Loading...
1/*
2 * Generic process-grouping system.
3 *
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
6 *
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
10 *
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
15 *
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
18 *
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
23 *
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
27 */
28
29#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
31#include "cgroup-internal.h"
32
33#include <linux/bpf-cgroup.h>
34#include <linux/cred.h>
35#include <linux/errno.h>
36#include <linux/init_task.h>
37#include <linux/kernel.h>
38#include <linux/magic.h>
39#include <linux/mutex.h>
40#include <linux/mount.h>
41#include <linux/pagemap.h>
42#include <linux/proc_fs.h>
43#include <linux/rcupdate.h>
44#include <linux/sched.h>
45#include <linux/sched/task.h>
46#include <linux/slab.h>
47#include <linux/spinlock.h>
48#include <linux/percpu-rwsem.h>
49#include <linux/string.h>
50#include <linux/hashtable.h>
51#include <linux/idr.h>
52#include <linux/kthread.h>
53#include <linux/atomic.h>
54#include <linux/cpuset.h>
55#include <linux/proc_ns.h>
56#include <linux/nsproxy.h>
57#include <linux/file.h>
58#include <linux/fs_parser.h>
59#include <linux/sched/cputime.h>
60#include <linux/sched/deadline.h>
61#include <linux/psi.h>
62#include <net/sock.h>
63
64#define CREATE_TRACE_POINTS
65#include <trace/events/cgroup.h>
66
67#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
68 MAX_CFTYPE_NAME + 2)
69/* let's not notify more than 100 times per second */
70#define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
71
72/*
73 * To avoid confusing the compiler (and generating warnings) with code
74 * that attempts to access what would be a 0-element array (i.e. sized
75 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this
76 * constant expression can be added.
77 */
78#define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0)
79
80/*
81 * cgroup_mutex is the master lock. Any modification to cgroup or its
82 * hierarchy must be performed while holding it.
83 *
84 * css_set_lock protects task->cgroups pointer, the list of css_set
85 * objects, and the chain of tasks off each css_set.
86 *
87 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
88 * cgroup.h can use them for lockdep annotations.
89 */
90DEFINE_MUTEX(cgroup_mutex);
91DEFINE_SPINLOCK(css_set_lock);
92
93#ifdef CONFIG_PROVE_RCU
94EXPORT_SYMBOL_GPL(cgroup_mutex);
95EXPORT_SYMBOL_GPL(css_set_lock);
96#endif
97
98DEFINE_SPINLOCK(trace_cgroup_path_lock);
99char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
100static bool cgroup_debug __read_mostly;
101
102/*
103 * Protects cgroup_idr and css_idr so that IDs can be released without
104 * grabbing cgroup_mutex.
105 */
106static DEFINE_SPINLOCK(cgroup_idr_lock);
107
108/*
109 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
110 * against file removal/re-creation across css hiding.
111 */
112static DEFINE_SPINLOCK(cgroup_file_kn_lock);
113
114DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
115
116#define cgroup_assert_mutex_or_rcu_locked() \
117 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
118 !lockdep_is_held(&cgroup_mutex), \
119 "cgroup_mutex or RCU read lock required");
120
121/*
122 * cgroup destruction makes heavy use of work items and there can be a lot
123 * of concurrent destructions. Use a separate workqueue so that cgroup
124 * destruction work items don't end up filling up max_active of system_wq
125 * which may lead to deadlock.
126 */
127static struct workqueue_struct *cgroup_destroy_wq;
128
129/* generate an array of cgroup subsystem pointers */
130#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
131struct cgroup_subsys *cgroup_subsys[] = {
132#include <linux/cgroup_subsys.h>
133};
134#undef SUBSYS
135
136/* array of cgroup subsystem names */
137#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
138static const char *cgroup_subsys_name[] = {
139#include <linux/cgroup_subsys.h>
140};
141#undef SUBSYS
142
143/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144#define SUBSYS(_x) \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
146 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
148 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
149#include <linux/cgroup_subsys.h>
150#undef SUBSYS
151
152#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
153static struct static_key_true *cgroup_subsys_enabled_key[] = {
154#include <linux/cgroup_subsys.h>
155};
156#undef SUBSYS
157
158#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
159static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
160#include <linux/cgroup_subsys.h>
161};
162#undef SUBSYS
163
164static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
165
166/* the default hierarchy */
167struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
168EXPORT_SYMBOL_GPL(cgrp_dfl_root);
169
170/*
171 * The default hierarchy always exists but is hidden until mounted for the
172 * first time. This is for backward compatibility.
173 */
174static bool cgrp_dfl_visible;
175
176/* some controllers are not supported in the default hierarchy */
177static u16 cgrp_dfl_inhibit_ss_mask;
178
179/* some controllers are implicitly enabled on the default hierarchy */
180static u16 cgrp_dfl_implicit_ss_mask;
181
182/* some controllers can be threaded on the default hierarchy */
183static u16 cgrp_dfl_threaded_ss_mask;
184
185/* The list of hierarchy roots */
186LIST_HEAD(cgroup_roots);
187static int cgroup_root_count;
188
189/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
190static DEFINE_IDR(cgroup_hierarchy_idr);
191
192/*
193 * Assign a monotonically increasing serial number to csses. It guarantees
194 * cgroups with bigger numbers are newer than those with smaller numbers.
195 * Also, as csses are always appended to the parent's ->children list, it
196 * guarantees that sibling csses are always sorted in the ascending serial
197 * number order on the list. Protected by cgroup_mutex.
198 */
199static u64 css_serial_nr_next = 1;
200
201/*
202 * These bitmasks identify subsystems with specific features to avoid
203 * having to do iterative checks repeatedly.
204 */
205static u16 have_fork_callback __read_mostly;
206static u16 have_exit_callback __read_mostly;
207static u16 have_release_callback __read_mostly;
208static u16 have_canfork_callback __read_mostly;
209
210static bool have_favordynmods __ro_after_init = IS_ENABLED(CONFIG_CGROUP_FAVOR_DYNMODS);
211
212/* cgroup namespace for init task */
213struct cgroup_namespace init_cgroup_ns = {
214 .ns.count = REFCOUNT_INIT(2),
215 .user_ns = &init_user_ns,
216 .ns.ops = &cgroupns_operations,
217 .ns.inum = PROC_CGROUP_INIT_INO,
218 .root_cset = &init_css_set,
219};
220
221static struct file_system_type cgroup2_fs_type;
222static struct cftype cgroup_base_files[];
223static struct cftype cgroup_psi_files[];
224
225/* cgroup optional features */
226enum cgroup_opt_features {
227#ifdef CONFIG_PSI
228 OPT_FEATURE_PRESSURE,
229#endif
230 OPT_FEATURE_COUNT
231};
232
233static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
234#ifdef CONFIG_PSI
235 "pressure",
236#endif
237};
238
239static u16 cgroup_feature_disable_mask __read_mostly;
240
241static int cgroup_apply_control(struct cgroup *cgrp);
242static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
243static void css_task_iter_skip(struct css_task_iter *it,
244 struct task_struct *task);
245static int cgroup_destroy_locked(struct cgroup *cgrp);
246static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
247 struct cgroup_subsys *ss);
248static void css_release(struct percpu_ref *ref);
249static void kill_css(struct cgroup_subsys_state *css);
250static int cgroup_addrm_files(struct cgroup_subsys_state *css,
251 struct cgroup *cgrp, struct cftype cfts[],
252 bool is_add);
253
254#ifdef CONFIG_DEBUG_CGROUP_REF
255#define CGROUP_REF_FN_ATTRS noinline
256#define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn);
257#include <linux/cgroup_refcnt.h>
258#endif
259
260/**
261 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
262 * @ssid: subsys ID of interest
263 *
264 * cgroup_subsys_enabled() can only be used with literal subsys names which
265 * is fine for individual subsystems but unsuitable for cgroup core. This
266 * is slower static_key_enabled() based test indexed by @ssid.
267 */
268bool cgroup_ssid_enabled(int ssid)
269{
270 if (!CGROUP_HAS_SUBSYS_CONFIG)
271 return false;
272
273 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
274}
275
276/**
277 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
278 * @cgrp: the cgroup of interest
279 *
280 * The default hierarchy is the v2 interface of cgroup and this function
281 * can be used to test whether a cgroup is on the default hierarchy for
282 * cases where a subsystem should behave differently depending on the
283 * interface version.
284 *
285 * List of changed behaviors:
286 *
287 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
288 * and "name" are disallowed.
289 *
290 * - When mounting an existing superblock, mount options should match.
291 *
292 * - rename(2) is disallowed.
293 *
294 * - "tasks" is removed. Everything should be at process granularity. Use
295 * "cgroup.procs" instead.
296 *
297 * - "cgroup.procs" is not sorted. pids will be unique unless they got
298 * recycled in-between reads.
299 *
300 * - "release_agent" and "notify_on_release" are removed. Replacement
301 * notification mechanism will be implemented.
302 *
303 * - "cgroup.clone_children" is removed.
304 *
305 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
306 * and its descendants contain no task; otherwise, 1. The file also
307 * generates kernfs notification which can be monitored through poll and
308 * [di]notify when the value of the file changes.
309 *
310 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
311 * take masks of ancestors with non-empty cpus/mems, instead of being
312 * moved to an ancestor.
313 *
314 * - cpuset: a task can be moved into an empty cpuset, and again it takes
315 * masks of ancestors.
316 *
317 * - blkcg: blk-throttle becomes properly hierarchical.
318 */
319bool cgroup_on_dfl(const struct cgroup *cgrp)
320{
321 return cgrp->root == &cgrp_dfl_root;
322}
323
324/* IDR wrappers which synchronize using cgroup_idr_lock */
325static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
326 gfp_t gfp_mask)
327{
328 int ret;
329
330 idr_preload(gfp_mask);
331 spin_lock_bh(&cgroup_idr_lock);
332 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
333 spin_unlock_bh(&cgroup_idr_lock);
334 idr_preload_end();
335 return ret;
336}
337
338static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
339{
340 void *ret;
341
342 spin_lock_bh(&cgroup_idr_lock);
343 ret = idr_replace(idr, ptr, id);
344 spin_unlock_bh(&cgroup_idr_lock);
345 return ret;
346}
347
348static void cgroup_idr_remove(struct idr *idr, int id)
349{
350 spin_lock_bh(&cgroup_idr_lock);
351 idr_remove(idr, id);
352 spin_unlock_bh(&cgroup_idr_lock);
353}
354
355static bool cgroup_has_tasks(struct cgroup *cgrp)
356{
357 return cgrp->nr_populated_csets;
358}
359
360static bool cgroup_is_threaded(struct cgroup *cgrp)
361{
362 return cgrp->dom_cgrp != cgrp;
363}
364
365/* can @cgrp host both domain and threaded children? */
366static bool cgroup_is_mixable(struct cgroup *cgrp)
367{
368 /*
369 * Root isn't under domain level resource control exempting it from
370 * the no-internal-process constraint, so it can serve as a thread
371 * root and a parent of resource domains at the same time.
372 */
373 return !cgroup_parent(cgrp);
374}
375
376/* can @cgrp become a thread root? Should always be true for a thread root */
377static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
378{
379 /* mixables don't care */
380 if (cgroup_is_mixable(cgrp))
381 return true;
382
383 /* domain roots can't be nested under threaded */
384 if (cgroup_is_threaded(cgrp))
385 return false;
386
387 /* can only have either domain or threaded children */
388 if (cgrp->nr_populated_domain_children)
389 return false;
390
391 /* and no domain controllers can be enabled */
392 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
393 return false;
394
395 return true;
396}
397
398/* is @cgrp root of a threaded subtree? */
399static bool cgroup_is_thread_root(struct cgroup *cgrp)
400{
401 /* thread root should be a domain */
402 if (cgroup_is_threaded(cgrp))
403 return false;
404
405 /* a domain w/ threaded children is a thread root */
406 if (cgrp->nr_threaded_children)
407 return true;
408
409 /*
410 * A domain which has tasks and explicit threaded controllers
411 * enabled is a thread root.
412 */
413 if (cgroup_has_tasks(cgrp) &&
414 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
415 return true;
416
417 return false;
418}
419
420/* a domain which isn't connected to the root w/o brekage can't be used */
421static bool cgroup_is_valid_domain(struct cgroup *cgrp)
422{
423 /* the cgroup itself can be a thread root */
424 if (cgroup_is_threaded(cgrp))
425 return false;
426
427 /* but the ancestors can't be unless mixable */
428 while ((cgrp = cgroup_parent(cgrp))) {
429 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
430 return false;
431 if (cgroup_is_threaded(cgrp))
432 return false;
433 }
434
435 return true;
436}
437
438/* subsystems visibly enabled on a cgroup */
439static u16 cgroup_control(struct cgroup *cgrp)
440{
441 struct cgroup *parent = cgroup_parent(cgrp);
442 u16 root_ss_mask = cgrp->root->subsys_mask;
443
444 if (parent) {
445 u16 ss_mask = parent->subtree_control;
446
447 /* threaded cgroups can only have threaded controllers */
448 if (cgroup_is_threaded(cgrp))
449 ss_mask &= cgrp_dfl_threaded_ss_mask;
450 return ss_mask;
451 }
452
453 if (cgroup_on_dfl(cgrp))
454 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
455 cgrp_dfl_implicit_ss_mask);
456 return root_ss_mask;
457}
458
459/* subsystems enabled on a cgroup */
460static u16 cgroup_ss_mask(struct cgroup *cgrp)
461{
462 struct cgroup *parent = cgroup_parent(cgrp);
463
464 if (parent) {
465 u16 ss_mask = parent->subtree_ss_mask;
466
467 /* threaded cgroups can only have threaded controllers */
468 if (cgroup_is_threaded(cgrp))
469 ss_mask &= cgrp_dfl_threaded_ss_mask;
470 return ss_mask;
471 }
472
473 return cgrp->root->subsys_mask;
474}
475
476/**
477 * cgroup_css - obtain a cgroup's css for the specified subsystem
478 * @cgrp: the cgroup of interest
479 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
480 *
481 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
482 * function must be called either under cgroup_mutex or rcu_read_lock() and
483 * the caller is responsible for pinning the returned css if it wants to
484 * keep accessing it outside the said locks. This function may return
485 * %NULL if @cgrp doesn't have @subsys_id enabled.
486 */
487static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
488 struct cgroup_subsys *ss)
489{
490 if (CGROUP_HAS_SUBSYS_CONFIG && ss)
491 return rcu_dereference_check(cgrp->subsys[ss->id],
492 lockdep_is_held(&cgroup_mutex));
493 else
494 return &cgrp->self;
495}
496
497/**
498 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
499 * @cgrp: the cgroup of interest
500 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
501 *
502 * Similar to cgroup_css() but returns the effective css, which is defined
503 * as the matching css of the nearest ancestor including self which has @ss
504 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
505 * function is guaranteed to return non-NULL css.
506 */
507static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
508 struct cgroup_subsys *ss)
509{
510 lockdep_assert_held(&cgroup_mutex);
511
512 if (!ss)
513 return &cgrp->self;
514
515 /*
516 * This function is used while updating css associations and thus
517 * can't test the csses directly. Test ss_mask.
518 */
519 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
520 cgrp = cgroup_parent(cgrp);
521 if (!cgrp)
522 return NULL;
523 }
524
525 return cgroup_css(cgrp, ss);
526}
527
528/**
529 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
530 * @cgrp: the cgroup of interest
531 * @ss: the subsystem of interest
532 *
533 * Find and get the effective css of @cgrp for @ss. The effective css is
534 * defined as the matching css of the nearest ancestor including self which
535 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
536 * the root css is returned, so this function always returns a valid css.
537 *
538 * The returned css is not guaranteed to be online, and therefore it is the
539 * callers responsibility to try get a reference for it.
540 */
541struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
542 struct cgroup_subsys *ss)
543{
544 struct cgroup_subsys_state *css;
545
546 if (!CGROUP_HAS_SUBSYS_CONFIG)
547 return NULL;
548
549 do {
550 css = cgroup_css(cgrp, ss);
551
552 if (css)
553 return css;
554 cgrp = cgroup_parent(cgrp);
555 } while (cgrp);
556
557 return init_css_set.subsys[ss->id];
558}
559
560/**
561 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
562 * @cgrp: the cgroup of interest
563 * @ss: the subsystem of interest
564 *
565 * Find and get the effective css of @cgrp for @ss. The effective css is
566 * defined as the matching css of the nearest ancestor including self which
567 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
568 * the root css is returned, so this function always returns a valid css.
569 * The returned css must be put using css_put().
570 */
571struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
572 struct cgroup_subsys *ss)
573{
574 struct cgroup_subsys_state *css;
575
576 if (!CGROUP_HAS_SUBSYS_CONFIG)
577 return NULL;
578
579 rcu_read_lock();
580
581 do {
582 css = cgroup_css(cgrp, ss);
583
584 if (css && css_tryget_online(css))
585 goto out_unlock;
586 cgrp = cgroup_parent(cgrp);
587 } while (cgrp);
588
589 css = init_css_set.subsys[ss->id];
590 css_get(css);
591out_unlock:
592 rcu_read_unlock();
593 return css;
594}
595EXPORT_SYMBOL_GPL(cgroup_get_e_css);
596
597static void cgroup_get_live(struct cgroup *cgrp)
598{
599 WARN_ON_ONCE(cgroup_is_dead(cgrp));
600 cgroup_get(cgrp);
601}
602
603/**
604 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
605 * is responsible for taking the css_set_lock.
606 * @cgrp: the cgroup in question
607 */
608int __cgroup_task_count(const struct cgroup *cgrp)
609{
610 int count = 0;
611 struct cgrp_cset_link *link;
612
613 lockdep_assert_held(&css_set_lock);
614
615 list_for_each_entry(link, &cgrp->cset_links, cset_link)
616 count += link->cset->nr_tasks;
617
618 return count;
619}
620
621/**
622 * cgroup_task_count - count the number of tasks in a cgroup.
623 * @cgrp: the cgroup in question
624 */
625int cgroup_task_count(const struct cgroup *cgrp)
626{
627 int count;
628
629 spin_lock_irq(&css_set_lock);
630 count = __cgroup_task_count(cgrp);
631 spin_unlock_irq(&css_set_lock);
632
633 return count;
634}
635
636struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
637{
638 struct cgroup *cgrp = of->kn->parent->priv;
639 struct cftype *cft = of_cft(of);
640
641 /*
642 * This is open and unprotected implementation of cgroup_css().
643 * seq_css() is only called from a kernfs file operation which has
644 * an active reference on the file. Because all the subsystem
645 * files are drained before a css is disassociated with a cgroup,
646 * the matching css from the cgroup's subsys table is guaranteed to
647 * be and stay valid until the enclosing operation is complete.
648 */
649 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
650 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
651 else
652 return &cgrp->self;
653}
654EXPORT_SYMBOL_GPL(of_css);
655
656/**
657 * for_each_css - iterate all css's of a cgroup
658 * @css: the iteration cursor
659 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
660 * @cgrp: the target cgroup to iterate css's of
661 *
662 * Should be called under cgroup_mutex.
663 */
664#define for_each_css(css, ssid, cgrp) \
665 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
666 if (!((css) = rcu_dereference_check( \
667 (cgrp)->subsys[(ssid)], \
668 lockdep_is_held(&cgroup_mutex)))) { } \
669 else
670
671/**
672 * do_each_subsys_mask - filter for_each_subsys with a bitmask
673 * @ss: the iteration cursor
674 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
675 * @ss_mask: the bitmask
676 *
677 * The block will only run for cases where the ssid-th bit (1 << ssid) of
678 * @ss_mask is set.
679 */
680#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
681 unsigned long __ss_mask = (ss_mask); \
682 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
683 (ssid) = 0; \
684 break; \
685 } \
686 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
687 (ss) = cgroup_subsys[ssid]; \
688 {
689
690#define while_each_subsys_mask() \
691 } \
692 } \
693} while (false)
694
695/* iterate over child cgrps, lock should be held throughout iteration */
696#define cgroup_for_each_live_child(child, cgrp) \
697 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
698 if (({ lockdep_assert_held(&cgroup_mutex); \
699 cgroup_is_dead(child); })) \
700 ; \
701 else
702
703/* walk live descendants in pre order */
704#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
705 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
706 if (({ lockdep_assert_held(&cgroup_mutex); \
707 (dsct) = (d_css)->cgroup; \
708 cgroup_is_dead(dsct); })) \
709 ; \
710 else
711
712/* walk live descendants in postorder */
713#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
714 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
715 if (({ lockdep_assert_held(&cgroup_mutex); \
716 (dsct) = (d_css)->cgroup; \
717 cgroup_is_dead(dsct); })) \
718 ; \
719 else
720
721/*
722 * The default css_set - used by init and its children prior to any
723 * hierarchies being mounted. It contains a pointer to the root state
724 * for each subsystem. Also used to anchor the list of css_sets. Not
725 * reference-counted, to improve performance when child cgroups
726 * haven't been created.
727 */
728struct css_set init_css_set = {
729 .refcount = REFCOUNT_INIT(1),
730 .dom_cset = &init_css_set,
731 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
732 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
733 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
734 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
735 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
736 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
737 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
738 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
739 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
740
741 /*
742 * The following field is re-initialized when this cset gets linked
743 * in cgroup_init(). However, let's initialize the field
744 * statically too so that the default cgroup can be accessed safely
745 * early during boot.
746 */
747 .dfl_cgrp = &cgrp_dfl_root.cgrp,
748};
749
750static int css_set_count = 1; /* 1 for init_css_set */
751
752static bool css_set_threaded(struct css_set *cset)
753{
754 return cset->dom_cset != cset;
755}
756
757/**
758 * css_set_populated - does a css_set contain any tasks?
759 * @cset: target css_set
760 *
761 * css_set_populated() should be the same as !!cset->nr_tasks at steady
762 * state. However, css_set_populated() can be called while a task is being
763 * added to or removed from the linked list before the nr_tasks is
764 * properly updated. Hence, we can't just look at ->nr_tasks here.
765 */
766static bool css_set_populated(struct css_set *cset)
767{
768 lockdep_assert_held(&css_set_lock);
769
770 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
771}
772
773/**
774 * cgroup_update_populated - update the populated count of a cgroup
775 * @cgrp: the target cgroup
776 * @populated: inc or dec populated count
777 *
778 * One of the css_sets associated with @cgrp is either getting its first
779 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
780 * count is propagated towards root so that a given cgroup's
781 * nr_populated_children is zero iff none of its descendants contain any
782 * tasks.
783 *
784 * @cgrp's interface file "cgroup.populated" is zero if both
785 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
786 * 1 otherwise. When the sum changes from or to zero, userland is notified
787 * that the content of the interface file has changed. This can be used to
788 * detect when @cgrp and its descendants become populated or empty.
789 */
790static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
791{
792 struct cgroup *child = NULL;
793 int adj = populated ? 1 : -1;
794
795 lockdep_assert_held(&css_set_lock);
796
797 do {
798 bool was_populated = cgroup_is_populated(cgrp);
799
800 if (!child) {
801 cgrp->nr_populated_csets += adj;
802 } else {
803 if (cgroup_is_threaded(child))
804 cgrp->nr_populated_threaded_children += adj;
805 else
806 cgrp->nr_populated_domain_children += adj;
807 }
808
809 if (was_populated == cgroup_is_populated(cgrp))
810 break;
811
812 cgroup1_check_for_release(cgrp);
813 TRACE_CGROUP_PATH(notify_populated, cgrp,
814 cgroup_is_populated(cgrp));
815 cgroup_file_notify(&cgrp->events_file);
816
817 child = cgrp;
818 cgrp = cgroup_parent(cgrp);
819 } while (cgrp);
820}
821
822/**
823 * css_set_update_populated - update populated state of a css_set
824 * @cset: target css_set
825 * @populated: whether @cset is populated or depopulated
826 *
827 * @cset is either getting the first task or losing the last. Update the
828 * populated counters of all associated cgroups accordingly.
829 */
830static void css_set_update_populated(struct css_set *cset, bool populated)
831{
832 struct cgrp_cset_link *link;
833
834 lockdep_assert_held(&css_set_lock);
835
836 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
837 cgroup_update_populated(link->cgrp, populated);
838}
839
840/*
841 * @task is leaving, advance task iterators which are pointing to it so
842 * that they can resume at the next position. Advancing an iterator might
843 * remove it from the list, use safe walk. See css_task_iter_skip() for
844 * details.
845 */
846static void css_set_skip_task_iters(struct css_set *cset,
847 struct task_struct *task)
848{
849 struct css_task_iter *it, *pos;
850
851 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
852 css_task_iter_skip(it, task);
853}
854
855/**
856 * css_set_move_task - move a task from one css_set to another
857 * @task: task being moved
858 * @from_cset: css_set @task currently belongs to (may be NULL)
859 * @to_cset: new css_set @task is being moved to (may be NULL)
860 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
861 *
862 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
863 * css_set, @from_cset can be NULL. If @task is being disassociated
864 * instead of moved, @to_cset can be NULL.
865 *
866 * This function automatically handles populated counter updates and
867 * css_task_iter adjustments but the caller is responsible for managing
868 * @from_cset and @to_cset's reference counts.
869 */
870static void css_set_move_task(struct task_struct *task,
871 struct css_set *from_cset, struct css_set *to_cset,
872 bool use_mg_tasks)
873{
874 lockdep_assert_held(&css_set_lock);
875
876 if (to_cset && !css_set_populated(to_cset))
877 css_set_update_populated(to_cset, true);
878
879 if (from_cset) {
880 WARN_ON_ONCE(list_empty(&task->cg_list));
881
882 css_set_skip_task_iters(from_cset, task);
883 list_del_init(&task->cg_list);
884 if (!css_set_populated(from_cset))
885 css_set_update_populated(from_cset, false);
886 } else {
887 WARN_ON_ONCE(!list_empty(&task->cg_list));
888 }
889
890 if (to_cset) {
891 /*
892 * We are synchronized through cgroup_threadgroup_rwsem
893 * against PF_EXITING setting such that we can't race
894 * against cgroup_exit()/cgroup_free() dropping the css_set.
895 */
896 WARN_ON_ONCE(task->flags & PF_EXITING);
897
898 cgroup_move_task(task, to_cset);
899 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
900 &to_cset->tasks);
901 }
902}
903
904/*
905 * hash table for cgroup groups. This improves the performance to find
906 * an existing css_set. This hash doesn't (currently) take into
907 * account cgroups in empty hierarchies.
908 */
909#define CSS_SET_HASH_BITS 7
910static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
911
912static unsigned long css_set_hash(struct cgroup_subsys_state **css)
913{
914 unsigned long key = 0UL;
915 struct cgroup_subsys *ss;
916 int i;
917
918 for_each_subsys(ss, i)
919 key += (unsigned long)css[i];
920 key = (key >> 16) ^ key;
921
922 return key;
923}
924
925void put_css_set_locked(struct css_set *cset)
926{
927 struct cgrp_cset_link *link, *tmp_link;
928 struct cgroup_subsys *ss;
929 int ssid;
930
931 lockdep_assert_held(&css_set_lock);
932
933 if (!refcount_dec_and_test(&cset->refcount))
934 return;
935
936 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
937
938 /* This css_set is dead. Unlink it and release cgroup and css refs */
939 for_each_subsys(ss, ssid) {
940 list_del(&cset->e_cset_node[ssid]);
941 css_put(cset->subsys[ssid]);
942 }
943 hash_del(&cset->hlist);
944 css_set_count--;
945
946 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
947 list_del(&link->cset_link);
948 list_del(&link->cgrp_link);
949 if (cgroup_parent(link->cgrp))
950 cgroup_put(link->cgrp);
951 kfree(link);
952 }
953
954 if (css_set_threaded(cset)) {
955 list_del(&cset->threaded_csets_node);
956 put_css_set_locked(cset->dom_cset);
957 }
958
959 kfree_rcu(cset, rcu_head);
960}
961
962/**
963 * compare_css_sets - helper function for find_existing_css_set().
964 * @cset: candidate css_set being tested
965 * @old_cset: existing css_set for a task
966 * @new_cgrp: cgroup that's being entered by the task
967 * @template: desired set of css pointers in css_set (pre-calculated)
968 *
969 * Returns true if "cset" matches "old_cset" except for the hierarchy
970 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
971 */
972static bool compare_css_sets(struct css_set *cset,
973 struct css_set *old_cset,
974 struct cgroup *new_cgrp,
975 struct cgroup_subsys_state *template[])
976{
977 struct cgroup *new_dfl_cgrp;
978 struct list_head *l1, *l2;
979
980 /*
981 * On the default hierarchy, there can be csets which are
982 * associated with the same set of cgroups but different csses.
983 * Let's first ensure that csses match.
984 */
985 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
986 return false;
987
988
989 /* @cset's domain should match the default cgroup's */
990 if (cgroup_on_dfl(new_cgrp))
991 new_dfl_cgrp = new_cgrp;
992 else
993 new_dfl_cgrp = old_cset->dfl_cgrp;
994
995 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
996 return false;
997
998 /*
999 * Compare cgroup pointers in order to distinguish between
1000 * different cgroups in hierarchies. As different cgroups may
1001 * share the same effective css, this comparison is always
1002 * necessary.
1003 */
1004 l1 = &cset->cgrp_links;
1005 l2 = &old_cset->cgrp_links;
1006 while (1) {
1007 struct cgrp_cset_link *link1, *link2;
1008 struct cgroup *cgrp1, *cgrp2;
1009
1010 l1 = l1->next;
1011 l2 = l2->next;
1012 /* See if we reached the end - both lists are equal length. */
1013 if (l1 == &cset->cgrp_links) {
1014 BUG_ON(l2 != &old_cset->cgrp_links);
1015 break;
1016 } else {
1017 BUG_ON(l2 == &old_cset->cgrp_links);
1018 }
1019 /* Locate the cgroups associated with these links. */
1020 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1021 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1022 cgrp1 = link1->cgrp;
1023 cgrp2 = link2->cgrp;
1024 /* Hierarchies should be linked in the same order. */
1025 BUG_ON(cgrp1->root != cgrp2->root);
1026
1027 /*
1028 * If this hierarchy is the hierarchy of the cgroup
1029 * that's changing, then we need to check that this
1030 * css_set points to the new cgroup; if it's any other
1031 * hierarchy, then this css_set should point to the
1032 * same cgroup as the old css_set.
1033 */
1034 if (cgrp1->root == new_cgrp->root) {
1035 if (cgrp1 != new_cgrp)
1036 return false;
1037 } else {
1038 if (cgrp1 != cgrp2)
1039 return false;
1040 }
1041 }
1042 return true;
1043}
1044
1045/**
1046 * find_existing_css_set - init css array and find the matching css_set
1047 * @old_cset: the css_set that we're using before the cgroup transition
1048 * @cgrp: the cgroup that we're moving into
1049 * @template: out param for the new set of csses, should be clear on entry
1050 */
1051static struct css_set *find_existing_css_set(struct css_set *old_cset,
1052 struct cgroup *cgrp,
1053 struct cgroup_subsys_state **template)
1054{
1055 struct cgroup_root *root = cgrp->root;
1056 struct cgroup_subsys *ss;
1057 struct css_set *cset;
1058 unsigned long key;
1059 int i;
1060
1061 /*
1062 * Build the set of subsystem state objects that we want to see in the
1063 * new css_set. While subsystems can change globally, the entries here
1064 * won't change, so no need for locking.
1065 */
1066 for_each_subsys(ss, i) {
1067 if (root->subsys_mask & (1UL << i)) {
1068 /*
1069 * @ss is in this hierarchy, so we want the
1070 * effective css from @cgrp.
1071 */
1072 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1073 } else {
1074 /*
1075 * @ss is not in this hierarchy, so we don't want
1076 * to change the css.
1077 */
1078 template[i] = old_cset->subsys[i];
1079 }
1080 }
1081
1082 key = css_set_hash(template);
1083 hash_for_each_possible(css_set_table, cset, hlist, key) {
1084 if (!compare_css_sets(cset, old_cset, cgrp, template))
1085 continue;
1086
1087 /* This css_set matches what we need */
1088 return cset;
1089 }
1090
1091 /* No existing cgroup group matched */
1092 return NULL;
1093}
1094
1095static void free_cgrp_cset_links(struct list_head *links_to_free)
1096{
1097 struct cgrp_cset_link *link, *tmp_link;
1098
1099 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1100 list_del(&link->cset_link);
1101 kfree(link);
1102 }
1103}
1104
1105/**
1106 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1107 * @count: the number of links to allocate
1108 * @tmp_links: list_head the allocated links are put on
1109 *
1110 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1111 * through ->cset_link. Returns 0 on success or -errno.
1112 */
1113static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1114{
1115 struct cgrp_cset_link *link;
1116 int i;
1117
1118 INIT_LIST_HEAD(tmp_links);
1119
1120 for (i = 0; i < count; i++) {
1121 link = kzalloc(sizeof(*link), GFP_KERNEL);
1122 if (!link) {
1123 free_cgrp_cset_links(tmp_links);
1124 return -ENOMEM;
1125 }
1126 list_add(&link->cset_link, tmp_links);
1127 }
1128 return 0;
1129}
1130
1131/**
1132 * link_css_set - a helper function to link a css_set to a cgroup
1133 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1134 * @cset: the css_set to be linked
1135 * @cgrp: the destination cgroup
1136 */
1137static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1138 struct cgroup *cgrp)
1139{
1140 struct cgrp_cset_link *link;
1141
1142 BUG_ON(list_empty(tmp_links));
1143
1144 if (cgroup_on_dfl(cgrp))
1145 cset->dfl_cgrp = cgrp;
1146
1147 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1148 link->cset = cset;
1149 link->cgrp = cgrp;
1150
1151 /*
1152 * Always add links to the tail of the lists so that the lists are
1153 * in chronological order.
1154 */
1155 list_move_tail(&link->cset_link, &cgrp->cset_links);
1156 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1157
1158 if (cgroup_parent(cgrp))
1159 cgroup_get_live(cgrp);
1160}
1161
1162/**
1163 * find_css_set - return a new css_set with one cgroup updated
1164 * @old_cset: the baseline css_set
1165 * @cgrp: the cgroup to be updated
1166 *
1167 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1168 * substituted into the appropriate hierarchy.
1169 */
1170static struct css_set *find_css_set(struct css_set *old_cset,
1171 struct cgroup *cgrp)
1172{
1173 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1174 struct css_set *cset;
1175 struct list_head tmp_links;
1176 struct cgrp_cset_link *link;
1177 struct cgroup_subsys *ss;
1178 unsigned long key;
1179 int ssid;
1180
1181 lockdep_assert_held(&cgroup_mutex);
1182
1183 /* First see if we already have a cgroup group that matches
1184 * the desired set */
1185 spin_lock_irq(&css_set_lock);
1186 cset = find_existing_css_set(old_cset, cgrp, template);
1187 if (cset)
1188 get_css_set(cset);
1189 spin_unlock_irq(&css_set_lock);
1190
1191 if (cset)
1192 return cset;
1193
1194 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1195 if (!cset)
1196 return NULL;
1197
1198 /* Allocate all the cgrp_cset_link objects that we'll need */
1199 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1200 kfree(cset);
1201 return NULL;
1202 }
1203
1204 refcount_set(&cset->refcount, 1);
1205 cset->dom_cset = cset;
1206 INIT_LIST_HEAD(&cset->tasks);
1207 INIT_LIST_HEAD(&cset->mg_tasks);
1208 INIT_LIST_HEAD(&cset->dying_tasks);
1209 INIT_LIST_HEAD(&cset->task_iters);
1210 INIT_LIST_HEAD(&cset->threaded_csets);
1211 INIT_HLIST_NODE(&cset->hlist);
1212 INIT_LIST_HEAD(&cset->cgrp_links);
1213 INIT_LIST_HEAD(&cset->mg_src_preload_node);
1214 INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1215 INIT_LIST_HEAD(&cset->mg_node);
1216
1217 /* Copy the set of subsystem state objects generated in
1218 * find_existing_css_set() */
1219 memcpy(cset->subsys, template, sizeof(cset->subsys));
1220
1221 spin_lock_irq(&css_set_lock);
1222 /* Add reference counts and links from the new css_set. */
1223 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1224 struct cgroup *c = link->cgrp;
1225
1226 if (c->root == cgrp->root)
1227 c = cgrp;
1228 link_css_set(&tmp_links, cset, c);
1229 }
1230
1231 BUG_ON(!list_empty(&tmp_links));
1232
1233 css_set_count++;
1234
1235 /* Add @cset to the hash table */
1236 key = css_set_hash(cset->subsys);
1237 hash_add(css_set_table, &cset->hlist, key);
1238
1239 for_each_subsys(ss, ssid) {
1240 struct cgroup_subsys_state *css = cset->subsys[ssid];
1241
1242 list_add_tail(&cset->e_cset_node[ssid],
1243 &css->cgroup->e_csets[ssid]);
1244 css_get(css);
1245 }
1246
1247 spin_unlock_irq(&css_set_lock);
1248
1249 /*
1250 * If @cset should be threaded, look up the matching dom_cset and
1251 * link them up. We first fully initialize @cset then look for the
1252 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1253 * to stay empty until we return.
1254 */
1255 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1256 struct css_set *dcset;
1257
1258 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1259 if (!dcset) {
1260 put_css_set(cset);
1261 return NULL;
1262 }
1263
1264 spin_lock_irq(&css_set_lock);
1265 cset->dom_cset = dcset;
1266 list_add_tail(&cset->threaded_csets_node,
1267 &dcset->threaded_csets);
1268 spin_unlock_irq(&css_set_lock);
1269 }
1270
1271 return cset;
1272}
1273
1274struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1275{
1276 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1277
1278 return root_cgrp->root;
1279}
1280
1281void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1282{
1283 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1284
1285 /* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
1286 if (favor && !favoring) {
1287 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1288 root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1289 } else if (!favor && favoring) {
1290 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1291 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1292 }
1293}
1294
1295static int cgroup_init_root_id(struct cgroup_root *root)
1296{
1297 int id;
1298
1299 lockdep_assert_held(&cgroup_mutex);
1300
1301 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1302 if (id < 0)
1303 return id;
1304
1305 root->hierarchy_id = id;
1306 return 0;
1307}
1308
1309static void cgroup_exit_root_id(struct cgroup_root *root)
1310{
1311 lockdep_assert_held(&cgroup_mutex);
1312
1313 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1314}
1315
1316void cgroup_free_root(struct cgroup_root *root)
1317{
1318 kfree_rcu(root, rcu);
1319}
1320
1321static void cgroup_destroy_root(struct cgroup_root *root)
1322{
1323 struct cgroup *cgrp = &root->cgrp;
1324 struct cgrp_cset_link *link, *tmp_link;
1325
1326 trace_cgroup_destroy_root(root);
1327
1328 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1329
1330 BUG_ON(atomic_read(&root->nr_cgrps));
1331 BUG_ON(!list_empty(&cgrp->self.children));
1332
1333 /* Rebind all subsystems back to the default hierarchy */
1334 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1335
1336 /*
1337 * Release all the links from cset_links to this hierarchy's
1338 * root cgroup
1339 */
1340 spin_lock_irq(&css_set_lock);
1341
1342 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1343 list_del(&link->cset_link);
1344 list_del(&link->cgrp_link);
1345 kfree(link);
1346 }
1347
1348 spin_unlock_irq(&css_set_lock);
1349
1350 WARN_ON_ONCE(list_empty(&root->root_list));
1351 list_del_rcu(&root->root_list);
1352 cgroup_root_count--;
1353
1354 if (!have_favordynmods)
1355 cgroup_favor_dynmods(root, false);
1356
1357 cgroup_exit_root_id(root);
1358
1359 cgroup_unlock();
1360
1361 cgroup_rstat_exit(cgrp);
1362 kernfs_destroy_root(root->kf_root);
1363 cgroup_free_root(root);
1364}
1365
1366/*
1367 * Returned cgroup is without refcount but it's valid as long as cset pins it.
1368 */
1369static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1370 struct cgroup_root *root)
1371{
1372 struct cgroup *res_cgroup = NULL;
1373
1374 if (cset == &init_css_set) {
1375 res_cgroup = &root->cgrp;
1376 } else if (root == &cgrp_dfl_root) {
1377 res_cgroup = cset->dfl_cgrp;
1378 } else {
1379 struct cgrp_cset_link *link;
1380 lockdep_assert_held(&css_set_lock);
1381
1382 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1383 struct cgroup *c = link->cgrp;
1384
1385 if (c->root == root) {
1386 res_cgroup = c;
1387 break;
1388 }
1389 }
1390 }
1391
1392 /*
1393 * If cgroup_mutex is not held, the cgrp_cset_link will be freed
1394 * before we remove the cgroup root from the root_list. Consequently,
1395 * when accessing a cgroup root, the cset_link may have already been
1396 * freed, resulting in a NULL res_cgroup. However, by holding the
1397 * cgroup_mutex, we ensure that res_cgroup can't be NULL.
1398 * If we don't hold cgroup_mutex in the caller, we must do the NULL
1399 * check.
1400 */
1401 return res_cgroup;
1402}
1403
1404/*
1405 * look up cgroup associated with current task's cgroup namespace on the
1406 * specified hierarchy
1407 */
1408static struct cgroup *
1409current_cgns_cgroup_from_root(struct cgroup_root *root)
1410{
1411 struct cgroup *res = NULL;
1412 struct css_set *cset;
1413
1414 lockdep_assert_held(&css_set_lock);
1415
1416 rcu_read_lock();
1417
1418 cset = current->nsproxy->cgroup_ns->root_cset;
1419 res = __cset_cgroup_from_root(cset, root);
1420
1421 rcu_read_unlock();
1422
1423 /*
1424 * The namespace_sem is held by current, so the root cgroup can't
1425 * be umounted. Therefore, we can ensure that the res is non-NULL.
1426 */
1427 WARN_ON_ONCE(!res);
1428 return res;
1429}
1430
1431/*
1432 * Look up cgroup associated with current task's cgroup namespace on the default
1433 * hierarchy.
1434 *
1435 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1436 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1437 * pointers.
1438 * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1439 * - As a bonus returned cgrp is pinned with the current because it cannot
1440 * switch cgroup_ns asynchronously.
1441 */
1442static struct cgroup *current_cgns_cgroup_dfl(void)
1443{
1444 struct css_set *cset;
1445
1446 if (current->nsproxy) {
1447 cset = current->nsproxy->cgroup_ns->root_cset;
1448 return __cset_cgroup_from_root(cset, &cgrp_dfl_root);
1449 } else {
1450 /*
1451 * NOTE: This function may be called from bpf_cgroup_from_id()
1452 * on a task which has already passed exit_task_namespaces() and
1453 * nsproxy == NULL. Fall back to cgrp_dfl_root which will make all
1454 * cgroups visible for lookups.
1455 */
1456 return &cgrp_dfl_root.cgrp;
1457 }
1458}
1459
1460/* look up cgroup associated with given css_set on the specified hierarchy */
1461static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1462 struct cgroup_root *root)
1463{
1464 lockdep_assert_held(&css_set_lock);
1465
1466 return __cset_cgroup_from_root(cset, root);
1467}
1468
1469/*
1470 * Return the cgroup for "task" from the given hierarchy. Must be
1471 * called with css_set_lock held to prevent task's groups from being modified.
1472 * Must be called with either cgroup_mutex or rcu read lock to prevent the
1473 * cgroup root from being destroyed.
1474 */
1475struct cgroup *task_cgroup_from_root(struct task_struct *task,
1476 struct cgroup_root *root)
1477{
1478 /*
1479 * No need to lock the task - since we hold css_set_lock the
1480 * task can't change groups.
1481 */
1482 return cset_cgroup_from_root(task_css_set(task), root);
1483}
1484
1485/*
1486 * A task must hold cgroup_mutex to modify cgroups.
1487 *
1488 * Any task can increment and decrement the count field without lock.
1489 * So in general, code holding cgroup_mutex can't rely on the count
1490 * field not changing. However, if the count goes to zero, then only
1491 * cgroup_attach_task() can increment it again. Because a count of zero
1492 * means that no tasks are currently attached, therefore there is no
1493 * way a task attached to that cgroup can fork (the other way to
1494 * increment the count). So code holding cgroup_mutex can safely
1495 * assume that if the count is zero, it will stay zero. Similarly, if
1496 * a task holds cgroup_mutex on a cgroup with zero count, it
1497 * knows that the cgroup won't be removed, as cgroup_rmdir()
1498 * needs that mutex.
1499 *
1500 * A cgroup can only be deleted if both its 'count' of using tasks
1501 * is zero, and its list of 'children' cgroups is empty. Since all
1502 * tasks in the system use _some_ cgroup, and since there is always at
1503 * least one task in the system (init, pid == 1), therefore, root cgroup
1504 * always has either children cgroups and/or using tasks. So we don't
1505 * need a special hack to ensure that root cgroup cannot be deleted.
1506 *
1507 * P.S. One more locking exception. RCU is used to guard the
1508 * update of a tasks cgroup pointer by cgroup_attach_task()
1509 */
1510
1511static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1512
1513static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1514 char *buf)
1515{
1516 struct cgroup_subsys *ss = cft->ss;
1517
1518 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1519 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1520 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1521
1522 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1523 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1524 cft->name);
1525 } else {
1526 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1527 }
1528 return buf;
1529}
1530
1531/**
1532 * cgroup_file_mode - deduce file mode of a control file
1533 * @cft: the control file in question
1534 *
1535 * S_IRUGO for read, S_IWUSR for write.
1536 */
1537static umode_t cgroup_file_mode(const struct cftype *cft)
1538{
1539 umode_t mode = 0;
1540
1541 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1542 mode |= S_IRUGO;
1543
1544 if (cft->write_u64 || cft->write_s64 || cft->write) {
1545 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1546 mode |= S_IWUGO;
1547 else
1548 mode |= S_IWUSR;
1549 }
1550
1551 return mode;
1552}
1553
1554/**
1555 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1556 * @subtree_control: the new subtree_control mask to consider
1557 * @this_ss_mask: available subsystems
1558 *
1559 * On the default hierarchy, a subsystem may request other subsystems to be
1560 * enabled together through its ->depends_on mask. In such cases, more
1561 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1562 *
1563 * This function calculates which subsystems need to be enabled if
1564 * @subtree_control is to be applied while restricted to @this_ss_mask.
1565 */
1566static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1567{
1568 u16 cur_ss_mask = subtree_control;
1569 struct cgroup_subsys *ss;
1570 int ssid;
1571
1572 lockdep_assert_held(&cgroup_mutex);
1573
1574 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1575
1576 while (true) {
1577 u16 new_ss_mask = cur_ss_mask;
1578
1579 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1580 new_ss_mask |= ss->depends_on;
1581 } while_each_subsys_mask();
1582
1583 /*
1584 * Mask out subsystems which aren't available. This can
1585 * happen only if some depended-upon subsystems were bound
1586 * to non-default hierarchies.
1587 */
1588 new_ss_mask &= this_ss_mask;
1589
1590 if (new_ss_mask == cur_ss_mask)
1591 break;
1592 cur_ss_mask = new_ss_mask;
1593 }
1594
1595 return cur_ss_mask;
1596}
1597
1598/**
1599 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1600 * @kn: the kernfs_node being serviced
1601 *
1602 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1603 * the method finishes if locking succeeded. Note that once this function
1604 * returns the cgroup returned by cgroup_kn_lock_live() may become
1605 * inaccessible any time. If the caller intends to continue to access the
1606 * cgroup, it should pin it before invoking this function.
1607 */
1608void cgroup_kn_unlock(struct kernfs_node *kn)
1609{
1610 struct cgroup *cgrp;
1611
1612 if (kernfs_type(kn) == KERNFS_DIR)
1613 cgrp = kn->priv;
1614 else
1615 cgrp = kn->parent->priv;
1616
1617 cgroup_unlock();
1618
1619 kernfs_unbreak_active_protection(kn);
1620 cgroup_put(cgrp);
1621}
1622
1623/**
1624 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1625 * @kn: the kernfs_node being serviced
1626 * @drain_offline: perform offline draining on the cgroup
1627 *
1628 * This helper is to be used by a cgroup kernfs method currently servicing
1629 * @kn. It breaks the active protection, performs cgroup locking and
1630 * verifies that the associated cgroup is alive. Returns the cgroup if
1631 * alive; otherwise, %NULL. A successful return should be undone by a
1632 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1633 * cgroup is drained of offlining csses before return.
1634 *
1635 * Any cgroup kernfs method implementation which requires locking the
1636 * associated cgroup should use this helper. It avoids nesting cgroup
1637 * locking under kernfs active protection and allows all kernfs operations
1638 * including self-removal.
1639 */
1640struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1641{
1642 struct cgroup *cgrp;
1643
1644 if (kernfs_type(kn) == KERNFS_DIR)
1645 cgrp = kn->priv;
1646 else
1647 cgrp = kn->parent->priv;
1648
1649 /*
1650 * We're gonna grab cgroup_mutex which nests outside kernfs
1651 * active_ref. cgroup liveliness check alone provides enough
1652 * protection against removal. Ensure @cgrp stays accessible and
1653 * break the active_ref protection.
1654 */
1655 if (!cgroup_tryget(cgrp))
1656 return NULL;
1657 kernfs_break_active_protection(kn);
1658
1659 if (drain_offline)
1660 cgroup_lock_and_drain_offline(cgrp);
1661 else
1662 cgroup_lock();
1663
1664 if (!cgroup_is_dead(cgrp))
1665 return cgrp;
1666
1667 cgroup_kn_unlock(kn);
1668 return NULL;
1669}
1670
1671static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1672{
1673 char name[CGROUP_FILE_NAME_MAX];
1674
1675 lockdep_assert_held(&cgroup_mutex);
1676
1677 if (cft->file_offset) {
1678 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1679 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1680
1681 spin_lock_irq(&cgroup_file_kn_lock);
1682 cfile->kn = NULL;
1683 spin_unlock_irq(&cgroup_file_kn_lock);
1684
1685 del_timer_sync(&cfile->notify_timer);
1686 }
1687
1688 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1689}
1690
1691/**
1692 * css_clear_dir - remove subsys files in a cgroup directory
1693 * @css: target css
1694 */
1695static void css_clear_dir(struct cgroup_subsys_state *css)
1696{
1697 struct cgroup *cgrp = css->cgroup;
1698 struct cftype *cfts;
1699
1700 if (!(css->flags & CSS_VISIBLE))
1701 return;
1702
1703 css->flags &= ~CSS_VISIBLE;
1704
1705 if (!css->ss) {
1706 if (cgroup_on_dfl(cgrp)) {
1707 cgroup_addrm_files(css, cgrp,
1708 cgroup_base_files, false);
1709 if (cgroup_psi_enabled())
1710 cgroup_addrm_files(css, cgrp,
1711 cgroup_psi_files, false);
1712 } else {
1713 cgroup_addrm_files(css, cgrp,
1714 cgroup1_base_files, false);
1715 }
1716 } else {
1717 list_for_each_entry(cfts, &css->ss->cfts, node)
1718 cgroup_addrm_files(css, cgrp, cfts, false);
1719 }
1720}
1721
1722/**
1723 * css_populate_dir - create subsys files in a cgroup directory
1724 * @css: target css
1725 *
1726 * On failure, no file is added.
1727 */
1728static int css_populate_dir(struct cgroup_subsys_state *css)
1729{
1730 struct cgroup *cgrp = css->cgroup;
1731 struct cftype *cfts, *failed_cfts;
1732 int ret;
1733
1734 if (css->flags & CSS_VISIBLE)
1735 return 0;
1736
1737 if (!css->ss) {
1738 if (cgroup_on_dfl(cgrp)) {
1739 ret = cgroup_addrm_files(css, cgrp,
1740 cgroup_base_files, true);
1741 if (ret < 0)
1742 return ret;
1743
1744 if (cgroup_psi_enabled()) {
1745 ret = cgroup_addrm_files(css, cgrp,
1746 cgroup_psi_files, true);
1747 if (ret < 0) {
1748 cgroup_addrm_files(css, cgrp,
1749 cgroup_base_files, false);
1750 return ret;
1751 }
1752 }
1753 } else {
1754 ret = cgroup_addrm_files(css, cgrp,
1755 cgroup1_base_files, true);
1756 if (ret < 0)
1757 return ret;
1758 }
1759 } else {
1760 list_for_each_entry(cfts, &css->ss->cfts, node) {
1761 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1762 if (ret < 0) {
1763 failed_cfts = cfts;
1764 goto err;
1765 }
1766 }
1767 }
1768
1769 css->flags |= CSS_VISIBLE;
1770
1771 return 0;
1772err:
1773 list_for_each_entry(cfts, &css->ss->cfts, node) {
1774 if (cfts == failed_cfts)
1775 break;
1776 cgroup_addrm_files(css, cgrp, cfts, false);
1777 }
1778 return ret;
1779}
1780
1781int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1782{
1783 struct cgroup *dcgrp = &dst_root->cgrp;
1784 struct cgroup_subsys *ss;
1785 int ssid, ret;
1786 u16 dfl_disable_ss_mask = 0;
1787
1788 lockdep_assert_held(&cgroup_mutex);
1789
1790 do_each_subsys_mask(ss, ssid, ss_mask) {
1791 /*
1792 * If @ss has non-root csses attached to it, can't move.
1793 * If @ss is an implicit controller, it is exempt from this
1794 * rule and can be stolen.
1795 */
1796 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1797 !ss->implicit_on_dfl)
1798 return -EBUSY;
1799
1800 /* can't move between two non-dummy roots either */
1801 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1802 return -EBUSY;
1803
1804 /*
1805 * Collect ssid's that need to be disabled from default
1806 * hierarchy.
1807 */
1808 if (ss->root == &cgrp_dfl_root)
1809 dfl_disable_ss_mask |= 1 << ssid;
1810
1811 } while_each_subsys_mask();
1812
1813 if (dfl_disable_ss_mask) {
1814 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1815
1816 /*
1817 * Controllers from default hierarchy that need to be rebound
1818 * are all disabled together in one go.
1819 */
1820 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1821 WARN_ON(cgroup_apply_control(scgrp));
1822 cgroup_finalize_control(scgrp, 0);
1823 }
1824
1825 do_each_subsys_mask(ss, ssid, ss_mask) {
1826 struct cgroup_root *src_root = ss->root;
1827 struct cgroup *scgrp = &src_root->cgrp;
1828 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1829 struct css_set *cset, *cset_pos;
1830 struct css_task_iter *it;
1831
1832 WARN_ON(!css || cgroup_css(dcgrp, ss));
1833
1834 if (src_root != &cgrp_dfl_root) {
1835 /* disable from the source */
1836 src_root->subsys_mask &= ~(1 << ssid);
1837 WARN_ON(cgroup_apply_control(scgrp));
1838 cgroup_finalize_control(scgrp, 0);
1839 }
1840
1841 /* rebind */
1842 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1843 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1844 ss->root = dst_root;
1845
1846 spin_lock_irq(&css_set_lock);
1847 css->cgroup = dcgrp;
1848 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id]));
1849 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id],
1850 e_cset_node[ss->id]) {
1851 list_move_tail(&cset->e_cset_node[ss->id],
1852 &dcgrp->e_csets[ss->id]);
1853 /*
1854 * all css_sets of scgrp together in same order to dcgrp,
1855 * patch in-flight iterators to preserve correct iteration.
1856 * since the iterator is always advanced right away and
1857 * finished when it->cset_pos meets it->cset_head, so only
1858 * update it->cset_head is enough here.
1859 */
1860 list_for_each_entry(it, &cset->task_iters, iters_node)
1861 if (it->cset_head == &scgrp->e_csets[ss->id])
1862 it->cset_head = &dcgrp->e_csets[ss->id];
1863 }
1864 spin_unlock_irq(&css_set_lock);
1865
1866 if (ss->css_rstat_flush) {
1867 list_del_rcu(&css->rstat_css_node);
1868 synchronize_rcu();
1869 list_add_rcu(&css->rstat_css_node,
1870 &dcgrp->rstat_css_list);
1871 }
1872
1873 /* default hierarchy doesn't enable controllers by default */
1874 dst_root->subsys_mask |= 1 << ssid;
1875 if (dst_root == &cgrp_dfl_root) {
1876 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1877 } else {
1878 dcgrp->subtree_control |= 1 << ssid;
1879 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1880 }
1881
1882 ret = cgroup_apply_control(dcgrp);
1883 if (ret)
1884 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1885 ss->name, ret);
1886
1887 if (ss->bind)
1888 ss->bind(css);
1889 } while_each_subsys_mask();
1890
1891 kernfs_activate(dcgrp->kn);
1892 return 0;
1893}
1894
1895int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1896 struct kernfs_root *kf_root)
1897{
1898 int len = 0;
1899 char *buf = NULL;
1900 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1901 struct cgroup *ns_cgroup;
1902
1903 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1904 if (!buf)
1905 return -ENOMEM;
1906
1907 spin_lock_irq(&css_set_lock);
1908 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1909 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1910 spin_unlock_irq(&css_set_lock);
1911
1912 if (len == -E2BIG)
1913 len = -ERANGE;
1914 else if (len > 0) {
1915 seq_escape(sf, buf, " \t\n\\");
1916 len = 0;
1917 }
1918 kfree(buf);
1919 return len;
1920}
1921
1922enum cgroup2_param {
1923 Opt_nsdelegate,
1924 Opt_favordynmods,
1925 Opt_memory_localevents,
1926 Opt_memory_recursiveprot,
1927 Opt_memory_hugetlb_accounting,
1928 Opt_pids_localevents,
1929 nr__cgroup2_params
1930};
1931
1932static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1933 fsparam_flag("nsdelegate", Opt_nsdelegate),
1934 fsparam_flag("favordynmods", Opt_favordynmods),
1935 fsparam_flag("memory_localevents", Opt_memory_localevents),
1936 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1937 fsparam_flag("memory_hugetlb_accounting", Opt_memory_hugetlb_accounting),
1938 fsparam_flag("pids_localevents", Opt_pids_localevents),
1939 {}
1940};
1941
1942static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1943{
1944 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1945 struct fs_parse_result result;
1946 int opt;
1947
1948 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1949 if (opt < 0)
1950 return opt;
1951
1952 switch (opt) {
1953 case Opt_nsdelegate:
1954 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1955 return 0;
1956 case Opt_favordynmods:
1957 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1958 return 0;
1959 case Opt_memory_localevents:
1960 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1961 return 0;
1962 case Opt_memory_recursiveprot:
1963 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1964 return 0;
1965 case Opt_memory_hugetlb_accounting:
1966 ctx->flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1967 return 0;
1968 case Opt_pids_localevents:
1969 ctx->flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS;
1970 return 0;
1971 }
1972 return -EINVAL;
1973}
1974
1975struct cgroup_of_peak *of_peak(struct kernfs_open_file *of)
1976{
1977 struct cgroup_file_ctx *ctx = of->priv;
1978
1979 return &ctx->peak;
1980}
1981
1982static void apply_cgroup_root_flags(unsigned int root_flags)
1983{
1984 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1985 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1986 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1987 else
1988 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1989
1990 cgroup_favor_dynmods(&cgrp_dfl_root,
1991 root_flags & CGRP_ROOT_FAVOR_DYNMODS);
1992
1993 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1994 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1995 else
1996 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1997
1998 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1999 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2000 else
2001 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
2002
2003 if (root_flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2004 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2005 else
2006 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
2007
2008 if (root_flags & CGRP_ROOT_PIDS_LOCAL_EVENTS)
2009 cgrp_dfl_root.flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS;
2010 else
2011 cgrp_dfl_root.flags &= ~CGRP_ROOT_PIDS_LOCAL_EVENTS;
2012 }
2013}
2014
2015static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
2016{
2017 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
2018 seq_puts(seq, ",nsdelegate");
2019 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
2020 seq_puts(seq, ",favordynmods");
2021 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
2022 seq_puts(seq, ",memory_localevents");
2023 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
2024 seq_puts(seq, ",memory_recursiveprot");
2025 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2026 seq_puts(seq, ",memory_hugetlb_accounting");
2027 if (cgrp_dfl_root.flags & CGRP_ROOT_PIDS_LOCAL_EVENTS)
2028 seq_puts(seq, ",pids_localevents");
2029 return 0;
2030}
2031
2032static int cgroup_reconfigure(struct fs_context *fc)
2033{
2034 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2035
2036 apply_cgroup_root_flags(ctx->flags);
2037 return 0;
2038}
2039
2040static void init_cgroup_housekeeping(struct cgroup *cgrp)
2041{
2042 struct cgroup_subsys *ss;
2043 int ssid;
2044
2045 INIT_LIST_HEAD(&cgrp->self.sibling);
2046 INIT_LIST_HEAD(&cgrp->self.children);
2047 INIT_LIST_HEAD(&cgrp->cset_links);
2048 INIT_LIST_HEAD(&cgrp->pidlists);
2049 mutex_init(&cgrp->pidlist_mutex);
2050 cgrp->self.cgroup = cgrp;
2051 cgrp->self.flags |= CSS_ONLINE;
2052 cgrp->dom_cgrp = cgrp;
2053 cgrp->max_descendants = INT_MAX;
2054 cgrp->max_depth = INT_MAX;
2055 INIT_LIST_HEAD(&cgrp->rstat_css_list);
2056 prev_cputime_init(&cgrp->prev_cputime);
2057
2058 for_each_subsys(ss, ssid)
2059 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2060
2061 init_waitqueue_head(&cgrp->offline_waitq);
2062 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2063}
2064
2065void init_cgroup_root(struct cgroup_fs_context *ctx)
2066{
2067 struct cgroup_root *root = ctx->root;
2068 struct cgroup *cgrp = &root->cgrp;
2069
2070 INIT_LIST_HEAD_RCU(&root->root_list);
2071 atomic_set(&root->nr_cgrps, 1);
2072 cgrp->root = root;
2073 init_cgroup_housekeeping(cgrp);
2074
2075 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2076 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2077 if (ctx->release_agent)
2078 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2079 if (ctx->name)
2080 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2081 if (ctx->cpuset_clone_children)
2082 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2083}
2084
2085int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2086{
2087 LIST_HEAD(tmp_links);
2088 struct cgroup *root_cgrp = &root->cgrp;
2089 struct kernfs_syscall_ops *kf_sops;
2090 struct css_set *cset;
2091 int i, ret;
2092
2093 lockdep_assert_held(&cgroup_mutex);
2094
2095 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2096 0, GFP_KERNEL);
2097 if (ret)
2098 goto out;
2099
2100 /*
2101 * We're accessing css_set_count without locking css_set_lock here,
2102 * but that's OK - it can only be increased by someone holding
2103 * cgroup_lock, and that's us. Later rebinding may disable
2104 * controllers on the default hierarchy and thus create new csets,
2105 * which can't be more than the existing ones. Allocate 2x.
2106 */
2107 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2108 if (ret)
2109 goto cancel_ref;
2110
2111 ret = cgroup_init_root_id(root);
2112 if (ret)
2113 goto cancel_ref;
2114
2115 kf_sops = root == &cgrp_dfl_root ?
2116 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2117
2118 root->kf_root = kernfs_create_root(kf_sops,
2119 KERNFS_ROOT_CREATE_DEACTIVATED |
2120 KERNFS_ROOT_SUPPORT_EXPORTOP |
2121 KERNFS_ROOT_SUPPORT_USER_XATTR,
2122 root_cgrp);
2123 if (IS_ERR(root->kf_root)) {
2124 ret = PTR_ERR(root->kf_root);
2125 goto exit_root_id;
2126 }
2127 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2128 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2129 root_cgrp->ancestors[0] = root_cgrp;
2130
2131 ret = css_populate_dir(&root_cgrp->self);
2132 if (ret)
2133 goto destroy_root;
2134
2135 ret = cgroup_rstat_init(root_cgrp);
2136 if (ret)
2137 goto destroy_root;
2138
2139 ret = rebind_subsystems(root, ss_mask);
2140 if (ret)
2141 goto exit_stats;
2142
2143 if (root == &cgrp_dfl_root) {
2144 ret = cgroup_bpf_inherit(root_cgrp);
2145 WARN_ON_ONCE(ret);
2146 }
2147
2148 trace_cgroup_setup_root(root);
2149
2150 /*
2151 * There must be no failure case after here, since rebinding takes
2152 * care of subsystems' refcounts, which are explicitly dropped in
2153 * the failure exit path.
2154 */
2155 list_add_rcu(&root->root_list, &cgroup_roots);
2156 cgroup_root_count++;
2157
2158 /*
2159 * Link the root cgroup in this hierarchy into all the css_set
2160 * objects.
2161 */
2162 spin_lock_irq(&css_set_lock);
2163 hash_for_each(css_set_table, i, cset, hlist) {
2164 link_css_set(&tmp_links, cset, root_cgrp);
2165 if (css_set_populated(cset))
2166 cgroup_update_populated(root_cgrp, true);
2167 }
2168 spin_unlock_irq(&css_set_lock);
2169
2170 BUG_ON(!list_empty(&root_cgrp->self.children));
2171 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2172
2173 ret = 0;
2174 goto out;
2175
2176exit_stats:
2177 cgroup_rstat_exit(root_cgrp);
2178destroy_root:
2179 kernfs_destroy_root(root->kf_root);
2180 root->kf_root = NULL;
2181exit_root_id:
2182 cgroup_exit_root_id(root);
2183cancel_ref:
2184 percpu_ref_exit(&root_cgrp->self.refcnt);
2185out:
2186 free_cgrp_cset_links(&tmp_links);
2187 return ret;
2188}
2189
2190int cgroup_do_get_tree(struct fs_context *fc)
2191{
2192 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2193 int ret;
2194
2195 ctx->kfc.root = ctx->root->kf_root;
2196 if (fc->fs_type == &cgroup2_fs_type)
2197 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2198 else
2199 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2200 ret = kernfs_get_tree(fc);
2201
2202 /*
2203 * In non-init cgroup namespace, instead of root cgroup's dentry,
2204 * we return the dentry corresponding to the cgroupns->root_cgrp.
2205 */
2206 if (!ret && ctx->ns != &init_cgroup_ns) {
2207 struct dentry *nsdentry;
2208 struct super_block *sb = fc->root->d_sb;
2209 struct cgroup *cgrp;
2210
2211 cgroup_lock();
2212 spin_lock_irq(&css_set_lock);
2213
2214 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2215
2216 spin_unlock_irq(&css_set_lock);
2217 cgroup_unlock();
2218
2219 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2220 dput(fc->root);
2221 if (IS_ERR(nsdentry)) {
2222 deactivate_locked_super(sb);
2223 ret = PTR_ERR(nsdentry);
2224 nsdentry = NULL;
2225 }
2226 fc->root = nsdentry;
2227 }
2228
2229 if (!ctx->kfc.new_sb_created)
2230 cgroup_put(&ctx->root->cgrp);
2231
2232 return ret;
2233}
2234
2235/*
2236 * Destroy a cgroup filesystem context.
2237 */
2238static void cgroup_fs_context_free(struct fs_context *fc)
2239{
2240 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2241
2242 kfree(ctx->name);
2243 kfree(ctx->release_agent);
2244 put_cgroup_ns(ctx->ns);
2245 kernfs_free_fs_context(fc);
2246 kfree(ctx);
2247}
2248
2249static int cgroup_get_tree(struct fs_context *fc)
2250{
2251 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2252 int ret;
2253
2254 WRITE_ONCE(cgrp_dfl_visible, true);
2255 cgroup_get_live(&cgrp_dfl_root.cgrp);
2256 ctx->root = &cgrp_dfl_root;
2257
2258 ret = cgroup_do_get_tree(fc);
2259 if (!ret)
2260 apply_cgroup_root_flags(ctx->flags);
2261 return ret;
2262}
2263
2264static const struct fs_context_operations cgroup_fs_context_ops = {
2265 .free = cgroup_fs_context_free,
2266 .parse_param = cgroup2_parse_param,
2267 .get_tree = cgroup_get_tree,
2268 .reconfigure = cgroup_reconfigure,
2269};
2270
2271static const struct fs_context_operations cgroup1_fs_context_ops = {
2272 .free = cgroup_fs_context_free,
2273 .parse_param = cgroup1_parse_param,
2274 .get_tree = cgroup1_get_tree,
2275 .reconfigure = cgroup1_reconfigure,
2276};
2277
2278/*
2279 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2280 * we select the namespace we're going to use.
2281 */
2282static int cgroup_init_fs_context(struct fs_context *fc)
2283{
2284 struct cgroup_fs_context *ctx;
2285
2286 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2287 if (!ctx)
2288 return -ENOMEM;
2289
2290 ctx->ns = current->nsproxy->cgroup_ns;
2291 get_cgroup_ns(ctx->ns);
2292 fc->fs_private = &ctx->kfc;
2293 if (fc->fs_type == &cgroup2_fs_type)
2294 fc->ops = &cgroup_fs_context_ops;
2295 else
2296 fc->ops = &cgroup1_fs_context_ops;
2297 put_user_ns(fc->user_ns);
2298 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2299 fc->global = true;
2300
2301 if (have_favordynmods)
2302 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2303
2304 return 0;
2305}
2306
2307static void cgroup_kill_sb(struct super_block *sb)
2308{
2309 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2310 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2311
2312 /*
2313 * If @root doesn't have any children, start killing it.
2314 * This prevents new mounts by disabling percpu_ref_tryget_live().
2315 *
2316 * And don't kill the default root.
2317 */
2318 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2319 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2320 percpu_ref_kill(&root->cgrp.self.refcnt);
2321 cgroup_put(&root->cgrp);
2322 kernfs_kill_sb(sb);
2323}
2324
2325struct file_system_type cgroup_fs_type = {
2326 .name = "cgroup",
2327 .init_fs_context = cgroup_init_fs_context,
2328 .parameters = cgroup1_fs_parameters,
2329 .kill_sb = cgroup_kill_sb,
2330 .fs_flags = FS_USERNS_MOUNT,
2331};
2332
2333static struct file_system_type cgroup2_fs_type = {
2334 .name = "cgroup2",
2335 .init_fs_context = cgroup_init_fs_context,
2336 .parameters = cgroup2_fs_parameters,
2337 .kill_sb = cgroup_kill_sb,
2338 .fs_flags = FS_USERNS_MOUNT,
2339};
2340
2341#ifdef CONFIG_CPUSETS_V1
2342static const struct fs_context_operations cpuset_fs_context_ops = {
2343 .get_tree = cgroup1_get_tree,
2344 .free = cgroup_fs_context_free,
2345};
2346
2347/*
2348 * This is ugly, but preserves the userspace API for existing cpuset
2349 * users. If someone tries to mount the "cpuset" filesystem, we
2350 * silently switch it to mount "cgroup" instead
2351 */
2352static int cpuset_init_fs_context(struct fs_context *fc)
2353{
2354 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2355 struct cgroup_fs_context *ctx;
2356 int err;
2357
2358 err = cgroup_init_fs_context(fc);
2359 if (err) {
2360 kfree(agent);
2361 return err;
2362 }
2363
2364 fc->ops = &cpuset_fs_context_ops;
2365
2366 ctx = cgroup_fc2context(fc);
2367 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2368 ctx->flags |= CGRP_ROOT_NOPREFIX;
2369 ctx->release_agent = agent;
2370
2371 get_filesystem(&cgroup_fs_type);
2372 put_filesystem(fc->fs_type);
2373 fc->fs_type = &cgroup_fs_type;
2374
2375 return 0;
2376}
2377
2378static struct file_system_type cpuset_fs_type = {
2379 .name = "cpuset",
2380 .init_fs_context = cpuset_init_fs_context,
2381 .fs_flags = FS_USERNS_MOUNT,
2382};
2383#endif
2384
2385int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2386 struct cgroup_namespace *ns)
2387{
2388 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2389
2390 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2391}
2392
2393int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2394 struct cgroup_namespace *ns)
2395{
2396 int ret;
2397
2398 cgroup_lock();
2399 spin_lock_irq(&css_set_lock);
2400
2401 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2402
2403 spin_unlock_irq(&css_set_lock);
2404 cgroup_unlock();
2405
2406 return ret;
2407}
2408EXPORT_SYMBOL_GPL(cgroup_path_ns);
2409
2410/**
2411 * cgroup_attach_lock - Lock for ->attach()
2412 * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2413 *
2414 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2415 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2416 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2417 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2418 * lead to deadlocks.
2419 *
2420 * Bringing up a CPU may involve creating and destroying tasks which requires
2421 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2422 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2423 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2424 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2425 * the threadgroup_rwsem to be released to create new tasks. For more details:
2426 *
2427 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2428 *
2429 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2430 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2431 * CPU hotplug is disabled on entry.
2432 */
2433void cgroup_attach_lock(bool lock_threadgroup)
2434{
2435 cpus_read_lock();
2436 if (lock_threadgroup)
2437 percpu_down_write(&cgroup_threadgroup_rwsem);
2438}
2439
2440/**
2441 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2442 * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2443 */
2444void cgroup_attach_unlock(bool lock_threadgroup)
2445{
2446 if (lock_threadgroup)
2447 percpu_up_write(&cgroup_threadgroup_rwsem);
2448 cpus_read_unlock();
2449}
2450
2451/**
2452 * cgroup_migrate_add_task - add a migration target task to a migration context
2453 * @task: target task
2454 * @mgctx: target migration context
2455 *
2456 * Add @task, which is a migration target, to @mgctx->tset. This function
2457 * becomes noop if @task doesn't need to be migrated. @task's css_set
2458 * should have been added as a migration source and @task->cg_list will be
2459 * moved from the css_set's tasks list to mg_tasks one.
2460 */
2461static void cgroup_migrate_add_task(struct task_struct *task,
2462 struct cgroup_mgctx *mgctx)
2463{
2464 struct css_set *cset;
2465
2466 lockdep_assert_held(&css_set_lock);
2467
2468 /* @task either already exited or can't exit until the end */
2469 if (task->flags & PF_EXITING)
2470 return;
2471
2472 /* cgroup_threadgroup_rwsem protects racing against forks */
2473 WARN_ON_ONCE(list_empty(&task->cg_list));
2474
2475 cset = task_css_set(task);
2476 if (!cset->mg_src_cgrp)
2477 return;
2478
2479 mgctx->tset.nr_tasks++;
2480
2481 list_move_tail(&task->cg_list, &cset->mg_tasks);
2482 if (list_empty(&cset->mg_node))
2483 list_add_tail(&cset->mg_node,
2484 &mgctx->tset.src_csets);
2485 if (list_empty(&cset->mg_dst_cset->mg_node))
2486 list_add_tail(&cset->mg_dst_cset->mg_node,
2487 &mgctx->tset.dst_csets);
2488}
2489
2490/**
2491 * cgroup_taskset_first - reset taskset and return the first task
2492 * @tset: taskset of interest
2493 * @dst_cssp: output variable for the destination css
2494 *
2495 * @tset iteration is initialized and the first task is returned.
2496 */
2497struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2498 struct cgroup_subsys_state **dst_cssp)
2499{
2500 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2501 tset->cur_task = NULL;
2502
2503 return cgroup_taskset_next(tset, dst_cssp);
2504}
2505
2506/**
2507 * cgroup_taskset_next - iterate to the next task in taskset
2508 * @tset: taskset of interest
2509 * @dst_cssp: output variable for the destination css
2510 *
2511 * Return the next task in @tset. Iteration must have been initialized
2512 * with cgroup_taskset_first().
2513 */
2514struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2515 struct cgroup_subsys_state **dst_cssp)
2516{
2517 struct css_set *cset = tset->cur_cset;
2518 struct task_struct *task = tset->cur_task;
2519
2520 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2521 if (!task)
2522 task = list_first_entry(&cset->mg_tasks,
2523 struct task_struct, cg_list);
2524 else
2525 task = list_next_entry(task, cg_list);
2526
2527 if (&task->cg_list != &cset->mg_tasks) {
2528 tset->cur_cset = cset;
2529 tset->cur_task = task;
2530
2531 /*
2532 * This function may be called both before and
2533 * after cgroup_migrate_execute(). The two cases
2534 * can be distinguished by looking at whether @cset
2535 * has its ->mg_dst_cset set.
2536 */
2537 if (cset->mg_dst_cset)
2538 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2539 else
2540 *dst_cssp = cset->subsys[tset->ssid];
2541
2542 return task;
2543 }
2544
2545 cset = list_next_entry(cset, mg_node);
2546 task = NULL;
2547 }
2548
2549 return NULL;
2550}
2551
2552/**
2553 * cgroup_migrate_execute - migrate a taskset
2554 * @mgctx: migration context
2555 *
2556 * Migrate tasks in @mgctx as setup by migration preparation functions.
2557 * This function fails iff one of the ->can_attach callbacks fails and
2558 * guarantees that either all or none of the tasks in @mgctx are migrated.
2559 * @mgctx is consumed regardless of success.
2560 */
2561static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2562{
2563 struct cgroup_taskset *tset = &mgctx->tset;
2564 struct cgroup_subsys *ss;
2565 struct task_struct *task, *tmp_task;
2566 struct css_set *cset, *tmp_cset;
2567 int ssid, failed_ssid, ret;
2568
2569 /* check that we can legitimately attach to the cgroup */
2570 if (tset->nr_tasks) {
2571 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2572 if (ss->can_attach) {
2573 tset->ssid = ssid;
2574 ret = ss->can_attach(tset);
2575 if (ret) {
2576 failed_ssid = ssid;
2577 goto out_cancel_attach;
2578 }
2579 }
2580 } while_each_subsys_mask();
2581 }
2582
2583 /*
2584 * Now that we're guaranteed success, proceed to move all tasks to
2585 * the new cgroup. There are no failure cases after here, so this
2586 * is the commit point.
2587 */
2588 spin_lock_irq(&css_set_lock);
2589 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2590 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2591 struct css_set *from_cset = task_css_set(task);
2592 struct css_set *to_cset = cset->mg_dst_cset;
2593
2594 get_css_set(to_cset);
2595 to_cset->nr_tasks++;
2596 css_set_move_task(task, from_cset, to_cset, true);
2597 from_cset->nr_tasks--;
2598 /*
2599 * If the source or destination cgroup is frozen,
2600 * the task might require to change its state.
2601 */
2602 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2603 to_cset->dfl_cgrp);
2604 put_css_set_locked(from_cset);
2605
2606 }
2607 }
2608 spin_unlock_irq(&css_set_lock);
2609
2610 /*
2611 * Migration is committed, all target tasks are now on dst_csets.
2612 * Nothing is sensitive to fork() after this point. Notify
2613 * controllers that migration is complete.
2614 */
2615 tset->csets = &tset->dst_csets;
2616
2617 if (tset->nr_tasks) {
2618 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2619 if (ss->attach) {
2620 tset->ssid = ssid;
2621 ss->attach(tset);
2622 }
2623 } while_each_subsys_mask();
2624 }
2625
2626 ret = 0;
2627 goto out_release_tset;
2628
2629out_cancel_attach:
2630 if (tset->nr_tasks) {
2631 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2632 if (ssid == failed_ssid)
2633 break;
2634 if (ss->cancel_attach) {
2635 tset->ssid = ssid;
2636 ss->cancel_attach(tset);
2637 }
2638 } while_each_subsys_mask();
2639 }
2640out_release_tset:
2641 spin_lock_irq(&css_set_lock);
2642 list_splice_init(&tset->dst_csets, &tset->src_csets);
2643 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2644 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2645 list_del_init(&cset->mg_node);
2646 }
2647 spin_unlock_irq(&css_set_lock);
2648
2649 /*
2650 * Re-initialize the cgroup_taskset structure in case it is reused
2651 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2652 * iteration.
2653 */
2654 tset->nr_tasks = 0;
2655 tset->csets = &tset->src_csets;
2656 return ret;
2657}
2658
2659/**
2660 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2661 * @dst_cgrp: destination cgroup to test
2662 *
2663 * On the default hierarchy, except for the mixable, (possible) thread root
2664 * and threaded cgroups, subtree_control must be zero for migration
2665 * destination cgroups with tasks so that child cgroups don't compete
2666 * against tasks.
2667 */
2668int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2669{
2670 /* v1 doesn't have any restriction */
2671 if (!cgroup_on_dfl(dst_cgrp))
2672 return 0;
2673
2674 /* verify @dst_cgrp can host resources */
2675 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2676 return -EOPNOTSUPP;
2677
2678 /*
2679 * If @dst_cgrp is already or can become a thread root or is
2680 * threaded, it doesn't matter.
2681 */
2682 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2683 return 0;
2684
2685 /* apply no-internal-process constraint */
2686 if (dst_cgrp->subtree_control)
2687 return -EBUSY;
2688
2689 return 0;
2690}
2691
2692/**
2693 * cgroup_migrate_finish - cleanup after attach
2694 * @mgctx: migration context
2695 *
2696 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2697 * those functions for details.
2698 */
2699void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2700{
2701 struct css_set *cset, *tmp_cset;
2702
2703 lockdep_assert_held(&cgroup_mutex);
2704
2705 spin_lock_irq(&css_set_lock);
2706
2707 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2708 mg_src_preload_node) {
2709 cset->mg_src_cgrp = NULL;
2710 cset->mg_dst_cgrp = NULL;
2711 cset->mg_dst_cset = NULL;
2712 list_del_init(&cset->mg_src_preload_node);
2713 put_css_set_locked(cset);
2714 }
2715
2716 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2717 mg_dst_preload_node) {
2718 cset->mg_src_cgrp = NULL;
2719 cset->mg_dst_cgrp = NULL;
2720 cset->mg_dst_cset = NULL;
2721 list_del_init(&cset->mg_dst_preload_node);
2722 put_css_set_locked(cset);
2723 }
2724
2725 spin_unlock_irq(&css_set_lock);
2726}
2727
2728/**
2729 * cgroup_migrate_add_src - add a migration source css_set
2730 * @src_cset: the source css_set to add
2731 * @dst_cgrp: the destination cgroup
2732 * @mgctx: migration context
2733 *
2734 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2735 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2736 * up by cgroup_migrate_finish().
2737 *
2738 * This function may be called without holding cgroup_threadgroup_rwsem
2739 * even if the target is a process. Threads may be created and destroyed
2740 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2741 * into play and the preloaded css_sets are guaranteed to cover all
2742 * migrations.
2743 */
2744void cgroup_migrate_add_src(struct css_set *src_cset,
2745 struct cgroup *dst_cgrp,
2746 struct cgroup_mgctx *mgctx)
2747{
2748 struct cgroup *src_cgrp;
2749
2750 lockdep_assert_held(&cgroup_mutex);
2751 lockdep_assert_held(&css_set_lock);
2752
2753 /*
2754 * If ->dead, @src_set is associated with one or more dead cgroups
2755 * and doesn't contain any migratable tasks. Ignore it early so
2756 * that the rest of migration path doesn't get confused by it.
2757 */
2758 if (src_cset->dead)
2759 return;
2760
2761 if (!list_empty(&src_cset->mg_src_preload_node))
2762 return;
2763
2764 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2765
2766 WARN_ON(src_cset->mg_src_cgrp);
2767 WARN_ON(src_cset->mg_dst_cgrp);
2768 WARN_ON(!list_empty(&src_cset->mg_tasks));
2769 WARN_ON(!list_empty(&src_cset->mg_node));
2770
2771 src_cset->mg_src_cgrp = src_cgrp;
2772 src_cset->mg_dst_cgrp = dst_cgrp;
2773 get_css_set(src_cset);
2774 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2775}
2776
2777/**
2778 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2779 * @mgctx: migration context
2780 *
2781 * Tasks are about to be moved and all the source css_sets have been
2782 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2783 * pins all destination css_sets, links each to its source, and append them
2784 * to @mgctx->preloaded_dst_csets.
2785 *
2786 * This function must be called after cgroup_migrate_add_src() has been
2787 * called on each migration source css_set. After migration is performed
2788 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2789 * @mgctx.
2790 */
2791int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2792{
2793 struct css_set *src_cset, *tmp_cset;
2794
2795 lockdep_assert_held(&cgroup_mutex);
2796
2797 /* look up the dst cset for each src cset and link it to src */
2798 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2799 mg_src_preload_node) {
2800 struct css_set *dst_cset;
2801 struct cgroup_subsys *ss;
2802 int ssid;
2803
2804 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2805 if (!dst_cset)
2806 return -ENOMEM;
2807
2808 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2809
2810 /*
2811 * If src cset equals dst, it's noop. Drop the src.
2812 * cgroup_migrate() will skip the cset too. Note that we
2813 * can't handle src == dst as some nodes are used by both.
2814 */
2815 if (src_cset == dst_cset) {
2816 src_cset->mg_src_cgrp = NULL;
2817 src_cset->mg_dst_cgrp = NULL;
2818 list_del_init(&src_cset->mg_src_preload_node);
2819 put_css_set(src_cset);
2820 put_css_set(dst_cset);
2821 continue;
2822 }
2823
2824 src_cset->mg_dst_cset = dst_cset;
2825
2826 if (list_empty(&dst_cset->mg_dst_preload_node))
2827 list_add_tail(&dst_cset->mg_dst_preload_node,
2828 &mgctx->preloaded_dst_csets);
2829 else
2830 put_css_set(dst_cset);
2831
2832 for_each_subsys(ss, ssid)
2833 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2834 mgctx->ss_mask |= 1 << ssid;
2835 }
2836
2837 return 0;
2838}
2839
2840/**
2841 * cgroup_migrate - migrate a process or task to a cgroup
2842 * @leader: the leader of the process or the task to migrate
2843 * @threadgroup: whether @leader points to the whole process or a single task
2844 * @mgctx: migration context
2845 *
2846 * Migrate a process or task denoted by @leader. If migrating a process,
2847 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2848 * responsible for invoking cgroup_migrate_add_src() and
2849 * cgroup_migrate_prepare_dst() on the targets before invoking this
2850 * function and following up with cgroup_migrate_finish().
2851 *
2852 * As long as a controller's ->can_attach() doesn't fail, this function is
2853 * guaranteed to succeed. This means that, excluding ->can_attach()
2854 * failure, when migrating multiple targets, the success or failure can be
2855 * decided for all targets by invoking group_migrate_prepare_dst() before
2856 * actually starting migrating.
2857 */
2858int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2859 struct cgroup_mgctx *mgctx)
2860{
2861 struct task_struct *task;
2862
2863 /*
2864 * The following thread iteration should be inside an RCU critical
2865 * section to prevent tasks from being freed while taking the snapshot.
2866 * spin_lock_irq() implies RCU critical section here.
2867 */
2868 spin_lock_irq(&css_set_lock);
2869 task = leader;
2870 do {
2871 cgroup_migrate_add_task(task, mgctx);
2872 if (!threadgroup)
2873 break;
2874 } while_each_thread(leader, task);
2875 spin_unlock_irq(&css_set_lock);
2876
2877 return cgroup_migrate_execute(mgctx);
2878}
2879
2880/**
2881 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2882 * @dst_cgrp: the cgroup to attach to
2883 * @leader: the task or the leader of the threadgroup to be attached
2884 * @threadgroup: attach the whole threadgroup?
2885 *
2886 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2887 */
2888int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2889 bool threadgroup)
2890{
2891 DEFINE_CGROUP_MGCTX(mgctx);
2892 struct task_struct *task;
2893 int ret = 0;
2894
2895 /* look up all src csets */
2896 spin_lock_irq(&css_set_lock);
2897 rcu_read_lock();
2898 task = leader;
2899 do {
2900 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2901 if (!threadgroup)
2902 break;
2903 } while_each_thread(leader, task);
2904 rcu_read_unlock();
2905 spin_unlock_irq(&css_set_lock);
2906
2907 /* prepare dst csets and commit */
2908 ret = cgroup_migrate_prepare_dst(&mgctx);
2909 if (!ret)
2910 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2911
2912 cgroup_migrate_finish(&mgctx);
2913
2914 if (!ret)
2915 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2916
2917 return ret;
2918}
2919
2920struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2921 bool *threadgroup_locked)
2922{
2923 struct task_struct *tsk;
2924 pid_t pid;
2925
2926 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2927 return ERR_PTR(-EINVAL);
2928
2929 /*
2930 * If we migrate a single thread, we don't care about threadgroup
2931 * stability. If the thread is `current`, it won't exit(2) under our
2932 * hands or change PID through exec(2). We exclude
2933 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2934 * callers by cgroup_mutex.
2935 * Therefore, we can skip the global lock.
2936 */
2937 lockdep_assert_held(&cgroup_mutex);
2938 *threadgroup_locked = pid || threadgroup;
2939 cgroup_attach_lock(*threadgroup_locked);
2940
2941 rcu_read_lock();
2942 if (pid) {
2943 tsk = find_task_by_vpid(pid);
2944 if (!tsk) {
2945 tsk = ERR_PTR(-ESRCH);
2946 goto out_unlock_threadgroup;
2947 }
2948 } else {
2949 tsk = current;
2950 }
2951
2952 if (threadgroup)
2953 tsk = tsk->group_leader;
2954
2955 /*
2956 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2957 * If userland migrates such a kthread to a non-root cgroup, it can
2958 * become trapped in a cpuset, or RT kthread may be born in a
2959 * cgroup with no rt_runtime allocated. Just say no.
2960 */
2961 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2962 tsk = ERR_PTR(-EINVAL);
2963 goto out_unlock_threadgroup;
2964 }
2965
2966 get_task_struct(tsk);
2967 goto out_unlock_rcu;
2968
2969out_unlock_threadgroup:
2970 cgroup_attach_unlock(*threadgroup_locked);
2971 *threadgroup_locked = false;
2972out_unlock_rcu:
2973 rcu_read_unlock();
2974 return tsk;
2975}
2976
2977void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2978{
2979 struct cgroup_subsys *ss;
2980 int ssid;
2981
2982 /* release reference from cgroup_procs_write_start() */
2983 put_task_struct(task);
2984
2985 cgroup_attach_unlock(threadgroup_locked);
2986
2987 for_each_subsys(ss, ssid)
2988 if (ss->post_attach)
2989 ss->post_attach();
2990}
2991
2992static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2993{
2994 struct cgroup_subsys *ss;
2995 bool printed = false;
2996 int ssid;
2997
2998 do_each_subsys_mask(ss, ssid, ss_mask) {
2999 if (printed)
3000 seq_putc(seq, ' ');
3001 seq_puts(seq, ss->name);
3002 printed = true;
3003 } while_each_subsys_mask();
3004 if (printed)
3005 seq_putc(seq, '\n');
3006}
3007
3008/* show controllers which are enabled from the parent */
3009static int cgroup_controllers_show(struct seq_file *seq, void *v)
3010{
3011 struct cgroup *cgrp = seq_css(seq)->cgroup;
3012
3013 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3014 return 0;
3015}
3016
3017/* show controllers which are enabled for a given cgroup's children */
3018static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3019{
3020 struct cgroup *cgrp = seq_css(seq)->cgroup;
3021
3022 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3023 return 0;
3024}
3025
3026/**
3027 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3028 * @cgrp: root of the subtree to update csses for
3029 *
3030 * @cgrp's control masks have changed and its subtree's css associations
3031 * need to be updated accordingly. This function looks up all css_sets
3032 * which are attached to the subtree, creates the matching updated css_sets
3033 * and migrates the tasks to the new ones.
3034 */
3035static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3036{
3037 DEFINE_CGROUP_MGCTX(mgctx);
3038 struct cgroup_subsys_state *d_css;
3039 struct cgroup *dsct;
3040 struct css_set *src_cset;
3041 bool has_tasks;
3042 int ret;
3043
3044 lockdep_assert_held(&cgroup_mutex);
3045
3046 /* look up all csses currently attached to @cgrp's subtree */
3047 spin_lock_irq(&css_set_lock);
3048 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3049 struct cgrp_cset_link *link;
3050
3051 /*
3052 * As cgroup_update_dfl_csses() is only called by
3053 * cgroup_apply_control(). The csses associated with the
3054 * given cgrp will not be affected by changes made to
3055 * its subtree_control file. We can skip them.
3056 */
3057 if (dsct == cgrp)
3058 continue;
3059
3060 list_for_each_entry(link, &dsct->cset_links, cset_link)
3061 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3062 }
3063 spin_unlock_irq(&css_set_lock);
3064
3065 /*
3066 * We need to write-lock threadgroup_rwsem while migrating tasks.
3067 * However, if there are no source csets for @cgrp, changing its
3068 * controllers isn't gonna produce any task migrations and the
3069 * write-locking can be skipped safely.
3070 */
3071 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3072 cgroup_attach_lock(has_tasks);
3073
3074 /* NULL dst indicates self on default hierarchy */
3075 ret = cgroup_migrate_prepare_dst(&mgctx);
3076 if (ret)
3077 goto out_finish;
3078
3079 spin_lock_irq(&css_set_lock);
3080 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3081 mg_src_preload_node) {
3082 struct task_struct *task, *ntask;
3083
3084 /* all tasks in src_csets need to be migrated */
3085 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3086 cgroup_migrate_add_task(task, &mgctx);
3087 }
3088 spin_unlock_irq(&css_set_lock);
3089
3090 ret = cgroup_migrate_execute(&mgctx);
3091out_finish:
3092 cgroup_migrate_finish(&mgctx);
3093 cgroup_attach_unlock(has_tasks);
3094 return ret;
3095}
3096
3097/**
3098 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3099 * @cgrp: root of the target subtree
3100 *
3101 * Because css offlining is asynchronous, userland may try to re-enable a
3102 * controller while the previous css is still around. This function grabs
3103 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3104 */
3105void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3106 __acquires(&cgroup_mutex)
3107{
3108 struct cgroup *dsct;
3109 struct cgroup_subsys_state *d_css;
3110 struct cgroup_subsys *ss;
3111 int ssid;
3112
3113restart:
3114 cgroup_lock();
3115
3116 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3117 for_each_subsys(ss, ssid) {
3118 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3119 DEFINE_WAIT(wait);
3120
3121 if (!css || !percpu_ref_is_dying(&css->refcnt))
3122 continue;
3123
3124 cgroup_get_live(dsct);
3125 prepare_to_wait(&dsct->offline_waitq, &wait,
3126 TASK_UNINTERRUPTIBLE);
3127
3128 cgroup_unlock();
3129 schedule();
3130 finish_wait(&dsct->offline_waitq, &wait);
3131
3132 cgroup_put(dsct);
3133 goto restart;
3134 }
3135 }
3136}
3137
3138/**
3139 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3140 * @cgrp: root of the target subtree
3141 *
3142 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3143 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3144 * itself.
3145 */
3146static void cgroup_save_control(struct cgroup *cgrp)
3147{
3148 struct cgroup *dsct;
3149 struct cgroup_subsys_state *d_css;
3150
3151 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3152 dsct->old_subtree_control = dsct->subtree_control;
3153 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3154 dsct->old_dom_cgrp = dsct->dom_cgrp;
3155 }
3156}
3157
3158/**
3159 * cgroup_propagate_control - refresh control masks of a subtree
3160 * @cgrp: root of the target subtree
3161 *
3162 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3163 * ->subtree_control and propagate controller availability through the
3164 * subtree so that descendants don't have unavailable controllers enabled.
3165 */
3166static void cgroup_propagate_control(struct cgroup *cgrp)
3167{
3168 struct cgroup *dsct;
3169 struct cgroup_subsys_state *d_css;
3170
3171 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3172 dsct->subtree_control &= cgroup_control(dsct);
3173 dsct->subtree_ss_mask =
3174 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3175 cgroup_ss_mask(dsct));
3176 }
3177}
3178
3179/**
3180 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3181 * @cgrp: root of the target subtree
3182 *
3183 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3184 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3185 * itself.
3186 */
3187static void cgroup_restore_control(struct cgroup *cgrp)
3188{
3189 struct cgroup *dsct;
3190 struct cgroup_subsys_state *d_css;
3191
3192 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3193 dsct->subtree_control = dsct->old_subtree_control;
3194 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3195 dsct->dom_cgrp = dsct->old_dom_cgrp;
3196 }
3197}
3198
3199static bool css_visible(struct cgroup_subsys_state *css)
3200{
3201 struct cgroup_subsys *ss = css->ss;
3202 struct cgroup *cgrp = css->cgroup;
3203
3204 if (cgroup_control(cgrp) & (1 << ss->id))
3205 return true;
3206 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3207 return false;
3208 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3209}
3210
3211/**
3212 * cgroup_apply_control_enable - enable or show csses according to control
3213 * @cgrp: root of the target subtree
3214 *
3215 * Walk @cgrp's subtree and create new csses or make the existing ones
3216 * visible. A css is created invisible if it's being implicitly enabled
3217 * through dependency. An invisible css is made visible when the userland
3218 * explicitly enables it.
3219 *
3220 * Returns 0 on success, -errno on failure. On failure, csses which have
3221 * been processed already aren't cleaned up. The caller is responsible for
3222 * cleaning up with cgroup_apply_control_disable().
3223 */
3224static int cgroup_apply_control_enable(struct cgroup *cgrp)
3225{
3226 struct cgroup *dsct;
3227 struct cgroup_subsys_state *d_css;
3228 struct cgroup_subsys *ss;
3229 int ssid, ret;
3230
3231 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3232 for_each_subsys(ss, ssid) {
3233 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3234
3235 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3236 continue;
3237
3238 if (!css) {
3239 css = css_create(dsct, ss);
3240 if (IS_ERR(css))
3241 return PTR_ERR(css);
3242 }
3243
3244 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3245
3246 if (css_visible(css)) {
3247 ret = css_populate_dir(css);
3248 if (ret)
3249 return ret;
3250 }
3251 }
3252 }
3253
3254 return 0;
3255}
3256
3257/**
3258 * cgroup_apply_control_disable - kill or hide csses according to control
3259 * @cgrp: root of the target subtree
3260 *
3261 * Walk @cgrp's subtree and kill and hide csses so that they match
3262 * cgroup_ss_mask() and cgroup_visible_mask().
3263 *
3264 * A css is hidden when the userland requests it to be disabled while other
3265 * subsystems are still depending on it. The css must not actively control
3266 * resources and be in the vanilla state if it's made visible again later.
3267 * Controllers which may be depended upon should provide ->css_reset() for
3268 * this purpose.
3269 */
3270static void cgroup_apply_control_disable(struct cgroup *cgrp)
3271{
3272 struct cgroup *dsct;
3273 struct cgroup_subsys_state *d_css;
3274 struct cgroup_subsys *ss;
3275 int ssid;
3276
3277 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3278 for_each_subsys(ss, ssid) {
3279 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3280
3281 if (!css)
3282 continue;
3283
3284 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3285
3286 if (css->parent &&
3287 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3288 kill_css(css);
3289 } else if (!css_visible(css)) {
3290 css_clear_dir(css);
3291 if (ss->css_reset)
3292 ss->css_reset(css);
3293 }
3294 }
3295 }
3296}
3297
3298/**
3299 * cgroup_apply_control - apply control mask updates to the subtree
3300 * @cgrp: root of the target subtree
3301 *
3302 * subsystems can be enabled and disabled in a subtree using the following
3303 * steps.
3304 *
3305 * 1. Call cgroup_save_control() to stash the current state.
3306 * 2. Update ->subtree_control masks in the subtree as desired.
3307 * 3. Call cgroup_apply_control() to apply the changes.
3308 * 4. Optionally perform other related operations.
3309 * 5. Call cgroup_finalize_control() to finish up.
3310 *
3311 * This function implements step 3 and propagates the mask changes
3312 * throughout @cgrp's subtree, updates csses accordingly and perform
3313 * process migrations.
3314 */
3315static int cgroup_apply_control(struct cgroup *cgrp)
3316{
3317 int ret;
3318
3319 cgroup_propagate_control(cgrp);
3320
3321 ret = cgroup_apply_control_enable(cgrp);
3322 if (ret)
3323 return ret;
3324
3325 /*
3326 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3327 * making the following cgroup_update_dfl_csses() properly update
3328 * css associations of all tasks in the subtree.
3329 */
3330 return cgroup_update_dfl_csses(cgrp);
3331}
3332
3333/**
3334 * cgroup_finalize_control - finalize control mask update
3335 * @cgrp: root of the target subtree
3336 * @ret: the result of the update
3337 *
3338 * Finalize control mask update. See cgroup_apply_control() for more info.
3339 */
3340static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3341{
3342 if (ret) {
3343 cgroup_restore_control(cgrp);
3344 cgroup_propagate_control(cgrp);
3345 }
3346
3347 cgroup_apply_control_disable(cgrp);
3348}
3349
3350static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3351{
3352 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3353
3354 /* if nothing is getting enabled, nothing to worry about */
3355 if (!enable)
3356 return 0;
3357
3358 /* can @cgrp host any resources? */
3359 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3360 return -EOPNOTSUPP;
3361
3362 /* mixables don't care */
3363 if (cgroup_is_mixable(cgrp))
3364 return 0;
3365
3366 if (domain_enable) {
3367 /* can't enable domain controllers inside a thread subtree */
3368 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3369 return -EOPNOTSUPP;
3370 } else {
3371 /*
3372 * Threaded controllers can handle internal competitions
3373 * and are always allowed inside a (prospective) thread
3374 * subtree.
3375 */
3376 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3377 return 0;
3378 }
3379
3380 /*
3381 * Controllers can't be enabled for a cgroup with tasks to avoid
3382 * child cgroups competing against tasks.
3383 */
3384 if (cgroup_has_tasks(cgrp))
3385 return -EBUSY;
3386
3387 return 0;
3388}
3389
3390/* change the enabled child controllers for a cgroup in the default hierarchy */
3391static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3392 char *buf, size_t nbytes,
3393 loff_t off)
3394{
3395 u16 enable = 0, disable = 0;
3396 struct cgroup *cgrp, *child;
3397 struct cgroup_subsys *ss;
3398 char *tok;
3399 int ssid, ret;
3400
3401 /*
3402 * Parse input - space separated list of subsystem names prefixed
3403 * with either + or -.
3404 */
3405 buf = strstrip(buf);
3406 while ((tok = strsep(&buf, " "))) {
3407 if (tok[0] == '\0')
3408 continue;
3409 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3410 if (!cgroup_ssid_enabled(ssid) ||
3411 strcmp(tok + 1, ss->name))
3412 continue;
3413
3414 if (*tok == '+') {
3415 enable |= 1 << ssid;
3416 disable &= ~(1 << ssid);
3417 } else if (*tok == '-') {
3418 disable |= 1 << ssid;
3419 enable &= ~(1 << ssid);
3420 } else {
3421 return -EINVAL;
3422 }
3423 break;
3424 } while_each_subsys_mask();
3425 if (ssid == CGROUP_SUBSYS_COUNT)
3426 return -EINVAL;
3427 }
3428
3429 cgrp = cgroup_kn_lock_live(of->kn, true);
3430 if (!cgrp)
3431 return -ENODEV;
3432
3433 for_each_subsys(ss, ssid) {
3434 if (enable & (1 << ssid)) {
3435 if (cgrp->subtree_control & (1 << ssid)) {
3436 enable &= ~(1 << ssid);
3437 continue;
3438 }
3439
3440 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3441 ret = -ENOENT;
3442 goto out_unlock;
3443 }
3444 } else if (disable & (1 << ssid)) {
3445 if (!(cgrp->subtree_control & (1 << ssid))) {
3446 disable &= ~(1 << ssid);
3447 continue;
3448 }
3449
3450 /* a child has it enabled? */
3451 cgroup_for_each_live_child(child, cgrp) {
3452 if (child->subtree_control & (1 << ssid)) {
3453 ret = -EBUSY;
3454 goto out_unlock;
3455 }
3456 }
3457 }
3458 }
3459
3460 if (!enable && !disable) {
3461 ret = 0;
3462 goto out_unlock;
3463 }
3464
3465 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3466 if (ret)
3467 goto out_unlock;
3468
3469 /* save and update control masks and prepare csses */
3470 cgroup_save_control(cgrp);
3471
3472 cgrp->subtree_control |= enable;
3473 cgrp->subtree_control &= ~disable;
3474
3475 ret = cgroup_apply_control(cgrp);
3476 cgroup_finalize_control(cgrp, ret);
3477 if (ret)
3478 goto out_unlock;
3479
3480 kernfs_activate(cgrp->kn);
3481out_unlock:
3482 cgroup_kn_unlock(of->kn);
3483 return ret ?: nbytes;
3484}
3485
3486/**
3487 * cgroup_enable_threaded - make @cgrp threaded
3488 * @cgrp: the target cgroup
3489 *
3490 * Called when "threaded" is written to the cgroup.type interface file and
3491 * tries to make @cgrp threaded and join the parent's resource domain.
3492 * This function is never called on the root cgroup as cgroup.type doesn't
3493 * exist on it.
3494 */
3495static int cgroup_enable_threaded(struct cgroup *cgrp)
3496{
3497 struct cgroup *parent = cgroup_parent(cgrp);
3498 struct cgroup *dom_cgrp = parent->dom_cgrp;
3499 struct cgroup *dsct;
3500 struct cgroup_subsys_state *d_css;
3501 int ret;
3502
3503 lockdep_assert_held(&cgroup_mutex);
3504
3505 /* noop if already threaded */
3506 if (cgroup_is_threaded(cgrp))
3507 return 0;
3508
3509 /*
3510 * If @cgroup is populated or has domain controllers enabled, it
3511 * can't be switched. While the below cgroup_can_be_thread_root()
3512 * test can catch the same conditions, that's only when @parent is
3513 * not mixable, so let's check it explicitly.
3514 */
3515 if (cgroup_is_populated(cgrp) ||
3516 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3517 return -EOPNOTSUPP;
3518
3519 /* we're joining the parent's domain, ensure its validity */
3520 if (!cgroup_is_valid_domain(dom_cgrp) ||
3521 !cgroup_can_be_thread_root(dom_cgrp))
3522 return -EOPNOTSUPP;
3523
3524 /*
3525 * The following shouldn't cause actual migrations and should
3526 * always succeed.
3527 */
3528 cgroup_save_control(cgrp);
3529
3530 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3531 if (dsct == cgrp || cgroup_is_threaded(dsct))
3532 dsct->dom_cgrp = dom_cgrp;
3533
3534 ret = cgroup_apply_control(cgrp);
3535 if (!ret)
3536 parent->nr_threaded_children++;
3537
3538 cgroup_finalize_control(cgrp, ret);
3539 return ret;
3540}
3541
3542static int cgroup_type_show(struct seq_file *seq, void *v)
3543{
3544 struct cgroup *cgrp = seq_css(seq)->cgroup;
3545
3546 if (cgroup_is_threaded(cgrp))
3547 seq_puts(seq, "threaded\n");
3548 else if (!cgroup_is_valid_domain(cgrp))
3549 seq_puts(seq, "domain invalid\n");
3550 else if (cgroup_is_thread_root(cgrp))
3551 seq_puts(seq, "domain threaded\n");
3552 else
3553 seq_puts(seq, "domain\n");
3554
3555 return 0;
3556}
3557
3558static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3559 size_t nbytes, loff_t off)
3560{
3561 struct cgroup *cgrp;
3562 int ret;
3563
3564 /* only switching to threaded mode is supported */
3565 if (strcmp(strstrip(buf), "threaded"))
3566 return -EINVAL;
3567
3568 /* drain dying csses before we re-apply (threaded) subtree control */
3569 cgrp = cgroup_kn_lock_live(of->kn, true);
3570 if (!cgrp)
3571 return -ENOENT;
3572
3573 /* threaded can only be enabled */
3574 ret = cgroup_enable_threaded(cgrp);
3575
3576 cgroup_kn_unlock(of->kn);
3577 return ret ?: nbytes;
3578}
3579
3580static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3581{
3582 struct cgroup *cgrp = seq_css(seq)->cgroup;
3583 int descendants = READ_ONCE(cgrp->max_descendants);
3584
3585 if (descendants == INT_MAX)
3586 seq_puts(seq, "max\n");
3587 else
3588 seq_printf(seq, "%d\n", descendants);
3589
3590 return 0;
3591}
3592
3593static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3594 char *buf, size_t nbytes, loff_t off)
3595{
3596 struct cgroup *cgrp;
3597 int descendants;
3598 ssize_t ret;
3599
3600 buf = strstrip(buf);
3601 if (!strcmp(buf, "max")) {
3602 descendants = INT_MAX;
3603 } else {
3604 ret = kstrtoint(buf, 0, &descendants);
3605 if (ret)
3606 return ret;
3607 }
3608
3609 if (descendants < 0)
3610 return -ERANGE;
3611
3612 cgrp = cgroup_kn_lock_live(of->kn, false);
3613 if (!cgrp)
3614 return -ENOENT;
3615
3616 cgrp->max_descendants = descendants;
3617
3618 cgroup_kn_unlock(of->kn);
3619
3620 return nbytes;
3621}
3622
3623static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3624{
3625 struct cgroup *cgrp = seq_css(seq)->cgroup;
3626 int depth = READ_ONCE(cgrp->max_depth);
3627
3628 if (depth == INT_MAX)
3629 seq_puts(seq, "max\n");
3630 else
3631 seq_printf(seq, "%d\n", depth);
3632
3633 return 0;
3634}
3635
3636static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3637 char *buf, size_t nbytes, loff_t off)
3638{
3639 struct cgroup *cgrp;
3640 ssize_t ret;
3641 int depth;
3642
3643 buf = strstrip(buf);
3644 if (!strcmp(buf, "max")) {
3645 depth = INT_MAX;
3646 } else {
3647 ret = kstrtoint(buf, 0, &depth);
3648 if (ret)
3649 return ret;
3650 }
3651
3652 if (depth < 0)
3653 return -ERANGE;
3654
3655 cgrp = cgroup_kn_lock_live(of->kn, false);
3656 if (!cgrp)
3657 return -ENOENT;
3658
3659 cgrp->max_depth = depth;
3660
3661 cgroup_kn_unlock(of->kn);
3662
3663 return nbytes;
3664}
3665
3666static int cgroup_events_show(struct seq_file *seq, void *v)
3667{
3668 struct cgroup *cgrp = seq_css(seq)->cgroup;
3669
3670 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3671 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3672
3673 return 0;
3674}
3675
3676static int cgroup_stat_show(struct seq_file *seq, void *v)
3677{
3678 struct cgroup *cgroup = seq_css(seq)->cgroup;
3679 struct cgroup_subsys_state *css;
3680 int dying_cnt[CGROUP_SUBSYS_COUNT];
3681 int ssid;
3682
3683 seq_printf(seq, "nr_descendants %d\n",
3684 cgroup->nr_descendants);
3685
3686 /*
3687 * Show the number of live and dying csses associated with each of
3688 * non-inhibited cgroup subsystems that is bound to cgroup v2.
3689 *
3690 * Without proper lock protection, racing is possible. So the
3691 * numbers may not be consistent when that happens.
3692 */
3693 rcu_read_lock();
3694 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) {
3695 dying_cnt[ssid] = -1;
3696 if ((BIT(ssid) & cgrp_dfl_inhibit_ss_mask) ||
3697 (cgroup_subsys[ssid]->root != &cgrp_dfl_root))
3698 continue;
3699 css = rcu_dereference_raw(cgroup->subsys[ssid]);
3700 dying_cnt[ssid] = cgroup->nr_dying_subsys[ssid];
3701 seq_printf(seq, "nr_subsys_%s %d\n", cgroup_subsys[ssid]->name,
3702 css ? (css->nr_descendants + 1) : 0);
3703 }
3704
3705 seq_printf(seq, "nr_dying_descendants %d\n",
3706 cgroup->nr_dying_descendants);
3707 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) {
3708 if (dying_cnt[ssid] >= 0)
3709 seq_printf(seq, "nr_dying_subsys_%s %d\n",
3710 cgroup_subsys[ssid]->name, dying_cnt[ssid]);
3711 }
3712 rcu_read_unlock();
3713 return 0;
3714}
3715
3716#ifdef CONFIG_CGROUP_SCHED
3717/**
3718 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
3719 * @cgrp: the cgroup of interest
3720 * @ss: the subsystem of interest
3721 *
3722 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
3723 * or is offline, %NULL is returned.
3724 */
3725static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
3726 struct cgroup_subsys *ss)
3727{
3728 struct cgroup_subsys_state *css;
3729
3730 rcu_read_lock();
3731 css = cgroup_css(cgrp, ss);
3732 if (css && !css_tryget_online(css))
3733 css = NULL;
3734 rcu_read_unlock();
3735
3736 return css;
3737}
3738
3739static int cgroup_extra_stat_show(struct seq_file *seq, int ssid)
3740{
3741 struct cgroup *cgrp = seq_css(seq)->cgroup;
3742 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3743 struct cgroup_subsys_state *css;
3744 int ret;
3745
3746 if (!ss->css_extra_stat_show)
3747 return 0;
3748
3749 css = cgroup_tryget_css(cgrp, ss);
3750 if (!css)
3751 return 0;
3752
3753 ret = ss->css_extra_stat_show(seq, css);
3754 css_put(css);
3755 return ret;
3756}
3757
3758static int cgroup_local_stat_show(struct seq_file *seq,
3759 struct cgroup *cgrp, int ssid)
3760{
3761 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3762 struct cgroup_subsys_state *css;
3763 int ret;
3764
3765 if (!ss->css_local_stat_show)
3766 return 0;
3767
3768 css = cgroup_tryget_css(cgrp, ss);
3769 if (!css)
3770 return 0;
3771
3772 ret = ss->css_local_stat_show(seq, css);
3773 css_put(css);
3774 return ret;
3775}
3776#endif
3777
3778static int cpu_stat_show(struct seq_file *seq, void *v)
3779{
3780 int ret = 0;
3781
3782 cgroup_base_stat_cputime_show(seq);
3783#ifdef CONFIG_CGROUP_SCHED
3784 ret = cgroup_extra_stat_show(seq, cpu_cgrp_id);
3785#endif
3786 return ret;
3787}
3788
3789static int cpu_local_stat_show(struct seq_file *seq, void *v)
3790{
3791 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3792 int ret = 0;
3793
3794#ifdef CONFIG_CGROUP_SCHED
3795 ret = cgroup_local_stat_show(seq, cgrp, cpu_cgrp_id);
3796#endif
3797 return ret;
3798}
3799
3800#ifdef CONFIG_PSI
3801static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3802{
3803 struct cgroup *cgrp = seq_css(seq)->cgroup;
3804 struct psi_group *psi = cgroup_psi(cgrp);
3805
3806 return psi_show(seq, psi, PSI_IO);
3807}
3808static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3809{
3810 struct cgroup *cgrp = seq_css(seq)->cgroup;
3811 struct psi_group *psi = cgroup_psi(cgrp);
3812
3813 return psi_show(seq, psi, PSI_MEM);
3814}
3815static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3816{
3817 struct cgroup *cgrp = seq_css(seq)->cgroup;
3818 struct psi_group *psi = cgroup_psi(cgrp);
3819
3820 return psi_show(seq, psi, PSI_CPU);
3821}
3822
3823static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3824 size_t nbytes, enum psi_res res)
3825{
3826 struct cgroup_file_ctx *ctx = of->priv;
3827 struct psi_trigger *new;
3828 struct cgroup *cgrp;
3829 struct psi_group *psi;
3830
3831 cgrp = cgroup_kn_lock_live(of->kn, false);
3832 if (!cgrp)
3833 return -ENODEV;
3834
3835 cgroup_get(cgrp);
3836 cgroup_kn_unlock(of->kn);
3837
3838 /* Allow only one trigger per file descriptor */
3839 if (ctx->psi.trigger) {
3840 cgroup_put(cgrp);
3841 return -EBUSY;
3842 }
3843
3844 psi = cgroup_psi(cgrp);
3845 new = psi_trigger_create(psi, buf, res, of->file, of);
3846 if (IS_ERR(new)) {
3847 cgroup_put(cgrp);
3848 return PTR_ERR(new);
3849 }
3850
3851 smp_store_release(&ctx->psi.trigger, new);
3852 cgroup_put(cgrp);
3853
3854 return nbytes;
3855}
3856
3857static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3858 char *buf, size_t nbytes,
3859 loff_t off)
3860{
3861 return pressure_write(of, buf, nbytes, PSI_IO);
3862}
3863
3864static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3865 char *buf, size_t nbytes,
3866 loff_t off)
3867{
3868 return pressure_write(of, buf, nbytes, PSI_MEM);
3869}
3870
3871static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3872 char *buf, size_t nbytes,
3873 loff_t off)
3874{
3875 return pressure_write(of, buf, nbytes, PSI_CPU);
3876}
3877
3878#ifdef CONFIG_IRQ_TIME_ACCOUNTING
3879static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
3880{
3881 struct cgroup *cgrp = seq_css(seq)->cgroup;
3882 struct psi_group *psi = cgroup_psi(cgrp);
3883
3884 return psi_show(seq, psi, PSI_IRQ);
3885}
3886
3887static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
3888 char *buf, size_t nbytes,
3889 loff_t off)
3890{
3891 return pressure_write(of, buf, nbytes, PSI_IRQ);
3892}
3893#endif
3894
3895static int cgroup_pressure_show(struct seq_file *seq, void *v)
3896{
3897 struct cgroup *cgrp = seq_css(seq)->cgroup;
3898 struct psi_group *psi = cgroup_psi(cgrp);
3899
3900 seq_printf(seq, "%d\n", psi->enabled);
3901
3902 return 0;
3903}
3904
3905static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
3906 char *buf, size_t nbytes,
3907 loff_t off)
3908{
3909 ssize_t ret;
3910 int enable;
3911 struct cgroup *cgrp;
3912 struct psi_group *psi;
3913
3914 ret = kstrtoint(strstrip(buf), 0, &enable);
3915 if (ret)
3916 return ret;
3917
3918 if (enable < 0 || enable > 1)
3919 return -ERANGE;
3920
3921 cgrp = cgroup_kn_lock_live(of->kn, false);
3922 if (!cgrp)
3923 return -ENOENT;
3924
3925 psi = cgroup_psi(cgrp);
3926 if (psi->enabled != enable) {
3927 int i;
3928
3929 /* show or hide {cpu,memory,io,irq}.pressure files */
3930 for (i = 0; i < NR_PSI_RESOURCES; i++)
3931 cgroup_file_show(&cgrp->psi_files[i], enable);
3932
3933 psi->enabled = enable;
3934 if (enable)
3935 psi_cgroup_restart(psi);
3936 }
3937
3938 cgroup_kn_unlock(of->kn);
3939
3940 return nbytes;
3941}
3942
3943static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3944 poll_table *pt)
3945{
3946 struct cgroup_file_ctx *ctx = of->priv;
3947
3948 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3949}
3950
3951static void cgroup_pressure_release(struct kernfs_open_file *of)
3952{
3953 struct cgroup_file_ctx *ctx = of->priv;
3954
3955 psi_trigger_destroy(ctx->psi.trigger);
3956}
3957
3958bool cgroup_psi_enabled(void)
3959{
3960 if (static_branch_likely(&psi_disabled))
3961 return false;
3962
3963 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3964}
3965
3966#else /* CONFIG_PSI */
3967bool cgroup_psi_enabled(void)
3968{
3969 return false;
3970}
3971
3972#endif /* CONFIG_PSI */
3973
3974static int cgroup_freeze_show(struct seq_file *seq, void *v)
3975{
3976 struct cgroup *cgrp = seq_css(seq)->cgroup;
3977
3978 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3979
3980 return 0;
3981}
3982
3983static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3984 char *buf, size_t nbytes, loff_t off)
3985{
3986 struct cgroup *cgrp;
3987 ssize_t ret;
3988 int freeze;
3989
3990 ret = kstrtoint(strstrip(buf), 0, &freeze);
3991 if (ret)
3992 return ret;
3993
3994 if (freeze < 0 || freeze > 1)
3995 return -ERANGE;
3996
3997 cgrp = cgroup_kn_lock_live(of->kn, false);
3998 if (!cgrp)
3999 return -ENOENT;
4000
4001 cgroup_freeze(cgrp, freeze);
4002
4003 cgroup_kn_unlock(of->kn);
4004
4005 return nbytes;
4006}
4007
4008static void __cgroup_kill(struct cgroup *cgrp)
4009{
4010 struct css_task_iter it;
4011 struct task_struct *task;
4012
4013 lockdep_assert_held(&cgroup_mutex);
4014
4015 spin_lock_irq(&css_set_lock);
4016 cgrp->kill_seq++;
4017 spin_unlock_irq(&css_set_lock);
4018
4019 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
4020 while ((task = css_task_iter_next(&it))) {
4021 /* Ignore kernel threads here. */
4022 if (task->flags & PF_KTHREAD)
4023 continue;
4024
4025 /* Skip tasks that are already dying. */
4026 if (__fatal_signal_pending(task))
4027 continue;
4028
4029 send_sig(SIGKILL, task, 0);
4030 }
4031 css_task_iter_end(&it);
4032}
4033
4034static void cgroup_kill(struct cgroup *cgrp)
4035{
4036 struct cgroup_subsys_state *css;
4037 struct cgroup *dsct;
4038
4039 lockdep_assert_held(&cgroup_mutex);
4040
4041 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
4042 __cgroup_kill(dsct);
4043}
4044
4045static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
4046 size_t nbytes, loff_t off)
4047{
4048 ssize_t ret = 0;
4049 int kill;
4050 struct cgroup *cgrp;
4051
4052 ret = kstrtoint(strstrip(buf), 0, &kill);
4053 if (ret)
4054 return ret;
4055
4056 if (kill != 1)
4057 return -ERANGE;
4058
4059 cgrp = cgroup_kn_lock_live(of->kn, false);
4060 if (!cgrp)
4061 return -ENOENT;
4062
4063 /*
4064 * Killing is a process directed operation, i.e. the whole thread-group
4065 * is taken down so act like we do for cgroup.procs and only make this
4066 * writable in non-threaded cgroups.
4067 */
4068 if (cgroup_is_threaded(cgrp))
4069 ret = -EOPNOTSUPP;
4070 else
4071 cgroup_kill(cgrp);
4072
4073 cgroup_kn_unlock(of->kn);
4074
4075 return ret ?: nbytes;
4076}
4077
4078static int cgroup_file_open(struct kernfs_open_file *of)
4079{
4080 struct cftype *cft = of_cft(of);
4081 struct cgroup_file_ctx *ctx;
4082 int ret;
4083
4084 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
4085 if (!ctx)
4086 return -ENOMEM;
4087
4088 ctx->ns = current->nsproxy->cgroup_ns;
4089 get_cgroup_ns(ctx->ns);
4090 of->priv = ctx;
4091
4092 if (!cft->open)
4093 return 0;
4094
4095 ret = cft->open(of);
4096 if (ret) {
4097 put_cgroup_ns(ctx->ns);
4098 kfree(ctx);
4099 }
4100 return ret;
4101}
4102
4103static void cgroup_file_release(struct kernfs_open_file *of)
4104{
4105 struct cftype *cft = of_cft(of);
4106 struct cgroup_file_ctx *ctx = of->priv;
4107
4108 if (cft->release)
4109 cft->release(of);
4110 put_cgroup_ns(ctx->ns);
4111 kfree(ctx);
4112}
4113
4114static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4115 size_t nbytes, loff_t off)
4116{
4117 struct cgroup_file_ctx *ctx = of->priv;
4118 struct cgroup *cgrp = of->kn->parent->priv;
4119 struct cftype *cft = of_cft(of);
4120 struct cgroup_subsys_state *css;
4121 int ret;
4122
4123 if (!nbytes)
4124 return 0;
4125
4126 /*
4127 * If namespaces are delegation boundaries, disallow writes to
4128 * files in an non-init namespace root from inside the namespace
4129 * except for the files explicitly marked delegatable -
4130 * eg. cgroup.procs, cgroup.threads and cgroup.subtree_control.
4131 */
4132 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4133 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4134 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4135 return -EPERM;
4136
4137 if (cft->write)
4138 return cft->write(of, buf, nbytes, off);
4139
4140 /*
4141 * kernfs guarantees that a file isn't deleted with operations in
4142 * flight, which means that the matching css is and stays alive and
4143 * doesn't need to be pinned. The RCU locking is not necessary
4144 * either. It's just for the convenience of using cgroup_css().
4145 */
4146 rcu_read_lock();
4147 css = cgroup_css(cgrp, cft->ss);
4148 rcu_read_unlock();
4149
4150 if (cft->write_u64) {
4151 unsigned long long v;
4152 ret = kstrtoull(buf, 0, &v);
4153 if (!ret)
4154 ret = cft->write_u64(css, cft, v);
4155 } else if (cft->write_s64) {
4156 long long v;
4157 ret = kstrtoll(buf, 0, &v);
4158 if (!ret)
4159 ret = cft->write_s64(css, cft, v);
4160 } else {
4161 ret = -EINVAL;
4162 }
4163
4164 return ret ?: nbytes;
4165}
4166
4167static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4168{
4169 struct cftype *cft = of_cft(of);
4170
4171 if (cft->poll)
4172 return cft->poll(of, pt);
4173
4174 return kernfs_generic_poll(of, pt);
4175}
4176
4177static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4178{
4179 return seq_cft(seq)->seq_start(seq, ppos);
4180}
4181
4182static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4183{
4184 return seq_cft(seq)->seq_next(seq, v, ppos);
4185}
4186
4187static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4188{
4189 if (seq_cft(seq)->seq_stop)
4190 seq_cft(seq)->seq_stop(seq, v);
4191}
4192
4193static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4194{
4195 struct cftype *cft = seq_cft(m);
4196 struct cgroup_subsys_state *css = seq_css(m);
4197
4198 if (cft->seq_show)
4199 return cft->seq_show(m, arg);
4200
4201 if (cft->read_u64)
4202 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4203 else if (cft->read_s64)
4204 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4205 else
4206 return -EINVAL;
4207 return 0;
4208}
4209
4210static struct kernfs_ops cgroup_kf_single_ops = {
4211 .atomic_write_len = PAGE_SIZE,
4212 .open = cgroup_file_open,
4213 .release = cgroup_file_release,
4214 .write = cgroup_file_write,
4215 .poll = cgroup_file_poll,
4216 .seq_show = cgroup_seqfile_show,
4217};
4218
4219static struct kernfs_ops cgroup_kf_ops = {
4220 .atomic_write_len = PAGE_SIZE,
4221 .open = cgroup_file_open,
4222 .release = cgroup_file_release,
4223 .write = cgroup_file_write,
4224 .poll = cgroup_file_poll,
4225 .seq_start = cgroup_seqfile_start,
4226 .seq_next = cgroup_seqfile_next,
4227 .seq_stop = cgroup_seqfile_stop,
4228 .seq_show = cgroup_seqfile_show,
4229};
4230
4231static void cgroup_file_notify_timer(struct timer_list *timer)
4232{
4233 cgroup_file_notify(container_of(timer, struct cgroup_file,
4234 notify_timer));
4235}
4236
4237static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4238 struct cftype *cft)
4239{
4240 char name[CGROUP_FILE_NAME_MAX];
4241 struct kernfs_node *kn;
4242 struct lock_class_key *key = NULL;
4243
4244#ifdef CONFIG_DEBUG_LOCK_ALLOC
4245 key = &cft->lockdep_key;
4246#endif
4247 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4248 cgroup_file_mode(cft),
4249 current_fsuid(), current_fsgid(),
4250 0, cft->kf_ops, cft,
4251 NULL, key);
4252 if (IS_ERR(kn))
4253 return PTR_ERR(kn);
4254
4255 if (cft->file_offset) {
4256 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4257
4258 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4259
4260 spin_lock_irq(&cgroup_file_kn_lock);
4261 cfile->kn = kn;
4262 spin_unlock_irq(&cgroup_file_kn_lock);
4263 }
4264
4265 return 0;
4266}
4267
4268/**
4269 * cgroup_addrm_files - add or remove files to a cgroup directory
4270 * @css: the target css
4271 * @cgrp: the target cgroup (usually css->cgroup)
4272 * @cfts: array of cftypes to be added
4273 * @is_add: whether to add or remove
4274 *
4275 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4276 * For removals, this function never fails.
4277 */
4278static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4279 struct cgroup *cgrp, struct cftype cfts[],
4280 bool is_add)
4281{
4282 struct cftype *cft, *cft_end = NULL;
4283 int ret = 0;
4284
4285 lockdep_assert_held(&cgroup_mutex);
4286
4287restart:
4288 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4289 /* does cft->flags tell us to skip this file on @cgrp? */
4290 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4291 continue;
4292 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4293 continue;
4294 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4295 continue;
4296 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4297 continue;
4298 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4299 continue;
4300 if (is_add) {
4301 ret = cgroup_add_file(css, cgrp, cft);
4302 if (ret) {
4303 pr_warn("%s: failed to add %s, err=%d\n",
4304 __func__, cft->name, ret);
4305 cft_end = cft;
4306 is_add = false;
4307 goto restart;
4308 }
4309 } else {
4310 cgroup_rm_file(cgrp, cft);
4311 }
4312 }
4313 return ret;
4314}
4315
4316static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4317{
4318 struct cgroup_subsys *ss = cfts[0].ss;
4319 struct cgroup *root = &ss->root->cgrp;
4320 struct cgroup_subsys_state *css;
4321 int ret = 0;
4322
4323 lockdep_assert_held(&cgroup_mutex);
4324
4325 /* add/rm files for all cgroups created before */
4326 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4327 struct cgroup *cgrp = css->cgroup;
4328
4329 if (!(css->flags & CSS_VISIBLE))
4330 continue;
4331
4332 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4333 if (ret)
4334 break;
4335 }
4336
4337 if (is_add && !ret)
4338 kernfs_activate(root->kn);
4339 return ret;
4340}
4341
4342static void cgroup_exit_cftypes(struct cftype *cfts)
4343{
4344 struct cftype *cft;
4345
4346 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4347 /* free copy for custom atomic_write_len, see init_cftypes() */
4348 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4349 kfree(cft->kf_ops);
4350 cft->kf_ops = NULL;
4351 cft->ss = NULL;
4352
4353 /* revert flags set by cgroup core while adding @cfts */
4354 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4355 __CFTYPE_ADDED);
4356 }
4357}
4358
4359static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4360{
4361 struct cftype *cft;
4362 int ret = 0;
4363
4364 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4365 struct kernfs_ops *kf_ops;
4366
4367 WARN_ON(cft->ss || cft->kf_ops);
4368
4369 if (cft->flags & __CFTYPE_ADDED) {
4370 ret = -EBUSY;
4371 break;
4372 }
4373
4374 if (cft->seq_start)
4375 kf_ops = &cgroup_kf_ops;
4376 else
4377 kf_ops = &cgroup_kf_single_ops;
4378
4379 /*
4380 * Ugh... if @cft wants a custom max_write_len, we need to
4381 * make a copy of kf_ops to set its atomic_write_len.
4382 */
4383 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4384 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4385 if (!kf_ops) {
4386 ret = -ENOMEM;
4387 break;
4388 }
4389 kf_ops->atomic_write_len = cft->max_write_len;
4390 }
4391
4392 cft->kf_ops = kf_ops;
4393 cft->ss = ss;
4394 cft->flags |= __CFTYPE_ADDED;
4395 }
4396
4397 if (ret)
4398 cgroup_exit_cftypes(cfts);
4399 return ret;
4400}
4401
4402static void cgroup_rm_cftypes_locked(struct cftype *cfts)
4403{
4404 lockdep_assert_held(&cgroup_mutex);
4405
4406 list_del(&cfts->node);
4407 cgroup_apply_cftypes(cfts, false);
4408 cgroup_exit_cftypes(cfts);
4409}
4410
4411/**
4412 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4413 * @cfts: zero-length name terminated array of cftypes
4414 *
4415 * Unregister @cfts. Files described by @cfts are removed from all
4416 * existing cgroups and all future cgroups won't have them either. This
4417 * function can be called anytime whether @cfts' subsys is attached or not.
4418 *
4419 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4420 * registered.
4421 */
4422int cgroup_rm_cftypes(struct cftype *cfts)
4423{
4424 if (!cfts || cfts[0].name[0] == '\0')
4425 return 0;
4426
4427 if (!(cfts[0].flags & __CFTYPE_ADDED))
4428 return -ENOENT;
4429
4430 cgroup_lock();
4431 cgroup_rm_cftypes_locked(cfts);
4432 cgroup_unlock();
4433 return 0;
4434}
4435
4436/**
4437 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4438 * @ss: target cgroup subsystem
4439 * @cfts: zero-length name terminated array of cftypes
4440 *
4441 * Register @cfts to @ss. Files described by @cfts are created for all
4442 * existing cgroups to which @ss is attached and all future cgroups will
4443 * have them too. This function can be called anytime whether @ss is
4444 * attached or not.
4445 *
4446 * Returns 0 on successful registration, -errno on failure. Note that this
4447 * function currently returns 0 as long as @cfts registration is successful
4448 * even if some file creation attempts on existing cgroups fail.
4449 */
4450static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4451{
4452 int ret;
4453
4454 if (!cgroup_ssid_enabled(ss->id))
4455 return 0;
4456
4457 if (!cfts || cfts[0].name[0] == '\0')
4458 return 0;
4459
4460 ret = cgroup_init_cftypes(ss, cfts);
4461 if (ret)
4462 return ret;
4463
4464 cgroup_lock();
4465
4466 list_add_tail(&cfts->node, &ss->cfts);
4467 ret = cgroup_apply_cftypes(cfts, true);
4468 if (ret)
4469 cgroup_rm_cftypes_locked(cfts);
4470
4471 cgroup_unlock();
4472 return ret;
4473}
4474
4475/**
4476 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4477 * @ss: target cgroup subsystem
4478 * @cfts: zero-length name terminated array of cftypes
4479 *
4480 * Similar to cgroup_add_cftypes() but the added files are only used for
4481 * the default hierarchy.
4482 */
4483int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4484{
4485 struct cftype *cft;
4486
4487 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4488 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4489 return cgroup_add_cftypes(ss, cfts);
4490}
4491
4492/**
4493 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4494 * @ss: target cgroup subsystem
4495 * @cfts: zero-length name terminated array of cftypes
4496 *
4497 * Similar to cgroup_add_cftypes() but the added files are only used for
4498 * the legacy hierarchies.
4499 */
4500int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4501{
4502 struct cftype *cft;
4503
4504 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4505 cft->flags |= __CFTYPE_NOT_ON_DFL;
4506 return cgroup_add_cftypes(ss, cfts);
4507}
4508
4509/**
4510 * cgroup_file_notify - generate a file modified event for a cgroup_file
4511 * @cfile: target cgroup_file
4512 *
4513 * @cfile must have been obtained by setting cftype->file_offset.
4514 */
4515void cgroup_file_notify(struct cgroup_file *cfile)
4516{
4517 unsigned long flags;
4518
4519 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4520 if (cfile->kn) {
4521 unsigned long last = cfile->notified_at;
4522 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4523
4524 if (time_in_range(jiffies, last, next)) {
4525 timer_reduce(&cfile->notify_timer, next);
4526 } else {
4527 kernfs_notify(cfile->kn);
4528 cfile->notified_at = jiffies;
4529 }
4530 }
4531 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4532}
4533
4534/**
4535 * cgroup_file_show - show or hide a hidden cgroup file
4536 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4537 * @show: whether to show or hide
4538 */
4539void cgroup_file_show(struct cgroup_file *cfile, bool show)
4540{
4541 struct kernfs_node *kn;
4542
4543 spin_lock_irq(&cgroup_file_kn_lock);
4544 kn = cfile->kn;
4545 kernfs_get(kn);
4546 spin_unlock_irq(&cgroup_file_kn_lock);
4547
4548 if (kn)
4549 kernfs_show(kn, show);
4550
4551 kernfs_put(kn);
4552}
4553
4554/**
4555 * css_next_child - find the next child of a given css
4556 * @pos: the current position (%NULL to initiate traversal)
4557 * @parent: css whose children to walk
4558 *
4559 * This function returns the next child of @parent and should be called
4560 * under either cgroup_mutex or RCU read lock. The only requirement is
4561 * that @parent and @pos are accessible. The next sibling is guaranteed to
4562 * be returned regardless of their states.
4563 *
4564 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4565 * css which finished ->css_online() is guaranteed to be visible in the
4566 * future iterations and will stay visible until the last reference is put.
4567 * A css which hasn't finished ->css_online() or already finished
4568 * ->css_offline() may show up during traversal. It's each subsystem's
4569 * responsibility to synchronize against on/offlining.
4570 */
4571struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4572 struct cgroup_subsys_state *parent)
4573{
4574 struct cgroup_subsys_state *next;
4575
4576 cgroup_assert_mutex_or_rcu_locked();
4577
4578 /*
4579 * @pos could already have been unlinked from the sibling list.
4580 * Once a cgroup is removed, its ->sibling.next is no longer
4581 * updated when its next sibling changes. CSS_RELEASED is set when
4582 * @pos is taken off list, at which time its next pointer is valid,
4583 * and, as releases are serialized, the one pointed to by the next
4584 * pointer is guaranteed to not have started release yet. This
4585 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4586 * critical section, the one pointed to by its next pointer is
4587 * guaranteed to not have finished its RCU grace period even if we
4588 * have dropped rcu_read_lock() in-between iterations.
4589 *
4590 * If @pos has CSS_RELEASED set, its next pointer can't be
4591 * dereferenced; however, as each css is given a monotonically
4592 * increasing unique serial number and always appended to the
4593 * sibling list, the next one can be found by walking the parent's
4594 * children until the first css with higher serial number than
4595 * @pos's. While this path can be slower, it happens iff iteration
4596 * races against release and the race window is very small.
4597 */
4598 if (!pos) {
4599 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4600 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4601 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4602 } else {
4603 list_for_each_entry_rcu(next, &parent->children, sibling,
4604 lockdep_is_held(&cgroup_mutex))
4605 if (next->serial_nr > pos->serial_nr)
4606 break;
4607 }
4608
4609 /*
4610 * @next, if not pointing to the head, can be dereferenced and is
4611 * the next sibling.
4612 */
4613 if (&next->sibling != &parent->children)
4614 return next;
4615 return NULL;
4616}
4617
4618/**
4619 * css_next_descendant_pre - find the next descendant for pre-order walk
4620 * @pos: the current position (%NULL to initiate traversal)
4621 * @root: css whose descendants to walk
4622 *
4623 * To be used by css_for_each_descendant_pre(). Find the next descendant
4624 * to visit for pre-order traversal of @root's descendants. @root is
4625 * included in the iteration and the first node to be visited.
4626 *
4627 * While this function requires cgroup_mutex or RCU read locking, it
4628 * doesn't require the whole traversal to be contained in a single critical
4629 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4630 * This function will return the correct next descendant as long as both @pos
4631 * and @root are accessible and @pos is a descendant of @root.
4632 *
4633 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4634 * css which finished ->css_online() is guaranteed to be visible in the
4635 * future iterations and will stay visible until the last reference is put.
4636 * A css which hasn't finished ->css_online() or already finished
4637 * ->css_offline() may show up during traversal. It's each subsystem's
4638 * responsibility to synchronize against on/offlining.
4639 */
4640struct cgroup_subsys_state *
4641css_next_descendant_pre(struct cgroup_subsys_state *pos,
4642 struct cgroup_subsys_state *root)
4643{
4644 struct cgroup_subsys_state *next;
4645
4646 cgroup_assert_mutex_or_rcu_locked();
4647
4648 /* if first iteration, visit @root */
4649 if (!pos)
4650 return root;
4651
4652 /* visit the first child if exists */
4653 next = css_next_child(NULL, pos);
4654 if (next)
4655 return next;
4656
4657 /* no child, visit my or the closest ancestor's next sibling */
4658 while (pos != root) {
4659 next = css_next_child(pos, pos->parent);
4660 if (next)
4661 return next;
4662 pos = pos->parent;
4663 }
4664
4665 return NULL;
4666}
4667EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4668
4669/**
4670 * css_rightmost_descendant - return the rightmost descendant of a css
4671 * @pos: css of interest
4672 *
4673 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4674 * is returned. This can be used during pre-order traversal to skip
4675 * subtree of @pos.
4676 *
4677 * While this function requires cgroup_mutex or RCU read locking, it
4678 * doesn't require the whole traversal to be contained in a single critical
4679 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4680 * This function will return the correct rightmost descendant as long as @pos
4681 * is accessible.
4682 */
4683struct cgroup_subsys_state *
4684css_rightmost_descendant(struct cgroup_subsys_state *pos)
4685{
4686 struct cgroup_subsys_state *last, *tmp;
4687
4688 cgroup_assert_mutex_or_rcu_locked();
4689
4690 do {
4691 last = pos;
4692 /* ->prev isn't RCU safe, walk ->next till the end */
4693 pos = NULL;
4694 css_for_each_child(tmp, last)
4695 pos = tmp;
4696 } while (pos);
4697
4698 return last;
4699}
4700
4701static struct cgroup_subsys_state *
4702css_leftmost_descendant(struct cgroup_subsys_state *pos)
4703{
4704 struct cgroup_subsys_state *last;
4705
4706 do {
4707 last = pos;
4708 pos = css_next_child(NULL, pos);
4709 } while (pos);
4710
4711 return last;
4712}
4713
4714/**
4715 * css_next_descendant_post - find the next descendant for post-order walk
4716 * @pos: the current position (%NULL to initiate traversal)
4717 * @root: css whose descendants to walk
4718 *
4719 * To be used by css_for_each_descendant_post(). Find the next descendant
4720 * to visit for post-order traversal of @root's descendants. @root is
4721 * included in the iteration and the last node to be visited.
4722 *
4723 * While this function requires cgroup_mutex or RCU read locking, it
4724 * doesn't require the whole traversal to be contained in a single critical
4725 * section. Additionally, it isn't necessary to hold onto a reference to @pos.
4726 * This function will return the correct next descendant as long as both @pos
4727 * and @cgroup are accessible and @pos is a descendant of @cgroup.
4728 *
4729 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4730 * css which finished ->css_online() is guaranteed to be visible in the
4731 * future iterations and will stay visible until the last reference is put.
4732 * A css which hasn't finished ->css_online() or already finished
4733 * ->css_offline() may show up during traversal. It's each subsystem's
4734 * responsibility to synchronize against on/offlining.
4735 */
4736struct cgroup_subsys_state *
4737css_next_descendant_post(struct cgroup_subsys_state *pos,
4738 struct cgroup_subsys_state *root)
4739{
4740 struct cgroup_subsys_state *next;
4741
4742 cgroup_assert_mutex_or_rcu_locked();
4743
4744 /* if first iteration, visit leftmost descendant which may be @root */
4745 if (!pos)
4746 return css_leftmost_descendant(root);
4747
4748 /* if we visited @root, we're done */
4749 if (pos == root)
4750 return NULL;
4751
4752 /* if there's an unvisited sibling, visit its leftmost descendant */
4753 next = css_next_child(pos, pos->parent);
4754 if (next)
4755 return css_leftmost_descendant(next);
4756
4757 /* no sibling left, visit parent */
4758 return pos->parent;
4759}
4760
4761/**
4762 * css_has_online_children - does a css have online children
4763 * @css: the target css
4764 *
4765 * Returns %true if @css has any online children; otherwise, %false. This
4766 * function can be called from any context but the caller is responsible
4767 * for synchronizing against on/offlining as necessary.
4768 */
4769bool css_has_online_children(struct cgroup_subsys_state *css)
4770{
4771 struct cgroup_subsys_state *child;
4772 bool ret = false;
4773
4774 rcu_read_lock();
4775 css_for_each_child(child, css) {
4776 if (child->flags & CSS_ONLINE) {
4777 ret = true;
4778 break;
4779 }
4780 }
4781 rcu_read_unlock();
4782 return ret;
4783}
4784
4785static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4786{
4787 struct list_head *l;
4788 struct cgrp_cset_link *link;
4789 struct css_set *cset;
4790
4791 lockdep_assert_held(&css_set_lock);
4792
4793 /* find the next threaded cset */
4794 if (it->tcset_pos) {
4795 l = it->tcset_pos->next;
4796
4797 if (l != it->tcset_head) {
4798 it->tcset_pos = l;
4799 return container_of(l, struct css_set,
4800 threaded_csets_node);
4801 }
4802
4803 it->tcset_pos = NULL;
4804 }
4805
4806 /* find the next cset */
4807 l = it->cset_pos;
4808 l = l->next;
4809 if (l == it->cset_head) {
4810 it->cset_pos = NULL;
4811 return NULL;
4812 }
4813
4814 if (it->ss) {
4815 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4816 } else {
4817 link = list_entry(l, struct cgrp_cset_link, cset_link);
4818 cset = link->cset;
4819 }
4820
4821 it->cset_pos = l;
4822
4823 /* initialize threaded css_set walking */
4824 if (it->flags & CSS_TASK_ITER_THREADED) {
4825 if (it->cur_dcset)
4826 put_css_set_locked(it->cur_dcset);
4827 it->cur_dcset = cset;
4828 get_css_set(cset);
4829
4830 it->tcset_head = &cset->threaded_csets;
4831 it->tcset_pos = &cset->threaded_csets;
4832 }
4833
4834 return cset;
4835}
4836
4837/**
4838 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4839 * @it: the iterator to advance
4840 *
4841 * Advance @it to the next css_set to walk.
4842 */
4843static void css_task_iter_advance_css_set(struct css_task_iter *it)
4844{
4845 struct css_set *cset;
4846
4847 lockdep_assert_held(&css_set_lock);
4848
4849 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4850 while ((cset = css_task_iter_next_css_set(it))) {
4851 if (!list_empty(&cset->tasks)) {
4852 it->cur_tasks_head = &cset->tasks;
4853 break;
4854 } else if (!list_empty(&cset->mg_tasks)) {
4855 it->cur_tasks_head = &cset->mg_tasks;
4856 break;
4857 } else if (!list_empty(&cset->dying_tasks)) {
4858 it->cur_tasks_head = &cset->dying_tasks;
4859 break;
4860 }
4861 }
4862 if (!cset) {
4863 it->task_pos = NULL;
4864 return;
4865 }
4866 it->task_pos = it->cur_tasks_head->next;
4867
4868 /*
4869 * We don't keep css_sets locked across iteration steps and thus
4870 * need to take steps to ensure that iteration can be resumed after
4871 * the lock is re-acquired. Iteration is performed at two levels -
4872 * css_sets and tasks in them.
4873 *
4874 * Once created, a css_set never leaves its cgroup lists, so a
4875 * pinned css_set is guaranteed to stay put and we can resume
4876 * iteration afterwards.
4877 *
4878 * Tasks may leave @cset across iteration steps. This is resolved
4879 * by registering each iterator with the css_set currently being
4880 * walked and making css_set_move_task() advance iterators whose
4881 * next task is leaving.
4882 */
4883 if (it->cur_cset) {
4884 list_del(&it->iters_node);
4885 put_css_set_locked(it->cur_cset);
4886 }
4887 get_css_set(cset);
4888 it->cur_cset = cset;
4889 list_add(&it->iters_node, &cset->task_iters);
4890}
4891
4892static void css_task_iter_skip(struct css_task_iter *it,
4893 struct task_struct *task)
4894{
4895 lockdep_assert_held(&css_set_lock);
4896
4897 if (it->task_pos == &task->cg_list) {
4898 it->task_pos = it->task_pos->next;
4899 it->flags |= CSS_TASK_ITER_SKIPPED;
4900 }
4901}
4902
4903static void css_task_iter_advance(struct css_task_iter *it)
4904{
4905 struct task_struct *task;
4906
4907 lockdep_assert_held(&css_set_lock);
4908repeat:
4909 if (it->task_pos) {
4910 /*
4911 * Advance iterator to find next entry. We go through cset
4912 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4913 * the next cset.
4914 */
4915 if (it->flags & CSS_TASK_ITER_SKIPPED)
4916 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4917 else
4918 it->task_pos = it->task_pos->next;
4919
4920 if (it->task_pos == &it->cur_cset->tasks) {
4921 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4922 it->task_pos = it->cur_tasks_head->next;
4923 }
4924 if (it->task_pos == &it->cur_cset->mg_tasks) {
4925 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4926 it->task_pos = it->cur_tasks_head->next;
4927 }
4928 if (it->task_pos == &it->cur_cset->dying_tasks)
4929 css_task_iter_advance_css_set(it);
4930 } else {
4931 /* called from start, proceed to the first cset */
4932 css_task_iter_advance_css_set(it);
4933 }
4934
4935 if (!it->task_pos)
4936 return;
4937
4938 task = list_entry(it->task_pos, struct task_struct, cg_list);
4939
4940 if (it->flags & CSS_TASK_ITER_PROCS) {
4941 /* if PROCS, skip over tasks which aren't group leaders */
4942 if (!thread_group_leader(task))
4943 goto repeat;
4944
4945 /* and dying leaders w/o live member threads */
4946 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4947 !atomic_read(&task->signal->live))
4948 goto repeat;
4949 } else {
4950 /* skip all dying ones */
4951 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4952 goto repeat;
4953 }
4954}
4955
4956/**
4957 * css_task_iter_start - initiate task iteration
4958 * @css: the css to walk tasks of
4959 * @flags: CSS_TASK_ITER_* flags
4960 * @it: the task iterator to use
4961 *
4962 * Initiate iteration through the tasks of @css. The caller can call
4963 * css_task_iter_next() to walk through the tasks until the function
4964 * returns NULL. On completion of iteration, css_task_iter_end() must be
4965 * called.
4966 */
4967void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4968 struct css_task_iter *it)
4969{
4970 unsigned long irqflags;
4971
4972 memset(it, 0, sizeof(*it));
4973
4974 spin_lock_irqsave(&css_set_lock, irqflags);
4975
4976 it->ss = css->ss;
4977 it->flags = flags;
4978
4979 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4980 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4981 else
4982 it->cset_pos = &css->cgroup->cset_links;
4983
4984 it->cset_head = it->cset_pos;
4985
4986 css_task_iter_advance(it);
4987
4988 spin_unlock_irqrestore(&css_set_lock, irqflags);
4989}
4990
4991/**
4992 * css_task_iter_next - return the next task for the iterator
4993 * @it: the task iterator being iterated
4994 *
4995 * The "next" function for task iteration. @it should have been
4996 * initialized via css_task_iter_start(). Returns NULL when the iteration
4997 * reaches the end.
4998 */
4999struct task_struct *css_task_iter_next(struct css_task_iter *it)
5000{
5001 unsigned long irqflags;
5002
5003 if (it->cur_task) {
5004 put_task_struct(it->cur_task);
5005 it->cur_task = NULL;
5006 }
5007
5008 spin_lock_irqsave(&css_set_lock, irqflags);
5009
5010 /* @it may be half-advanced by skips, finish advancing */
5011 if (it->flags & CSS_TASK_ITER_SKIPPED)
5012 css_task_iter_advance(it);
5013
5014 if (it->task_pos) {
5015 it->cur_task = list_entry(it->task_pos, struct task_struct,
5016 cg_list);
5017 get_task_struct(it->cur_task);
5018 css_task_iter_advance(it);
5019 }
5020
5021 spin_unlock_irqrestore(&css_set_lock, irqflags);
5022
5023 return it->cur_task;
5024}
5025
5026/**
5027 * css_task_iter_end - finish task iteration
5028 * @it: the task iterator to finish
5029 *
5030 * Finish task iteration started by css_task_iter_start().
5031 */
5032void css_task_iter_end(struct css_task_iter *it)
5033{
5034 unsigned long irqflags;
5035
5036 if (it->cur_cset) {
5037 spin_lock_irqsave(&css_set_lock, irqflags);
5038 list_del(&it->iters_node);
5039 put_css_set_locked(it->cur_cset);
5040 spin_unlock_irqrestore(&css_set_lock, irqflags);
5041 }
5042
5043 if (it->cur_dcset)
5044 put_css_set(it->cur_dcset);
5045
5046 if (it->cur_task)
5047 put_task_struct(it->cur_task);
5048}
5049
5050static void cgroup_procs_release(struct kernfs_open_file *of)
5051{
5052 struct cgroup_file_ctx *ctx = of->priv;
5053
5054 if (ctx->procs.started)
5055 css_task_iter_end(&ctx->procs.iter);
5056}
5057
5058static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
5059{
5060 struct kernfs_open_file *of = s->private;
5061 struct cgroup_file_ctx *ctx = of->priv;
5062
5063 if (pos)
5064 (*pos)++;
5065
5066 return css_task_iter_next(&ctx->procs.iter);
5067}
5068
5069static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5070 unsigned int iter_flags)
5071{
5072 struct kernfs_open_file *of = s->private;
5073 struct cgroup *cgrp = seq_css(s)->cgroup;
5074 struct cgroup_file_ctx *ctx = of->priv;
5075 struct css_task_iter *it = &ctx->procs.iter;
5076
5077 /*
5078 * When a seq_file is seeked, it's always traversed sequentially
5079 * from position 0, so we can simply keep iterating on !0 *pos.
5080 */
5081 if (!ctx->procs.started) {
5082 if (WARN_ON_ONCE((*pos)))
5083 return ERR_PTR(-EINVAL);
5084 css_task_iter_start(&cgrp->self, iter_flags, it);
5085 ctx->procs.started = true;
5086 } else if (!(*pos)) {
5087 css_task_iter_end(it);
5088 css_task_iter_start(&cgrp->self, iter_flags, it);
5089 } else
5090 return it->cur_task;
5091
5092 return cgroup_procs_next(s, NULL, NULL);
5093}
5094
5095static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5096{
5097 struct cgroup *cgrp = seq_css(s)->cgroup;
5098
5099 /*
5100 * All processes of a threaded subtree belong to the domain cgroup
5101 * of the subtree. Only threads can be distributed across the
5102 * subtree. Reject reads on cgroup.procs in the subtree proper.
5103 * They're always empty anyway.
5104 */
5105 if (cgroup_is_threaded(cgrp))
5106 return ERR_PTR(-EOPNOTSUPP);
5107
5108 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5109 CSS_TASK_ITER_THREADED);
5110}
5111
5112static int cgroup_procs_show(struct seq_file *s, void *v)
5113{
5114 seq_printf(s, "%d\n", task_pid_vnr(v));
5115 return 0;
5116}
5117
5118static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5119{
5120 int ret;
5121 struct inode *inode;
5122
5123 lockdep_assert_held(&cgroup_mutex);
5124
5125 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5126 if (!inode)
5127 return -ENOMEM;
5128
5129 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5130 iput(inode);
5131 return ret;
5132}
5133
5134static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5135 struct cgroup *dst_cgrp,
5136 struct super_block *sb,
5137 struct cgroup_namespace *ns)
5138{
5139 struct cgroup *com_cgrp = src_cgrp;
5140 int ret;
5141
5142 lockdep_assert_held(&cgroup_mutex);
5143
5144 /* find the common ancestor */
5145 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5146 com_cgrp = cgroup_parent(com_cgrp);
5147
5148 /* %current should be authorized to migrate to the common ancestor */
5149 ret = cgroup_may_write(com_cgrp, sb);
5150 if (ret)
5151 return ret;
5152
5153 /*
5154 * If namespaces are delegation boundaries, %current must be able
5155 * to see both source and destination cgroups from its namespace.
5156 */
5157 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5158 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5159 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5160 return -ENOENT;
5161
5162 return 0;
5163}
5164
5165static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5166 struct cgroup *dst_cgrp,
5167 struct super_block *sb, bool threadgroup,
5168 struct cgroup_namespace *ns)
5169{
5170 int ret = 0;
5171
5172 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5173 if (ret)
5174 return ret;
5175
5176 ret = cgroup_migrate_vet_dst(dst_cgrp);
5177 if (ret)
5178 return ret;
5179
5180 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5181 ret = -EOPNOTSUPP;
5182
5183 return ret;
5184}
5185
5186static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5187 bool threadgroup)
5188{
5189 struct cgroup_file_ctx *ctx = of->priv;
5190 struct cgroup *src_cgrp, *dst_cgrp;
5191 struct task_struct *task;
5192 const struct cred *saved_cred;
5193 ssize_t ret;
5194 bool threadgroup_locked;
5195
5196 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5197 if (!dst_cgrp)
5198 return -ENODEV;
5199
5200 task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
5201 ret = PTR_ERR_OR_ZERO(task);
5202 if (ret)
5203 goto out_unlock;
5204
5205 /* find the source cgroup */
5206 spin_lock_irq(&css_set_lock);
5207 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5208 spin_unlock_irq(&css_set_lock);
5209
5210 /*
5211 * Process and thread migrations follow same delegation rule. Check
5212 * permissions using the credentials from file open to protect against
5213 * inherited fd attacks.
5214 */
5215 saved_cred = override_creds(of->file->f_cred);
5216 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5217 of->file->f_path.dentry->d_sb,
5218 threadgroup, ctx->ns);
5219 revert_creds(saved_cred);
5220 if (ret)
5221 goto out_finish;
5222
5223 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5224
5225out_finish:
5226 cgroup_procs_write_finish(task, threadgroup_locked);
5227out_unlock:
5228 cgroup_kn_unlock(of->kn);
5229
5230 return ret;
5231}
5232
5233static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5234 char *buf, size_t nbytes, loff_t off)
5235{
5236 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5237}
5238
5239static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5240{
5241 return __cgroup_procs_start(s, pos, 0);
5242}
5243
5244static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5245 char *buf, size_t nbytes, loff_t off)
5246{
5247 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5248}
5249
5250/* cgroup core interface files for the default hierarchy */
5251static struct cftype cgroup_base_files[] = {
5252 {
5253 .name = "cgroup.type",
5254 .flags = CFTYPE_NOT_ON_ROOT,
5255 .seq_show = cgroup_type_show,
5256 .write = cgroup_type_write,
5257 },
5258 {
5259 .name = "cgroup.procs",
5260 .flags = CFTYPE_NS_DELEGATABLE,
5261 .file_offset = offsetof(struct cgroup, procs_file),
5262 .release = cgroup_procs_release,
5263 .seq_start = cgroup_procs_start,
5264 .seq_next = cgroup_procs_next,
5265 .seq_show = cgroup_procs_show,
5266 .write = cgroup_procs_write,
5267 },
5268 {
5269 .name = "cgroup.threads",
5270 .flags = CFTYPE_NS_DELEGATABLE,
5271 .release = cgroup_procs_release,
5272 .seq_start = cgroup_threads_start,
5273 .seq_next = cgroup_procs_next,
5274 .seq_show = cgroup_procs_show,
5275 .write = cgroup_threads_write,
5276 },
5277 {
5278 .name = "cgroup.controllers",
5279 .seq_show = cgroup_controllers_show,
5280 },
5281 {
5282 .name = "cgroup.subtree_control",
5283 .flags = CFTYPE_NS_DELEGATABLE,
5284 .seq_show = cgroup_subtree_control_show,
5285 .write = cgroup_subtree_control_write,
5286 },
5287 {
5288 .name = "cgroup.events",
5289 .flags = CFTYPE_NOT_ON_ROOT,
5290 .file_offset = offsetof(struct cgroup, events_file),
5291 .seq_show = cgroup_events_show,
5292 },
5293 {
5294 .name = "cgroup.max.descendants",
5295 .seq_show = cgroup_max_descendants_show,
5296 .write = cgroup_max_descendants_write,
5297 },
5298 {
5299 .name = "cgroup.max.depth",
5300 .seq_show = cgroup_max_depth_show,
5301 .write = cgroup_max_depth_write,
5302 },
5303 {
5304 .name = "cgroup.stat",
5305 .seq_show = cgroup_stat_show,
5306 },
5307 {
5308 .name = "cgroup.freeze",
5309 .flags = CFTYPE_NOT_ON_ROOT,
5310 .seq_show = cgroup_freeze_show,
5311 .write = cgroup_freeze_write,
5312 },
5313 {
5314 .name = "cgroup.kill",
5315 .flags = CFTYPE_NOT_ON_ROOT,
5316 .write = cgroup_kill_write,
5317 },
5318 {
5319 .name = "cpu.stat",
5320 .seq_show = cpu_stat_show,
5321 },
5322 {
5323 .name = "cpu.stat.local",
5324 .seq_show = cpu_local_stat_show,
5325 },
5326 { } /* terminate */
5327};
5328
5329static struct cftype cgroup_psi_files[] = {
5330#ifdef CONFIG_PSI
5331 {
5332 .name = "io.pressure",
5333 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5334 .seq_show = cgroup_io_pressure_show,
5335 .write = cgroup_io_pressure_write,
5336 .poll = cgroup_pressure_poll,
5337 .release = cgroup_pressure_release,
5338 },
5339 {
5340 .name = "memory.pressure",
5341 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5342 .seq_show = cgroup_memory_pressure_show,
5343 .write = cgroup_memory_pressure_write,
5344 .poll = cgroup_pressure_poll,
5345 .release = cgroup_pressure_release,
5346 },
5347 {
5348 .name = "cpu.pressure",
5349 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5350 .seq_show = cgroup_cpu_pressure_show,
5351 .write = cgroup_cpu_pressure_write,
5352 .poll = cgroup_pressure_poll,
5353 .release = cgroup_pressure_release,
5354 },
5355#ifdef CONFIG_IRQ_TIME_ACCOUNTING
5356 {
5357 .name = "irq.pressure",
5358 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5359 .seq_show = cgroup_irq_pressure_show,
5360 .write = cgroup_irq_pressure_write,
5361 .poll = cgroup_pressure_poll,
5362 .release = cgroup_pressure_release,
5363 },
5364#endif
5365 {
5366 .name = "cgroup.pressure",
5367 .seq_show = cgroup_pressure_show,
5368 .write = cgroup_pressure_write,
5369 },
5370#endif /* CONFIG_PSI */
5371 { } /* terminate */
5372};
5373
5374/*
5375 * css destruction is four-stage process.
5376 *
5377 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5378 * Implemented in kill_css().
5379 *
5380 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5381 * and thus css_tryget_online() is guaranteed to fail, the css can be
5382 * offlined by invoking offline_css(). After offlining, the base ref is
5383 * put. Implemented in css_killed_work_fn().
5384 *
5385 * 3. When the percpu_ref reaches zero, the only possible remaining
5386 * accessors are inside RCU read sections. css_release() schedules the
5387 * RCU callback.
5388 *
5389 * 4. After the grace period, the css can be freed. Implemented in
5390 * css_free_rwork_fn().
5391 *
5392 * It is actually hairier because both step 2 and 4 require process context
5393 * and thus involve punting to css->destroy_work adding two additional
5394 * steps to the already complex sequence.
5395 */
5396static void css_free_rwork_fn(struct work_struct *work)
5397{
5398 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5399 struct cgroup_subsys_state, destroy_rwork);
5400 struct cgroup_subsys *ss = css->ss;
5401 struct cgroup *cgrp = css->cgroup;
5402
5403 percpu_ref_exit(&css->refcnt);
5404
5405 if (ss) {
5406 /* css free path */
5407 struct cgroup_subsys_state *parent = css->parent;
5408 int id = css->id;
5409
5410 ss->css_free(css);
5411 cgroup_idr_remove(&ss->css_idr, id);
5412 cgroup_put(cgrp);
5413
5414 if (parent)
5415 css_put(parent);
5416 } else {
5417 /* cgroup free path */
5418 atomic_dec(&cgrp->root->nr_cgrps);
5419 if (!cgroup_on_dfl(cgrp))
5420 cgroup1_pidlist_destroy_all(cgrp);
5421 cancel_work_sync(&cgrp->release_agent_work);
5422 bpf_cgrp_storage_free(cgrp);
5423
5424 if (cgroup_parent(cgrp)) {
5425 /*
5426 * We get a ref to the parent, and put the ref when
5427 * this cgroup is being freed, so it's guaranteed
5428 * that the parent won't be destroyed before its
5429 * children.
5430 */
5431 cgroup_put(cgroup_parent(cgrp));
5432 kernfs_put(cgrp->kn);
5433 psi_cgroup_free(cgrp);
5434 cgroup_rstat_exit(cgrp);
5435 kfree(cgrp);
5436 } else {
5437 /*
5438 * This is root cgroup's refcnt reaching zero,
5439 * which indicates that the root should be
5440 * released.
5441 */
5442 cgroup_destroy_root(cgrp->root);
5443 }
5444 }
5445}
5446
5447static void css_release_work_fn(struct work_struct *work)
5448{
5449 struct cgroup_subsys_state *css =
5450 container_of(work, struct cgroup_subsys_state, destroy_work);
5451 struct cgroup_subsys *ss = css->ss;
5452 struct cgroup *cgrp = css->cgroup;
5453
5454 cgroup_lock();
5455
5456 css->flags |= CSS_RELEASED;
5457 list_del_rcu(&css->sibling);
5458
5459 if (ss) {
5460 struct cgroup *parent_cgrp;
5461
5462 /* css release path */
5463 if (!list_empty(&css->rstat_css_node)) {
5464 cgroup_rstat_flush(cgrp);
5465 list_del_rcu(&css->rstat_css_node);
5466 }
5467
5468 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5469 if (ss->css_released)
5470 ss->css_released(css);
5471
5472 cgrp->nr_dying_subsys[ss->id]--;
5473 /*
5474 * When a css is released and ready to be freed, its
5475 * nr_descendants must be zero. However, the corresponding
5476 * cgrp->nr_dying_subsys[ss->id] may not be 0 if a subsystem
5477 * is activated and deactivated multiple times with one or
5478 * more of its previous activation leaving behind dying csses.
5479 */
5480 WARN_ON_ONCE(css->nr_descendants);
5481 parent_cgrp = cgroup_parent(cgrp);
5482 while (parent_cgrp) {
5483 parent_cgrp->nr_dying_subsys[ss->id]--;
5484 parent_cgrp = cgroup_parent(parent_cgrp);
5485 }
5486 } else {
5487 struct cgroup *tcgrp;
5488
5489 /* cgroup release path */
5490 TRACE_CGROUP_PATH(release, cgrp);
5491
5492 cgroup_rstat_flush(cgrp);
5493
5494 spin_lock_irq(&css_set_lock);
5495 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5496 tcgrp = cgroup_parent(tcgrp))
5497 tcgrp->nr_dying_descendants--;
5498 spin_unlock_irq(&css_set_lock);
5499
5500 /*
5501 * There are two control paths which try to determine
5502 * cgroup from dentry without going through kernfs -
5503 * cgroupstats_build() and css_tryget_online_from_dir().
5504 * Those are supported by RCU protecting clearing of
5505 * cgrp->kn->priv backpointer.
5506 */
5507 if (cgrp->kn)
5508 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5509 NULL);
5510 }
5511
5512 cgroup_unlock();
5513
5514 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5515 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5516}
5517
5518static void css_release(struct percpu_ref *ref)
5519{
5520 struct cgroup_subsys_state *css =
5521 container_of(ref, struct cgroup_subsys_state, refcnt);
5522
5523 INIT_WORK(&css->destroy_work, css_release_work_fn);
5524 queue_work(cgroup_destroy_wq, &css->destroy_work);
5525}
5526
5527static void init_and_link_css(struct cgroup_subsys_state *css,
5528 struct cgroup_subsys *ss, struct cgroup *cgrp)
5529{
5530 lockdep_assert_held(&cgroup_mutex);
5531
5532 cgroup_get_live(cgrp);
5533
5534 memset(css, 0, sizeof(*css));
5535 css->cgroup = cgrp;
5536 css->ss = ss;
5537 css->id = -1;
5538 INIT_LIST_HEAD(&css->sibling);
5539 INIT_LIST_HEAD(&css->children);
5540 INIT_LIST_HEAD(&css->rstat_css_node);
5541 css->serial_nr = css_serial_nr_next++;
5542 atomic_set(&css->online_cnt, 0);
5543
5544 if (cgroup_parent(cgrp)) {
5545 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5546 css_get(css->parent);
5547 }
5548
5549 if (ss->css_rstat_flush)
5550 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5551
5552 BUG_ON(cgroup_css(cgrp, ss));
5553}
5554
5555/* invoke ->css_online() on a new CSS and mark it online if successful */
5556static int online_css(struct cgroup_subsys_state *css)
5557{
5558 struct cgroup_subsys *ss = css->ss;
5559 int ret = 0;
5560
5561 lockdep_assert_held(&cgroup_mutex);
5562
5563 if (ss->css_online)
5564 ret = ss->css_online(css);
5565 if (!ret) {
5566 css->flags |= CSS_ONLINE;
5567 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5568
5569 atomic_inc(&css->online_cnt);
5570 if (css->parent) {
5571 atomic_inc(&css->parent->online_cnt);
5572 while ((css = css->parent))
5573 css->nr_descendants++;
5574 }
5575 }
5576 return ret;
5577}
5578
5579/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5580static void offline_css(struct cgroup_subsys_state *css)
5581{
5582 struct cgroup_subsys *ss = css->ss;
5583
5584 lockdep_assert_held(&cgroup_mutex);
5585
5586 if (!(css->flags & CSS_ONLINE))
5587 return;
5588
5589 if (ss->css_offline)
5590 ss->css_offline(css);
5591
5592 css->flags &= ~CSS_ONLINE;
5593 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5594
5595 wake_up_all(&css->cgroup->offline_waitq);
5596
5597 css->cgroup->nr_dying_subsys[ss->id]++;
5598 /*
5599 * Parent css and cgroup cannot be freed until after the freeing
5600 * of child css, see css_free_rwork_fn().
5601 */
5602 while ((css = css->parent)) {
5603 css->nr_descendants--;
5604 css->cgroup->nr_dying_subsys[ss->id]++;
5605 }
5606}
5607
5608/**
5609 * css_create - create a cgroup_subsys_state
5610 * @cgrp: the cgroup new css will be associated with
5611 * @ss: the subsys of new css
5612 *
5613 * Create a new css associated with @cgrp - @ss pair. On success, the new
5614 * css is online and installed in @cgrp. This function doesn't create the
5615 * interface files. Returns 0 on success, -errno on failure.
5616 */
5617static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5618 struct cgroup_subsys *ss)
5619{
5620 struct cgroup *parent = cgroup_parent(cgrp);
5621 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5622 struct cgroup_subsys_state *css;
5623 int err;
5624
5625 lockdep_assert_held(&cgroup_mutex);
5626
5627 css = ss->css_alloc(parent_css);
5628 if (!css)
5629 css = ERR_PTR(-ENOMEM);
5630 if (IS_ERR(css))
5631 return css;
5632
5633 init_and_link_css(css, ss, cgrp);
5634
5635 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5636 if (err)
5637 goto err_free_css;
5638
5639 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5640 if (err < 0)
5641 goto err_free_css;
5642 css->id = err;
5643
5644 /* @css is ready to be brought online now, make it visible */
5645 list_add_tail_rcu(&css->sibling, &parent_css->children);
5646 cgroup_idr_replace(&ss->css_idr, css, css->id);
5647
5648 err = online_css(css);
5649 if (err)
5650 goto err_list_del;
5651
5652 return css;
5653
5654err_list_del:
5655 list_del_rcu(&css->sibling);
5656err_free_css:
5657 list_del_rcu(&css->rstat_css_node);
5658 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5659 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5660 return ERR_PTR(err);
5661}
5662
5663/*
5664 * The returned cgroup is fully initialized including its control mask, but
5665 * it doesn't have the control mask applied.
5666 */
5667static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5668 umode_t mode)
5669{
5670 struct cgroup_root *root = parent->root;
5671 struct cgroup *cgrp, *tcgrp;
5672 struct kernfs_node *kn;
5673 int level = parent->level + 1;
5674 int ret;
5675
5676 /* allocate the cgroup and its ID, 0 is reserved for the root */
5677 cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5678 if (!cgrp)
5679 return ERR_PTR(-ENOMEM);
5680
5681 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5682 if (ret)
5683 goto out_free_cgrp;
5684
5685 ret = cgroup_rstat_init(cgrp);
5686 if (ret)
5687 goto out_cancel_ref;
5688
5689 /* create the directory */
5690 kn = kernfs_create_dir_ns(parent->kn, name, mode,
5691 current_fsuid(), current_fsgid(),
5692 cgrp, NULL);
5693 if (IS_ERR(kn)) {
5694 ret = PTR_ERR(kn);
5695 goto out_stat_exit;
5696 }
5697 cgrp->kn = kn;
5698
5699 init_cgroup_housekeeping(cgrp);
5700
5701 cgrp->self.parent = &parent->self;
5702 cgrp->root = root;
5703 cgrp->level = level;
5704
5705 ret = psi_cgroup_alloc(cgrp);
5706 if (ret)
5707 goto out_kernfs_remove;
5708
5709 if (cgrp->root == &cgrp_dfl_root) {
5710 ret = cgroup_bpf_inherit(cgrp);
5711 if (ret)
5712 goto out_psi_free;
5713 }
5714
5715 /*
5716 * New cgroup inherits effective freeze counter, and
5717 * if the parent has to be frozen, the child has too.
5718 */
5719 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5720 if (cgrp->freezer.e_freeze) {
5721 /*
5722 * Set the CGRP_FREEZE flag, so when a process will be
5723 * attached to the child cgroup, it will become frozen.
5724 * At this point the new cgroup is unpopulated, so we can
5725 * consider it frozen immediately.
5726 */
5727 set_bit(CGRP_FREEZE, &cgrp->flags);
5728 set_bit(CGRP_FROZEN, &cgrp->flags);
5729 }
5730
5731 spin_lock_irq(&css_set_lock);
5732 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5733 cgrp->ancestors[tcgrp->level] = tcgrp;
5734
5735 if (tcgrp != cgrp) {
5736 tcgrp->nr_descendants++;
5737
5738 /*
5739 * If the new cgroup is frozen, all ancestor cgroups
5740 * get a new frozen descendant, but their state can't
5741 * change because of this.
5742 */
5743 if (cgrp->freezer.e_freeze)
5744 tcgrp->freezer.nr_frozen_descendants++;
5745 }
5746 }
5747 spin_unlock_irq(&css_set_lock);
5748
5749 if (notify_on_release(parent))
5750 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5751
5752 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5753 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5754
5755 cgrp->self.serial_nr = css_serial_nr_next++;
5756
5757 /* allocation complete, commit to creation */
5758 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5759 atomic_inc(&root->nr_cgrps);
5760 cgroup_get_live(parent);
5761
5762 /*
5763 * On the default hierarchy, a child doesn't automatically inherit
5764 * subtree_control from the parent. Each is configured manually.
5765 */
5766 if (!cgroup_on_dfl(cgrp))
5767 cgrp->subtree_control = cgroup_control(cgrp);
5768
5769 cgroup_propagate_control(cgrp);
5770
5771 return cgrp;
5772
5773out_psi_free:
5774 psi_cgroup_free(cgrp);
5775out_kernfs_remove:
5776 kernfs_remove(cgrp->kn);
5777out_stat_exit:
5778 cgroup_rstat_exit(cgrp);
5779out_cancel_ref:
5780 percpu_ref_exit(&cgrp->self.refcnt);
5781out_free_cgrp:
5782 kfree(cgrp);
5783 return ERR_PTR(ret);
5784}
5785
5786static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5787{
5788 struct cgroup *cgroup;
5789 int ret = false;
5790 int level = 0;
5791
5792 lockdep_assert_held(&cgroup_mutex);
5793
5794 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5795 if (cgroup->nr_descendants >= cgroup->max_descendants)
5796 goto fail;
5797
5798 if (level >= cgroup->max_depth)
5799 goto fail;
5800
5801 level++;
5802 }
5803
5804 ret = true;
5805fail:
5806 return ret;
5807}
5808
5809int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5810{
5811 struct cgroup *parent, *cgrp;
5812 int ret;
5813
5814 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5815 if (strchr(name, '\n'))
5816 return -EINVAL;
5817
5818 parent = cgroup_kn_lock_live(parent_kn, false);
5819 if (!parent)
5820 return -ENODEV;
5821
5822 if (!cgroup_check_hierarchy_limits(parent)) {
5823 ret = -EAGAIN;
5824 goto out_unlock;
5825 }
5826
5827 cgrp = cgroup_create(parent, name, mode);
5828 if (IS_ERR(cgrp)) {
5829 ret = PTR_ERR(cgrp);
5830 goto out_unlock;
5831 }
5832
5833 /*
5834 * This extra ref will be put in cgroup_free_fn() and guarantees
5835 * that @cgrp->kn is always accessible.
5836 */
5837 kernfs_get(cgrp->kn);
5838
5839 ret = css_populate_dir(&cgrp->self);
5840 if (ret)
5841 goto out_destroy;
5842
5843 ret = cgroup_apply_control_enable(cgrp);
5844 if (ret)
5845 goto out_destroy;
5846
5847 TRACE_CGROUP_PATH(mkdir, cgrp);
5848
5849 /* let's create and online css's */
5850 kernfs_activate(cgrp->kn);
5851
5852 ret = 0;
5853 goto out_unlock;
5854
5855out_destroy:
5856 cgroup_destroy_locked(cgrp);
5857out_unlock:
5858 cgroup_kn_unlock(parent_kn);
5859 return ret;
5860}
5861
5862/*
5863 * This is called when the refcnt of a css is confirmed to be killed.
5864 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5865 * initiate destruction and put the css ref from kill_css().
5866 */
5867static void css_killed_work_fn(struct work_struct *work)
5868{
5869 struct cgroup_subsys_state *css =
5870 container_of(work, struct cgroup_subsys_state, destroy_work);
5871
5872 cgroup_lock();
5873
5874 do {
5875 offline_css(css);
5876 css_put(css);
5877 /* @css can't go away while we're holding cgroup_mutex */
5878 css = css->parent;
5879 } while (css && atomic_dec_and_test(&css->online_cnt));
5880
5881 cgroup_unlock();
5882}
5883
5884/* css kill confirmation processing requires process context, bounce */
5885static void css_killed_ref_fn(struct percpu_ref *ref)
5886{
5887 struct cgroup_subsys_state *css =
5888 container_of(ref, struct cgroup_subsys_state, refcnt);
5889
5890 if (atomic_dec_and_test(&css->online_cnt)) {
5891 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5892 queue_work(cgroup_destroy_wq, &css->destroy_work);
5893 }
5894}
5895
5896/**
5897 * kill_css - destroy a css
5898 * @css: css to destroy
5899 *
5900 * This function initiates destruction of @css by removing cgroup interface
5901 * files and putting its base reference. ->css_offline() will be invoked
5902 * asynchronously once css_tryget_online() is guaranteed to fail and when
5903 * the reference count reaches zero, @css will be released.
5904 */
5905static void kill_css(struct cgroup_subsys_state *css)
5906{
5907 lockdep_assert_held(&cgroup_mutex);
5908
5909 if (css->flags & CSS_DYING)
5910 return;
5911
5912 css->flags |= CSS_DYING;
5913
5914 /*
5915 * This must happen before css is disassociated with its cgroup.
5916 * See seq_css() for details.
5917 */
5918 css_clear_dir(css);
5919
5920 /*
5921 * Killing would put the base ref, but we need to keep it alive
5922 * until after ->css_offline().
5923 */
5924 css_get(css);
5925
5926 /*
5927 * cgroup core guarantees that, by the time ->css_offline() is
5928 * invoked, no new css reference will be given out via
5929 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5930 * proceed to offlining css's because percpu_ref_kill() doesn't
5931 * guarantee that the ref is seen as killed on all CPUs on return.
5932 *
5933 * Use percpu_ref_kill_and_confirm() to get notifications as each
5934 * css is confirmed to be seen as killed on all CPUs.
5935 */
5936 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5937}
5938
5939/**
5940 * cgroup_destroy_locked - the first stage of cgroup destruction
5941 * @cgrp: cgroup to be destroyed
5942 *
5943 * css's make use of percpu refcnts whose killing latency shouldn't be
5944 * exposed to userland and are RCU protected. Also, cgroup core needs to
5945 * guarantee that css_tryget_online() won't succeed by the time
5946 * ->css_offline() is invoked. To satisfy all the requirements,
5947 * destruction is implemented in the following two steps.
5948 *
5949 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5950 * userland visible parts and start killing the percpu refcnts of
5951 * css's. Set up so that the next stage will be kicked off once all
5952 * the percpu refcnts are confirmed to be killed.
5953 *
5954 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5955 * rest of destruction. Once all cgroup references are gone, the
5956 * cgroup is RCU-freed.
5957 *
5958 * This function implements s1. After this step, @cgrp is gone as far as
5959 * the userland is concerned and a new cgroup with the same name may be
5960 * created. As cgroup doesn't care about the names internally, this
5961 * doesn't cause any problem.
5962 */
5963static int cgroup_destroy_locked(struct cgroup *cgrp)
5964 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5965{
5966 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5967 struct cgroup_subsys_state *css;
5968 struct cgrp_cset_link *link;
5969 int ssid;
5970
5971 lockdep_assert_held(&cgroup_mutex);
5972
5973 /*
5974 * Only migration can raise populated from zero and we're already
5975 * holding cgroup_mutex.
5976 */
5977 if (cgroup_is_populated(cgrp))
5978 return -EBUSY;
5979
5980 /*
5981 * Make sure there's no live children. We can't test emptiness of
5982 * ->self.children as dead children linger on it while being
5983 * drained; otherwise, "rmdir parent/child parent" may fail.
5984 */
5985 if (css_has_online_children(&cgrp->self))
5986 return -EBUSY;
5987
5988 /*
5989 * Mark @cgrp and the associated csets dead. The former prevents
5990 * further task migration and child creation by disabling
5991 * cgroup_kn_lock_live(). The latter makes the csets ignored by
5992 * the migration path.
5993 */
5994 cgrp->self.flags &= ~CSS_ONLINE;
5995
5996 spin_lock_irq(&css_set_lock);
5997 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5998 link->cset->dead = true;
5999 spin_unlock_irq(&css_set_lock);
6000
6001 /* initiate massacre of all css's */
6002 for_each_css(css, ssid, cgrp)
6003 kill_css(css);
6004
6005 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
6006 css_clear_dir(&cgrp->self);
6007 kernfs_remove(cgrp->kn);
6008
6009 if (cgroup_is_threaded(cgrp))
6010 parent->nr_threaded_children--;
6011
6012 spin_lock_irq(&css_set_lock);
6013 for (tcgrp = parent; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
6014 tcgrp->nr_descendants--;
6015 tcgrp->nr_dying_descendants++;
6016 /*
6017 * If the dying cgroup is frozen, decrease frozen descendants
6018 * counters of ancestor cgroups.
6019 */
6020 if (test_bit(CGRP_FROZEN, &cgrp->flags))
6021 tcgrp->freezer.nr_frozen_descendants--;
6022 }
6023 spin_unlock_irq(&css_set_lock);
6024
6025 cgroup1_check_for_release(parent);
6026
6027 if (cgrp->root == &cgrp_dfl_root)
6028 cgroup_bpf_offline(cgrp);
6029
6030 /* put the base reference */
6031 percpu_ref_kill(&cgrp->self.refcnt);
6032
6033 return 0;
6034};
6035
6036int cgroup_rmdir(struct kernfs_node *kn)
6037{
6038 struct cgroup *cgrp;
6039 int ret = 0;
6040
6041 cgrp = cgroup_kn_lock_live(kn, false);
6042 if (!cgrp)
6043 return 0;
6044
6045 ret = cgroup_destroy_locked(cgrp);
6046 if (!ret)
6047 TRACE_CGROUP_PATH(rmdir, cgrp);
6048
6049 cgroup_kn_unlock(kn);
6050 return ret;
6051}
6052
6053static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
6054 .show_options = cgroup_show_options,
6055 .mkdir = cgroup_mkdir,
6056 .rmdir = cgroup_rmdir,
6057 .show_path = cgroup_show_path,
6058};
6059
6060static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
6061{
6062 struct cgroup_subsys_state *css;
6063
6064 pr_debug("Initializing cgroup subsys %s\n", ss->name);
6065
6066 cgroup_lock();
6067
6068 idr_init(&ss->css_idr);
6069 INIT_LIST_HEAD(&ss->cfts);
6070
6071 /* Create the root cgroup state for this subsystem */
6072 ss->root = &cgrp_dfl_root;
6073 css = ss->css_alloc(NULL);
6074 /* We don't handle early failures gracefully */
6075 BUG_ON(IS_ERR(css));
6076 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
6077
6078 /*
6079 * Root csses are never destroyed and we can't initialize
6080 * percpu_ref during early init. Disable refcnting.
6081 */
6082 css->flags |= CSS_NO_REF;
6083
6084 if (early) {
6085 /* allocation can't be done safely during early init */
6086 css->id = 1;
6087 } else {
6088 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
6089 BUG_ON(css->id < 0);
6090 }
6091
6092 /* Update the init_css_set to contain a subsys
6093 * pointer to this state - since the subsystem is
6094 * newly registered, all tasks and hence the
6095 * init_css_set is in the subsystem's root cgroup. */
6096 init_css_set.subsys[ss->id] = css;
6097
6098 have_fork_callback |= (bool)ss->fork << ss->id;
6099 have_exit_callback |= (bool)ss->exit << ss->id;
6100 have_release_callback |= (bool)ss->release << ss->id;
6101 have_canfork_callback |= (bool)ss->can_fork << ss->id;
6102
6103 /* At system boot, before all subsystems have been
6104 * registered, no tasks have been forked, so we don't
6105 * need to invoke fork callbacks here. */
6106 BUG_ON(!list_empty(&init_task.tasks));
6107
6108 BUG_ON(online_css(css));
6109
6110 cgroup_unlock();
6111}
6112
6113/**
6114 * cgroup_init_early - cgroup initialization at system boot
6115 *
6116 * Initialize cgroups at system boot, and initialize any
6117 * subsystems that request early init.
6118 */
6119int __init cgroup_init_early(void)
6120{
6121 static struct cgroup_fs_context __initdata ctx;
6122 struct cgroup_subsys *ss;
6123 int i;
6124
6125 ctx.root = &cgrp_dfl_root;
6126 init_cgroup_root(&ctx);
6127 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6128
6129 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6130
6131 for_each_subsys(ss, i) {
6132 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6133 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6134 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6135 ss->id, ss->name);
6136 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6137 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6138
6139 ss->id = i;
6140 ss->name = cgroup_subsys_name[i];
6141 if (!ss->legacy_name)
6142 ss->legacy_name = cgroup_subsys_name[i];
6143
6144 if (ss->early_init)
6145 cgroup_init_subsys(ss, true);
6146 }
6147 return 0;
6148}
6149
6150/**
6151 * cgroup_init - cgroup initialization
6152 *
6153 * Register cgroup filesystem and /proc file, and initialize
6154 * any subsystems that didn't request early init.
6155 */
6156int __init cgroup_init(void)
6157{
6158 struct cgroup_subsys *ss;
6159 int ssid;
6160
6161 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6162 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6163 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6164 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6165
6166 cgroup_rstat_boot();
6167
6168 get_user_ns(init_cgroup_ns.user_ns);
6169
6170 cgroup_lock();
6171
6172 /*
6173 * Add init_css_set to the hash table so that dfl_root can link to
6174 * it during init.
6175 */
6176 hash_add(css_set_table, &init_css_set.hlist,
6177 css_set_hash(init_css_set.subsys));
6178
6179 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6180
6181 cgroup_unlock();
6182
6183 for_each_subsys(ss, ssid) {
6184 if (ss->early_init) {
6185 struct cgroup_subsys_state *css =
6186 init_css_set.subsys[ss->id];
6187
6188 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6189 GFP_KERNEL);
6190 BUG_ON(css->id < 0);
6191 } else {
6192 cgroup_init_subsys(ss, false);
6193 }
6194
6195 list_add_tail(&init_css_set.e_cset_node[ssid],
6196 &cgrp_dfl_root.cgrp.e_csets[ssid]);
6197
6198 /*
6199 * Setting dfl_root subsys_mask needs to consider the
6200 * disabled flag and cftype registration needs kmalloc,
6201 * both of which aren't available during early_init.
6202 */
6203 if (!cgroup_ssid_enabled(ssid))
6204 continue;
6205
6206 if (cgroup1_ssid_disabled(ssid))
6207 pr_info("Disabling %s control group subsystem in v1 mounts\n",
6208 ss->legacy_name);
6209
6210 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6211
6212 /* implicit controllers must be threaded too */
6213 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6214
6215 if (ss->implicit_on_dfl)
6216 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6217 else if (!ss->dfl_cftypes)
6218 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6219
6220 if (ss->threaded)
6221 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6222
6223 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6224 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6225 } else {
6226 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6227 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6228 }
6229
6230 if (ss->bind)
6231 ss->bind(init_css_set.subsys[ssid]);
6232
6233 cgroup_lock();
6234 css_populate_dir(init_css_set.subsys[ssid]);
6235 cgroup_unlock();
6236 }
6237
6238 /* init_css_set.subsys[] has been updated, re-hash */
6239 hash_del(&init_css_set.hlist);
6240 hash_add(css_set_table, &init_css_set.hlist,
6241 css_set_hash(init_css_set.subsys));
6242
6243 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6244 WARN_ON(register_filesystem(&cgroup_fs_type));
6245 WARN_ON(register_filesystem(&cgroup2_fs_type));
6246 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6247#ifdef CONFIG_CPUSETS_V1
6248 WARN_ON(register_filesystem(&cpuset_fs_type));
6249#endif
6250
6251 return 0;
6252}
6253
6254static int __init cgroup_wq_init(void)
6255{
6256 /*
6257 * There isn't much point in executing destruction path in
6258 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6259 * Use 1 for @max_active.
6260 *
6261 * We would prefer to do this in cgroup_init() above, but that
6262 * is called before init_workqueues(): so leave this until after.
6263 */
6264 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
6265 BUG_ON(!cgroup_destroy_wq);
6266 return 0;
6267}
6268core_initcall(cgroup_wq_init);
6269
6270void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6271{
6272 struct kernfs_node *kn;
6273
6274 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6275 if (!kn)
6276 return;
6277 kernfs_path(kn, buf, buflen);
6278 kernfs_put(kn);
6279}
6280
6281/*
6282 * cgroup_get_from_id : get the cgroup associated with cgroup id
6283 * @id: cgroup id
6284 * On success return the cgrp or ERR_PTR on failure
6285 * Only cgroups within current task's cgroup NS are valid.
6286 */
6287struct cgroup *cgroup_get_from_id(u64 id)
6288{
6289 struct kernfs_node *kn;
6290 struct cgroup *cgrp, *root_cgrp;
6291
6292 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6293 if (!kn)
6294 return ERR_PTR(-ENOENT);
6295
6296 if (kernfs_type(kn) != KERNFS_DIR) {
6297 kernfs_put(kn);
6298 return ERR_PTR(-ENOENT);
6299 }
6300
6301 rcu_read_lock();
6302
6303 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6304 if (cgrp && !cgroup_tryget(cgrp))
6305 cgrp = NULL;
6306
6307 rcu_read_unlock();
6308 kernfs_put(kn);
6309
6310 if (!cgrp)
6311 return ERR_PTR(-ENOENT);
6312
6313 root_cgrp = current_cgns_cgroup_dfl();
6314 if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6315 cgroup_put(cgrp);
6316 return ERR_PTR(-ENOENT);
6317 }
6318
6319 return cgrp;
6320}
6321EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6322
6323/*
6324 * proc_cgroup_show()
6325 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6326 * - Used for /proc/<pid>/cgroup.
6327 */
6328int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6329 struct pid *pid, struct task_struct *tsk)
6330{
6331 char *buf;
6332 int retval;
6333 struct cgroup_root *root;
6334
6335 retval = -ENOMEM;
6336 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6337 if (!buf)
6338 goto out;
6339
6340 rcu_read_lock();
6341 spin_lock_irq(&css_set_lock);
6342
6343 for_each_root(root) {
6344 struct cgroup_subsys *ss;
6345 struct cgroup *cgrp;
6346 int ssid, count = 0;
6347
6348 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6349 continue;
6350
6351 cgrp = task_cgroup_from_root(tsk, root);
6352 /* The root has already been unmounted. */
6353 if (!cgrp)
6354 continue;
6355
6356 seq_printf(m, "%d:", root->hierarchy_id);
6357 if (root != &cgrp_dfl_root)
6358 for_each_subsys(ss, ssid)
6359 if (root->subsys_mask & (1 << ssid))
6360 seq_printf(m, "%s%s", count++ ? "," : "",
6361 ss->legacy_name);
6362 if (strlen(root->name))
6363 seq_printf(m, "%sname=%s", count ? "," : "",
6364 root->name);
6365 seq_putc(m, ':');
6366 /*
6367 * On traditional hierarchies, all zombie tasks show up as
6368 * belonging to the root cgroup. On the default hierarchy,
6369 * while a zombie doesn't show up in "cgroup.procs" and
6370 * thus can't be migrated, its /proc/PID/cgroup keeps
6371 * reporting the cgroup it belonged to before exiting. If
6372 * the cgroup is removed before the zombie is reaped,
6373 * " (deleted)" is appended to the cgroup path.
6374 */
6375 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6376 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6377 current->nsproxy->cgroup_ns);
6378 if (retval == -E2BIG)
6379 retval = -ENAMETOOLONG;
6380 if (retval < 0)
6381 goto out_unlock;
6382
6383 seq_puts(m, buf);
6384 } else {
6385 seq_puts(m, "/");
6386 }
6387
6388 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6389 seq_puts(m, " (deleted)\n");
6390 else
6391 seq_putc(m, '\n');
6392 }
6393
6394 retval = 0;
6395out_unlock:
6396 spin_unlock_irq(&css_set_lock);
6397 rcu_read_unlock();
6398 kfree(buf);
6399out:
6400 return retval;
6401}
6402
6403/**
6404 * cgroup_fork - initialize cgroup related fields during copy_process()
6405 * @child: pointer to task_struct of forking parent process.
6406 *
6407 * A task is associated with the init_css_set until cgroup_post_fork()
6408 * attaches it to the target css_set.
6409 */
6410void cgroup_fork(struct task_struct *child)
6411{
6412 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6413 INIT_LIST_HEAD(&child->cg_list);
6414}
6415
6416/**
6417 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6418 * @f: file corresponding to cgroup_dir
6419 *
6420 * Find the cgroup from a file pointer associated with a cgroup directory.
6421 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6422 * cgroup cannot be found.
6423 */
6424static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6425{
6426 struct cgroup_subsys_state *css;
6427
6428 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6429 if (IS_ERR(css))
6430 return ERR_CAST(css);
6431
6432 return css->cgroup;
6433}
6434
6435/**
6436 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6437 * cgroup2.
6438 * @f: file corresponding to cgroup2_dir
6439 */
6440static struct cgroup *cgroup_get_from_file(struct file *f)
6441{
6442 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6443
6444 if (IS_ERR(cgrp))
6445 return ERR_CAST(cgrp);
6446
6447 if (!cgroup_on_dfl(cgrp)) {
6448 cgroup_put(cgrp);
6449 return ERR_PTR(-EBADF);
6450 }
6451
6452 return cgrp;
6453}
6454
6455/**
6456 * cgroup_css_set_fork - find or create a css_set for a child process
6457 * @kargs: the arguments passed to create the child process
6458 *
6459 * This functions finds or creates a new css_set which the child
6460 * process will be attached to in cgroup_post_fork(). By default,
6461 * the child process will be given the same css_set as its parent.
6462 *
6463 * If CLONE_INTO_CGROUP is specified this function will try to find an
6464 * existing css_set which includes the requested cgroup and if not create
6465 * a new css_set that the child will be attached to later. If this function
6466 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6467 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6468 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6469 * to the target cgroup.
6470 */
6471static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6472 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6473{
6474 int ret;
6475 struct cgroup *dst_cgrp = NULL;
6476 struct css_set *cset;
6477 struct super_block *sb;
6478
6479 if (kargs->flags & CLONE_INTO_CGROUP)
6480 cgroup_lock();
6481
6482 cgroup_threadgroup_change_begin(current);
6483
6484 spin_lock_irq(&css_set_lock);
6485 cset = task_css_set(current);
6486 get_css_set(cset);
6487 if (kargs->cgrp)
6488 kargs->kill_seq = kargs->cgrp->kill_seq;
6489 else
6490 kargs->kill_seq = cset->dfl_cgrp->kill_seq;
6491 spin_unlock_irq(&css_set_lock);
6492
6493 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6494 kargs->cset = cset;
6495 return 0;
6496 }
6497
6498 CLASS(fd_raw, f)(kargs->cgroup);
6499 if (fd_empty(f)) {
6500 ret = -EBADF;
6501 goto err;
6502 }
6503 sb = fd_file(f)->f_path.dentry->d_sb;
6504
6505 dst_cgrp = cgroup_get_from_file(fd_file(f));
6506 if (IS_ERR(dst_cgrp)) {
6507 ret = PTR_ERR(dst_cgrp);
6508 dst_cgrp = NULL;
6509 goto err;
6510 }
6511
6512 if (cgroup_is_dead(dst_cgrp)) {
6513 ret = -ENODEV;
6514 goto err;
6515 }
6516
6517 /*
6518 * Verify that we the target cgroup is writable for us. This is
6519 * usually done by the vfs layer but since we're not going through
6520 * the vfs layer here we need to do it "manually".
6521 */
6522 ret = cgroup_may_write(dst_cgrp, sb);
6523 if (ret)
6524 goto err;
6525
6526 /*
6527 * Spawning a task directly into a cgroup works by passing a file
6528 * descriptor to the target cgroup directory. This can even be an O_PATH
6529 * file descriptor. But it can never be a cgroup.procs file descriptor.
6530 * This was done on purpose so spawning into a cgroup could be
6531 * conceptualized as an atomic
6532 *
6533 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6534 * write(fd, <child-pid>, ...);
6535 *
6536 * sequence, i.e. it's a shorthand for the caller opening and writing
6537 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6538 * to always use the caller's credentials.
6539 */
6540 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6541 !(kargs->flags & CLONE_THREAD),
6542 current->nsproxy->cgroup_ns);
6543 if (ret)
6544 goto err;
6545
6546 kargs->cset = find_css_set(cset, dst_cgrp);
6547 if (!kargs->cset) {
6548 ret = -ENOMEM;
6549 goto err;
6550 }
6551
6552 put_css_set(cset);
6553 kargs->cgrp = dst_cgrp;
6554 return ret;
6555
6556err:
6557 cgroup_threadgroup_change_end(current);
6558 cgroup_unlock();
6559 if (dst_cgrp)
6560 cgroup_put(dst_cgrp);
6561 put_css_set(cset);
6562 if (kargs->cset)
6563 put_css_set(kargs->cset);
6564 return ret;
6565}
6566
6567/**
6568 * cgroup_css_set_put_fork - drop references we took during fork
6569 * @kargs: the arguments passed to create the child process
6570 *
6571 * Drop references to the prepared css_set and target cgroup if
6572 * CLONE_INTO_CGROUP was requested.
6573 */
6574static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6575 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6576{
6577 struct cgroup *cgrp = kargs->cgrp;
6578 struct css_set *cset = kargs->cset;
6579
6580 cgroup_threadgroup_change_end(current);
6581
6582 if (cset) {
6583 put_css_set(cset);
6584 kargs->cset = NULL;
6585 }
6586
6587 if (kargs->flags & CLONE_INTO_CGROUP) {
6588 cgroup_unlock();
6589 if (cgrp) {
6590 cgroup_put(cgrp);
6591 kargs->cgrp = NULL;
6592 }
6593 }
6594}
6595
6596/**
6597 * cgroup_can_fork - called on a new task before the process is exposed
6598 * @child: the child process
6599 * @kargs: the arguments passed to create the child process
6600 *
6601 * This prepares a new css_set for the child process which the child will
6602 * be attached to in cgroup_post_fork().
6603 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6604 * callback returns an error, the fork aborts with that error code. This
6605 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6606 */
6607int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6608{
6609 struct cgroup_subsys *ss;
6610 int i, j, ret;
6611
6612 ret = cgroup_css_set_fork(kargs);
6613 if (ret)
6614 return ret;
6615
6616 do_each_subsys_mask(ss, i, have_canfork_callback) {
6617 ret = ss->can_fork(child, kargs->cset);
6618 if (ret)
6619 goto out_revert;
6620 } while_each_subsys_mask();
6621
6622 return 0;
6623
6624out_revert:
6625 for_each_subsys(ss, j) {
6626 if (j >= i)
6627 break;
6628 if (ss->cancel_fork)
6629 ss->cancel_fork(child, kargs->cset);
6630 }
6631
6632 cgroup_css_set_put_fork(kargs);
6633
6634 return ret;
6635}
6636
6637/**
6638 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6639 * @child: the child process
6640 * @kargs: the arguments passed to create the child process
6641 *
6642 * This calls the cancel_fork() callbacks if a fork failed *after*
6643 * cgroup_can_fork() succeeded and cleans up references we took to
6644 * prepare a new css_set for the child process in cgroup_can_fork().
6645 */
6646void cgroup_cancel_fork(struct task_struct *child,
6647 struct kernel_clone_args *kargs)
6648{
6649 struct cgroup_subsys *ss;
6650 int i;
6651
6652 for_each_subsys(ss, i)
6653 if (ss->cancel_fork)
6654 ss->cancel_fork(child, kargs->cset);
6655
6656 cgroup_css_set_put_fork(kargs);
6657}
6658
6659/**
6660 * cgroup_post_fork - finalize cgroup setup for the child process
6661 * @child: the child process
6662 * @kargs: the arguments passed to create the child process
6663 *
6664 * Attach the child process to its css_set calling the subsystem fork()
6665 * callbacks.
6666 */
6667void cgroup_post_fork(struct task_struct *child,
6668 struct kernel_clone_args *kargs)
6669 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6670{
6671 unsigned int cgrp_kill_seq = 0;
6672 unsigned long cgrp_flags = 0;
6673 bool kill = false;
6674 struct cgroup_subsys *ss;
6675 struct css_set *cset;
6676 int i;
6677
6678 cset = kargs->cset;
6679 kargs->cset = NULL;
6680
6681 spin_lock_irq(&css_set_lock);
6682
6683 /* init tasks are special, only link regular threads */
6684 if (likely(child->pid)) {
6685 if (kargs->cgrp) {
6686 cgrp_flags = kargs->cgrp->flags;
6687 cgrp_kill_seq = kargs->cgrp->kill_seq;
6688 } else {
6689 cgrp_flags = cset->dfl_cgrp->flags;
6690 cgrp_kill_seq = cset->dfl_cgrp->kill_seq;
6691 }
6692
6693 WARN_ON_ONCE(!list_empty(&child->cg_list));
6694 cset->nr_tasks++;
6695 css_set_move_task(child, NULL, cset, false);
6696 } else {
6697 put_css_set(cset);
6698 cset = NULL;
6699 }
6700
6701 if (!(child->flags & PF_KTHREAD)) {
6702 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6703 /*
6704 * If the cgroup has to be frozen, the new task has
6705 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6706 * get the task into the frozen state.
6707 */
6708 spin_lock(&child->sighand->siglock);
6709 WARN_ON_ONCE(child->frozen);
6710 child->jobctl |= JOBCTL_TRAP_FREEZE;
6711 spin_unlock(&child->sighand->siglock);
6712
6713 /*
6714 * Calling cgroup_update_frozen() isn't required here,
6715 * because it will be called anyway a bit later from
6716 * do_freezer_trap(). So we avoid cgroup's transient
6717 * switch from the frozen state and back.
6718 */
6719 }
6720
6721 /*
6722 * If the cgroup is to be killed notice it now and take the
6723 * child down right after we finished preparing it for
6724 * userspace.
6725 */
6726 kill = kargs->kill_seq != cgrp_kill_seq;
6727 }
6728
6729 spin_unlock_irq(&css_set_lock);
6730
6731 /*
6732 * Call ss->fork(). This must happen after @child is linked on
6733 * css_set; otherwise, @child might change state between ->fork()
6734 * and addition to css_set.
6735 */
6736 do_each_subsys_mask(ss, i, have_fork_callback) {
6737 ss->fork(child);
6738 } while_each_subsys_mask();
6739
6740 /* Make the new cset the root_cset of the new cgroup namespace. */
6741 if (kargs->flags & CLONE_NEWCGROUP) {
6742 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6743
6744 get_css_set(cset);
6745 child->nsproxy->cgroup_ns->root_cset = cset;
6746 put_css_set(rcset);
6747 }
6748
6749 /* Cgroup has to be killed so take down child immediately. */
6750 if (unlikely(kill))
6751 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6752
6753 cgroup_css_set_put_fork(kargs);
6754}
6755
6756/**
6757 * cgroup_exit - detach cgroup from exiting task
6758 * @tsk: pointer to task_struct of exiting process
6759 *
6760 * Description: Detach cgroup from @tsk.
6761 *
6762 */
6763void cgroup_exit(struct task_struct *tsk)
6764{
6765 struct cgroup_subsys *ss;
6766 struct css_set *cset;
6767 int i;
6768
6769 spin_lock_irq(&css_set_lock);
6770
6771 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6772 cset = task_css_set(tsk);
6773 css_set_move_task(tsk, cset, NULL, false);
6774 cset->nr_tasks--;
6775 /* matches the signal->live check in css_task_iter_advance() */
6776 if (thread_group_leader(tsk) && atomic_read(&tsk->signal->live))
6777 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6778
6779 if (dl_task(tsk))
6780 dec_dl_tasks_cs(tsk);
6781
6782 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6783 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6784 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6785 cgroup_update_frozen(task_dfl_cgroup(tsk));
6786
6787 spin_unlock_irq(&css_set_lock);
6788
6789 /* see cgroup_post_fork() for details */
6790 do_each_subsys_mask(ss, i, have_exit_callback) {
6791 ss->exit(tsk);
6792 } while_each_subsys_mask();
6793}
6794
6795void cgroup_release(struct task_struct *task)
6796{
6797 struct cgroup_subsys *ss;
6798 int ssid;
6799
6800 do_each_subsys_mask(ss, ssid, have_release_callback) {
6801 ss->release(task);
6802 } while_each_subsys_mask();
6803
6804 if (!list_empty(&task->cg_list)) {
6805 spin_lock_irq(&css_set_lock);
6806 css_set_skip_task_iters(task_css_set(task), task);
6807 list_del_init(&task->cg_list);
6808 spin_unlock_irq(&css_set_lock);
6809 }
6810}
6811
6812void cgroup_free(struct task_struct *task)
6813{
6814 struct css_set *cset = task_css_set(task);
6815 put_css_set(cset);
6816}
6817
6818static int __init cgroup_disable(char *str)
6819{
6820 struct cgroup_subsys *ss;
6821 char *token;
6822 int i;
6823
6824 while ((token = strsep(&str, ",")) != NULL) {
6825 if (!*token)
6826 continue;
6827
6828 for_each_subsys(ss, i) {
6829 if (strcmp(token, ss->name) &&
6830 strcmp(token, ss->legacy_name))
6831 continue;
6832
6833 static_branch_disable(cgroup_subsys_enabled_key[i]);
6834 pr_info("Disabling %s control group subsystem\n",
6835 ss->name);
6836 }
6837
6838 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6839 if (strcmp(token, cgroup_opt_feature_names[i]))
6840 continue;
6841 cgroup_feature_disable_mask |= 1 << i;
6842 pr_info("Disabling %s control group feature\n",
6843 cgroup_opt_feature_names[i]);
6844 break;
6845 }
6846 }
6847 return 1;
6848}
6849__setup("cgroup_disable=", cgroup_disable);
6850
6851void __init __weak enable_debug_cgroup(void) { }
6852
6853static int __init enable_cgroup_debug(char *str)
6854{
6855 cgroup_debug = true;
6856 enable_debug_cgroup();
6857 return 1;
6858}
6859__setup("cgroup_debug", enable_cgroup_debug);
6860
6861static int __init cgroup_favordynmods_setup(char *str)
6862{
6863 return (kstrtobool(str, &have_favordynmods) == 0);
6864}
6865__setup("cgroup_favordynmods=", cgroup_favordynmods_setup);
6866
6867/**
6868 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6869 * @dentry: directory dentry of interest
6870 * @ss: subsystem of interest
6871 *
6872 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6873 * to get the corresponding css and return it. If such css doesn't exist
6874 * or can't be pinned, an ERR_PTR value is returned.
6875 */
6876struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6877 struct cgroup_subsys *ss)
6878{
6879 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6880 struct file_system_type *s_type = dentry->d_sb->s_type;
6881 struct cgroup_subsys_state *css = NULL;
6882 struct cgroup *cgrp;
6883
6884 /* is @dentry a cgroup dir? */
6885 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6886 !kn || kernfs_type(kn) != KERNFS_DIR)
6887 return ERR_PTR(-EBADF);
6888
6889 rcu_read_lock();
6890
6891 /*
6892 * This path doesn't originate from kernfs and @kn could already
6893 * have been or be removed at any point. @kn->priv is RCU
6894 * protected for this access. See css_release_work_fn() for details.
6895 */
6896 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6897 if (cgrp)
6898 css = cgroup_css(cgrp, ss);
6899
6900 if (!css || !css_tryget_online(css))
6901 css = ERR_PTR(-ENOENT);
6902
6903 rcu_read_unlock();
6904 return css;
6905}
6906
6907/**
6908 * css_from_id - lookup css by id
6909 * @id: the cgroup id
6910 * @ss: cgroup subsys to be looked into
6911 *
6912 * Returns the css if there's valid one with @id, otherwise returns NULL.
6913 * Should be called under rcu_read_lock().
6914 */
6915struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6916{
6917 WARN_ON_ONCE(!rcu_read_lock_held());
6918 return idr_find(&ss->css_idr, id);
6919}
6920
6921/**
6922 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6923 * @path: path on the default hierarchy
6924 *
6925 * Find the cgroup at @path on the default hierarchy, increment its
6926 * reference count and return it. Returns pointer to the found cgroup on
6927 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6928 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6929 */
6930struct cgroup *cgroup_get_from_path(const char *path)
6931{
6932 struct kernfs_node *kn;
6933 struct cgroup *cgrp = ERR_PTR(-ENOENT);
6934 struct cgroup *root_cgrp;
6935
6936 root_cgrp = current_cgns_cgroup_dfl();
6937 kn = kernfs_walk_and_get(root_cgrp->kn, path);
6938 if (!kn)
6939 goto out;
6940
6941 if (kernfs_type(kn) != KERNFS_DIR) {
6942 cgrp = ERR_PTR(-ENOTDIR);
6943 goto out_kernfs;
6944 }
6945
6946 rcu_read_lock();
6947
6948 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6949 if (!cgrp || !cgroup_tryget(cgrp))
6950 cgrp = ERR_PTR(-ENOENT);
6951
6952 rcu_read_unlock();
6953
6954out_kernfs:
6955 kernfs_put(kn);
6956out:
6957 return cgrp;
6958}
6959EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6960
6961/**
6962 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
6963 * @fd: fd obtained by open(cgroup_dir)
6964 *
6965 * Find the cgroup from a fd which should be obtained
6966 * by opening a cgroup directory. Returns a pointer to the
6967 * cgroup on success. ERR_PTR is returned if the cgroup
6968 * cannot be found.
6969 */
6970struct cgroup *cgroup_v1v2_get_from_fd(int fd)
6971{
6972 CLASS(fd_raw, f)(fd);
6973 if (fd_empty(f))
6974 return ERR_PTR(-EBADF);
6975
6976 return cgroup_v1v2_get_from_file(fd_file(f));
6977}
6978
6979/**
6980 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
6981 * cgroup2.
6982 * @fd: fd obtained by open(cgroup2_dir)
6983 */
6984struct cgroup *cgroup_get_from_fd(int fd)
6985{
6986 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
6987
6988 if (IS_ERR(cgrp))
6989 return ERR_CAST(cgrp);
6990
6991 if (!cgroup_on_dfl(cgrp)) {
6992 cgroup_put(cgrp);
6993 return ERR_PTR(-EBADF);
6994 }
6995 return cgrp;
6996}
6997EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6998
6999static u64 power_of_ten(int power)
7000{
7001 u64 v = 1;
7002 while (power--)
7003 v *= 10;
7004 return v;
7005}
7006
7007/**
7008 * cgroup_parse_float - parse a floating number
7009 * @input: input string
7010 * @dec_shift: number of decimal digits to shift
7011 * @v: output
7012 *
7013 * Parse a decimal floating point number in @input and store the result in
7014 * @v with decimal point right shifted @dec_shift times. For example, if
7015 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
7016 * Returns 0 on success, -errno otherwise.
7017 *
7018 * There's nothing cgroup specific about this function except that it's
7019 * currently the only user.
7020 */
7021int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
7022{
7023 s64 whole, frac = 0;
7024 int fstart = 0, fend = 0, flen;
7025
7026 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
7027 return -EINVAL;
7028 if (frac < 0)
7029 return -EINVAL;
7030
7031 flen = fend > fstart ? fend - fstart : 0;
7032 if (flen < dec_shift)
7033 frac *= power_of_ten(dec_shift - flen);
7034 else
7035 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
7036
7037 *v = whole * power_of_ten(dec_shift) + frac;
7038 return 0;
7039}
7040
7041/*
7042 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
7043 * definition in cgroup-defs.h.
7044 */
7045#ifdef CONFIG_SOCK_CGROUP_DATA
7046
7047void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
7048{
7049 struct cgroup *cgroup;
7050
7051 rcu_read_lock();
7052 /* Don't associate the sock with unrelated interrupted task's cgroup. */
7053 if (in_interrupt()) {
7054 cgroup = &cgrp_dfl_root.cgrp;
7055 cgroup_get(cgroup);
7056 goto out;
7057 }
7058
7059 while (true) {
7060 struct css_set *cset;
7061
7062 cset = task_css_set(current);
7063 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
7064 cgroup = cset->dfl_cgrp;
7065 break;
7066 }
7067 cpu_relax();
7068 }
7069out:
7070 skcd->cgroup = cgroup;
7071 cgroup_bpf_get(cgroup);
7072 rcu_read_unlock();
7073}
7074
7075void cgroup_sk_clone(struct sock_cgroup_data *skcd)
7076{
7077 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7078
7079 /*
7080 * We might be cloning a socket which is left in an empty
7081 * cgroup and the cgroup might have already been rmdir'd.
7082 * Don't use cgroup_get_live().
7083 */
7084 cgroup_get(cgrp);
7085 cgroup_bpf_get(cgrp);
7086}
7087
7088void cgroup_sk_free(struct sock_cgroup_data *skcd)
7089{
7090 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7091
7092 cgroup_bpf_put(cgrp);
7093 cgroup_put(cgrp);
7094}
7095
7096#endif /* CONFIG_SOCK_CGROUP_DATA */
7097
7098#ifdef CONFIG_SYSFS
7099static ssize_t show_delegatable_files(struct cftype *files, char *buf,
7100 ssize_t size, const char *prefix)
7101{
7102 struct cftype *cft;
7103 ssize_t ret = 0;
7104
7105 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
7106 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
7107 continue;
7108
7109 if (prefix)
7110 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
7111
7112 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
7113
7114 if (WARN_ON(ret >= size))
7115 break;
7116 }
7117
7118 return ret;
7119}
7120
7121static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7122 char *buf)
7123{
7124 struct cgroup_subsys *ss;
7125 int ssid;
7126 ssize_t ret = 0;
7127
7128 ret = show_delegatable_files(cgroup_base_files, buf + ret,
7129 PAGE_SIZE - ret, NULL);
7130 if (cgroup_psi_enabled())
7131 ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7132 PAGE_SIZE - ret, NULL);
7133
7134 for_each_subsys(ss, ssid)
7135 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7136 PAGE_SIZE - ret,
7137 cgroup_subsys_name[ssid]);
7138
7139 return ret;
7140}
7141static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7142
7143static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7144 char *buf)
7145{
7146 return snprintf(buf, PAGE_SIZE,
7147 "nsdelegate\n"
7148 "favordynmods\n"
7149 "memory_localevents\n"
7150 "memory_recursiveprot\n"
7151 "memory_hugetlb_accounting\n"
7152 "pids_localevents\n");
7153}
7154static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7155
7156static struct attribute *cgroup_sysfs_attrs[] = {
7157 &cgroup_delegate_attr.attr,
7158 &cgroup_features_attr.attr,
7159 NULL,
7160};
7161
7162static const struct attribute_group cgroup_sysfs_attr_group = {
7163 .attrs = cgroup_sysfs_attrs,
7164 .name = "cgroup",
7165};
7166
7167static int __init cgroup_sysfs_init(void)
7168{
7169 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7170}
7171subsys_initcall(cgroup_sysfs_init);
7172
7173#endif /* CONFIG_SYSFS */
1/*
2 * Generic process-grouping system.
3 *
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
6 *
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
10 *
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
15 *
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
18 *
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
23 *
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
27 */
28
29#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
31#include "cgroup-internal.h"
32
33#include <linux/cred.h>
34#include <linux/errno.h>
35#include <linux/init_task.h>
36#include <linux/kernel.h>
37#include <linux/magic.h>
38#include <linux/mutex.h>
39#include <linux/mount.h>
40#include <linux/pagemap.h>
41#include <linux/proc_fs.h>
42#include <linux/rcupdate.h>
43#include <linux/sched.h>
44#include <linux/sched/task.h>
45#include <linux/slab.h>
46#include <linux/spinlock.h>
47#include <linux/percpu-rwsem.h>
48#include <linux/string.h>
49#include <linux/hashtable.h>
50#include <linux/idr.h>
51#include <linux/kthread.h>
52#include <linux/atomic.h>
53#include <linux/cpuset.h>
54#include <linux/proc_ns.h>
55#include <linux/nsproxy.h>
56#include <linux/file.h>
57#include <linux/fs_parser.h>
58#include <linux/sched/cputime.h>
59#include <linux/psi.h>
60#include <net/sock.h>
61
62#define CREATE_TRACE_POINTS
63#include <trace/events/cgroup.h>
64
65#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
66 MAX_CFTYPE_NAME + 2)
67/* let's not notify more than 100 times per second */
68#define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
69
70/*
71 * cgroup_mutex is the master lock. Any modification to cgroup or its
72 * hierarchy must be performed while holding it.
73 *
74 * css_set_lock protects task->cgroups pointer, the list of css_set
75 * objects, and the chain of tasks off each css_set.
76 *
77 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
78 * cgroup.h can use them for lockdep annotations.
79 */
80DEFINE_MUTEX(cgroup_mutex);
81DEFINE_SPINLOCK(css_set_lock);
82
83#ifdef CONFIG_PROVE_RCU
84EXPORT_SYMBOL_GPL(cgroup_mutex);
85EXPORT_SYMBOL_GPL(css_set_lock);
86#endif
87
88DEFINE_SPINLOCK(trace_cgroup_path_lock);
89char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
90bool cgroup_debug __read_mostly;
91
92/*
93 * Protects cgroup_idr and css_idr so that IDs can be released without
94 * grabbing cgroup_mutex.
95 */
96static DEFINE_SPINLOCK(cgroup_idr_lock);
97
98/*
99 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
100 * against file removal/re-creation across css hiding.
101 */
102static DEFINE_SPINLOCK(cgroup_file_kn_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_wq
115 * which may lead to deadlock.
116 */
117static struct workqueue_struct *cgroup_destroy_wq;
118
119/* generate an array of cgroup subsystem pointers */
120#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
121struct cgroup_subsys *cgroup_subsys[] = {
122#include <linux/cgroup_subsys.h>
123};
124#undef SUBSYS
125
126/* array of cgroup subsystem names */
127#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
128static const char *cgroup_subsys_name[] = {
129#include <linux/cgroup_subsys.h>
130};
131#undef SUBSYS
132
133/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
134#define SUBSYS(_x) \
135 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
136 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
137 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
138 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
139#include <linux/cgroup_subsys.h>
140#undef SUBSYS
141
142#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
143static struct static_key_true *cgroup_subsys_enabled_key[] = {
144#include <linux/cgroup_subsys.h>
145};
146#undef SUBSYS
147
148#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
149static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
150#include <linux/cgroup_subsys.h>
151};
152#undef SUBSYS
153
154static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
155
156/* the default hierarchy */
157struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
158EXPORT_SYMBOL_GPL(cgrp_dfl_root);
159
160/*
161 * The default hierarchy always exists but is hidden until mounted for the
162 * first time. This is for backward compatibility.
163 */
164static bool cgrp_dfl_visible;
165
166/* some controllers are not supported in the default hierarchy */
167static u16 cgrp_dfl_inhibit_ss_mask;
168
169/* some controllers are implicitly enabled on the default hierarchy */
170static u16 cgrp_dfl_implicit_ss_mask;
171
172/* some controllers can be threaded on the default hierarchy */
173static u16 cgrp_dfl_threaded_ss_mask;
174
175/* The list of hierarchy roots */
176LIST_HEAD(cgroup_roots);
177static int cgroup_root_count;
178
179/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
180static DEFINE_IDR(cgroup_hierarchy_idr);
181
182/*
183 * Assign a monotonically increasing serial number to csses. It guarantees
184 * cgroups with bigger numbers are newer than those with smaller numbers.
185 * Also, as csses are always appended to the parent's ->children list, it
186 * guarantees that sibling csses are always sorted in the ascending serial
187 * number order on the list. Protected by cgroup_mutex.
188 */
189static u64 css_serial_nr_next = 1;
190
191/*
192 * These bitmasks identify subsystems with specific features to avoid
193 * having to do iterative checks repeatedly.
194 */
195static u16 have_fork_callback __read_mostly;
196static u16 have_exit_callback __read_mostly;
197static u16 have_release_callback __read_mostly;
198static u16 have_canfork_callback __read_mostly;
199
200/* cgroup namespace for init task */
201struct cgroup_namespace init_cgroup_ns = {
202 .ns.count = REFCOUNT_INIT(2),
203 .user_ns = &init_user_ns,
204 .ns.ops = &cgroupns_operations,
205 .ns.inum = PROC_CGROUP_INIT_INO,
206 .root_cset = &init_css_set,
207};
208
209static struct file_system_type cgroup2_fs_type;
210static struct cftype cgroup_base_files[];
211
212/* cgroup optional features */
213enum cgroup_opt_features {
214#ifdef CONFIG_PSI
215 OPT_FEATURE_PRESSURE,
216#endif
217 OPT_FEATURE_COUNT
218};
219
220static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
221#ifdef CONFIG_PSI
222 "pressure",
223#endif
224};
225
226static u16 cgroup_feature_disable_mask __read_mostly;
227
228static int cgroup_apply_control(struct cgroup *cgrp);
229static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
230static void css_task_iter_skip(struct css_task_iter *it,
231 struct task_struct *task);
232static int cgroup_destroy_locked(struct cgroup *cgrp);
233static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
234 struct cgroup_subsys *ss);
235static void css_release(struct percpu_ref *ref);
236static void kill_css(struct cgroup_subsys_state *css);
237static int cgroup_addrm_files(struct cgroup_subsys_state *css,
238 struct cgroup *cgrp, struct cftype cfts[],
239 bool is_add);
240
241/**
242 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
243 * @ssid: subsys ID of interest
244 *
245 * cgroup_subsys_enabled() can only be used with literal subsys names which
246 * is fine for individual subsystems but unsuitable for cgroup core. This
247 * is slower static_key_enabled() based test indexed by @ssid.
248 */
249bool cgroup_ssid_enabled(int ssid)
250{
251 if (CGROUP_SUBSYS_COUNT == 0)
252 return false;
253
254 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
255}
256
257/**
258 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
259 * @cgrp: the cgroup of interest
260 *
261 * The default hierarchy is the v2 interface of cgroup and this function
262 * can be used to test whether a cgroup is on the default hierarchy for
263 * cases where a subsystem should behave differently depending on the
264 * interface version.
265 *
266 * List of changed behaviors:
267 *
268 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
269 * and "name" are disallowed.
270 *
271 * - When mounting an existing superblock, mount options should match.
272 *
273 * - Remount is disallowed.
274 *
275 * - rename(2) is disallowed.
276 *
277 * - "tasks" is removed. Everything should be at process granularity. Use
278 * "cgroup.procs" instead.
279 *
280 * - "cgroup.procs" is not sorted. pids will be unique unless they got
281 * recycled in-between reads.
282 *
283 * - "release_agent" and "notify_on_release" are removed. Replacement
284 * notification mechanism will be implemented.
285 *
286 * - "cgroup.clone_children" is removed.
287 *
288 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
289 * and its descendants contain no task; otherwise, 1. The file also
290 * generates kernfs notification which can be monitored through poll and
291 * [di]notify when the value of the file changes.
292 *
293 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
294 * take masks of ancestors with non-empty cpus/mems, instead of being
295 * moved to an ancestor.
296 *
297 * - cpuset: a task can be moved into an empty cpuset, and again it takes
298 * masks of ancestors.
299 *
300 * - blkcg: blk-throttle becomes properly hierarchical.
301 *
302 * - debug: disallowed on the default hierarchy.
303 */
304bool cgroup_on_dfl(const struct cgroup *cgrp)
305{
306 return cgrp->root == &cgrp_dfl_root;
307}
308
309/* IDR wrappers which synchronize using cgroup_idr_lock */
310static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
311 gfp_t gfp_mask)
312{
313 int ret;
314
315 idr_preload(gfp_mask);
316 spin_lock_bh(&cgroup_idr_lock);
317 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
318 spin_unlock_bh(&cgroup_idr_lock);
319 idr_preload_end();
320 return ret;
321}
322
323static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
324{
325 void *ret;
326
327 spin_lock_bh(&cgroup_idr_lock);
328 ret = idr_replace(idr, ptr, id);
329 spin_unlock_bh(&cgroup_idr_lock);
330 return ret;
331}
332
333static void cgroup_idr_remove(struct idr *idr, int id)
334{
335 spin_lock_bh(&cgroup_idr_lock);
336 idr_remove(idr, id);
337 spin_unlock_bh(&cgroup_idr_lock);
338}
339
340static bool cgroup_has_tasks(struct cgroup *cgrp)
341{
342 return cgrp->nr_populated_csets;
343}
344
345bool cgroup_is_threaded(struct cgroup *cgrp)
346{
347 return cgrp->dom_cgrp != cgrp;
348}
349
350/* can @cgrp host both domain and threaded children? */
351static bool cgroup_is_mixable(struct cgroup *cgrp)
352{
353 /*
354 * Root isn't under domain level resource control exempting it from
355 * the no-internal-process constraint, so it can serve as a thread
356 * root and a parent of resource domains at the same time.
357 */
358 return !cgroup_parent(cgrp);
359}
360
361/* can @cgrp become a thread root? Should always be true for a thread root */
362static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
363{
364 /* mixables don't care */
365 if (cgroup_is_mixable(cgrp))
366 return true;
367
368 /* domain roots can't be nested under threaded */
369 if (cgroup_is_threaded(cgrp))
370 return false;
371
372 /* can only have either domain or threaded children */
373 if (cgrp->nr_populated_domain_children)
374 return false;
375
376 /* and no domain controllers can be enabled */
377 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
378 return false;
379
380 return true;
381}
382
383/* is @cgrp root of a threaded subtree? */
384bool cgroup_is_thread_root(struct cgroup *cgrp)
385{
386 /* thread root should be a domain */
387 if (cgroup_is_threaded(cgrp))
388 return false;
389
390 /* a domain w/ threaded children is a thread root */
391 if (cgrp->nr_threaded_children)
392 return true;
393
394 /*
395 * A domain which has tasks and explicit threaded controllers
396 * enabled is a thread root.
397 */
398 if (cgroup_has_tasks(cgrp) &&
399 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
400 return true;
401
402 return false;
403}
404
405/* a domain which isn't connected to the root w/o brekage can't be used */
406static bool cgroup_is_valid_domain(struct cgroup *cgrp)
407{
408 /* the cgroup itself can be a thread root */
409 if (cgroup_is_threaded(cgrp))
410 return false;
411
412 /* but the ancestors can't be unless mixable */
413 while ((cgrp = cgroup_parent(cgrp))) {
414 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
415 return false;
416 if (cgroup_is_threaded(cgrp))
417 return false;
418 }
419
420 return true;
421}
422
423/* subsystems visibly enabled on a cgroup */
424static u16 cgroup_control(struct cgroup *cgrp)
425{
426 struct cgroup *parent = cgroup_parent(cgrp);
427 u16 root_ss_mask = cgrp->root->subsys_mask;
428
429 if (parent) {
430 u16 ss_mask = parent->subtree_control;
431
432 /* threaded cgroups can only have threaded controllers */
433 if (cgroup_is_threaded(cgrp))
434 ss_mask &= cgrp_dfl_threaded_ss_mask;
435 return ss_mask;
436 }
437
438 if (cgroup_on_dfl(cgrp))
439 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
440 cgrp_dfl_implicit_ss_mask);
441 return root_ss_mask;
442}
443
444/* subsystems enabled on a cgroup */
445static u16 cgroup_ss_mask(struct cgroup *cgrp)
446{
447 struct cgroup *parent = cgroup_parent(cgrp);
448
449 if (parent) {
450 u16 ss_mask = parent->subtree_ss_mask;
451
452 /* threaded cgroups can only have threaded controllers */
453 if (cgroup_is_threaded(cgrp))
454 ss_mask &= cgrp_dfl_threaded_ss_mask;
455 return ss_mask;
456 }
457
458 return cgrp->root->subsys_mask;
459}
460
461/**
462 * cgroup_css - obtain a cgroup's css for the specified subsystem
463 * @cgrp: the cgroup of interest
464 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
465 *
466 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
467 * function must be called either under cgroup_mutex or rcu_read_lock() and
468 * the caller is responsible for pinning the returned css if it wants to
469 * keep accessing it outside the said locks. This function may return
470 * %NULL if @cgrp doesn't have @subsys_id enabled.
471 */
472static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
473 struct cgroup_subsys *ss)
474{
475 if (ss)
476 return rcu_dereference_check(cgrp->subsys[ss->id],
477 lockdep_is_held(&cgroup_mutex));
478 else
479 return &cgrp->self;
480}
481
482/**
483 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
484 * @cgrp: the cgroup of interest
485 * @ss: the subsystem of interest
486 *
487 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
488 * or is offline, %NULL is returned.
489 */
490static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
491 struct cgroup_subsys *ss)
492{
493 struct cgroup_subsys_state *css;
494
495 rcu_read_lock();
496 css = cgroup_css(cgrp, ss);
497 if (css && !css_tryget_online(css))
498 css = NULL;
499 rcu_read_unlock();
500
501 return css;
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 do {
554 css = cgroup_css(cgrp, ss);
555
556 if (css)
557 return css;
558 cgrp = cgroup_parent(cgrp);
559 } while (cgrp);
560
561 return init_css_set.subsys[ss->id];
562}
563
564/**
565 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
566 * @cgrp: the cgroup of interest
567 * @ss: the subsystem of interest
568 *
569 * Find and get the effective css of @cgrp for @ss. The effective css is
570 * defined as the matching css of the nearest ancestor including self which
571 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
572 * the root css is returned, so this function always returns a valid css.
573 * The returned css must be put using css_put().
574 */
575struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
576 struct cgroup_subsys *ss)
577{
578 struct cgroup_subsys_state *css;
579
580 rcu_read_lock();
581
582 do {
583 css = cgroup_css(cgrp, ss);
584
585 if (css && css_tryget_online(css))
586 goto out_unlock;
587 cgrp = cgroup_parent(cgrp);
588 } while (cgrp);
589
590 css = init_css_set.subsys[ss->id];
591 css_get(css);
592out_unlock:
593 rcu_read_unlock();
594 return css;
595}
596EXPORT_SYMBOL_GPL(cgroup_get_e_css);
597
598static void cgroup_get_live(struct cgroup *cgrp)
599{
600 WARN_ON_ONCE(cgroup_is_dead(cgrp));
601 css_get(&cgrp->self);
602}
603
604/**
605 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
606 * is responsible for taking the css_set_lock.
607 * @cgrp: the cgroup in question
608 */
609int __cgroup_task_count(const struct cgroup *cgrp)
610{
611 int count = 0;
612 struct cgrp_cset_link *link;
613
614 lockdep_assert_held(&css_set_lock);
615
616 list_for_each_entry(link, &cgrp->cset_links, cset_link)
617 count += link->cset->nr_tasks;
618
619 return count;
620}
621
622/**
623 * cgroup_task_count - count the number of tasks in a cgroup.
624 * @cgrp: the cgroup in question
625 */
626int cgroup_task_count(const struct cgroup *cgrp)
627{
628 int count;
629
630 spin_lock_irq(&css_set_lock);
631 count = __cgroup_task_count(cgrp);
632 spin_unlock_irq(&css_set_lock);
633
634 return count;
635}
636
637struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
638{
639 struct cgroup *cgrp = of->kn->parent->priv;
640 struct cftype *cft = of_cft(of);
641
642 /*
643 * This is open and unprotected implementation of cgroup_css().
644 * seq_css() is only called from a kernfs file operation which has
645 * an active reference on the file. Because all the subsystem
646 * files are drained before a css is disassociated with a cgroup,
647 * the matching css from the cgroup's subsys table is guaranteed to
648 * be and stay valid until the enclosing operation is complete.
649 */
650 if (cft->ss)
651 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
652 else
653 return &cgrp->self;
654}
655EXPORT_SYMBOL_GPL(of_css);
656
657/**
658 * for_each_css - iterate all css's of a cgroup
659 * @css: the iteration cursor
660 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
661 * @cgrp: the target cgroup to iterate css's of
662 *
663 * Should be called under cgroup_[tree_]mutex.
664 */
665#define for_each_css(css, ssid, cgrp) \
666 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
667 if (!((css) = rcu_dereference_check( \
668 (cgrp)->subsys[(ssid)], \
669 lockdep_is_held(&cgroup_mutex)))) { } \
670 else
671
672/**
673 * for_each_e_css - iterate all effective css's of a cgroup
674 * @css: the iteration cursor
675 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
676 * @cgrp: the target cgroup to iterate css's of
677 *
678 * Should be called under cgroup_[tree_]mutex.
679 */
680#define for_each_e_css(css, ssid, cgrp) \
681 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
682 if (!((css) = cgroup_e_css_by_mask(cgrp, \
683 cgroup_subsys[(ssid)]))) \
684 ; \
685 else
686
687/**
688 * do_each_subsys_mask - filter for_each_subsys with a bitmask
689 * @ss: the iteration cursor
690 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
691 * @ss_mask: the bitmask
692 *
693 * The block will only run for cases where the ssid-th bit (1 << ssid) of
694 * @ss_mask is set.
695 */
696#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
697 unsigned long __ss_mask = (ss_mask); \
698 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
699 (ssid) = 0; \
700 break; \
701 } \
702 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
703 (ss) = cgroup_subsys[ssid]; \
704 {
705
706#define while_each_subsys_mask() \
707 } \
708 } \
709} while (false)
710
711/* iterate over child cgrps, lock should be held throughout iteration */
712#define cgroup_for_each_live_child(child, cgrp) \
713 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
714 if (({ lockdep_assert_held(&cgroup_mutex); \
715 cgroup_is_dead(child); })) \
716 ; \
717 else
718
719/* walk live descendants in pre order */
720#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
721 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
722 if (({ lockdep_assert_held(&cgroup_mutex); \
723 (dsct) = (d_css)->cgroup; \
724 cgroup_is_dead(dsct); })) \
725 ; \
726 else
727
728/* walk live descendants in postorder */
729#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
730 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
731 if (({ lockdep_assert_held(&cgroup_mutex); \
732 (dsct) = (d_css)->cgroup; \
733 cgroup_is_dead(dsct); })) \
734 ; \
735 else
736
737/*
738 * The default css_set - used by init and its children prior to any
739 * hierarchies being mounted. It contains a pointer to the root state
740 * for each subsystem. Also used to anchor the list of css_sets. Not
741 * reference-counted, to improve performance when child cgroups
742 * haven't been created.
743 */
744struct css_set init_css_set = {
745 .refcount = REFCOUNT_INIT(1),
746 .dom_cset = &init_css_set,
747 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
748 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
749 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
750 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
751 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
752 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
753 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
754 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
755
756 /*
757 * The following field is re-initialized when this cset gets linked
758 * in cgroup_init(). However, let's initialize the field
759 * statically too so that the default cgroup can be accessed safely
760 * early during boot.
761 */
762 .dfl_cgrp = &cgrp_dfl_root.cgrp,
763};
764
765static int css_set_count = 1; /* 1 for init_css_set */
766
767static bool css_set_threaded(struct css_set *cset)
768{
769 return cset->dom_cset != cset;
770}
771
772/**
773 * css_set_populated - does a css_set contain any tasks?
774 * @cset: target css_set
775 *
776 * css_set_populated() should be the same as !!cset->nr_tasks at steady
777 * state. However, css_set_populated() can be called while a task is being
778 * added to or removed from the linked list before the nr_tasks is
779 * properly updated. Hence, we can't just look at ->nr_tasks here.
780 */
781static bool css_set_populated(struct css_set *cset)
782{
783 lockdep_assert_held(&css_set_lock);
784
785 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
786}
787
788/**
789 * cgroup_update_populated - update the populated count of a cgroup
790 * @cgrp: the target cgroup
791 * @populated: inc or dec populated count
792 *
793 * One of the css_sets associated with @cgrp is either getting its first
794 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
795 * count is propagated towards root so that a given cgroup's
796 * nr_populated_children is zero iff none of its descendants contain any
797 * tasks.
798 *
799 * @cgrp's interface file "cgroup.populated" is zero if both
800 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
801 * 1 otherwise. When the sum changes from or to zero, userland is notified
802 * that the content of the interface file has changed. This can be used to
803 * detect when @cgrp and its descendants become populated or empty.
804 */
805static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
806{
807 struct cgroup *child = NULL;
808 int adj = populated ? 1 : -1;
809
810 lockdep_assert_held(&css_set_lock);
811
812 do {
813 bool was_populated = cgroup_is_populated(cgrp);
814
815 if (!child) {
816 cgrp->nr_populated_csets += adj;
817 } else {
818 if (cgroup_is_threaded(child))
819 cgrp->nr_populated_threaded_children += adj;
820 else
821 cgrp->nr_populated_domain_children += adj;
822 }
823
824 if (was_populated == cgroup_is_populated(cgrp))
825 break;
826
827 cgroup1_check_for_release(cgrp);
828 TRACE_CGROUP_PATH(notify_populated, cgrp,
829 cgroup_is_populated(cgrp));
830 cgroup_file_notify(&cgrp->events_file);
831
832 child = cgrp;
833 cgrp = cgroup_parent(cgrp);
834 } while (cgrp);
835}
836
837/**
838 * css_set_update_populated - update populated state of a css_set
839 * @cset: target css_set
840 * @populated: whether @cset is populated or depopulated
841 *
842 * @cset is either getting the first task or losing the last. Update the
843 * populated counters of all associated cgroups accordingly.
844 */
845static void css_set_update_populated(struct css_set *cset, bool populated)
846{
847 struct cgrp_cset_link *link;
848
849 lockdep_assert_held(&css_set_lock);
850
851 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
852 cgroup_update_populated(link->cgrp, populated);
853}
854
855/*
856 * @task is leaving, advance task iterators which are pointing to it so
857 * that they can resume at the next position. Advancing an iterator might
858 * remove it from the list, use safe walk. See css_task_iter_skip() for
859 * details.
860 */
861static void css_set_skip_task_iters(struct css_set *cset,
862 struct task_struct *task)
863{
864 struct css_task_iter *it, *pos;
865
866 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
867 css_task_iter_skip(it, task);
868}
869
870/**
871 * css_set_move_task - move a task from one css_set to another
872 * @task: task being moved
873 * @from_cset: css_set @task currently belongs to (may be NULL)
874 * @to_cset: new css_set @task is being moved to (may be NULL)
875 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
876 *
877 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
878 * css_set, @from_cset can be NULL. If @task is being disassociated
879 * instead of moved, @to_cset can be NULL.
880 *
881 * This function automatically handles populated counter updates and
882 * css_task_iter adjustments but the caller is responsible for managing
883 * @from_cset and @to_cset's reference counts.
884 */
885static void css_set_move_task(struct task_struct *task,
886 struct css_set *from_cset, struct css_set *to_cset,
887 bool use_mg_tasks)
888{
889 lockdep_assert_held(&css_set_lock);
890
891 if (to_cset && !css_set_populated(to_cset))
892 css_set_update_populated(to_cset, true);
893
894 if (from_cset) {
895 WARN_ON_ONCE(list_empty(&task->cg_list));
896
897 css_set_skip_task_iters(from_cset, task);
898 list_del_init(&task->cg_list);
899 if (!css_set_populated(from_cset))
900 css_set_update_populated(from_cset, false);
901 } else {
902 WARN_ON_ONCE(!list_empty(&task->cg_list));
903 }
904
905 if (to_cset) {
906 /*
907 * We are synchronized through cgroup_threadgroup_rwsem
908 * against PF_EXITING setting such that we can't race
909 * against cgroup_exit()/cgroup_free() dropping the css_set.
910 */
911 WARN_ON_ONCE(task->flags & PF_EXITING);
912
913 cgroup_move_task(task, to_cset);
914 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
915 &to_cset->tasks);
916 }
917}
918
919/*
920 * hash table for cgroup groups. This improves the performance to find
921 * an existing css_set. This hash doesn't (currently) take into
922 * account cgroups in empty hierarchies.
923 */
924#define CSS_SET_HASH_BITS 7
925static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
926
927static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
928{
929 unsigned long key = 0UL;
930 struct cgroup_subsys *ss;
931 int i;
932
933 for_each_subsys(ss, i)
934 key += (unsigned long)css[i];
935 key = (key >> 16) ^ key;
936
937 return key;
938}
939
940void put_css_set_locked(struct css_set *cset)
941{
942 struct cgrp_cset_link *link, *tmp_link;
943 struct cgroup_subsys *ss;
944 int ssid;
945
946 lockdep_assert_held(&css_set_lock);
947
948 if (!refcount_dec_and_test(&cset->refcount))
949 return;
950
951 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
952
953 /* This css_set is dead. Unlink it and release cgroup and css refs */
954 for_each_subsys(ss, ssid) {
955 list_del(&cset->e_cset_node[ssid]);
956 css_put(cset->subsys[ssid]);
957 }
958 hash_del(&cset->hlist);
959 css_set_count--;
960
961 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
962 list_del(&link->cset_link);
963 list_del(&link->cgrp_link);
964 if (cgroup_parent(link->cgrp))
965 cgroup_put(link->cgrp);
966 kfree(link);
967 }
968
969 if (css_set_threaded(cset)) {
970 list_del(&cset->threaded_csets_node);
971 put_css_set_locked(cset->dom_cset);
972 }
973
974 kfree_rcu(cset, rcu_head);
975}
976
977/**
978 * compare_css_sets - helper function for find_existing_css_set().
979 * @cset: candidate css_set being tested
980 * @old_cset: existing css_set for a task
981 * @new_cgrp: cgroup that's being entered by the task
982 * @template: desired set of css pointers in css_set (pre-calculated)
983 *
984 * Returns true if "cset" matches "old_cset" except for the hierarchy
985 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
986 */
987static bool compare_css_sets(struct css_set *cset,
988 struct css_set *old_cset,
989 struct cgroup *new_cgrp,
990 struct cgroup_subsys_state *template[])
991{
992 struct cgroup *new_dfl_cgrp;
993 struct list_head *l1, *l2;
994
995 /*
996 * On the default hierarchy, there can be csets which are
997 * associated with the same set of cgroups but different csses.
998 * Let's first ensure that csses match.
999 */
1000 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1001 return false;
1002
1003
1004 /* @cset's domain should match the default cgroup's */
1005 if (cgroup_on_dfl(new_cgrp))
1006 new_dfl_cgrp = new_cgrp;
1007 else
1008 new_dfl_cgrp = old_cset->dfl_cgrp;
1009
1010 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1011 return false;
1012
1013 /*
1014 * Compare cgroup pointers in order to distinguish between
1015 * different cgroups in hierarchies. As different cgroups may
1016 * share the same effective css, this comparison is always
1017 * necessary.
1018 */
1019 l1 = &cset->cgrp_links;
1020 l2 = &old_cset->cgrp_links;
1021 while (1) {
1022 struct cgrp_cset_link *link1, *link2;
1023 struct cgroup *cgrp1, *cgrp2;
1024
1025 l1 = l1->next;
1026 l2 = l2->next;
1027 /* See if we reached the end - both lists are equal length. */
1028 if (l1 == &cset->cgrp_links) {
1029 BUG_ON(l2 != &old_cset->cgrp_links);
1030 break;
1031 } else {
1032 BUG_ON(l2 == &old_cset->cgrp_links);
1033 }
1034 /* Locate the cgroups associated with these links. */
1035 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1036 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1037 cgrp1 = link1->cgrp;
1038 cgrp2 = link2->cgrp;
1039 /* Hierarchies should be linked in the same order. */
1040 BUG_ON(cgrp1->root != cgrp2->root);
1041
1042 /*
1043 * If this hierarchy is the hierarchy of the cgroup
1044 * that's changing, then we need to check that this
1045 * css_set points to the new cgroup; if it's any other
1046 * hierarchy, then this css_set should point to the
1047 * same cgroup as the old css_set.
1048 */
1049 if (cgrp1->root == new_cgrp->root) {
1050 if (cgrp1 != new_cgrp)
1051 return false;
1052 } else {
1053 if (cgrp1 != cgrp2)
1054 return false;
1055 }
1056 }
1057 return true;
1058}
1059
1060/**
1061 * find_existing_css_set - init css array and find the matching css_set
1062 * @old_cset: the css_set that we're using before the cgroup transition
1063 * @cgrp: the cgroup that we're moving into
1064 * @template: out param for the new set of csses, should be clear on entry
1065 */
1066static struct css_set *find_existing_css_set(struct css_set *old_cset,
1067 struct cgroup *cgrp,
1068 struct cgroup_subsys_state *template[])
1069{
1070 struct cgroup_root *root = cgrp->root;
1071 struct cgroup_subsys *ss;
1072 struct css_set *cset;
1073 unsigned long key;
1074 int i;
1075
1076 /*
1077 * Build the set of subsystem state objects that we want to see in the
1078 * new css_set. While subsystems can change globally, the entries here
1079 * won't change, so no need for locking.
1080 */
1081 for_each_subsys(ss, i) {
1082 if (root->subsys_mask & (1UL << i)) {
1083 /*
1084 * @ss is in this hierarchy, so we want the
1085 * effective css from @cgrp.
1086 */
1087 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1088 } else {
1089 /*
1090 * @ss is not in this hierarchy, so we don't want
1091 * to change the css.
1092 */
1093 template[i] = old_cset->subsys[i];
1094 }
1095 }
1096
1097 key = css_set_hash(template);
1098 hash_for_each_possible(css_set_table, cset, hlist, key) {
1099 if (!compare_css_sets(cset, old_cset, cgrp, template))
1100 continue;
1101
1102 /* This css_set matches what we need */
1103 return cset;
1104 }
1105
1106 /* No existing cgroup group matched */
1107 return NULL;
1108}
1109
1110static void free_cgrp_cset_links(struct list_head *links_to_free)
1111{
1112 struct cgrp_cset_link *link, *tmp_link;
1113
1114 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1115 list_del(&link->cset_link);
1116 kfree(link);
1117 }
1118}
1119
1120/**
1121 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1122 * @count: the number of links to allocate
1123 * @tmp_links: list_head the allocated links are put on
1124 *
1125 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1126 * through ->cset_link. Returns 0 on success or -errno.
1127 */
1128static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1129{
1130 struct cgrp_cset_link *link;
1131 int i;
1132
1133 INIT_LIST_HEAD(tmp_links);
1134
1135 for (i = 0; i < count; i++) {
1136 link = kzalloc(sizeof(*link), GFP_KERNEL);
1137 if (!link) {
1138 free_cgrp_cset_links(tmp_links);
1139 return -ENOMEM;
1140 }
1141 list_add(&link->cset_link, tmp_links);
1142 }
1143 return 0;
1144}
1145
1146/**
1147 * link_css_set - a helper function to link a css_set to a cgroup
1148 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1149 * @cset: the css_set to be linked
1150 * @cgrp: the destination cgroup
1151 */
1152static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1153 struct cgroup *cgrp)
1154{
1155 struct cgrp_cset_link *link;
1156
1157 BUG_ON(list_empty(tmp_links));
1158
1159 if (cgroup_on_dfl(cgrp))
1160 cset->dfl_cgrp = cgrp;
1161
1162 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1163 link->cset = cset;
1164 link->cgrp = cgrp;
1165
1166 /*
1167 * Always add links to the tail of the lists so that the lists are
1168 * in chronological order.
1169 */
1170 list_move_tail(&link->cset_link, &cgrp->cset_links);
1171 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1172
1173 if (cgroup_parent(cgrp))
1174 cgroup_get_live(cgrp);
1175}
1176
1177/**
1178 * find_css_set - return a new css_set with one cgroup updated
1179 * @old_cset: the baseline css_set
1180 * @cgrp: the cgroup to be updated
1181 *
1182 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1183 * substituted into the appropriate hierarchy.
1184 */
1185static struct css_set *find_css_set(struct css_set *old_cset,
1186 struct cgroup *cgrp)
1187{
1188 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1189 struct css_set *cset;
1190 struct list_head tmp_links;
1191 struct cgrp_cset_link *link;
1192 struct cgroup_subsys *ss;
1193 unsigned long key;
1194 int ssid;
1195
1196 lockdep_assert_held(&cgroup_mutex);
1197
1198 /* First see if we already have a cgroup group that matches
1199 * the desired set */
1200 spin_lock_irq(&css_set_lock);
1201 cset = find_existing_css_set(old_cset, cgrp, template);
1202 if (cset)
1203 get_css_set(cset);
1204 spin_unlock_irq(&css_set_lock);
1205
1206 if (cset)
1207 return cset;
1208
1209 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1210 if (!cset)
1211 return NULL;
1212
1213 /* Allocate all the cgrp_cset_link objects that we'll need */
1214 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1215 kfree(cset);
1216 return NULL;
1217 }
1218
1219 refcount_set(&cset->refcount, 1);
1220 cset->dom_cset = cset;
1221 INIT_LIST_HEAD(&cset->tasks);
1222 INIT_LIST_HEAD(&cset->mg_tasks);
1223 INIT_LIST_HEAD(&cset->dying_tasks);
1224 INIT_LIST_HEAD(&cset->task_iters);
1225 INIT_LIST_HEAD(&cset->threaded_csets);
1226 INIT_HLIST_NODE(&cset->hlist);
1227 INIT_LIST_HEAD(&cset->cgrp_links);
1228 INIT_LIST_HEAD(&cset->mg_preload_node);
1229 INIT_LIST_HEAD(&cset->mg_node);
1230
1231 /* Copy the set of subsystem state objects generated in
1232 * find_existing_css_set() */
1233 memcpy(cset->subsys, template, sizeof(cset->subsys));
1234
1235 spin_lock_irq(&css_set_lock);
1236 /* Add reference counts and links from the new css_set. */
1237 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1238 struct cgroup *c = link->cgrp;
1239
1240 if (c->root == cgrp->root)
1241 c = cgrp;
1242 link_css_set(&tmp_links, cset, c);
1243 }
1244
1245 BUG_ON(!list_empty(&tmp_links));
1246
1247 css_set_count++;
1248
1249 /* Add @cset to the hash table */
1250 key = css_set_hash(cset->subsys);
1251 hash_add(css_set_table, &cset->hlist, key);
1252
1253 for_each_subsys(ss, ssid) {
1254 struct cgroup_subsys_state *css = cset->subsys[ssid];
1255
1256 list_add_tail(&cset->e_cset_node[ssid],
1257 &css->cgroup->e_csets[ssid]);
1258 css_get(css);
1259 }
1260
1261 spin_unlock_irq(&css_set_lock);
1262
1263 /*
1264 * If @cset should be threaded, look up the matching dom_cset and
1265 * link them up. We first fully initialize @cset then look for the
1266 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1267 * to stay empty until we return.
1268 */
1269 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1270 struct css_set *dcset;
1271
1272 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1273 if (!dcset) {
1274 put_css_set(cset);
1275 return NULL;
1276 }
1277
1278 spin_lock_irq(&css_set_lock);
1279 cset->dom_cset = dcset;
1280 list_add_tail(&cset->threaded_csets_node,
1281 &dcset->threaded_csets);
1282 spin_unlock_irq(&css_set_lock);
1283 }
1284
1285 return cset;
1286}
1287
1288struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1289{
1290 struct cgroup *root_cgrp = kf_root->kn->priv;
1291
1292 return root_cgrp->root;
1293}
1294
1295static int cgroup_init_root_id(struct cgroup_root *root)
1296{
1297 int id;
1298
1299 lockdep_assert_held(&cgroup_mutex);
1300
1301 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1302 if (id < 0)
1303 return id;
1304
1305 root->hierarchy_id = id;
1306 return 0;
1307}
1308
1309static void cgroup_exit_root_id(struct cgroup_root *root)
1310{
1311 lockdep_assert_held(&cgroup_mutex);
1312
1313 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1314}
1315
1316void cgroup_free_root(struct cgroup_root *root)
1317{
1318 kfree(root);
1319}
1320
1321static void cgroup_destroy_root(struct cgroup_root *root)
1322{
1323 struct cgroup *cgrp = &root->cgrp;
1324 struct cgrp_cset_link *link, *tmp_link;
1325
1326 trace_cgroup_destroy_root(root);
1327
1328 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1329
1330 BUG_ON(atomic_read(&root->nr_cgrps));
1331 BUG_ON(!list_empty(&cgrp->self.children));
1332
1333 /* Rebind all subsystems back to the default hierarchy */
1334 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1335
1336 /*
1337 * Release all the links from cset_links to this hierarchy's
1338 * root cgroup
1339 */
1340 spin_lock_irq(&css_set_lock);
1341
1342 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1343 list_del(&link->cset_link);
1344 list_del(&link->cgrp_link);
1345 kfree(link);
1346 }
1347
1348 spin_unlock_irq(&css_set_lock);
1349
1350 if (!list_empty(&root->root_list)) {
1351 list_del(&root->root_list);
1352 cgroup_root_count--;
1353 }
1354
1355 cgroup_exit_root_id(root);
1356
1357 mutex_unlock(&cgroup_mutex);
1358
1359 cgroup_rstat_exit(cgrp);
1360 kernfs_destroy_root(root->kf_root);
1361 cgroup_free_root(root);
1362}
1363
1364/*
1365 * look up cgroup associated with current task's cgroup namespace on the
1366 * specified hierarchy
1367 */
1368static struct cgroup *
1369current_cgns_cgroup_from_root(struct cgroup_root *root)
1370{
1371 struct cgroup *res = NULL;
1372 struct css_set *cset;
1373
1374 lockdep_assert_held(&css_set_lock);
1375
1376 rcu_read_lock();
1377
1378 cset = current->nsproxy->cgroup_ns->root_cset;
1379 if (cset == &init_css_set) {
1380 res = &root->cgrp;
1381 } else if (root == &cgrp_dfl_root) {
1382 res = cset->dfl_cgrp;
1383 } else {
1384 struct cgrp_cset_link *link;
1385
1386 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1387 struct cgroup *c = link->cgrp;
1388
1389 if (c->root == root) {
1390 res = c;
1391 break;
1392 }
1393 }
1394 }
1395 rcu_read_unlock();
1396
1397 BUG_ON(!res);
1398 return res;
1399}
1400
1401/* look up cgroup associated with given css_set on the specified hierarchy */
1402static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1403 struct cgroup_root *root)
1404{
1405 struct cgroup *res = NULL;
1406
1407 lockdep_assert_held(&cgroup_mutex);
1408 lockdep_assert_held(&css_set_lock);
1409
1410 if (cset == &init_css_set) {
1411 res = &root->cgrp;
1412 } else if (root == &cgrp_dfl_root) {
1413 res = cset->dfl_cgrp;
1414 } else {
1415 struct cgrp_cset_link *link;
1416
1417 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1418 struct cgroup *c = link->cgrp;
1419
1420 if (c->root == root) {
1421 res = c;
1422 break;
1423 }
1424 }
1425 }
1426
1427 BUG_ON(!res);
1428 return res;
1429}
1430
1431/*
1432 * Return the cgroup for "task" from the given hierarchy. Must be
1433 * called with cgroup_mutex and css_set_lock held.
1434 */
1435struct cgroup *task_cgroup_from_root(struct task_struct *task,
1436 struct cgroup_root *root)
1437{
1438 /*
1439 * No need to lock the task - since we hold css_set_lock the
1440 * task can't change groups.
1441 */
1442 return cset_cgroup_from_root(task_css_set(task), root);
1443}
1444
1445/*
1446 * A task must hold cgroup_mutex to modify cgroups.
1447 *
1448 * Any task can increment and decrement the count field without lock.
1449 * So in general, code holding cgroup_mutex can't rely on the count
1450 * field not changing. However, if the count goes to zero, then only
1451 * cgroup_attach_task() can increment it again. Because a count of zero
1452 * means that no tasks are currently attached, therefore there is no
1453 * way a task attached to that cgroup can fork (the other way to
1454 * increment the count). So code holding cgroup_mutex can safely
1455 * assume that if the count is zero, it will stay zero. Similarly, if
1456 * a task holds cgroup_mutex on a cgroup with zero count, it
1457 * knows that the cgroup won't be removed, as cgroup_rmdir()
1458 * needs that mutex.
1459 *
1460 * A cgroup can only be deleted if both its 'count' of using tasks
1461 * is zero, and its list of 'children' cgroups is empty. Since all
1462 * tasks in the system use _some_ cgroup, and since there is always at
1463 * least one task in the system (init, pid == 1), therefore, root cgroup
1464 * always has either children cgroups and/or using tasks. So we don't
1465 * need a special hack to ensure that root cgroup cannot be deleted.
1466 *
1467 * P.S. One more locking exception. RCU is used to guard the
1468 * update of a tasks cgroup pointer by cgroup_attach_task()
1469 */
1470
1471static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1472
1473static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1474 char *buf)
1475{
1476 struct cgroup_subsys *ss = cft->ss;
1477
1478 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1479 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1480 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1481
1482 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1483 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1484 cft->name);
1485 } else {
1486 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1487 }
1488 return buf;
1489}
1490
1491/**
1492 * cgroup_file_mode - deduce file mode of a control file
1493 * @cft: the control file in question
1494 *
1495 * S_IRUGO for read, S_IWUSR for write.
1496 */
1497static umode_t cgroup_file_mode(const struct cftype *cft)
1498{
1499 umode_t mode = 0;
1500
1501 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1502 mode |= S_IRUGO;
1503
1504 if (cft->write_u64 || cft->write_s64 || cft->write) {
1505 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1506 mode |= S_IWUGO;
1507 else
1508 mode |= S_IWUSR;
1509 }
1510
1511 return mode;
1512}
1513
1514/**
1515 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1516 * @subtree_control: the new subtree_control mask to consider
1517 * @this_ss_mask: available subsystems
1518 *
1519 * On the default hierarchy, a subsystem may request other subsystems to be
1520 * enabled together through its ->depends_on mask. In such cases, more
1521 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1522 *
1523 * This function calculates which subsystems need to be enabled if
1524 * @subtree_control is to be applied while restricted to @this_ss_mask.
1525 */
1526static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1527{
1528 u16 cur_ss_mask = subtree_control;
1529 struct cgroup_subsys *ss;
1530 int ssid;
1531
1532 lockdep_assert_held(&cgroup_mutex);
1533
1534 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1535
1536 while (true) {
1537 u16 new_ss_mask = cur_ss_mask;
1538
1539 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1540 new_ss_mask |= ss->depends_on;
1541 } while_each_subsys_mask();
1542
1543 /*
1544 * Mask out subsystems which aren't available. This can
1545 * happen only if some depended-upon subsystems were bound
1546 * to non-default hierarchies.
1547 */
1548 new_ss_mask &= this_ss_mask;
1549
1550 if (new_ss_mask == cur_ss_mask)
1551 break;
1552 cur_ss_mask = new_ss_mask;
1553 }
1554
1555 return cur_ss_mask;
1556}
1557
1558/**
1559 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1560 * @kn: the kernfs_node being serviced
1561 *
1562 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1563 * the method finishes if locking succeeded. Note that once this function
1564 * returns the cgroup returned by cgroup_kn_lock_live() may become
1565 * inaccessible any time. If the caller intends to continue to access the
1566 * cgroup, it should pin it before invoking this function.
1567 */
1568void cgroup_kn_unlock(struct kernfs_node *kn)
1569{
1570 struct cgroup *cgrp;
1571
1572 if (kernfs_type(kn) == KERNFS_DIR)
1573 cgrp = kn->priv;
1574 else
1575 cgrp = kn->parent->priv;
1576
1577 mutex_unlock(&cgroup_mutex);
1578
1579 kernfs_unbreak_active_protection(kn);
1580 cgroup_put(cgrp);
1581}
1582
1583/**
1584 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1585 * @kn: the kernfs_node being serviced
1586 * @drain_offline: perform offline draining on the cgroup
1587 *
1588 * This helper is to be used by a cgroup kernfs method currently servicing
1589 * @kn. It breaks the active protection, performs cgroup locking and
1590 * verifies that the associated cgroup is alive. Returns the cgroup if
1591 * alive; otherwise, %NULL. A successful return should be undone by a
1592 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1593 * cgroup is drained of offlining csses before return.
1594 *
1595 * Any cgroup kernfs method implementation which requires locking the
1596 * associated cgroup should use this helper. It avoids nesting cgroup
1597 * locking under kernfs active protection and allows all kernfs operations
1598 * including self-removal.
1599 */
1600struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1601{
1602 struct cgroup *cgrp;
1603
1604 if (kernfs_type(kn) == KERNFS_DIR)
1605 cgrp = kn->priv;
1606 else
1607 cgrp = kn->parent->priv;
1608
1609 /*
1610 * We're gonna grab cgroup_mutex which nests outside kernfs
1611 * active_ref. cgroup liveliness check alone provides enough
1612 * protection against removal. Ensure @cgrp stays accessible and
1613 * break the active_ref protection.
1614 */
1615 if (!cgroup_tryget(cgrp))
1616 return NULL;
1617 kernfs_break_active_protection(kn);
1618
1619 if (drain_offline)
1620 cgroup_lock_and_drain_offline(cgrp);
1621 else
1622 mutex_lock(&cgroup_mutex);
1623
1624 if (!cgroup_is_dead(cgrp))
1625 return cgrp;
1626
1627 cgroup_kn_unlock(kn);
1628 return NULL;
1629}
1630
1631static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1632{
1633 char name[CGROUP_FILE_NAME_MAX];
1634
1635 lockdep_assert_held(&cgroup_mutex);
1636
1637 if (cft->file_offset) {
1638 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1639 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1640
1641 spin_lock_irq(&cgroup_file_kn_lock);
1642 cfile->kn = NULL;
1643 spin_unlock_irq(&cgroup_file_kn_lock);
1644
1645 del_timer_sync(&cfile->notify_timer);
1646 }
1647
1648 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1649}
1650
1651/**
1652 * css_clear_dir - remove subsys files in a cgroup directory
1653 * @css: target css
1654 */
1655static void css_clear_dir(struct cgroup_subsys_state *css)
1656{
1657 struct cgroup *cgrp = css->cgroup;
1658 struct cftype *cfts;
1659
1660 if (!(css->flags & CSS_VISIBLE))
1661 return;
1662
1663 css->flags &= ~CSS_VISIBLE;
1664
1665 if (!css->ss) {
1666 if (cgroup_on_dfl(cgrp))
1667 cfts = cgroup_base_files;
1668 else
1669 cfts = cgroup1_base_files;
1670
1671 cgroup_addrm_files(css, cgrp, cfts, false);
1672 } else {
1673 list_for_each_entry(cfts, &css->ss->cfts, node)
1674 cgroup_addrm_files(css, cgrp, cfts, false);
1675 }
1676}
1677
1678/**
1679 * css_populate_dir - create subsys files in a cgroup directory
1680 * @css: target css
1681 *
1682 * On failure, no file is added.
1683 */
1684static int css_populate_dir(struct cgroup_subsys_state *css)
1685{
1686 struct cgroup *cgrp = css->cgroup;
1687 struct cftype *cfts, *failed_cfts;
1688 int ret;
1689
1690 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1691 return 0;
1692
1693 if (!css->ss) {
1694 if (cgroup_on_dfl(cgrp))
1695 cfts = cgroup_base_files;
1696 else
1697 cfts = cgroup1_base_files;
1698
1699 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1700 if (ret < 0)
1701 return ret;
1702 } else {
1703 list_for_each_entry(cfts, &css->ss->cfts, node) {
1704 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1705 if (ret < 0) {
1706 failed_cfts = cfts;
1707 goto err;
1708 }
1709 }
1710 }
1711
1712 css->flags |= CSS_VISIBLE;
1713
1714 return 0;
1715err:
1716 list_for_each_entry(cfts, &css->ss->cfts, node) {
1717 if (cfts == failed_cfts)
1718 break;
1719 cgroup_addrm_files(css, cgrp, cfts, false);
1720 }
1721 return ret;
1722}
1723
1724int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1725{
1726 struct cgroup *dcgrp = &dst_root->cgrp;
1727 struct cgroup_subsys *ss;
1728 int ssid, i, ret;
1729
1730 lockdep_assert_held(&cgroup_mutex);
1731
1732 do_each_subsys_mask(ss, ssid, ss_mask) {
1733 /*
1734 * If @ss has non-root csses attached to it, can't move.
1735 * If @ss is an implicit controller, it is exempt from this
1736 * rule and can be stolen.
1737 */
1738 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1739 !ss->implicit_on_dfl)
1740 return -EBUSY;
1741
1742 /* can't move between two non-dummy roots either */
1743 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1744 return -EBUSY;
1745 } while_each_subsys_mask();
1746
1747 do_each_subsys_mask(ss, ssid, ss_mask) {
1748 struct cgroup_root *src_root = ss->root;
1749 struct cgroup *scgrp = &src_root->cgrp;
1750 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1751 struct css_set *cset;
1752
1753 WARN_ON(!css || cgroup_css(dcgrp, ss));
1754
1755 /* disable from the source */
1756 src_root->subsys_mask &= ~(1 << ssid);
1757 WARN_ON(cgroup_apply_control(scgrp));
1758 cgroup_finalize_control(scgrp, 0);
1759
1760 /* rebind */
1761 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1762 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1763 ss->root = dst_root;
1764 css->cgroup = dcgrp;
1765
1766 spin_lock_irq(&css_set_lock);
1767 hash_for_each(css_set_table, i, cset, hlist)
1768 list_move_tail(&cset->e_cset_node[ss->id],
1769 &dcgrp->e_csets[ss->id]);
1770 spin_unlock_irq(&css_set_lock);
1771
1772 if (ss->css_rstat_flush) {
1773 list_del_rcu(&css->rstat_css_node);
1774 list_add_rcu(&css->rstat_css_node,
1775 &dcgrp->rstat_css_list);
1776 }
1777
1778 /* default hierarchy doesn't enable controllers by default */
1779 dst_root->subsys_mask |= 1 << ssid;
1780 if (dst_root == &cgrp_dfl_root) {
1781 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1782 } else {
1783 dcgrp->subtree_control |= 1 << ssid;
1784 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1785 }
1786
1787 ret = cgroup_apply_control(dcgrp);
1788 if (ret)
1789 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1790 ss->name, ret);
1791
1792 if (ss->bind)
1793 ss->bind(css);
1794 } while_each_subsys_mask();
1795
1796 kernfs_activate(dcgrp->kn);
1797 return 0;
1798}
1799
1800int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1801 struct kernfs_root *kf_root)
1802{
1803 int len = 0;
1804 char *buf = NULL;
1805 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1806 struct cgroup *ns_cgroup;
1807
1808 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1809 if (!buf)
1810 return -ENOMEM;
1811
1812 spin_lock_irq(&css_set_lock);
1813 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1814 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1815 spin_unlock_irq(&css_set_lock);
1816
1817 if (len >= PATH_MAX)
1818 len = -ERANGE;
1819 else if (len > 0) {
1820 seq_escape(sf, buf, " \t\n\\");
1821 len = 0;
1822 }
1823 kfree(buf);
1824 return len;
1825}
1826
1827enum cgroup2_param {
1828 Opt_nsdelegate,
1829 Opt_memory_localevents,
1830 Opt_memory_recursiveprot,
1831 nr__cgroup2_params
1832};
1833
1834static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1835 fsparam_flag("nsdelegate", Opt_nsdelegate),
1836 fsparam_flag("memory_localevents", Opt_memory_localevents),
1837 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1838 {}
1839};
1840
1841static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1842{
1843 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1844 struct fs_parse_result result;
1845 int opt;
1846
1847 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1848 if (opt < 0)
1849 return opt;
1850
1851 switch (opt) {
1852 case Opt_nsdelegate:
1853 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1854 return 0;
1855 case Opt_memory_localevents:
1856 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1857 return 0;
1858 case Opt_memory_recursiveprot:
1859 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1860 return 0;
1861 }
1862 return -EINVAL;
1863}
1864
1865static void apply_cgroup_root_flags(unsigned int root_flags)
1866{
1867 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1868 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1869 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1870 else
1871 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1872
1873 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1874 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1875 else
1876 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1877
1878 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1879 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1880 else
1881 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1882 }
1883}
1884
1885static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1886{
1887 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1888 seq_puts(seq, ",nsdelegate");
1889 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1890 seq_puts(seq, ",memory_localevents");
1891 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1892 seq_puts(seq, ",memory_recursiveprot");
1893 return 0;
1894}
1895
1896static int cgroup_reconfigure(struct fs_context *fc)
1897{
1898 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1899
1900 apply_cgroup_root_flags(ctx->flags);
1901 return 0;
1902}
1903
1904static void init_cgroup_housekeeping(struct cgroup *cgrp)
1905{
1906 struct cgroup_subsys *ss;
1907 int ssid;
1908
1909 INIT_LIST_HEAD(&cgrp->self.sibling);
1910 INIT_LIST_HEAD(&cgrp->self.children);
1911 INIT_LIST_HEAD(&cgrp->cset_links);
1912 INIT_LIST_HEAD(&cgrp->pidlists);
1913 mutex_init(&cgrp->pidlist_mutex);
1914 cgrp->self.cgroup = cgrp;
1915 cgrp->self.flags |= CSS_ONLINE;
1916 cgrp->dom_cgrp = cgrp;
1917 cgrp->max_descendants = INT_MAX;
1918 cgrp->max_depth = INT_MAX;
1919 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1920 prev_cputime_init(&cgrp->prev_cputime);
1921
1922 for_each_subsys(ss, ssid)
1923 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1924
1925 init_waitqueue_head(&cgrp->offline_waitq);
1926 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1927}
1928
1929void init_cgroup_root(struct cgroup_fs_context *ctx)
1930{
1931 struct cgroup_root *root = ctx->root;
1932 struct cgroup *cgrp = &root->cgrp;
1933
1934 INIT_LIST_HEAD(&root->root_list);
1935 atomic_set(&root->nr_cgrps, 1);
1936 cgrp->root = root;
1937 init_cgroup_housekeeping(cgrp);
1938
1939 root->flags = ctx->flags;
1940 if (ctx->release_agent)
1941 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1942 if (ctx->name)
1943 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1944 if (ctx->cpuset_clone_children)
1945 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1946}
1947
1948int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1949{
1950 LIST_HEAD(tmp_links);
1951 struct cgroup *root_cgrp = &root->cgrp;
1952 struct kernfs_syscall_ops *kf_sops;
1953 struct css_set *cset;
1954 int i, ret;
1955
1956 lockdep_assert_held(&cgroup_mutex);
1957
1958 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1959 0, GFP_KERNEL);
1960 if (ret)
1961 goto out;
1962
1963 /*
1964 * We're accessing css_set_count without locking css_set_lock here,
1965 * but that's OK - it can only be increased by someone holding
1966 * cgroup_lock, and that's us. Later rebinding may disable
1967 * controllers on the default hierarchy and thus create new csets,
1968 * which can't be more than the existing ones. Allocate 2x.
1969 */
1970 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1971 if (ret)
1972 goto cancel_ref;
1973
1974 ret = cgroup_init_root_id(root);
1975 if (ret)
1976 goto cancel_ref;
1977
1978 kf_sops = root == &cgrp_dfl_root ?
1979 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1980
1981 root->kf_root = kernfs_create_root(kf_sops,
1982 KERNFS_ROOT_CREATE_DEACTIVATED |
1983 KERNFS_ROOT_SUPPORT_EXPORTOP |
1984 KERNFS_ROOT_SUPPORT_USER_XATTR,
1985 root_cgrp);
1986 if (IS_ERR(root->kf_root)) {
1987 ret = PTR_ERR(root->kf_root);
1988 goto exit_root_id;
1989 }
1990 root_cgrp->kn = root->kf_root->kn;
1991 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
1992 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
1993
1994 ret = css_populate_dir(&root_cgrp->self);
1995 if (ret)
1996 goto destroy_root;
1997
1998 ret = cgroup_rstat_init(root_cgrp);
1999 if (ret)
2000 goto destroy_root;
2001
2002 ret = rebind_subsystems(root, ss_mask);
2003 if (ret)
2004 goto exit_stats;
2005
2006 ret = cgroup_bpf_inherit(root_cgrp);
2007 WARN_ON_ONCE(ret);
2008
2009 trace_cgroup_setup_root(root);
2010
2011 /*
2012 * There must be no failure case after here, since rebinding takes
2013 * care of subsystems' refcounts, which are explicitly dropped in
2014 * the failure exit path.
2015 */
2016 list_add(&root->root_list, &cgroup_roots);
2017 cgroup_root_count++;
2018
2019 /*
2020 * Link the root cgroup in this hierarchy into all the css_set
2021 * objects.
2022 */
2023 spin_lock_irq(&css_set_lock);
2024 hash_for_each(css_set_table, i, cset, hlist) {
2025 link_css_set(&tmp_links, cset, root_cgrp);
2026 if (css_set_populated(cset))
2027 cgroup_update_populated(root_cgrp, true);
2028 }
2029 spin_unlock_irq(&css_set_lock);
2030
2031 BUG_ON(!list_empty(&root_cgrp->self.children));
2032 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2033
2034 ret = 0;
2035 goto out;
2036
2037exit_stats:
2038 cgroup_rstat_exit(root_cgrp);
2039destroy_root:
2040 kernfs_destroy_root(root->kf_root);
2041 root->kf_root = NULL;
2042exit_root_id:
2043 cgroup_exit_root_id(root);
2044cancel_ref:
2045 percpu_ref_exit(&root_cgrp->self.refcnt);
2046out:
2047 free_cgrp_cset_links(&tmp_links);
2048 return ret;
2049}
2050
2051int cgroup_do_get_tree(struct fs_context *fc)
2052{
2053 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2054 int ret;
2055
2056 ctx->kfc.root = ctx->root->kf_root;
2057 if (fc->fs_type == &cgroup2_fs_type)
2058 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2059 else
2060 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2061 ret = kernfs_get_tree(fc);
2062
2063 /*
2064 * In non-init cgroup namespace, instead of root cgroup's dentry,
2065 * we return the dentry corresponding to the cgroupns->root_cgrp.
2066 */
2067 if (!ret && ctx->ns != &init_cgroup_ns) {
2068 struct dentry *nsdentry;
2069 struct super_block *sb = fc->root->d_sb;
2070 struct cgroup *cgrp;
2071
2072 mutex_lock(&cgroup_mutex);
2073 spin_lock_irq(&css_set_lock);
2074
2075 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2076
2077 spin_unlock_irq(&css_set_lock);
2078 mutex_unlock(&cgroup_mutex);
2079
2080 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2081 dput(fc->root);
2082 if (IS_ERR(nsdentry)) {
2083 deactivate_locked_super(sb);
2084 ret = PTR_ERR(nsdentry);
2085 nsdentry = NULL;
2086 }
2087 fc->root = nsdentry;
2088 }
2089
2090 if (!ctx->kfc.new_sb_created)
2091 cgroup_put(&ctx->root->cgrp);
2092
2093 return ret;
2094}
2095
2096/*
2097 * Destroy a cgroup filesystem context.
2098 */
2099static void cgroup_fs_context_free(struct fs_context *fc)
2100{
2101 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2102
2103 kfree(ctx->name);
2104 kfree(ctx->release_agent);
2105 put_cgroup_ns(ctx->ns);
2106 kernfs_free_fs_context(fc);
2107 kfree(ctx);
2108}
2109
2110static int cgroup_get_tree(struct fs_context *fc)
2111{
2112 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2113 int ret;
2114
2115 cgrp_dfl_visible = true;
2116 cgroup_get_live(&cgrp_dfl_root.cgrp);
2117 ctx->root = &cgrp_dfl_root;
2118
2119 ret = cgroup_do_get_tree(fc);
2120 if (!ret)
2121 apply_cgroup_root_flags(ctx->flags);
2122 return ret;
2123}
2124
2125static const struct fs_context_operations cgroup_fs_context_ops = {
2126 .free = cgroup_fs_context_free,
2127 .parse_param = cgroup2_parse_param,
2128 .get_tree = cgroup_get_tree,
2129 .reconfigure = cgroup_reconfigure,
2130};
2131
2132static const struct fs_context_operations cgroup1_fs_context_ops = {
2133 .free = cgroup_fs_context_free,
2134 .parse_param = cgroup1_parse_param,
2135 .get_tree = cgroup1_get_tree,
2136 .reconfigure = cgroup1_reconfigure,
2137};
2138
2139/*
2140 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2141 * we select the namespace we're going to use.
2142 */
2143static int cgroup_init_fs_context(struct fs_context *fc)
2144{
2145 struct cgroup_fs_context *ctx;
2146
2147 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2148 if (!ctx)
2149 return -ENOMEM;
2150
2151 ctx->ns = current->nsproxy->cgroup_ns;
2152 get_cgroup_ns(ctx->ns);
2153 fc->fs_private = &ctx->kfc;
2154 if (fc->fs_type == &cgroup2_fs_type)
2155 fc->ops = &cgroup_fs_context_ops;
2156 else
2157 fc->ops = &cgroup1_fs_context_ops;
2158 put_user_ns(fc->user_ns);
2159 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2160 fc->global = true;
2161 return 0;
2162}
2163
2164static void cgroup_kill_sb(struct super_block *sb)
2165{
2166 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2167 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2168
2169 /*
2170 * If @root doesn't have any children, start killing it.
2171 * This prevents new mounts by disabling percpu_ref_tryget_live().
2172 * cgroup_mount() may wait for @root's release.
2173 *
2174 * And don't kill the default root.
2175 */
2176 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2177 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2178 percpu_ref_kill(&root->cgrp.self.refcnt);
2179 cgroup_put(&root->cgrp);
2180 kernfs_kill_sb(sb);
2181}
2182
2183struct file_system_type cgroup_fs_type = {
2184 .name = "cgroup",
2185 .init_fs_context = cgroup_init_fs_context,
2186 .parameters = cgroup1_fs_parameters,
2187 .kill_sb = cgroup_kill_sb,
2188 .fs_flags = FS_USERNS_MOUNT,
2189};
2190
2191static struct file_system_type cgroup2_fs_type = {
2192 .name = "cgroup2",
2193 .init_fs_context = cgroup_init_fs_context,
2194 .parameters = cgroup2_fs_parameters,
2195 .kill_sb = cgroup_kill_sb,
2196 .fs_flags = FS_USERNS_MOUNT,
2197};
2198
2199#ifdef CONFIG_CPUSETS
2200static const struct fs_context_operations cpuset_fs_context_ops = {
2201 .get_tree = cgroup1_get_tree,
2202 .free = cgroup_fs_context_free,
2203};
2204
2205/*
2206 * This is ugly, but preserves the userspace API for existing cpuset
2207 * users. If someone tries to mount the "cpuset" filesystem, we
2208 * silently switch it to mount "cgroup" instead
2209 */
2210static int cpuset_init_fs_context(struct fs_context *fc)
2211{
2212 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2213 struct cgroup_fs_context *ctx;
2214 int err;
2215
2216 err = cgroup_init_fs_context(fc);
2217 if (err) {
2218 kfree(agent);
2219 return err;
2220 }
2221
2222 fc->ops = &cpuset_fs_context_ops;
2223
2224 ctx = cgroup_fc2context(fc);
2225 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2226 ctx->flags |= CGRP_ROOT_NOPREFIX;
2227 ctx->release_agent = agent;
2228
2229 get_filesystem(&cgroup_fs_type);
2230 put_filesystem(fc->fs_type);
2231 fc->fs_type = &cgroup_fs_type;
2232
2233 return 0;
2234}
2235
2236static struct file_system_type cpuset_fs_type = {
2237 .name = "cpuset",
2238 .init_fs_context = cpuset_init_fs_context,
2239 .fs_flags = FS_USERNS_MOUNT,
2240};
2241#endif
2242
2243int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2244 struct cgroup_namespace *ns)
2245{
2246 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2247
2248 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2249}
2250
2251int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2252 struct cgroup_namespace *ns)
2253{
2254 int ret;
2255
2256 mutex_lock(&cgroup_mutex);
2257 spin_lock_irq(&css_set_lock);
2258
2259 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2260
2261 spin_unlock_irq(&css_set_lock);
2262 mutex_unlock(&cgroup_mutex);
2263
2264 return ret;
2265}
2266EXPORT_SYMBOL_GPL(cgroup_path_ns);
2267
2268/**
2269 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2270 * @task: target task
2271 * @buf: the buffer to write the path into
2272 * @buflen: the length of the buffer
2273 *
2274 * Determine @task's cgroup on the first (the one with the lowest non-zero
2275 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2276 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2277 * cgroup controller callbacks.
2278 *
2279 * Return value is the same as kernfs_path().
2280 */
2281int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2282{
2283 struct cgroup_root *root;
2284 struct cgroup *cgrp;
2285 int hierarchy_id = 1;
2286 int ret;
2287
2288 mutex_lock(&cgroup_mutex);
2289 spin_lock_irq(&css_set_lock);
2290
2291 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2292
2293 if (root) {
2294 cgrp = task_cgroup_from_root(task, root);
2295 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2296 } else {
2297 /* if no hierarchy exists, everyone is in "/" */
2298 ret = strlcpy(buf, "/", buflen);
2299 }
2300
2301 spin_unlock_irq(&css_set_lock);
2302 mutex_unlock(&cgroup_mutex);
2303 return ret;
2304}
2305EXPORT_SYMBOL_GPL(task_cgroup_path);
2306
2307/**
2308 * cgroup_migrate_add_task - add a migration target task to a migration context
2309 * @task: target task
2310 * @mgctx: target migration context
2311 *
2312 * Add @task, which is a migration target, to @mgctx->tset. This function
2313 * becomes noop if @task doesn't need to be migrated. @task's css_set
2314 * should have been added as a migration source and @task->cg_list will be
2315 * moved from the css_set's tasks list to mg_tasks one.
2316 */
2317static void cgroup_migrate_add_task(struct task_struct *task,
2318 struct cgroup_mgctx *mgctx)
2319{
2320 struct css_set *cset;
2321
2322 lockdep_assert_held(&css_set_lock);
2323
2324 /* @task either already exited or can't exit until the end */
2325 if (task->flags & PF_EXITING)
2326 return;
2327
2328 /* cgroup_threadgroup_rwsem protects racing against forks */
2329 WARN_ON_ONCE(list_empty(&task->cg_list));
2330
2331 cset = task_css_set(task);
2332 if (!cset->mg_src_cgrp)
2333 return;
2334
2335 mgctx->tset.nr_tasks++;
2336
2337 list_move_tail(&task->cg_list, &cset->mg_tasks);
2338 if (list_empty(&cset->mg_node))
2339 list_add_tail(&cset->mg_node,
2340 &mgctx->tset.src_csets);
2341 if (list_empty(&cset->mg_dst_cset->mg_node))
2342 list_add_tail(&cset->mg_dst_cset->mg_node,
2343 &mgctx->tset.dst_csets);
2344}
2345
2346/**
2347 * cgroup_taskset_first - reset taskset and return the first task
2348 * @tset: taskset of interest
2349 * @dst_cssp: output variable for the destination css
2350 *
2351 * @tset iteration is initialized and the first task is returned.
2352 */
2353struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2354 struct cgroup_subsys_state **dst_cssp)
2355{
2356 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2357 tset->cur_task = NULL;
2358
2359 return cgroup_taskset_next(tset, dst_cssp);
2360}
2361
2362/**
2363 * cgroup_taskset_next - iterate to the next task in taskset
2364 * @tset: taskset of interest
2365 * @dst_cssp: output variable for the destination css
2366 *
2367 * Return the next task in @tset. Iteration must have been initialized
2368 * with cgroup_taskset_first().
2369 */
2370struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2371 struct cgroup_subsys_state **dst_cssp)
2372{
2373 struct css_set *cset = tset->cur_cset;
2374 struct task_struct *task = tset->cur_task;
2375
2376 while (&cset->mg_node != tset->csets) {
2377 if (!task)
2378 task = list_first_entry(&cset->mg_tasks,
2379 struct task_struct, cg_list);
2380 else
2381 task = list_next_entry(task, cg_list);
2382
2383 if (&task->cg_list != &cset->mg_tasks) {
2384 tset->cur_cset = cset;
2385 tset->cur_task = task;
2386
2387 /*
2388 * This function may be called both before and
2389 * after cgroup_taskset_migrate(). The two cases
2390 * can be distinguished by looking at whether @cset
2391 * has its ->mg_dst_cset set.
2392 */
2393 if (cset->mg_dst_cset)
2394 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2395 else
2396 *dst_cssp = cset->subsys[tset->ssid];
2397
2398 return task;
2399 }
2400
2401 cset = list_next_entry(cset, mg_node);
2402 task = NULL;
2403 }
2404
2405 return NULL;
2406}
2407
2408/**
2409 * cgroup_migrate_execute - migrate a taskset
2410 * @mgctx: migration context
2411 *
2412 * Migrate tasks in @mgctx as setup by migration preparation functions.
2413 * This function fails iff one of the ->can_attach callbacks fails and
2414 * guarantees that either all or none of the tasks in @mgctx are migrated.
2415 * @mgctx is consumed regardless of success.
2416 */
2417static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2418{
2419 struct cgroup_taskset *tset = &mgctx->tset;
2420 struct cgroup_subsys *ss;
2421 struct task_struct *task, *tmp_task;
2422 struct css_set *cset, *tmp_cset;
2423 int ssid, failed_ssid, ret;
2424
2425 /* check that we can legitimately attach to the cgroup */
2426 if (tset->nr_tasks) {
2427 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2428 if (ss->can_attach) {
2429 tset->ssid = ssid;
2430 ret = ss->can_attach(tset);
2431 if (ret) {
2432 failed_ssid = ssid;
2433 goto out_cancel_attach;
2434 }
2435 }
2436 } while_each_subsys_mask();
2437 }
2438
2439 /*
2440 * Now that we're guaranteed success, proceed to move all tasks to
2441 * the new cgroup. There are no failure cases after here, so this
2442 * is the commit point.
2443 */
2444 spin_lock_irq(&css_set_lock);
2445 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2446 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2447 struct css_set *from_cset = task_css_set(task);
2448 struct css_set *to_cset = cset->mg_dst_cset;
2449
2450 get_css_set(to_cset);
2451 to_cset->nr_tasks++;
2452 css_set_move_task(task, from_cset, to_cset, true);
2453 from_cset->nr_tasks--;
2454 /*
2455 * If the source or destination cgroup is frozen,
2456 * the task might require to change its state.
2457 */
2458 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2459 to_cset->dfl_cgrp);
2460 put_css_set_locked(from_cset);
2461
2462 }
2463 }
2464 spin_unlock_irq(&css_set_lock);
2465
2466 /*
2467 * Migration is committed, all target tasks are now on dst_csets.
2468 * Nothing is sensitive to fork() after this point. Notify
2469 * controllers that migration is complete.
2470 */
2471 tset->csets = &tset->dst_csets;
2472
2473 if (tset->nr_tasks) {
2474 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2475 if (ss->attach) {
2476 tset->ssid = ssid;
2477 ss->attach(tset);
2478 }
2479 } while_each_subsys_mask();
2480 }
2481
2482 ret = 0;
2483 goto out_release_tset;
2484
2485out_cancel_attach:
2486 if (tset->nr_tasks) {
2487 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2488 if (ssid == failed_ssid)
2489 break;
2490 if (ss->cancel_attach) {
2491 tset->ssid = ssid;
2492 ss->cancel_attach(tset);
2493 }
2494 } while_each_subsys_mask();
2495 }
2496out_release_tset:
2497 spin_lock_irq(&css_set_lock);
2498 list_splice_init(&tset->dst_csets, &tset->src_csets);
2499 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2500 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2501 list_del_init(&cset->mg_node);
2502 }
2503 spin_unlock_irq(&css_set_lock);
2504
2505 /*
2506 * Re-initialize the cgroup_taskset structure in case it is reused
2507 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2508 * iteration.
2509 */
2510 tset->nr_tasks = 0;
2511 tset->csets = &tset->src_csets;
2512 return ret;
2513}
2514
2515/**
2516 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2517 * @dst_cgrp: destination cgroup to test
2518 *
2519 * On the default hierarchy, except for the mixable, (possible) thread root
2520 * and threaded cgroups, subtree_control must be zero for migration
2521 * destination cgroups with tasks so that child cgroups don't compete
2522 * against tasks.
2523 */
2524int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2525{
2526 /* v1 doesn't have any restriction */
2527 if (!cgroup_on_dfl(dst_cgrp))
2528 return 0;
2529
2530 /* verify @dst_cgrp can host resources */
2531 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2532 return -EOPNOTSUPP;
2533
2534 /* mixables don't care */
2535 if (cgroup_is_mixable(dst_cgrp))
2536 return 0;
2537
2538 /*
2539 * If @dst_cgrp is already or can become a thread root or is
2540 * threaded, it doesn't matter.
2541 */
2542 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2543 return 0;
2544
2545 /* apply no-internal-process constraint */
2546 if (dst_cgrp->subtree_control)
2547 return -EBUSY;
2548
2549 return 0;
2550}
2551
2552/**
2553 * cgroup_migrate_finish - cleanup after attach
2554 * @mgctx: migration context
2555 *
2556 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2557 * those functions for details.
2558 */
2559void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2560{
2561 LIST_HEAD(preloaded);
2562 struct css_set *cset, *tmp_cset;
2563
2564 lockdep_assert_held(&cgroup_mutex);
2565
2566 spin_lock_irq(&css_set_lock);
2567
2568 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2569 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2570
2571 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2572 cset->mg_src_cgrp = NULL;
2573 cset->mg_dst_cgrp = NULL;
2574 cset->mg_dst_cset = NULL;
2575 list_del_init(&cset->mg_preload_node);
2576 put_css_set_locked(cset);
2577 }
2578
2579 spin_unlock_irq(&css_set_lock);
2580}
2581
2582/**
2583 * cgroup_migrate_add_src - add a migration source css_set
2584 * @src_cset: the source css_set to add
2585 * @dst_cgrp: the destination cgroup
2586 * @mgctx: migration context
2587 *
2588 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2589 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2590 * up by cgroup_migrate_finish().
2591 *
2592 * This function may be called without holding cgroup_threadgroup_rwsem
2593 * even if the target is a process. Threads may be created and destroyed
2594 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2595 * into play and the preloaded css_sets are guaranteed to cover all
2596 * migrations.
2597 */
2598void cgroup_migrate_add_src(struct css_set *src_cset,
2599 struct cgroup *dst_cgrp,
2600 struct cgroup_mgctx *mgctx)
2601{
2602 struct cgroup *src_cgrp;
2603
2604 lockdep_assert_held(&cgroup_mutex);
2605 lockdep_assert_held(&css_set_lock);
2606
2607 /*
2608 * If ->dead, @src_set is associated with one or more dead cgroups
2609 * and doesn't contain any migratable tasks. Ignore it early so
2610 * that the rest of migration path doesn't get confused by it.
2611 */
2612 if (src_cset->dead)
2613 return;
2614
2615 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2616
2617 if (!list_empty(&src_cset->mg_preload_node))
2618 return;
2619
2620 WARN_ON(src_cset->mg_src_cgrp);
2621 WARN_ON(src_cset->mg_dst_cgrp);
2622 WARN_ON(!list_empty(&src_cset->mg_tasks));
2623 WARN_ON(!list_empty(&src_cset->mg_node));
2624
2625 src_cset->mg_src_cgrp = src_cgrp;
2626 src_cset->mg_dst_cgrp = dst_cgrp;
2627 get_css_set(src_cset);
2628 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2629}
2630
2631/**
2632 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2633 * @mgctx: migration context
2634 *
2635 * Tasks are about to be moved and all the source css_sets have been
2636 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2637 * pins all destination css_sets, links each to its source, and append them
2638 * to @mgctx->preloaded_dst_csets.
2639 *
2640 * This function must be called after cgroup_migrate_add_src() has been
2641 * called on each migration source css_set. After migration is performed
2642 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2643 * @mgctx.
2644 */
2645int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2646{
2647 struct css_set *src_cset, *tmp_cset;
2648
2649 lockdep_assert_held(&cgroup_mutex);
2650
2651 /* look up the dst cset for each src cset and link it to src */
2652 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2653 mg_preload_node) {
2654 struct css_set *dst_cset;
2655 struct cgroup_subsys *ss;
2656 int ssid;
2657
2658 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2659 if (!dst_cset)
2660 return -ENOMEM;
2661
2662 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2663
2664 /*
2665 * If src cset equals dst, it's noop. Drop the src.
2666 * cgroup_migrate() will skip the cset too. Note that we
2667 * can't handle src == dst as some nodes are used by both.
2668 */
2669 if (src_cset == dst_cset) {
2670 src_cset->mg_src_cgrp = NULL;
2671 src_cset->mg_dst_cgrp = NULL;
2672 list_del_init(&src_cset->mg_preload_node);
2673 put_css_set(src_cset);
2674 put_css_set(dst_cset);
2675 continue;
2676 }
2677
2678 src_cset->mg_dst_cset = dst_cset;
2679
2680 if (list_empty(&dst_cset->mg_preload_node))
2681 list_add_tail(&dst_cset->mg_preload_node,
2682 &mgctx->preloaded_dst_csets);
2683 else
2684 put_css_set(dst_cset);
2685
2686 for_each_subsys(ss, ssid)
2687 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2688 mgctx->ss_mask |= 1 << ssid;
2689 }
2690
2691 return 0;
2692}
2693
2694/**
2695 * cgroup_migrate - migrate a process or task to a cgroup
2696 * @leader: the leader of the process or the task to migrate
2697 * @threadgroup: whether @leader points to the whole process or a single task
2698 * @mgctx: migration context
2699 *
2700 * Migrate a process or task denoted by @leader. If migrating a process,
2701 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2702 * responsible for invoking cgroup_migrate_add_src() and
2703 * cgroup_migrate_prepare_dst() on the targets before invoking this
2704 * function and following up with cgroup_migrate_finish().
2705 *
2706 * As long as a controller's ->can_attach() doesn't fail, this function is
2707 * guaranteed to succeed. This means that, excluding ->can_attach()
2708 * failure, when migrating multiple targets, the success or failure can be
2709 * decided for all targets by invoking group_migrate_prepare_dst() before
2710 * actually starting migrating.
2711 */
2712int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2713 struct cgroup_mgctx *mgctx)
2714{
2715 struct task_struct *task;
2716
2717 /*
2718 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2719 * already PF_EXITING could be freed from underneath us unless we
2720 * take an rcu_read_lock.
2721 */
2722 spin_lock_irq(&css_set_lock);
2723 rcu_read_lock();
2724 task = leader;
2725 do {
2726 cgroup_migrate_add_task(task, mgctx);
2727 if (!threadgroup)
2728 break;
2729 } while_each_thread(leader, task);
2730 rcu_read_unlock();
2731 spin_unlock_irq(&css_set_lock);
2732
2733 return cgroup_migrate_execute(mgctx);
2734}
2735
2736/**
2737 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2738 * @dst_cgrp: the cgroup to attach to
2739 * @leader: the task or the leader of the threadgroup to be attached
2740 * @threadgroup: attach the whole threadgroup?
2741 *
2742 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2743 */
2744int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2745 bool threadgroup)
2746{
2747 DEFINE_CGROUP_MGCTX(mgctx);
2748 struct task_struct *task;
2749 int ret = 0;
2750
2751 /* look up all src csets */
2752 spin_lock_irq(&css_set_lock);
2753 rcu_read_lock();
2754 task = leader;
2755 do {
2756 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2757 if (!threadgroup)
2758 break;
2759 } while_each_thread(leader, task);
2760 rcu_read_unlock();
2761 spin_unlock_irq(&css_set_lock);
2762
2763 /* prepare dst csets and commit */
2764 ret = cgroup_migrate_prepare_dst(&mgctx);
2765 if (!ret)
2766 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2767
2768 cgroup_migrate_finish(&mgctx);
2769
2770 if (!ret)
2771 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2772
2773 return ret;
2774}
2775
2776struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2777 bool *locked)
2778 __acquires(&cgroup_threadgroup_rwsem)
2779{
2780 struct task_struct *tsk;
2781 pid_t pid;
2782
2783 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2784 return ERR_PTR(-EINVAL);
2785
2786 /*
2787 * If we migrate a single thread, we don't care about threadgroup
2788 * stability. If the thread is `current`, it won't exit(2) under our
2789 * hands or change PID through exec(2). We exclude
2790 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2791 * callers by cgroup_mutex.
2792 * Therefore, we can skip the global lock.
2793 */
2794 lockdep_assert_held(&cgroup_mutex);
2795 if (pid || threadgroup) {
2796 percpu_down_write(&cgroup_threadgroup_rwsem);
2797 *locked = true;
2798 } else {
2799 *locked = false;
2800 }
2801
2802 rcu_read_lock();
2803 if (pid) {
2804 tsk = find_task_by_vpid(pid);
2805 if (!tsk) {
2806 tsk = ERR_PTR(-ESRCH);
2807 goto out_unlock_threadgroup;
2808 }
2809 } else {
2810 tsk = current;
2811 }
2812
2813 if (threadgroup)
2814 tsk = tsk->group_leader;
2815
2816 /*
2817 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2818 * If userland migrates such a kthread to a non-root cgroup, it can
2819 * become trapped in a cpuset, or RT kthread may be born in a
2820 * cgroup with no rt_runtime allocated. Just say no.
2821 */
2822 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2823 tsk = ERR_PTR(-EINVAL);
2824 goto out_unlock_threadgroup;
2825 }
2826
2827 get_task_struct(tsk);
2828 goto out_unlock_rcu;
2829
2830out_unlock_threadgroup:
2831 if (*locked) {
2832 percpu_up_write(&cgroup_threadgroup_rwsem);
2833 *locked = false;
2834 }
2835out_unlock_rcu:
2836 rcu_read_unlock();
2837 return tsk;
2838}
2839
2840void cgroup_procs_write_finish(struct task_struct *task, bool locked)
2841 __releases(&cgroup_threadgroup_rwsem)
2842{
2843 struct cgroup_subsys *ss;
2844 int ssid;
2845
2846 /* release reference from cgroup_procs_write_start() */
2847 put_task_struct(task);
2848
2849 if (locked)
2850 percpu_up_write(&cgroup_threadgroup_rwsem);
2851 for_each_subsys(ss, ssid)
2852 if (ss->post_attach)
2853 ss->post_attach();
2854}
2855
2856static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2857{
2858 struct cgroup_subsys *ss;
2859 bool printed = false;
2860 int ssid;
2861
2862 do_each_subsys_mask(ss, ssid, ss_mask) {
2863 if (printed)
2864 seq_putc(seq, ' ');
2865 seq_puts(seq, ss->name);
2866 printed = true;
2867 } while_each_subsys_mask();
2868 if (printed)
2869 seq_putc(seq, '\n');
2870}
2871
2872/* show controllers which are enabled from the parent */
2873static int cgroup_controllers_show(struct seq_file *seq, void *v)
2874{
2875 struct cgroup *cgrp = seq_css(seq)->cgroup;
2876
2877 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2878 return 0;
2879}
2880
2881/* show controllers which are enabled for a given cgroup's children */
2882static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2883{
2884 struct cgroup *cgrp = seq_css(seq)->cgroup;
2885
2886 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2887 return 0;
2888}
2889
2890/**
2891 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2892 * @cgrp: root of the subtree to update csses for
2893 *
2894 * @cgrp's control masks have changed and its subtree's css associations
2895 * need to be updated accordingly. This function looks up all css_sets
2896 * which are attached to the subtree, creates the matching updated css_sets
2897 * and migrates the tasks to the new ones.
2898 */
2899static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2900{
2901 DEFINE_CGROUP_MGCTX(mgctx);
2902 struct cgroup_subsys_state *d_css;
2903 struct cgroup *dsct;
2904 struct css_set *src_cset;
2905 int ret;
2906
2907 lockdep_assert_held(&cgroup_mutex);
2908
2909 percpu_down_write(&cgroup_threadgroup_rwsem);
2910
2911 /* look up all csses currently attached to @cgrp's subtree */
2912 spin_lock_irq(&css_set_lock);
2913 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2914 struct cgrp_cset_link *link;
2915
2916 list_for_each_entry(link, &dsct->cset_links, cset_link)
2917 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2918 }
2919 spin_unlock_irq(&css_set_lock);
2920
2921 /* NULL dst indicates self on default hierarchy */
2922 ret = cgroup_migrate_prepare_dst(&mgctx);
2923 if (ret)
2924 goto out_finish;
2925
2926 spin_lock_irq(&css_set_lock);
2927 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2928 struct task_struct *task, *ntask;
2929
2930 /* all tasks in src_csets need to be migrated */
2931 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2932 cgroup_migrate_add_task(task, &mgctx);
2933 }
2934 spin_unlock_irq(&css_set_lock);
2935
2936 ret = cgroup_migrate_execute(&mgctx);
2937out_finish:
2938 cgroup_migrate_finish(&mgctx);
2939 percpu_up_write(&cgroup_threadgroup_rwsem);
2940 return ret;
2941}
2942
2943/**
2944 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2945 * @cgrp: root of the target subtree
2946 *
2947 * Because css offlining is asynchronous, userland may try to re-enable a
2948 * controller while the previous css is still around. This function grabs
2949 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2950 */
2951void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2952 __acquires(&cgroup_mutex)
2953{
2954 struct cgroup *dsct;
2955 struct cgroup_subsys_state *d_css;
2956 struct cgroup_subsys *ss;
2957 int ssid;
2958
2959restart:
2960 mutex_lock(&cgroup_mutex);
2961
2962 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2963 for_each_subsys(ss, ssid) {
2964 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2965 DEFINE_WAIT(wait);
2966
2967 if (!css || !percpu_ref_is_dying(&css->refcnt))
2968 continue;
2969
2970 cgroup_get_live(dsct);
2971 prepare_to_wait(&dsct->offline_waitq, &wait,
2972 TASK_UNINTERRUPTIBLE);
2973
2974 mutex_unlock(&cgroup_mutex);
2975 schedule();
2976 finish_wait(&dsct->offline_waitq, &wait);
2977
2978 cgroup_put(dsct);
2979 goto restart;
2980 }
2981 }
2982}
2983
2984/**
2985 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2986 * @cgrp: root of the target subtree
2987 *
2988 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2989 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2990 * itself.
2991 */
2992static void cgroup_save_control(struct cgroup *cgrp)
2993{
2994 struct cgroup *dsct;
2995 struct cgroup_subsys_state *d_css;
2996
2997 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2998 dsct->old_subtree_control = dsct->subtree_control;
2999 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3000 dsct->old_dom_cgrp = dsct->dom_cgrp;
3001 }
3002}
3003
3004/**
3005 * cgroup_propagate_control - refresh control masks of a subtree
3006 * @cgrp: root of the target subtree
3007 *
3008 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3009 * ->subtree_control and propagate controller availability through the
3010 * subtree so that descendants don't have unavailable controllers enabled.
3011 */
3012static void cgroup_propagate_control(struct cgroup *cgrp)
3013{
3014 struct cgroup *dsct;
3015 struct cgroup_subsys_state *d_css;
3016
3017 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3018 dsct->subtree_control &= cgroup_control(dsct);
3019 dsct->subtree_ss_mask =
3020 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3021 cgroup_ss_mask(dsct));
3022 }
3023}
3024
3025/**
3026 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3027 * @cgrp: root of the target subtree
3028 *
3029 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3030 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3031 * itself.
3032 */
3033static void cgroup_restore_control(struct cgroup *cgrp)
3034{
3035 struct cgroup *dsct;
3036 struct cgroup_subsys_state *d_css;
3037
3038 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3039 dsct->subtree_control = dsct->old_subtree_control;
3040 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3041 dsct->dom_cgrp = dsct->old_dom_cgrp;
3042 }
3043}
3044
3045static bool css_visible(struct cgroup_subsys_state *css)
3046{
3047 struct cgroup_subsys *ss = css->ss;
3048 struct cgroup *cgrp = css->cgroup;
3049
3050 if (cgroup_control(cgrp) & (1 << ss->id))
3051 return true;
3052 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3053 return false;
3054 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3055}
3056
3057/**
3058 * cgroup_apply_control_enable - enable or show csses according to control
3059 * @cgrp: root of the target subtree
3060 *
3061 * Walk @cgrp's subtree and create new csses or make the existing ones
3062 * visible. A css is created invisible if it's being implicitly enabled
3063 * through dependency. An invisible css is made visible when the userland
3064 * explicitly enables it.
3065 *
3066 * Returns 0 on success, -errno on failure. On failure, csses which have
3067 * been processed already aren't cleaned up. The caller is responsible for
3068 * cleaning up with cgroup_apply_control_disable().
3069 */
3070static int cgroup_apply_control_enable(struct cgroup *cgrp)
3071{
3072 struct cgroup *dsct;
3073 struct cgroup_subsys_state *d_css;
3074 struct cgroup_subsys *ss;
3075 int ssid, ret;
3076
3077 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3078 for_each_subsys(ss, ssid) {
3079 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3080
3081 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3082 continue;
3083
3084 if (!css) {
3085 css = css_create(dsct, ss);
3086 if (IS_ERR(css))
3087 return PTR_ERR(css);
3088 }
3089
3090 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3091
3092 if (css_visible(css)) {
3093 ret = css_populate_dir(css);
3094 if (ret)
3095 return ret;
3096 }
3097 }
3098 }
3099
3100 return 0;
3101}
3102
3103/**
3104 * cgroup_apply_control_disable - kill or hide csses according to control
3105 * @cgrp: root of the target subtree
3106 *
3107 * Walk @cgrp's subtree and kill and hide csses so that they match
3108 * cgroup_ss_mask() and cgroup_visible_mask().
3109 *
3110 * A css is hidden when the userland requests it to be disabled while other
3111 * subsystems are still depending on it. The css must not actively control
3112 * resources and be in the vanilla state if it's made visible again later.
3113 * Controllers which may be depended upon should provide ->css_reset() for
3114 * this purpose.
3115 */
3116static void cgroup_apply_control_disable(struct cgroup *cgrp)
3117{
3118 struct cgroup *dsct;
3119 struct cgroup_subsys_state *d_css;
3120 struct cgroup_subsys *ss;
3121 int ssid;
3122
3123 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3124 for_each_subsys(ss, ssid) {
3125 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3126
3127 if (!css)
3128 continue;
3129
3130 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3131
3132 if (css->parent &&
3133 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3134 kill_css(css);
3135 } else if (!css_visible(css)) {
3136 css_clear_dir(css);
3137 if (ss->css_reset)
3138 ss->css_reset(css);
3139 }
3140 }
3141 }
3142}
3143
3144/**
3145 * cgroup_apply_control - apply control mask updates to the subtree
3146 * @cgrp: root of the target subtree
3147 *
3148 * subsystems can be enabled and disabled in a subtree using the following
3149 * steps.
3150 *
3151 * 1. Call cgroup_save_control() to stash the current state.
3152 * 2. Update ->subtree_control masks in the subtree as desired.
3153 * 3. Call cgroup_apply_control() to apply the changes.
3154 * 4. Optionally perform other related operations.
3155 * 5. Call cgroup_finalize_control() to finish up.
3156 *
3157 * This function implements step 3 and propagates the mask changes
3158 * throughout @cgrp's subtree, updates csses accordingly and perform
3159 * process migrations.
3160 */
3161static int cgroup_apply_control(struct cgroup *cgrp)
3162{
3163 int ret;
3164
3165 cgroup_propagate_control(cgrp);
3166
3167 ret = cgroup_apply_control_enable(cgrp);
3168 if (ret)
3169 return ret;
3170
3171 /*
3172 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3173 * making the following cgroup_update_dfl_csses() properly update
3174 * css associations of all tasks in the subtree.
3175 */
3176 ret = cgroup_update_dfl_csses(cgrp);
3177 if (ret)
3178 return ret;
3179
3180 return 0;
3181}
3182
3183/**
3184 * cgroup_finalize_control - finalize control mask update
3185 * @cgrp: root of the target subtree
3186 * @ret: the result of the update
3187 *
3188 * Finalize control mask update. See cgroup_apply_control() for more info.
3189 */
3190static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3191{
3192 if (ret) {
3193 cgroup_restore_control(cgrp);
3194 cgroup_propagate_control(cgrp);
3195 }
3196
3197 cgroup_apply_control_disable(cgrp);
3198}
3199
3200static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3201{
3202 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3203
3204 /* if nothing is getting enabled, nothing to worry about */
3205 if (!enable)
3206 return 0;
3207
3208 /* can @cgrp host any resources? */
3209 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3210 return -EOPNOTSUPP;
3211
3212 /* mixables don't care */
3213 if (cgroup_is_mixable(cgrp))
3214 return 0;
3215
3216 if (domain_enable) {
3217 /* can't enable domain controllers inside a thread subtree */
3218 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3219 return -EOPNOTSUPP;
3220 } else {
3221 /*
3222 * Threaded controllers can handle internal competitions
3223 * and are always allowed inside a (prospective) thread
3224 * subtree.
3225 */
3226 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3227 return 0;
3228 }
3229
3230 /*
3231 * Controllers can't be enabled for a cgroup with tasks to avoid
3232 * child cgroups competing against tasks.
3233 */
3234 if (cgroup_has_tasks(cgrp))
3235 return -EBUSY;
3236
3237 return 0;
3238}
3239
3240/* change the enabled child controllers for a cgroup in the default hierarchy */
3241static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3242 char *buf, size_t nbytes,
3243 loff_t off)
3244{
3245 u16 enable = 0, disable = 0;
3246 struct cgroup *cgrp, *child;
3247 struct cgroup_subsys *ss;
3248 char *tok;
3249 int ssid, ret;
3250
3251 /*
3252 * Parse input - space separated list of subsystem names prefixed
3253 * with either + or -.
3254 */
3255 buf = strstrip(buf);
3256 while ((tok = strsep(&buf, " "))) {
3257 if (tok[0] == '\0')
3258 continue;
3259 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3260 if (!cgroup_ssid_enabled(ssid) ||
3261 strcmp(tok + 1, ss->name))
3262 continue;
3263
3264 if (*tok == '+') {
3265 enable |= 1 << ssid;
3266 disable &= ~(1 << ssid);
3267 } else if (*tok == '-') {
3268 disable |= 1 << ssid;
3269 enable &= ~(1 << ssid);
3270 } else {
3271 return -EINVAL;
3272 }
3273 break;
3274 } while_each_subsys_mask();
3275 if (ssid == CGROUP_SUBSYS_COUNT)
3276 return -EINVAL;
3277 }
3278
3279 cgrp = cgroup_kn_lock_live(of->kn, true);
3280 if (!cgrp)
3281 return -ENODEV;
3282
3283 for_each_subsys(ss, ssid) {
3284 if (enable & (1 << ssid)) {
3285 if (cgrp->subtree_control & (1 << ssid)) {
3286 enable &= ~(1 << ssid);
3287 continue;
3288 }
3289
3290 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3291 ret = -ENOENT;
3292 goto out_unlock;
3293 }
3294 } else if (disable & (1 << ssid)) {
3295 if (!(cgrp->subtree_control & (1 << ssid))) {
3296 disable &= ~(1 << ssid);
3297 continue;
3298 }
3299
3300 /* a child has it enabled? */
3301 cgroup_for_each_live_child(child, cgrp) {
3302 if (child->subtree_control & (1 << ssid)) {
3303 ret = -EBUSY;
3304 goto out_unlock;
3305 }
3306 }
3307 }
3308 }
3309
3310 if (!enable && !disable) {
3311 ret = 0;
3312 goto out_unlock;
3313 }
3314
3315 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3316 if (ret)
3317 goto out_unlock;
3318
3319 /* save and update control masks and prepare csses */
3320 cgroup_save_control(cgrp);
3321
3322 cgrp->subtree_control |= enable;
3323 cgrp->subtree_control &= ~disable;
3324
3325 ret = cgroup_apply_control(cgrp);
3326 cgroup_finalize_control(cgrp, ret);
3327 if (ret)
3328 goto out_unlock;
3329
3330 kernfs_activate(cgrp->kn);
3331out_unlock:
3332 cgroup_kn_unlock(of->kn);
3333 return ret ?: nbytes;
3334}
3335
3336/**
3337 * cgroup_enable_threaded - make @cgrp threaded
3338 * @cgrp: the target cgroup
3339 *
3340 * Called when "threaded" is written to the cgroup.type interface file and
3341 * tries to make @cgrp threaded and join the parent's resource domain.
3342 * This function is never called on the root cgroup as cgroup.type doesn't
3343 * exist on it.
3344 */
3345static int cgroup_enable_threaded(struct cgroup *cgrp)
3346{
3347 struct cgroup *parent = cgroup_parent(cgrp);
3348 struct cgroup *dom_cgrp = parent->dom_cgrp;
3349 struct cgroup *dsct;
3350 struct cgroup_subsys_state *d_css;
3351 int ret;
3352
3353 lockdep_assert_held(&cgroup_mutex);
3354
3355 /* noop if already threaded */
3356 if (cgroup_is_threaded(cgrp))
3357 return 0;
3358
3359 /*
3360 * If @cgroup is populated or has domain controllers enabled, it
3361 * can't be switched. While the below cgroup_can_be_thread_root()
3362 * test can catch the same conditions, that's only when @parent is
3363 * not mixable, so let's check it explicitly.
3364 */
3365 if (cgroup_is_populated(cgrp) ||
3366 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3367 return -EOPNOTSUPP;
3368
3369 /* we're joining the parent's domain, ensure its validity */
3370 if (!cgroup_is_valid_domain(dom_cgrp) ||
3371 !cgroup_can_be_thread_root(dom_cgrp))
3372 return -EOPNOTSUPP;
3373
3374 /*
3375 * The following shouldn't cause actual migrations and should
3376 * always succeed.
3377 */
3378 cgroup_save_control(cgrp);
3379
3380 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3381 if (dsct == cgrp || cgroup_is_threaded(dsct))
3382 dsct->dom_cgrp = dom_cgrp;
3383
3384 ret = cgroup_apply_control(cgrp);
3385 if (!ret)
3386 parent->nr_threaded_children++;
3387
3388 cgroup_finalize_control(cgrp, ret);
3389 return ret;
3390}
3391
3392static int cgroup_type_show(struct seq_file *seq, void *v)
3393{
3394 struct cgroup *cgrp = seq_css(seq)->cgroup;
3395
3396 if (cgroup_is_threaded(cgrp))
3397 seq_puts(seq, "threaded\n");
3398 else if (!cgroup_is_valid_domain(cgrp))
3399 seq_puts(seq, "domain invalid\n");
3400 else if (cgroup_is_thread_root(cgrp))
3401 seq_puts(seq, "domain threaded\n");
3402 else
3403 seq_puts(seq, "domain\n");
3404
3405 return 0;
3406}
3407
3408static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3409 size_t nbytes, loff_t off)
3410{
3411 struct cgroup *cgrp;
3412 int ret;
3413
3414 /* only switching to threaded mode is supported */
3415 if (strcmp(strstrip(buf), "threaded"))
3416 return -EINVAL;
3417
3418 /* drain dying csses before we re-apply (threaded) subtree control */
3419 cgrp = cgroup_kn_lock_live(of->kn, true);
3420 if (!cgrp)
3421 return -ENOENT;
3422
3423 /* threaded can only be enabled */
3424 ret = cgroup_enable_threaded(cgrp);
3425
3426 cgroup_kn_unlock(of->kn);
3427 return ret ?: nbytes;
3428}
3429
3430static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3431{
3432 struct cgroup *cgrp = seq_css(seq)->cgroup;
3433 int descendants = READ_ONCE(cgrp->max_descendants);
3434
3435 if (descendants == INT_MAX)
3436 seq_puts(seq, "max\n");
3437 else
3438 seq_printf(seq, "%d\n", descendants);
3439
3440 return 0;
3441}
3442
3443static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3444 char *buf, size_t nbytes, loff_t off)
3445{
3446 struct cgroup *cgrp;
3447 int descendants;
3448 ssize_t ret;
3449
3450 buf = strstrip(buf);
3451 if (!strcmp(buf, "max")) {
3452 descendants = INT_MAX;
3453 } else {
3454 ret = kstrtoint(buf, 0, &descendants);
3455 if (ret)
3456 return ret;
3457 }
3458
3459 if (descendants < 0)
3460 return -ERANGE;
3461
3462 cgrp = cgroup_kn_lock_live(of->kn, false);
3463 if (!cgrp)
3464 return -ENOENT;
3465
3466 cgrp->max_descendants = descendants;
3467
3468 cgroup_kn_unlock(of->kn);
3469
3470 return nbytes;
3471}
3472
3473static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3474{
3475 struct cgroup *cgrp = seq_css(seq)->cgroup;
3476 int depth = READ_ONCE(cgrp->max_depth);
3477
3478 if (depth == INT_MAX)
3479 seq_puts(seq, "max\n");
3480 else
3481 seq_printf(seq, "%d\n", depth);
3482
3483 return 0;
3484}
3485
3486static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3487 char *buf, size_t nbytes, loff_t off)
3488{
3489 struct cgroup *cgrp;
3490 ssize_t ret;
3491 int depth;
3492
3493 buf = strstrip(buf);
3494 if (!strcmp(buf, "max")) {
3495 depth = INT_MAX;
3496 } else {
3497 ret = kstrtoint(buf, 0, &depth);
3498 if (ret)
3499 return ret;
3500 }
3501
3502 if (depth < 0)
3503 return -ERANGE;
3504
3505 cgrp = cgroup_kn_lock_live(of->kn, false);
3506 if (!cgrp)
3507 return -ENOENT;
3508
3509 cgrp->max_depth = depth;
3510
3511 cgroup_kn_unlock(of->kn);
3512
3513 return nbytes;
3514}
3515
3516static int cgroup_events_show(struct seq_file *seq, void *v)
3517{
3518 struct cgroup *cgrp = seq_css(seq)->cgroup;
3519
3520 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3521 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3522
3523 return 0;
3524}
3525
3526static int cgroup_stat_show(struct seq_file *seq, void *v)
3527{
3528 struct cgroup *cgroup = seq_css(seq)->cgroup;
3529
3530 seq_printf(seq, "nr_descendants %d\n",
3531 cgroup->nr_descendants);
3532 seq_printf(seq, "nr_dying_descendants %d\n",
3533 cgroup->nr_dying_descendants);
3534
3535 return 0;
3536}
3537
3538static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3539 struct cgroup *cgrp, int ssid)
3540{
3541 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3542 struct cgroup_subsys_state *css;
3543 int ret;
3544
3545 if (!ss->css_extra_stat_show)
3546 return 0;
3547
3548 css = cgroup_tryget_css(cgrp, ss);
3549 if (!css)
3550 return 0;
3551
3552 ret = ss->css_extra_stat_show(seq, css);
3553 css_put(css);
3554 return ret;
3555}
3556
3557static int cpu_stat_show(struct seq_file *seq, void *v)
3558{
3559 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3560 int ret = 0;
3561
3562 cgroup_base_stat_cputime_show(seq);
3563#ifdef CONFIG_CGROUP_SCHED
3564 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3565#endif
3566 return ret;
3567}
3568
3569#ifdef CONFIG_PSI
3570static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3571{
3572 struct cgroup *cgrp = seq_css(seq)->cgroup;
3573 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3574
3575 return psi_show(seq, psi, PSI_IO);
3576}
3577static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3578{
3579 struct cgroup *cgrp = seq_css(seq)->cgroup;
3580 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3581
3582 return psi_show(seq, psi, PSI_MEM);
3583}
3584static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3585{
3586 struct cgroup *cgrp = seq_css(seq)->cgroup;
3587 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3588
3589 return psi_show(seq, psi, PSI_CPU);
3590}
3591
3592static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3593 size_t nbytes, enum psi_res res)
3594{
3595 struct psi_trigger *new;
3596 struct cgroup *cgrp;
3597 struct psi_group *psi;
3598
3599 cgrp = cgroup_kn_lock_live(of->kn, false);
3600 if (!cgrp)
3601 return -ENODEV;
3602
3603 cgroup_get(cgrp);
3604 cgroup_kn_unlock(of->kn);
3605
3606 psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3607 new = psi_trigger_create(psi, buf, nbytes, res);
3608 if (IS_ERR(new)) {
3609 cgroup_put(cgrp);
3610 return PTR_ERR(new);
3611 }
3612
3613 psi_trigger_replace(&of->priv, new);
3614
3615 cgroup_put(cgrp);
3616
3617 return nbytes;
3618}
3619
3620static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3621 char *buf, size_t nbytes,
3622 loff_t off)
3623{
3624 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3625}
3626
3627static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3628 char *buf, size_t nbytes,
3629 loff_t off)
3630{
3631 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3632}
3633
3634static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3635 char *buf, size_t nbytes,
3636 loff_t off)
3637{
3638 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3639}
3640
3641static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3642 poll_table *pt)
3643{
3644 return psi_trigger_poll(&of->priv, of->file, pt);
3645}
3646
3647static void cgroup_pressure_release(struct kernfs_open_file *of)
3648{
3649 psi_trigger_replace(&of->priv, NULL);
3650}
3651
3652bool cgroup_psi_enabled(void)
3653{
3654 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3655}
3656
3657#else /* CONFIG_PSI */
3658bool cgroup_psi_enabled(void)
3659{
3660 return false;
3661}
3662
3663#endif /* CONFIG_PSI */
3664
3665static int cgroup_freeze_show(struct seq_file *seq, void *v)
3666{
3667 struct cgroup *cgrp = seq_css(seq)->cgroup;
3668
3669 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3670
3671 return 0;
3672}
3673
3674static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3675 char *buf, size_t nbytes, loff_t off)
3676{
3677 struct cgroup *cgrp;
3678 ssize_t ret;
3679 int freeze;
3680
3681 ret = kstrtoint(strstrip(buf), 0, &freeze);
3682 if (ret)
3683 return ret;
3684
3685 if (freeze < 0 || freeze > 1)
3686 return -ERANGE;
3687
3688 cgrp = cgroup_kn_lock_live(of->kn, false);
3689 if (!cgrp)
3690 return -ENOENT;
3691
3692 cgroup_freeze(cgrp, freeze);
3693
3694 cgroup_kn_unlock(of->kn);
3695
3696 return nbytes;
3697}
3698
3699static void __cgroup_kill(struct cgroup *cgrp)
3700{
3701 struct css_task_iter it;
3702 struct task_struct *task;
3703
3704 lockdep_assert_held(&cgroup_mutex);
3705
3706 spin_lock_irq(&css_set_lock);
3707 set_bit(CGRP_KILL, &cgrp->flags);
3708 spin_unlock_irq(&css_set_lock);
3709
3710 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3711 while ((task = css_task_iter_next(&it))) {
3712 /* Ignore kernel threads here. */
3713 if (task->flags & PF_KTHREAD)
3714 continue;
3715
3716 /* Skip tasks that are already dying. */
3717 if (__fatal_signal_pending(task))
3718 continue;
3719
3720 send_sig(SIGKILL, task, 0);
3721 }
3722 css_task_iter_end(&it);
3723
3724 spin_lock_irq(&css_set_lock);
3725 clear_bit(CGRP_KILL, &cgrp->flags);
3726 spin_unlock_irq(&css_set_lock);
3727}
3728
3729static void cgroup_kill(struct cgroup *cgrp)
3730{
3731 struct cgroup_subsys_state *css;
3732 struct cgroup *dsct;
3733
3734 lockdep_assert_held(&cgroup_mutex);
3735
3736 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3737 __cgroup_kill(dsct);
3738}
3739
3740static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3741 size_t nbytes, loff_t off)
3742{
3743 ssize_t ret = 0;
3744 int kill;
3745 struct cgroup *cgrp;
3746
3747 ret = kstrtoint(strstrip(buf), 0, &kill);
3748 if (ret)
3749 return ret;
3750
3751 if (kill != 1)
3752 return -ERANGE;
3753
3754 cgrp = cgroup_kn_lock_live(of->kn, false);
3755 if (!cgrp)
3756 return -ENOENT;
3757
3758 /*
3759 * Killing is a process directed operation, i.e. the whole thread-group
3760 * is taken down so act like we do for cgroup.procs and only make this
3761 * writable in non-threaded cgroups.
3762 */
3763 if (cgroup_is_threaded(cgrp))
3764 ret = -EOPNOTSUPP;
3765 else
3766 cgroup_kill(cgrp);
3767
3768 cgroup_kn_unlock(of->kn);
3769
3770 return ret ?: nbytes;
3771}
3772
3773static int cgroup_file_open(struct kernfs_open_file *of)
3774{
3775 struct cftype *cft = of_cft(of);
3776
3777 if (cft->open)
3778 return cft->open(of);
3779 return 0;
3780}
3781
3782static void cgroup_file_release(struct kernfs_open_file *of)
3783{
3784 struct cftype *cft = of_cft(of);
3785
3786 if (cft->release)
3787 cft->release(of);
3788}
3789
3790static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3791 size_t nbytes, loff_t off)
3792{
3793 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3794 struct cgroup *cgrp = of->kn->parent->priv;
3795 struct cftype *cft = of_cft(of);
3796 struct cgroup_subsys_state *css;
3797 int ret;
3798
3799 if (!nbytes)
3800 return 0;
3801
3802 /*
3803 * If namespaces are delegation boundaries, disallow writes to
3804 * files in an non-init namespace root from inside the namespace
3805 * except for the files explicitly marked delegatable -
3806 * cgroup.procs and cgroup.subtree_control.
3807 */
3808 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3809 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3810 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3811 return -EPERM;
3812
3813 if (cft->write)
3814 return cft->write(of, buf, nbytes, off);
3815
3816 /*
3817 * kernfs guarantees that a file isn't deleted with operations in
3818 * flight, which means that the matching css is and stays alive and
3819 * doesn't need to be pinned. The RCU locking is not necessary
3820 * either. It's just for the convenience of using cgroup_css().
3821 */
3822 rcu_read_lock();
3823 css = cgroup_css(cgrp, cft->ss);
3824 rcu_read_unlock();
3825
3826 if (cft->write_u64) {
3827 unsigned long long v;
3828 ret = kstrtoull(buf, 0, &v);
3829 if (!ret)
3830 ret = cft->write_u64(css, cft, v);
3831 } else if (cft->write_s64) {
3832 long long v;
3833 ret = kstrtoll(buf, 0, &v);
3834 if (!ret)
3835 ret = cft->write_s64(css, cft, v);
3836 } else {
3837 ret = -EINVAL;
3838 }
3839
3840 return ret ?: nbytes;
3841}
3842
3843static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3844{
3845 struct cftype *cft = of_cft(of);
3846
3847 if (cft->poll)
3848 return cft->poll(of, pt);
3849
3850 return kernfs_generic_poll(of, pt);
3851}
3852
3853static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3854{
3855 return seq_cft(seq)->seq_start(seq, ppos);
3856}
3857
3858static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3859{
3860 return seq_cft(seq)->seq_next(seq, v, ppos);
3861}
3862
3863static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3864{
3865 if (seq_cft(seq)->seq_stop)
3866 seq_cft(seq)->seq_stop(seq, v);
3867}
3868
3869static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3870{
3871 struct cftype *cft = seq_cft(m);
3872 struct cgroup_subsys_state *css = seq_css(m);
3873
3874 if (cft->seq_show)
3875 return cft->seq_show(m, arg);
3876
3877 if (cft->read_u64)
3878 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3879 else if (cft->read_s64)
3880 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3881 else
3882 return -EINVAL;
3883 return 0;
3884}
3885
3886static struct kernfs_ops cgroup_kf_single_ops = {
3887 .atomic_write_len = PAGE_SIZE,
3888 .open = cgroup_file_open,
3889 .release = cgroup_file_release,
3890 .write = cgroup_file_write,
3891 .poll = cgroup_file_poll,
3892 .seq_show = cgroup_seqfile_show,
3893};
3894
3895static struct kernfs_ops cgroup_kf_ops = {
3896 .atomic_write_len = PAGE_SIZE,
3897 .open = cgroup_file_open,
3898 .release = cgroup_file_release,
3899 .write = cgroup_file_write,
3900 .poll = cgroup_file_poll,
3901 .seq_start = cgroup_seqfile_start,
3902 .seq_next = cgroup_seqfile_next,
3903 .seq_stop = cgroup_seqfile_stop,
3904 .seq_show = cgroup_seqfile_show,
3905};
3906
3907/* set uid and gid of cgroup dirs and files to that of the creator */
3908static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3909{
3910 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3911 .ia_uid = current_fsuid(),
3912 .ia_gid = current_fsgid(), };
3913
3914 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3915 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3916 return 0;
3917
3918 return kernfs_setattr(kn, &iattr);
3919}
3920
3921static void cgroup_file_notify_timer(struct timer_list *timer)
3922{
3923 cgroup_file_notify(container_of(timer, struct cgroup_file,
3924 notify_timer));
3925}
3926
3927static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3928 struct cftype *cft)
3929{
3930 char name[CGROUP_FILE_NAME_MAX];
3931 struct kernfs_node *kn;
3932 struct lock_class_key *key = NULL;
3933 int ret;
3934
3935#ifdef CONFIG_DEBUG_LOCK_ALLOC
3936 key = &cft->lockdep_key;
3937#endif
3938 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3939 cgroup_file_mode(cft),
3940 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3941 0, cft->kf_ops, cft,
3942 NULL, key);
3943 if (IS_ERR(kn))
3944 return PTR_ERR(kn);
3945
3946 ret = cgroup_kn_set_ugid(kn);
3947 if (ret) {
3948 kernfs_remove(kn);
3949 return ret;
3950 }
3951
3952 if (cft->file_offset) {
3953 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3954
3955 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3956
3957 spin_lock_irq(&cgroup_file_kn_lock);
3958 cfile->kn = kn;
3959 spin_unlock_irq(&cgroup_file_kn_lock);
3960 }
3961
3962 return 0;
3963}
3964
3965/**
3966 * cgroup_addrm_files - add or remove files to a cgroup directory
3967 * @css: the target css
3968 * @cgrp: the target cgroup (usually css->cgroup)
3969 * @cfts: array of cftypes to be added
3970 * @is_add: whether to add or remove
3971 *
3972 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3973 * For removals, this function never fails.
3974 */
3975static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3976 struct cgroup *cgrp, struct cftype cfts[],
3977 bool is_add)
3978{
3979 struct cftype *cft, *cft_end = NULL;
3980 int ret = 0;
3981
3982 lockdep_assert_held(&cgroup_mutex);
3983
3984restart:
3985 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3986 /* does cft->flags tell us to skip this file on @cgrp? */
3987 if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
3988 continue;
3989 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3990 continue;
3991 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3992 continue;
3993 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3994 continue;
3995 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3996 continue;
3997 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3998 continue;
3999 if (is_add) {
4000 ret = cgroup_add_file(css, cgrp, cft);
4001 if (ret) {
4002 pr_warn("%s: failed to add %s, err=%d\n",
4003 __func__, cft->name, ret);
4004 cft_end = cft;
4005 is_add = false;
4006 goto restart;
4007 }
4008 } else {
4009 cgroup_rm_file(cgrp, cft);
4010 }
4011 }
4012 return ret;
4013}
4014
4015static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4016{
4017 struct cgroup_subsys *ss = cfts[0].ss;
4018 struct cgroup *root = &ss->root->cgrp;
4019 struct cgroup_subsys_state *css;
4020 int ret = 0;
4021
4022 lockdep_assert_held(&cgroup_mutex);
4023
4024 /* add/rm files for all cgroups created before */
4025 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4026 struct cgroup *cgrp = css->cgroup;
4027
4028 if (!(css->flags & CSS_VISIBLE))
4029 continue;
4030
4031 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4032 if (ret)
4033 break;
4034 }
4035
4036 if (is_add && !ret)
4037 kernfs_activate(root->kn);
4038 return ret;
4039}
4040
4041static void cgroup_exit_cftypes(struct cftype *cfts)
4042{
4043 struct cftype *cft;
4044
4045 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4046 /* free copy for custom atomic_write_len, see init_cftypes() */
4047 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4048 kfree(cft->kf_ops);
4049 cft->kf_ops = NULL;
4050 cft->ss = NULL;
4051
4052 /* revert flags set by cgroup core while adding @cfts */
4053 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
4054 }
4055}
4056
4057static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4058{
4059 struct cftype *cft;
4060
4061 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4062 struct kernfs_ops *kf_ops;
4063
4064 WARN_ON(cft->ss || cft->kf_ops);
4065
4066 if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
4067 continue;
4068
4069 if (cft->seq_start)
4070 kf_ops = &cgroup_kf_ops;
4071 else
4072 kf_ops = &cgroup_kf_single_ops;
4073
4074 /*
4075 * Ugh... if @cft wants a custom max_write_len, we need to
4076 * make a copy of kf_ops to set its atomic_write_len.
4077 */
4078 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4079 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4080 if (!kf_ops) {
4081 cgroup_exit_cftypes(cfts);
4082 return -ENOMEM;
4083 }
4084 kf_ops->atomic_write_len = cft->max_write_len;
4085 }
4086
4087 cft->kf_ops = kf_ops;
4088 cft->ss = ss;
4089 }
4090
4091 return 0;
4092}
4093
4094static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4095{
4096 lockdep_assert_held(&cgroup_mutex);
4097
4098 if (!cfts || !cfts[0].ss)
4099 return -ENOENT;
4100
4101 list_del(&cfts->node);
4102 cgroup_apply_cftypes(cfts, false);
4103 cgroup_exit_cftypes(cfts);
4104 return 0;
4105}
4106
4107/**
4108 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4109 * @cfts: zero-length name terminated array of cftypes
4110 *
4111 * Unregister @cfts. Files described by @cfts are removed from all
4112 * existing cgroups and all future cgroups won't have them either. This
4113 * function can be called anytime whether @cfts' subsys is attached or not.
4114 *
4115 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4116 * registered.
4117 */
4118int cgroup_rm_cftypes(struct cftype *cfts)
4119{
4120 int ret;
4121
4122 mutex_lock(&cgroup_mutex);
4123 ret = cgroup_rm_cftypes_locked(cfts);
4124 mutex_unlock(&cgroup_mutex);
4125 return ret;
4126}
4127
4128/**
4129 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4130 * @ss: target cgroup subsystem
4131 * @cfts: zero-length name terminated array of cftypes
4132 *
4133 * Register @cfts to @ss. Files described by @cfts are created for all
4134 * existing cgroups to which @ss is attached and all future cgroups will
4135 * have them too. This function can be called anytime whether @ss is
4136 * attached or not.
4137 *
4138 * Returns 0 on successful registration, -errno on failure. Note that this
4139 * function currently returns 0 as long as @cfts registration is successful
4140 * even if some file creation attempts on existing cgroups fail.
4141 */
4142static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4143{
4144 int ret;
4145
4146 if (!cgroup_ssid_enabled(ss->id))
4147 return 0;
4148
4149 if (!cfts || cfts[0].name[0] == '\0')
4150 return 0;
4151
4152 ret = cgroup_init_cftypes(ss, cfts);
4153 if (ret)
4154 return ret;
4155
4156 mutex_lock(&cgroup_mutex);
4157
4158 list_add_tail(&cfts->node, &ss->cfts);
4159 ret = cgroup_apply_cftypes(cfts, true);
4160 if (ret)
4161 cgroup_rm_cftypes_locked(cfts);
4162
4163 mutex_unlock(&cgroup_mutex);
4164 return ret;
4165}
4166
4167/**
4168 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4169 * @ss: target cgroup subsystem
4170 * @cfts: zero-length name terminated array of cftypes
4171 *
4172 * Similar to cgroup_add_cftypes() but the added files are only used for
4173 * the default hierarchy.
4174 */
4175int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4176{
4177 struct cftype *cft;
4178
4179 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4180 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4181 return cgroup_add_cftypes(ss, cfts);
4182}
4183
4184/**
4185 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4186 * @ss: target cgroup subsystem
4187 * @cfts: zero-length name terminated array of cftypes
4188 *
4189 * Similar to cgroup_add_cftypes() but the added files are only used for
4190 * the legacy hierarchies.
4191 */
4192int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4193{
4194 struct cftype *cft;
4195
4196 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4197 cft->flags |= __CFTYPE_NOT_ON_DFL;
4198 return cgroup_add_cftypes(ss, cfts);
4199}
4200
4201/**
4202 * cgroup_file_notify - generate a file modified event for a cgroup_file
4203 * @cfile: target cgroup_file
4204 *
4205 * @cfile must have been obtained by setting cftype->file_offset.
4206 */
4207void cgroup_file_notify(struct cgroup_file *cfile)
4208{
4209 unsigned long flags;
4210
4211 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4212 if (cfile->kn) {
4213 unsigned long last = cfile->notified_at;
4214 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4215
4216 if (time_in_range(jiffies, last, next)) {
4217 timer_reduce(&cfile->notify_timer, next);
4218 } else {
4219 kernfs_notify(cfile->kn);
4220 cfile->notified_at = jiffies;
4221 }
4222 }
4223 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4224}
4225
4226/**
4227 * css_next_child - find the next child of a given css
4228 * @pos: the current position (%NULL to initiate traversal)
4229 * @parent: css whose children to walk
4230 *
4231 * This function returns the next child of @parent and should be called
4232 * under either cgroup_mutex or RCU read lock. The only requirement is
4233 * that @parent and @pos are accessible. The next sibling is guaranteed to
4234 * be returned regardless of their states.
4235 *
4236 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4237 * css which finished ->css_online() is guaranteed to be visible in the
4238 * future iterations and will stay visible until the last reference is put.
4239 * A css which hasn't finished ->css_online() or already finished
4240 * ->css_offline() may show up during traversal. It's each subsystem's
4241 * responsibility to synchronize against on/offlining.
4242 */
4243struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4244 struct cgroup_subsys_state *parent)
4245{
4246 struct cgroup_subsys_state *next;
4247
4248 cgroup_assert_mutex_or_rcu_locked();
4249
4250 /*
4251 * @pos could already have been unlinked from the sibling list.
4252 * Once a cgroup is removed, its ->sibling.next is no longer
4253 * updated when its next sibling changes. CSS_RELEASED is set when
4254 * @pos is taken off list, at which time its next pointer is valid,
4255 * and, as releases are serialized, the one pointed to by the next
4256 * pointer is guaranteed to not have started release yet. This
4257 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4258 * critical section, the one pointed to by its next pointer is
4259 * guaranteed to not have finished its RCU grace period even if we
4260 * have dropped rcu_read_lock() in-between iterations.
4261 *
4262 * If @pos has CSS_RELEASED set, its next pointer can't be
4263 * dereferenced; however, as each css is given a monotonically
4264 * increasing unique serial number and always appended to the
4265 * sibling list, the next one can be found by walking the parent's
4266 * children until the first css with higher serial number than
4267 * @pos's. While this path can be slower, it happens iff iteration
4268 * races against release and the race window is very small.
4269 */
4270 if (!pos) {
4271 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4272 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4273 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4274 } else {
4275 list_for_each_entry_rcu(next, &parent->children, sibling,
4276 lockdep_is_held(&cgroup_mutex))
4277 if (next->serial_nr > pos->serial_nr)
4278 break;
4279 }
4280
4281 /*
4282 * @next, if not pointing to the head, can be dereferenced and is
4283 * the next sibling.
4284 */
4285 if (&next->sibling != &parent->children)
4286 return next;
4287 return NULL;
4288}
4289
4290/**
4291 * css_next_descendant_pre - find the next descendant for pre-order walk
4292 * @pos: the current position (%NULL to initiate traversal)
4293 * @root: css whose descendants to walk
4294 *
4295 * To be used by css_for_each_descendant_pre(). Find the next descendant
4296 * to visit for pre-order traversal of @root's descendants. @root is
4297 * included in the iteration and the first node to be visited.
4298 *
4299 * While this function requires cgroup_mutex or RCU read locking, it
4300 * doesn't require the whole traversal to be contained in a single critical
4301 * section. This function will return the correct next descendant as long
4302 * as both @pos and @root are accessible and @pos is a descendant of @root.
4303 *
4304 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4305 * css which finished ->css_online() is guaranteed to be visible in the
4306 * future iterations and will stay visible until the last reference is put.
4307 * A css which hasn't finished ->css_online() or already finished
4308 * ->css_offline() may show up during traversal. It's each subsystem's
4309 * responsibility to synchronize against on/offlining.
4310 */
4311struct cgroup_subsys_state *
4312css_next_descendant_pre(struct cgroup_subsys_state *pos,
4313 struct cgroup_subsys_state *root)
4314{
4315 struct cgroup_subsys_state *next;
4316
4317 cgroup_assert_mutex_or_rcu_locked();
4318
4319 /* if first iteration, visit @root */
4320 if (!pos)
4321 return root;
4322
4323 /* visit the first child if exists */
4324 next = css_next_child(NULL, pos);
4325 if (next)
4326 return next;
4327
4328 /* no child, visit my or the closest ancestor's next sibling */
4329 while (pos != root) {
4330 next = css_next_child(pos, pos->parent);
4331 if (next)
4332 return next;
4333 pos = pos->parent;
4334 }
4335
4336 return NULL;
4337}
4338EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4339
4340/**
4341 * css_rightmost_descendant - return the rightmost descendant of a css
4342 * @pos: css of interest
4343 *
4344 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4345 * is returned. This can be used during pre-order traversal to skip
4346 * subtree of @pos.
4347 *
4348 * While this function requires cgroup_mutex or RCU read locking, it
4349 * doesn't require the whole traversal to be contained in a single critical
4350 * section. This function will return the correct rightmost descendant as
4351 * long as @pos is accessible.
4352 */
4353struct cgroup_subsys_state *
4354css_rightmost_descendant(struct cgroup_subsys_state *pos)
4355{
4356 struct cgroup_subsys_state *last, *tmp;
4357
4358 cgroup_assert_mutex_or_rcu_locked();
4359
4360 do {
4361 last = pos;
4362 /* ->prev isn't RCU safe, walk ->next till the end */
4363 pos = NULL;
4364 css_for_each_child(tmp, last)
4365 pos = tmp;
4366 } while (pos);
4367
4368 return last;
4369}
4370
4371static struct cgroup_subsys_state *
4372css_leftmost_descendant(struct cgroup_subsys_state *pos)
4373{
4374 struct cgroup_subsys_state *last;
4375
4376 do {
4377 last = pos;
4378 pos = css_next_child(NULL, pos);
4379 } while (pos);
4380
4381 return last;
4382}
4383
4384/**
4385 * css_next_descendant_post - find the next descendant for post-order walk
4386 * @pos: the current position (%NULL to initiate traversal)
4387 * @root: css whose descendants to walk
4388 *
4389 * To be used by css_for_each_descendant_post(). Find the next descendant
4390 * to visit for post-order traversal of @root's descendants. @root is
4391 * included in the iteration and the last node to be visited.
4392 *
4393 * While this function requires cgroup_mutex or RCU read locking, it
4394 * doesn't require the whole traversal to be contained in a single critical
4395 * section. This function will return the correct next descendant as long
4396 * as both @pos and @cgroup are accessible and @pos is a descendant of
4397 * @cgroup.
4398 *
4399 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4400 * css which finished ->css_online() is guaranteed to be visible in the
4401 * future iterations and will stay visible until the last reference is put.
4402 * A css which hasn't finished ->css_online() or already finished
4403 * ->css_offline() may show up during traversal. It's each subsystem's
4404 * responsibility to synchronize against on/offlining.
4405 */
4406struct cgroup_subsys_state *
4407css_next_descendant_post(struct cgroup_subsys_state *pos,
4408 struct cgroup_subsys_state *root)
4409{
4410 struct cgroup_subsys_state *next;
4411
4412 cgroup_assert_mutex_or_rcu_locked();
4413
4414 /* if first iteration, visit leftmost descendant which may be @root */
4415 if (!pos)
4416 return css_leftmost_descendant(root);
4417
4418 /* if we visited @root, we're done */
4419 if (pos == root)
4420 return NULL;
4421
4422 /* if there's an unvisited sibling, visit its leftmost descendant */
4423 next = css_next_child(pos, pos->parent);
4424 if (next)
4425 return css_leftmost_descendant(next);
4426
4427 /* no sibling left, visit parent */
4428 return pos->parent;
4429}
4430
4431/**
4432 * css_has_online_children - does a css have online children
4433 * @css: the target css
4434 *
4435 * Returns %true if @css has any online children; otherwise, %false. This
4436 * function can be called from any context but the caller is responsible
4437 * for synchronizing against on/offlining as necessary.
4438 */
4439bool css_has_online_children(struct cgroup_subsys_state *css)
4440{
4441 struct cgroup_subsys_state *child;
4442 bool ret = false;
4443
4444 rcu_read_lock();
4445 css_for_each_child(child, css) {
4446 if (child->flags & CSS_ONLINE) {
4447 ret = true;
4448 break;
4449 }
4450 }
4451 rcu_read_unlock();
4452 return ret;
4453}
4454
4455static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4456{
4457 struct list_head *l;
4458 struct cgrp_cset_link *link;
4459 struct css_set *cset;
4460
4461 lockdep_assert_held(&css_set_lock);
4462
4463 /* find the next threaded cset */
4464 if (it->tcset_pos) {
4465 l = it->tcset_pos->next;
4466
4467 if (l != it->tcset_head) {
4468 it->tcset_pos = l;
4469 return container_of(l, struct css_set,
4470 threaded_csets_node);
4471 }
4472
4473 it->tcset_pos = NULL;
4474 }
4475
4476 /* find the next cset */
4477 l = it->cset_pos;
4478 l = l->next;
4479 if (l == it->cset_head) {
4480 it->cset_pos = NULL;
4481 return NULL;
4482 }
4483
4484 if (it->ss) {
4485 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4486 } else {
4487 link = list_entry(l, struct cgrp_cset_link, cset_link);
4488 cset = link->cset;
4489 }
4490
4491 it->cset_pos = l;
4492
4493 /* initialize threaded css_set walking */
4494 if (it->flags & CSS_TASK_ITER_THREADED) {
4495 if (it->cur_dcset)
4496 put_css_set_locked(it->cur_dcset);
4497 it->cur_dcset = cset;
4498 get_css_set(cset);
4499
4500 it->tcset_head = &cset->threaded_csets;
4501 it->tcset_pos = &cset->threaded_csets;
4502 }
4503
4504 return cset;
4505}
4506
4507/**
4508 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4509 * @it: the iterator to advance
4510 *
4511 * Advance @it to the next css_set to walk.
4512 */
4513static void css_task_iter_advance_css_set(struct css_task_iter *it)
4514{
4515 struct css_set *cset;
4516
4517 lockdep_assert_held(&css_set_lock);
4518
4519 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4520 while ((cset = css_task_iter_next_css_set(it))) {
4521 if (!list_empty(&cset->tasks)) {
4522 it->cur_tasks_head = &cset->tasks;
4523 break;
4524 } else if (!list_empty(&cset->mg_tasks)) {
4525 it->cur_tasks_head = &cset->mg_tasks;
4526 break;
4527 } else if (!list_empty(&cset->dying_tasks)) {
4528 it->cur_tasks_head = &cset->dying_tasks;
4529 break;
4530 }
4531 }
4532 if (!cset) {
4533 it->task_pos = NULL;
4534 return;
4535 }
4536 it->task_pos = it->cur_tasks_head->next;
4537
4538 /*
4539 * We don't keep css_sets locked across iteration steps and thus
4540 * need to take steps to ensure that iteration can be resumed after
4541 * the lock is re-acquired. Iteration is performed at two levels -
4542 * css_sets and tasks in them.
4543 *
4544 * Once created, a css_set never leaves its cgroup lists, so a
4545 * pinned css_set is guaranteed to stay put and we can resume
4546 * iteration afterwards.
4547 *
4548 * Tasks may leave @cset across iteration steps. This is resolved
4549 * by registering each iterator with the css_set currently being
4550 * walked and making css_set_move_task() advance iterators whose
4551 * next task is leaving.
4552 */
4553 if (it->cur_cset) {
4554 list_del(&it->iters_node);
4555 put_css_set_locked(it->cur_cset);
4556 }
4557 get_css_set(cset);
4558 it->cur_cset = cset;
4559 list_add(&it->iters_node, &cset->task_iters);
4560}
4561
4562static void css_task_iter_skip(struct css_task_iter *it,
4563 struct task_struct *task)
4564{
4565 lockdep_assert_held(&css_set_lock);
4566
4567 if (it->task_pos == &task->cg_list) {
4568 it->task_pos = it->task_pos->next;
4569 it->flags |= CSS_TASK_ITER_SKIPPED;
4570 }
4571}
4572
4573static void css_task_iter_advance(struct css_task_iter *it)
4574{
4575 struct task_struct *task;
4576
4577 lockdep_assert_held(&css_set_lock);
4578repeat:
4579 if (it->task_pos) {
4580 /*
4581 * Advance iterator to find next entry. We go through cset
4582 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4583 * the next cset.
4584 */
4585 if (it->flags & CSS_TASK_ITER_SKIPPED)
4586 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4587 else
4588 it->task_pos = it->task_pos->next;
4589
4590 if (it->task_pos == &it->cur_cset->tasks) {
4591 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4592 it->task_pos = it->cur_tasks_head->next;
4593 }
4594 if (it->task_pos == &it->cur_cset->mg_tasks) {
4595 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4596 it->task_pos = it->cur_tasks_head->next;
4597 }
4598 if (it->task_pos == &it->cur_cset->dying_tasks)
4599 css_task_iter_advance_css_set(it);
4600 } else {
4601 /* called from start, proceed to the first cset */
4602 css_task_iter_advance_css_set(it);
4603 }
4604
4605 if (!it->task_pos)
4606 return;
4607
4608 task = list_entry(it->task_pos, struct task_struct, cg_list);
4609
4610 if (it->flags & CSS_TASK_ITER_PROCS) {
4611 /* if PROCS, skip over tasks which aren't group leaders */
4612 if (!thread_group_leader(task))
4613 goto repeat;
4614
4615 /* and dying leaders w/o live member threads */
4616 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4617 !atomic_read(&task->signal->live))
4618 goto repeat;
4619 } else {
4620 /* skip all dying ones */
4621 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4622 goto repeat;
4623 }
4624}
4625
4626/**
4627 * css_task_iter_start - initiate task iteration
4628 * @css: the css to walk tasks of
4629 * @flags: CSS_TASK_ITER_* flags
4630 * @it: the task iterator to use
4631 *
4632 * Initiate iteration through the tasks of @css. The caller can call
4633 * css_task_iter_next() to walk through the tasks until the function
4634 * returns NULL. On completion of iteration, css_task_iter_end() must be
4635 * called.
4636 */
4637void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4638 struct css_task_iter *it)
4639{
4640 memset(it, 0, sizeof(*it));
4641
4642 spin_lock_irq(&css_set_lock);
4643
4644 it->ss = css->ss;
4645 it->flags = flags;
4646
4647 if (it->ss)
4648 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4649 else
4650 it->cset_pos = &css->cgroup->cset_links;
4651
4652 it->cset_head = it->cset_pos;
4653
4654 css_task_iter_advance(it);
4655
4656 spin_unlock_irq(&css_set_lock);
4657}
4658
4659/**
4660 * css_task_iter_next - return the next task for the iterator
4661 * @it: the task iterator being iterated
4662 *
4663 * The "next" function for task iteration. @it should have been
4664 * initialized via css_task_iter_start(). Returns NULL when the iteration
4665 * reaches the end.
4666 */
4667struct task_struct *css_task_iter_next(struct css_task_iter *it)
4668{
4669 if (it->cur_task) {
4670 put_task_struct(it->cur_task);
4671 it->cur_task = NULL;
4672 }
4673
4674 spin_lock_irq(&css_set_lock);
4675
4676 /* @it may be half-advanced by skips, finish advancing */
4677 if (it->flags & CSS_TASK_ITER_SKIPPED)
4678 css_task_iter_advance(it);
4679
4680 if (it->task_pos) {
4681 it->cur_task = list_entry(it->task_pos, struct task_struct,
4682 cg_list);
4683 get_task_struct(it->cur_task);
4684 css_task_iter_advance(it);
4685 }
4686
4687 spin_unlock_irq(&css_set_lock);
4688
4689 return it->cur_task;
4690}
4691
4692/**
4693 * css_task_iter_end - finish task iteration
4694 * @it: the task iterator to finish
4695 *
4696 * Finish task iteration started by css_task_iter_start().
4697 */
4698void css_task_iter_end(struct css_task_iter *it)
4699{
4700 if (it->cur_cset) {
4701 spin_lock_irq(&css_set_lock);
4702 list_del(&it->iters_node);
4703 put_css_set_locked(it->cur_cset);
4704 spin_unlock_irq(&css_set_lock);
4705 }
4706
4707 if (it->cur_dcset)
4708 put_css_set(it->cur_dcset);
4709
4710 if (it->cur_task)
4711 put_task_struct(it->cur_task);
4712}
4713
4714static void cgroup_procs_release(struct kernfs_open_file *of)
4715{
4716 if (of->priv) {
4717 css_task_iter_end(of->priv);
4718 kfree(of->priv);
4719 }
4720}
4721
4722static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4723{
4724 struct kernfs_open_file *of = s->private;
4725 struct css_task_iter *it = of->priv;
4726
4727 if (pos)
4728 (*pos)++;
4729
4730 return css_task_iter_next(it);
4731}
4732
4733static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4734 unsigned int iter_flags)
4735{
4736 struct kernfs_open_file *of = s->private;
4737 struct cgroup *cgrp = seq_css(s)->cgroup;
4738 struct css_task_iter *it = of->priv;
4739
4740 /*
4741 * When a seq_file is seeked, it's always traversed sequentially
4742 * from position 0, so we can simply keep iterating on !0 *pos.
4743 */
4744 if (!it) {
4745 if (WARN_ON_ONCE((*pos)))
4746 return ERR_PTR(-EINVAL);
4747
4748 it = kzalloc(sizeof(*it), GFP_KERNEL);
4749 if (!it)
4750 return ERR_PTR(-ENOMEM);
4751 of->priv = it;
4752 css_task_iter_start(&cgrp->self, iter_flags, it);
4753 } else if (!(*pos)) {
4754 css_task_iter_end(it);
4755 css_task_iter_start(&cgrp->self, iter_flags, it);
4756 } else
4757 return it->cur_task;
4758
4759 return cgroup_procs_next(s, NULL, NULL);
4760}
4761
4762static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4763{
4764 struct cgroup *cgrp = seq_css(s)->cgroup;
4765
4766 /*
4767 * All processes of a threaded subtree belong to the domain cgroup
4768 * of the subtree. Only threads can be distributed across the
4769 * subtree. Reject reads on cgroup.procs in the subtree proper.
4770 * They're always empty anyway.
4771 */
4772 if (cgroup_is_threaded(cgrp))
4773 return ERR_PTR(-EOPNOTSUPP);
4774
4775 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4776 CSS_TASK_ITER_THREADED);
4777}
4778
4779static int cgroup_procs_show(struct seq_file *s, void *v)
4780{
4781 seq_printf(s, "%d\n", task_pid_vnr(v));
4782 return 0;
4783}
4784
4785static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
4786{
4787 int ret;
4788 struct inode *inode;
4789
4790 lockdep_assert_held(&cgroup_mutex);
4791
4792 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
4793 if (!inode)
4794 return -ENOMEM;
4795
4796 ret = inode_permission(&init_user_ns, inode, MAY_WRITE);
4797 iput(inode);
4798 return ret;
4799}
4800
4801static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4802 struct cgroup *dst_cgrp,
4803 struct super_block *sb)
4804{
4805 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4806 struct cgroup *com_cgrp = src_cgrp;
4807 int ret;
4808
4809 lockdep_assert_held(&cgroup_mutex);
4810
4811 /* find the common ancestor */
4812 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4813 com_cgrp = cgroup_parent(com_cgrp);
4814
4815 /* %current should be authorized to migrate to the common ancestor */
4816 ret = cgroup_may_write(com_cgrp, sb);
4817 if (ret)
4818 return ret;
4819
4820 /*
4821 * If namespaces are delegation boundaries, %current must be able
4822 * to see both source and destination cgroups from its namespace.
4823 */
4824 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4825 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4826 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4827 return -ENOENT;
4828
4829 return 0;
4830}
4831
4832static int cgroup_attach_permissions(struct cgroup *src_cgrp,
4833 struct cgroup *dst_cgrp,
4834 struct super_block *sb, bool threadgroup)
4835{
4836 int ret = 0;
4837
4838 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb);
4839 if (ret)
4840 return ret;
4841
4842 ret = cgroup_migrate_vet_dst(dst_cgrp);
4843 if (ret)
4844 return ret;
4845
4846 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
4847 ret = -EOPNOTSUPP;
4848
4849 return ret;
4850}
4851
4852static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
4853 bool threadgroup)
4854{
4855 struct cgroup *src_cgrp, *dst_cgrp;
4856 struct task_struct *task;
4857 ssize_t ret;
4858 bool locked;
4859
4860 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4861 if (!dst_cgrp)
4862 return -ENODEV;
4863
4864 task = cgroup_procs_write_start(buf, threadgroup, &locked);
4865 ret = PTR_ERR_OR_ZERO(task);
4866 if (ret)
4867 goto out_unlock;
4868
4869 /* find the source cgroup */
4870 spin_lock_irq(&css_set_lock);
4871 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4872 spin_unlock_irq(&css_set_lock);
4873
4874 /* process and thread migrations follow same delegation rule */
4875 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
4876 of->file->f_path.dentry->d_sb, threadgroup);
4877 if (ret)
4878 goto out_finish;
4879
4880 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
4881
4882out_finish:
4883 cgroup_procs_write_finish(task, locked);
4884out_unlock:
4885 cgroup_kn_unlock(of->kn);
4886
4887 return ret;
4888}
4889
4890static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4891 char *buf, size_t nbytes, loff_t off)
4892{
4893 return __cgroup_procs_write(of, buf, true) ?: nbytes;
4894}
4895
4896static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4897{
4898 return __cgroup_procs_start(s, pos, 0);
4899}
4900
4901static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4902 char *buf, size_t nbytes, loff_t off)
4903{
4904 return __cgroup_procs_write(of, buf, false) ?: nbytes;
4905}
4906
4907/* cgroup core interface files for the default hierarchy */
4908static struct cftype cgroup_base_files[] = {
4909 {
4910 .name = "cgroup.type",
4911 .flags = CFTYPE_NOT_ON_ROOT,
4912 .seq_show = cgroup_type_show,
4913 .write = cgroup_type_write,
4914 },
4915 {
4916 .name = "cgroup.procs",
4917 .flags = CFTYPE_NS_DELEGATABLE,
4918 .file_offset = offsetof(struct cgroup, procs_file),
4919 .release = cgroup_procs_release,
4920 .seq_start = cgroup_procs_start,
4921 .seq_next = cgroup_procs_next,
4922 .seq_show = cgroup_procs_show,
4923 .write = cgroup_procs_write,
4924 },
4925 {
4926 .name = "cgroup.threads",
4927 .flags = CFTYPE_NS_DELEGATABLE,
4928 .release = cgroup_procs_release,
4929 .seq_start = cgroup_threads_start,
4930 .seq_next = cgroup_procs_next,
4931 .seq_show = cgroup_procs_show,
4932 .write = cgroup_threads_write,
4933 },
4934 {
4935 .name = "cgroup.controllers",
4936 .seq_show = cgroup_controllers_show,
4937 },
4938 {
4939 .name = "cgroup.subtree_control",
4940 .flags = CFTYPE_NS_DELEGATABLE,
4941 .seq_show = cgroup_subtree_control_show,
4942 .write = cgroup_subtree_control_write,
4943 },
4944 {
4945 .name = "cgroup.events",
4946 .flags = CFTYPE_NOT_ON_ROOT,
4947 .file_offset = offsetof(struct cgroup, events_file),
4948 .seq_show = cgroup_events_show,
4949 },
4950 {
4951 .name = "cgroup.max.descendants",
4952 .seq_show = cgroup_max_descendants_show,
4953 .write = cgroup_max_descendants_write,
4954 },
4955 {
4956 .name = "cgroup.max.depth",
4957 .seq_show = cgroup_max_depth_show,
4958 .write = cgroup_max_depth_write,
4959 },
4960 {
4961 .name = "cgroup.stat",
4962 .seq_show = cgroup_stat_show,
4963 },
4964 {
4965 .name = "cgroup.freeze",
4966 .flags = CFTYPE_NOT_ON_ROOT,
4967 .seq_show = cgroup_freeze_show,
4968 .write = cgroup_freeze_write,
4969 },
4970 {
4971 .name = "cgroup.kill",
4972 .flags = CFTYPE_NOT_ON_ROOT,
4973 .write = cgroup_kill_write,
4974 },
4975 {
4976 .name = "cpu.stat",
4977 .seq_show = cpu_stat_show,
4978 },
4979#ifdef CONFIG_PSI
4980 {
4981 .name = "io.pressure",
4982 .flags = CFTYPE_PRESSURE,
4983 .seq_show = cgroup_io_pressure_show,
4984 .write = cgroup_io_pressure_write,
4985 .poll = cgroup_pressure_poll,
4986 .release = cgroup_pressure_release,
4987 },
4988 {
4989 .name = "memory.pressure",
4990 .flags = CFTYPE_PRESSURE,
4991 .seq_show = cgroup_memory_pressure_show,
4992 .write = cgroup_memory_pressure_write,
4993 .poll = cgroup_pressure_poll,
4994 .release = cgroup_pressure_release,
4995 },
4996 {
4997 .name = "cpu.pressure",
4998 .flags = CFTYPE_PRESSURE,
4999 .seq_show = cgroup_cpu_pressure_show,
5000 .write = cgroup_cpu_pressure_write,
5001 .poll = cgroup_pressure_poll,
5002 .release = cgroup_pressure_release,
5003 },
5004#endif /* CONFIG_PSI */
5005 { } /* terminate */
5006};
5007
5008/*
5009 * css destruction is four-stage process.
5010 *
5011 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5012 * Implemented in kill_css().
5013 *
5014 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5015 * and thus css_tryget_online() is guaranteed to fail, the css can be
5016 * offlined by invoking offline_css(). After offlining, the base ref is
5017 * put. Implemented in css_killed_work_fn().
5018 *
5019 * 3. When the percpu_ref reaches zero, the only possible remaining
5020 * accessors are inside RCU read sections. css_release() schedules the
5021 * RCU callback.
5022 *
5023 * 4. After the grace period, the css can be freed. Implemented in
5024 * css_free_work_fn().
5025 *
5026 * It is actually hairier because both step 2 and 4 require process context
5027 * and thus involve punting to css->destroy_work adding two additional
5028 * steps to the already complex sequence.
5029 */
5030static void css_free_rwork_fn(struct work_struct *work)
5031{
5032 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5033 struct cgroup_subsys_state, destroy_rwork);
5034 struct cgroup_subsys *ss = css->ss;
5035 struct cgroup *cgrp = css->cgroup;
5036
5037 percpu_ref_exit(&css->refcnt);
5038
5039 if (ss) {
5040 /* css free path */
5041 struct cgroup_subsys_state *parent = css->parent;
5042 int id = css->id;
5043
5044 ss->css_free(css);
5045 cgroup_idr_remove(&ss->css_idr, id);
5046 cgroup_put(cgrp);
5047
5048 if (parent)
5049 css_put(parent);
5050 } else {
5051 /* cgroup free path */
5052 atomic_dec(&cgrp->root->nr_cgrps);
5053 cgroup1_pidlist_destroy_all(cgrp);
5054 cancel_work_sync(&cgrp->release_agent_work);
5055
5056 if (cgroup_parent(cgrp)) {
5057 /*
5058 * We get a ref to the parent, and put the ref when
5059 * this cgroup is being freed, so it's guaranteed
5060 * that the parent won't be destroyed before its
5061 * children.
5062 */
5063 cgroup_put(cgroup_parent(cgrp));
5064 kernfs_put(cgrp->kn);
5065 psi_cgroup_free(cgrp);
5066 cgroup_rstat_exit(cgrp);
5067 kfree(cgrp);
5068 } else {
5069 /*
5070 * This is root cgroup's refcnt reaching zero,
5071 * which indicates that the root should be
5072 * released.
5073 */
5074 cgroup_destroy_root(cgrp->root);
5075 }
5076 }
5077}
5078
5079static void css_release_work_fn(struct work_struct *work)
5080{
5081 struct cgroup_subsys_state *css =
5082 container_of(work, struct cgroup_subsys_state, destroy_work);
5083 struct cgroup_subsys *ss = css->ss;
5084 struct cgroup *cgrp = css->cgroup;
5085
5086 mutex_lock(&cgroup_mutex);
5087
5088 css->flags |= CSS_RELEASED;
5089 list_del_rcu(&css->sibling);
5090
5091 if (ss) {
5092 /* css release path */
5093 if (!list_empty(&css->rstat_css_node)) {
5094 cgroup_rstat_flush(cgrp);
5095 list_del_rcu(&css->rstat_css_node);
5096 }
5097
5098 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5099 if (ss->css_released)
5100 ss->css_released(css);
5101 } else {
5102 struct cgroup *tcgrp;
5103
5104 /* cgroup release path */
5105 TRACE_CGROUP_PATH(release, cgrp);
5106
5107 cgroup_rstat_flush(cgrp);
5108
5109 spin_lock_irq(&css_set_lock);
5110 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5111 tcgrp = cgroup_parent(tcgrp))
5112 tcgrp->nr_dying_descendants--;
5113 spin_unlock_irq(&css_set_lock);
5114
5115 /*
5116 * There are two control paths which try to determine
5117 * cgroup from dentry without going through kernfs -
5118 * cgroupstats_build() and css_tryget_online_from_dir().
5119 * Those are supported by RCU protecting clearing of
5120 * cgrp->kn->priv backpointer.
5121 */
5122 if (cgrp->kn)
5123 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5124 NULL);
5125 }
5126
5127 mutex_unlock(&cgroup_mutex);
5128
5129 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5130 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5131}
5132
5133static void css_release(struct percpu_ref *ref)
5134{
5135 struct cgroup_subsys_state *css =
5136 container_of(ref, struct cgroup_subsys_state, refcnt);
5137
5138 INIT_WORK(&css->destroy_work, css_release_work_fn);
5139 queue_work(cgroup_destroy_wq, &css->destroy_work);
5140}
5141
5142static void init_and_link_css(struct cgroup_subsys_state *css,
5143 struct cgroup_subsys *ss, struct cgroup *cgrp)
5144{
5145 lockdep_assert_held(&cgroup_mutex);
5146
5147 cgroup_get_live(cgrp);
5148
5149 memset(css, 0, sizeof(*css));
5150 css->cgroup = cgrp;
5151 css->ss = ss;
5152 css->id = -1;
5153 INIT_LIST_HEAD(&css->sibling);
5154 INIT_LIST_HEAD(&css->children);
5155 INIT_LIST_HEAD(&css->rstat_css_node);
5156 css->serial_nr = css_serial_nr_next++;
5157 atomic_set(&css->online_cnt, 0);
5158
5159 if (cgroup_parent(cgrp)) {
5160 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5161 css_get(css->parent);
5162 }
5163
5164 if (ss->css_rstat_flush)
5165 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5166
5167 BUG_ON(cgroup_css(cgrp, ss));
5168}
5169
5170/* invoke ->css_online() on a new CSS and mark it online if successful */
5171static int online_css(struct cgroup_subsys_state *css)
5172{
5173 struct cgroup_subsys *ss = css->ss;
5174 int ret = 0;
5175
5176 lockdep_assert_held(&cgroup_mutex);
5177
5178 if (ss->css_online)
5179 ret = ss->css_online(css);
5180 if (!ret) {
5181 css->flags |= CSS_ONLINE;
5182 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5183
5184 atomic_inc(&css->online_cnt);
5185 if (css->parent)
5186 atomic_inc(&css->parent->online_cnt);
5187 }
5188 return ret;
5189}
5190
5191/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5192static void offline_css(struct cgroup_subsys_state *css)
5193{
5194 struct cgroup_subsys *ss = css->ss;
5195
5196 lockdep_assert_held(&cgroup_mutex);
5197
5198 if (!(css->flags & CSS_ONLINE))
5199 return;
5200
5201 if (ss->css_offline)
5202 ss->css_offline(css);
5203
5204 css->flags &= ~CSS_ONLINE;
5205 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5206
5207 wake_up_all(&css->cgroup->offline_waitq);
5208}
5209
5210/**
5211 * css_create - create a cgroup_subsys_state
5212 * @cgrp: the cgroup new css will be associated with
5213 * @ss: the subsys of new css
5214 *
5215 * Create a new css associated with @cgrp - @ss pair. On success, the new
5216 * css is online and installed in @cgrp. This function doesn't create the
5217 * interface files. Returns 0 on success, -errno on failure.
5218 */
5219static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5220 struct cgroup_subsys *ss)
5221{
5222 struct cgroup *parent = cgroup_parent(cgrp);
5223 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5224 struct cgroup_subsys_state *css;
5225 int err;
5226
5227 lockdep_assert_held(&cgroup_mutex);
5228
5229 css = ss->css_alloc(parent_css);
5230 if (!css)
5231 css = ERR_PTR(-ENOMEM);
5232 if (IS_ERR(css))
5233 return css;
5234
5235 init_and_link_css(css, ss, cgrp);
5236
5237 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5238 if (err)
5239 goto err_free_css;
5240
5241 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5242 if (err < 0)
5243 goto err_free_css;
5244 css->id = err;
5245
5246 /* @css is ready to be brought online now, make it visible */
5247 list_add_tail_rcu(&css->sibling, &parent_css->children);
5248 cgroup_idr_replace(&ss->css_idr, css, css->id);
5249
5250 err = online_css(css);
5251 if (err)
5252 goto err_list_del;
5253
5254 return css;
5255
5256err_list_del:
5257 list_del_rcu(&css->sibling);
5258err_free_css:
5259 list_del_rcu(&css->rstat_css_node);
5260 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5261 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5262 return ERR_PTR(err);
5263}
5264
5265/*
5266 * The returned cgroup is fully initialized including its control mask, but
5267 * it isn't associated with its kernfs_node and doesn't have the control
5268 * mask applied.
5269 */
5270static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5271 umode_t mode)
5272{
5273 struct cgroup_root *root = parent->root;
5274 struct cgroup *cgrp, *tcgrp;
5275 struct kernfs_node *kn;
5276 int level = parent->level + 1;
5277 int ret;
5278
5279 /* allocate the cgroup and its ID, 0 is reserved for the root */
5280 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5281 GFP_KERNEL);
5282 if (!cgrp)
5283 return ERR_PTR(-ENOMEM);
5284
5285 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5286 if (ret)
5287 goto out_free_cgrp;
5288
5289 ret = cgroup_rstat_init(cgrp);
5290 if (ret)
5291 goto out_cancel_ref;
5292
5293 /* create the directory */
5294 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5295 if (IS_ERR(kn)) {
5296 ret = PTR_ERR(kn);
5297 goto out_stat_exit;
5298 }
5299 cgrp->kn = kn;
5300
5301 init_cgroup_housekeeping(cgrp);
5302
5303 cgrp->self.parent = &parent->self;
5304 cgrp->root = root;
5305 cgrp->level = level;
5306
5307 ret = psi_cgroup_alloc(cgrp);
5308 if (ret)
5309 goto out_kernfs_remove;
5310
5311 ret = cgroup_bpf_inherit(cgrp);
5312 if (ret)
5313 goto out_psi_free;
5314
5315 /*
5316 * New cgroup inherits effective freeze counter, and
5317 * if the parent has to be frozen, the child has too.
5318 */
5319 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5320 if (cgrp->freezer.e_freeze) {
5321 /*
5322 * Set the CGRP_FREEZE flag, so when a process will be
5323 * attached to the child cgroup, it will become frozen.
5324 * At this point the new cgroup is unpopulated, so we can
5325 * consider it frozen immediately.
5326 */
5327 set_bit(CGRP_FREEZE, &cgrp->flags);
5328 set_bit(CGRP_FROZEN, &cgrp->flags);
5329 }
5330
5331 spin_lock_irq(&css_set_lock);
5332 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5333 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp);
5334
5335 if (tcgrp != cgrp) {
5336 tcgrp->nr_descendants++;
5337
5338 /*
5339 * If the new cgroup is frozen, all ancestor cgroups
5340 * get a new frozen descendant, but their state can't
5341 * change because of this.
5342 */
5343 if (cgrp->freezer.e_freeze)
5344 tcgrp->freezer.nr_frozen_descendants++;
5345 }
5346 }
5347 spin_unlock_irq(&css_set_lock);
5348
5349 if (notify_on_release(parent))
5350 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5351
5352 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5353 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5354
5355 cgrp->self.serial_nr = css_serial_nr_next++;
5356
5357 /* allocation complete, commit to creation */
5358 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5359 atomic_inc(&root->nr_cgrps);
5360 cgroup_get_live(parent);
5361
5362 /*
5363 * On the default hierarchy, a child doesn't automatically inherit
5364 * subtree_control from the parent. Each is configured manually.
5365 */
5366 if (!cgroup_on_dfl(cgrp))
5367 cgrp->subtree_control = cgroup_control(cgrp);
5368
5369 cgroup_propagate_control(cgrp);
5370
5371 return cgrp;
5372
5373out_psi_free:
5374 psi_cgroup_free(cgrp);
5375out_kernfs_remove:
5376 kernfs_remove(cgrp->kn);
5377out_stat_exit:
5378 cgroup_rstat_exit(cgrp);
5379out_cancel_ref:
5380 percpu_ref_exit(&cgrp->self.refcnt);
5381out_free_cgrp:
5382 kfree(cgrp);
5383 return ERR_PTR(ret);
5384}
5385
5386static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5387{
5388 struct cgroup *cgroup;
5389 int ret = false;
5390 int level = 1;
5391
5392 lockdep_assert_held(&cgroup_mutex);
5393
5394 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5395 if (cgroup->nr_descendants >= cgroup->max_descendants)
5396 goto fail;
5397
5398 if (level > cgroup->max_depth)
5399 goto fail;
5400
5401 level++;
5402 }
5403
5404 ret = true;
5405fail:
5406 return ret;
5407}
5408
5409int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5410{
5411 struct cgroup *parent, *cgrp;
5412 int ret;
5413
5414 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5415 if (strchr(name, '\n'))
5416 return -EINVAL;
5417
5418 parent = cgroup_kn_lock_live(parent_kn, false);
5419 if (!parent)
5420 return -ENODEV;
5421
5422 if (!cgroup_check_hierarchy_limits(parent)) {
5423 ret = -EAGAIN;
5424 goto out_unlock;
5425 }
5426
5427 cgrp = cgroup_create(parent, name, mode);
5428 if (IS_ERR(cgrp)) {
5429 ret = PTR_ERR(cgrp);
5430 goto out_unlock;
5431 }
5432
5433 /*
5434 * This extra ref will be put in cgroup_free_fn() and guarantees
5435 * that @cgrp->kn is always accessible.
5436 */
5437 kernfs_get(cgrp->kn);
5438
5439 ret = cgroup_kn_set_ugid(cgrp->kn);
5440 if (ret)
5441 goto out_destroy;
5442
5443 ret = css_populate_dir(&cgrp->self);
5444 if (ret)
5445 goto out_destroy;
5446
5447 ret = cgroup_apply_control_enable(cgrp);
5448 if (ret)
5449 goto out_destroy;
5450
5451 TRACE_CGROUP_PATH(mkdir, cgrp);
5452
5453 /* let's create and online css's */
5454 kernfs_activate(cgrp->kn);
5455
5456 ret = 0;
5457 goto out_unlock;
5458
5459out_destroy:
5460 cgroup_destroy_locked(cgrp);
5461out_unlock:
5462 cgroup_kn_unlock(parent_kn);
5463 return ret;
5464}
5465
5466/*
5467 * This is called when the refcnt of a css is confirmed to be killed.
5468 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5469 * initiate destruction and put the css ref from kill_css().
5470 */
5471static void css_killed_work_fn(struct work_struct *work)
5472{
5473 struct cgroup_subsys_state *css =
5474 container_of(work, struct cgroup_subsys_state, destroy_work);
5475
5476 mutex_lock(&cgroup_mutex);
5477
5478 do {
5479 offline_css(css);
5480 css_put(css);
5481 /* @css can't go away while we're holding cgroup_mutex */
5482 css = css->parent;
5483 } while (css && atomic_dec_and_test(&css->online_cnt));
5484
5485 mutex_unlock(&cgroup_mutex);
5486}
5487
5488/* css kill confirmation processing requires process context, bounce */
5489static void css_killed_ref_fn(struct percpu_ref *ref)
5490{
5491 struct cgroup_subsys_state *css =
5492 container_of(ref, struct cgroup_subsys_state, refcnt);
5493
5494 if (atomic_dec_and_test(&css->online_cnt)) {
5495 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5496 queue_work(cgroup_destroy_wq, &css->destroy_work);
5497 }
5498}
5499
5500/**
5501 * kill_css - destroy a css
5502 * @css: css to destroy
5503 *
5504 * This function initiates destruction of @css by removing cgroup interface
5505 * files and putting its base reference. ->css_offline() will be invoked
5506 * asynchronously once css_tryget_online() is guaranteed to fail and when
5507 * the reference count reaches zero, @css will be released.
5508 */
5509static void kill_css(struct cgroup_subsys_state *css)
5510{
5511 lockdep_assert_held(&cgroup_mutex);
5512
5513 if (css->flags & CSS_DYING)
5514 return;
5515
5516 css->flags |= CSS_DYING;
5517
5518 /*
5519 * This must happen before css is disassociated with its cgroup.
5520 * See seq_css() for details.
5521 */
5522 css_clear_dir(css);
5523
5524 /*
5525 * Killing would put the base ref, but we need to keep it alive
5526 * until after ->css_offline().
5527 */
5528 css_get(css);
5529
5530 /*
5531 * cgroup core guarantees that, by the time ->css_offline() is
5532 * invoked, no new css reference will be given out via
5533 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5534 * proceed to offlining css's because percpu_ref_kill() doesn't
5535 * guarantee that the ref is seen as killed on all CPUs on return.
5536 *
5537 * Use percpu_ref_kill_and_confirm() to get notifications as each
5538 * css is confirmed to be seen as killed on all CPUs.
5539 */
5540 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5541}
5542
5543/**
5544 * cgroup_destroy_locked - the first stage of cgroup destruction
5545 * @cgrp: cgroup to be destroyed
5546 *
5547 * css's make use of percpu refcnts whose killing latency shouldn't be
5548 * exposed to userland and are RCU protected. Also, cgroup core needs to
5549 * guarantee that css_tryget_online() won't succeed by the time
5550 * ->css_offline() is invoked. To satisfy all the requirements,
5551 * destruction is implemented in the following two steps.
5552 *
5553 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5554 * userland visible parts and start killing the percpu refcnts of
5555 * css's. Set up so that the next stage will be kicked off once all
5556 * the percpu refcnts are confirmed to be killed.
5557 *
5558 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5559 * rest of destruction. Once all cgroup references are gone, the
5560 * cgroup is RCU-freed.
5561 *
5562 * This function implements s1. After this step, @cgrp is gone as far as
5563 * the userland is concerned and a new cgroup with the same name may be
5564 * created. As cgroup doesn't care about the names internally, this
5565 * doesn't cause any problem.
5566 */
5567static int cgroup_destroy_locked(struct cgroup *cgrp)
5568 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5569{
5570 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5571 struct cgroup_subsys_state *css;
5572 struct cgrp_cset_link *link;
5573 int ssid;
5574
5575 lockdep_assert_held(&cgroup_mutex);
5576
5577 /*
5578 * Only migration can raise populated from zero and we're already
5579 * holding cgroup_mutex.
5580 */
5581 if (cgroup_is_populated(cgrp))
5582 return -EBUSY;
5583
5584 /*
5585 * Make sure there's no live children. We can't test emptiness of
5586 * ->self.children as dead children linger on it while being
5587 * drained; otherwise, "rmdir parent/child parent" may fail.
5588 */
5589 if (css_has_online_children(&cgrp->self))
5590 return -EBUSY;
5591
5592 /*
5593 * Mark @cgrp and the associated csets dead. The former prevents
5594 * further task migration and child creation by disabling
5595 * cgroup_lock_live_group(). The latter makes the csets ignored by
5596 * the migration path.
5597 */
5598 cgrp->self.flags &= ~CSS_ONLINE;
5599
5600 spin_lock_irq(&css_set_lock);
5601 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5602 link->cset->dead = true;
5603 spin_unlock_irq(&css_set_lock);
5604
5605 /* initiate massacre of all css's */
5606 for_each_css(css, ssid, cgrp)
5607 kill_css(css);
5608
5609 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5610 css_clear_dir(&cgrp->self);
5611 kernfs_remove(cgrp->kn);
5612
5613 if (parent && cgroup_is_threaded(cgrp))
5614 parent->nr_threaded_children--;
5615
5616 spin_lock_irq(&css_set_lock);
5617 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5618 tcgrp->nr_descendants--;
5619 tcgrp->nr_dying_descendants++;
5620 /*
5621 * If the dying cgroup is frozen, decrease frozen descendants
5622 * counters of ancestor cgroups.
5623 */
5624 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5625 tcgrp->freezer.nr_frozen_descendants--;
5626 }
5627 spin_unlock_irq(&css_set_lock);
5628
5629 cgroup1_check_for_release(parent);
5630
5631 cgroup_bpf_offline(cgrp);
5632
5633 /* put the base reference */
5634 percpu_ref_kill(&cgrp->self.refcnt);
5635
5636 return 0;
5637};
5638
5639int cgroup_rmdir(struct kernfs_node *kn)
5640{
5641 struct cgroup *cgrp;
5642 int ret = 0;
5643
5644 cgrp = cgroup_kn_lock_live(kn, false);
5645 if (!cgrp)
5646 return 0;
5647
5648 ret = cgroup_destroy_locked(cgrp);
5649 if (!ret)
5650 TRACE_CGROUP_PATH(rmdir, cgrp);
5651
5652 cgroup_kn_unlock(kn);
5653 return ret;
5654}
5655
5656static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5657 .show_options = cgroup_show_options,
5658 .mkdir = cgroup_mkdir,
5659 .rmdir = cgroup_rmdir,
5660 .show_path = cgroup_show_path,
5661};
5662
5663static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5664{
5665 struct cgroup_subsys_state *css;
5666
5667 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5668
5669 mutex_lock(&cgroup_mutex);
5670
5671 idr_init(&ss->css_idr);
5672 INIT_LIST_HEAD(&ss->cfts);
5673
5674 /* Create the root cgroup state for this subsystem */
5675 ss->root = &cgrp_dfl_root;
5676 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5677 /* We don't handle early failures gracefully */
5678 BUG_ON(IS_ERR(css));
5679 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5680
5681 /*
5682 * Root csses are never destroyed and we can't initialize
5683 * percpu_ref during early init. Disable refcnting.
5684 */
5685 css->flags |= CSS_NO_REF;
5686
5687 if (early) {
5688 /* allocation can't be done safely during early init */
5689 css->id = 1;
5690 } else {
5691 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5692 BUG_ON(css->id < 0);
5693 }
5694
5695 /* Update the init_css_set to contain a subsys
5696 * pointer to this state - since the subsystem is
5697 * newly registered, all tasks and hence the
5698 * init_css_set is in the subsystem's root cgroup. */
5699 init_css_set.subsys[ss->id] = css;
5700
5701 have_fork_callback |= (bool)ss->fork << ss->id;
5702 have_exit_callback |= (bool)ss->exit << ss->id;
5703 have_release_callback |= (bool)ss->release << ss->id;
5704 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5705
5706 /* At system boot, before all subsystems have been
5707 * registered, no tasks have been forked, so we don't
5708 * need to invoke fork callbacks here. */
5709 BUG_ON(!list_empty(&init_task.tasks));
5710
5711 BUG_ON(online_css(css));
5712
5713 mutex_unlock(&cgroup_mutex);
5714}
5715
5716/**
5717 * cgroup_init_early - cgroup initialization at system boot
5718 *
5719 * Initialize cgroups at system boot, and initialize any
5720 * subsystems that request early init.
5721 */
5722int __init cgroup_init_early(void)
5723{
5724 static struct cgroup_fs_context __initdata ctx;
5725 struct cgroup_subsys *ss;
5726 int i;
5727
5728 ctx.root = &cgrp_dfl_root;
5729 init_cgroup_root(&ctx);
5730 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5731
5732 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5733
5734 for_each_subsys(ss, i) {
5735 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5736 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5737 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5738 ss->id, ss->name);
5739 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5740 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5741
5742 ss->id = i;
5743 ss->name = cgroup_subsys_name[i];
5744 if (!ss->legacy_name)
5745 ss->legacy_name = cgroup_subsys_name[i];
5746
5747 if (ss->early_init)
5748 cgroup_init_subsys(ss, true);
5749 }
5750 return 0;
5751}
5752
5753/**
5754 * cgroup_init - cgroup initialization
5755 *
5756 * Register cgroup filesystem and /proc file, and initialize
5757 * any subsystems that didn't request early init.
5758 */
5759int __init cgroup_init(void)
5760{
5761 struct cgroup_subsys *ss;
5762 int ssid;
5763
5764 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5765 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5766 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5767
5768 cgroup_rstat_boot();
5769
5770 /*
5771 * The latency of the synchronize_rcu() is too high for cgroups,
5772 * avoid it at the cost of forcing all readers into the slow path.
5773 */
5774 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5775
5776 get_user_ns(init_cgroup_ns.user_ns);
5777
5778 mutex_lock(&cgroup_mutex);
5779
5780 /*
5781 * Add init_css_set to the hash table so that dfl_root can link to
5782 * it during init.
5783 */
5784 hash_add(css_set_table, &init_css_set.hlist,
5785 css_set_hash(init_css_set.subsys));
5786
5787 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5788
5789 mutex_unlock(&cgroup_mutex);
5790
5791 for_each_subsys(ss, ssid) {
5792 if (ss->early_init) {
5793 struct cgroup_subsys_state *css =
5794 init_css_set.subsys[ss->id];
5795
5796 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5797 GFP_KERNEL);
5798 BUG_ON(css->id < 0);
5799 } else {
5800 cgroup_init_subsys(ss, false);
5801 }
5802
5803 list_add_tail(&init_css_set.e_cset_node[ssid],
5804 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5805
5806 /*
5807 * Setting dfl_root subsys_mask needs to consider the
5808 * disabled flag and cftype registration needs kmalloc,
5809 * both of which aren't available during early_init.
5810 */
5811 if (!cgroup_ssid_enabled(ssid))
5812 continue;
5813
5814 if (cgroup1_ssid_disabled(ssid))
5815 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5816 ss->name);
5817
5818 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5819
5820 /* implicit controllers must be threaded too */
5821 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5822
5823 if (ss->implicit_on_dfl)
5824 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5825 else if (!ss->dfl_cftypes)
5826 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5827
5828 if (ss->threaded)
5829 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5830
5831 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5832 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5833 } else {
5834 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5835 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5836 }
5837
5838 if (ss->bind)
5839 ss->bind(init_css_set.subsys[ssid]);
5840
5841 mutex_lock(&cgroup_mutex);
5842 css_populate_dir(init_css_set.subsys[ssid]);
5843 mutex_unlock(&cgroup_mutex);
5844 }
5845
5846 /* init_css_set.subsys[] has been updated, re-hash */
5847 hash_del(&init_css_set.hlist);
5848 hash_add(css_set_table, &init_css_set.hlist,
5849 css_set_hash(init_css_set.subsys));
5850
5851 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5852 WARN_ON(register_filesystem(&cgroup_fs_type));
5853 WARN_ON(register_filesystem(&cgroup2_fs_type));
5854 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5855#ifdef CONFIG_CPUSETS
5856 WARN_ON(register_filesystem(&cpuset_fs_type));
5857#endif
5858
5859 return 0;
5860}
5861
5862static int __init cgroup_wq_init(void)
5863{
5864 /*
5865 * There isn't much point in executing destruction path in
5866 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5867 * Use 1 for @max_active.
5868 *
5869 * We would prefer to do this in cgroup_init() above, but that
5870 * is called before init_workqueues(): so leave this until after.
5871 */
5872 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5873 BUG_ON(!cgroup_destroy_wq);
5874 return 0;
5875}
5876core_initcall(cgroup_wq_init);
5877
5878void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
5879{
5880 struct kernfs_node *kn;
5881
5882 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
5883 if (!kn)
5884 return;
5885 kernfs_path(kn, buf, buflen);
5886 kernfs_put(kn);
5887}
5888
5889/*
5890 * cgroup_get_from_id : get the cgroup associated with cgroup id
5891 * @id: cgroup id
5892 * On success return the cgrp, on failure return NULL
5893 */
5894struct cgroup *cgroup_get_from_id(u64 id)
5895{
5896 struct kernfs_node *kn;
5897 struct cgroup *cgrp = NULL;
5898
5899 mutex_lock(&cgroup_mutex);
5900 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
5901 if (!kn)
5902 goto out_unlock;
5903
5904 cgrp = kn->priv;
5905 if (cgroup_is_dead(cgrp) || !cgroup_tryget(cgrp))
5906 cgrp = NULL;
5907 kernfs_put(kn);
5908out_unlock:
5909 mutex_unlock(&cgroup_mutex);
5910 return cgrp;
5911}
5912EXPORT_SYMBOL_GPL(cgroup_get_from_id);
5913
5914/*
5915 * proc_cgroup_show()
5916 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5917 * - Used for /proc/<pid>/cgroup.
5918 */
5919int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5920 struct pid *pid, struct task_struct *tsk)
5921{
5922 char *buf;
5923 int retval;
5924 struct cgroup_root *root;
5925
5926 retval = -ENOMEM;
5927 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5928 if (!buf)
5929 goto out;
5930
5931 mutex_lock(&cgroup_mutex);
5932 spin_lock_irq(&css_set_lock);
5933
5934 for_each_root(root) {
5935 struct cgroup_subsys *ss;
5936 struct cgroup *cgrp;
5937 int ssid, count = 0;
5938
5939 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5940 continue;
5941
5942 seq_printf(m, "%d:", root->hierarchy_id);
5943 if (root != &cgrp_dfl_root)
5944 for_each_subsys(ss, ssid)
5945 if (root->subsys_mask & (1 << ssid))
5946 seq_printf(m, "%s%s", count++ ? "," : "",
5947 ss->legacy_name);
5948 if (strlen(root->name))
5949 seq_printf(m, "%sname=%s", count ? "," : "",
5950 root->name);
5951 seq_putc(m, ':');
5952
5953 cgrp = task_cgroup_from_root(tsk, root);
5954
5955 /*
5956 * On traditional hierarchies, all zombie tasks show up as
5957 * belonging to the root cgroup. On the default hierarchy,
5958 * while a zombie doesn't show up in "cgroup.procs" and
5959 * thus can't be migrated, its /proc/PID/cgroup keeps
5960 * reporting the cgroup it belonged to before exiting. If
5961 * the cgroup is removed before the zombie is reaped,
5962 * " (deleted)" is appended to the cgroup path.
5963 */
5964 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5965 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5966 current->nsproxy->cgroup_ns);
5967 if (retval >= PATH_MAX)
5968 retval = -ENAMETOOLONG;
5969 if (retval < 0)
5970 goto out_unlock;
5971
5972 seq_puts(m, buf);
5973 } else {
5974 seq_puts(m, "/");
5975 }
5976
5977 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5978 seq_puts(m, " (deleted)\n");
5979 else
5980 seq_putc(m, '\n');
5981 }
5982
5983 retval = 0;
5984out_unlock:
5985 spin_unlock_irq(&css_set_lock);
5986 mutex_unlock(&cgroup_mutex);
5987 kfree(buf);
5988out:
5989 return retval;
5990}
5991
5992/**
5993 * cgroup_fork - initialize cgroup related fields during copy_process()
5994 * @child: pointer to task_struct of forking parent process.
5995 *
5996 * A task is associated with the init_css_set until cgroup_post_fork()
5997 * attaches it to the target css_set.
5998 */
5999void cgroup_fork(struct task_struct *child)
6000{
6001 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6002 INIT_LIST_HEAD(&child->cg_list);
6003}
6004
6005static struct cgroup *cgroup_get_from_file(struct file *f)
6006{
6007 struct cgroup_subsys_state *css;
6008 struct cgroup *cgrp;
6009
6010 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6011 if (IS_ERR(css))
6012 return ERR_CAST(css);
6013
6014 cgrp = css->cgroup;
6015 if (!cgroup_on_dfl(cgrp)) {
6016 cgroup_put(cgrp);
6017 return ERR_PTR(-EBADF);
6018 }
6019
6020 return cgrp;
6021}
6022
6023/**
6024 * cgroup_css_set_fork - find or create a css_set for a child process
6025 * @kargs: the arguments passed to create the child process
6026 *
6027 * This functions finds or creates a new css_set which the child
6028 * process will be attached to in cgroup_post_fork(). By default,
6029 * the child process will be given the same css_set as its parent.
6030 *
6031 * If CLONE_INTO_CGROUP is specified this function will try to find an
6032 * existing css_set which includes the requested cgroup and if not create
6033 * a new css_set that the child will be attached to later. If this function
6034 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6035 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6036 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6037 * to the target cgroup.
6038 */
6039static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6040 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6041{
6042 int ret;
6043 struct cgroup *dst_cgrp = NULL;
6044 struct css_set *cset;
6045 struct super_block *sb;
6046 struct file *f;
6047
6048 if (kargs->flags & CLONE_INTO_CGROUP)
6049 mutex_lock(&cgroup_mutex);
6050
6051 cgroup_threadgroup_change_begin(current);
6052
6053 spin_lock_irq(&css_set_lock);
6054 cset = task_css_set(current);
6055 get_css_set(cset);
6056 spin_unlock_irq(&css_set_lock);
6057
6058 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6059 kargs->cset = cset;
6060 return 0;
6061 }
6062
6063 f = fget_raw(kargs->cgroup);
6064 if (!f) {
6065 ret = -EBADF;
6066 goto err;
6067 }
6068 sb = f->f_path.dentry->d_sb;
6069
6070 dst_cgrp = cgroup_get_from_file(f);
6071 if (IS_ERR(dst_cgrp)) {
6072 ret = PTR_ERR(dst_cgrp);
6073 dst_cgrp = NULL;
6074 goto err;
6075 }
6076
6077 if (cgroup_is_dead(dst_cgrp)) {
6078 ret = -ENODEV;
6079 goto err;
6080 }
6081
6082 /*
6083 * Verify that we the target cgroup is writable for us. This is
6084 * usually done by the vfs layer but since we're not going through
6085 * the vfs layer here we need to do it "manually".
6086 */
6087 ret = cgroup_may_write(dst_cgrp, sb);
6088 if (ret)
6089 goto err;
6090
6091 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6092 !(kargs->flags & CLONE_THREAD));
6093 if (ret)
6094 goto err;
6095
6096 kargs->cset = find_css_set(cset, dst_cgrp);
6097 if (!kargs->cset) {
6098 ret = -ENOMEM;
6099 goto err;
6100 }
6101
6102 put_css_set(cset);
6103 fput(f);
6104 kargs->cgrp = dst_cgrp;
6105 return ret;
6106
6107err:
6108 cgroup_threadgroup_change_end(current);
6109 mutex_unlock(&cgroup_mutex);
6110 if (f)
6111 fput(f);
6112 if (dst_cgrp)
6113 cgroup_put(dst_cgrp);
6114 put_css_set(cset);
6115 if (kargs->cset)
6116 put_css_set(kargs->cset);
6117 return ret;
6118}
6119
6120/**
6121 * cgroup_css_set_put_fork - drop references we took during fork
6122 * @kargs: the arguments passed to create the child process
6123 *
6124 * Drop references to the prepared css_set and target cgroup if
6125 * CLONE_INTO_CGROUP was requested.
6126 */
6127static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6128 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6129{
6130 cgroup_threadgroup_change_end(current);
6131
6132 if (kargs->flags & CLONE_INTO_CGROUP) {
6133 struct cgroup *cgrp = kargs->cgrp;
6134 struct css_set *cset = kargs->cset;
6135
6136 mutex_unlock(&cgroup_mutex);
6137
6138 if (cset) {
6139 put_css_set(cset);
6140 kargs->cset = NULL;
6141 }
6142
6143 if (cgrp) {
6144 cgroup_put(cgrp);
6145 kargs->cgrp = NULL;
6146 }
6147 }
6148}
6149
6150/**
6151 * cgroup_can_fork - called on a new task before the process is exposed
6152 * @child: the child process
6153 *
6154 * This prepares a new css_set for the child process which the child will
6155 * be attached to in cgroup_post_fork().
6156 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6157 * callback returns an error, the fork aborts with that error code. This
6158 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6159 */
6160int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6161{
6162 struct cgroup_subsys *ss;
6163 int i, j, ret;
6164
6165 ret = cgroup_css_set_fork(kargs);
6166 if (ret)
6167 return ret;
6168
6169 do_each_subsys_mask(ss, i, have_canfork_callback) {
6170 ret = ss->can_fork(child, kargs->cset);
6171 if (ret)
6172 goto out_revert;
6173 } while_each_subsys_mask();
6174
6175 return 0;
6176
6177out_revert:
6178 for_each_subsys(ss, j) {
6179 if (j >= i)
6180 break;
6181 if (ss->cancel_fork)
6182 ss->cancel_fork(child, kargs->cset);
6183 }
6184
6185 cgroup_css_set_put_fork(kargs);
6186
6187 return ret;
6188}
6189
6190/**
6191 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6192 * @child: the child process
6193 * @kargs: the arguments passed to create the child process
6194 *
6195 * This calls the cancel_fork() callbacks if a fork failed *after*
6196 * cgroup_can_fork() succeeded and cleans up references we took to
6197 * prepare a new css_set for the child process in cgroup_can_fork().
6198 */
6199void cgroup_cancel_fork(struct task_struct *child,
6200 struct kernel_clone_args *kargs)
6201{
6202 struct cgroup_subsys *ss;
6203 int i;
6204
6205 for_each_subsys(ss, i)
6206 if (ss->cancel_fork)
6207 ss->cancel_fork(child, kargs->cset);
6208
6209 cgroup_css_set_put_fork(kargs);
6210}
6211
6212/**
6213 * cgroup_post_fork - finalize cgroup setup for the child process
6214 * @child: the child process
6215 *
6216 * Attach the child process to its css_set calling the subsystem fork()
6217 * callbacks.
6218 */
6219void cgroup_post_fork(struct task_struct *child,
6220 struct kernel_clone_args *kargs)
6221 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6222{
6223 unsigned long cgrp_flags = 0;
6224 bool kill = false;
6225 struct cgroup_subsys *ss;
6226 struct css_set *cset;
6227 int i;
6228
6229 cset = kargs->cset;
6230 kargs->cset = NULL;
6231
6232 spin_lock_irq(&css_set_lock);
6233
6234 /* init tasks are special, only link regular threads */
6235 if (likely(child->pid)) {
6236 if (kargs->cgrp)
6237 cgrp_flags = kargs->cgrp->flags;
6238 else
6239 cgrp_flags = cset->dfl_cgrp->flags;
6240
6241 WARN_ON_ONCE(!list_empty(&child->cg_list));
6242 cset->nr_tasks++;
6243 css_set_move_task(child, NULL, cset, false);
6244 } else {
6245 put_css_set(cset);
6246 cset = NULL;
6247 }
6248
6249 if (!(child->flags & PF_KTHREAD)) {
6250 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6251 /*
6252 * If the cgroup has to be frozen, the new task has
6253 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6254 * get the task into the frozen state.
6255 */
6256 spin_lock(&child->sighand->siglock);
6257 WARN_ON_ONCE(child->frozen);
6258 child->jobctl |= JOBCTL_TRAP_FREEZE;
6259 spin_unlock(&child->sighand->siglock);
6260
6261 /*
6262 * Calling cgroup_update_frozen() isn't required here,
6263 * because it will be called anyway a bit later from
6264 * do_freezer_trap(). So we avoid cgroup's transient
6265 * switch from the frozen state and back.
6266 */
6267 }
6268
6269 /*
6270 * If the cgroup is to be killed notice it now and take the
6271 * child down right after we finished preparing it for
6272 * userspace.
6273 */
6274 kill = test_bit(CGRP_KILL, &cgrp_flags);
6275 }
6276
6277 spin_unlock_irq(&css_set_lock);
6278
6279 /*
6280 * Call ss->fork(). This must happen after @child is linked on
6281 * css_set; otherwise, @child might change state between ->fork()
6282 * and addition to css_set.
6283 */
6284 do_each_subsys_mask(ss, i, have_fork_callback) {
6285 ss->fork(child);
6286 } while_each_subsys_mask();
6287
6288 /* Make the new cset the root_cset of the new cgroup namespace. */
6289 if (kargs->flags & CLONE_NEWCGROUP) {
6290 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6291
6292 get_css_set(cset);
6293 child->nsproxy->cgroup_ns->root_cset = cset;
6294 put_css_set(rcset);
6295 }
6296
6297 /* Cgroup has to be killed so take down child immediately. */
6298 if (unlikely(kill))
6299 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6300
6301 cgroup_css_set_put_fork(kargs);
6302}
6303
6304/**
6305 * cgroup_exit - detach cgroup from exiting task
6306 * @tsk: pointer to task_struct of exiting process
6307 *
6308 * Description: Detach cgroup from @tsk.
6309 *
6310 */
6311void cgroup_exit(struct task_struct *tsk)
6312{
6313 struct cgroup_subsys *ss;
6314 struct css_set *cset;
6315 int i;
6316
6317 spin_lock_irq(&css_set_lock);
6318
6319 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6320 cset = task_css_set(tsk);
6321 css_set_move_task(tsk, cset, NULL, false);
6322 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6323 cset->nr_tasks--;
6324
6325 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6326 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6327 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6328 cgroup_update_frozen(task_dfl_cgroup(tsk));
6329
6330 spin_unlock_irq(&css_set_lock);
6331
6332 /* see cgroup_post_fork() for details */
6333 do_each_subsys_mask(ss, i, have_exit_callback) {
6334 ss->exit(tsk);
6335 } while_each_subsys_mask();
6336}
6337
6338void cgroup_release(struct task_struct *task)
6339{
6340 struct cgroup_subsys *ss;
6341 int ssid;
6342
6343 do_each_subsys_mask(ss, ssid, have_release_callback) {
6344 ss->release(task);
6345 } while_each_subsys_mask();
6346
6347 spin_lock_irq(&css_set_lock);
6348 css_set_skip_task_iters(task_css_set(task), task);
6349 list_del_init(&task->cg_list);
6350 spin_unlock_irq(&css_set_lock);
6351}
6352
6353void cgroup_free(struct task_struct *task)
6354{
6355 struct css_set *cset = task_css_set(task);
6356 put_css_set(cset);
6357}
6358
6359static int __init cgroup_disable(char *str)
6360{
6361 struct cgroup_subsys *ss;
6362 char *token;
6363 int i;
6364
6365 while ((token = strsep(&str, ",")) != NULL) {
6366 if (!*token)
6367 continue;
6368
6369 for_each_subsys(ss, i) {
6370 if (strcmp(token, ss->name) &&
6371 strcmp(token, ss->legacy_name))
6372 continue;
6373
6374 static_branch_disable(cgroup_subsys_enabled_key[i]);
6375 pr_info("Disabling %s control group subsystem\n",
6376 ss->name);
6377 }
6378
6379 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6380 if (strcmp(token, cgroup_opt_feature_names[i]))
6381 continue;
6382 cgroup_feature_disable_mask |= 1 << i;
6383 pr_info("Disabling %s control group feature\n",
6384 cgroup_opt_feature_names[i]);
6385 break;
6386 }
6387 }
6388 return 1;
6389}
6390__setup("cgroup_disable=", cgroup_disable);
6391
6392void __init __weak enable_debug_cgroup(void) { }
6393
6394static int __init enable_cgroup_debug(char *str)
6395{
6396 cgroup_debug = true;
6397 enable_debug_cgroup();
6398 return 1;
6399}
6400__setup("cgroup_debug", enable_cgroup_debug);
6401
6402/**
6403 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6404 * @dentry: directory dentry of interest
6405 * @ss: subsystem of interest
6406 *
6407 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6408 * to get the corresponding css and return it. If such css doesn't exist
6409 * or can't be pinned, an ERR_PTR value is returned.
6410 */
6411struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6412 struct cgroup_subsys *ss)
6413{
6414 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6415 struct file_system_type *s_type = dentry->d_sb->s_type;
6416 struct cgroup_subsys_state *css = NULL;
6417 struct cgroup *cgrp;
6418
6419 /* is @dentry a cgroup dir? */
6420 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6421 !kn || kernfs_type(kn) != KERNFS_DIR)
6422 return ERR_PTR(-EBADF);
6423
6424 rcu_read_lock();
6425
6426 /*
6427 * This path doesn't originate from kernfs and @kn could already
6428 * have been or be removed at any point. @kn->priv is RCU
6429 * protected for this access. See css_release_work_fn() for details.
6430 */
6431 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6432 if (cgrp)
6433 css = cgroup_css(cgrp, ss);
6434
6435 if (!css || !css_tryget_online(css))
6436 css = ERR_PTR(-ENOENT);
6437
6438 rcu_read_unlock();
6439 return css;
6440}
6441
6442/**
6443 * css_from_id - lookup css by id
6444 * @id: the cgroup id
6445 * @ss: cgroup subsys to be looked into
6446 *
6447 * Returns the css if there's valid one with @id, otherwise returns NULL.
6448 * Should be called under rcu_read_lock().
6449 */
6450struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6451{
6452 WARN_ON_ONCE(!rcu_read_lock_held());
6453 return idr_find(&ss->css_idr, id);
6454}
6455
6456/**
6457 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6458 * @path: path on the default hierarchy
6459 *
6460 * Find the cgroup at @path on the default hierarchy, increment its
6461 * reference count and return it. Returns pointer to the found cgroup on
6462 * success, ERR_PTR(-ENOENT) if @path doesn't exist and ERR_PTR(-ENOTDIR)
6463 * if @path points to a non-directory.
6464 */
6465struct cgroup *cgroup_get_from_path(const char *path)
6466{
6467 struct kernfs_node *kn;
6468 struct cgroup *cgrp;
6469
6470 mutex_lock(&cgroup_mutex);
6471
6472 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6473 if (kn) {
6474 if (kernfs_type(kn) == KERNFS_DIR) {
6475 cgrp = kn->priv;
6476 cgroup_get_live(cgrp);
6477 } else {
6478 cgrp = ERR_PTR(-ENOTDIR);
6479 }
6480 kernfs_put(kn);
6481 } else {
6482 cgrp = ERR_PTR(-ENOENT);
6483 }
6484
6485 mutex_unlock(&cgroup_mutex);
6486 return cgrp;
6487}
6488EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6489
6490/**
6491 * cgroup_get_from_fd - get a cgroup pointer from a fd
6492 * @fd: fd obtained by open(cgroup2_dir)
6493 *
6494 * Find the cgroup from a fd which should be obtained
6495 * by opening a cgroup directory. Returns a pointer to the
6496 * cgroup on success. ERR_PTR is returned if the cgroup
6497 * cannot be found.
6498 */
6499struct cgroup *cgroup_get_from_fd(int fd)
6500{
6501 struct cgroup *cgrp;
6502 struct file *f;
6503
6504 f = fget_raw(fd);
6505 if (!f)
6506 return ERR_PTR(-EBADF);
6507
6508 cgrp = cgroup_get_from_file(f);
6509 fput(f);
6510 return cgrp;
6511}
6512EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6513
6514static u64 power_of_ten(int power)
6515{
6516 u64 v = 1;
6517 while (power--)
6518 v *= 10;
6519 return v;
6520}
6521
6522/**
6523 * cgroup_parse_float - parse a floating number
6524 * @input: input string
6525 * @dec_shift: number of decimal digits to shift
6526 * @v: output
6527 *
6528 * Parse a decimal floating point number in @input and store the result in
6529 * @v with decimal point right shifted @dec_shift times. For example, if
6530 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6531 * Returns 0 on success, -errno otherwise.
6532 *
6533 * There's nothing cgroup specific about this function except that it's
6534 * currently the only user.
6535 */
6536int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6537{
6538 s64 whole, frac = 0;
6539 int fstart = 0, fend = 0, flen;
6540
6541 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6542 return -EINVAL;
6543 if (frac < 0)
6544 return -EINVAL;
6545
6546 flen = fend > fstart ? fend - fstart : 0;
6547 if (flen < dec_shift)
6548 frac *= power_of_ten(dec_shift - flen);
6549 else
6550 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6551
6552 *v = whole * power_of_ten(dec_shift) + frac;
6553 return 0;
6554}
6555
6556/*
6557 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6558 * definition in cgroup-defs.h.
6559 */
6560#ifdef CONFIG_SOCK_CGROUP_DATA
6561
6562#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6563
6564DEFINE_SPINLOCK(cgroup_sk_update_lock);
6565static bool cgroup_sk_alloc_disabled __read_mostly;
6566
6567void cgroup_sk_alloc_disable(void)
6568{
6569 if (cgroup_sk_alloc_disabled)
6570 return;
6571 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6572 cgroup_sk_alloc_disabled = true;
6573}
6574
6575#else
6576
6577#define cgroup_sk_alloc_disabled false
6578
6579#endif
6580
6581void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6582{
6583 if (cgroup_sk_alloc_disabled) {
6584 skcd->no_refcnt = 1;
6585 return;
6586 }
6587
6588 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6589 if (in_interrupt())
6590 return;
6591
6592 rcu_read_lock();
6593
6594 while (true) {
6595 struct css_set *cset;
6596
6597 cset = task_css_set(current);
6598 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6599 skcd->val = (unsigned long)cset->dfl_cgrp;
6600 cgroup_bpf_get(cset->dfl_cgrp);
6601 break;
6602 }
6603 cpu_relax();
6604 }
6605
6606 rcu_read_unlock();
6607}
6608
6609void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6610{
6611 if (skcd->val) {
6612 if (skcd->no_refcnt)
6613 return;
6614 /*
6615 * We might be cloning a socket which is left in an empty
6616 * cgroup and the cgroup might have already been rmdir'd.
6617 * Don't use cgroup_get_live().
6618 */
6619 cgroup_get(sock_cgroup_ptr(skcd));
6620 cgroup_bpf_get(sock_cgroup_ptr(skcd));
6621 }
6622}
6623
6624void cgroup_sk_free(struct sock_cgroup_data *skcd)
6625{
6626 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6627
6628 if (skcd->no_refcnt)
6629 return;
6630 cgroup_bpf_put(cgrp);
6631 cgroup_put(cgrp);
6632}
6633
6634#endif /* CONFIG_SOCK_CGROUP_DATA */
6635
6636#ifdef CONFIG_CGROUP_BPF
6637int cgroup_bpf_attach(struct cgroup *cgrp,
6638 struct bpf_prog *prog, struct bpf_prog *replace_prog,
6639 struct bpf_cgroup_link *link,
6640 enum bpf_attach_type type,
6641 u32 flags)
6642{
6643 int ret;
6644
6645 mutex_lock(&cgroup_mutex);
6646 ret = __cgroup_bpf_attach(cgrp, prog, replace_prog, link, type, flags);
6647 mutex_unlock(&cgroup_mutex);
6648 return ret;
6649}
6650
6651int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6652 enum bpf_attach_type type)
6653{
6654 int ret;
6655
6656 mutex_lock(&cgroup_mutex);
6657 ret = __cgroup_bpf_detach(cgrp, prog, NULL, type);
6658 mutex_unlock(&cgroup_mutex);
6659 return ret;
6660}
6661
6662int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6663 union bpf_attr __user *uattr)
6664{
6665 int ret;
6666
6667 mutex_lock(&cgroup_mutex);
6668 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6669 mutex_unlock(&cgroup_mutex);
6670 return ret;
6671}
6672#endif /* CONFIG_CGROUP_BPF */
6673
6674#ifdef CONFIG_SYSFS
6675static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6676 ssize_t size, const char *prefix)
6677{
6678 struct cftype *cft;
6679 ssize_t ret = 0;
6680
6681 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6682 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6683 continue;
6684
6685 if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
6686 continue;
6687
6688 if (prefix)
6689 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6690
6691 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6692
6693 if (WARN_ON(ret >= size))
6694 break;
6695 }
6696
6697 return ret;
6698}
6699
6700static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6701 char *buf)
6702{
6703 struct cgroup_subsys *ss;
6704 int ssid;
6705 ssize_t ret = 0;
6706
6707 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6708 NULL);
6709
6710 for_each_subsys(ss, ssid)
6711 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6712 PAGE_SIZE - ret,
6713 cgroup_subsys_name[ssid]);
6714
6715 return ret;
6716}
6717static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6718
6719static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6720 char *buf)
6721{
6722 return snprintf(buf, PAGE_SIZE,
6723 "nsdelegate\n"
6724 "memory_localevents\n"
6725 "memory_recursiveprot\n");
6726}
6727static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6728
6729static struct attribute *cgroup_sysfs_attrs[] = {
6730 &cgroup_delegate_attr.attr,
6731 &cgroup_features_attr.attr,
6732 NULL,
6733};
6734
6735static const struct attribute_group cgroup_sysfs_attr_group = {
6736 .attrs = cgroup_sysfs_attrs,
6737 .name = "cgroup",
6738};
6739
6740static int __init cgroup_sysfs_init(void)
6741{
6742 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6743}
6744subsys_initcall(cgroup_sysfs_init);
6745
6746#endif /* CONFIG_SYSFS */