoids nesting cgroup
1694 * locking under kernfs active protection and allows all kernfs operations
1695 * including self-removal.
1696 */
1697struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1698{
1699 struct cgroup *cgrp;
1700
1701 if (kernfs_type(kn) == KERNFS_DIR)
1702 cgrp = kn->priv;
1703 else
1704 cgrp = kn_priv(kn);
1705
1706 /*
1707 * We're gonna grab cgroup_mutex which nests outside kernfs
1708 * active_ref. cgroup liveliness check alone provides enough
1709 * protection against removal. Ensure @cgrp stays accessible and
1710 * break the active_ref protection.
1711 */
1712 if (!cgroup_tryget(cgrp))
1713 return NULL;
1714 kernfs_break_active_protection(kn);
1715
1716 if (drain_offline)
1717 cgroup_lock_and_drain_offline(cgrp);
1718 else
1719 cgroup_lock();
1720
1721 if (!cgroup_is_dead(cgrp))
1722 return cgrp;
1723
1724 cgroup_kn_unlock(kn);
1725 return NULL;
1726}
1727
1728static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1729{
1730 char name[CGROUP_FILE_NAME_MAX];
1731
1732 lockdep_assert_held(&cgroup_mutex);
1733
1734 if (cft->file_offset) {
1735 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1736 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1737
1738 spin_lock_irq(&cfile->lock);
1739 WRITE_ONCE(cfile->kn, NULL);
1740 spin_unlock_irq(&cfile->lock);
1741
1742 timer_delete_sync(&cfile->notify_timer);
1743 }
1744
1745 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1746}
1747
1748/**
1749 * css_clear_dir - remove subsys files in a cgroup directory
1750 * @css: target css
1751 */
1752static void css_clear_dir(struct cgroup_subsys_state *css)
1753{
1754 struct cgroup *cgrp = css->cgroup;
1755 struct cftype *cfts;
1756
1757 if (!(css->flags & CSS_VISIBLE))
1758 return;
1759
1760 css->flags &= ~CSS_VISIBLE;
1761
1762 if (css_is_self(css)) {
1763 if (cgroup_on_dfl(cgrp)) {
1764 cgroup_addrm_files(css, cgrp,
1765 cgroup_base_files, false);
1766 if (cgroup_psi_enabled())
1767 cgroup_addrm_files(css, cgrp,
1768 cgroup_psi_files, false);
1769 } else {
1770 cgroup_addrm_files(css, cgrp,
1771 cgroup1_base_files, false);
1772 }
1773 } else {
1774 list_for_each_entry(cfts, &css->ss->cfts, node)
1775 cgroup_addrm_files(css, cgrp, cfts, false);
1776 }
1777}
1778
1779/**
1780 * css_populate_dir - create subsys files in a cgroup directory
1781 * @css: target css
1782 *
1783 * On failure, no file is added.
1784 */
1785static int css_populate_dir(struct cgroup_subsys_state *css)
1786{
1787 struct cgroup *cgrp = css->cgroup;
1788 struct cftype *cfts, *failed_cfts;
1789 int ret;
1790
1791 if (css->flags & CSS_VISIBLE)
1792 return 0;
1793
1794 if (css_is_self(css)) {
1795 if (cgroup_on_dfl(cgrp)) {
1796 ret = cgroup_addrm_files(css, cgrp,
1797 cgroup_base_files, true);
1798 if (ret < 0)
1799 return ret;
1800
1801 if (cgroup_psi_enabled()) {
1802 ret = cgroup_addrm_files(css, cgrp,
1803 cgroup_psi_files, true);
1804 if (ret < 0) {
1805 cgroup_addrm_files(css, cgrp,
1806 cgroup_base_files, false);
1807 return ret;
1808 }
1809 }
1810 } else {
1811 ret = cgroup_addrm_files(css, cgrp,
1812 cgroup1_base_files, true);
1813 if (ret < 0)
1814 return ret;
1815 }
1816 } else {
1817 list_for_each_entry(cfts, &css->ss->cfts, node) {
1818 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1819 if (ret < 0) {
1820 failed_cfts = cfts;
1821 goto err;
1822 }
1823 }
1824 }
1825
1826 css->flags |= CSS_VISIBLE;
1827
1828 return 0;
1829err:
1830 list_for_each_entry(cfts, &css->ss->cfts, node) {
1831 if (cfts == failed_cfts)
1832 break;
1833 cgroup_addrm_files(css, cgrp, cfts, false);
1834 }
1835 return ret;
1836}
1837
1838int rebind_subsystems(struct cgroup_root *dst_root, u32 ss_mask)
1839{
1840 struct cgroup *dcgrp = &dst_root->cgrp;
1841 struct cgroup_subsys *ss;
1842 int ssid, ret;
1843 u32 dfl_disable_ss_mask = 0;
1844
1845 lockdep_assert_held(&cgroup_mutex);
1846
1847 do_each_subsys_mask(ss, ssid, ss_mask) {
1848 /*
1849 * If @ss has non-root csses attached to it, can't move.
1850 * If @ss is an implicit controller, it is exempt from this
1851 * rule and can be stolen.
1852 */
1853 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1854 !ss->implicit_on_dfl)
1855 return -EBUSY;
1856
1857 /* can't move between two non-dummy roots either */
1858 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1859 return -EBUSY;
1860
1861 /*
1862 * Collect ssid's that need to be disabled from default
1863 * hierarchy.
1864 */
1865 if (ss->root == &cgrp_dfl_root)
1866 dfl_disable_ss_mask |= 1 << ssid;
1867
1868 } while_each_subsys_mask();
1869
1870 if (dfl_disable_ss_mask) {
1871 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1872
1873 /*
1874 * Controllers leaving the default hierarchy are disabled
1875 * together. cgroup_rebind_ss_mask makes cgroup_apply_control()
1876 * migrate their tasks to the root css, so the per-cgroup csses
1877 * are unpopulated when cgroup_finalize_control() kills them.
1878 * Clear it before cgroup_finalize_control(), which does no
1879 * css_set lookup.
1880 */
1881 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1882 cgroup_rebind_ss_mask = dfl_disable_ss_mask;
1883 WARN_ON(cgroup_apply_control(scgrp));
1884 cgroup_rebind_ss_mask = 0;
1885 cgroup_finalize_control(scgrp, 0);
1886 }
1887
1888 do_each_subsys_mask(ss, ssid, ss_mask) {
1889 struct cgroup_root *src_root = ss->root;
1890 struct cgroup *scgrp = &src_root->cgrp;
1891 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1892 struct css_set *cset, *cset_pos;
1893 struct css_task_iter *it;
1894
1895 WARN_ON(!css || cgroup_css(dcgrp, ss));
1896
1897 if (src_root != &cgrp_dfl_root) {
1898 /*
1899 * Disable from the source, migrating its tasks to the
1900 * root css first (see cgroup_rebind_ss_mask).
1901 */
1902 src_root->subsys_mask &= ~(1 << ssid);
1903 cgroup_rebind_ss_mask = 1 << ssid;
1904 WARN_ON(cgroup_apply_control(scgrp));
1905 cgroup_rebind_ss_mask = 0;
1906 cgroup_finalize_control(scgrp, 0);
1907 }
1908
1909 /* rebind */
1910 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1911 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1912 ss->root = dst_root;
1913
1914 spin_lock_irq(&css_set_lock);
1915 css->cgroup = dcgrp;
1916 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id]));
1917 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id],
1918 e_cset_node[ss->id]) {
1919 list_move_tail(&cset->e_cset_node[ss->id],
1920 &dcgrp->e_csets[ss->id]);
1921 /*
1922 * all css_sets of scgrp together in same order to dcgrp,
1923 * patch in-flight iterators to preserve correct iteration.
1924 * since the iterator is always advanced right away and
1925 * finished when it->cset_pos meets it->cset_head, so only
1926 * update it->cset_head is enough here.
1927 */
1928 list_for_each_entry(it, &cset->task_iters, iters_node)
1929 if (it->cset_head == &scgrp->e_csets[ss->id])
1930 it->cset_head = &dcgrp->e_csets[ss->id];
1931 }
1932 spin_unlock_irq(&css_set_lock);
1933
1934 /* default hierarchy doesn't enable controllers by default */
1935 dst_root->subsys_mask |= 1 << ssid;
1936 if (dst_root == &cgrp_dfl_root) {
1937 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1938 } else {
1939 dcgrp->subtree_control |= 1 << ssid;
1940 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1941 }
1942
1943 ret = cgroup_apply_control(dcgrp);
1944 if (ret)
1945 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1946 ss->name, ret);
1947
1948 if (ss->bind)
1949 ss->bind(css);
1950 } while_each_subsys_mask();
1951
1952 kernfs_activate(dcgrp->kn);
1953 return 0;
1954}
1955
1956int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1957 struct kernfs_root *kf_root)
1958{
1959 int len = 0;
1960 char *buf = NULL;
1961 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1962 struct cgroup *ns_cgroup;
1963
1964 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1965 if (!buf)
1966 return -ENOMEM;
1967
1968 spin_lock_irq(&css_set_lock);
1969 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1970 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1971 spin_unlock_irq(&css_set_lock);
1972
1973 if (len == -E2BIG)
1974 len = -ERANGE;
1975 else if (len > 0) {
1976 seq_escape(sf, buf, " \t\n\\");
1977 len = 0;
1978 }
1979 kfree(buf);
1980 return len;
1981}
1982
1983enum cgroup2_param {
1984 Opt_nsdelegate,
1985 Opt_favordynmods,
1986 Opt_memory_localevents,
1987 Opt_memory_recursiveprot,
1988 Opt_memory_hugetlb_accounting,
1989 Opt_pids_localevents,
1990 nr__cgroup2_params
1991};
1992
1993static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1994 fsparam_flag("nsdelegate", Opt_nsdelegate),
1995 fsparam_flag("favordynmods", Opt_favordynmods),
1996 fsparam_flag("memory_localevents", Opt_memory_localevents),
1997 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1998 fsparam_flag("memory_hugetlb_accounting", Opt_memory_hugetlb_accounting),
1999 fsparam_flag("pids_localevents", Opt_pids_localevents),
2000 {}
2001};
2002
2003static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
2004{
2005 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2006 struct fs_parse_result result;
2007 int opt;
2008
2009 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
2010 if (opt < 0)
2011 return opt;
2012
2013 switch (opt) {
2014 case Opt_nsdelegate:
2015 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
2016 return 0;
2017 case Opt_favordynmods:
2018 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2019 return 0;
2020 case Opt_memory_localevents:
2021 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
2022 return 0;
2023 case Opt_memory_recursiveprot:
2024 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2025 return 0;
2026 case Opt_memory_hugetlb_accounting:
2027 ctx->flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2028 return 0;
2029 case Opt_pids_localevents:
2030 ctx->flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS;
2031 return 0;
2032 }
2033 return -EINVAL;
2034}
2035
2036struct cgroup_of_peak *of_peak(struct kernfs_open_file *of)
2037{
2038 struct cgroup_file_ctx *ctx = of->priv;
2039
2040 return &ctx->peak;
2041}
2042
2043static void apply_cgroup_root_flags(unsigned int root_flags)
2044{
2045 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
2046 if (root_flags & CGRP_ROOT_NS_DELEGATE)
2047 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
2048 else
2049 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
2050
2051 cgroup_favor_dynmods(&cgrp_dfl_root,
2052 root_flags & CGRP_ROOT_FAVOR_DYNMODS);
2053
2054 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
2055 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
2056 else
2057 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
2058
2059 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
2060 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2061 else
2062 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2063
2064 if (root_flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2065 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2066 else
2067 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2068
2069 if (root_flags & CGRP_ROOT_PIDS_LOCAL_EVENTS)
2070 cgrp_dfl_root.flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS;
2071 else
2072 cgrp_dfl_root.flags &= ~CGRP_ROOT_PIDS_LOCAL_EVENTS;
2073 }
2074}
2075
2076static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
2077{
2078 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
2079 seq_puts(seq, ",nsdelegate");
2080 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
2081 seq_puts(seq, ",favordynmods");
2082 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
2083 seq_puts(seq, ",memory_localevents");
2084 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
2085 seq_puts(seq, ",memory_recursiveprot");
2086 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2087 seq_puts(seq, ",memory_hugetlb_accounting");
2088 if (cgrp_dfl_root.flags & CGRP_ROOT_PIDS_LOCAL_EVENTS)
2089 seq_puts(seq, ",pids_localevents");
2090 return 0;
2091}
2092
2093static int cgroup_reconfigure(struct fs_context *fc)
2094{
2095 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2096
2097 apply_cgroup_root_flags(ctx->flags);
2098 return 0;
2099}
2100
2101static void init_cgroup_housekeeping(struct cgroup *cgrp)
2102{
2103 struct cgroup_subsys *ss;
2104 int ssid;
2105
2106 INIT_LIST_HEAD(&cgrp->self.sibling);
2107 INIT_LIST_HEAD(&cgrp->self.children);
2108 INIT_LIST_HEAD(&cgrp->cset_links);
2109 INIT_LIST_HEAD(&cgrp->pidlists);
2110 mutex_init(&cgrp->pidlist_mutex);
2111 cgrp->self.cgroup = cgrp;
2112 cgrp->self.flags |= CSS_ONLINE;
2113 cgrp->dom_cgrp = cgrp;
2114 cgrp->max_descendants = INT_MAX;
2115 cgrp->max_depth = INT_MAX;
2116 prev_cputime_init(&cgrp->prev_cputime);
2117
2118 for_each_subsys(ss, ssid)
2119 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2120
2121#ifdef CONFIG_CGROUP_BPF
2122 for (int i = 0; i < ARRAY_SIZE(cgrp->bpf.revisions); i++)
2123 cgrp->bpf.revisions[i] = 1;
2124#endif
2125
2126 init_waitqueue_head(&cgrp->offline_waitq);
2127 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2128}
2129
2130void init_cgroup_root(struct cgroup_fs_context *ctx)
2131{
2132 struct cgroup_root *root = ctx->root;
2133 struct cgroup *cgrp = &root->cgrp;
2134
2135 INIT_LIST_HEAD_RCU(&root->root_list);
2136 atomic_set(&root->nr_cgrps, 1);
2137 cgrp->root = root;
2138 init_cgroup_housekeeping(cgrp);
2139
2140 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2141 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2142 if (ctx->release_agent)
2143 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2144 if (ctx->name)
2145 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2146 if (ctx->cpuset_clone_children)
2147 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2148}
2149
2150int cgroup_setup_root(struct cgroup_root *root, u32 ss_mask)
2151{
2152 LIST_HEAD(tmp_links);
2153 struct cgroup *root_cgrp = &root->cgrp;
2154 struct kernfs_syscall_ops *kf_sops;
2155 struct css_set *cset;
2156 int i, ret;
2157
2158 lockdep_assert_held(&cgroup_mutex);
2159
2160 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2161 0, GFP_KERNEL);
2162 if (ret)
2163 goto out;
2164
2165 /*
2166 * We're accessing css_set_count without locking css_set_lock here,
2167 * but that's OK - it can only be increased by someone holding
2168 * cgroup_lock, and that's us. Later rebinding may disable
2169 * controllers on the default hierarchy and thus create new csets,
2170 * which can't be more than the existing ones. Allocate 2x.
2171 */
2172 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2173 if (ret)
2174 goto cancel_ref;
2175
2176 ret = cgroup_init_root_id(root);
2177 if (ret)
2178 goto cancel_ref;
2179
2180 kf_sops = root == &cgrp_dfl_root ?
2181 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2182
2183 root->kf_root = kernfs_create_root(kf_sops,
2184 KERNFS_ROOT_CREATE_DEACTIVATED |
2185 KERNFS_ROOT_SUPPORT_EXPORTOP |
2186 KERNFS_ROOT_SUPPORT_USER_XATTR |
2187 KERNFS_ROOT_INVARIANT_PARENT,
2188 root_cgrp);
2189 if (IS_ERR(root->kf_root)) {
2190 ret = PTR_ERR(root->kf_root);
2191 goto exit_root_id;
2192 }
2193 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2194 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2195 root_cgrp->ancestors[0] = root_cgrp;
2196
2197 ret = css_populate_dir(&root_cgrp->self);
2198 if (ret)
2199 goto destroy_root;
2200
2201 ret = css_rstat_init(&root_cgrp->self);
2202 if (ret)
2203 goto destroy_root;
2204
2205 ret = rebind_subsystems(root, ss_mask);
2206 if (ret)
2207 goto exit_stats;
2208
2209 ret = blocking_notifier_call_chain(&cgroup_lifetime_notifier,
2210 CGROUP_LIFETIME_ONLINE, root_cgrp);
2211 WARN_ON_ONCE(notifier_to_errno(ret));
2212
2213 trace_cgroup_setup_root(root);
2214
2215 /*
2216 * There must be no failure case after here, since rebinding takes
2217 * care of subsystems' refcounts, which are explicitly dropped in
2218 * the failure exit path.
2219 */
2220 list_add_rcu(&root->root_list, &cgroup_roots);
2221 cgroup_root_count++;
2222
2223 /*
2224 * Link the root cgroup in this hierarchy into all the css_set
2225 * objects.
2226 */
2227 spin_lock_irq(&css_set_lock);
2228 hash_for_each(css_set_table, i, cset, hlist) {
2229 link_css_set(&tmp_links, cset, root_cgrp);
2230 if (css_set_populated(cset))
2231 css_update_populated(&root_cgrp->self, true);
2232 }
2233 spin_unlock_irq(&css_set_lock);
2234
2235 BUG_ON(!list_empty(&root_cgrp->self.children));
2236 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2237
2238 ret = 0;
2239 goto out;
2240
2241exit_stats:
2242 css_rstat_exit(&root_cgrp->self);
2243destroy_root:
2244 kernfs_destroy_root(root->kf_root);
2245 root->kf_root = NULL;
2246exit_root_id:
2247 cgroup_exit_root_id(root);
2248cancel_ref:
2249 percpu_ref_exit(&root_cgrp->self.refcnt);
2250out:
2251 free_cgrp_cset_links(&tmp_links);
2252 return ret;
2253}
2254
2255int cgroup_do_get_tree(struct fs_context *fc)
2256{
2257 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2258 int ret;
2259
2260 ctx->kfc.root = ctx->root->kf_root;
2261 if (fc->fs_type == &cgroup2_fs_type)
2262 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2263 else
2264 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2265 ret = kernfs_get_tree(fc);
2266
2267 /*
2268 * In non-init cgroup namespace, instead of root cgroup's dentry,
2269 * we return the dentry corresponding to the cgroupns->root_cgrp.
2270 */
2271 if (!ret && ctx->ns != &init_cgroup_ns) {
2272 struct dentry *nsdentry;
2273 struct super_block *sb = fc->root->d_sb;
2274 struct cgroup *cgrp;
2275
2276 cgroup_lock();
2277 spin_lock_irq(&css_set_lock);
2278
2279 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2280
2281 spin_unlock_irq(&css_set_lock);
2282 cgroup_unlock();
2283
2284 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2285 dput(fc->root);
2286 if (IS_ERR(nsdentry)) {
2287 deactivate_locked_super(sb);
2288 ret = PTR_ERR(nsdentry);
2289 nsdentry = NULL;
2290 }
2291 fc->root = nsdentry;
2292 }
2293
2294 if (!ctx->kfc.new_sb_created)
2295 cgroup_put(&ctx->root->cgrp);
2296
2297 return ret;
2298}
2299
2300/*
2301 * Destroy a cgroup filesystem context.
2302 */
2303static void cgroup_fs_context_free(struct fs_context *fc)
2304{
2305 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2306
2307 kfree(ctx->name);
2308 kfree(ctx->release_agent);
2309 put_cgroup_ns(ctx->ns);
2310 kernfs_free_fs_context(fc);
2311 kfree(ctx);
2312}
2313
2314static int cgroup_get_tree(struct fs_context *fc)
2315{
2316 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2317 int ret;
2318
2319 WRITE_ONCE(cgrp_dfl_visible, true);
2320 cgroup_get_live(&cgrp_dfl_root.cgrp);
2321 ctx->root = &cgrp_dfl_root;
2322
2323 ret = cgroup_do_get_tree(fc);
2324 if (!ret)
2325 apply_cgroup_root_flags(ctx->flags);
2326 return ret;
2327}
2328
2329static const struct fs_context_operations cgroup_fs_context_ops = {
2330 .free = cgroup_fs_context_free,
2331 .parse_param = cgroup2_parse_param,
2332 .get_tree = cgroup_get_tree,
2333 .reconfigure = cgroup_reconfigure,
2334};
2335
2336static const struct fs_context_operations cgroup1_fs_context_ops = {
2337 .free = cgroup_fs_context_free,
2338 .parse_param = cgroup1_parse_param,
2339 .get_tree = cgroup1_get_tree,
2340 .reconfigure = cgroup1_reconfigure,
2341};
2342
2343/*
2344 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2345 * we select the namespace we're going to use.
2346 */
2347static int cgroup_init_fs_context(struct fs_context *fc)
2348{
2349 struct cgroup_fs_context *ctx;
2350
2351 ctx = kzalloc_obj(struct cgroup_fs_context);
2352 if (!ctx)
2353 return -ENOMEM;
2354
2355 ctx->ns = current->nsproxy->cgroup_ns;
2356 get_cgroup_ns(ctx->ns);
2357 fc->fs_private = &ctx->kfc;
2358 if (fc->fs_type == &cgroup2_fs_type)
2359 fc->ops = &cgroup_fs_context_ops;
2360 else
2361 fc->ops = &cgroup1_fs_context_ops;
2362 put_user_ns(fc->user_ns);
2363 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2364 fc->global = true;
2365
2366 if (have_favordynmods)
2367 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2368
2369 return 0;
2370}
2371
2372static void cgroup_kill_sb(struct super_block *sb)
2373{
2374 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2375 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2376
2377 /*
2378 * If @root doesn't have any children, start killing it.
2379 * This prevents new mounts by disabling percpu_ref_tryget_live().
2380 *
2381 * And don't kill the default root.
2382 */
2383 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2384 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2385 percpu_ref_kill(&root->cgrp.self.refcnt);
2386 cgroup_put(&root->cgrp);
2387 kernfs_kill_sb(sb);
2388}
2389
2390struct file_system_type cgroup_fs_type = {
2391 .name = "cgroup",
2392 .init_fs_context = cgroup_init_fs_context,
2393 .parameters = cgroup1_fs_parameters,
2394 .kill_sb = cgroup_kill_sb,
2395 .fs_flags = FS_USERNS_MOUNT,
2396};
2397
2398static struct file_system_type cgroup2_fs_type = {
2399 .name = "cgroup2",
2400 .init_fs_context = cgroup_init_fs_context,
2401 .parameters = cgroup2_fs_parameters,
2402 .kill_sb = cgroup_kill_sb,
2403 .fs_flags = FS_USERNS_MOUNT,
2404};
2405
2406#ifdef CONFIG_CPUSETS_V1
2407enum cpuset_param {
2408 Opt_cpuset_v2_mode,
2409};
2410
2411static const struct fs_parameter_spec cpuset_fs_parameters[] = {
2412 fsparam_flag ("cpuset_v2_mode", Opt_cpuset_v2_mode),
2413 {}
2414};
2415
2416static int cpuset_parse_param(struct fs_context *fc, struct fs_parameter *param)
2417{
2418 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2419 struct fs_parse_result result;
2420 int opt;
2421
2422 opt = fs_parse(fc, cpuset_fs_parameters, param, &result);
2423 if (opt < 0)
2424 return opt;
2425
2426 switch (opt) {
2427 case Opt_cpuset_v2_mode:
2428 ctx->flags |= CGRP_ROOT_CPUSET_V2_MODE;
2429 return 0;
2430 }
2431 return -EINVAL;
2432}
2433
2434static const struct fs_context_operations cpuset_fs_context_ops = {
2435 .get_tree = cgroup1_get_tree,
2436 .free = cgroup_fs_context_free,
2437 .parse_param = cpuset_parse_param,
2438};
2439
2440/*
2441 * This is ugly, but preserves the userspace API for existing cpuset
2442 * users. If someone tries to mount the "cpuset" filesystem, we
2443 * silently switch it to mount "cgroup" instead
2444 */
2445static int cpuset_init_fs_context(struct fs_context *fc)
2446{
2447 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2448 struct cgroup_fs_context *ctx;
2449 int err;
2450
2451 err = cgroup_init_fs_context(fc);
2452 if (err) {
2453 kfree(agent);
2454 return err;
2455 }
2456
2457 fc->ops = &cpuset_fs_context_ops;
2458
2459 ctx = cgroup_fc2context(fc);
2460 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2461 ctx->flags |= CGRP_ROOT_NOPREFIX;
2462 ctx->release_agent = agent;
2463
2464 get_filesystem(&cgroup_fs_type);
2465 put_filesystem(fc->fs_type);
2466 fc->fs_type = &cgroup_fs_type;
2467
2468 return 0;
2469}
2470
2471static struct file_system_type cpuset_fs_type = {
2472 .name = "cpuset",
2473 .init_fs_context = cpuset_init_fs_context,
2474 .parameters = cpuset_fs_parameters,
2475 .fs_flags = FS_USERNS_MOUNT,
2476};
2477#endif
2478
2479int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2480 struct cgroup_namespace *ns)
2481{
2482 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2483
2484 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2485}
2486
2487int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2488 struct cgroup_namespace *ns)
2489{
2490 int ret;
2491
2492 cgroup_lock();
2493 spin_lock_irq(&css_set_lock);
2494
2495 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2496
2497 spin_unlock_irq(&css_set_lock);
2498 cgroup_unlock();
2499
2500 return ret;
2501}
2502EXPORT_SYMBOL_GPL(cgroup_path_ns);
2503
2504/**
2505 * cgroup_attach_lock - Lock for ->attach()
2506 * @lock_mode: whether acquire and acquire which rwsem
2507 * @tsk: thread group to lock
2508 *
2509 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2510 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2511 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2512 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2513 * lead to deadlocks.
2514 *
2515 * Bringing up a CPU may involve creating and destroying tasks which requires
2516 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2517 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2518 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2519 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2520 * the threadgroup_rwsem to be released to create new tasks. For more details:
2521 *
2522 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2523 *
2524 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2525 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2526 * CPU hotplug is disabled on entry.
2527 *
2528 * When favordynmods is enabled, take per threadgroup rwsem to reduce overhead
2529 * on dynamic cgroup modifications. see the comment above
2530 * CGRP_ROOT_FAVOR_DYNMODS definition.
2531 *
2532 * tsk is not NULL only when writing to cgroup.procs.
2533 */
2534void cgroup_attach_lock(enum cgroup_attach_lock_mode lock_mode,
2535 struct task_struct *tsk)
2536{
2537 cpus_read_lock();
2538
2539 switch (lock_mode) {
2540 case CGRP_ATTACH_LOCK_NONE:
2541 break;
2542 case CGRP_ATTACH_LOCK_GLOBAL:
2543 percpu_down_write(&cgroup_threadgroup_rwsem);
2544 break;
2545 case CGRP_ATTACH_LOCK_PER_THREADGROUP:
2546 down_write(&tsk->signal->cgroup_threadgroup_rwsem);
2547 break;
2548 default:
2549 pr_warn("cgroup: Unexpected attach lock mode.");
2550 break;
2551 }
2552}
2553
2554/**
2555 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2556 * @lock_mode: whether release and release which rwsem
2557 * @tsk: thread group to lock
2558 */
2559void cgroup_attach_unlock(enum cgroup_attach_lock_mode lock_mode,
2560 struct task_struct *tsk)
2561{
2562 switch (lock_mode) {
2563 case CGRP_ATTACH_LOCK_NONE:
2564 break;
2565 case CGRP_ATTACH_LOCK_GLOBAL:
2566 percpu_up_write(&cgroup_threadgroup_rwsem);
2567 break;
2568 case CGRP_ATTACH_LOCK_PER_THREADGROUP:
2569 up_write(&tsk->signal->cgroup_threadgroup_rwsem);
2570 break;
2571 default:
2572 pr_warn("cgroup: Unexpected attach lock mode.");
2573 break;
2574 }
2575
2576 cpus_read_unlock();
2577}
2578
2579/**
2580 * cgroup_migrate_add_task - add a migration target task to a migration context
2581 * @task: target task
2582 * @mgctx: target migration context
2583 *
2584 * Add @task, which is a migration target, to @mgctx->tset. This function
2585 * becomes noop if @task doesn't need to be migrated. @task's css_set
2586 * should have been added as a migration source and @task->cg_list will be
2587 * moved from the css_set's tasks list to mg_tasks one.
2588 */
2589static void cgroup_migrate_add_task(struct task_struct *task,
2590 struct cgroup_mgctx *mgctx)
2591{
2592 struct css_set *cset;
2593
2594 lockdep_assert_held(&css_set_lock);
2595
2596 /* @task either already exited or can't exit until the end */
2597 if (task->flags & PF_EXITING)
2598 return;
2599
2600 /* cgroup_threadgroup_rwsem protects racing against forks */
2601 WARN_ON_ONCE(list_empty(&task->cg_list));
2602
2603 cset = task_css_set(task);
2604 if (!cset->mg_src_cgrp)
2605 return;
2606
2607 mgctx->tset.nr_tasks++;
2608
2609 css_set_skip_task_iters(cset, task);
2610 list_move_tail(&task->cg_list, &cset->mg_tasks);
2611 if (list_empty(&cset->mg_node))
2612 list_add_tail(&cset->mg_node,
2613 &mgctx->tset.src_csets);
2614 if (list_empty(&cset->mg_dst_cset->mg_node))
2615 list_add_tail(&cset->mg_dst_cset->mg_node,
2616 &mgctx->tset.dst_csets);
2617}
2618
2619/**
2620 * cgroup_taskset_first - reset taskset and return the first task
2621 * @tset: taskset of interest
2622 * @dst_cssp: output variable for the destination css
2623 *
2624 * @tset iteration is initialized and the first task is returned.
2625 */
2626struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2627 struct cgroup_subsys_state **dst_cssp)
2628{
2629 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2630 tset->cur_task = NULL;
2631
2632 return cgroup_taskset_next(tset, dst_cssp);
2633}
2634
2635/**
2636 * cgroup_taskset_next - iterate to the next task in taskset
2637 * @tset: taskset of interest
2638 * @dst_cssp: output variable for the destination css
2639 *
2640 * Return the next task in @tset. Iteration must have been initialized
2641 * with cgroup_taskset_first().
2642 */
2643struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2644 struct cgroup_subsys_state **dst_cssp)
2645{
2646 struct css_set *cset = tset->cur_cset;
2647 struct task_struct *task = tset->cur_task;
2648
2649 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2650 if (!task)
2651 task = list_first_entry(&cset->mg_tasks,
2652 struct task_struct, cg_list);
2653 else
2654 task = list_next_entry(task, cg_list);
2655
2656 if (&task->cg_list != &cset->mg_tasks) {
2657 tset->cur_cset = cset;
2658 tset->cur_task = task;
2659
2660 /*
2661 * This function may be called both before and
2662 * after cgroup_migrate_execute(). The two cases
2663 * can be distinguished by looking at whether @cset
2664 * has its ->mg_dst_cset set.
2665 */
2666 if (cset->mg_dst_cset)
2667 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2668 else
2669 *dst_cssp = cset->subsys[tset->ssid];
2670
2671 return task;
2672 }
2673
2674 cset = list_next_entry(cset, mg_node);
2675 task = NULL;
2676 }
2677
2678 return NULL;
2679}
2680
2681static void cgroup_migrate_notify_canceled(struct css_set *src_cset,
2682 struct task_struct *task)
2683{
2684 struct cgroup_task_migrate_ctx ctx = {
2685 .task = task,
2686 .src_dcgrp = src_cset->dfl_cgrp,
2687 .dst_dcgrp = src_cset->mg_dst_cset->dfl_cgrp,
2688 };
2689
2690 blocking_notifier_call_chain(&cgroup_task_notifier,
2691 CGROUP_TASK_MIGRATE_CANCELED, &ctx);
2692}
2693
2694/**
2695 * cgroup_migrate_execute - migrate a taskset
2696 * @mgctx: migration context
2697 *
2698 * Migrate tasks in @mgctx as setup by migration preparation functions. This
2699 * function fails iff one of the ->can_attach callbacks or CGROUP_TASK_MIGRATING
2700 * notifications fails and guarantees that either all or none of the tasks in
2701 * @mgctx are migrated. @mgctx is consumed regardless of success.
2702 */
2703static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2704{
2705 struct cgroup_taskset *tset = &mgctx->tset;
2706 struct cgroup_subsys *ss;
2707 struct task_struct *task, *tmp_task;
2708 struct css_set *cset, *tmp_cset;
2709 bool dfl_migration = false;
2710 int ssid, failed_ssid, ret;
2711
2712 /* check that we can legitimately attach to the cgroup */
2713 if (tset->nr_tasks) {
2714 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2715 if (ss->can_attach) {
2716 tset->ssid = ssid;
2717 ret = ss->can_attach(tset);
2718 if (ret) {
2719 failed_ssid = ssid;
2720 goto out_cancel_attach;
2721 }
2722 }
2723 } while_each_subsys_mask();
2724 }
2725
2726 /*
2727 * Notify each task about the impending migration. An error return fails
2728 * the migration. Only migrations on the default hierarchy are reported:
2729 * a migration modifies either every moved task's dfl cgroup or, on
2730 * cgroup1 or for subtree_control writes, none.
2731 */
2732 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2733 if (cset->dfl_cgrp == cset->mg_dst_cset->dfl_cgrp)
2734 continue;
2735 dfl_migration = true;
2736 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
2737 struct cgroup_task_migrate_ctx ctx = {
2738 .task = task,
2739 .src_dcgrp = cset->dfl_cgrp,
2740 .dst_dcgrp = cset->mg_dst_cset->dfl_cgrp,
2741 };
2742
2743 ret = blocking_notifier_call_chain_robust(&cgroup_task_notifier,
2744 CGROUP_TASK_MIGRATING,
2745 CGROUP_TASK_MIGRATE_CANCELED,
2746 &ctx);
2747 ret = notifier_to_errno(ret);
2748 if (ret)
2749 goto out_cancel_migrating;
2750 }
2751 }
2752
2753 /*
2754 * Now that we're guaranteed success, proceed to move all tasks to
2755 * the new cgroup. There are no failure cases after here, so this
2756 * is the commit point.
2757 */
2758 spin_lock_irq(&css_set_lock);
2759 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2760 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2761 struct css_set *from_cset = task_css_set(task);
2762 struct css_set *to_cset = cset->mg_dst_cset;
2763
2764 get_css_set(to_cset);
2765 to_cset->nr_tasks++;
2766 css_set_move_task(task, from_cset, to_cset, true);
2767 from_cset->nr_tasks--;
2768 /*
2769 * If the source or destination cgroup is frozen,
2770 * the task might require to change its state.
2771 */
2772 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2773 to_cset->dfl_cgrp);
2774 put_css_set_locked(from_cset);
2775
2776 }
2777 }
2778 spin_unlock_irq(&css_set_lock);
2779
2780 /*
2781 * Migration is committed, all target tasks are now on dst_csets.
2782 * Nothing is sensitive to fork() after this point. Notify
2783 * controllers that migration is complete.
2784 */
2785 tset->csets = &tset->dst_csets;
2786
2787 if (tset->nr_tasks) {
2788 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2789 if (ss->attach) {
2790 tset->ssid = ssid;
2791 ss->attach(tset);
2792 }
2793 } while_each_subsys_mask();
2794 }
2795
2796 /*
2797 * Notify each task after successful migration. The operation can no
2798 * longer fail and the return value is ignored. The MIGRATING loop
2799 * above explains why only dfl migrations are reported. Per-task
2800 * sources are not tracked past the commit point, so src_dcgrp is
2801 * NULL.
2802 */
2803 if (dfl_migration) {
2804 list_for_each_entry(cset, &tset->dst_csets, mg_node) {
2805 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
2806 struct cgroup_task_migrate_ctx ctx = {
2807 .task = task,
2808 .dst_dcgrp = cset->dfl_cgrp,
2809 };
2810
2811 blocking_notifier_call_chain(
2812 &cgroup_task_notifier,
2813 CGROUP_TASK_MIGRATED, &ctx);
2814 }
2815 }
2816 }
2817
2818 ret = 0;
2819 goto out_release_tset;
2820
2821out_cancel_migrating:
2822 list_for_each_entry_continue_reverse(task, &cset->mg_tasks, cg_list)
2823 cgroup_migrate_notify_canceled(cset, task);
2824 list_for_each_entry_continue_reverse(cset, &tset->src_csets, mg_node) {
2825 if (cset->dfl_cgrp == cset->mg_dst_cset->dfl_cgrp)
2826 continue;
2827 list_for_each_entry_reverse(task, &cset->mg_tasks, cg_list)
2828 cgroup_migrate_notify_canceled(cset, task);
2829 }
2830 failed_ssid = CGROUP_SUBSYS_COUNT;
2831out_cancel_attach:
2832 if (tset->nr_tasks) {
2833 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2834 if (ssid == failed_ssid)
2835 break;
2836 if (ss->cancel_attach) {
2837 tset->ssid = ssid;
2838 ss->cancel_attach(tset);
2839 }
2840 } while_each_subsys_mask();
2841 }
2842out_release_tset:
2843 spin_lock_irq(&css_set_lock);
2844 list_splice_init(&tset->dst_csets, &tset->src_csets);
2845 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2846 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2847 list_del_init(&cset->mg_node);
2848 }
2849 spin_unlock_irq(&css_set_lock);
2850
2851 /*
2852 * Re-initialize the cgroup_taskset structure in case it is reused
2853 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2854 * iteration.
2855 */
2856 tset->nr_tasks = 0;
2857 tset->csets = &tset->src_csets;
2858 return ret;
2859}
2860
2861/**
2862 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2863 * @dst_cgrp: destination cgroup to test
2864 *
2865 * On the default hierarchy, except for the mixable, (possible) thread root
2866 * and threaded cgroups, subtree_control must be zero for migration
2867 * destination cgroups with tasks so that child cgroups don't compete
2868 * against tasks.
2869 */
2870int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2871{
2872 /* v1 doesn't have any restriction */
2873 if (!cgroup_on_dfl(dst_cgrp))
2874 return 0;
2875
2876 /* verify @dst_cgrp can host resources */
2877 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2878 return -EOPNOTSUPP;
2879
2880 /*
2881 * If @dst_cgrp is already or can become a thread root or is
2882 * threaded, it doesn't matter.
2883 */
2884 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2885 return 0;
2886
2887 /* apply no-internal-process constraint */
2888 if (dst_cgrp->subtree_control)
2889 return -EBUSY;
2890
2891 return 0;
2892}
2893
2894/**
2895 * cgroup_migrate_finish - cleanup after attach
2896 * @mgctx: migration context
2897 *
2898 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2899 * those functions for details.
2900 */
2901void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2902{
2903 struct css_set *cset, *tmp_cset;
2904
2905 lockdep_assert_held(&cgroup_mutex);
2906
2907 spin_lock_irq(&css_set_lock);
2908
2909 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2910 mg_src_preload_node) {
2911 cset->mg_src_cgrp = NULL;
2912 cset->mg_dst_cgrp = NULL;
2913 cset->mg_dst_cset = NULL;
2914 list_del_init(&cset->mg_src_preload_node);
2915 put_css_set_locked(cset);
2916 }
2917
2918 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2919 mg_dst_preload_node) {
2920 cset->mg_src_cgrp = NULL;
2921 cset->mg_dst_cgrp = NULL;
2922 cset->mg_dst_cset = NULL;
2923 list_del_init(&cset->mg_dst_preload_node);
2924 put_css_set_locked(cset);
2925 }
2926
2927 spin_unlock_irq(&css_set_lock);
2928}
2929
2930/**
2931 * cgroup_migrate_add_src - add a migration source css_set
2932 * @src_cset: the source css_set to add
2933 * @dst_cgrp: the destination cgroup
2934 * @mgctx: migration context
2935 *
2936 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2937 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2938 * up by cgroup_migrate_finish().
2939 *
2940 * This function may be called without holding cgroup_threadgroup_rwsem
2941 * even if the target is a process. Threads may be created and destroyed
2942 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2943 * into play and the preloaded css_sets are guaranteed to cover all
2944 * migrations.
2945 */
2946void cgroup_migrate_add_src(struct css_set *src_cset,
2947 struct cgroup *dst_cgrp,
2948 struct cgroup_mgctx *mgctx)
2949{
2950 struct cgroup *src_cgrp;
2951
2952 lockdep_assert_held(&cgroup_mutex);
2953 lockdep_assert_held(&css_set_lock);
2954
2955 /*
2956 * If ->dead, @src_set is associated with one or more dead cgroups
2957 * and doesn't contain any migratable tasks. Ignore it early so
2958 * that the rest of migration path doesn't get confused by it.
2959 */
2960 if (src_cset->dead)
2961 return;
2962
2963 if (!list_empty(&src_cset->mg_src_preload_node))
2964 return;
2965
2966 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2967
2968 WARN_ON(src_cset->mg_src_cgrp);
2969 WARN_ON(src_cset->mg_dst_cgrp);
2970 WARN_ON(!list_empty(&src_cset->mg_tasks));
2971 WARN_ON(!list_empty(&src_cset->mg_node));
2972
2973 src_cset->mg_src_cgrp = src_cgrp;
2974 src_cset->mg_dst_cgrp = dst_cgrp;
2975 get_css_set(src_cset);
2976 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2977}
2978
2979/**
2980 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2981 * @mgctx: migration context
2982 *
2983 * Tasks are about to be moved and all the source css_sets have been
2984 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2985 * pins all destination css_sets, links each to its source, and append them
2986 * to @mgctx->preloaded_dst_csets.
2987 *
2988 * This function must be called after cgroup_migrate_add_src() has been
2989 * called on each migration source css_set. After migration is performed
2990 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2991 * @mgctx.
2992 */
2993int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2994{
2995 struct css_set *src_cset, *tmp_cset;
2996
2997 lockdep_assert_held(&cgroup_mutex);
2998
2999 /* look up the dst cset for each src cset and link it to src */
3000 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
3001 mg_src_preload_node) {
3002 struct css_set *dst_cset;
3003 struct cgroup_subsys *ss;
3004 int ssid;
3005
3006 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
3007 if (!dst_cset)
3008 return -ENOMEM;
3009
3010 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
3011
3012 /*
3013 * If src cset equals dst, it's noop. Drop the src.
3014 * cgroup_migrate() will skip the cset too. Note that we
3015 * can't handle src == dst as some nodes are used by both.
3016 */
3017 if (src_cset == dst_cset) {
3018 src_cset->mg_src_cgrp = NULL;
3019 src_cset->mg_dst_cgrp = NULL;
3020 list_del_init(&src_cset->mg_src_preload_node);
3021 put_css_set(src_cset);
3022 put_css_set(dst_cset);
3023 continue;
3024 }
3025
3026 src_cset->mg_dst_cset = dst_cset;
3027
3028 if (list_empty(&dst_cset->mg_dst_preload_node))
3029 list_add_tail(&dst_cset->mg_dst_preload_node,
3030 &mgctx->preloaded_dst_csets);
3031 else
3032 put_css_set(dst_cset);
3033
3034 for_each_subsys(ss, ssid)
3035 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
3036 mgctx->ss_mask |= 1 << ssid;
3037 }
3038
3039 return 0;
3040}
3041
3042/**
3043 * cgroup_migrate - migrate a process or task to a cgroup
3044 * @leader: the leader of the process or the task to migrate
3045 * @threadgroup: whether @leader points to the whole process or a single task
3046 * @mgctx: migration context
3047 *
3048 * Migrate a process or task denoted by @leader. If migrating a process,
3049 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
3050 * responsible for invoking cgroup_migrate_add_src() and
3051 * cgroup_migrate_prepare_dst() on the targets before invoking this
3052 * function and following up with cgroup_migrate_finish().
3053 *
3054 * As long as a controller's ->can_attach() or a CGROUP_TASK_MIGRATING
3055 * notification doesn't fail, this function is guaranteed to succeed. This
3056 * means that, excluding those failures, when migrating multiple targets,
3057 * the success or failure can be decided for all targets by invoking
3058 * group_migrate_prepare_dst() before actually starting migrating.
3059 */
3060int cgroup_migrate(struct task_struct *leader, bool threadgroup,
3061 struct cgroup_mgctx *mgctx)
3062{
3063 struct task_struct *task;
3064
3065 /*
3066 * The following thread iteration should be inside an RCU critical
3067 * section to prevent tasks from being freed while taking the snapshot.
3068 * spin_lock_irq() implies RCU critical section here.
3069 */
3070 spin_lock_irq(&css_set_lock);
3071 task = leader;
3072 do {
3073 cgroup_migrate_add_task(task, mgctx);
3074 if (!threadgroup)
3075 break;
3076 } while_each_thread(leader, task);
3077 spin_unlock_irq(&css_set_lock);
3078
3079 return cgroup_migrate_execute(mgctx);
3080}
3081
3082/**
3083 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
3084 * @dst_cgrp: the cgroup to attach to
3085 * @leader: the task or the leader of the threadgroup to be attached
3086 * @threadgroup: attach the whole threadgroup?
3087 *
3088 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
3089 */
3090int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
3091 bool threadgroup)
3092{
3093 DEFINE_CGROUP_MGCTX(mgctx);
3094 struct task_struct *task;
3095 int ret = 0;
3096
3097 /* look up all src csets */
3098 spin_lock_irq(&css_set_lock);
3099 task = leader;
3100 do {
3101 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
3102 if (!threadgroup)
3103 break;
3104 } while_each_thread(leader, task);
3105 spin_unlock_irq(&css_set_lock);
3106
3107 /* prepare dst csets and commit */
3108 ret = cgroup_migrate_prepare_dst(&mgctx);
3109 if (!ret)
3110 ret = cgroup_migrate(leader, threadgroup, &mgctx);
3111
3112 cgroup_migrate_finish(&mgctx);
3113
3114 if (!ret)
3115 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
3116
3117 return ret;
3118}
3119
3120struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
3121 enum cgroup_attach_lock_mode *lock_mode)
3122{
3123 struct task_struct *tsk;
3124 pid_t pid;
3125
3126 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
3127 return ERR_PTR(-EINVAL);
3128
3129retry_find_task:
3130 rcu_read_lock();
3131 if (pid) {
3132 tsk = find_task_by_vpid(pid);
3133 if (!tsk) {
3134 tsk = ERR_PTR(-ESRCH);
3135 goto out_unlock_rcu;
3136 }
3137 } else {
3138 tsk = current;
3139 }
3140
3141 if (threadgroup)
3142 tsk = tsk->group_leader;
3143
3144 /*
3145 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
3146 * If userland migrates such a kthread to a non-root cgroup, it can
3147 * become trapped in a cpuset, or RT kthread may be born in a
3148 * cgroup with no rt_runtime allocated. Just say no.
3149 */
3150 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
3151 tsk = ERR_PTR(-EINVAL);
3152 goto out_unlock_rcu;
3153 }
3154 get_task_struct(tsk);
3155 rcu_read_unlock();
3156
3157 /*
3158 * If we migrate a single thread, we don't care about threadgroup
3159 * stability. If the thread is `current`, it won't exit(2) under our
3160 * hands or change PID through exec(2). We exclude
3161 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write callers
3162 * by cgroup_mutex. Therefore, we can skip the global lock.
3163 */
3164 lockdep_assert_held(&cgroup_mutex);
3165
3166 if (pid || threadgroup) {
3167 if (cgroup_enable_per_threadgroup_rwsem)
3168 *lock_mode = CGRP_ATTACH_LOCK_PER_THREADGROUP;
3169 else
3170 *lock_mode = CGRP_ATTACH_LOCK_GLOBAL;
3171 } else {
3172 *lock_mode = CGRP_ATTACH_LOCK_NONE;
3173 }
3174
3175 cgroup_attach_lock(*lock_mode, tsk);
3176
3177 if (threadgroup) {
3178 if (!thread_group_leader(tsk)) {
3179 /*
3180 * A race with de_thread from another thread's exec()
3181 * may strip us of our leadership. If this happens,
3182 * throw this task away and try again.
3183 */
3184 cgroup_attach_unlock(*lock_mode, tsk);
3185 put_task_struct(tsk);
3186 goto retry_find_task;
3187 }
3188 }
3189
3190 return tsk;
3191
3192out_unlock_rcu:
3193 rcu_read_unlock();
3194 return tsk;
3195}
3196
3197void cgroup_procs_write_finish(struct task_struct *task,
3198 enum cgroup_attach_lock_mode lock_mode)
3199{
3200 cgroup_attach_unlock(lock_mode, task);
3201
3202 /* release reference from cgroup_procs_write_start() */
3203 put_task_struct(task);
3204}
3205
3206static void cgroup_print_ss_mask(struct seq_file *seq, u32 ss_mask)
3207{
3208 struct cgroup_subsys *ss;
3209 bool printed = false;
3210 int ssid;
3211
3212 do_each_subsys_mask(ss, ssid, ss_mask) {
3213 if (printed)
3214 seq_putc(seq, ' ');
3215 seq_puts(seq, ss->name);
3216 printed = true;
3217 } while_each_subsys_mask();
3218 if (printed)
3219 seq_putc(seq, '\n');
3220}
3221
3222/* show controllers which are enabled from the parent */
3223static int cgroup_controllers_show(struct seq_file *seq, void *v)
3224{
3225 struct cgroup *cgrp = seq_css(seq)->cgroup;
3226
3227 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3228 return 0;
3229}
3230
3231/* show controllers which are enabled for a given cgroup's children */
3232static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3233{
3234 struct cgroup *cgrp = seq_css(seq)->cgroup;
3235
3236 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3237 return 0;
3238}
3239
3240/**
3241 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3242 * @cgrp: root of the subtree to update csses for
3243 *
3244 * @cgrp's control masks have changed and its subtree's css associations
3245 * need to be updated accordingly. This function looks up all css_sets
3246 * which are attached to the subtree, creates the matching updated css_sets
3247 * and migrates the tasks to the new ones.
3248 */
3249static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3250{
3251 DEFINE_CGROUP_MGCTX(mgctx);
3252 struct cgroup_subsys_state *d_css;
3253 struct cgroup *dsct;
3254 struct css_set *src_cset;
3255 enum cgroup_attach_lock_mode lock_mode;
3256 bool has_tasks;
3257 int ret;
3258
3259 lockdep_assert_held(&cgroup_mutex);
3260
3261 /* look up all csses currently attached to @cgrp's subtree */
3262 spin_lock_irq(&css_set_lock);
3263 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3264 struct cgrp_cset_link *link;
3265
3266 /*
3267 * As cgroup_update_dfl_csses() is only called by
3268 * cgroup_apply_control(). The csses associated with the
3269 * given cgrp will not be affected by changes made to
3270 * its subtree_control file. We can skip them.
3271 */
3272 if (dsct == cgrp)
3273 continue;
3274
3275 list_for_each_entry(link, &dsct->cset_links, cset_link)
3276 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3277 }
3278 spin_unlock_irq(&css_set_lock);
3279
3280 /*
3281 * We need to write-lock threadgroup_rwsem while migrating tasks.
3282 * However, if there are no source csets for @cgrp, changing its
3283 * controllers isn't gonna produce any task migrations and the
3284 * write-locking can be skipped safely.
3285 */
3286 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3287
3288 if (has_tasks)
3289 lock_mode = CGRP_ATTACH_LOCK_GLOBAL;
3290 else
3291 lock_mode = CGRP_ATTACH_LOCK_NONE;
3292
3293 cgroup_attach_lock(lock_mode, NULL);
3294
3295 /* NULL dst indicates self on default hierarchy */
3296 ret = cgroup_migrate_prepare_dst(&mgctx);
3297 if (ret)
3298 goto out_finish;
3299
3300 spin_lock_irq(&css_set_lock);
3301 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3302 mg_src_preload_node) {
3303 struct task_struct *task, *ntask;
3304
3305 /* all tasks in src_csets need to be migrated */
3306 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3307 cgroup_migrate_add_task(task, &mgctx);
3308 }
3309 spin_unlock_irq(&css_set_lock);
3310
3311 ret = cgroup_migrate_execute(&mgctx);
3312out_finish:
3313 cgroup_migrate_finish(&mgctx);
3314 cgroup_attach_unlock(lock_mode, NULL);
3315 return ret;
3316}
3317
3318/**
3319 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3320 * @cgrp: root of the target subtree
3321 *
3322 * Because css offlining is asynchronous, userland may try to re-enable a
3323 * controller while the previous css is still around. This function grabs
3324 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3325 */
3326void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3327 __acquires(&cgroup_mutex)
3328{
3329 struct cgroup *dsct;
3330 struct cgroup_subsys_state *d_css;
3331 struct cgroup_subsys *ss;
3332 int ssid;
3333
3334restart:
3335 cgroup_lock();
3336
3337 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3338 for_each_subsys(ss, ssid) {
3339 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3340 DEFINE_WAIT(wait);
3341
3342 if (!css || !css_is_dying(css))
3343 continue;
3344
3345 cgroup_get_live(dsct);
3346 prepare_to_wait(&dsct->offline_waitq, &wait,
3347 TASK_UNINTERRUPTIBLE);
3348
3349 cgroup_unlock();
3350 schedule();
3351 finish_wait(&dsct->offline_waitq, &wait);
3352
3353 cgroup_put(dsct);
3354 goto restart;
3355 }
3356 }
3357}
3358
3359/**
3360 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3361 * @cgrp: root of the target subtree
3362 *
3363 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3364 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3365 * itself.
3366 */
3367static void cgroup_save_control(struct cgroup *cgrp)
3368{
3369 struct cgroup *dsct;
3370 struct cgroup_subsys_state *d_css;
3371
3372 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3373 dsct->old_subtree_control = dsct->subtree_control;
3374 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3375 dsct->old_dom_cgrp = dsct->dom_cgrp;
3376 }
3377}
3378
3379/**
3380 * cgroup_propagate_control - refresh control masks of a subtree
3381 * @cgrp: root of the target subtree
3382 *
3383 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3384 * ->subtree_control and propagate controller availability through the
3385 * subtree so that descendants don't have unavailable controllers enabled.
3386 */
3387static void cgroup_propagate_control(struct cgroup *cgrp)
3388{
3389 struct cgroup *dsct;
3390 struct cgroup_subsys_state *d_css;
3391
3392 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3393 dsct->subtree_control &= cgroup_control(dsct);
3394 dsct->subtree_ss_mask =
3395 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3396 cgroup_ss_mask(dsct));
3397 }
3398}
3399
3400/**
3401 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3402 * @cgrp: root of the target subtree
3403 *
3404 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3405 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3406 * itself.
3407 */
3408static void cgroup_restore_control(struct cgroup *cgrp)
3409{
3410 struct cgroup *dsct;
3411 struct cgroup_subsys_state *d_css;
3412
3413 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3414 dsct->subtree_control = dsct->old_subtree_control;
3415 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3416 dsct->dom_cgrp = dsct->old_dom_cgrp;
3417 }
3418}
3419
3420static bool css_visible(struct cgroup_subsys_state *css)
3421{
3422 struct cgroup_subsys *ss = css->ss;
3423 struct cgroup *cgrp = css->cgroup;
3424
3425 if (cgroup_control(cgrp) & (1 << ss->id))
3426 return true;
3427 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3428 return false;
3429 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3430}
3431
3432/**
3433 * cgroup_apply_control_enable - enable or show csses according to control
3434 * @cgrp: root of the target subtree
3435 *
3436 * Walk @cgrp's subtree and create new csses or make the existing ones
3437 * visible. A css is created invisible if it's being implicitly enabled
3438 * through dependency. An invisible css is made visible when the userland
3439 * explicitly enables it.
3440 *
3441 * Returns 0 on success, -errno on failure. On failure, csses which have
3442 * been processed already aren't cleaned up. The caller is responsible for
3443 * cleaning up with cgroup_apply_control_disable().
3444 */
3445static int cgroup_apply_control_enable(struct cgroup *cgrp)
3446{
3447 struct cgroup *dsct;
3448 struct cgroup_subsys_state *d_css;
3449 struct cgroup_subsys *ss;
3450 int ssid, ret;
3451
3452 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3453 for_each_subsys(ss, ssid) {
3454 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3455
3456 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3457 continue;
3458
3459 if (!css) {
3460 css = css_create(dsct, ss);
3461 if (IS_ERR(css))
3462 return PTR_ERR(css);
3463 }
3464
3465 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3466
3467 if (css_visible(css)) {
3468 ret = css_populate_dir(css);
3469 if (ret)
3470 return ret;
3471 }
3472 }
3473 }
3474
3475 return 0;
3476}
3477
3478/**
3479 * cgroup_apply_control_disable - kill or hide csses according to control
3480 * @cgrp: root of the target subtree
3481 *
3482 * Walk @cgrp's subtree and kill and hide csses so that they match
3483 * cgroup_ss_mask() and cgroup_visible_mask().
3484 *
3485 * A css is hidden when the userland requests it to be disabled while other
3486 * subsystems are still depending on it. The css must not actively control
3487 * resources and be in the vanilla state if it's made visible again later.
3488 * Controllers which may be depended upon should provide ->css_reset() for
3489 * this purpose.
3490 */
3491static void cgroup_apply_control_disable(struct cgroup *cgrp)
3492{
3493 struct cgroup *dsct;
3494 struct cgroup_subsys_state *d_css;
3495 struct cgroup_subsys *ss;
3496 int ssid;
3497
3498 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3499 for_each_subsys(ss, ssid) {
3500 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3501
3502 if (!css)
3503 continue;
3504
3505 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3506
3507 if (css->parent &&
3508 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3509 kill_css_sync(css);
3510 if (!css_is_populated(css))
3511 kill_css_finish(css);
3512 } else if (!css_visible(css)) {
3513 css_clear_dir(css);
3514 if (ss->css_reset)
3515 ss->css_reset(css);
3516 }
3517 }
3518 }
3519}
3520
3521/**
3522 * cgroup_apply_control - apply control mask updates to the subtree
3523 * @cgrp: root of the target subtree
3524 *
3525 * subsystems can be enabled and disabled in a subtree using the following
3526 * steps.
3527 *
3528 * 1. Call cgroup_save_control() to stash the current state.
3529 * 2. Update ->subtree_control masks in the subtree as desired.
3530 * 3. Call cgroup_apply_control() to apply the changes.
3531 * 4. Optionally perform other related operations.
3532 * 5. Call cgroup_finalize_control() to finish up.
3533 *
3534 * This function implements step 3 and propagates the mask changes
3535 * throughout @cgrp's subtree, updates csses accordingly and perform
3536 * process migrations.
3537 */
3538static int cgroup_apply_control(struct cgroup *cgrp)
3539{
3540 int ret;
3541
3542 cgroup_propagate_control(cgrp);
3543
3544 ret = cgroup_apply_control_enable(cgrp);
3545 if (ret)
3546 return ret;
3547
3548 /*
3549 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3550 * making the following cgroup_update_dfl_csses() properly update
3551 * css associations of all tasks in the subtree.
3552 */
3553 return cgroup_update_dfl_csses(cgrp);
3554}
3555
3556/**
3557 * cgroup_finalize_control - finalize control mask update
3558 * @cgrp: root of the target subtree
3559 * @ret: the result of the update
3560 *
3561 * Finalize control mask update. See cgroup_apply_control() for more info.
3562 */
3563static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3564{
3565 if (ret) {
3566 cgroup_restore_control(cgrp);
3567 cgroup_propagate_control(cgrp);
3568 }
3569
3570 cgroup_apply_control_disable(cgrp);
3571}
3572
3573static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u32 enable)
3574{
3575 u32 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3576
3577 /* if nothing is getting enabled, nothing to worry about */
3578 if (!enable)
3579 return 0;
3580
3581 /* can @cgrp host any resources? */
3582 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3583 return -EOPNOTSUPP;
3584
3585 /* mixables don't care */
3586 if (cgroup_is_mixable(cgrp))
3587 return 0;
3588
3589 if (domain_enable) {
3590 /* can't enable domain controllers inside a thread subtree */
3591 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3592 return -EOPNOTSUPP;
3593 } else {
3594 /*
3595 * Threaded controllers can handle internal competitions
3596 * and are always allowed inside a (prospective) thread
3597 * subtree.
3598 */
3599 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3600 return 0;
3601 }
3602
3603 /*
3604 * Controllers can't be enabled for a cgroup with tasks to avoid
3605 * child cgroups competing against tasks.
3606 */
3607 if (cgroup_has_tasks(cgrp))
3608 return -EBUSY;
3609
3610 return 0;
3611}
3612
3613/* change the enabled child controllers for a cgroup in the default hierarchy */
3614static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3615 char *buf, size_t nbytes,
3616 loff_t off)
3617{
3618 u32 enable = 0, disable = 0;
3619 struct cgroup *cgrp, *child;
3620 struct cgroup_subsys *ss;
3621 char *tok;
3622 int ssid, ret;
3623
3624 /*
3625 * Parse input - space separated list of subsystem names prefixed
3626 * with either + or -.
3627 */
3628 buf = strstrip(buf);
3629 while ((tok = strsep(&buf, " "))) {
3630 if (tok[0] == '\0')
3631 continue;
3632 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3633 if (!cgroup_ssid_enabled(ssid) ||
3634 strcmp(tok + 1, ss->name))
3635 continue;
3636
3637 if (*tok == '+') {
3638 enable |= 1 << ssid;
3639 disable &= ~(1 << ssid);
3640 } else if (*tok == '-') {
3641 disable |= 1 << ssid;
3642 enable &= ~(1 << ssid);
3643 } else {
3644 return -EINVAL;
3645 }
3646 break;
3647 } while_each_subsys_mask();
3648 if (ssid == CGROUP_SUBSYS_COUNT)
3649 return -EINVAL;
3650 }
3651
3652 cgrp = cgroup_kn_lock_live(of->kn, true);
3653 if (!cgrp)
3654 return -ENODEV;
3655
3656 for_each_subsys(ss, ssid) {
3657 if (enable & (1 << ssid)) {
3658 if (cgrp->subtree_control & (1 << ssid)) {
3659 enable &= ~(1 << ssid);
3660 continue;
3661 }
3662
3663 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3664 ret = -ENOENT;
3665 goto out_unlock;
3666 }
3667 } else if (disable & (1 << ssid)) {
3668 if (!(cgrp->subtree_control & (1 << ssid))) {
3669 disable &= ~(1 << ssid);
3670 continue;
3671 }
3672
3673 /* a child has it enabled? */
3674 cgroup_for_each_live_child(child, cgrp) {
3675 if (child->subtree_control & (1 << ssid)) {
3676 ret = -EBUSY;
3677 goto out_unlock;
3678 }
3679 }
3680 }
3681 }
3682
3683 if (!enable && !disable) {
3684 ret = 0;
3685 goto out_unlock;
3686 }
3687
3688 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3689 if (ret)
3690 goto out_unlock;
3691
3692 /* save and update control masks and prepare csses */
3693 cgroup_save_control(cgrp);
3694
3695 cgrp->subtree_control |= enable;
3696 cgrp->subtree_control &= ~disable;
3697
3698 ret = cgroup_apply_control(cgrp);
3699 cgroup_finalize_control(cgrp, ret);
3700 if (ret)
3701 goto out_unlock;
3702
3703 kernfs_activate(cgrp->kn);
3704out_unlock:
3705 cgroup_kn_unlock(of->kn);
3706 return ret ?: nbytes;
3707}
3708
3709/**
3710 * cgroup_enable_threaded - make @cgrp threaded
3711 * @cgrp: the target cgroup
3712 *
3713 * Called when "threaded" is written to the cgroup.type interface file and
3714 * tries to make @cgrp threaded and join the parent's resource domain.
3715 * This function is never called on the root cgroup as cgroup.type doesn't
3716 * exist on it.
3717 */
3718static int cgroup_enable_threaded(struct cgroup *cgrp)
3719{
3720 struct cgroup *parent = cgroup_parent(cgrp);
3721 struct cgroup *dom_cgrp = parent->dom_cgrp;
3722 struct cgroup *dsct;
3723 struct cgroup_subsys_state *d_css;
3724 int ret;
3725
3726 lockdep_assert_held(&cgroup_mutex);
3727
3728 /* noop if already threaded */
3729 if (cgroup_is_threaded(cgrp))
3730 return 0;
3731
3732 /*
3733 * If @cgroup is populated or has domain controllers enabled, it
3734 * can't be switched. While the below cgroup_can_be_thread_root()
3735 * test can catch the same conditions, that's only when @parent is
3736 * not mixable, so let's check it explicitly.
3737 */
3738 if (cgroup_is_populated(cgrp) ||
3739 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3740 return -EOPNOTSUPP;
3741
3742 /* we're joining the parent's domain, ensure its validity */
3743 if (!cgroup_is_valid_domain(dom_cgrp) ||
3744 !cgroup_can_be_thread_root(dom_cgrp))
3745 return -EOPNOTSUPP;
3746
3747 /*
3748 * The following shouldn't cause actual migrations and should
3749 * always succeed.
3750 */
3751 cgroup_save_control(cgrp);
3752
3753 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3754 if (dsct == cgrp || cgroup_is_threaded(dsct))
3755 dsct->dom_cgrp = dom_cgrp;
3756
3757 ret = cgroup_apply_control(cgrp);
3758 if (!ret)
3759 parent->nr_threaded_children++;
3760
3761 cgroup_finalize_control(cgrp, ret);
3762 return ret;
3763}
3764
3765static int cgroup_type_show(struct seq_file *seq, void *v)
3766{
3767 struct cgroup *cgrp = seq_css(seq)->cgroup;
3768
3769 if (cgroup_is_threaded(cgrp))
3770 seq_puts(seq, "threaded\n");
3771 else if (!cgroup_is_valid_domain(cgrp))
3772 seq_puts(seq, "domain invalid\n");
3773 else if (cgroup_is_thread_root(cgrp))
3774 seq_puts(seq, "domain threaded\n");
3775 else
3776 seq_puts(seq, "domain\n");
3777
3778 return 0;
3779}
3780
3781static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3782 size_t nbytes, loff_t off)
3783{
3784 struct cgroup *cgrp;
3785 int ret;
3786
3787 /* only switching to threaded mode is supported */
3788 if (strcmp(strstrip(buf), "threaded"))
3789 return -EINVAL;
3790
3791 /* drain dying csses before we re-apply (threaded) subtree control */
3792 cgrp = cgroup_kn_lock_live(of->kn, true);
3793 if (!cgrp)
3794 return -ENOENT;
3795
3796 /* threaded can only be enabled */
3797 ret = cgroup_enable_threaded(cgrp);
3798
3799 cgroup_kn_unlock(of->kn);
3800 return ret ?: nbytes;
3801}
3802
3803static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3804{
3805 struct cgroup *cgrp = seq_css(seq)->cgroup;
3806 int descendants = READ_ONCE(cgrp->max_descendants);
3807
3808 if (descendants == INT_MAX)
3809 seq_puts(seq, "max\n");
3810 else
3811 seq_printf(seq, "%d\n", descendants);
3812
3813 return 0;
3814}
3815
3816static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3817 char *buf, size_t nbytes, loff_t off)
3818{
3819 struct cgroup *cgrp;
3820 int descendants;
3821 ssize_t ret;
3822
3823 buf = strstrip(buf);
3824 if (!strcmp(buf, "max")) {
3825 descendants = INT_MAX;
3826 } else {
3827 ret = kstrtoint(buf, 0, &descendants);
3828 if (ret)
3829 return ret;
3830 }
3831
3832 if (descendants < 0)
3833 return -ERANGE;
3834
3835 cgrp = cgroup_kn_lock_live(of->kn, false);
3836 if (!cgrp)
3837 return -ENOENT;
3838
3839 WRITE_ONCE(cgrp->max_descendants, descendants);
3840
3841 cgroup_kn_unlock(of->kn);
3842
3843 return nbytes;
3844}
3845
3846static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3847{
3848 struct cgroup *cgrp = seq_css(seq)->cgroup;
3849 int depth = READ_ONCE(cgrp->max_depth);
3850
3851 if (depth == INT_MAX)
3852 seq_puts(seq, "max\n");
3853 else
3854 seq_printf(seq, "%d\n", depth);
3855
3856 return 0;
3857}
3858
3859static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3860 char *buf, size_t nbytes, loff_t off)
3861{
3862 struct cgroup *cgrp;
3863 ssize_t ret;
3864 int depth;
3865
3866 buf = strstrip(buf);
3867 if (!strcmp(buf, "max")) {
3868 depth = INT_MAX;
3869 } else {
3870 ret = kstrtoint(buf, 0, &depth);
3871 if (ret)
3872 return ret;
3873 }
3874
3875 if (depth < 0)
3876 return -ERANGE;
3877
3878 cgrp = cgroup_kn_lock_live(of->kn, false);
3879 if (!cgrp)
3880 return -ENOENT;
3881
3882 WRITE_ONCE(cgrp->max_depth, depth);
3883
3884 cgroup_kn_unlock(of->kn);
3885
3886 return nbytes;
3887}
3888
3889static int cgroup_events_show(struct seq_file *seq, void *v)
3890{
3891 struct cgroup *cgrp = seq_css(seq)->cgroup;
3892
3893 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3894 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3895
3896 return 0;
3897}
3898
3899static int cgroup_stat_show(struct seq_file *seq, void *v)
3900{
3901 struct cgroup *cgroup = seq_css(seq)->cgroup;
3902 struct cgroup_subsys_state *css;
3903 int dying_cnt[CGROUP_SUBSYS_COUNT];
3904 int ssid;
3905
3906 seq_printf(seq, "nr_descendants %d\n",
3907 cgroup->nr_descendants);
3908
3909 /*
3910 * Show the number of live and dying csses associated with each of
3911 * non-inhibited cgroup subsystems that is bound to cgroup v2.
3912 *
3913 * Without proper lock protection, racing is possible. So the
3914 * numbers may not be consistent when that happens.
3915 */
3916 rcu_read_lock();
3917 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) {
3918 dying_cnt[ssid] = -1;
3919 if ((BIT(ssid) & cgrp_dfl_inhibit_ss_mask) ||
3920 (cgroup_subsys[ssid]->root != &cgrp_dfl_root))
3921 continue;
3922 css = rcu_dereference_raw(cgroup->subsys[ssid]);
3923 dying_cnt[ssid] = cgroup->nr_dying_subsys[ssid];
3924 seq_printf(seq, "nr_subsys_%s %d\n", cgroup_subsys[ssid]->name,
3925 css ? (css->nr_descendants + 1) : 0);
3926 }
3927
3928 seq_printf(seq, "nr_dying_descendants %d\n",
3929 cgroup->nr_dying_descendants);
3930 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) {
3931 if (dying_cnt[ssid] >= 0)
3932 seq_printf(seq, "nr_dying_subsys_%s %d\n",
3933 cgroup_subsys[ssid]->name, dying_cnt[ssid]);
3934 }
3935 rcu_read_unlock();
3936 return 0;
3937}
3938
3939static int cgroup_core_local_stat_show(struct seq_file *seq, void *v)
3940{
3941 struct cgroup *cgrp = seq_css(seq)->cgroup;
3942 unsigned int sequence;
3943 u64 freeze_time;
3944
3945 do {
3946 sequence = read_seqcount_begin(&cgrp->freezer.freeze_seq);
3947 freeze_time = cgrp->freezer.frozen_nsec;
3948 /* Add in current freezer interval if the cgroup is freezing. */
3949 if (test_bit(CGRP_FREEZE, &cgrp->flags))
3950 freeze_time += (ktime_get_ns() -
3951 cgrp->freezer.freeze_start_nsec);
3952 } while (read_seqcount_retry(&cgrp->freezer.freeze_seq, sequence));
3953
3954 do_div(freeze_time, NSEC_PER_USEC);
3955 seq_printf(seq, "frozen_usec %llu\n", freeze_time);
3956
3957 return 0;
3958}
3959
3960#ifdef CONFIG_CGROUP_SCHED
3961/**
3962 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
3963 * @cgrp: the cgroup of interest
3964 * @ss: the subsystem of interest
3965 *
3966 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
3967 * or is offline, %NULL is returned.
3968 */
3969static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
3970 struct cgroup_subsys *ss)
3971{
3972 struct cgroup_subsys_state *css;
3973
3974 rcu_read_lock();
3975 css = cgroup_css(cgrp, ss);
3976 if (css && !css_tryget_online(css))
3977 css = NULL;
3978 rcu_read_unlock();
3979
3980 return css;
3981}
3982
3983static int cgroup_extra_stat_show(struct seq_file *seq, int ssid)
3984{
3985 struct cgroup *cgrp = seq_css(seq)->cgroup;
3986 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3987 struct cgroup_subsys_state *css;
3988 int ret;
3989
3990 if (!ss->css_extra_stat_show)
3991 return 0;
3992
3993 css = cgroup_tryget_css(cgrp, ss);
3994 if (!css)
3995 return 0;
3996
3997 ret = ss->css_extra_stat_show(seq, css);
3998 css_put(css);
3999 return ret;
4000}
4001
4002static int cgroup_local_stat_show(struct seq_file *seq,
4003 struct cgroup *cgrp, int ssid)
4004{
4005 struct cgroup_subsys *ss = cgroup_subsys[ssid];
4006 struct cgroup_subsys_state *css;
4007 int ret;
4008
4009 if (!ss->css_local_stat_show)
4010 return 0;
4011
4012 css = cgroup_tryget_css(cgrp, ss);
4013 if (!css)
4014 return 0;
4015
4016 ret = ss->css_local_stat_show(seq, css);
4017 css_put(css);
4018 return ret;
4019}
4020#endif
4021
4022static int cpu_stat_show(struct seq_file *seq, void *v)
4023{
4024 int ret = 0;
4025
4026 cgroup_base_stat_cputime_show(seq);
4027#ifdef CONFIG_CGROUP_SCHED
4028 ret = cgroup_extra_stat_show(seq, cpu_cgrp_id);
4029#endif
4030 return ret;
4031}
4032
4033static int cpu_local_stat_show(struct seq_file *seq, void *v)
4034{
4035 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
4036 int ret = 0;
4037
4038#ifdef CONFIG_CGROUP_SCHED
4039 ret = cgroup_local_stat_show(seq, cgrp, cpu_cgrp_id);
4040#endif
4041 return ret;
4042}
4043
4044#ifdef CONFIG_PSI
4045static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
4046{
4047 struct cgroup *cgrp = seq_css(seq)->cgroup;
4048 struct psi_group *psi = cgroup_psi(cgrp);
4049
4050 return psi_show(seq, psi, PSI_IO);
4051}
4052static int cgroup_memory_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_MEM);
4058}
4059static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
4060{
4061 struct cgroup *cgrp = seq_css(seq)->cgroup;
4062 struct psi_group *psi = cgroup_psi(cgrp);
4063
4064 return psi_show(seq, psi, PSI_CPU);
4065}
4066
4067static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
4068 size_t nbytes, enum psi_res res)
4069{
4070 struct cgroup_file_ctx *ctx;
4071 struct psi_trigger *new;
4072 struct cgroup *cgrp;
4073 struct psi_group *psi;
4074 bool need_rtpoll_worker;
4075 ssize_t ret = 0;
4076
4077 cgrp = cgroup_kn_lock_live(of->kn, false);
4078 if (!cgrp)
4079 return -ENODEV;
4080
4081 ctx = of->priv;
4082 if (!ctx) {
4083 ret = -ENODEV;
4084 goto out_unlock;
4085 }
4086
4087 /* Allow only one trigger per file descriptor */
4088 if (ctx->psi.trigger) {
4089 ret = -EBUSY;
4090 goto out_unlock;
4091 }
4092
4093 psi = cgroup_psi(cgrp);
4094 new = psi_trigger_create(psi, buf, res, of->file, of,
4095 &need_rtpoll_worker);
4096 if (IS_ERR(new)) {
4097 ret = PTR_ERR(new);
4098 goto out_unlock;
4099 }
4100
4101 /*
4102 * The worker fork must run with neither cgroup_mutex nor the file's
4103 * kernfs active reference held. The latter is broken since
4104 * cgroup_kn_lock_live(). @of->priv may be released while unlocked, so
4105 * recheck before publishing @new.
4106 */
4107 if (need_rtpoll_worker) {
4108 cgroup_unlock();
4109 ret = psi_trigger_create_rtpoll_worker(psi);
4110 cgroup_lock();
4111
4112 if (!ret && !of->priv)
4113 ret = -ENODEV;
4114 if (ret) {
4115 psi_trigger_destroy(new);
4116 goto out_unlock;
4117 }
4118 }
4119
4120 smp_store_release(&ctx->psi.trigger, new);
4121
4122out_unlock:
4123 cgroup_kn_unlock(of->kn);
4124 if (ret)
4125 return ret;
4126
4127 return nbytes;
4128}
4129
4130static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
4131 char *buf, size_t nbytes,
4132 loff_t off)
4133{
4134 return pressure_write(of, buf, nbytes, PSI_IO);
4135}
4136
4137static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
4138 char *buf, size_t nbytes,
4139 loff_t off)
4140{
4141 return pressure_write(of, buf, nbytes, PSI_MEM);
4142}
4143
4144static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
4145 char *buf, size_t nbytes,
4146 loff_t off)
4147{
4148 return pressure_write(of, buf, nbytes, PSI_CPU);
4149}
4150
4151#ifdef CONFIG_IRQ_TIME_ACCOUNTING
4152static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
4153{
4154 struct cgroup *cgrp = seq_css(seq)->cgroup;
4155 struct psi_group *psi = cgroup_psi(cgrp);
4156
4157 return psi_show(seq, psi, PSI_IRQ);
4158}
4159
4160static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
4161 char *buf, size_t nbytes,
4162 loff_t off)
4163{
4164 return pressure_write(of, buf, nbytes, PSI_IRQ);
4165}
4166#endif
4167
4168static int cgroup_pressure_show(struct seq_file *seq, void *v)
4169{
4170 struct cgroup *cgrp = seq_css(seq)->cgroup;
4171 struct psi_group *psi = cgroup_psi(cgrp);
4172
4173 seq_printf(seq, "%d\n", psi->enabled);
4174
4175 return 0;
4176}
4177
4178static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
4179 char *buf, size_t nbytes,
4180 loff_t off)
4181{
4182 ssize_t ret;
4183 int enable;
4184 struct cgroup *cgrp;
4185 struct psi_group *psi;
4186
4187 ret = kstrtoint(strstrip(buf), 0, &enable);
4188 if (ret)
4189 return ret;
4190
4191 if (enable < 0 || enable > 1)
4192 return -ERANGE;
4193
4194 cgrp = cgroup_kn_lock_live(of->kn, false);
4195 if (!cgrp)
4196 return -ENOENT;
4197
4198 psi = cgroup_psi(cgrp);
4199 if (psi->enabled != enable) {
4200 int i;
4201
4202 /* show or hide {cpu,memory,io,irq}.pressure files */
4203 for (i = 0; i < NR_PSI_RESOURCES; i++)
4204 cgroup_file_show(&cgrp->psi_files[i], enable);
4205
4206 psi->enabled = enable;
4207 if (enable)
4208 psi_cgroup_restart(psi);
4209 }
4210
4211 cgroup_kn_unlock(of->kn);
4212
4213 return nbytes;
4214}
4215
4216static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
4217 poll_table *pt)
4218{
4219 struct cgroup_file_ctx *ctx = of->priv;
4220
4221 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
4222}
4223
4224static void cgroup_pressure_release(struct kernfs_open_file *of)
4225{
4226 struct cgroup_file_ctx *ctx = of->priv;
4227
4228 psi_trigger_destroy(ctx->psi.trigger);
4229}
4230
4231bool cgroup_psi_enabled(void)
4232{
4233 if (static_branch_likely(&psi_disabled))
4234 return false;
4235
4236 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
4237}
4238
4239#else /* CONFIG_PSI */
4240bool cgroup_psi_enabled(void)
4241{
4242 return false;
4243}
4244
4245#endif /* CONFIG_PSI */
4246
4247static int cgroup_freeze_show(struct seq_file *seq, void *v)
4248{
4249 struct cgroup *cgrp = seq_css(seq)->cgroup;
4250
4251 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
4252
4253 return 0;
4254}
4255
4256static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
4257 char *buf, size_t nbytes, loff_t off)
4258{
4259 struct cgroup *cgrp;
4260 ssize_t ret;
4261 int freeze;
4262
4263 ret = kstrtoint(strstrip(buf), 0, &freeze);
4264 if (ret)
4265 return ret;
4266
4267 if (freeze < 0 || freeze > 1)
4268 return -ERANGE;
4269
4270 cgrp = cgroup_kn_lock_live(of->kn, false);
4271 if (!cgrp)
4272 return -ENOENT;
4273
4274 cgroup_freeze(cgrp, freeze);
4275
4276 cgroup_kn_unlock(of->kn);
4277
4278 return nbytes;
4279}
4280
4281static void __cgroup_kill(struct cgroup *cgrp)
4282{
4283 struct css_task_iter it;
4284 struct task_struct *task;
4285
4286 lockdep_assert_held(&cgroup_mutex);
4287
4288 spin_lock_irq(&css_set_lock);
4289 cgrp->kill_seq++;
4290 spin_unlock_irq(&css_set_lock);
4291
4292 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
4293 while ((task = css_task_iter_next(&it))) {
4294 /* Ignore kernel threads here. */
4295 if (task->flags & PF_KTHREAD)
4296 continue;
4297
4298 /* Skip tasks that are already dying. */
4299 if (__fatal_signal_pending(task))
4300 continue;
4301
4302 send_sig(SIGKILL, task, 0);
4303 }
4304 css_task_iter_end(&it);
4305}
4306
4307static void cgroup_kill(struct cgroup *cgrp)
4308{
4309 struct cgroup_subsys_state *css;
4310 struct cgroup *dsct;
4311
4312 lockdep_assert_held(&cgroup_mutex);
4313
4314 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
4315 __cgroup_kill(dsct);
4316}
4317
4318static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
4319 size_t nbytes, loff_t off)
4320{
4321 ssize_t ret = 0;
4322 int kill;
4323 struct cgroup *cgrp;
4324
4325 ret = kstrtoint(strstrip(buf), 0, &kill);
4326 if (ret)
4327 return ret;
4328
4329 if (kill != 1)
4330 return -ERANGE;
4331
4332 cgrp = cgroup_kn_lock_live(of->kn, false);
4333 if (!cgrp)
4334 return -ENOENT;
4335
4336 /*
4337 * Killing is a process directed operation, i.e. the whole thread-group
4338 * is taken down so act like we do for cgroup.procs and only make this
4339 * writable in non-threaded cgroups.
4340 */
4341 if (cgroup_is_threaded(cgrp))
4342 ret = -EOPNOTSUPP;
4343 else
4344 cgroup_kill(cgrp);
4345
4346 cgroup_kn_unlock(of->kn);
4347
4348 return ret ?: nbytes;
4349}
4350
4351static int cgroup_file_open(struct kernfs_open_file *of)
4352{
4353 struct cftype *cft = of_cft(of);
4354 struct cgroup_file_ctx *ctx;
4355 int ret;
4356
4357 ctx = kzalloc_obj(*ctx);
4358 if (!ctx)
4359 return -ENOMEM;
4360
4361 ctx->ns = current->nsproxy->cgroup_ns;
4362 get_cgroup_ns(ctx->ns);
4363 of->priv = ctx;
4364
4365 if (!cft->open)
4366 return 0;
4367
4368 ret = cft->open(of);
4369 if (ret) {
4370 put_cgroup_ns(ctx->ns);
4371 kfree(ctx);
4372 }
4373 return ret;
4374}
4375
4376static void cgroup_file_release(struct kernfs_open_file *of)
4377{
4378 struct cftype *cft = of_cft(of);
4379 struct cgroup_file_ctx *ctx = of->priv;
4380
4381 if (cft->release)
4382 cft->release(of);
4383 put_cgroup_ns(ctx->ns);
4384 kfree(ctx);
4385 of->priv = NULL;
4386}
4387
4388static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4389 size_t nbytes, loff_t off)
4390{
4391 struct cgroup_file_ctx *ctx = of->priv;
4392 struct cgroup *cgrp = kn_priv(of->kn);
4393 struct cftype *cft = of_cft(of);
4394 struct cgroup_subsys_state *css;
4395 int ret;
4396
4397 if (!nbytes)
4398 return 0;
4399
4400 /*
4401 * If namespaces are delegation boundaries, disallow writes to
4402 * files in an non-init namespace root from inside the namespace
4403 * except for the files explicitly marked delegatable -
4404 * eg. cgroup.procs, cgroup.threads and cgroup.subtree_control.
4405 */
4406 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4407 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4408 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4409 return -EPERM;
4410
4411 if (cft->write)
4412 return cft->write(of, buf, nbytes, off);
4413
4414 /*
4415 * kernfs guarantees that a file isn't deleted with operations in
4416 * flight, which means that the matching css is and stays alive and
4417 * doesn't need to be pinned. The RCU locking is not necessary
4418 * either. It's just for the convenience of using cgroup_css().
4419 */
4420 rcu_read_lock();
4421 css = cgroup_css(cgrp, cft->ss);
4422 rcu_read_unlock();
4423
4424 if (cft->write_u64) {
4425 unsigned long long v;
4426 ret = kstrtoull(buf, 0, &v);
4427 if (!ret)
4428 ret = cft->write_u64(css, cft, v);
4429 } else if (cft->write_s64) {
4430 long long v;
4431 ret = kstrtoll(buf, 0, &v);
4432 if (!ret)
4433 ret = cft->write_s64(css, cft, v);
4434 } else {
4435 ret = -EINVAL;
4436 }
4437
4438 return ret ?: nbytes;
4439}
4440
4441static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4442{
4443 struct cftype *cft = of_cft(of);
4444
4445 if (cft->poll)
4446 return cft->poll(of, pt);
4447
4448 return kernfs_generic_poll(of, pt);
4449}
4450
4451static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4452{
4453 return seq_cft(seq)->seq_start(seq, ppos);
4454}
4455
4456static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4457{
4458 return seq_cft(seq)->seq_next(seq, v, ppos);
4459}
4460
4461static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4462{
4463 if (seq_cft(seq)->seq_stop)
4464 seq_cft(seq)->seq_stop(seq, v);
4465}
4466
4467static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4468{
4469 struct cftype *cft = seq_cft(m);
4470 struct cgroup_subsys_state *css = seq_css(m);
4471
4472 if (cft->seq_show)
4473 return cft->seq_show(m, arg);
4474
4475 if (cft->read_u64)
4476 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4477 else if (cft->read_s64)
4478 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4479 else
4480 return -EINVAL;
4481 return 0;
4482}
4483
4484static struct kernfs_ops cgroup_kf_single_ops = {
4485 .atomic_write_len = PAGE_SIZE,
4486 .open = cgroup_file_open,
4487 .release = cgroup_file_release,
4488 .write = cgroup_file_write,
4489 .poll = cgroup_file_poll,
4490 .seq_show = cgroup_seqfile_show,
4491};
4492
4493static struct kernfs_ops cgroup_kf_ops = {
4494 .atomic_write_len = PAGE_SIZE,
4495 .open = cgroup_file_open,
4496 .release = cgroup_file_release,
4497 .write = cgroup_file_write,
4498 .poll = cgroup_file_poll,
4499 .seq_start = cgroup_seqfile_start,
4500 .seq_next = cgroup_seqfile_next,
4501 .seq_stop = cgroup_seqfile_stop,
4502 .seq_show = cgroup_seqfile_show,
4503};
4504
4505static void cgroup_file_notify_timer(struct timer_list *timer)
4506{
4507 cgroup_file_notify(container_of(timer, struct cgroup_file,
4508 notify_timer));
4509}
4510
4511static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4512 struct cftype *cft)
4513{
4514 char name[CGROUP_FILE_NAME_MAX];
4515 struct kernfs_node *kn;
4516 struct lock_class_key *key = NULL;
4517
4518#ifdef CONFIG_DEBUG_LOCK_ALLOC
4519 key = &cft->lockdep_key;
4520#endif
4521 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4522 cgroup_file_mode(cft),
4523 current_fsuid(), current_fsgid(),
4524 0, cft->kf_ops, cft,
4525 NULL, key);
4526 if (IS_ERR(kn))
4527 return PTR_ERR(kn);
4528
4529 if (cft->file_offset) {
4530 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4531
4532 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4533 spin_lock_init(&cfile->lock);
4534 cfile->kn = kn;
4535 }
4536
4537 return 0;
4538}
4539
4540/**
4541 * cgroup_addrm_files - add or remove files to a cgroup directory
4542 * @css: the target css
4543 * @cgrp: the target cgroup (usually css->cgroup)
4544 * @cfts: array of cftypes to be added
4545 * @is_add: whether to add or remove
4546 *
4547 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4548 * For removals, this function never fails.
4549 */
4550static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4551 struct cgroup *cgrp, struct cftype cfts[],
4552 bool is_add)
4553{
4554 struct cftype *cft, *cft_end = NULL;
4555 int ret = 0;
4556
4557 lockdep_assert_held(&cgroup_mutex);
4558
4559restart:
4560 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4561 /* does cft->flags tell us to skip this file on @cgrp? */
4562 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4563 continue;
4564 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4565 continue;
4566 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4567 continue;
4568 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4569 continue;
4570 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4571 continue;
4572 if (is_add) {
4573 ret = cgroup_add_file(css, cgrp, cft);
4574 if (ret) {
4575 pr_warn("%s: failed to add %s, err=%d\n",
4576 __func__, cft->name, ret);
4577 cft_end = cft;
4578 is_add = false;
4579 goto restart;
4580 }
4581 } else {
4582 cgroup_rm_file(cgrp, cft);
4583 }
4584 }
4585 return ret;
4586}
4587
4588static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4589{
4590 struct cgroup_subsys *ss = cfts[0].ss;
4591 struct cgroup *root = &ss->root->cgrp;
4592 struct cgroup_subsys_state *css;
4593 int ret = 0;
4594
4595 lockdep_assert_held(&cgroup_mutex);
4596
4597 /* add/rm files for all cgroups created before */
4598 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4599 struct cgroup *cgrp = css->cgroup;
4600
4601 if (!(css->flags & CSS_VISIBLE))
4602 continue;
4603
4604 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4605 if (ret)
4606 break;
4607 }
4608
4609 if (is_add && !ret)
4610 kernfs_activate(root->kn);
4611 return ret;
4612}
4613
4614static void cgroup_exit_cftypes(struct cftype *cfts)
4615{
4616 struct cftype *cft;
4617
4618 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4619 /* free copy for custom atomic_write_len, see init_cftypes() */
4620 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4621 kfree(cft->kf_ops);
4622 cft->kf_ops = NULL;
4623 cft->ss = NULL;
4624
4625 /* revert flags set by cgroup core while adding @cfts */
4626 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4627 __CFTYPE_ADDED);
4628 }
4629}
4630
4631static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4632{
4633 struct cftype *cft;
4634 int ret = 0;
4635
4636 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4637 struct kernfs_ops *kf_ops;
4638
4639 WARN_ON(cft->ss || cft->kf_ops);
4640
4641 if (cft->flags & __CFTYPE_ADDED) {
4642 ret = -EBUSY;
4643 break;
4644 }
4645
4646 if (cft->seq_start)
4647 kf_ops = &cgroup_kf_ops;
4648 else
4649 kf_ops = &cgroup_kf_single_ops;
4650
4651 /*
4652 * Ugh... if @cft wants a custom max_write_len, we need to
4653 * make a copy of kf_ops to set its atomic_write_len.
4654 */
4655 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4656 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4657 if (!kf_ops) {
4658 ret = -ENOMEM;
4659 break;
4660 }
4661 kf_ops->atomic_write_len = cft->max_write_len;
4662 }
4663
4664 cft->kf_ops = kf_ops;
4665 cft->ss = ss;
4666 cft->flags |= __CFTYPE_ADDED;
4667 }
4668
4669 if (ret)
4670 cgroup_exit_cftypes(cfts);
4671 return ret;
4672}
4673
4674static void cgroup_rm_cftypes_locked(struct cftype *cfts)
4675{
4676 lockdep_assert_held(&cgroup_mutex);
4677
4678 list_del(&cfts->node);
4679 cgroup_apply_cftypes(cfts, false);
4680 cgroup_exit_cftypes(cfts);
4681}
4682
4683/**
4684 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4685 * @cfts: zero-length name terminated array of cftypes
4686 *
4687 * Unregister @cfts. Files described by @cfts are removed from all
4688 * existing cgroups and all future cgroups won't have them either. This
4689 * function can be called anytime whether @cfts' subsys is attached or not.
4690 *
4691 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4692 * registered.
4693 */
4694int cgroup_rm_cftypes(struct cftype *cfts)
4695{
4696 if (!cfts || cfts[0].name[0] == '\0')
4697 return 0;
4698
4699 if (!(cfts[0].flags & __CFTYPE_ADDED))
4700 return -ENOENT;
4701
4702 cgroup_lock();
4703 cgroup_rm_cftypes_locked(cfts);
4704 cgroup_unlock();
4705 return 0;
4706}
4707
4708/**
4709 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4710 * @ss: target cgroup subsystem
4711 * @cfts: zero-length name terminated array of cftypes
4712 *
4713 * Register @cfts to @ss. Files described by @cfts are created for all
4714 * existing cgroups to which @ss is attached and all future cgroups will
4715 * have them too. This function can be called anytime whether @ss is
4716 * attached or not.
4717 *
4718 * Returns 0 on successful registration, -errno on failure. Note that this
4719 * function currently returns 0 as long as @cfts registration is successful
4720 * even if some file creation attempts on existing cgroups fail.
4721 */
4722int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4723{
4724 int ret;
4725
4726 if (!cgroup_ssid_enabled(ss->id))
4727 return 0;
4728
4729 if (!cfts || cfts[0].name[0] == '\0')
4730 return 0;
4731
4732 ret = cgroup_init_cftypes(ss, cfts);
4733 if (ret)
4734 return ret;
4735
4736 cgroup_lock();
4737
4738 list_add_tail(&cfts->node, &ss->cfts);
4739 ret = cgroup_apply_cftypes(cfts, true);
4740 if (ret)
4741 cgroup_rm_cftypes_locked(cfts);
4742
4743 cgroup_unlock();
4744 return ret;
4745}
4746
4747/**
4748 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4749 * @ss: target cgroup subsystem
4750 * @cfts: zero-length name terminated array of cftypes
4751 *
4752 * Similar to cgroup_add_cftypes() but the added files are only used for
4753 * the default hierarchy.
4754 */
4755int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4756{
4757 struct cftype *cft;
4758
4759 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4760 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4761 return cgroup_add_cftypes(ss, cfts);
4762}
4763
4764/**
4765 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4766 * @ss: target cgroup subsystem
4767 * @cfts: zero-length name terminated array of cftypes
4768 *
4769 * Similar to cgroup_add_cftypes() but the added files are only used for
4770 * the legacy hierarchies.
4771 */
4772int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4773{
4774 struct cftype *cft;
4775
4776 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4777 cft->flags |= __CFTYPE_NOT_ON_DFL;
4778 return cgroup_add_cftypes(ss, cfts);
4779}
4780
4781/**
4782 * cgroup_file_notify - generate a file modified event for a cgroup_file
4783 * @cfile: target cgroup_file
4784 *
4785 * @cfile must have been obtained by setting cftype->file_offset.
4786 */
4787void cgroup_file_notify(struct cgroup_file *cfile)
4788{
4789 unsigned long flags, last, next;
4790 struct kernfs_node *kn = NULL;
4791
4792 if (!READ_ONCE(cfile->kn))
4793 return;
4794
4795 last = READ_ONCE(cfile->notified_at);
4796 next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4797 if (time_in_range(jiffies, last, next)) {
4798 timer_reduce(&cfile->notify_timer, next);
4799 if (timer_pending(&cfile->notify_timer))
4800 return;
4801 }
4802
4803 spin_lock_irqsave(&cfile->lock, flags);
4804 if (cfile->kn) {
4805 kn = cfile->kn;
4806 kernfs_get(kn);
4807 WRITE_ONCE(cfile->notified_at, jiffies);
4808 }
4809 spin_unlock_irqrestore(&cfile->lock, flags);
4810
4811 if (kn) {
4812 kernfs_notify(kn);
4813 kernfs_put(kn);
4814 }
4815}
4816EXPORT_SYMBOL_GPL(cgroup_file_notify);
4817
4818/**
4819 * cgroup_file_show - show or hide a hidden cgroup file
4820 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4821 * @show: whether to show or hide
4822 */
4823void cgroup_file_show(struct cgroup_file *cfile, bool show)
4824{
4825 struct kernfs_node *kn;
4826
4827 spin_lock_irq(&cfile->lock);
4828 kn = cfile->kn;
4829 kernfs_get(kn);
4830 spin_unlock_irq(&cfile->lock);
4831
4832 if (kn)
4833 kernfs_show(kn, show);
4834
4835 kernfs_put(kn);
4836}
4837
4838/**
4839 * css_next_child - find the next child of a given css
4840 * @pos: the current position (%NULL to initiate traversal)
4841 * @parent: css whose children to walk
4842 *
4843 * This function returns the next child of @parent and should be called
4844 * under either cgroup_mutex or RCU read lock. The only requirement is
4845 * that @parent and @pos are accessible. The next sibling is guaranteed to
4846 * be returned regardless of their states.
4847 *
4848 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4849 * css which finished ->css_online() is guaranteed to be visible in the
4850 * future iterations and will stay visible until the last reference is put.
4851 * A css which hasn't finished ->css_online() or already finished
4852 * ->css_offline() may show up during traversal. It's each subsystem's
4853 * responsibility to synchronize against on/offlining.
4854 */
4855struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4856 struct cgroup_subsys_state *parent)
4857{
4858 struct cgroup_subsys_state *next;
4859
4860 cgroup_assert_mutex_or_rcu_locked();
4861
4862 /*
4863 * @pos could already have been unlinked from the sibling list.
4864 * Once a cgroup is removed, its ->sibling.next is no longer
4865 * updated when its next sibling changes. CSS_RELEASED is set when
4866 * @pos is taken off list, at which time its next pointer is valid,
4867 * and, as releases are serialized, the one pointed to by the next
4868 * pointer is guaranteed to not have started release yet. This
4869 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4870 * critical section, the one pointed to by its next pointer is
4871 * guaranteed to not have finished its RCU grace period even if we
4872 * have dropped rcu_read_lock() in-between iterations.
4873 *
4874 * If @pos has CSS_RELEASED set, its next pointer can't be
4875 * dereferenced; however, as each css is given a monotonically
4876 * increasing unique serial number and always appended to the
4877 * sibling list, the next one can be found by walking the parent's
4878 * children until the first css with higher serial number than
4879 * @pos's. While this path can be slower, it happens iff iteration
4880 * races against release and the race window is very small.
4881 */
4882 if (!pos) {
4883 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4884 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4885 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4886 } else {
4887 list_for_each_entry_rcu(next, &parent->children, sibling,
4888 lockdep_is_held(&cgroup_mutex))
4889 if (next->serial_nr > pos->serial_nr)
4890 break;
4891 }
4892
4893 /*
4894 * @next, if not pointing to the head, can be dereferenced and is
4895 * the next sibling.
4896 */
4897 if (&next->sibling != &parent->children)
4898 return next;
4899 return NULL;
4900}
4901
4902/**
4903 * css_next_descendant_pre - find the next descendant for pre-order walk
4904 * @pos: the current position (%NULL to initiate traversal)
4905 * @root: css whose descendants to walk
4906 *
4907 * To be used by css_for_each_descendant_pre(). Find the next descendant
4908 * to visit for pre-order traversal of @root's descendants. @root is
4909 * included in the iteration and the first node to be visited.
4910 *
4911 * While this function requires cgroup_mutex or RCU read locking, it
4912 * doesn't require the whole traversal to be contained in a single critical
4913 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4914 * This function will return the correct next descendant as long as both @pos
4915 * and @root are accessible and @pos is a descendant of @root.
4916 *
4917 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4918 * css which finished ->css_online() is guaranteed to be visible in the
4919 * future iterations and will stay visible until the last reference is put.
4920 * A css which hasn't finished ->css_online() or already finished
4921 * ->css_offline() may show up during traversal. It's each subsystem's
4922 * responsibility to synchronize against on/offlining.
4923 */
4924struct cgroup_subsys_state *
4925css_next_descendant_pre(struct cgroup_subsys_state *pos,
4926 struct cgroup_subsys_state *root)
4927{
4928 struct cgroup_subsys_state *next;
4929
4930 cgroup_assert_mutex_or_rcu_locked();
4931
4932 /* if first iteration, visit @root */
4933 if (!pos)
4934 return root;
4935
4936 /* visit the first child if exists */
4937 next = css_next_child(NULL, pos);
4938 if (next)
4939 return next;
4940
4941 /* no child, visit my or the closest ancestor's next sibling */
4942 while (pos != root) {
4943 next = css_next_child(pos, pos->parent);
4944 if (next)
4945 return next;
4946 pos = pos->parent;
4947 }
4948
4949 return NULL;
4950}
4951EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4952
4953/**
4954 * css_rightmost_descendant - return the rightmost descendant of a css
4955 * @pos: css of interest
4956 *
4957 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4958 * is returned. This can be used during pre-order traversal to skip
4959 * subtree of @pos.
4960 *
4961 * While this function requires cgroup_mutex or RCU read locking, it
4962 * doesn't require the whole traversal to be contained in a single critical
4963 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4964 * This function will return the correct rightmost descendant as long as @pos
4965 * is accessible.
4966 */
4967struct cgroup_subsys_state *
4968css_rightmost_descendant(struct cgroup_subsys_state *pos)
4969{
4970 struct cgroup_subsys_state *last, *tmp;
4971
4972 cgroup_assert_mutex_or_rcu_locked();
4973
4974 do {
4975 last = pos;
4976 /* ->prev isn't RCU safe, walk ->next till the end */
4977 pos = NULL;
4978 css_for_each_child(tmp, last)
4979 pos = tmp;
4980 } while (pos);
4981
4982 return last;
4983}
4984
4985static struct cgroup_subsys_state *
4986css_leftmost_descendant(struct cgroup_subsys_state *pos)
4987{
4988 struct cgroup_subsys_state *last;
4989
4990 do {
4991 last = pos;
4992 pos = css_next_child(NULL, pos);
4993 } while (pos);
4994
4995 return last;
4996}
4997
4998/**
4999 * css_next_descendant_post - find the next descendant for post-order walk
5000 * @pos: the current position (%NULL to initiate traversal)
5001 * @root: css whose descendants to walk
5002 *
5003 * To be used by css_for_each_descendant_post(). Find the next descendant
5004 * to visit for post-order traversal of @root's descendants. @root is
5005 * included in the iteration and the last node to be visited.
5006 *
5007 * While this function requires cgroup_mutex or RCU read locking, it
5008 * doesn't require the whole traversal to be contained in a single critical
5009 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
5010 * This function will return the correct next descendant as long as both @pos
5011 * and @cgroup are accessible and @pos is a descendant of @cgroup.
5012 *
5013 * If a subsystem synchronizes ->css_online() and the start of iteration, a
5014 * css which finished ->css_online() is guaranteed to be visible in the
5015 * future iterations and will stay visible until the last reference is put.
5016 * A css which hasn't finished ->css_online() or already finished
5017 * ->css_offline() may show up during traversal. It's each subsystem's
5018 * responsibility to synchronize against on/offlining.
5019 */
5020struct cgroup_subsys_state *
5021css_next_descendant_post(struct cgroup_subsys_state *pos,
5022 struct cgroup_subsys_state *root)
5023{
5024 struct cgroup_subsys_state *next;
5025
5026 cgroup_assert_mutex_or_rcu_locked();
5027
5028 /* if first iteration, visit leftmost descendant which may be @root */
5029 if (!pos)
5030 return css_leftmost_descendant(root);
5031
5032 /* if we visited @root, we're done */
5033 if (pos == root)
5034 return NULL;
5035
5036 /* if there's an unvisited sibling, visit its leftmost descendant */
5037 next = css_next_child(pos, pos->parent);
5038 if (next)
5039 return css_leftmost_descendant(next);
5040
5041 /* no sibling left, visit parent */
5042 return pos->parent;
5043}
5044
5045/**
5046 * css_has_online_children - does a css have online children
5047 * @css: the target css
5048 *
5049 * Returns %true if @css has any online children; otherwise, %false. This
5050 * function can be called from any context but the caller is responsible
5051 * for synchronizing against on/offlining as necessary.
5052 */
5053bool css_has_online_children(struct cgroup_subsys_state *css)
5054{
5055 struct cgroup_subsys_state *child;
5056 bool ret = false;
5057
5058 rcu_read_lock();
5059 css_for_each_child(child, css) {
5060 if (css_is_online(child)) {
5061 ret = true;
5062 break;
5063 }
5064 }
5065 rcu_read_unlock();
5066 return ret;
5067}
5068
5069static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
5070{
5071 struct list_head *l;
5072 struct cgrp_cset_link *link;
5073 struct css_set *cset;
5074
5075 lockdep_assert_held(&css_set_lock);
5076
5077 /* find the next threaded cset */
5078 if (it->tcset_pos) {
5079 l = it->tcset_pos->next;
5080
5081 if (l != it->tcset_head) {
5082 it->tcset_pos = l;
5083 return container_of(l, struct css_set,
5084 threaded_csets_node);
5085 }
5086
5087 it->tcset_pos = NULL;
5088 }
5089
5090 /* find the next cset */
5091 l = it->cset_pos;
5092 l = l->next;
5093 if (l == it->cset_head) {
5094 it->cset_pos = NULL;
5095 return NULL;
5096 }
5097
5098 if (it->ss) {
5099 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
5100 } else {
5101 link = list_entry(l, struct cgrp_cset_link, cset_link);
5102 cset = link->cset;
5103 }
5104
5105 it->cset_pos = l;
5106
5107 /* initialize threaded css_set walking */
5108 if (it->flags & CSS_TASK_ITER_THREADED) {
5109 if (it->cur_dcset)
5110 put_css_set_locked(it->cur_dcset);
5111 it->cur_dcset = cset;
5112 get_css_set(cset);
5113
5114 it->tcset_head = &cset->threaded_csets;
5115 it->tcset_pos = &cset->threaded_csets;
5116 }
5117
5118 return cset;
5119}
5120
5121/**
5122 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
5123 * @it: the iterator to advance
5124 *
5125 * Advance @it to the next css_set to walk.
5126 */
5127static void css_task_iter_advance_css_set(struct css_task_iter *it)
5128{
5129 struct css_set *cset;
5130
5131 lockdep_assert_held(&css_set_lock);
5132
5133 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
5134 while ((cset = css_task_iter_next_css_set(it))) {
5135 if (!list_empty(&cset->tasks)) {
5136 it->cur_tasks_head = &cset->tasks;
5137 break;
5138 } else if (!list_empty(&cset->mg_tasks)) {
5139 it->cur_tasks_head = &cset->mg_tasks;
5140 break;
5141 } else if (!list_empty(&cset->dying_tasks)) {
5142 it->cur_tasks_head = &cset->dying_tasks;
5143 break;
5144 }
5145 }
5146 if (!cset) {
5147 it->task_pos = NULL;
5148 return;
5149 }
5150 it->task_pos = it->cur_tasks_head->next;
5151
5152 /*
5153 * We don't keep css_sets locked across iteration steps and thus
5154 * need to take steps to ensure that iteration can be resumed after
5155 * the lock is re-acquired. Iteration is performed at two levels -
5156 * css_sets and tasks in them.
5157 *
5158 * Once created, a css_set never leaves its cgroup lists, so a
5159 * pinned css_set is guaranteed to stay put and we can resume
5160 * iteration afterwards.
5161 *
5162 * Tasks may leave @cset across iteration steps. This is resolved
5163 * by registering each iterator with the css_set currently being
5164 * walked and making css_set_move_task() advance iterators whose
5165 * next task is leaving.
5166 */
5167 if (it->cur_cset) {
5168 list_del(&it->iters_node);
5169 put_css_set_locked(it->cur_cset);
5170 }
5171 get_css_set(cset);
5172 it->cur_cset = cset;
5173 list_add(&it->iters_node, &cset->task_iters);
5174}
5175
5176static void css_task_iter_skip(struct css_task_iter *it,
5177 struct task_struct *task)
5178{
5179 lockdep_assert_held(&css_set_lock);
5180
5181 if (it->task_pos == &task->cg_list) {
5182 it->task_pos = it->task_pos->next;
5183 it->flags |= CSS_TASK_ITER_SKIPPED;
5184 }
5185}
5186
5187static void css_task_iter_advance(struct css_task_iter *it)
5188{
5189 struct task_struct *task;
5190
5191 lockdep_assert_held(&css_set_lock);
5192repeat:
5193 if (it->task_pos) {
5194 /*
5195 * Advance iterator to find next entry. We go through cset
5196 * tasks, mg_tasks and dying_tasks, when consumed we move onto
5197 * the next cset.
5198 */
5199 if (it->flags & CSS_TASK_ITER_SKIPPED)
5200 it->flags &= ~CSS_TASK_ITER_SKIPPED;
5201 else
5202 it->task_pos = it->task_pos->next;
5203
5204 if (it->task_pos == &it->cur_cset->tasks) {
5205 it->cur_tasks_head = &it->cur_cset->mg_tasks;
5206 it->task_pos = it->cur_tasks_head->next;
5207 }
5208 if (it->task_pos == &it->cur_cset->mg_tasks) {
5209 it->cur_tasks_head = &it->cur_cset->dying_tasks;
5210 it->task_pos = it->cur_tasks_head->next;
5211 }
5212 if (it->task_pos == &it->cur_cset->dying_tasks)
5213 css_task_iter_advance_css_set(it);
5214 } else {
5215 /* called from start, proceed to the first cset */
5216 css_task_iter_advance_css_set(it);
5217 }
5218
5219 if (!it->task_pos)
5220 return;
5221
5222 task = list_entry(it->task_pos, struct task_struct, cg_list);
5223 /*
5224 * Hide tasks that are exiting but not yet removed by default. Keep
5225 * zombie leaders with live threads visible. Usages that need to walk
5226 * every existing task can opt out via CSS_TASK_ITER_WITH_DEAD.
5227 */
5228 if (!(it->flags & CSS_TASK_ITER_WITH_DEAD) &&
5229 (task->flags & PF_EXITING) && !atomic_read(&task->signal->live))
5230 goto repeat;
5231
5232 if (it->flags & CSS_TASK_ITER_PROCS) {
5233 /* if PROCS, skip over tasks which aren't group leaders */
5234 if (!thread_group_leader(task))
5235 goto repeat;
5236
5237 /* and dying leaders w/o live member threads */
5238 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
5239 !atomic_read(&task->signal->live))
5240 goto repeat;
5241 } else {
5242 /* skip all dying ones */
5243 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
5244 goto repeat;
5245 }
5246}
5247
5248/**
5249 * css_task_iter_start - initiate task iteration
5250 * @css: the css to walk tasks of
5251 * @flags: CSS_TASK_ITER_* flags
5252 * @it: the task iterator to use
5253 *
5254 * Initiate iteration through the tasks of @css. The caller can call
5255 * css_task_iter_next() to walk through the tasks until the function
5256 * returns NULL. On completion of iteration, css_task_iter_end() must be
5257 * called.
5258 */
5259void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
5260 struct css_task_iter *it)
5261{
5262 unsigned long irqflags;
5263
5264 memset(it, 0, sizeof(*it));
5265
5266 spin_lock_irqsave(&css_set_lock, irqflags);
5267
5268 it->ss = css->ss;
5269 it->flags = flags;
5270
5271 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
5272 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
5273 else
5274 it->cset_pos = &css->cgroup->cset_links;
5275
5276 it->cset_head = it->cset_pos;
5277
5278 css_task_iter_advance(it);
5279
5280 spin_unlock_irqrestore(&css_set_lock, irqflags);
5281}
5282
5283/**
5284 * css_task_iter_next - return the next task for the iterator
5285 * @it: the task iterator being iterated
5286 *
5287 * The "next" function for task iteration. @it should have been
5288 * initialized via css_task_iter_start(). Returns NULL when the iteration
5289 * reaches the end.
5290 */
5291struct task_struct *css_task_iter_next(struct css_task_iter *it)
5292{
5293 unsigned long irqflags;
5294
5295 if (it->cur_task) {
5296 put_task_struct(it->cur_task);
5297 it->cur_task = NULL;
5298 }
5299
5300 spin_lock_irqsave(&css_set_lock, irqflags);
5301
5302 /* @it may be half-advanced by skips, finish advancing */
5303 if (it->flags & CSS_TASK_ITER_SKIPPED)
5304 css_task_iter_advance(it);
5305
5306 while (it->task_pos && !it->cur_task) {
5307 it->cur_task = list_entry(it->task_pos, struct task_struct,
5308 cg_list);
5309 /* a task on dying_tasks with zero refcount is only valid for
5310 * RCU readers, not even interesting for
5311 * CSS_TASK_ITER_WITH_DEAD, find another one */
5312 it->cur_task = tryget_task_struct(it->cur_task);
5313 css_task_iter_advance(it);
5314 }
5315
5316 spin_unlock_irqrestore(&css_set_lock, irqflags);
5317
5318 return it->cur_task;
5319}
5320
5321/**
5322 * css_task_iter_end - finish task iteration
5323 * @it: the task iterator to finish
5324 *
5325 * Finish task iteration started by css_task_iter_start().
5326 */
5327void css_task_iter_end(struct css_task_iter *it)
5328{
5329 unsigned long irqflags;
5330
5331 if (it->cur_cset) {
5332 spin_lock_irqsave(&css_set_lock, irqflags);
5333 list_del(&it->iters_node);
5334 put_css_set_locked(it->cur_cset);
5335 spin_unlock_irqrestore(&css_set_lock, irqflags);
5336 }
5337
5338 if (it->cur_dcset)
5339 put_css_set(it->cur_dcset);
5340
5341 if (it->cur_task)
5342 put_task_struct(it->cur_task);
5343}
5344
5345static void cgroup_procs_release(struct kernfs_open_file *of)
5346{
5347 struct cgroup_file_ctx *ctx = of->priv;
5348
5349 if (ctx->procs.started)
5350 css_task_iter_end(&ctx->procs.iter);
5351}
5352
5353static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
5354{
5355 struct kernfs_open_file *of = s->private;
5356 struct cgroup_file_ctx *ctx = of->priv;
5357
5358 if (pos)
5359 (*pos)++;
5360
5361 return css_task_iter_next(&ctx->procs.iter);
5362}
5363
5364static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5365 unsigned int iter_flags)
5366{
5367 struct kernfs_open_file *of = s->private;
5368 struct cgroup *cgrp = seq_css(s)->cgroup;
5369 struct cgroup_file_ctx *ctx = of->priv;
5370 struct css_task_iter *it = &ctx->procs.iter;
5371
5372 /*
5373 * When a seq_file is seeked, it's always traversed sequentially
5374 * from position 0, so we can simply keep iterating on !0 *pos.
5375 */
5376 if (!ctx->procs.started) {
5377 if (WARN_ON_ONCE((*pos)))
5378 return ERR_PTR(-EINVAL);
5379 css_task_iter_start(&cgrp->self, iter_flags, it);
5380 ctx->procs.started = true;
5381 } else if (!(*pos)) {
5382 css_task_iter_end(it);
5383 css_task_iter_start(&cgrp->self, iter_flags, it);
5384 } else
5385 return it->cur_task;
5386
5387 return cgroup_procs_next(s, NULL, NULL);
5388}
5389
5390static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5391{
5392 struct cgroup *cgrp = seq_css(s)->cgroup;
5393
5394 /*
5395 * All processes of a threaded subtree belong to the domain cgroup
5396 * of the subtree. Only threads can be distributed across the
5397 * subtree. Reject reads on cgroup.procs in the subtree proper.
5398 * They're always empty anyway.
5399 */
5400 if (cgroup_is_threaded(cgrp))
5401 return ERR_PTR(-EOPNOTSUPP);
5402
5403 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5404 CSS_TASK_ITER_THREADED);
5405}
5406
5407static int cgroup_procs_show(struct seq_file *s, void *v)
5408{
5409 seq_printf(s, "%d\n", task_pid_vnr(v));
5410 return 0;
5411}
5412
5413static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5414{
5415 int ret;
5416 struct inode *inode;
5417
5418 lockdep_assert_held(&cgroup_mutex);
5419
5420 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5421 if (!inode)
5422 return -ENOMEM;
5423
5424 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5425 iput(inode);
5426 return ret;
5427}
5428
5429static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5430 struct cgroup *dst_cgrp,
5431 struct super_block *sb,
5432 struct cgroup_namespace *ns)
5433{
5434 struct cgroup *com_cgrp = src_cgrp;
5435 int ret;
5436
5437 lockdep_assert_held(&cgroup_mutex);
5438
5439 /* find the common ancestor */
5440 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5441 com_cgrp = cgroup_parent(com_cgrp);
5442
5443 /* %current should be authorized to migrate to the common ancestor */
5444 ret = cgroup_may_write(com_cgrp, sb);
5445 if (ret)
5446 return ret;
5447
5448 /*
5449 * If namespaces are delegation boundaries, %current must be able
5450 * to see both source and destination cgroups from its namespace.
5451 */
5452 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5453 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5454 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5455 return -ENOENT;
5456
5457 return 0;
5458}
5459
5460static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5461 struct cgroup *dst_cgrp,
5462 struct super_block *sb, bool threadgroup,
5463 struct cgroup_namespace *ns)
5464{
5465 int ret = 0;
5466
5467 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5468 if (ret)
5469 return ret;
5470
5471 ret = cgroup_migrate_vet_dst(dst_cgrp);
5472 if (ret)
5473 return ret;
5474
5475 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5476 ret = -EOPNOTSUPP;
5477
5478 return ret;
5479}
5480
5481static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5482 bool threadgroup)
5483{
5484 struct cgroup_file_ctx *ctx = of->priv;
5485 struct cgroup *src_cgrp, *dst_cgrp;
5486 struct task_struct *task;
5487 ssize_t ret;
5488 enum cgroup_attach_lock_mode lock_mode;
5489
5490 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5491 if (!dst_cgrp)
5492 return -ENODEV;
5493
5494 task = cgroup_procs_write_start(buf, threadgroup, &lock_mode);
5495 ret = PTR_ERR_OR_ZERO(task);
5496 if (ret)
5497 goto out_unlock;
5498
5499 /* find the source cgroup */
5500 spin_lock_irq(&css_set_lock);
5501 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5502 spin_unlock_irq(&css_set_lock);
5503
5504 /*
5505 * Process and thread migrations follow same delegation rule. Check
5506 * permissions using the credentials from file open to protect against
5507 * inherited fd attacks.
5508 */
5509 scoped_with_creds(of->file->f_cred)
5510 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5511 of->file->f_path.dentry->d_sb,
5512 threadgroup, ctx->ns);
5513 if (ret)
5514 goto out_finish;
5515
5516 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5517
5518out_finish:
5519 cgroup_procs_write_finish(task, lock_mode);
5520out_unlock:
5521 cgroup_kn_unlock(of->kn);
5522
5523 return ret;
5524}
5525
5526static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5527 char *buf, size_t nbytes, loff_t off)
5528{
5529 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5530}
5531
5532static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5533{
5534 return __cgroup_procs_start(s, pos, 0);
5535}
5536
5537static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5538 char *buf, size_t nbytes, loff_t off)
5539{
5540 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5541}
5542
5543/* cgroup core interface files for the default hierarchy */
5544static struct cftype cgroup_base_files[] = {
5545 {
5546 .name = "cgroup.type",
5547 .flags = CFTYPE_NOT_ON_ROOT,
5548 .seq_show = cgroup_type_show,
5549 .write = cgroup_type_write,
5550 },
5551 {
5552 .name = "cgroup.procs",
5553 .flags = CFTYPE_NS_DELEGATABLE,
5554 .file_offset = offsetof(struct cgroup, procs_file),
5555 .release = cgroup_procs_release,
5556 .seq_start = cgroup_procs_start,
5557 .seq_next = cgroup_procs_next,
5558 .seq_show = cgroup_procs_show,
5559 .write = cgroup_procs_write,
5560 },
5561 {
5562 .name = "cgroup.threads",
5563 .flags = CFTYPE_NS_DELEGATABLE,
5564 .release = cgroup_procs_release,
5565 .seq_start = cgroup_threads_start,
5566 .seq_next = cgroup_procs_next,
5567 .seq_show = cgroup_procs_show,
5568 .write = cgroup_threads_write,
5569 },
5570 {
5571 .name = "cgroup.controllers",
5572 .seq_show = cgroup_controllers_show,
5573 },
5574 {
5575 .name = "cgroup.subtree_control",
5576 .flags = CFTYPE_NS_DELEGATABLE,
5577 .seq_show = cgroup_subtree_control_show,
5578 .write = cgroup_subtree_control_write,
5579 },
5580 {
5581 .name = "cgroup.events",
5582 .flags = CFTYPE_NOT_ON_ROOT,
5583 .file_offset = offsetof(struct cgroup, events_file),
5584 .seq_show = cgroup_events_show,
5585 },
5586 {
5587 .name = "cgroup.max.descendants",
5588 .seq_show = cgroup_max_descendants_show,
5589 .write = cgroup_max_descendants_write,
5590 },
5591 {
5592 .name = "cgroup.max.depth",
5593 .seq_show = cgroup_max_depth_show,
5594 .write = cgroup_max_depth_write,
5595 },
5596 {
5597 .name = "cgroup.stat",
5598 .seq_show = cgroup_stat_show,
5599 },
5600 {
5601 .name = "cgroup.stat.local",
5602 .flags = CFTYPE_NOT_ON_ROOT,
5603 .seq_show = cgroup_core_local_stat_show,
5604 },
5605 {
5606 .name = "cgroup.freeze",
5607 .flags = CFTYPE_NOT_ON_ROOT,
5608 .seq_show = cgroup_freeze_show,
5609 .write = cgroup_freeze_write,
5610 },
5611 {
5612 .name = "cgroup.kill",
5613 .flags = CFTYPE_NOT_ON_ROOT,
5614 .write = cgroup_kill_write,
5615 },
5616 {
5617 .name = "cpu.stat",
5618 .seq_show = cpu_stat_show,
5619 },
5620 {
5621 .name = "cpu.stat.local",
5622 .seq_show = cpu_local_stat_show,
5623 },
5624 { } /* terminate */
5625};
5626
5627static struct cftype cgroup_psi_files[] = {
5628#ifdef CONFIG_PSI
5629 {
5630 .name = "io.pressure",
5631 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5632 .seq_show = cgroup_io_pressure_show,
5633 .write = cgroup_io_pressure_write,
5634 .poll = cgroup_pressure_poll,
5635 .release = cgroup_pressure_release,
5636 },
5637 {
5638 .name = "memory.pressure",
5639 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5640 .seq_show = cgroup_memory_pressure_show,
5641 .write = cgroup_memory_pressure_write,
5642 .poll = cgroup_pressure_poll,
5643 .release = cgroup_pressure_release,
5644 },
5645 {
5646 .name = "cpu.pressure",
5647 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5648 .seq_show = cgroup_cpu_pressure_show,
5649 .write = cgroup_cpu_pressure_write,
5650 .poll = cgroup_pressure_poll,
5651 .release = cgroup_pressure_release,
5652 },
5653#ifdef CONFIG_IRQ_TIME_ACCOUNTING
5654 {
5655 .name = "irq.pressure",
5656 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5657 .seq_show = cgroup_irq_pressure_show,
5658 .write = cgroup_irq_pressure_write,
5659 .poll = cgroup_pressure_poll,
5660 .release = cgroup_pressure_release,
5661 },
5662#endif
5663 {
5664 .name = "cgroup.pressure",
5665 .seq_show = cgroup_pressure_show,
5666 .write = cgroup_pressure_write,
5667 },
5668#endif /* CONFIG_PSI */
5669 { } /* terminate */
5670};
5671
5672/*
5673 * css destruction is four-stage process.
5674 *
5675 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5676 * Implemented in kill_css_finish().
5677 *
5678 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5679 * and thus css_tryget_online() is guaranteed to fail, the css can be
5680 * offlined by invoking offline_css(). After offlining, the base ref is
5681 * put. Implemented in css_killed_work_fn().
5682 *
5683 * 3. When the percpu_ref reaches zero, the only possible remaining
5684 * accessors are inside RCU read sections. css_release() schedules the
5685 * RCU callback.
5686 *
5687 * 4. After the grace period, the css can be freed. Implemented in
5688 * css_free_rwork_fn().
5689 *
5690 * It is actually hairier because both step 2 and 4 require process context
5691 * and thus involve punting to css->destroy_work adding two additional
5692 * steps to the already complex sequence.
5693 */
5694static void css_free_rwork_fn(struct work_struct *work)
5695{
5696 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5697 struct cgroup_subsys_state, destroy_rwork);
5698 struct cgroup_subsys *ss = css->ss;
5699 struct cgroup *cgrp = css->cgroup;
5700
5701 percpu_ref_exit(&css->refcnt);
5702 css_rstat_exit(css);
5703
5704 if (!css_is_self(css)) {
5705 /* css free path */
5706 struct cgroup_subsys_state *parent = css->parent;
5707 int id = css->id;
5708
5709 ss->css_free(css);
5710 cgroup_idr_remove(&ss->css_idr, id);
5711 cgroup_put(cgrp);
5712
5713 if (parent)
5714 css_put(parent);
5715 } else {
5716 /* cgroup free path */
5717 atomic_dec(&cgrp->root->nr_cgrps);
5718 if (!cgroup_on_dfl(cgrp))
5719 cgroup1_pidlist_destroy_all(cgrp);
5720 cancel_work_sync(&cgrp->release_agent_work);
5721 bpf_cgrp_storage_free(cgrp);
5722
5723 if (cgroup_parent(cgrp)) {
5724 /*
5725 * We get a ref to the parent, and put the ref when
5726 * this cgroup is being freed, so it's guaranteed
5727 * that the parent won't be destroyed before its
5728 * children.
5729 */
5730 cgroup_put(cgroup_parent(cgrp));
5731 kernfs_put(cgrp->kn);
5732 psi_cgroup_free(cgrp);
5733 kfree(cgrp);
5734 } else {
5735 /*
5736 * This is root cgroup's refcnt reaching zero,
5737 * which indicates that the root should be
5738 * released.
5739 */
5740 cgroup_destroy_root(cgrp->root);
5741 }
5742 }
5743}
5744
5745static void css_release_work_fn(struct work_struct *work)
5746{
5747 struct cgroup_subsys_state *css =
5748 container_of(work, struct cgroup_subsys_state, destroy_work);
5749 struct cgroup_subsys *ss = css->ss;
5750 struct cgroup *cgrp = css->cgroup;
5751
5752 cgroup_lock();
5753
5754 css->flags |= CSS_RELEASED;
5755 list_del_rcu(&css->sibling);
5756
5757 if (!css_is_self(css)) {
5758 struct cgroup *parent_cgrp;
5759
5760 css_rstat_flush(css);
5761
5762 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5763 if (ss->css_released)
5764 ss->css_released(css);
5765
5766 cgrp->nr_dying_subsys[ss->id]--;
5767 /*
5768 * When a css is released and ready to be freed, its
5769 * nr_descendants must be zero. However, the corresponding
5770 * cgrp->nr_dying_subsys[ss->id] may not be 0 if a subsystem
5771 * is activated and deactivated multiple times with one or
5772 * more of its previous activation leaving behind dying csses.
5773 */
5774 WARN_ON_ONCE(css->nr_descendants);
5775 parent_cgrp = cgroup_parent(cgrp);
5776 while (parent_cgrp) {
5777 parent_cgrp->nr_dying_subsys[ss->id]--;
5778 parent_cgrp = cgroup_parent(parent_cgrp);
5779 }
5780 } else {
5781 struct cgroup *tcgrp;
5782
5783 /* cgroup release path */
5784 TRACE_CGROUP_PATH(release, cgrp);
5785
5786 css_rstat_flush(&cgrp->self);
5787
5788 spin_lock_irq(&css_set_lock);
5789 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5790 tcgrp = cgroup_parent(tcgrp))
5791 tcgrp->nr_dying_descendants--;
5792 spin_unlock_irq(&css_set_lock);
5793
5794 /*
5795 * There are two control paths which try to determine
5796 * cgroup from dentry without going through kernfs -
5797 * cgroupstats_build() and css_tryget_online_from_dir().
5798 * Those are supported by RCU protecting clearing of
5799 * cgrp->kn->priv backpointer.
5800 */
5801 if (cgrp->kn)
5802 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5803 NULL);
5804 }
5805
5806 cgroup_unlock();
5807
5808 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5809 queue_rcu_work(cgroup_free_wq, &css->destroy_rwork);
5810}
5811
5812static void css_release(struct percpu_ref *ref)
5813{
5814 struct cgroup_subsys_state *css =
5815 container_of(ref, struct cgroup_subsys_state, refcnt);
5816
5817 INIT_WORK(&css->destroy_work, css_release_work_fn);
5818 queue_work(cgroup_release_wq, &css->destroy_work);
5819}
5820
5821/*
5822 * Deferred kill_css_finish() fired from css_update_populated() once a dying
5823 * css's hierarchical populated state drops to zero. Pinned by css_get() at the
5824 * queue site; matched by css_put() here.
5825 */
5826static void kill_css_finish_work_fn(struct work_struct *work)
5827{
5828 struct cgroup_subsys_state *css =
5829 container_of(work, struct cgroup_subsys_state, kill_finish_work);
5830
5831 cgroup_lock();
5832 kill_css_finish(css);
5833 cgroup_unlock();
5834 css_put(css);
5835}
5836
5837static void init_and_link_css(struct cgroup_subsys_state *css,
5838 struct cgroup_subsys *ss, struct cgroup *cgrp)
5839{
5840 lockdep_assert_held(&cgroup_mutex);
5841
5842 cgroup_get_live(cgrp);
5843
5844 memset(css, 0, sizeof(*css));
5845 css->cgroup = cgrp;
5846 css->ss = ss;
5847 css->id = -1;
5848 INIT_LIST_HEAD(&css->sibling);
5849 INIT_LIST_HEAD(&css->children);
5850 INIT_WORK(&css->kill_finish_work, kill_css_finish_work_fn);
5851 css->serial_nr = css_serial_nr_next++;
5852 atomic_set(&css->online_cnt, 0);
5853
5854 if (cgroup_parent(cgrp)) {
5855 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5856 css_get(css->parent);
5857 }
5858
5859 BUG_ON(cgroup_css(cgrp, ss));
5860}
5861
5862/* invoke ->css_online() on a new CSS and mark it online if successful */
5863static int online_css(struct cgroup_subsys_state *css)
5864{
5865 struct cgroup_subsys *ss = css->ss;
5866 int ret = 0;
5867
5868 lockdep_assert_held(&cgroup_mutex);
5869
5870 if (ss->css_online)
5871 ret = ss->css_online(css);
5872 if (!ret) {
5873 css->flags |= CSS_ONLINE;
5874 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5875
5876 atomic_inc(&css->online_cnt);
5877 if (css->parent) {
5878 atomic_inc(&css->parent->online_cnt);
5879 while ((css = css->parent))
5880 css->nr_descendants++;
5881 }
5882 }
5883 return ret;
5884}
5885
5886/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5887static void offline_css(struct cgroup_subsys_state *css)
5888{
5889 struct cgroup_subsys *ss = css->ss;
5890
5891 lockdep_assert_held(&cgroup_mutex);
5892
5893 if (!css_is_online(css))
5894 return;
5895
5896 if (ss->css_offline)
5897 ss->css_offline(css);
5898
5899 css->flags &= ~CSS_ONLINE;
5900 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5901
5902 wake_up_all(&css->cgroup->offline_waitq);
5903}
5904
5905/**
5906 * css_create - create a cgroup_subsys_state
5907 * @cgrp: the cgroup new css will be associated with
5908 * @ss: the subsys of new css
5909 *
5910 * Create a new css associated with @cgrp - @ss pair. On success, the new
5911 * css is online and installed in @cgrp. This function doesn't create the
5912 * interface files. Returns 0 on success, -errno on failure.
5913 */
5914static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5915 struct cgroup_subsys *ss)
5916{
5917 struct cgroup *parent = cgroup_parent(cgrp);
5918 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5919 struct cgroup_subsys_state *css;
5920 int err;
5921
5922 lockdep_assert_held(&cgroup_mutex);
5923
5924 css = ss->css_alloc(parent_css);
5925 if (!css)
5926 css = ERR_PTR(-ENOMEM);
5927 if (IS_ERR(css))
5928 return css;
5929
5930 init_and_link_css(css, ss, cgrp);
5931
5932 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5933 if (err)
5934 goto err_free_css;
5935
5936 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5937 if (err < 0)
5938 goto err_free_css;
5939 css->id = err;
5940
5941 err = css_rstat_init(css);
5942 if (err)
5943 goto err_free_css;
5944
5945 /* @css is ready to be brought online now, make it visible */
5946 list_add_tail_rcu(&css->sibling, &parent_css->children);
5947 cgroup_idr_replace(&ss->css_idr, css, css->id);
5948
5949 err = online_css(css);
5950 if (err)
5951 goto err_list_del;
5952
5953 return css;
5954
5955err_list_del:
5956 list_del_rcu(&css->sibling);
5957err_free_css:
5958 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5959 queue_rcu_work(cgroup_free_wq, &css->destroy_rwork);
5960 return ERR_PTR(err);
5961}
5962
5963/*
5964 * The returned cgroup is fully initialized including its control mask, but
5965 * it doesn't have the control mask applied.
5966 */
5967static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5968 umode_t mode)
5969{
5970 struct cgroup_root *root = parent->root;
5971 struct cgroup *cgrp, *tcgrp;
5972 struct kernfs_node *kn;
5973 int i, level = parent->level + 1;
5974 int ret;
5975
5976 /* allocate the cgroup and its ID, 0 is reserved for the root */
5977 cgrp = kzalloc_flex(*cgrp, _low_ancestors, level);
5978 if (!cgrp)
5979 return ERR_PTR(-ENOMEM);
5980
5981 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5982 if (ret)
5983 goto out_free_cgrp;
5984
5985 /* create the directory */
5986 kn = kernfs_create_dir_ns(parent->kn, name, mode,
5987 current_fsuid(), current_fsgid(),
5988 cgrp, NULL);
5989 if (IS_ERR(kn)) {
5990 ret = PTR_ERR(kn);
5991 goto out_cancel_ref;
5992 }
5993 cgrp->kn = kn;
5994
5995 init_cgroup_housekeeping(cgrp);
5996
5997 cgrp->self.parent = &parent->self;
5998 cgrp->root = root;
5999 cgrp->level = level;
6000
6001 /*
6002 * Now that init_cgroup_housekeeping() has been called and cgrp->self
6003 * is setup, it is safe to perform rstat initialization on it.
6004 */
6005 ret = css_rstat_init(&cgrp->self);
6006 if (ret)
6007 goto out_kernfs_remove;
6008
6009 ret = psi_cgroup_alloc(cgrp);
6010 if (ret)
6011 goto out_stat_exit;
6012
6013 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
6014 cgrp->ancestors[tcgrp->level] = tcgrp;
6015
6016 /*
6017 * New cgroup inherits effective freeze counter, and
6018 * if the parent has to be frozen, the child has too.
6019 */
6020 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
6021 seqcount_spinlock_init(&cgrp->freezer.freeze_seq, &css_set_lock);
6022 if (cgrp->freezer.e_freeze) {
6023 /*
6024 * Set the CGRP_FREEZE flag, so when a process will be
6025 * attached to the child cgroup, it will become frozen.
6026 * At this point the new cgroup is unpopulated, so we can
6027 * consider it frozen immediately.
6028 */
6029 set_bit(CGRP_FREEZE, &cgrp->flags);
6030 cgrp->freezer.freeze_start_nsec = ktime_get_ns();
6031 set_bit(CGRP_FROZEN, &cgrp->flags);
6032 }
6033
6034 if (notify_on_release(parent))
6035 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
6036
6037 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
6038 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
6039
6040 cgrp->self.serial_nr = css_serial_nr_next++;
6041
6042 ret = blocking_notifier_call_chain_robust(&cgroup_lifetime_notifier,
6043 CGROUP_LIFETIME_ONLINE,
6044 CGROUP_LIFETIME_OFFLINE, cgrp);
6045 ret = notifier_to_errno(ret);
6046 if (ret)
6047 goto out_psi_free;
6048
6049 /* allocation complete, commit to creation */
6050 spin_lock_irq(&css_set_lock);
6051 for (i = 0; i < level; i++) {
6052 tcgrp = cgrp->ancestors[i];
6053 tcgrp->nr_descendants++;
6054
6055 /*
6056 * If the new cgroup is frozen, all ancestor cgroups get a new
6057 * frozen descendant, but their state can't change because of
6058 * this.
6059 */
6060 if (cgrp->freezer.e_freeze)
6061 tcgrp->freezer.nr_frozen_descendants++;
6062 }
6063 spin_unlock_irq(&css_set_lock);
6064
6065 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
6066 atomic_inc(&root->nr_cgrps);
6067 cgroup_get_live(parent);
6068
6069 /*
6070 * On the default hierarchy, a child doesn't automatically inherit
6071 * subtree_control from the parent. Each is configured manually.
6072 */
6073 if (!cgroup_on_dfl(cgrp))
6074 cgrp->subtree_control = cgroup_control(cgrp);
6075
6076 cgroup_propagate_control(cgrp);
6077
6078 return cgrp;
6079
6080out_psi_free:
6081 psi_cgroup_free(cgrp);
6082out_stat_exit:
6083 css_rstat_exit(&cgrp->self);
6084out_kernfs_remove:
6085 kernfs_remove(cgrp->kn);
6086out_cancel_ref:
6087 percpu_ref_exit(&cgrp->self.refcnt);
6088out_free_cgrp:
6089 kfree(cgrp);
6090 return ERR_PTR(ret);
6091}
6092
6093static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
6094{
6095 struct cgroup *cgroup;
6096 int ret = false;
6097 int level = 0;
6098
6099 lockdep_assert_held(&cgroup_mutex);
6100
6101 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
6102 if (cgroup->nr_descendants >= cgroup->max_descendants)
6103 goto fail;
6104
6105 if (level >= cgroup->max_depth)
6106 goto fail;
6107
6108 level++;
6109 }
6110
6111 ret = true;
6112fail:
6113 return ret;
6114}
6115
6116int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
6117{
6118 struct cgroup *parent, *cgrp;
6119 int ret;
6120
6121 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
6122 if (strchr(name, '\n'))
6123 return -EINVAL;
6124
6125 parent = cgroup_kn_lock_live(parent_kn, false);
6126 if (!parent)
6127 return -ENODEV;
6128
6129 if (!cgroup_check_hierarchy_limits(parent)) {
6130 ret = -EAGAIN;
6131 goto out_unlock;
6132 }
6133
6134 cgrp = cgroup_create(parent, name, mode);
6135 if (IS_ERR(cgrp)) {
6136 ret = PTR_ERR(cgrp);
6137 goto out_unlock;
6138 }
6139
6140 /*
6141 * This extra ref will be put in css_free_rwork_fn() and guarantees
6142 * that @cgrp->kn is always accessible.
6143 */
6144 kernfs_get(cgrp->kn);
6145
6146 ret = css_populate_dir(&cgrp->self);
6147 if (ret)
6148 goto out_destroy;
6149
6150 ret = cgroup_apply_control_enable(cgrp);
6151 if (ret)
6152 goto out_destroy;
6153
6154 TRACE_CGROUP_PATH(mkdir, cgrp);
6155
6156 /* let's create and online css's */
6157 kernfs_activate(cgrp->kn);
6158
6159 ret = 0;
6160 goto out_unlock;
6161
6162out_destroy:
6163 cgroup_destroy_locked(cgrp);
6164out_unlock:
6165 cgroup_kn_unlock(parent_kn);
6166 return ret;
6167}
6168
6169/*
6170 * This is called when the refcnt of a css is confirmed to be killed.
6171 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
6172 * initiate destruction and put the css ref from kill_css_finish().
6173 */
6174static void css_killed_work_fn(struct work_struct *work)
6175{
6176 struct cgroup_subsys_state *css;
6177
6178 css = container_of(to_rcu_work(work), struct cgroup_subsys_state, destroy_rwork);
6179
6180 cgroup_lock();
6181
6182 do {
6183 offline_css(css);
6184 css_put(css);
6185 /* @css can't go away while we're holding cgroup_mutex */
6186 css = css->parent;
6187 } while (css && atomic_dec_and_test(&css->online_cnt));
6188
6189 cgroup_unlock();
6190}
6191
6192/* css kill confirmation processing requires process context, bounce */
6193static void css_killed_ref_fn(struct percpu_ref *ref)
6194{
6195 struct cgroup_subsys_state *css =
6196 container_of(ref, struct cgroup_subsys_state, refcnt);
6197
6198 if (atomic_dec_and_test(&css->online_cnt)) {
6199 INIT_RCU_WORK(&css->destroy_rwork, css_killed_work_fn);
6200 queue_rcu_work(cgroup_offline_wq, &css->destroy_rwork);
6201 }
6202}
6203
6204/**
6205 * kill_css_sync - synchronous half of css teardown
6206 * @css: css being killed
6207 *
6208 * See cgroup_destroy_locked().
6209 */
6210static void kill_css_sync(struct cgroup_subsys_state *css)
6211{
6212 struct cgroup_subsys *ss = css->ss;
6213
6214 lockdep_assert_held(&cgroup_mutex);
6215
6216 if (css->flags & CSS_DYING)
6217 return;
6218
6219 /*
6220 * Call css_killed(), if defined, before setting the CSS_DYING flag
6221 */
6222 if (css->ss->css_killed)
6223 css->ss->css_killed(css);
6224
6225 css->flags |= CSS_DYING;
6226
6227 /*
6228 * Pair with smp_mb() in css_update_populated(). Either our
6229 * caller observes the walker's decrement and fires
6230 * synchronously, or the walker observes CSS_DYING and queues.
6231 */
6232 smp_mb();
6233
6234 /*
6235 * This must happen before css is disassociated with its cgroup.
6236 * See seq_css() for details.
6237 */
6238 css_clear_dir(css);
6239
6240 css->cgroup->nr_dying_subsys[ss->id]++;
6241 /*
6242 * Parent css and cgroup cannot be freed until after the freeing
6243 * of child css, see css_free_rwork_fn().
6244 */
6245 while ((css = css->parent)) {
6246 css->nr_descendants--;
6247 css->cgroup->nr_dying_subsys[ss->id]++;
6248 }
6249}
6250
6251/**
6252 * kill_css_finish - deferred half of css teardown
6253 * @css: css being killed
6254 *
6255 * See cgroup_destroy_locked().
6256 */
6257static void kill_css_finish(struct cgroup_subsys_state *css)
6258{
6259 lockdep_assert_held(&cgroup_mutex);
6260
6261 /*
6262 * Skip on re-entry: cgroup_apply_control_disable() may have killed @css
6263 * earlier. cgroup_destroy_locked() can still walk it because
6264 * offline_css() (which NULLs cgrp->subsys[ssid]) runs async.
6265 */
6266 if (percpu_ref_is_dying(&css->refcnt))
6267 return;
6268
6269 /*
6270 * Killing would put the base ref, but we need to keep it alive until
6271 * after ->css_offline().
6272 */
6273 css_get(css);
6274
6275 /*
6276 * cgroup core guarantees that, by the time ->css_offline() is invoked,
6277 * no new css reference will be given out via css_tryget_online(). We
6278 * can't simply call percpu_ref_kill() and proceed to offlining css's
6279 * because percpu_ref_kill() doesn't guarantee that the ref is seen as
6280 * killed on all CPUs on return.
6281 *
6282 * Use percpu_ref_kill_and_confirm() to get notifications as each css is
6283 * confirmed to be seen as killed on all CPUs.
6284 */
6285 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
6286}
6287
6288/**
6289 * cgroup_destroy_locked - destroy @cgrp (called on rmdir)
6290 * @cgrp: cgroup to be destroyed
6291 *
6292 * Tear down @cgrp on behalf of rmdir. Constraints:
6293 *
6294 * - Userspace: rmdir must succeed when cgroup.procs and friends are empty.
6295 *
6296 * - Kernel: subsystem ->css_offline() must not run while any task in @cgrp's
6297 * subtree is still doing kernel work. A task hidden from cgroup.procs (past
6298 * exit_signals() with signal->live cleared) can still schedule, allocate, and
6299 * consume resources until its final context switch. Dying descendants in the
6300 * subtree can host such tasks too.
6301 *
6302 * - Kernel: css_tryget_online() must fail by the time ->css_offline() runs.
6303 *
6304 * The destruction runs in three parts:
6305 *
6306 * - This function: synchronous user-visible state teardown plus kill_css_sync()
6307 * on each subsystem css.
6308 *
6309 * - For each subsys css: fire kill_css_finish() synchronously if the subtree is
6310 * already drained, otherwise rely on css_update_populated() to queue
6311 * kill_finish_work when the last populated cset under the css empties.
6312 *
6313 * - The percpu_ref kill chain: css_killed_ref_fn -> css_killed_work_fn ->
6314 * ->css_offline() -> release/free.
6315 *
6316 * Return 0 on success, -EBUSY if a userspace-visible task or an online child
6317 * remains.
6318 */
6319static int cgroup_destroy_locked(struct cgroup *cgrp)
6320{
6321 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
6322 struct cgroup_subsys_state *css;
6323 struct cgrp_cset_link *link;
6324 struct css_task_iter it;
6325 struct task_struct *task;
6326 int ssid, ret;
6327
6328 lockdep_assert_held(&cgroup_mutex);
6329
6330 css_task_iter_start(&cgrp->self, 0, &it);
6331 task = css_task_iter_next(&it);
6332 css_task_iter_end(&it);
6333 if (task)
6334 return -EBUSY;
6335
6336 /*
6337 * Make sure there's no live children. We can't test emptiness of
6338 * ->self.children as dead children linger on it while being
6339 * drained; otherwise, "rmdir parent/child parent" may fail.
6340 */
6341 if (css_has_online_children(&cgrp->self))
6342 return -EBUSY;
6343
6344 /*
6345 * Mark @cgrp and the associated csets dead. The former prevents
6346 * further task migration and child creation by disabling
6347 * cgroup_kn_lock_live(). The latter makes the csets ignored by
6348 * the migration path.
6349 */
6350 cgrp->self.flags &= ~CSS_ONLINE;
6351
6352 spin_lock_irq(&css_set_lock);
6353 list_for_each_entry(link, &cgrp->cset_links, cset_link)
6354 link->cset->dead = true;
6355 spin_unlock_irq(&css_set_lock);
6356
6357 for_each_css(css, ssid, cgrp)
6358 kill_css_sync(css);
6359
6360 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
6361 css_clear_dir(&cgrp->self);
6362 kernfs_remove(cgrp->kn);
6363
6364 if (cgroup_is_threaded(cgrp))
6365 parent->nr_threaded_children--;
6366
6367 spin_lock_irq(&css_set_lock);
6368 for (tcgrp = parent; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
6369 tcgrp->nr_descendants--;
6370 tcgrp->nr_dying_descendants++;
6371 /*
6372 * If the dying cgroup is frozen, decrease frozen descendants
6373 * counters of ancestor cgroups.
6374 */
6375 if (test_bit(CGRP_FROZEN, &cgrp->flags))
6376 tcgrp->freezer.nr_frozen_descendants--;
6377 }
6378 spin_unlock_irq(&css_set_lock);
6379
6380 cgroup1_check_for_release(parent);
6381
6382 ret = blocking_notifier_call_chain(&cgroup_lifetime_notifier,
6383 CGROUP_LIFETIME_OFFLINE, cgrp);
6384 WARN_ON_ONCE(notifier_to_errno(ret));
6385
6386 /* put the base reference */
6387 percpu_ref_kill(&cgrp->self.refcnt);
6388
6389 for_each_css(css, ssid, cgrp) {
6390 if (!css_is_populated(css))
6391 kill_css_finish(css);
6392 }
6393
6394 return 0;
6395};
6396
6397int cgroup_rmdir(struct kernfs_node *kn)
6398{
6399 struct cgroup *cgrp;
6400 int ret = 0;
6401
6402 cgrp = cgroup_kn_lock_live(kn, false);
6403 if (!cgrp)
6404 return 0;
6405
6406 ret = cgroup_destroy_locked(cgrp);
6407 if (!ret)
6408 TRACE_CGROUP_PATH(rmdir, cgrp);
6409
6410 cgroup_kn_unlock(kn);
6411 return ret;
6412}
6413
6414static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
6415 .show_options = cgroup_show_options,
6416 .mkdir = cgroup_mkdir,
6417 .rmdir = cgroup_rmdir,
6418 .show_path = cgroup_show_path,
6419};
6420
6421static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
6422{
6423 struct cgroup_subsys_state *css;
6424
6425 pr_debug("Initializing cgroup subsys %s\n", ss->name);
6426
6427 cgroup_lock();
6428
6429 idr_init(&ss->css_idr);
6430 INIT_LIST_HEAD(&ss->cfts);
6431
6432 /* Create the root cgroup state for this subsystem */
6433 ss->root = &cgrp_dfl_root;
6434 css = ss->css_alloc(NULL);
6435 /* We don't handle early failures gracefully */
6436 BUG_ON(IS_ERR(css));
6437 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
6438
6439 /*
6440 * Root csses are never destroyed and we can't initialize
6441 * percpu_ref during early init. Disable refcnting.
6442 */
6443 css->flags |= CSS_NO_REF;
6444
6445 if (early) {
6446 /* allocation can't be done safely during early init */
6447 css->id = 1;
6448 } else {
6449 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
6450 BUG_ON(css->id < 0);
6451
6452 BUG_ON(ss_rstat_init(ss));
6453 BUG_ON(css_rstat_init(css));
6454 }
6455
6456 /* Update the init_css_set to contain a subsys
6457 * pointer to this state - since the subsystem is
6458 * newly registered, all tasks and hence the
6459 * init_css_set is in the subsystem's root cgroup. */
6460 init_css_set.subsys[ss->id] = css;
6461
6462 have_fork_callback |= (bool)ss->fork << ss->id;
6463 have_exit_callback |= (bool)ss->exit << ss->id;
6464 have_release_callback |= (bool)ss->release << ss->id;
6465 have_canfork_callback |= (bool)ss->can_fork << ss->id;
6466
6467 /* At system boot, before all subsystems have been
6468 * registered, no tasks have been forked, so we don't
6469 * need to invoke fork callbacks here. */
6470 BUG_ON(!list_empty(&init_task.tasks));
6471
6472 BUG_ON(online_css(css));
6473
6474 cgroup_unlock();
6475}
6476
6477/**
6478 * cgroup_init_early - cgroup initialization at system boot
6479 *
6480 * Initialize cgroups at system boot, and initialize any
6481 * subsystems that request early init.
6482 */
6483int __init cgroup_init_early(void)
6484{
6485 static struct cgroup_fs_context __initdata ctx;
6486 struct cgroup_subsys *ss;
6487 int i;
6488
6489 ctx.root = &cgrp_dfl_root;
6490 init_cgroup_root(&ctx);
6491 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6492
6493 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6494
6495 for_each_subsys(ss, i) {
6496 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6497 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6498 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6499 ss->id, ss->name);
6500 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6501 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6502 WARN(ss->early_init && ss->css_rstat_flush,
6503 "cgroup rstat cannot be used with early init subsystem\n");
6504
6505 ss->id = i;
6506 ss->name = cgroup_subsys_name[i];
6507 if (!ss->legacy_name)
6508 ss->legacy_name = cgroup_subsys_name[i];
6509
6510 if (ss->early_init)
6511 cgroup_init_subsys(ss, true);
6512 }
6513 return 0;
6514}
6515
6516/**
6517 * cgroup_init - cgroup initialization
6518 *
6519 * Register cgroup filesystem and /proc file, and initialize
6520 * any subsystems that didn't request early init.
6521 */
6522int __init cgroup_init(void)
6523{
6524 struct cgroup_subsys *ss;
6525 int ssid;
6526
6527 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 32);
6528 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6529 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6530 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6531
6532 BUG_ON(ss_rstat_init(NULL));
6533
6534 get_user_ns(init_cgroup_ns.user_ns);
6535 cgroup_rt_init();
6536
6537 cgroup_lock();
6538
6539 /*
6540 * Add init_css_set to the hash table so that dfl_root can link to
6541 * it during init.
6542 */
6543 hash_add(css_set_table, &init_css_set.hlist,
6544 css_set_hash(init_css_set.subsys));
6545
6546 cgroup_bpf_lifetime_notifier_init();
6547
6548 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6549
6550 cgroup_unlock();
6551
6552 for_each_subsys(ss, ssid) {
6553 if (ss->early_init) {
6554 struct cgroup_subsys_state *css =
6555 init_css_set.subsys[ss->id];
6556
6557 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6558 GFP_KERNEL);
6559 BUG_ON(css->id < 0);
6560 } else {
6561 cgroup_init_subsys(ss, false);
6562 }
6563
6564 list_add_tail(&init_css_set.e_cset_node[ssid],
6565 &cgrp_dfl_root.cgrp.e_csets[ssid]);
6566
6567 /*
6568 * Setting dfl_root subsys_mask needs to consider the
6569 * disabled flag and cftype registration needs kmalloc,
6570 * both of which aren't available during early_init.
6571 */
6572 if (!cgroup_ssid_enabled(ssid))
6573 continue;
6574
6575 if (cgroup1_ssid_disabled(ssid))
6576 pr_info("Disabling %s control group subsystem in v1 mounts\n",
6577 ss->legacy_name);
6578
6579 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6580
6581 /* implicit controllers must be threaded too */
6582 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6583
6584 if (ss->implicit_on_dfl)
6585 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6586 else if (!ss->dfl_cftypes)
6587 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6588
6589 if (ss->threaded)
6590 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6591
6592 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6593 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6594 } else {
6595 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6596 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6597 }
6598
6599 if (ss->bind)
6600 ss->bind(init_css_set.subsys[ssid]);
6601
6602 cgroup_lock();
6603 css_populate_dir(init_css_set.subsys[ssid]);
6604 cgroup_unlock();
6605 }
6606
6607 /* init_css_set.subsys[] has been updated, re-hash */
6608 hash_del(&init_css_set.hlist);
6609 hash_add(css_set_table, &init_css_set.hlist,
6610 css_set_hash(init_css_set.subsys));
6611
6612 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6613 WARN_ON(register_filesystem(&cgroup_fs_type));
6614 WARN_ON(register_filesystem(&cgroup2_fs_type));
6615 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6616#ifdef CONFIG_CPUSETS_V1
6617 WARN_ON(register_filesystem(&cpuset_fs_type));
6618#endif
6619
6620 ns_tree_add(&init_cgroup_ns);
6621 return 0;
6622}
6623
6624static int __init cgroup_wq_init(void)
6625{
6626 /*
6627 * There isn't much point in executing destruction path in
6628 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6629 * Use 1 for @max_active.
6630 *
6631 * We would prefer to do this in cgroup_init() above, but that
6632 * is called before init_workqueues(): so leave this until after.
6633 */
6634 cgroup_offline_wq = alloc_workqueue("cgroup_offline", WQ_PERCPU, 1);
6635 BUG_ON(!cgroup_offline_wq);
6636
6637 cgroup_release_wq = alloc_workqueue("cgroup_release", WQ_PERCPU, 1);
6638 BUG_ON(!cgroup_release_wq);
6639
6640 cgroup_free_wq = alloc_workqueue("cgroup_free", WQ_PERCPU, 1);
6641 BUG_ON(!cgroup_free_wq);
6642 return 0;
6643}
6644core_initcall(cgroup_wq_init);
6645
6646void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6647{
6648 struct kernfs_node *kn;
6649
6650 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6651 if (!kn)
6652 return;
6653 kernfs_path(kn, buf, buflen);
6654 kernfs_put(kn);
6655}
6656
6657/*
6658 * __cgroup_get_from_id : get the cgroup associated with cgroup id
6659 * @id: cgroup id
6660 * On success return the cgrp or ERR_PTR on failure
6661 * There are no cgroup NS restrictions.
6662 */
6663struct cgroup *__cgroup_get_from_id(u64 id)
6664{
6665 struct kernfs_node *kn;
6666 struct cgroup *cgrp;
6667
6668 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6669 if (!kn)
6670 return ERR_PTR(-ENOENT);
6671
6672 if (kernfs_type(kn) != KERNFS_DIR) {
6673 kernfs_put(kn);
6674 return ERR_PTR(-ENOENT);
6675 }
6676
6677 rcu_read_lock();
6678
6679 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6680 if (cgrp && !cgroup_tryget(cgrp))
6681 cgrp = NULL;
6682
6683 rcu_read_unlock();
6684 kernfs_put(kn);
6685
6686 if (!cgrp)
6687 return ERR_PTR(-ENOENT);
6688 return cgrp;
6689}
6690
6691/*
6692 * cgroup_get_from_id : get the cgroup associated with cgroup id
6693 * @id: cgroup id
6694 * On success return the cgrp or ERR_PTR on failure
6695 * Only cgroups within current task's cgroup NS are valid.
6696 */
6697struct cgroup *cgroup_get_from_id(u64 id)
6698{
6699 struct cgroup *cgrp, *root_cgrp;
6700
6701 cgrp = __cgroup_get_from_id(id);
6702 if (IS_ERR(cgrp))
6703 return cgrp;
6704
6705 root_cgrp = current_cgns_cgroup_dfl();
6706 if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6707 cgroup_put(cgrp);
6708 return ERR_PTR(-ENOENT);
6709 }
6710
6711 return cgrp;
6712}
6713EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6714
6715/*
6716 * proc_cgroup_show()
6717 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6718 * - Used for /proc/<pid>/cgroup.
6719 */
6720int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6721 struct pid *pid, struct task_struct *tsk)
6722{
6723 char *buf;
6724 int retval;
6725 struct cgroup_root *root;
6726
6727 retval = -ENOMEM;
6728 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6729 if (!buf)
6730 goto out;
6731
6732 rcu_read_lock();
6733 spin_lock_irq(&css_set_lock);
6734
6735 for_each_root(root) {
6736 struct cgroup_subsys *ss;
6737 struct cgroup *cgrp;
6738 int ssid, count = 0;
6739
6740 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6741 continue;
6742
6743 cgrp = task_cgroup_from_root(tsk, root);
6744 /* The root has already been unmounted. */
6745 if (!cgrp)
6746 continue;
6747
6748 seq_printf(m, "%d:", root->hierarchy_id);
6749 if (root != &cgrp_dfl_root)
6750 for_each_subsys(ss, ssid)
6751 if (root->subsys_mask & (1 << ssid))
6752 seq_printf(m, "%s%s", count++ ? "," : "",
6753 ss->legacy_name);
6754 if (strlen(root->name))
6755 seq_printf(m, "%sname=%s", count ? "," : "",
6756 root->name);
6757 seq_putc(m, ':');
6758 /*
6759 * On traditional hierarchies, all zombie tasks show up as
6760 * belonging to the root cgroup. On the default hierarchy,
6761 * while a zombie doesn't show up in "cgroup.procs" and
6762 * thus can't be migrated, its /proc/PID/cgroup keeps
6763 * reporting the cgroup it belonged to before exiting. If
6764 * the cgroup is removed before the zombie is reaped,
6765 * " (deleted)" is appended to the cgroup path.
6766 */
6767 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6768 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6769 current->nsproxy->cgroup_ns);
6770 if (retval == -E2BIG)
6771 retval = -ENAMETOOLONG;
6772 if (retval < 0)
6773 goto out_unlock;
6774
6775 seq_puts(m, buf);
6776 } else {
6777 seq_puts(m, "/");
6778 }
6779
6780 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6781 seq_puts(m, " (deleted)\n");
6782 else
6783 seq_putc(m, '\n');
6784 }
6785
6786 retval = 0;
6787out_unlock:
6788 spin_unlock_irq(&css_set_lock);
6789 rcu_read_unlock();
6790 kfree(buf);
6791out:
6792 return retval;
6793}
6794
6795/**
6796 * cgroup_fork - initialize cgroup related fields during copy_process()
6797 * @child: pointer to task_struct of forking parent process.
6798 *
6799 * A task is associated with the init_css_set until cgroup_post_fork()
6800 * attaches it to the target css_set.
6801 */
6802void cgroup_fork(struct task_struct *child)
6803{
6804 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6805 INIT_LIST_HEAD(&child->cg_list);
6806}
6807
6808/**
6809 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6810 * @f: file corresponding to cgroup_dir
6811 *
6812 * Find the cgroup from a file pointer associated with a cgroup directory.
6813 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6814 * cgroup cannot be found.
6815 */
6816static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6817{
6818 struct cgroup_subsys_state *css;
6819
6820 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6821 if (IS_ERR(css))
6822 return ERR_CAST(css);
6823
6824 return css->cgroup;
6825}
6826
6827/**
6828 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6829 * cgroup2.
6830 * @f: file corresponding to cgroup2_dir
6831 */
6832static struct cgroup *cgroup_get_from_file(struct file *f)
6833{
6834 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6835
6836 if (IS_ERR(cgrp))
6837 return ERR_CAST(cgrp);
6838
6839 if (!cgroup_on_dfl(cgrp)) {
6840 cgroup_put(cgrp);
6841 return ERR_PTR(-EBADF);
6842 }
6843
6844 return cgrp;
6845}
6846
6847/**
6848 * cgroup_css_set_fork - find or create a css_set for a child process
6849 * @kargs: the arguments passed to create the child process
6850 *
6851 * This functions finds or creates a new css_set which the child
6852 * process will be attached to in cgroup_post_fork(). By default,
6853 * the child process will be given the same css_set as its parent.
6854 *
6855 * If CLONE_INTO_CGROUP is specified this function will try to find an
6856 * existing css_set which includes the requested cgroup and if not create
6857 * a new css_set that the child will be attached to later. If this function
6858 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6859 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6860 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6861 * to the target cgroup.
6862 */
6863static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6864 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6865{
6866 int ret;
6867 struct cgroup *dst_cgrp = NULL;
6868 struct css_set *cset;
6869 struct super_block *sb;
6870
6871 if (kargs->flags & CLONE_INTO_CGROUP)
6872 cgroup_lock();
6873
6874 cgroup_threadgroup_change_begin(current);
6875
6876 spin_lock_irq(&css_set_lock);
6877 cset = task_css_set(current);
6878 get_css_set(cset);
6879 kargs->kill_seq = cset->dfl_cgrp->kill_seq;
6880 spin_unlock_irq(&css_set_lock);
6881
6882 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6883 kargs->cset = cset;
6884 return 0;
6885 }
6886
6887 CLASS(fd_raw, f)(kargs->cgroup);
6888 if (fd_empty(f)) {
6889 ret = -EBADF;
6890 goto err;
6891 }
6892 sb = fd_file(f)->f_path.dentry->d_sb;
6893
6894 dst_cgrp = cgroup_get_from_file(fd_file(f));
6895 if (IS_ERR(dst_cgrp)) {
6896 ret = PTR_ERR(dst_cgrp);
6897 dst_cgrp = NULL;
6898 goto err;
6899 }
6900
6901 if (cgroup_is_dead(dst_cgrp)) {
6902 ret = -ENODEV;
6903 goto err;
6904 }
6905
6906 /*
6907 * Verify that we the target cgroup is writable for us. This is
6908 * usually done by the vfs layer but since we're not going through
6909 * the vfs layer here we need to do it "manually".
6910 */
6911 ret = cgroup_may_write(dst_cgrp, sb);
6912 if (ret)
6913 goto err;
6914
6915 /*
6916 * Spawning a task directly into a cgroup works by passing a file
6917 * descriptor to the target cgroup directory. This can even be an O_PATH
6918 * file descriptor. But it can never be a cgroup.procs file descriptor.
6919 * This was done on purpose so spawning into a cgroup could be
6920 * conceptualized as an atomic
6921 *
6922 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6923 * write(fd, <child-pid>, ...);
6924 *
6925 * sequence, i.e. it's a shorthand for the caller opening and writing
6926 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6927 * to always use the caller's credentials.
6928 */
6929 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6930 !(kargs->flags & CLONE_THREAD),
6931 current->nsproxy->cgroup_ns);
6932 if (ret)
6933 goto err;
6934
6935 kargs->cset = find_css_set(cset, dst_cgrp);
6936 if (!kargs->cset) {
6937 ret = -ENOMEM;
6938 goto err;
6939 }
6940
6941 put_css_set(cset);
6942 kargs->cgrp = dst_cgrp;
6943 kargs->kill_seq = dst_cgrp->kill_seq;
6944 return ret;
6945
6946err:
6947 cgroup_threadgroup_change_end(current);
6948 cgroup_unlock();
6949 if (dst_cgrp)
6950 cgroup_put(dst_cgrp);
6951 put_css_set(cset);
6952 if (kargs->cset)
6953 put_css_set(kargs->cset);
6954 return ret;
6955}
6956
6957/**
6958 * cgroup_css_set_put_fork - drop references we took during fork
6959 * @kargs: the arguments passed to create the child process
6960 *
6961 * Drop references to the prepared css_set and target cgroup if
6962 * CLONE_INTO_CGROUP was requested.
6963 */
6964static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6965 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6966{
6967 struct cgroup *cgrp = kargs->cgrp;
6968 struct css_set *cset = kargs->cset;
6969
6970 cgroup_threadgroup_change_end(current);
6971
6972 if (cset) {
6973 put_css_set(cset);
6974 kargs->cset = NULL;
6975 }
6976
6977 if (kargs->flags & CLONE_INTO_CGROUP) {
6978 cgroup_unlock();
6979 if (cgrp) {
6980 cgroup_put(cgrp);
6981 kargs->cgrp = NULL;
6982 }
6983 }
6984}
6985
6986/**
6987 * cgroup_can_fork - called on a new task before the process is exposed
6988 * @child: the child process
6989 * @kargs: the arguments passed to create the child process
6990 *
6991 * This prepares a new css_set for the child process which the child will
6992 * be attached to in cgroup_post_fork().
6993 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6994 * callback returns an error, the fork aborts with that error code. This
6995 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6996 */
6997int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6998{
6999 struct cgroup_subsys *ss;
7000 int i, j, ret;
7001
7002 ret = cgroup_css_set_fork(kargs);
7003 if (ret)
7004 return ret;
7005
7006 do_each_subsys_mask(ss, i, have_canfork_callback) {
7007 ret = ss->can_fork(child, kargs->cset);
7008 if (ret)
7009 goto out_revert;
7010 } while_each_subsys_mask();
7011
7012 return 0;
7013
7014out_revert:
7015 for_each_subsys(ss, j) {
7016 if (j >= i)
7017 break;
7018 if (ss->cancel_fork)
7019 ss->cancel_fork(child, kargs->cset);
7020 }
7021
7022 cgroup_css_set_put_fork(kargs);
7023
7024 return ret;
7025}
7026
7027/**
7028 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
7029 * @child: the child process
7030 * @kargs: the arguments passed to create the child process
7031 *
7032 * This calls the cancel_fork() callbacks if a fork failed *after*
7033 * cgroup_can_fork() succeeded and cleans up references we took to
7034 * prepare a new css_set for the child process in cgroup_can_fork().
7035 */
7036void cgroup_cancel_fork(struct task_struct *child,
7037 struct kernel_clone_args *kargs)
7038{
7039 struct cgroup_subsys *ss;
7040 int i;
7041
7042 for_each_subsys(ss, i)
7043 if (ss->cancel_fork)
7044 ss->cancel_fork(child, kargs->cset);
7045
7046 cgroup_css_set_put_fork(kargs);
7047}
7048
7049/**
7050 * cgroup_post_fork - finalize cgroup setup for the child process
7051 * @child: the child process
7052 * @kargs: the arguments passed to create the child process
7053 *
7054 * Attach the child process to its css_set calling the subsystem fork()
7055 * callbacks.
7056 */
7057void cgroup_post_fork(struct task_struct *child,
7058 struct kernel_clone_args *kargs)
7059 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
7060{
7061 unsigned int cgrp_kill_seq = 0;
7062 unsigned long cgrp_flags = 0;
7063 bool kill = false;
7064 struct cgroup_subsys *ss;
7065 struct css_set *cset;
7066 int i;
7067
7068 cset = kargs->cset;
7069 kargs->cset = NULL;
7070
7071 spin_lock_irq(&css_set_lock);
7072
7073 /* init tasks are special, only link regular threads */
7074 if (likely(child->pid)) {
7075 if (kargs->cgrp) {
7076 cgrp_flags = kargs->cgrp->flags;
7077 cgrp_kill_seq = kargs->cgrp->kill_seq;
7078 } else {
7079 cgrp_flags = cset->dfl_cgrp->flags;
7080 cgrp_kill_seq = cset->dfl_cgrp->kill_seq;
7081 }
7082
7083 WARN_ON_ONCE(!list_empty(&child->cg_list));
7084 cset->nr_tasks++;
7085 css_set_move_task(child, NULL, cset, false);
7086 } else {
7087 put_css_set(cset);
7088 cset = NULL;
7089 }
7090
7091 if (!(child->flags & PF_KTHREAD)) {
7092 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
7093 /*
7094 * If the cgroup has to be frozen, the new task has
7095 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
7096 * get the task into the frozen state.
7097 */
7098 spin_lock(&child->sighand->siglock);
7099 WARN_ON_ONCE(child->frozen);
7100 child->jobctl |= JOBCTL_TRAP_FREEZE;
7101 spin_unlock(&child->sighand->siglock);
7102
7103 /*
7104 * Calling cgroup_update_frozen() isn't required here,
7105 * because it will be called anyway a bit later from
7106 * do_freezer_trap(). So we avoid cgroup's transient
7107 * switch from the frozen state and back.
7108 */
7109 }
7110
7111 /*
7112 * If the cgroup is to be killed notice it now and take the
7113 * child down right after we finished preparing it for
7114 * userspace.
7115 */
7116 kill = kargs->kill_seq != cgrp_kill_seq;
7117 }
7118
7119 spin_unlock_irq(&css_set_lock);
7120
7121 /*
7122 * Call ss->fork(). This must happen after @child is linked on
7123 * css_set; otherwise, @child might change state between ->fork()
7124 * and addition to css_set.
7125 */
7126 do_each_subsys_mask(ss, i, have_fork_callback) {
7127 ss->fork(child);
7128 } while_each_subsys_mask();
7129
7130 /* Make the new cset the root_cset of the new cgroup namespace. */
7131 if (kargs->flags & CLONE_NEWCGROUP) {
7132 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
7133
7134 get_css_set(cset);
7135 child->nsproxy->cgroup_ns->root_cset = cset;
7136 put_css_set(rcset);
7137 }
7138
7139 /* Cgroup has to be killed so take down child immediately. */
7140 if (unlikely(kill))
7141 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
7142
7143 cgroup_css_set_put_fork(kargs);
7144}
7145
7146/**
7147 * cgroup_task_exit - detach cgroup from exiting task
7148 * @tsk: pointer to task_struct of exiting process
7149 *
7150 * Description: Detach cgroup from @tsk.
7151 *
7152 */
7153void cgroup_task_exit(struct task_struct *tsk)
7154{
7155 struct cgroup_subsys *ss;
7156 int i;
7157
7158 /* see cgroup_post_fork() for details */
7159 do_each_subsys_mask(ss, i, have_exit_callback) {
7160 ss->exit(tsk);
7161 } while_each_subsys_mask();
7162}
7163
7164static void do_cgroup_task_dead(struct task_struct *tsk)
7165{
7166 struct css_set *cset;
7167 unsigned long flags;
7168
7169 spin_lock_irqsave(&css_set_lock, flags);
7170
7171 WARN_ON_ONCE(list_empty(&tsk->cg_list));
7172 cset = task_css_set(tsk);
7173 css_set_move_task(tsk, cset, NULL, false);
7174 cset->nr_tasks--;
7175 /* matches the signal->live check in css_task_iter_advance() */
7176 if (thread_group_leader(tsk) && atomic_read(&tsk->signal->live))
7177 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
7178
7179 if (dl_task(tsk))
7180 dec_dl_tasks_cs(tsk);
7181
7182 WARN_ON_ONCE(cgroup_task_frozen(tsk));
7183 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
7184 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
7185 cgroup_update_frozen(task_dfl_cgroup(tsk));
7186
7187 spin_unlock_irqrestore(&css_set_lock, flags);
7188}
7189
7190#ifdef CONFIG_PREEMPT_RT
7191/*
7192 * cgroup_task_dead() is called from finish_task_switch() which doesn't allow
7193 * scheduling even in RT. As the task_dead path requires grabbing css_set_lock,
7194 * this lead to sleeping in the invalid context warning bug. css_set_lock is too
7195 * big to become a raw_spinlock. The task_dead path doesn't need to run
7196 * synchronously but can't be delayed indefinitely either as the dead task pins
7197 * the cgroup and task_struct can be pinned indefinitely. Bounce through lazy
7198 * irq_work to allow batching while ensuring timely completion.
7199 */
7200static DEFINE_PER_CPU(struct llist_head, cgrp_dead_tasks);
7201static DEFINE_PER_CPU(struct irq_work, cgrp_dead_tasks_iwork);
7202
7203static void cgrp_dead_tasks_iwork_fn(struct irq_work *iwork)
7204{
7205 struct llist_node *lnode;
7206 struct task_struct *task, *next;
7207
7208 lnode = llist_del_all(this_cpu_ptr(&cgrp_dead_tasks));
7209 llist_for_each_entry_safe(task, next, lnode, cg_dead_lnode) {
7210 do_cgroup_task_dead(task);
7211 put_task_struct(task);
7212 }
7213}
7214
7215static void __init cgroup_rt_init(void)
7216{
7217 int cpu;
7218
7219 for_each_possible_cpu(cpu) {
7220 init_llist_head(per_cpu_ptr(&cgrp_dead_tasks, cpu));
7221 per_cpu(cgrp_dead_tasks_iwork, cpu) =
7222 IRQ_WORK_INIT_LAZY(cgrp_dead_tasks_iwork_fn);
7223 }
7224}
7225
7226void cgroup_task_dead(struct task_struct *task)
7227{
7228 get_task_struct(task);
7229 llist_add(&task->cg_dead_lnode, this_cpu_ptr(&cgrp_dead_tasks));
7230 irq_work_queue(this_cpu_ptr(&cgrp_dead_tasks_iwork));
7231}
7232#else /* CONFIG_PREEMPT_RT */
7233static void __init cgroup_rt_init(void) {}
7234
7235void cgroup_task_dead(struct task_struct *task)
7236{
7237 do_cgroup_task_dead(task);
7238}
7239#endif /* CONFIG_PREEMPT_RT */
7240
7241void cgroup_task_release(struct task_struct *task)
7242{
7243 struct cgroup_subsys *ss;
7244 int ssid;
7245
7246 do_each_subsys_mask(ss, ssid, have_release_callback) {
7247 ss->release(task);
7248 } while_each_subsys_mask();
7249}
7250
7251void cgroup_task_free(struct task_struct *task)
7252{
7253 struct css_set *cset = task_css_set(task);
7254
7255 if (!list_empty(&task->cg_list)) {
7256 spin_lock_irq(&css_set_lock);
7257 css_set_skip_task_iters(task_css_set(task), task);
7258 list_del_init(&task->cg_list);
7259 spin_unlock_irq(&css_set_lock);
7260 }
7261
7262 put_css_set(cset);
7263}
7264
7265static int __init cgroup_disable(char *str)
7266{
7267 struct cgroup_subsys *ss;
7268 char *token;
7269 int i;
7270
7271 while ((token = strsep(&str, ",")) != NULL) {
7272 if (!*token)
7273 continue;
7274
7275 for_each_subsys(ss, i) {
7276 if (strcmp(token, ss->name) &&
7277 strcmp(token, ss->legacy_name))
7278 continue;
7279
7280 static_branch_disable(cgroup_subsys_enabled_key[i]);
7281 pr_info("Disabling %s control group subsystem\n",
7282 ss->name);
7283 }
7284
7285 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
7286 if (strcmp(token, cgroup_opt_feature_names[i]))
7287 continue;
7288 cgroup_feature_disable_mask |= 1 << i;
7289 pr_info("Disabling %s control group feature\n",
7290 cgroup_opt_feature_names[i]);
7291 break;
7292 }
7293 }
7294 return 1;
7295}
7296__setup("cgroup_disable=", cgroup_disable);
7297
7298void __init __weak enable_debug_cgroup(void) { }
7299
7300static int __init enable_cgroup_debug(char *str)
7301{
7302 cgroup_debug = true;
7303 enable_debug_cgroup();
7304 return 1;
7305}
7306__setup("cgroup_debug", enable_cgroup_debug);
7307
7308static int __init cgroup_favordynmods_setup(char *str)
7309{
7310 return (kstrtobool(str, &have_favordynmods) == 0);
7311}
7312__setup("cgroup_favordynmods=", cgroup_favordynmods_setup);
7313
7314/**
7315 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
7316 * @dentry: directory dentry of interest
7317 * @ss: subsystem of interest
7318 *
7319 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
7320 * to get the corresponding css and return it. If such css doesn't exist
7321 * or can't be pinned, an ERR_PTR value is returned.
7322 */
7323struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
7324 struct cgroup_subsys *ss)
7325{
7326 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
7327 struct file_system_type *s_type = dentry->d_sb->s_type;
7328 struct cgroup_subsys_state *css = NULL;
7329 struct cgroup *cgrp;
7330
7331 /* is @dentry a cgroup dir? */
7332 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
7333 !kn || kernfs_type(kn) != KERNFS_DIR)
7334 return ERR_PTR(-EBADF);
7335
7336 rcu_read_lock();
7337
7338 /*
7339 * This path doesn't originate from kernfs and @kn could already
7340 * have been or be removed at any point. @kn->priv is RCU
7341 * protected for this access. See css_release_work_fn() for details.
7342 */
7343 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
7344 if (cgrp)
7345 css = cgroup_css(cgrp, ss);
7346
7347 if (!css || !css_tryget_online(css))
7348 css = ERR_PTR(-ENOENT);
7349
7350 rcu_read_unlock();
7351 return css;
7352}
7353
7354/**
7355 * css_from_id - lookup css by id
7356 * @id: the cgroup id
7357 * @ss: cgroup subsys to be looked into
7358 *
7359 * Returns the css if there's valid one with @id, otherwise returns NULL.
7360 * Should be called under rcu_read_lock().
7361 */
7362struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
7363{
7364 WARN_ON_ONCE(!rcu_read_lock_held());
7365 return idr_find(&ss->css_idr, id);
7366}
7367
7368/**
7369 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
7370 * @path: path on the default hierarchy
7371 *
7372 * Find the cgroup at @path on the default hierarchy, increment its
7373 * reference count and return it. Returns pointer to the found cgroup on
7374 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
7375 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
7376 */
7377struct cgroup *cgroup_get_from_path(const char *path)
7378{
7379 struct kernfs_node *kn;
7380 struct cgroup *cgrp = ERR_PTR(-ENOENT);
7381 struct cgroup *root_cgrp;
7382
7383 root_cgrp = current_cgns_cgroup_dfl();
7384 kn = kernfs_walk_and_get(root_cgrp->kn, path);
7385 if (!kn)
7386 goto out;
7387
7388 if (kernfs_type(kn) != KERNFS_DIR) {
7389 cgrp = ERR_PTR(-ENOTDIR);
7390 goto out_kernfs;
7391 }
7392
7393 rcu_read_lock();
7394
7395 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
7396 if (!cgrp || !cgroup_tryget(cgrp))
7397 cgrp = ERR_PTR(-ENOENT);
7398
7399 rcu_read_unlock();
7400
7401out_kernfs:
7402 kernfs_put(kn);
7403out:
7404 return cgrp;
7405}
7406EXPORT_SYMBOL_GPL(cgroup_get_from_path);
7407
7408/**
7409 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
7410 * @fd: fd obtained by open(cgroup_dir)
7411 *
7412 * Find the cgroup from a fd which should be obtained
7413 * by opening a cgroup directory. Returns a pointer to the
7414 * cgroup on success. ERR_PTR is returned if the cgroup
7415 * cannot be found.
7416 */
7417struct cgroup *cgroup_v1v2_get_from_fd(int fd)
7418{
7419 CLASS(fd_raw, f)(fd);
7420 if (fd_empty(f))
7421 return ERR_PTR(-EBADF);
7422
7423 return cgroup_v1v2_get_from_file(fd_file(f));
7424}
7425
7426/**
7427 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
7428 * cgroup2.
7429 * @fd: fd obtained by open(cgroup2_dir)
7430 */
7431struct cgroup *cgroup_get_from_fd(int fd)
7432{
7433 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
7434
7435 if (IS_ERR(cgrp))
7436 return ERR_CAST(cgrp);
7437
7438 if (!cgroup_on_dfl(cgrp)) {
7439 cgroup_put(cgrp);
7440 return ERR_PTR(-EBADF);
7441 }
7442 return cgrp;
7443}
7444EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
7445
7446static u64 power_of_ten(int power)
7447{
7448 u64 v = 1;
7449 while (power--)
7450 v *= 10;
7451 return v;
7452}
7453
7454/**
7455 * cgroup_parse_float - parse a floating number
7456 * @input: input string
7457 * @dec_shift: number of decimal digits to shift
7458 * @v: output
7459 *
7460 * Parse a decimal floating point number in @input and store the result in
7461 * @v with decimal point right shifted @dec_shift times. For example, if
7462 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
7463 * Returns 0 on success, -errno otherwise.
7464 *
7465 * There's nothing cgroup specific about this function except that it's
7466 * currently the only user.
7467 */
7468int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
7469{
7470 s64 whole, frac = 0;
7471 int fstart = 0, fend = 0, flen;
7472
7473 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
7474 return -EINVAL;
7475 if (frac < 0)
7476 return -EINVAL;
7477
7478 flen = fend > fstart ? fend - fstart : 0;
7479 if (flen < dec_shift)
7480 frac *= power_of_ten(dec_shift - flen);
7481 else
7482 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
7483
7484 *v = whole * power_of_ten(dec_shift) + frac;
7485 return 0;
7486}
7487
7488/*
7489 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
7490 * definition in cgroup-defs.h.
7491 */
7492#ifdef CONFIG_SOCK_CGROUP_DATA
7493
7494void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
7495{
7496 struct cgroup *cgroup;
7497
7498 rcu_read_lock();
7499 /* Don't associate the sock with unrelated interrupted task's cgroup. */
7500 if (in_interrupt()) {
7501 cgroup = &cgrp_dfl_root.cgrp;
7502 cgroup_get(cgroup);
7503 goto out;
7504 }
7505
7506 while (true) {
7507 struct css_set *cset;
7508
7509 cset = task_css_set(current);
7510 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
7511 cgroup = cset->dfl_cgrp;
7512 break;
7513 }
7514 cpu_relax();
7515 }
7516out:
7517 skcd->cgroup = cgroup;
7518 cgroup_bpf_get(cgroup);
7519 rcu_read_unlock();
7520}
7521
7522void cgroup_sk_clone(struct sock_cgroup_data *skcd)
7523{
7524 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7525
7526 /*
7527 * We might be cloning a socket which is left in an empty
7528 * cgroup and the cgroup might have already been rmdir'd.
7529 * Don't use cgroup_get_live().
7530 */
7531 cgroup_get(cgrp);
7532 cgroup_bpf_get(cgrp);
7533}
7534
7535void cgroup_sk_free(struct sock_cgroup_data *skcd)
7536{
7537 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7538
7539 cgroup_bpf_put(cgrp);
7540 cgroup_put(cgrp);
7541}
7542
7543#endif /* CONFIG_SOCK_CGROUP_DATA */
7544
7545#ifdef CONFIG_SYSFS
7546static ssize_t show_delegatable_files(struct cftype *files, char *buf,
7547 ssize_t size, const char *prefix)
7548{
7549 struct cftype *cft;
7550 ssize_t ret = 0;
7551
7552 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
7553 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
7554 continue;
7555
7556 if (prefix)
7557 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
7558
7559 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
7560
7561 if (WARN_ON(ret >= size))
7562 break;
7563 }
7564
7565 return ret;
7566}
7567
7568static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7569 char *buf)
7570{
7571 struct cgroup_subsys *ss;
7572 int ssid;
7573 ssize_t ret = 0;
7574
7575 ret = show_delegatable_files(cgroup_base_files, buf + ret,
7576 PAGE_SIZE - ret, NULL);
7577 if (cgroup_psi_enabled())
7578 ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7579 PAGE_SIZE - ret, NULL);
7580
7581 for_each_subsys(ss, ssid)
7582 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7583 PAGE_SIZE - ret,
7584 cgroup_subsys_name[ssid]);
7585
7586 return ret;
7587}
7588static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7589
7590static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7591 char *buf)
7592{
7593 return snprintf(buf, PAGE_SIZE,
7594 "nsdelegate\n"
7595 "favordynmods\n"
7596 "memory_localevents\n"
7597 "memory_recursiveprot\n"
7598 "memory_hugetlb_accounting\n"
7599 "pids_localevents\n");
7600}
7601static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7602
7603static struct attribute *cgroup_sysfs_attrs[] = {
7604 &cgroup_delegate_attr.attr,
7605 &cgroup_features_attr.attr,
7606 NULL,
7607};
7608
7609static const struct attribute_group cgroup_sysfs_attr_group = {
7610 .attrs = cgroup_sysfs_attrs,
7611 .name = "cgroup",
7612};
7613
7614static int __init cgroup_sysfs_init(void)
7615{
7616 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7617}
7618subsys_initcall(cgroup_sysfs_init);
7619
7620#endif /* CONFIG_SYSFS */