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/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/*
157 * The default hierarchy, reserved for the subsystems that are otherwise
158 * unattached - it never has more than a single cgroup, and all tasks are
159 * part of that cgroup.
160 */
161struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
162EXPORT_SYMBOL_GPL(cgrp_dfl_root);
163
164/*
165 * The default hierarchy always exists but is hidden until mounted for the
166 * first time. This is for backward compatibility.
167 */
168static bool cgrp_dfl_visible;
169
170/* some controllers are not supported in the default hierarchy */
171static u16 cgrp_dfl_inhibit_ss_mask;
172
173/* some controllers are implicitly enabled on the default hierarchy */
174static u16 cgrp_dfl_implicit_ss_mask;
175
176/* some controllers can be threaded on the default hierarchy */
177static u16 cgrp_dfl_threaded_ss_mask;
178
179/* The list of hierarchy roots */
180LIST_HEAD(cgroup_roots);
181static int cgroup_root_count;
182
183/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
184static DEFINE_IDR(cgroup_hierarchy_idr);
185
186/*
187 * Assign a monotonically increasing serial number to csses. It guarantees
188 * cgroups with bigger numbers are newer than those with smaller numbers.
189 * Also, as csses are always appended to the parent's ->children list, it
190 * guarantees that sibling csses are always sorted in the ascending serial
191 * number order on the list. Protected by cgroup_mutex.
192 */
193static u64 css_serial_nr_next = 1;
194
195/*
196 * These bitmasks identify subsystems with specific features to avoid
197 * having to do iterative checks repeatedly.
198 */
199static u16 have_fork_callback __read_mostly;
200static u16 have_exit_callback __read_mostly;
201static u16 have_release_callback __read_mostly;
202static u16 have_canfork_callback __read_mostly;
203
204/* cgroup namespace for init task */
205struct cgroup_namespace init_cgroup_ns = {
206 .count = REFCOUNT_INIT(2),
207 .user_ns = &init_user_ns,
208 .ns.ops = &cgroupns_operations,
209 .ns.inum = PROC_CGROUP_INIT_INO,
210 .root_cset = &init_css_set,
211};
212
213static struct file_system_type cgroup2_fs_type;
214static struct cftype cgroup_base_files[];
215
216static int cgroup_apply_control(struct cgroup *cgrp);
217static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
218static void css_task_iter_skip(struct css_task_iter *it,
219 struct task_struct *task);
220static int cgroup_destroy_locked(struct cgroup *cgrp);
221static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
222 struct cgroup_subsys *ss);
223static void css_release(struct percpu_ref *ref);
224static void kill_css(struct cgroup_subsys_state *css);
225static int cgroup_addrm_files(struct cgroup_subsys_state *css,
226 struct cgroup *cgrp, struct cftype cfts[],
227 bool is_add);
228
229/**
230 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
231 * @ssid: subsys ID of interest
232 *
233 * cgroup_subsys_enabled() can only be used with literal subsys names which
234 * is fine for individual subsystems but unsuitable for cgroup core. This
235 * is slower static_key_enabled() based test indexed by @ssid.
236 */
237bool cgroup_ssid_enabled(int ssid)
238{
239 if (CGROUP_SUBSYS_COUNT == 0)
240 return false;
241
242 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
243}
244
245/**
246 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
247 * @cgrp: the cgroup of interest
248 *
249 * The default hierarchy is the v2 interface of cgroup and this function
250 * can be used to test whether a cgroup is on the default hierarchy for
251 * cases where a subsystem should behave differnetly depending on the
252 * interface version.
253 *
254 * The set of behaviors which change on the default hierarchy are still
255 * being determined and the mount option is prefixed with __DEVEL__.
256 *
257 * List of changed behaviors:
258 *
259 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
260 * and "name" are disallowed.
261 *
262 * - When mounting an existing superblock, mount options should match.
263 *
264 * - Remount is disallowed.
265 *
266 * - rename(2) is disallowed.
267 *
268 * - "tasks" is removed. Everything should be at process granularity. Use
269 * "cgroup.procs" instead.
270 *
271 * - "cgroup.procs" is not sorted. pids will be unique unless they got
272 * recycled inbetween reads.
273 *
274 * - "release_agent" and "notify_on_release" are removed. Replacement
275 * notification mechanism will be implemented.
276 *
277 * - "cgroup.clone_children" is removed.
278 *
279 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
280 * and its descendants contain no task; otherwise, 1. The file also
281 * generates kernfs notification which can be monitored through poll and
282 * [di]notify when the value of the file changes.
283 *
284 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
285 * take masks of ancestors with non-empty cpus/mems, instead of being
286 * moved to an ancestor.
287 *
288 * - cpuset: a task can be moved into an empty cpuset, and again it takes
289 * masks of ancestors.
290 *
291 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
292 * is not created.
293 *
294 * - blkcg: blk-throttle becomes properly hierarchical.
295 *
296 * - debug: disallowed on the default hierarchy.
297 */
298bool cgroup_on_dfl(const struct cgroup *cgrp)
299{
300 return cgrp->root == &cgrp_dfl_root;
301}
302
303/* IDR wrappers which synchronize using cgroup_idr_lock */
304static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
305 gfp_t gfp_mask)
306{
307 int ret;
308
309 idr_preload(gfp_mask);
310 spin_lock_bh(&cgroup_idr_lock);
311 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
312 spin_unlock_bh(&cgroup_idr_lock);
313 idr_preload_end();
314 return ret;
315}
316
317static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
318{
319 void *ret;
320
321 spin_lock_bh(&cgroup_idr_lock);
322 ret = idr_replace(idr, ptr, id);
323 spin_unlock_bh(&cgroup_idr_lock);
324 return ret;
325}
326
327static void cgroup_idr_remove(struct idr *idr, int id)
328{
329 spin_lock_bh(&cgroup_idr_lock);
330 idr_remove(idr, id);
331 spin_unlock_bh(&cgroup_idr_lock);
332}
333
334static bool cgroup_has_tasks(struct cgroup *cgrp)
335{
336 return cgrp->nr_populated_csets;
337}
338
339bool cgroup_is_threaded(struct cgroup *cgrp)
340{
341 return cgrp->dom_cgrp != cgrp;
342}
343
344/* can @cgrp host both domain and threaded children? */
345static bool cgroup_is_mixable(struct cgroup *cgrp)
346{
347 /*
348 * Root isn't under domain level resource control exempting it from
349 * the no-internal-process constraint, so it can serve as a thread
350 * root and a parent of resource domains at the same time.
351 */
352 return !cgroup_parent(cgrp);
353}
354
355/* can @cgrp become a thread root? should always be true for a thread root */
356static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
357{
358 /* mixables don't care */
359 if (cgroup_is_mixable(cgrp))
360 return true;
361
362 /* domain roots can't be nested under threaded */
363 if (cgroup_is_threaded(cgrp))
364 return false;
365
366 /* can only have either domain or threaded children */
367 if (cgrp->nr_populated_domain_children)
368 return false;
369
370 /* and no domain controllers can be enabled */
371 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
372 return false;
373
374 return true;
375}
376
377/* is @cgrp root of a threaded subtree? */
378bool cgroup_is_thread_root(struct cgroup *cgrp)
379{
380 /* thread root should be a domain */
381 if (cgroup_is_threaded(cgrp))
382 return false;
383
384 /* a domain w/ threaded children is a thread root */
385 if (cgrp->nr_threaded_children)
386 return true;
387
388 /*
389 * A domain which has tasks and explicit threaded controllers
390 * enabled is a thread root.
391 */
392 if (cgroup_has_tasks(cgrp) &&
393 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
394 return true;
395
396 return false;
397}
398
399/* a domain which isn't connected to the root w/o brekage can't be used */
400static bool cgroup_is_valid_domain(struct cgroup *cgrp)
401{
402 /* the cgroup itself can be a thread root */
403 if (cgroup_is_threaded(cgrp))
404 return false;
405
406 /* but the ancestors can't be unless mixable */
407 while ((cgrp = cgroup_parent(cgrp))) {
408 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
409 return false;
410 if (cgroup_is_threaded(cgrp))
411 return false;
412 }
413
414 return true;
415}
416
417/* subsystems visibly enabled on a cgroup */
418static u16 cgroup_control(struct cgroup *cgrp)
419{
420 struct cgroup *parent = cgroup_parent(cgrp);
421 u16 root_ss_mask = cgrp->root->subsys_mask;
422
423 if (parent) {
424 u16 ss_mask = parent->subtree_control;
425
426 /* threaded cgroups can only have threaded controllers */
427 if (cgroup_is_threaded(cgrp))
428 ss_mask &= cgrp_dfl_threaded_ss_mask;
429 return ss_mask;
430 }
431
432 if (cgroup_on_dfl(cgrp))
433 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
434 cgrp_dfl_implicit_ss_mask);
435 return root_ss_mask;
436}
437
438/* subsystems enabled on a cgroup */
439static u16 cgroup_ss_mask(struct cgroup *cgrp)
440{
441 struct cgroup *parent = cgroup_parent(cgrp);
442
443 if (parent) {
444 u16 ss_mask = parent->subtree_ss_mask;
445
446 /* threaded cgroups can only have threaded controllers */
447 if (cgroup_is_threaded(cgrp))
448 ss_mask &= cgrp_dfl_threaded_ss_mask;
449 return ss_mask;
450 }
451
452 return cgrp->root->subsys_mask;
453}
454
455/**
456 * cgroup_css - obtain a cgroup's css for the specified subsystem
457 * @cgrp: the cgroup of interest
458 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
459 *
460 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
461 * function must be called either under cgroup_mutex or rcu_read_lock() and
462 * the caller is responsible for pinning the returned css if it wants to
463 * keep accessing it outside the said locks. This function may return
464 * %NULL if @cgrp doesn't have @subsys_id enabled.
465 */
466static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
467 struct cgroup_subsys *ss)
468{
469 if (ss)
470 return rcu_dereference_check(cgrp->subsys[ss->id],
471 lockdep_is_held(&cgroup_mutex));
472 else
473 return &cgrp->self;
474}
475
476/**
477 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
478 * @cgrp: the cgroup of interest
479 * @ss: the subsystem of interest
480 *
481 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
482 * or is offline, %NULL is returned.
483 */
484static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
485 struct cgroup_subsys *ss)
486{
487 struct cgroup_subsys_state *css;
488
489 rcu_read_lock();
490 css = cgroup_css(cgrp, ss);
491 if (css && !css_tryget_online(css))
492 css = NULL;
493 rcu_read_unlock();
494
495 return css;
496}
497
498/**
499 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
500 * @cgrp: the cgroup of interest
501 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
502 *
503 * Similar to cgroup_css() but returns the effective css, which is defined
504 * as the matching css of the nearest ancestor including self which has @ss
505 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
506 * function is guaranteed to return non-NULL css.
507 */
508static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
509 struct cgroup_subsys *ss)
510{
511 lockdep_assert_held(&cgroup_mutex);
512
513 if (!ss)
514 return &cgrp->self;
515
516 /*
517 * This function is used while updating css associations and thus
518 * can't test the csses directly. Test ss_mask.
519 */
520 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
521 cgrp = cgroup_parent(cgrp);
522 if (!cgrp)
523 return NULL;
524 }
525
526 return cgroup_css(cgrp, ss);
527}
528
529/**
530 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
531 * @cgrp: the cgroup of interest
532 * @ss: the subsystem of interest
533 *
534 * Find and get the effective css of @cgrp for @ss. The effective css is
535 * defined as the matching css of the nearest ancestor including self which
536 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
537 * the root css is returned, so this function always returns a valid css.
538 *
539 * The returned css is not guaranteed to be online, and therefore it is the
540 * callers responsiblity to tryget a reference for it.
541 */
542struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
543 struct cgroup_subsys *ss)
544{
545 struct cgroup_subsys_state *css;
546
547 do {
548 css = cgroup_css(cgrp, ss);
549
550 if (css)
551 return css;
552 cgrp = cgroup_parent(cgrp);
553 } while (cgrp);
554
555 return init_css_set.subsys[ss->id];
556}
557
558/**
559 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
560 * @cgrp: the cgroup of interest
561 * @ss: the subsystem of interest
562 *
563 * Find and get the effective css of @cgrp for @ss. The effective css is
564 * defined as the matching css of the nearest ancestor including self which
565 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
566 * the root css is returned, so this function always returns a valid css.
567 * The returned css must be put using css_put().
568 */
569struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
570 struct cgroup_subsys *ss)
571{
572 struct cgroup_subsys_state *css;
573
574 rcu_read_lock();
575
576 do {
577 css = cgroup_css(cgrp, ss);
578
579 if (css && css_tryget_online(css))
580 goto out_unlock;
581 cgrp = cgroup_parent(cgrp);
582 } while (cgrp);
583
584 css = init_css_set.subsys[ss->id];
585 css_get(css);
586out_unlock:
587 rcu_read_unlock();
588 return css;
589}
590
591static void cgroup_get_live(struct cgroup *cgrp)
592{
593 WARN_ON_ONCE(cgroup_is_dead(cgrp));
594 css_get(&cgrp->self);
595}
596
597/**
598 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
599 * is responsible for taking the css_set_lock.
600 * @cgrp: the cgroup in question
601 */
602int __cgroup_task_count(const struct cgroup *cgrp)
603{
604 int count = 0;
605 struct cgrp_cset_link *link;
606
607 lockdep_assert_held(&css_set_lock);
608
609 list_for_each_entry(link, &cgrp->cset_links, cset_link)
610 count += link->cset->nr_tasks;
611
612 return count;
613}
614
615/**
616 * cgroup_task_count - count the number of tasks in a cgroup.
617 * @cgrp: the cgroup in question
618 */
619int cgroup_task_count(const struct cgroup *cgrp)
620{
621 int count;
622
623 spin_lock_irq(&css_set_lock);
624 count = __cgroup_task_count(cgrp);
625 spin_unlock_irq(&css_set_lock);
626
627 return count;
628}
629
630struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
631{
632 struct cgroup *cgrp = of->kn->parent->priv;
633 struct cftype *cft = of_cft(of);
634
635 /*
636 * This is open and unprotected implementation of cgroup_css().
637 * seq_css() is only called from a kernfs file operation which has
638 * an active reference on the file. Because all the subsystem
639 * files are drained before a css is disassociated with a cgroup,
640 * the matching css from the cgroup's subsys table is guaranteed to
641 * be and stay valid until the enclosing operation is complete.
642 */
643 if (cft->ss)
644 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
645 else
646 return &cgrp->self;
647}
648EXPORT_SYMBOL_GPL(of_css);
649
650/**
651 * for_each_css - iterate all css's of a cgroup
652 * @css: the iteration cursor
653 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
654 * @cgrp: the target cgroup to iterate css's of
655 *
656 * Should be called under cgroup_[tree_]mutex.
657 */
658#define for_each_css(css, ssid, cgrp) \
659 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
660 if (!((css) = rcu_dereference_check( \
661 (cgrp)->subsys[(ssid)], \
662 lockdep_is_held(&cgroup_mutex)))) { } \
663 else
664
665/**
666 * for_each_e_css - iterate all effective css's of a cgroup
667 * @css: the iteration cursor
668 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
669 * @cgrp: the target cgroup to iterate css's of
670 *
671 * Should be called under cgroup_[tree_]mutex.
672 */
673#define for_each_e_css(css, ssid, cgrp) \
674 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
675 if (!((css) = cgroup_e_css_by_mask(cgrp, \
676 cgroup_subsys[(ssid)]))) \
677 ; \
678 else
679
680/**
681 * do_each_subsys_mask - filter for_each_subsys with a bitmask
682 * @ss: the iteration cursor
683 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
684 * @ss_mask: the bitmask
685 *
686 * The block will only run for cases where the ssid-th bit (1 << ssid) of
687 * @ss_mask is set.
688 */
689#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
690 unsigned long __ss_mask = (ss_mask); \
691 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
692 (ssid) = 0; \
693 break; \
694 } \
695 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
696 (ss) = cgroup_subsys[ssid]; \
697 {
698
699#define while_each_subsys_mask() \
700 } \
701 } \
702} while (false)
703
704/* iterate over child cgrps, lock should be held throughout iteration */
705#define cgroup_for_each_live_child(child, cgrp) \
706 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
707 if (({ lockdep_assert_held(&cgroup_mutex); \
708 cgroup_is_dead(child); })) \
709 ; \
710 else
711
712/* walk live descendants in preorder */
713#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
714 css_for_each_descendant_pre((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/* walk live descendants in postorder */
722#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
723 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
724 if (({ lockdep_assert_held(&cgroup_mutex); \
725 (dsct) = (d_css)->cgroup; \
726 cgroup_is_dead(dsct); })) \
727 ; \
728 else
729
730/*
731 * The default css_set - used by init and its children prior to any
732 * hierarchies being mounted. It contains a pointer to the root state
733 * for each subsystem. Also used to anchor the list of css_sets. Not
734 * reference-counted, to improve performance when child cgroups
735 * haven't been created.
736 */
737struct css_set init_css_set = {
738 .refcount = REFCOUNT_INIT(1),
739 .dom_cset = &init_css_set,
740 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
741 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
742 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
743 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
744 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
745 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
746 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
747 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
748
749 /*
750 * The following field is re-initialized when this cset gets linked
751 * in cgroup_init(). However, let's initialize the field
752 * statically too so that the default cgroup can be accessed safely
753 * early during boot.
754 */
755 .dfl_cgrp = &cgrp_dfl_root.cgrp,
756};
757
758static int css_set_count = 1; /* 1 for init_css_set */
759
760static bool css_set_threaded(struct css_set *cset)
761{
762 return cset->dom_cset != cset;
763}
764
765/**
766 * css_set_populated - does a css_set contain any tasks?
767 * @cset: target css_set
768 *
769 * css_set_populated() should be the same as !!cset->nr_tasks at steady
770 * state. However, css_set_populated() can be called while a task is being
771 * added to or removed from the linked list before the nr_tasks is
772 * properly updated. Hence, we can't just look at ->nr_tasks here.
773 */
774static bool css_set_populated(struct css_set *cset)
775{
776 lockdep_assert_held(&css_set_lock);
777
778 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
779}
780
781/**
782 * cgroup_update_populated - update the populated count of a cgroup
783 * @cgrp: the target cgroup
784 * @populated: inc or dec populated count
785 *
786 * One of the css_sets associated with @cgrp is either getting its first
787 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
788 * count is propagated towards root so that a given cgroup's
789 * nr_populated_children is zero iff none of its descendants contain any
790 * tasks.
791 *
792 * @cgrp's interface file "cgroup.populated" is zero if both
793 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
794 * 1 otherwise. When the sum changes from or to zero, userland is notified
795 * that the content of the interface file has changed. This can be used to
796 * detect when @cgrp and its descendants become populated or empty.
797 */
798static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
799{
800 struct cgroup *child = NULL;
801 int adj = populated ? 1 : -1;
802
803 lockdep_assert_held(&css_set_lock);
804
805 do {
806 bool was_populated = cgroup_is_populated(cgrp);
807
808 if (!child) {
809 cgrp->nr_populated_csets += adj;
810 } else {
811 if (cgroup_is_threaded(child))
812 cgrp->nr_populated_threaded_children += adj;
813 else
814 cgrp->nr_populated_domain_children += adj;
815 }
816
817 if (was_populated == cgroup_is_populated(cgrp))
818 break;
819
820 cgroup1_check_for_release(cgrp);
821 TRACE_CGROUP_PATH(notify_populated, cgrp,
822 cgroup_is_populated(cgrp));
823 cgroup_file_notify(&cgrp->events_file);
824
825 child = cgrp;
826 cgrp = cgroup_parent(cgrp);
827 } while (cgrp);
828}
829
830/**
831 * css_set_update_populated - update populated state of a css_set
832 * @cset: target css_set
833 * @populated: whether @cset is populated or depopulated
834 *
835 * @cset is either getting the first task or losing the last. Update the
836 * populated counters of all associated cgroups accordingly.
837 */
838static void css_set_update_populated(struct css_set *cset, bool populated)
839{
840 struct cgrp_cset_link *link;
841
842 lockdep_assert_held(&css_set_lock);
843
844 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
845 cgroup_update_populated(link->cgrp, populated);
846}
847
848/*
849 * @task is leaving, advance task iterators which are pointing to it so
850 * that they can resume at the next position. Advancing an iterator might
851 * remove it from the list, use safe walk. See css_task_iter_skip() for
852 * details.
853 */
854static void css_set_skip_task_iters(struct css_set *cset,
855 struct task_struct *task)
856{
857 struct css_task_iter *it, *pos;
858
859 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
860 css_task_iter_skip(it, task);
861}
862
863/**
864 * css_set_move_task - move a task from one css_set to another
865 * @task: task being moved
866 * @from_cset: css_set @task currently belongs to (may be NULL)
867 * @to_cset: new css_set @task is being moved to (may be NULL)
868 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
869 *
870 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
871 * css_set, @from_cset can be NULL. If @task is being disassociated
872 * instead of moved, @to_cset can be NULL.
873 *
874 * This function automatically handles populated counter updates and
875 * css_task_iter adjustments but the caller is responsible for managing
876 * @from_cset and @to_cset's reference counts.
877 */
878static void css_set_move_task(struct task_struct *task,
879 struct css_set *from_cset, struct css_set *to_cset,
880 bool use_mg_tasks)
881{
882 lockdep_assert_held(&css_set_lock);
883
884 if (to_cset && !css_set_populated(to_cset))
885 css_set_update_populated(to_cset, true);
886
887 if (from_cset) {
888 WARN_ON_ONCE(list_empty(&task->cg_list));
889
890 css_set_skip_task_iters(from_cset, task);
891 list_del_init(&task->cg_list);
892 if (!css_set_populated(from_cset))
893 css_set_update_populated(from_cset, false);
894 } else {
895 WARN_ON_ONCE(!list_empty(&task->cg_list));
896 }
897
898 if (to_cset) {
899 /*
900 * We are synchronized through cgroup_threadgroup_rwsem
901 * against PF_EXITING setting such that we can't race
902 * against cgroup_exit() changing the css_set to
903 * init_css_set and dropping the old one.
904 */
905 WARN_ON_ONCE(task->flags & PF_EXITING);
906
907 cgroup_move_task(task, to_cset);
908 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
909 &to_cset->tasks);
910 }
911}
912
913/*
914 * hash table for cgroup groups. This improves the performance to find
915 * an existing css_set. This hash doesn't (currently) take into
916 * account cgroups in empty hierarchies.
917 */
918#define CSS_SET_HASH_BITS 7
919static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
920
921static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
922{
923 unsigned long key = 0UL;
924 struct cgroup_subsys *ss;
925 int i;
926
927 for_each_subsys(ss, i)
928 key += (unsigned long)css[i];
929 key = (key >> 16) ^ key;
930
931 return key;
932}
933
934void put_css_set_locked(struct css_set *cset)
935{
936 struct cgrp_cset_link *link, *tmp_link;
937 struct cgroup_subsys *ss;
938 int ssid;
939
940 lockdep_assert_held(&css_set_lock);
941
942 if (!refcount_dec_and_test(&cset->refcount))
943 return;
944
945 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
946
947 /* This css_set is dead. unlink it and release cgroup and css refs */
948 for_each_subsys(ss, ssid) {
949 list_del(&cset->e_cset_node[ssid]);
950 css_put(cset->subsys[ssid]);
951 }
952 hash_del(&cset->hlist);
953 css_set_count--;
954
955 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
956 list_del(&link->cset_link);
957 list_del(&link->cgrp_link);
958 if (cgroup_parent(link->cgrp))
959 cgroup_put(link->cgrp);
960 kfree(link);
961 }
962
963 if (css_set_threaded(cset)) {
964 list_del(&cset->threaded_csets_node);
965 put_css_set_locked(cset->dom_cset);
966 }
967
968 kfree_rcu(cset, rcu_head);
969}
970
971/**
972 * compare_css_sets - helper function for find_existing_css_set().
973 * @cset: candidate css_set being tested
974 * @old_cset: existing css_set for a task
975 * @new_cgrp: cgroup that's being entered by the task
976 * @template: desired set of css pointers in css_set (pre-calculated)
977 *
978 * Returns true if "cset" matches "old_cset" except for the hierarchy
979 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
980 */
981static bool compare_css_sets(struct css_set *cset,
982 struct css_set *old_cset,
983 struct cgroup *new_cgrp,
984 struct cgroup_subsys_state *template[])
985{
986 struct cgroup *new_dfl_cgrp;
987 struct list_head *l1, *l2;
988
989 /*
990 * On the default hierarchy, there can be csets which are
991 * associated with the same set of cgroups but different csses.
992 * Let's first ensure that csses match.
993 */
994 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
995 return false;
996
997
998 /* @cset's domain should match the default cgroup's */
999 if (cgroup_on_dfl(new_cgrp))
1000 new_dfl_cgrp = new_cgrp;
1001 else
1002 new_dfl_cgrp = old_cset->dfl_cgrp;
1003
1004 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1005 return false;
1006
1007 /*
1008 * Compare cgroup pointers in order to distinguish between
1009 * different cgroups in hierarchies. As different cgroups may
1010 * share the same effective css, this comparison is always
1011 * necessary.
1012 */
1013 l1 = &cset->cgrp_links;
1014 l2 = &old_cset->cgrp_links;
1015 while (1) {
1016 struct cgrp_cset_link *link1, *link2;
1017 struct cgroup *cgrp1, *cgrp2;
1018
1019 l1 = l1->next;
1020 l2 = l2->next;
1021 /* See if we reached the end - both lists are equal length. */
1022 if (l1 == &cset->cgrp_links) {
1023 BUG_ON(l2 != &old_cset->cgrp_links);
1024 break;
1025 } else {
1026 BUG_ON(l2 == &old_cset->cgrp_links);
1027 }
1028 /* Locate the cgroups associated with these links. */
1029 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1030 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1031 cgrp1 = link1->cgrp;
1032 cgrp2 = link2->cgrp;
1033 /* Hierarchies should be linked in the same order. */
1034 BUG_ON(cgrp1->root != cgrp2->root);
1035
1036 /*
1037 * If this hierarchy is the hierarchy of the cgroup
1038 * that's changing, then we need to check that this
1039 * css_set points to the new cgroup; if it's any other
1040 * hierarchy, then this css_set should point to the
1041 * same cgroup as the old css_set.
1042 */
1043 if (cgrp1->root == new_cgrp->root) {
1044 if (cgrp1 != new_cgrp)
1045 return false;
1046 } else {
1047 if (cgrp1 != cgrp2)
1048 return false;
1049 }
1050 }
1051 return true;
1052}
1053
1054/**
1055 * find_existing_css_set - init css array and find the matching css_set
1056 * @old_cset: the css_set that we're using before the cgroup transition
1057 * @cgrp: the cgroup that we're moving into
1058 * @template: out param for the new set of csses, should be clear on entry
1059 */
1060static struct css_set *find_existing_css_set(struct css_set *old_cset,
1061 struct cgroup *cgrp,
1062 struct cgroup_subsys_state *template[])
1063{
1064 struct cgroup_root *root = cgrp->root;
1065 struct cgroup_subsys *ss;
1066 struct css_set *cset;
1067 unsigned long key;
1068 int i;
1069
1070 /*
1071 * Build the set of subsystem state objects that we want to see in the
1072 * new css_set. while subsystems can change globally, the entries here
1073 * won't change, so no need for locking.
1074 */
1075 for_each_subsys(ss, i) {
1076 if (root->subsys_mask & (1UL << i)) {
1077 /*
1078 * @ss is in this hierarchy, so we want the
1079 * effective css from @cgrp.
1080 */
1081 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1082 } else {
1083 /*
1084 * @ss is not in this hierarchy, so we don't want
1085 * to change the css.
1086 */
1087 template[i] = old_cset->subsys[i];
1088 }
1089 }
1090
1091 key = css_set_hash(template);
1092 hash_for_each_possible(css_set_table, cset, hlist, key) {
1093 if (!compare_css_sets(cset, old_cset, cgrp, template))
1094 continue;
1095
1096 /* This css_set matches what we need */
1097 return cset;
1098 }
1099
1100 /* No existing cgroup group matched */
1101 return NULL;
1102}
1103
1104static void free_cgrp_cset_links(struct list_head *links_to_free)
1105{
1106 struct cgrp_cset_link *link, *tmp_link;
1107
1108 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1109 list_del(&link->cset_link);
1110 kfree(link);
1111 }
1112}
1113
1114/**
1115 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1116 * @count: the number of links to allocate
1117 * @tmp_links: list_head the allocated links are put on
1118 *
1119 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1120 * through ->cset_link. Returns 0 on success or -errno.
1121 */
1122static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1123{
1124 struct cgrp_cset_link *link;
1125 int i;
1126
1127 INIT_LIST_HEAD(tmp_links);
1128
1129 for (i = 0; i < count; i++) {
1130 link = kzalloc(sizeof(*link), GFP_KERNEL);
1131 if (!link) {
1132 free_cgrp_cset_links(tmp_links);
1133 return -ENOMEM;
1134 }
1135 list_add(&link->cset_link, tmp_links);
1136 }
1137 return 0;
1138}
1139
1140/**
1141 * link_css_set - a helper function to link a css_set to a cgroup
1142 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1143 * @cset: the css_set to be linked
1144 * @cgrp: the destination cgroup
1145 */
1146static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1147 struct cgroup *cgrp)
1148{
1149 struct cgrp_cset_link *link;
1150
1151 BUG_ON(list_empty(tmp_links));
1152
1153 if (cgroup_on_dfl(cgrp))
1154 cset->dfl_cgrp = cgrp;
1155
1156 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1157 link->cset = cset;
1158 link->cgrp = cgrp;
1159
1160 /*
1161 * Always add links to the tail of the lists so that the lists are
1162 * in choronological order.
1163 */
1164 list_move_tail(&link->cset_link, &cgrp->cset_links);
1165 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1166
1167 if (cgroup_parent(cgrp))
1168 cgroup_get_live(cgrp);
1169}
1170
1171/**
1172 * find_css_set - return a new css_set with one cgroup updated
1173 * @old_cset: the baseline css_set
1174 * @cgrp: the cgroup to be updated
1175 *
1176 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1177 * substituted into the appropriate hierarchy.
1178 */
1179static struct css_set *find_css_set(struct css_set *old_cset,
1180 struct cgroup *cgrp)
1181{
1182 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1183 struct css_set *cset;
1184 struct list_head tmp_links;
1185 struct cgrp_cset_link *link;
1186 struct cgroup_subsys *ss;
1187 unsigned long key;
1188 int ssid;
1189
1190 lockdep_assert_held(&cgroup_mutex);
1191
1192 /* First see if we already have a cgroup group that matches
1193 * the desired set */
1194 spin_lock_irq(&css_set_lock);
1195 cset = find_existing_css_set(old_cset, cgrp, template);
1196 if (cset)
1197 get_css_set(cset);
1198 spin_unlock_irq(&css_set_lock);
1199
1200 if (cset)
1201 return cset;
1202
1203 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1204 if (!cset)
1205 return NULL;
1206
1207 /* Allocate all the cgrp_cset_link objects that we'll need */
1208 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1209 kfree(cset);
1210 return NULL;
1211 }
1212
1213 refcount_set(&cset->refcount, 1);
1214 cset->dom_cset = cset;
1215 INIT_LIST_HEAD(&cset->tasks);
1216 INIT_LIST_HEAD(&cset->mg_tasks);
1217 INIT_LIST_HEAD(&cset->dying_tasks);
1218 INIT_LIST_HEAD(&cset->task_iters);
1219 INIT_LIST_HEAD(&cset->threaded_csets);
1220 INIT_HLIST_NODE(&cset->hlist);
1221 INIT_LIST_HEAD(&cset->cgrp_links);
1222 INIT_LIST_HEAD(&cset->mg_preload_node);
1223 INIT_LIST_HEAD(&cset->mg_node);
1224
1225 /* Copy the set of subsystem state objects generated in
1226 * find_existing_css_set() */
1227 memcpy(cset->subsys, template, sizeof(cset->subsys));
1228
1229 spin_lock_irq(&css_set_lock);
1230 /* Add reference counts and links from the new css_set. */
1231 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1232 struct cgroup *c = link->cgrp;
1233
1234 if (c->root == cgrp->root)
1235 c = cgrp;
1236 link_css_set(&tmp_links, cset, c);
1237 }
1238
1239 BUG_ON(!list_empty(&tmp_links));
1240
1241 css_set_count++;
1242
1243 /* Add @cset to the hash table */
1244 key = css_set_hash(cset->subsys);
1245 hash_add(css_set_table, &cset->hlist, key);
1246
1247 for_each_subsys(ss, ssid) {
1248 struct cgroup_subsys_state *css = cset->subsys[ssid];
1249
1250 list_add_tail(&cset->e_cset_node[ssid],
1251 &css->cgroup->e_csets[ssid]);
1252 css_get(css);
1253 }
1254
1255 spin_unlock_irq(&css_set_lock);
1256
1257 /*
1258 * If @cset should be threaded, look up the matching dom_cset and
1259 * link them up. We first fully initialize @cset then look for the
1260 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1261 * to stay empty until we return.
1262 */
1263 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1264 struct css_set *dcset;
1265
1266 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1267 if (!dcset) {
1268 put_css_set(cset);
1269 return NULL;
1270 }
1271
1272 spin_lock_irq(&css_set_lock);
1273 cset->dom_cset = dcset;
1274 list_add_tail(&cset->threaded_csets_node,
1275 &dcset->threaded_csets);
1276 spin_unlock_irq(&css_set_lock);
1277 }
1278
1279 return cset;
1280}
1281
1282struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1283{
1284 struct cgroup *root_cgrp = kf_root->kn->priv;
1285
1286 return root_cgrp->root;
1287}
1288
1289static int cgroup_init_root_id(struct cgroup_root *root)
1290{
1291 int id;
1292
1293 lockdep_assert_held(&cgroup_mutex);
1294
1295 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1296 if (id < 0)
1297 return id;
1298
1299 root->hierarchy_id = id;
1300 return 0;
1301}
1302
1303static void cgroup_exit_root_id(struct cgroup_root *root)
1304{
1305 lockdep_assert_held(&cgroup_mutex);
1306
1307 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1308}
1309
1310void cgroup_free_root(struct cgroup_root *root)
1311{
1312 if (root) {
1313 idr_destroy(&root->cgroup_idr);
1314 kfree(root);
1315 }
1316}
1317
1318static void cgroup_destroy_root(struct cgroup_root *root)
1319{
1320 struct cgroup *cgrp = &root->cgrp;
1321 struct cgrp_cset_link *link, *tmp_link;
1322
1323 trace_cgroup_destroy_root(root);
1324
1325 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1326
1327 BUG_ON(atomic_read(&root->nr_cgrps));
1328 BUG_ON(!list_empty(&cgrp->self.children));
1329
1330 /* Rebind all subsystems back to the default hierarchy */
1331 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1332
1333 /*
1334 * Release all the links from cset_links to this hierarchy's
1335 * root cgroup
1336 */
1337 spin_lock_irq(&css_set_lock);
1338
1339 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1340 list_del(&link->cset_link);
1341 list_del(&link->cgrp_link);
1342 kfree(link);
1343 }
1344
1345 spin_unlock_irq(&css_set_lock);
1346
1347 if (!list_empty(&root->root_list)) {
1348 list_del(&root->root_list);
1349 cgroup_root_count--;
1350 }
1351
1352 cgroup_exit_root_id(root);
1353
1354 mutex_unlock(&cgroup_mutex);
1355
1356 kernfs_destroy_root(root->kf_root);
1357 cgroup_free_root(root);
1358}
1359
1360/*
1361 * look up cgroup associated with current task's cgroup namespace on the
1362 * specified hierarchy
1363 */
1364static struct cgroup *
1365current_cgns_cgroup_from_root(struct cgroup_root *root)
1366{
1367 struct cgroup *res = NULL;
1368 struct css_set *cset;
1369
1370 lockdep_assert_held(&css_set_lock);
1371
1372 rcu_read_lock();
1373
1374 cset = current->nsproxy->cgroup_ns->root_cset;
1375 if (cset == &init_css_set) {
1376 res = &root->cgrp;
1377 } else {
1378 struct cgrp_cset_link *link;
1379
1380 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1381 struct cgroup *c = link->cgrp;
1382
1383 if (c->root == root) {
1384 res = c;
1385 break;
1386 }
1387 }
1388 }
1389 rcu_read_unlock();
1390
1391 BUG_ON(!res);
1392 return res;
1393}
1394
1395/* look up cgroup associated with given css_set on the specified hierarchy */
1396static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1397 struct cgroup_root *root)
1398{
1399 struct cgroup *res = NULL;
1400
1401 lockdep_assert_held(&cgroup_mutex);
1402 lockdep_assert_held(&css_set_lock);
1403
1404 if (cset == &init_css_set) {
1405 res = &root->cgrp;
1406 } else if (root == &cgrp_dfl_root) {
1407 res = cset->dfl_cgrp;
1408 } else {
1409 struct cgrp_cset_link *link;
1410
1411 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1412 struct cgroup *c = link->cgrp;
1413
1414 if (c->root == root) {
1415 res = c;
1416 break;
1417 }
1418 }
1419 }
1420
1421 BUG_ON(!res);
1422 return res;
1423}
1424
1425/*
1426 * Return the cgroup for "task" from the given hierarchy. Must be
1427 * called with cgroup_mutex and css_set_lock held.
1428 */
1429struct cgroup *task_cgroup_from_root(struct task_struct *task,
1430 struct cgroup_root *root)
1431{
1432 /*
1433 * No need to lock the task - since we hold cgroup_mutex the
1434 * task can't change groups, so the only thing that can happen
1435 * is that it exits and its css is set back to init_css_set.
1436 */
1437 return cset_cgroup_from_root(task_css_set(task), root);
1438}
1439
1440/*
1441 * A task must hold cgroup_mutex to modify cgroups.
1442 *
1443 * Any task can increment and decrement the count field without lock.
1444 * So in general, code holding cgroup_mutex can't rely on the count
1445 * field not changing. However, if the count goes to zero, then only
1446 * cgroup_attach_task() can increment it again. Because a count of zero
1447 * means that no tasks are currently attached, therefore there is no
1448 * way a task attached to that cgroup can fork (the other way to
1449 * increment the count). So code holding cgroup_mutex can safely
1450 * assume that if the count is zero, it will stay zero. Similarly, if
1451 * a task holds cgroup_mutex on a cgroup with zero count, it
1452 * knows that the cgroup won't be removed, as cgroup_rmdir()
1453 * needs that mutex.
1454 *
1455 * A cgroup can only be deleted if both its 'count' of using tasks
1456 * is zero, and its list of 'children' cgroups is empty. Since all
1457 * tasks in the system use _some_ cgroup, and since there is always at
1458 * least one task in the system (init, pid == 1), therefore, root cgroup
1459 * always has either children cgroups and/or using tasks. So we don't
1460 * need a special hack to ensure that root cgroup cannot be deleted.
1461 *
1462 * P.S. One more locking exception. RCU is used to guard the
1463 * update of a tasks cgroup pointer by cgroup_attach_task()
1464 */
1465
1466static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1467
1468static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1469 char *buf)
1470{
1471 struct cgroup_subsys *ss = cft->ss;
1472
1473 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1474 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1475 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1476
1477 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1478 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1479 cft->name);
1480 } else {
1481 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1482 }
1483 return buf;
1484}
1485
1486/**
1487 * cgroup_file_mode - deduce file mode of a control file
1488 * @cft: the control file in question
1489 *
1490 * S_IRUGO for read, S_IWUSR for write.
1491 */
1492static umode_t cgroup_file_mode(const struct cftype *cft)
1493{
1494 umode_t mode = 0;
1495
1496 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1497 mode |= S_IRUGO;
1498
1499 if (cft->write_u64 || cft->write_s64 || cft->write) {
1500 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1501 mode |= S_IWUGO;
1502 else
1503 mode |= S_IWUSR;
1504 }
1505
1506 return mode;
1507}
1508
1509/**
1510 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1511 * @subtree_control: the new subtree_control mask to consider
1512 * @this_ss_mask: available subsystems
1513 *
1514 * On the default hierarchy, a subsystem may request other subsystems to be
1515 * enabled together through its ->depends_on mask. In such cases, more
1516 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1517 *
1518 * This function calculates which subsystems need to be enabled if
1519 * @subtree_control is to be applied while restricted to @this_ss_mask.
1520 */
1521static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1522{
1523 u16 cur_ss_mask = subtree_control;
1524 struct cgroup_subsys *ss;
1525 int ssid;
1526
1527 lockdep_assert_held(&cgroup_mutex);
1528
1529 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1530
1531 while (true) {
1532 u16 new_ss_mask = cur_ss_mask;
1533
1534 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1535 new_ss_mask |= ss->depends_on;
1536 } while_each_subsys_mask();
1537
1538 /*
1539 * Mask out subsystems which aren't available. This can
1540 * happen only if some depended-upon subsystems were bound
1541 * to non-default hierarchies.
1542 */
1543 new_ss_mask &= this_ss_mask;
1544
1545 if (new_ss_mask == cur_ss_mask)
1546 break;
1547 cur_ss_mask = new_ss_mask;
1548 }
1549
1550 return cur_ss_mask;
1551}
1552
1553/**
1554 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1555 * @kn: the kernfs_node being serviced
1556 *
1557 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1558 * the method finishes if locking succeeded. Note that once this function
1559 * returns the cgroup returned by cgroup_kn_lock_live() may become
1560 * inaccessible any time. If the caller intends to continue to access the
1561 * cgroup, it should pin it before invoking this function.
1562 */
1563void cgroup_kn_unlock(struct kernfs_node *kn)
1564{
1565 struct cgroup *cgrp;
1566
1567 if (kernfs_type(kn) == KERNFS_DIR)
1568 cgrp = kn->priv;
1569 else
1570 cgrp = kn->parent->priv;
1571
1572 mutex_unlock(&cgroup_mutex);
1573
1574 kernfs_unbreak_active_protection(kn);
1575 cgroup_put(cgrp);
1576}
1577
1578/**
1579 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1580 * @kn: the kernfs_node being serviced
1581 * @drain_offline: perform offline draining on the cgroup
1582 *
1583 * This helper is to be used by a cgroup kernfs method currently servicing
1584 * @kn. It breaks the active protection, performs cgroup locking and
1585 * verifies that the associated cgroup is alive. Returns the cgroup if
1586 * alive; otherwise, %NULL. A successful return should be undone by a
1587 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1588 * cgroup is drained of offlining csses before return.
1589 *
1590 * Any cgroup kernfs method implementation which requires locking the
1591 * associated cgroup should use this helper. It avoids nesting cgroup
1592 * locking under kernfs active protection and allows all kernfs operations
1593 * including self-removal.
1594 */
1595struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1596{
1597 struct cgroup *cgrp;
1598
1599 if (kernfs_type(kn) == KERNFS_DIR)
1600 cgrp = kn->priv;
1601 else
1602 cgrp = kn->parent->priv;
1603
1604 /*
1605 * We're gonna grab cgroup_mutex which nests outside kernfs
1606 * active_ref. cgroup liveliness check alone provides enough
1607 * protection against removal. Ensure @cgrp stays accessible and
1608 * break the active_ref protection.
1609 */
1610 if (!cgroup_tryget(cgrp))
1611 return NULL;
1612 kernfs_break_active_protection(kn);
1613
1614 if (drain_offline)
1615 cgroup_lock_and_drain_offline(cgrp);
1616 else
1617 mutex_lock(&cgroup_mutex);
1618
1619 if (!cgroup_is_dead(cgrp))
1620 return cgrp;
1621
1622 cgroup_kn_unlock(kn);
1623 return NULL;
1624}
1625
1626static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1627{
1628 char name[CGROUP_FILE_NAME_MAX];
1629
1630 lockdep_assert_held(&cgroup_mutex);
1631
1632 if (cft->file_offset) {
1633 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1634 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1635
1636 spin_lock_irq(&cgroup_file_kn_lock);
1637 cfile->kn = NULL;
1638 spin_unlock_irq(&cgroup_file_kn_lock);
1639
1640 del_timer_sync(&cfile->notify_timer);
1641 }
1642
1643 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1644}
1645
1646/**
1647 * css_clear_dir - remove subsys files in a cgroup directory
1648 * @css: taget css
1649 */
1650static void css_clear_dir(struct cgroup_subsys_state *css)
1651{
1652 struct cgroup *cgrp = css->cgroup;
1653 struct cftype *cfts;
1654
1655 if (!(css->flags & CSS_VISIBLE))
1656 return;
1657
1658 css->flags &= ~CSS_VISIBLE;
1659
1660 if (!css->ss) {
1661 if (cgroup_on_dfl(cgrp))
1662 cfts = cgroup_base_files;
1663 else
1664 cfts = cgroup1_base_files;
1665
1666 cgroup_addrm_files(css, cgrp, cfts, false);
1667 } else {
1668 list_for_each_entry(cfts, &css->ss->cfts, node)
1669 cgroup_addrm_files(css, cgrp, cfts, false);
1670 }
1671}
1672
1673/**
1674 * css_populate_dir - create subsys files in a cgroup directory
1675 * @css: target css
1676 *
1677 * On failure, no file is added.
1678 */
1679static int css_populate_dir(struct cgroup_subsys_state *css)
1680{
1681 struct cgroup *cgrp = css->cgroup;
1682 struct cftype *cfts, *failed_cfts;
1683 int ret;
1684
1685 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1686 return 0;
1687
1688 if (!css->ss) {
1689 if (cgroup_on_dfl(cgrp))
1690 cfts = cgroup_base_files;
1691 else
1692 cfts = cgroup1_base_files;
1693
1694 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1695 if (ret < 0)
1696 return ret;
1697 } else {
1698 list_for_each_entry(cfts, &css->ss->cfts, node) {
1699 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1700 if (ret < 0) {
1701 failed_cfts = cfts;
1702 goto err;
1703 }
1704 }
1705 }
1706
1707 css->flags |= CSS_VISIBLE;
1708
1709 return 0;
1710err:
1711 list_for_each_entry(cfts, &css->ss->cfts, node) {
1712 if (cfts == failed_cfts)
1713 break;
1714 cgroup_addrm_files(css, cgrp, cfts, false);
1715 }
1716 return ret;
1717}
1718
1719int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1720{
1721 struct cgroup *dcgrp = &dst_root->cgrp;
1722 struct cgroup_subsys *ss;
1723 int ssid, i, ret;
1724
1725 lockdep_assert_held(&cgroup_mutex);
1726
1727 do_each_subsys_mask(ss, ssid, ss_mask) {
1728 /*
1729 * If @ss has non-root csses attached to it, can't move.
1730 * If @ss is an implicit controller, it is exempt from this
1731 * rule and can be stolen.
1732 */
1733 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1734 !ss->implicit_on_dfl)
1735 return -EBUSY;
1736
1737 /* can't move between two non-dummy roots either */
1738 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1739 return -EBUSY;
1740 } while_each_subsys_mask();
1741
1742 do_each_subsys_mask(ss, ssid, ss_mask) {
1743 struct cgroup_root *src_root = ss->root;
1744 struct cgroup *scgrp = &src_root->cgrp;
1745 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1746 struct css_set *cset;
1747
1748 WARN_ON(!css || cgroup_css(dcgrp, ss));
1749
1750 /* disable from the source */
1751 src_root->subsys_mask &= ~(1 << ssid);
1752 WARN_ON(cgroup_apply_control(scgrp));
1753 cgroup_finalize_control(scgrp, 0);
1754
1755 /* rebind */
1756 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1757 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1758 ss->root = dst_root;
1759 css->cgroup = dcgrp;
1760
1761 spin_lock_irq(&css_set_lock);
1762 hash_for_each(css_set_table, i, cset, hlist)
1763 list_move_tail(&cset->e_cset_node[ss->id],
1764 &dcgrp->e_csets[ss->id]);
1765 spin_unlock_irq(&css_set_lock);
1766
1767 /* default hierarchy doesn't enable controllers by default */
1768 dst_root->subsys_mask |= 1 << ssid;
1769 if (dst_root == &cgrp_dfl_root) {
1770 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1771 } else {
1772 dcgrp->subtree_control |= 1 << ssid;
1773 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1774 }
1775
1776 ret = cgroup_apply_control(dcgrp);
1777 if (ret)
1778 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1779 ss->name, ret);
1780
1781 if (ss->bind)
1782 ss->bind(css);
1783 } while_each_subsys_mask();
1784
1785 kernfs_activate(dcgrp->kn);
1786 return 0;
1787}
1788
1789int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1790 struct kernfs_root *kf_root)
1791{
1792 int len = 0;
1793 char *buf = NULL;
1794 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1795 struct cgroup *ns_cgroup;
1796
1797 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1798 if (!buf)
1799 return -ENOMEM;
1800
1801 spin_lock_irq(&css_set_lock);
1802 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1803 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1804 spin_unlock_irq(&css_set_lock);
1805
1806 if (len >= PATH_MAX)
1807 len = -ERANGE;
1808 else if (len > 0) {
1809 seq_escape(sf, buf, " \t\n\\");
1810 len = 0;
1811 }
1812 kfree(buf);
1813 return len;
1814}
1815
1816enum cgroup2_param {
1817 Opt_nsdelegate,
1818 Opt_memory_localevents,
1819 nr__cgroup2_params
1820};
1821
1822static const struct fs_parameter_spec cgroup2_param_specs[] = {
1823 fsparam_flag("nsdelegate", Opt_nsdelegate),
1824 fsparam_flag("memory_localevents", Opt_memory_localevents),
1825 {}
1826};
1827
1828static const struct fs_parameter_description cgroup2_fs_parameters = {
1829 .name = "cgroup2",
1830 .specs = cgroup2_param_specs,
1831};
1832
1833static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1834{
1835 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1836 struct fs_parse_result result;
1837 int opt;
1838
1839 opt = fs_parse(fc, &cgroup2_fs_parameters, param, &result);
1840 if (opt < 0)
1841 return opt;
1842
1843 switch (opt) {
1844 case Opt_nsdelegate:
1845 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1846 return 0;
1847 case Opt_memory_localevents:
1848 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1849 return 0;
1850 }
1851 return -EINVAL;
1852}
1853
1854static void apply_cgroup_root_flags(unsigned int root_flags)
1855{
1856 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1857 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1858 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1859 else
1860 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1861
1862 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1863 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1864 else
1865 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1866 }
1867}
1868
1869static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1870{
1871 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1872 seq_puts(seq, ",nsdelegate");
1873 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1874 seq_puts(seq, ",memory_localevents");
1875 return 0;
1876}
1877
1878static int cgroup_reconfigure(struct fs_context *fc)
1879{
1880 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1881
1882 apply_cgroup_root_flags(ctx->flags);
1883 return 0;
1884}
1885
1886/*
1887 * To reduce the fork() overhead for systems that are not actually using
1888 * their cgroups capability, we don't maintain the lists running through
1889 * each css_set to its tasks until we see the list actually used - in other
1890 * words after the first mount.
1891 */
1892static bool use_task_css_set_links __read_mostly;
1893
1894void cgroup_enable_task_cg_lists(void)
1895{
1896 struct task_struct *p, *g;
1897
1898 /*
1899 * We need tasklist_lock because RCU is not safe against
1900 * while_each_thread(). Besides, a forking task that has passed
1901 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1902 * is not guaranteed to have its child immediately visible in the
1903 * tasklist if we walk through it with RCU.
1904 */
1905 read_lock(&tasklist_lock);
1906 spin_lock_irq(&css_set_lock);
1907
1908 if (use_task_css_set_links)
1909 goto out_unlock;
1910
1911 use_task_css_set_links = true;
1912
1913 do_each_thread(g, p) {
1914 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1915 task_css_set(p) != &init_css_set);
1916
1917 /*
1918 * We should check if the process is exiting, otherwise
1919 * it will race with cgroup_exit() in that the list
1920 * entry won't be deleted though the process has exited.
1921 * Do it while holding siglock so that we don't end up
1922 * racing against cgroup_exit().
1923 *
1924 * Interrupts were already disabled while acquiring
1925 * the css_set_lock, so we do not need to disable it
1926 * again when acquiring the sighand->siglock here.
1927 */
1928 spin_lock(&p->sighand->siglock);
1929 if (!(p->flags & PF_EXITING)) {
1930 struct css_set *cset = task_css_set(p);
1931
1932 if (!css_set_populated(cset))
1933 css_set_update_populated(cset, true);
1934 list_add_tail(&p->cg_list, &cset->tasks);
1935 get_css_set(cset);
1936 cset->nr_tasks++;
1937 }
1938 spin_unlock(&p->sighand->siglock);
1939 } while_each_thread(g, p);
1940out_unlock:
1941 spin_unlock_irq(&css_set_lock);
1942 read_unlock(&tasklist_lock);
1943}
1944
1945static void init_cgroup_housekeeping(struct cgroup *cgrp)
1946{
1947 struct cgroup_subsys *ss;
1948 int ssid;
1949
1950 INIT_LIST_HEAD(&cgrp->self.sibling);
1951 INIT_LIST_HEAD(&cgrp->self.children);
1952 INIT_LIST_HEAD(&cgrp->cset_links);
1953 INIT_LIST_HEAD(&cgrp->pidlists);
1954 mutex_init(&cgrp->pidlist_mutex);
1955 cgrp->self.cgroup = cgrp;
1956 cgrp->self.flags |= CSS_ONLINE;
1957 cgrp->dom_cgrp = cgrp;
1958 cgrp->max_descendants = INT_MAX;
1959 cgrp->max_depth = INT_MAX;
1960 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1961 prev_cputime_init(&cgrp->prev_cputime);
1962
1963 for_each_subsys(ss, ssid)
1964 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1965
1966 init_waitqueue_head(&cgrp->offline_waitq);
1967 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1968}
1969
1970void init_cgroup_root(struct cgroup_fs_context *ctx)
1971{
1972 struct cgroup_root *root = ctx->root;
1973 struct cgroup *cgrp = &root->cgrp;
1974
1975 INIT_LIST_HEAD(&root->root_list);
1976 atomic_set(&root->nr_cgrps, 1);
1977 cgrp->root = root;
1978 init_cgroup_housekeeping(cgrp);
1979 idr_init(&root->cgroup_idr);
1980
1981 root->flags = ctx->flags;
1982 if (ctx->release_agent)
1983 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1984 if (ctx->name)
1985 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1986 if (ctx->cpuset_clone_children)
1987 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1988}
1989
1990int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1991{
1992 LIST_HEAD(tmp_links);
1993 struct cgroup *root_cgrp = &root->cgrp;
1994 struct kernfs_syscall_ops *kf_sops;
1995 struct css_set *cset;
1996 int i, ret;
1997
1998 lockdep_assert_held(&cgroup_mutex);
1999
2000 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
2001 if (ret < 0)
2002 goto out;
2003 root_cgrp->id = ret;
2004 root_cgrp->ancestor_ids[0] = ret;
2005
2006 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2007 0, GFP_KERNEL);
2008 if (ret)
2009 goto out;
2010
2011 /*
2012 * We're accessing css_set_count without locking css_set_lock here,
2013 * but that's OK - it can only be increased by someone holding
2014 * cgroup_lock, and that's us. Later rebinding may disable
2015 * controllers on the default hierarchy and thus create new csets,
2016 * which can't be more than the existing ones. Allocate 2x.
2017 */
2018 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2019 if (ret)
2020 goto cancel_ref;
2021
2022 ret = cgroup_init_root_id(root);
2023 if (ret)
2024 goto cancel_ref;
2025
2026 kf_sops = root == &cgrp_dfl_root ?
2027 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2028
2029 root->kf_root = kernfs_create_root(kf_sops,
2030 KERNFS_ROOT_CREATE_DEACTIVATED |
2031 KERNFS_ROOT_SUPPORT_EXPORTOP,
2032 root_cgrp);
2033 if (IS_ERR(root->kf_root)) {
2034 ret = PTR_ERR(root->kf_root);
2035 goto exit_root_id;
2036 }
2037 root_cgrp->kn = root->kf_root->kn;
2038
2039 ret = css_populate_dir(&root_cgrp->self);
2040 if (ret)
2041 goto destroy_root;
2042
2043 ret = rebind_subsystems(root, ss_mask);
2044 if (ret)
2045 goto destroy_root;
2046
2047 ret = cgroup_bpf_inherit(root_cgrp);
2048 WARN_ON_ONCE(ret);
2049
2050 trace_cgroup_setup_root(root);
2051
2052 /*
2053 * There must be no failure case after here, since rebinding takes
2054 * care of subsystems' refcounts, which are explicitly dropped in
2055 * the failure exit path.
2056 */
2057 list_add(&root->root_list, &cgroup_roots);
2058 cgroup_root_count++;
2059
2060 /*
2061 * Link the root cgroup in this hierarchy into all the css_set
2062 * objects.
2063 */
2064 spin_lock_irq(&css_set_lock);
2065 hash_for_each(css_set_table, i, cset, hlist) {
2066 link_css_set(&tmp_links, cset, root_cgrp);
2067 if (css_set_populated(cset))
2068 cgroup_update_populated(root_cgrp, true);
2069 }
2070 spin_unlock_irq(&css_set_lock);
2071
2072 BUG_ON(!list_empty(&root_cgrp->self.children));
2073 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2074
2075 kernfs_activate(root_cgrp->kn);
2076 ret = 0;
2077 goto out;
2078
2079destroy_root:
2080 kernfs_destroy_root(root->kf_root);
2081 root->kf_root = NULL;
2082exit_root_id:
2083 cgroup_exit_root_id(root);
2084cancel_ref:
2085 percpu_ref_exit(&root_cgrp->self.refcnt);
2086out:
2087 free_cgrp_cset_links(&tmp_links);
2088 return ret;
2089}
2090
2091int cgroup_do_get_tree(struct fs_context *fc)
2092{
2093 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2094 int ret;
2095
2096 ctx->kfc.root = ctx->root->kf_root;
2097 if (fc->fs_type == &cgroup2_fs_type)
2098 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2099 else
2100 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2101 ret = kernfs_get_tree(fc);
2102
2103 /*
2104 * In non-init cgroup namespace, instead of root cgroup's dentry,
2105 * we return the dentry corresponding to the cgroupns->root_cgrp.
2106 */
2107 if (!ret && ctx->ns != &init_cgroup_ns) {
2108 struct dentry *nsdentry;
2109 struct super_block *sb = fc->root->d_sb;
2110 struct cgroup *cgrp;
2111
2112 mutex_lock(&cgroup_mutex);
2113 spin_lock_irq(&css_set_lock);
2114
2115 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2116
2117 spin_unlock_irq(&css_set_lock);
2118 mutex_unlock(&cgroup_mutex);
2119
2120 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2121 dput(fc->root);
2122 if (IS_ERR(nsdentry)) {
2123 deactivate_locked_super(sb);
2124 ret = PTR_ERR(nsdentry);
2125 nsdentry = NULL;
2126 }
2127 fc->root = nsdentry;
2128 }
2129
2130 if (!ctx->kfc.new_sb_created)
2131 cgroup_put(&ctx->root->cgrp);
2132
2133 return ret;
2134}
2135
2136/*
2137 * Destroy a cgroup filesystem context.
2138 */
2139static void cgroup_fs_context_free(struct fs_context *fc)
2140{
2141 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2142
2143 kfree(ctx->name);
2144 kfree(ctx->release_agent);
2145 put_cgroup_ns(ctx->ns);
2146 kernfs_free_fs_context(fc);
2147 kfree(ctx);
2148}
2149
2150static int cgroup_get_tree(struct fs_context *fc)
2151{
2152 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2153 int ret;
2154
2155 cgrp_dfl_visible = true;
2156 cgroup_get_live(&cgrp_dfl_root.cgrp);
2157 ctx->root = &cgrp_dfl_root;
2158
2159 ret = cgroup_do_get_tree(fc);
2160 if (!ret)
2161 apply_cgroup_root_flags(ctx->flags);
2162 return ret;
2163}
2164
2165static const struct fs_context_operations cgroup_fs_context_ops = {
2166 .free = cgroup_fs_context_free,
2167 .parse_param = cgroup2_parse_param,
2168 .get_tree = cgroup_get_tree,
2169 .reconfigure = cgroup_reconfigure,
2170};
2171
2172static const struct fs_context_operations cgroup1_fs_context_ops = {
2173 .free = cgroup_fs_context_free,
2174 .parse_param = cgroup1_parse_param,
2175 .get_tree = cgroup1_get_tree,
2176 .reconfigure = cgroup1_reconfigure,
2177};
2178
2179/*
2180 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2181 * we select the namespace we're going to use.
2182 */
2183static int cgroup_init_fs_context(struct fs_context *fc)
2184{
2185 struct cgroup_fs_context *ctx;
2186
2187 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2188 if (!ctx)
2189 return -ENOMEM;
2190
2191 /*
2192 * The first time anyone tries to mount a cgroup, enable the list
2193 * linking each css_set to its tasks and fix up all existing tasks.
2194 */
2195 if (!use_task_css_set_links)
2196 cgroup_enable_task_cg_lists();
2197
2198 ctx->ns = current->nsproxy->cgroup_ns;
2199 get_cgroup_ns(ctx->ns);
2200 fc->fs_private = &ctx->kfc;
2201 if (fc->fs_type == &cgroup2_fs_type)
2202 fc->ops = &cgroup_fs_context_ops;
2203 else
2204 fc->ops = &cgroup1_fs_context_ops;
2205 put_user_ns(fc->user_ns);
2206 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2207 fc->global = true;
2208 return 0;
2209}
2210
2211static void cgroup_kill_sb(struct super_block *sb)
2212{
2213 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2214 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2215
2216 /*
2217 * If @root doesn't have any children, start killing it.
2218 * This prevents new mounts by disabling percpu_ref_tryget_live().
2219 * cgroup_mount() may wait for @root's release.
2220 *
2221 * And don't kill the default root.
2222 */
2223 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2224 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2225 percpu_ref_kill(&root->cgrp.self.refcnt);
2226 cgroup_put(&root->cgrp);
2227 kernfs_kill_sb(sb);
2228}
2229
2230struct file_system_type cgroup_fs_type = {
2231 .name = "cgroup",
2232 .init_fs_context = cgroup_init_fs_context,
2233 .parameters = &cgroup1_fs_parameters,
2234 .kill_sb = cgroup_kill_sb,
2235 .fs_flags = FS_USERNS_MOUNT,
2236};
2237
2238static struct file_system_type cgroup2_fs_type = {
2239 .name = "cgroup2",
2240 .init_fs_context = cgroup_init_fs_context,
2241 .parameters = &cgroup2_fs_parameters,
2242 .kill_sb = cgroup_kill_sb,
2243 .fs_flags = FS_USERNS_MOUNT,
2244};
2245
2246#ifdef CONFIG_CPUSETS
2247static const struct fs_context_operations cpuset_fs_context_ops = {
2248 .get_tree = cgroup1_get_tree,
2249 .free = cgroup_fs_context_free,
2250};
2251
2252/*
2253 * This is ugly, but preserves the userspace API for existing cpuset
2254 * users. If someone tries to mount the "cpuset" filesystem, we
2255 * silently switch it to mount "cgroup" instead
2256 */
2257static int cpuset_init_fs_context(struct fs_context *fc)
2258{
2259 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2260 struct cgroup_fs_context *ctx;
2261 int err;
2262
2263 err = cgroup_init_fs_context(fc);
2264 if (err) {
2265 kfree(agent);
2266 return err;
2267 }
2268
2269 fc->ops = &cpuset_fs_context_ops;
2270
2271 ctx = cgroup_fc2context(fc);
2272 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2273 ctx->flags |= CGRP_ROOT_NOPREFIX;
2274 ctx->release_agent = agent;
2275
2276 get_filesystem(&cgroup_fs_type);
2277 put_filesystem(fc->fs_type);
2278 fc->fs_type = &cgroup_fs_type;
2279
2280 return 0;
2281}
2282
2283static struct file_system_type cpuset_fs_type = {
2284 .name = "cpuset",
2285 .init_fs_context = cpuset_init_fs_context,
2286 .fs_flags = FS_USERNS_MOUNT,
2287};
2288#endif
2289
2290int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2291 struct cgroup_namespace *ns)
2292{
2293 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2294
2295 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2296}
2297
2298int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2299 struct cgroup_namespace *ns)
2300{
2301 int ret;
2302
2303 mutex_lock(&cgroup_mutex);
2304 spin_lock_irq(&css_set_lock);
2305
2306 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2307
2308 spin_unlock_irq(&css_set_lock);
2309 mutex_unlock(&cgroup_mutex);
2310
2311 return ret;
2312}
2313EXPORT_SYMBOL_GPL(cgroup_path_ns);
2314
2315/**
2316 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2317 * @task: target task
2318 * @buf: the buffer to write the path into
2319 * @buflen: the length of the buffer
2320 *
2321 * Determine @task's cgroup on the first (the one with the lowest non-zero
2322 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2323 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2324 * cgroup controller callbacks.
2325 *
2326 * Return value is the same as kernfs_path().
2327 */
2328int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2329{
2330 struct cgroup_root *root;
2331 struct cgroup *cgrp;
2332 int hierarchy_id = 1;
2333 int ret;
2334
2335 mutex_lock(&cgroup_mutex);
2336 spin_lock_irq(&css_set_lock);
2337
2338 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2339
2340 if (root) {
2341 cgrp = task_cgroup_from_root(task, root);
2342 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2343 } else {
2344 /* if no hierarchy exists, everyone is in "/" */
2345 ret = strlcpy(buf, "/", buflen);
2346 }
2347
2348 spin_unlock_irq(&css_set_lock);
2349 mutex_unlock(&cgroup_mutex);
2350 return ret;
2351}
2352EXPORT_SYMBOL_GPL(task_cgroup_path);
2353
2354/**
2355 * cgroup_migrate_add_task - add a migration target task to a migration context
2356 * @task: target task
2357 * @mgctx: target migration context
2358 *
2359 * Add @task, which is a migration target, to @mgctx->tset. This function
2360 * becomes noop if @task doesn't need to be migrated. @task's css_set
2361 * should have been added as a migration source and @task->cg_list will be
2362 * moved from the css_set's tasks list to mg_tasks one.
2363 */
2364static void cgroup_migrate_add_task(struct task_struct *task,
2365 struct cgroup_mgctx *mgctx)
2366{
2367 struct css_set *cset;
2368
2369 lockdep_assert_held(&css_set_lock);
2370
2371 /* @task either already exited or can't exit until the end */
2372 if (task->flags & PF_EXITING)
2373 return;
2374
2375 /* leave @task alone if post_fork() hasn't linked it yet */
2376 if (list_empty(&task->cg_list))
2377 return;
2378
2379 cset = task_css_set(task);
2380 if (!cset->mg_src_cgrp)
2381 return;
2382
2383 mgctx->tset.nr_tasks++;
2384
2385 list_move_tail(&task->cg_list, &cset->mg_tasks);
2386 if (list_empty(&cset->mg_node))
2387 list_add_tail(&cset->mg_node,
2388 &mgctx->tset.src_csets);
2389 if (list_empty(&cset->mg_dst_cset->mg_node))
2390 list_add_tail(&cset->mg_dst_cset->mg_node,
2391 &mgctx->tset.dst_csets);
2392}
2393
2394/**
2395 * cgroup_taskset_first - reset taskset and return the first task
2396 * @tset: taskset of interest
2397 * @dst_cssp: output variable for the destination css
2398 *
2399 * @tset iteration is initialized and the first task is returned.
2400 */
2401struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2402 struct cgroup_subsys_state **dst_cssp)
2403{
2404 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2405 tset->cur_task = NULL;
2406
2407 return cgroup_taskset_next(tset, dst_cssp);
2408}
2409
2410/**
2411 * cgroup_taskset_next - iterate to the next task in taskset
2412 * @tset: taskset of interest
2413 * @dst_cssp: output variable for the destination css
2414 *
2415 * Return the next task in @tset. Iteration must have been initialized
2416 * with cgroup_taskset_first().
2417 */
2418struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2419 struct cgroup_subsys_state **dst_cssp)
2420{
2421 struct css_set *cset = tset->cur_cset;
2422 struct task_struct *task = tset->cur_task;
2423
2424 while (&cset->mg_node != tset->csets) {
2425 if (!task)
2426 task = list_first_entry(&cset->mg_tasks,
2427 struct task_struct, cg_list);
2428 else
2429 task = list_next_entry(task, cg_list);
2430
2431 if (&task->cg_list != &cset->mg_tasks) {
2432 tset->cur_cset = cset;
2433 tset->cur_task = task;
2434
2435 /*
2436 * This function may be called both before and
2437 * after cgroup_taskset_migrate(). The two cases
2438 * can be distinguished by looking at whether @cset
2439 * has its ->mg_dst_cset set.
2440 */
2441 if (cset->mg_dst_cset)
2442 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2443 else
2444 *dst_cssp = cset->subsys[tset->ssid];
2445
2446 return task;
2447 }
2448
2449 cset = list_next_entry(cset, mg_node);
2450 task = NULL;
2451 }
2452
2453 return NULL;
2454}
2455
2456/**
2457 * cgroup_taskset_migrate - migrate a taskset
2458 * @mgctx: migration context
2459 *
2460 * Migrate tasks in @mgctx as setup by migration preparation functions.
2461 * This function fails iff one of the ->can_attach callbacks fails and
2462 * guarantees that either all or none of the tasks in @mgctx are migrated.
2463 * @mgctx is consumed regardless of success.
2464 */
2465static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2466{
2467 struct cgroup_taskset *tset = &mgctx->tset;
2468 struct cgroup_subsys *ss;
2469 struct task_struct *task, *tmp_task;
2470 struct css_set *cset, *tmp_cset;
2471 int ssid, failed_ssid, ret;
2472
2473 /* check that we can legitimately attach to the cgroup */
2474 if (tset->nr_tasks) {
2475 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2476 if (ss->can_attach) {
2477 tset->ssid = ssid;
2478 ret = ss->can_attach(tset);
2479 if (ret) {
2480 failed_ssid = ssid;
2481 goto out_cancel_attach;
2482 }
2483 }
2484 } while_each_subsys_mask();
2485 }
2486
2487 /*
2488 * Now that we're guaranteed success, proceed to move all tasks to
2489 * the new cgroup. There are no failure cases after here, so this
2490 * is the commit point.
2491 */
2492 spin_lock_irq(&css_set_lock);
2493 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2494 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2495 struct css_set *from_cset = task_css_set(task);
2496 struct css_set *to_cset = cset->mg_dst_cset;
2497
2498 get_css_set(to_cset);
2499 to_cset->nr_tasks++;
2500 css_set_move_task(task, from_cset, to_cset, true);
2501 from_cset->nr_tasks--;
2502 /*
2503 * If the source or destination cgroup is frozen,
2504 * the task might require to change its state.
2505 */
2506 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2507 to_cset->dfl_cgrp);
2508 put_css_set_locked(from_cset);
2509
2510 }
2511 }
2512 spin_unlock_irq(&css_set_lock);
2513
2514 /*
2515 * Migration is committed, all target tasks are now on dst_csets.
2516 * Nothing is sensitive to fork() after this point. Notify
2517 * controllers that migration is complete.
2518 */
2519 tset->csets = &tset->dst_csets;
2520
2521 if (tset->nr_tasks) {
2522 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2523 if (ss->attach) {
2524 tset->ssid = ssid;
2525 ss->attach(tset);
2526 }
2527 } while_each_subsys_mask();
2528 }
2529
2530 ret = 0;
2531 goto out_release_tset;
2532
2533out_cancel_attach:
2534 if (tset->nr_tasks) {
2535 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2536 if (ssid == failed_ssid)
2537 break;
2538 if (ss->cancel_attach) {
2539 tset->ssid = ssid;
2540 ss->cancel_attach(tset);
2541 }
2542 } while_each_subsys_mask();
2543 }
2544out_release_tset:
2545 spin_lock_irq(&css_set_lock);
2546 list_splice_init(&tset->dst_csets, &tset->src_csets);
2547 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2548 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2549 list_del_init(&cset->mg_node);
2550 }
2551 spin_unlock_irq(&css_set_lock);
2552
2553 /*
2554 * Re-initialize the cgroup_taskset structure in case it is reused
2555 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2556 * iteration.
2557 */
2558 tset->nr_tasks = 0;
2559 tset->csets = &tset->src_csets;
2560 return ret;
2561}
2562
2563/**
2564 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2565 * @dst_cgrp: destination cgroup to test
2566 *
2567 * On the default hierarchy, except for the mixable, (possible) thread root
2568 * and threaded cgroups, subtree_control must be zero for migration
2569 * destination cgroups with tasks so that child cgroups don't compete
2570 * against tasks.
2571 */
2572int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2573{
2574 /* v1 doesn't have any restriction */
2575 if (!cgroup_on_dfl(dst_cgrp))
2576 return 0;
2577
2578 /* verify @dst_cgrp can host resources */
2579 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2580 return -EOPNOTSUPP;
2581
2582 /* mixables don't care */
2583 if (cgroup_is_mixable(dst_cgrp))
2584 return 0;
2585
2586 /*
2587 * If @dst_cgrp is already or can become a thread root or is
2588 * threaded, it doesn't matter.
2589 */
2590 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2591 return 0;
2592
2593 /* apply no-internal-process constraint */
2594 if (dst_cgrp->subtree_control)
2595 return -EBUSY;
2596
2597 return 0;
2598}
2599
2600/**
2601 * cgroup_migrate_finish - cleanup after attach
2602 * @mgctx: migration context
2603 *
2604 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2605 * those functions for details.
2606 */
2607void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2608{
2609 LIST_HEAD(preloaded);
2610 struct css_set *cset, *tmp_cset;
2611
2612 lockdep_assert_held(&cgroup_mutex);
2613
2614 spin_lock_irq(&css_set_lock);
2615
2616 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2617 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2618
2619 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2620 cset->mg_src_cgrp = NULL;
2621 cset->mg_dst_cgrp = NULL;
2622 cset->mg_dst_cset = NULL;
2623 list_del_init(&cset->mg_preload_node);
2624 put_css_set_locked(cset);
2625 }
2626
2627 spin_unlock_irq(&css_set_lock);
2628}
2629
2630/**
2631 * cgroup_migrate_add_src - add a migration source css_set
2632 * @src_cset: the source css_set to add
2633 * @dst_cgrp: the destination cgroup
2634 * @mgctx: migration context
2635 *
2636 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2637 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2638 * up by cgroup_migrate_finish().
2639 *
2640 * This function may be called without holding cgroup_threadgroup_rwsem
2641 * even if the target is a process. Threads may be created and destroyed
2642 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2643 * into play and the preloaded css_sets are guaranteed to cover all
2644 * migrations.
2645 */
2646void cgroup_migrate_add_src(struct css_set *src_cset,
2647 struct cgroup *dst_cgrp,
2648 struct cgroup_mgctx *mgctx)
2649{
2650 struct cgroup *src_cgrp;
2651
2652 lockdep_assert_held(&cgroup_mutex);
2653 lockdep_assert_held(&css_set_lock);
2654
2655 /*
2656 * If ->dead, @src_set is associated with one or more dead cgroups
2657 * and doesn't contain any migratable tasks. Ignore it early so
2658 * that the rest of migration path doesn't get confused by it.
2659 */
2660 if (src_cset->dead)
2661 return;
2662
2663 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2664
2665 if (!list_empty(&src_cset->mg_preload_node))
2666 return;
2667
2668 WARN_ON(src_cset->mg_src_cgrp);
2669 WARN_ON(src_cset->mg_dst_cgrp);
2670 WARN_ON(!list_empty(&src_cset->mg_tasks));
2671 WARN_ON(!list_empty(&src_cset->mg_node));
2672
2673 src_cset->mg_src_cgrp = src_cgrp;
2674 src_cset->mg_dst_cgrp = dst_cgrp;
2675 get_css_set(src_cset);
2676 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2677}
2678
2679/**
2680 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2681 * @mgctx: migration context
2682 *
2683 * Tasks are about to be moved and all the source css_sets have been
2684 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2685 * pins all destination css_sets, links each to its source, and append them
2686 * to @mgctx->preloaded_dst_csets.
2687 *
2688 * This function must be called after cgroup_migrate_add_src() has been
2689 * called on each migration source css_set. After migration is performed
2690 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2691 * @mgctx.
2692 */
2693int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2694{
2695 struct css_set *src_cset, *tmp_cset;
2696
2697 lockdep_assert_held(&cgroup_mutex);
2698
2699 /* look up the dst cset for each src cset and link it to src */
2700 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2701 mg_preload_node) {
2702 struct css_set *dst_cset;
2703 struct cgroup_subsys *ss;
2704 int ssid;
2705
2706 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2707 if (!dst_cset)
2708 return -ENOMEM;
2709
2710 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2711
2712 /*
2713 * If src cset equals dst, it's noop. Drop the src.
2714 * cgroup_migrate() will skip the cset too. Note that we
2715 * can't handle src == dst as some nodes are used by both.
2716 */
2717 if (src_cset == dst_cset) {
2718 src_cset->mg_src_cgrp = NULL;
2719 src_cset->mg_dst_cgrp = NULL;
2720 list_del_init(&src_cset->mg_preload_node);
2721 put_css_set(src_cset);
2722 put_css_set(dst_cset);
2723 continue;
2724 }
2725
2726 src_cset->mg_dst_cset = dst_cset;
2727
2728 if (list_empty(&dst_cset->mg_preload_node))
2729 list_add_tail(&dst_cset->mg_preload_node,
2730 &mgctx->preloaded_dst_csets);
2731 else
2732 put_css_set(dst_cset);
2733
2734 for_each_subsys(ss, ssid)
2735 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2736 mgctx->ss_mask |= 1 << ssid;
2737 }
2738
2739 return 0;
2740}
2741
2742/**
2743 * cgroup_migrate - migrate a process or task to a cgroup
2744 * @leader: the leader of the process or the task to migrate
2745 * @threadgroup: whether @leader points to the whole process or a single task
2746 * @mgctx: migration context
2747 *
2748 * Migrate a process or task denoted by @leader. If migrating a process,
2749 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2750 * responsible for invoking cgroup_migrate_add_src() and
2751 * cgroup_migrate_prepare_dst() on the targets before invoking this
2752 * function and following up with cgroup_migrate_finish().
2753 *
2754 * As long as a controller's ->can_attach() doesn't fail, this function is
2755 * guaranteed to succeed. This means that, excluding ->can_attach()
2756 * failure, when migrating multiple targets, the success or failure can be
2757 * decided for all targets by invoking group_migrate_prepare_dst() before
2758 * actually starting migrating.
2759 */
2760int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2761 struct cgroup_mgctx *mgctx)
2762{
2763 struct task_struct *task;
2764
2765 /*
2766 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2767 * already PF_EXITING could be freed from underneath us unless we
2768 * take an rcu_read_lock.
2769 */
2770 spin_lock_irq(&css_set_lock);
2771 rcu_read_lock();
2772 task = leader;
2773 do {
2774 cgroup_migrate_add_task(task, mgctx);
2775 if (!threadgroup)
2776 break;
2777 } while_each_thread(leader, task);
2778 rcu_read_unlock();
2779 spin_unlock_irq(&css_set_lock);
2780
2781 return cgroup_migrate_execute(mgctx);
2782}
2783
2784/**
2785 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2786 * @dst_cgrp: the cgroup to attach to
2787 * @leader: the task or the leader of the threadgroup to be attached
2788 * @threadgroup: attach the whole threadgroup?
2789 *
2790 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2791 */
2792int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2793 bool threadgroup)
2794{
2795 DEFINE_CGROUP_MGCTX(mgctx);
2796 struct task_struct *task;
2797 int ret;
2798
2799 ret = cgroup_migrate_vet_dst(dst_cgrp);
2800 if (ret)
2801 return ret;
2802
2803 /* look up all src csets */
2804 spin_lock_irq(&css_set_lock);
2805 rcu_read_lock();
2806 task = leader;
2807 do {
2808 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2809 if (!threadgroup)
2810 break;
2811 } while_each_thread(leader, task);
2812 rcu_read_unlock();
2813 spin_unlock_irq(&css_set_lock);
2814
2815 /* prepare dst csets and commit */
2816 ret = cgroup_migrate_prepare_dst(&mgctx);
2817 if (!ret)
2818 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2819
2820 cgroup_migrate_finish(&mgctx);
2821
2822 if (!ret)
2823 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2824
2825 return ret;
2826}
2827
2828struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2829 __acquires(&cgroup_threadgroup_rwsem)
2830{
2831 struct task_struct *tsk;
2832 pid_t pid;
2833
2834 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2835 return ERR_PTR(-EINVAL);
2836
2837 percpu_down_write(&cgroup_threadgroup_rwsem);
2838
2839 rcu_read_lock();
2840 if (pid) {
2841 tsk = find_task_by_vpid(pid);
2842 if (!tsk) {
2843 tsk = ERR_PTR(-ESRCH);
2844 goto out_unlock_threadgroup;
2845 }
2846 } else {
2847 tsk = current;
2848 }
2849
2850 if (threadgroup)
2851 tsk = tsk->group_leader;
2852
2853 /*
2854 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2855 * If userland migrates such a kthread to a non-root cgroup, it can
2856 * become trapped in a cpuset, or RT kthread may be born in a
2857 * cgroup with no rt_runtime allocated. Just say no.
2858 */
2859 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2860 tsk = ERR_PTR(-EINVAL);
2861 goto out_unlock_threadgroup;
2862 }
2863
2864 get_task_struct(tsk);
2865 goto out_unlock_rcu;
2866
2867out_unlock_threadgroup:
2868 percpu_up_write(&cgroup_threadgroup_rwsem);
2869out_unlock_rcu:
2870 rcu_read_unlock();
2871 return tsk;
2872}
2873
2874void cgroup_procs_write_finish(struct task_struct *task)
2875 __releases(&cgroup_threadgroup_rwsem)
2876{
2877 struct cgroup_subsys *ss;
2878 int ssid;
2879
2880 /* release reference from cgroup_procs_write_start() */
2881 put_task_struct(task);
2882
2883 percpu_up_write(&cgroup_threadgroup_rwsem);
2884 for_each_subsys(ss, ssid)
2885 if (ss->post_attach)
2886 ss->post_attach();
2887}
2888
2889static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2890{
2891 struct cgroup_subsys *ss;
2892 bool printed = false;
2893 int ssid;
2894
2895 do_each_subsys_mask(ss, ssid, ss_mask) {
2896 if (printed)
2897 seq_putc(seq, ' ');
2898 seq_puts(seq, ss->name);
2899 printed = true;
2900 } while_each_subsys_mask();
2901 if (printed)
2902 seq_putc(seq, '\n');
2903}
2904
2905/* show controllers which are enabled from the parent */
2906static int cgroup_controllers_show(struct seq_file *seq, void *v)
2907{
2908 struct cgroup *cgrp = seq_css(seq)->cgroup;
2909
2910 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2911 return 0;
2912}
2913
2914/* show controllers which are enabled for a given cgroup's children */
2915static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2916{
2917 struct cgroup *cgrp = seq_css(seq)->cgroup;
2918
2919 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2920 return 0;
2921}
2922
2923/**
2924 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2925 * @cgrp: root of the subtree to update csses for
2926 *
2927 * @cgrp's control masks have changed and its subtree's css associations
2928 * need to be updated accordingly. This function looks up all css_sets
2929 * which are attached to the subtree, creates the matching updated css_sets
2930 * and migrates the tasks to the new ones.
2931 */
2932static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2933{
2934 DEFINE_CGROUP_MGCTX(mgctx);
2935 struct cgroup_subsys_state *d_css;
2936 struct cgroup *dsct;
2937 struct css_set *src_cset;
2938 int ret;
2939
2940 lockdep_assert_held(&cgroup_mutex);
2941
2942 percpu_down_write(&cgroup_threadgroup_rwsem);
2943
2944 /* look up all csses currently attached to @cgrp's subtree */
2945 spin_lock_irq(&css_set_lock);
2946 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2947 struct cgrp_cset_link *link;
2948
2949 list_for_each_entry(link, &dsct->cset_links, cset_link)
2950 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2951 }
2952 spin_unlock_irq(&css_set_lock);
2953
2954 /* NULL dst indicates self on default hierarchy */
2955 ret = cgroup_migrate_prepare_dst(&mgctx);
2956 if (ret)
2957 goto out_finish;
2958
2959 spin_lock_irq(&css_set_lock);
2960 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2961 struct task_struct *task, *ntask;
2962
2963 /* all tasks in src_csets need to be migrated */
2964 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2965 cgroup_migrate_add_task(task, &mgctx);
2966 }
2967 spin_unlock_irq(&css_set_lock);
2968
2969 ret = cgroup_migrate_execute(&mgctx);
2970out_finish:
2971 cgroup_migrate_finish(&mgctx);
2972 percpu_up_write(&cgroup_threadgroup_rwsem);
2973 return ret;
2974}
2975
2976/**
2977 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2978 * @cgrp: root of the target subtree
2979 *
2980 * Because css offlining is asynchronous, userland may try to re-enable a
2981 * controller while the previous css is still around. This function grabs
2982 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2983 */
2984void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2985 __acquires(&cgroup_mutex)
2986{
2987 struct cgroup *dsct;
2988 struct cgroup_subsys_state *d_css;
2989 struct cgroup_subsys *ss;
2990 int ssid;
2991
2992restart:
2993 mutex_lock(&cgroup_mutex);
2994
2995 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2996 for_each_subsys(ss, ssid) {
2997 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2998 DEFINE_WAIT(wait);
2999
3000 if (!css || !percpu_ref_is_dying(&css->refcnt))
3001 continue;
3002
3003 cgroup_get_live(dsct);
3004 prepare_to_wait(&dsct->offline_waitq, &wait,
3005 TASK_UNINTERRUPTIBLE);
3006
3007 mutex_unlock(&cgroup_mutex);
3008 schedule();
3009 finish_wait(&dsct->offline_waitq, &wait);
3010
3011 cgroup_put(dsct);
3012 goto restart;
3013 }
3014 }
3015}
3016
3017/**
3018 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3019 * @cgrp: root of the target subtree
3020 *
3021 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3022 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3023 * itself.
3024 */
3025static void cgroup_save_control(struct cgroup *cgrp)
3026{
3027 struct cgroup *dsct;
3028 struct cgroup_subsys_state *d_css;
3029
3030 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3031 dsct->old_subtree_control = dsct->subtree_control;
3032 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3033 dsct->old_dom_cgrp = dsct->dom_cgrp;
3034 }
3035}
3036
3037/**
3038 * cgroup_propagate_control - refresh control masks of a subtree
3039 * @cgrp: root of the target subtree
3040 *
3041 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3042 * ->subtree_control and propagate controller availability through the
3043 * subtree so that descendants don't have unavailable controllers enabled.
3044 */
3045static void cgroup_propagate_control(struct cgroup *cgrp)
3046{
3047 struct cgroup *dsct;
3048 struct cgroup_subsys_state *d_css;
3049
3050 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3051 dsct->subtree_control &= cgroup_control(dsct);
3052 dsct->subtree_ss_mask =
3053 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3054 cgroup_ss_mask(dsct));
3055 }
3056}
3057
3058/**
3059 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3060 * @cgrp: root of the target subtree
3061 *
3062 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3063 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3064 * itself.
3065 */
3066static void cgroup_restore_control(struct cgroup *cgrp)
3067{
3068 struct cgroup *dsct;
3069 struct cgroup_subsys_state *d_css;
3070
3071 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3072 dsct->subtree_control = dsct->old_subtree_control;
3073 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3074 dsct->dom_cgrp = dsct->old_dom_cgrp;
3075 }
3076}
3077
3078static bool css_visible(struct cgroup_subsys_state *css)
3079{
3080 struct cgroup_subsys *ss = css->ss;
3081 struct cgroup *cgrp = css->cgroup;
3082
3083 if (cgroup_control(cgrp) & (1 << ss->id))
3084 return true;
3085 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3086 return false;
3087 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3088}
3089
3090/**
3091 * cgroup_apply_control_enable - enable or show csses according to control
3092 * @cgrp: root of the target subtree
3093 *
3094 * Walk @cgrp's subtree and create new csses or make the existing ones
3095 * visible. A css is created invisible if it's being implicitly enabled
3096 * through dependency. An invisible css is made visible when the userland
3097 * explicitly enables it.
3098 *
3099 * Returns 0 on success, -errno on failure. On failure, csses which have
3100 * been processed already aren't cleaned up. The caller is responsible for
3101 * cleaning up with cgroup_apply_control_disable().
3102 */
3103static int cgroup_apply_control_enable(struct cgroup *cgrp)
3104{
3105 struct cgroup *dsct;
3106 struct cgroup_subsys_state *d_css;
3107 struct cgroup_subsys *ss;
3108 int ssid, ret;
3109
3110 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3111 for_each_subsys(ss, ssid) {
3112 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3113
3114 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3115
3116 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3117 continue;
3118
3119 if (!css) {
3120 css = css_create(dsct, ss);
3121 if (IS_ERR(css))
3122 return PTR_ERR(css);
3123 }
3124
3125 if (css_visible(css)) {
3126 ret = css_populate_dir(css);
3127 if (ret)
3128 return ret;
3129 }
3130 }
3131 }
3132
3133 return 0;
3134}
3135
3136/**
3137 * cgroup_apply_control_disable - kill or hide csses according to control
3138 * @cgrp: root of the target subtree
3139 *
3140 * Walk @cgrp's subtree and kill and hide csses so that they match
3141 * cgroup_ss_mask() and cgroup_visible_mask().
3142 *
3143 * A css is hidden when the userland requests it to be disabled while other
3144 * subsystems are still depending on it. The css must not actively control
3145 * resources and be in the vanilla state if it's made visible again later.
3146 * Controllers which may be depended upon should provide ->css_reset() for
3147 * this purpose.
3148 */
3149static void cgroup_apply_control_disable(struct cgroup *cgrp)
3150{
3151 struct cgroup *dsct;
3152 struct cgroup_subsys_state *d_css;
3153 struct cgroup_subsys *ss;
3154 int ssid;
3155
3156 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3157 for_each_subsys(ss, ssid) {
3158 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3159
3160 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3161
3162 if (!css)
3163 continue;
3164
3165 if (css->parent &&
3166 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3167 kill_css(css);
3168 } else if (!css_visible(css)) {
3169 css_clear_dir(css);
3170 if (ss->css_reset)
3171 ss->css_reset(css);
3172 }
3173 }
3174 }
3175}
3176
3177/**
3178 * cgroup_apply_control - apply control mask updates to the subtree
3179 * @cgrp: root of the target subtree
3180 *
3181 * subsystems can be enabled and disabled in a subtree using the following
3182 * steps.
3183 *
3184 * 1. Call cgroup_save_control() to stash the current state.
3185 * 2. Update ->subtree_control masks in the subtree as desired.
3186 * 3. Call cgroup_apply_control() to apply the changes.
3187 * 4. Optionally perform other related operations.
3188 * 5. Call cgroup_finalize_control() to finish up.
3189 *
3190 * This function implements step 3 and propagates the mask changes
3191 * throughout @cgrp's subtree, updates csses accordingly and perform
3192 * process migrations.
3193 */
3194static int cgroup_apply_control(struct cgroup *cgrp)
3195{
3196 int ret;
3197
3198 cgroup_propagate_control(cgrp);
3199
3200 ret = cgroup_apply_control_enable(cgrp);
3201 if (ret)
3202 return ret;
3203
3204 /*
3205 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3206 * making the following cgroup_update_dfl_csses() properly update
3207 * css associations of all tasks in the subtree.
3208 */
3209 ret = cgroup_update_dfl_csses(cgrp);
3210 if (ret)
3211 return ret;
3212
3213 return 0;
3214}
3215
3216/**
3217 * cgroup_finalize_control - finalize control mask update
3218 * @cgrp: root of the target subtree
3219 * @ret: the result of the update
3220 *
3221 * Finalize control mask update. See cgroup_apply_control() for more info.
3222 */
3223static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3224{
3225 if (ret) {
3226 cgroup_restore_control(cgrp);
3227 cgroup_propagate_control(cgrp);
3228 }
3229
3230 cgroup_apply_control_disable(cgrp);
3231}
3232
3233static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3234{
3235 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3236
3237 /* if nothing is getting enabled, nothing to worry about */
3238 if (!enable)
3239 return 0;
3240
3241 /* can @cgrp host any resources? */
3242 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3243 return -EOPNOTSUPP;
3244
3245 /* mixables don't care */
3246 if (cgroup_is_mixable(cgrp))
3247 return 0;
3248
3249 if (domain_enable) {
3250 /* can't enable domain controllers inside a thread subtree */
3251 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3252 return -EOPNOTSUPP;
3253 } else {
3254 /*
3255 * Threaded controllers can handle internal competitions
3256 * and are always allowed inside a (prospective) thread
3257 * subtree.
3258 */
3259 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3260 return 0;
3261 }
3262
3263 /*
3264 * Controllers can't be enabled for a cgroup with tasks to avoid
3265 * child cgroups competing against tasks.
3266 */
3267 if (cgroup_has_tasks(cgrp))
3268 return -EBUSY;
3269
3270 return 0;
3271}
3272
3273/* change the enabled child controllers for a cgroup in the default hierarchy */
3274static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3275 char *buf, size_t nbytes,
3276 loff_t off)
3277{
3278 u16 enable = 0, disable = 0;
3279 struct cgroup *cgrp, *child;
3280 struct cgroup_subsys *ss;
3281 char *tok;
3282 int ssid, ret;
3283
3284 /*
3285 * Parse input - space separated list of subsystem names prefixed
3286 * with either + or -.
3287 */
3288 buf = strstrip(buf);
3289 while ((tok = strsep(&buf, " "))) {
3290 if (tok[0] == '\0')
3291 continue;
3292 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3293 if (!cgroup_ssid_enabled(ssid) ||
3294 strcmp(tok + 1, ss->name))
3295 continue;
3296
3297 if (*tok == '+') {
3298 enable |= 1 << ssid;
3299 disable &= ~(1 << ssid);
3300 } else if (*tok == '-') {
3301 disable |= 1 << ssid;
3302 enable &= ~(1 << ssid);
3303 } else {
3304 return -EINVAL;
3305 }
3306 break;
3307 } while_each_subsys_mask();
3308 if (ssid == CGROUP_SUBSYS_COUNT)
3309 return -EINVAL;
3310 }
3311
3312 cgrp = cgroup_kn_lock_live(of->kn, true);
3313 if (!cgrp)
3314 return -ENODEV;
3315
3316 for_each_subsys(ss, ssid) {
3317 if (enable & (1 << ssid)) {
3318 if (cgrp->subtree_control & (1 << ssid)) {
3319 enable &= ~(1 << ssid);
3320 continue;
3321 }
3322
3323 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3324 ret = -ENOENT;
3325 goto out_unlock;
3326 }
3327 } else if (disable & (1 << ssid)) {
3328 if (!(cgrp->subtree_control & (1 << ssid))) {
3329 disable &= ~(1 << ssid);
3330 continue;
3331 }
3332
3333 /* a child has it enabled? */
3334 cgroup_for_each_live_child(child, cgrp) {
3335 if (child->subtree_control & (1 << ssid)) {
3336 ret = -EBUSY;
3337 goto out_unlock;
3338 }
3339 }
3340 }
3341 }
3342
3343 if (!enable && !disable) {
3344 ret = 0;
3345 goto out_unlock;
3346 }
3347
3348 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3349 if (ret)
3350 goto out_unlock;
3351
3352 /* save and update control masks and prepare csses */
3353 cgroup_save_control(cgrp);
3354
3355 cgrp->subtree_control |= enable;
3356 cgrp->subtree_control &= ~disable;
3357
3358 ret = cgroup_apply_control(cgrp);
3359 cgroup_finalize_control(cgrp, ret);
3360 if (ret)
3361 goto out_unlock;
3362
3363 kernfs_activate(cgrp->kn);
3364out_unlock:
3365 cgroup_kn_unlock(of->kn);
3366 return ret ?: nbytes;
3367}
3368
3369/**
3370 * cgroup_enable_threaded - make @cgrp threaded
3371 * @cgrp: the target cgroup
3372 *
3373 * Called when "threaded" is written to the cgroup.type interface file and
3374 * tries to make @cgrp threaded and join the parent's resource domain.
3375 * This function is never called on the root cgroup as cgroup.type doesn't
3376 * exist on it.
3377 */
3378static int cgroup_enable_threaded(struct cgroup *cgrp)
3379{
3380 struct cgroup *parent = cgroup_parent(cgrp);
3381 struct cgroup *dom_cgrp = parent->dom_cgrp;
3382 struct cgroup *dsct;
3383 struct cgroup_subsys_state *d_css;
3384 int ret;
3385
3386 lockdep_assert_held(&cgroup_mutex);
3387
3388 /* noop if already threaded */
3389 if (cgroup_is_threaded(cgrp))
3390 return 0;
3391
3392 /*
3393 * If @cgroup is populated or has domain controllers enabled, it
3394 * can't be switched. While the below cgroup_can_be_thread_root()
3395 * test can catch the same conditions, that's only when @parent is
3396 * not mixable, so let's check it explicitly.
3397 */
3398 if (cgroup_is_populated(cgrp) ||
3399 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3400 return -EOPNOTSUPP;
3401
3402 /* we're joining the parent's domain, ensure its validity */
3403 if (!cgroup_is_valid_domain(dom_cgrp) ||
3404 !cgroup_can_be_thread_root(dom_cgrp))
3405 return -EOPNOTSUPP;
3406
3407 /*
3408 * The following shouldn't cause actual migrations and should
3409 * always succeed.
3410 */
3411 cgroup_save_control(cgrp);
3412
3413 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3414 if (dsct == cgrp || cgroup_is_threaded(dsct))
3415 dsct->dom_cgrp = dom_cgrp;
3416
3417 ret = cgroup_apply_control(cgrp);
3418 if (!ret)
3419 parent->nr_threaded_children++;
3420
3421 cgroup_finalize_control(cgrp, ret);
3422 return ret;
3423}
3424
3425static int cgroup_type_show(struct seq_file *seq, void *v)
3426{
3427 struct cgroup *cgrp = seq_css(seq)->cgroup;
3428
3429 if (cgroup_is_threaded(cgrp))
3430 seq_puts(seq, "threaded\n");
3431 else if (!cgroup_is_valid_domain(cgrp))
3432 seq_puts(seq, "domain invalid\n");
3433 else if (cgroup_is_thread_root(cgrp))
3434 seq_puts(seq, "domain threaded\n");
3435 else
3436 seq_puts(seq, "domain\n");
3437
3438 return 0;
3439}
3440
3441static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3442 size_t nbytes, loff_t off)
3443{
3444 struct cgroup *cgrp;
3445 int ret;
3446
3447 /* only switching to threaded mode is supported */
3448 if (strcmp(strstrip(buf), "threaded"))
3449 return -EINVAL;
3450
3451 cgrp = cgroup_kn_lock_live(of->kn, false);
3452 if (!cgrp)
3453 return -ENOENT;
3454
3455 /* threaded can only be enabled */
3456 ret = cgroup_enable_threaded(cgrp);
3457
3458 cgroup_kn_unlock(of->kn);
3459 return ret ?: nbytes;
3460}
3461
3462static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3463{
3464 struct cgroup *cgrp = seq_css(seq)->cgroup;
3465 int descendants = READ_ONCE(cgrp->max_descendants);
3466
3467 if (descendants == INT_MAX)
3468 seq_puts(seq, "max\n");
3469 else
3470 seq_printf(seq, "%d\n", descendants);
3471
3472 return 0;
3473}
3474
3475static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3476 char *buf, size_t nbytes, loff_t off)
3477{
3478 struct cgroup *cgrp;
3479 int descendants;
3480 ssize_t ret;
3481
3482 buf = strstrip(buf);
3483 if (!strcmp(buf, "max")) {
3484 descendants = INT_MAX;
3485 } else {
3486 ret = kstrtoint(buf, 0, &descendants);
3487 if (ret)
3488 return ret;
3489 }
3490
3491 if (descendants < 0)
3492 return -ERANGE;
3493
3494 cgrp = cgroup_kn_lock_live(of->kn, false);
3495 if (!cgrp)
3496 return -ENOENT;
3497
3498 cgrp->max_descendants = descendants;
3499
3500 cgroup_kn_unlock(of->kn);
3501
3502 return nbytes;
3503}
3504
3505static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3506{
3507 struct cgroup *cgrp = seq_css(seq)->cgroup;
3508 int depth = READ_ONCE(cgrp->max_depth);
3509
3510 if (depth == INT_MAX)
3511 seq_puts(seq, "max\n");
3512 else
3513 seq_printf(seq, "%d\n", depth);
3514
3515 return 0;
3516}
3517
3518static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3519 char *buf, size_t nbytes, loff_t off)
3520{
3521 struct cgroup *cgrp;
3522 ssize_t ret;
3523 int depth;
3524
3525 buf = strstrip(buf);
3526 if (!strcmp(buf, "max")) {
3527 depth = INT_MAX;
3528 } else {
3529 ret = kstrtoint(buf, 0, &depth);
3530 if (ret)
3531 return ret;
3532 }
3533
3534 if (depth < 0)
3535 return -ERANGE;
3536
3537 cgrp = cgroup_kn_lock_live(of->kn, false);
3538 if (!cgrp)
3539 return -ENOENT;
3540
3541 cgrp->max_depth = depth;
3542
3543 cgroup_kn_unlock(of->kn);
3544
3545 return nbytes;
3546}
3547
3548static int cgroup_events_show(struct seq_file *seq, void *v)
3549{
3550 struct cgroup *cgrp = seq_css(seq)->cgroup;
3551
3552 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3553 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3554
3555 return 0;
3556}
3557
3558static int cgroup_stat_show(struct seq_file *seq, void *v)
3559{
3560 struct cgroup *cgroup = seq_css(seq)->cgroup;
3561
3562 seq_printf(seq, "nr_descendants %d\n",
3563 cgroup->nr_descendants);
3564 seq_printf(seq, "nr_dying_descendants %d\n",
3565 cgroup->nr_dying_descendants);
3566
3567 return 0;
3568}
3569
3570static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3571 struct cgroup *cgrp, int ssid)
3572{
3573 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3574 struct cgroup_subsys_state *css;
3575 int ret;
3576
3577 if (!ss->css_extra_stat_show)
3578 return 0;
3579
3580 css = cgroup_tryget_css(cgrp, ss);
3581 if (!css)
3582 return 0;
3583
3584 ret = ss->css_extra_stat_show(seq, css);
3585 css_put(css);
3586 return ret;
3587}
3588
3589static int cpu_stat_show(struct seq_file *seq, void *v)
3590{
3591 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3592 int ret = 0;
3593
3594 cgroup_base_stat_cputime_show(seq);
3595#ifdef CONFIG_CGROUP_SCHED
3596 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3597#endif
3598 return ret;
3599}
3600
3601#ifdef CONFIG_PSI
3602static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3603{
3604 struct cgroup *cgroup = seq_css(seq)->cgroup;
3605 struct psi_group *psi = cgroup->id == 1 ? &psi_system : &cgroup->psi;
3606
3607 return psi_show(seq, psi, PSI_IO);
3608}
3609static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3610{
3611 struct cgroup *cgroup = seq_css(seq)->cgroup;
3612 struct psi_group *psi = cgroup->id == 1 ? &psi_system : &cgroup->psi;
3613
3614 return psi_show(seq, psi, PSI_MEM);
3615}
3616static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3617{
3618 struct cgroup *cgroup = seq_css(seq)->cgroup;
3619 struct psi_group *psi = cgroup->id == 1 ? &psi_system : &cgroup->psi;
3620
3621 return psi_show(seq, psi, PSI_CPU);
3622}
3623
3624static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3625 size_t nbytes, enum psi_res res)
3626{
3627 struct psi_trigger *new;
3628 struct cgroup *cgrp;
3629
3630 cgrp = cgroup_kn_lock_live(of->kn, false);
3631 if (!cgrp)
3632 return -ENODEV;
3633
3634 cgroup_get(cgrp);
3635 cgroup_kn_unlock(of->kn);
3636
3637 new = psi_trigger_create(&cgrp->psi, buf, nbytes, res);
3638 if (IS_ERR(new)) {
3639 cgroup_put(cgrp);
3640 return PTR_ERR(new);
3641 }
3642
3643 psi_trigger_replace(&of->priv, new);
3644
3645 cgroup_put(cgrp);
3646
3647 return nbytes;
3648}
3649
3650static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3651 char *buf, size_t nbytes,
3652 loff_t off)
3653{
3654 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3655}
3656
3657static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3658 char *buf, size_t nbytes,
3659 loff_t off)
3660{
3661 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3662}
3663
3664static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3665 char *buf, size_t nbytes,
3666 loff_t off)
3667{
3668 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3669}
3670
3671static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3672 poll_table *pt)
3673{
3674 return psi_trigger_poll(&of->priv, of->file, pt);
3675}
3676
3677static void cgroup_pressure_release(struct kernfs_open_file *of)
3678{
3679 psi_trigger_replace(&of->priv, NULL);
3680}
3681#endif /* CONFIG_PSI */
3682
3683static int cgroup_freeze_show(struct seq_file *seq, void *v)
3684{
3685 struct cgroup *cgrp = seq_css(seq)->cgroup;
3686
3687 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3688
3689 return 0;
3690}
3691
3692static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3693 char *buf, size_t nbytes, loff_t off)
3694{
3695 struct cgroup *cgrp;
3696 ssize_t ret;
3697 int freeze;
3698
3699 ret = kstrtoint(strstrip(buf), 0, &freeze);
3700 if (ret)
3701 return ret;
3702
3703 if (freeze < 0 || freeze > 1)
3704 return -ERANGE;
3705
3706 cgrp = cgroup_kn_lock_live(of->kn, false);
3707 if (!cgrp)
3708 return -ENOENT;
3709
3710 cgroup_freeze(cgrp, freeze);
3711
3712 cgroup_kn_unlock(of->kn);
3713
3714 return nbytes;
3715}
3716
3717static int cgroup_file_open(struct kernfs_open_file *of)
3718{
3719 struct cftype *cft = of->kn->priv;
3720
3721 if (cft->open)
3722 return cft->open(of);
3723 return 0;
3724}
3725
3726static void cgroup_file_release(struct kernfs_open_file *of)
3727{
3728 struct cftype *cft = of->kn->priv;
3729
3730 if (cft->release)
3731 cft->release(of);
3732}
3733
3734static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3735 size_t nbytes, loff_t off)
3736{
3737 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3738 struct cgroup *cgrp = of->kn->parent->priv;
3739 struct cftype *cft = of->kn->priv;
3740 struct cgroup_subsys_state *css;
3741 int ret;
3742
3743 /*
3744 * If namespaces are delegation boundaries, disallow writes to
3745 * files in an non-init namespace root from inside the namespace
3746 * except for the files explicitly marked delegatable -
3747 * cgroup.procs and cgroup.subtree_control.
3748 */
3749 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3750 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3751 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3752 return -EPERM;
3753
3754 if (cft->write)
3755 return cft->write(of, buf, nbytes, off);
3756
3757 /*
3758 * kernfs guarantees that a file isn't deleted with operations in
3759 * flight, which means that the matching css is and stays alive and
3760 * doesn't need to be pinned. The RCU locking is not necessary
3761 * either. It's just for the convenience of using cgroup_css().
3762 */
3763 rcu_read_lock();
3764 css = cgroup_css(cgrp, cft->ss);
3765 rcu_read_unlock();
3766
3767 if (cft->write_u64) {
3768 unsigned long long v;
3769 ret = kstrtoull(buf, 0, &v);
3770 if (!ret)
3771 ret = cft->write_u64(css, cft, v);
3772 } else if (cft->write_s64) {
3773 long long v;
3774 ret = kstrtoll(buf, 0, &v);
3775 if (!ret)
3776 ret = cft->write_s64(css, cft, v);
3777 } else {
3778 ret = -EINVAL;
3779 }
3780
3781 return ret ?: nbytes;
3782}
3783
3784static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3785{
3786 struct cftype *cft = of->kn->priv;
3787
3788 if (cft->poll)
3789 return cft->poll(of, pt);
3790
3791 return kernfs_generic_poll(of, pt);
3792}
3793
3794static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3795{
3796 return seq_cft(seq)->seq_start(seq, ppos);
3797}
3798
3799static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3800{
3801 return seq_cft(seq)->seq_next(seq, v, ppos);
3802}
3803
3804static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3805{
3806 if (seq_cft(seq)->seq_stop)
3807 seq_cft(seq)->seq_stop(seq, v);
3808}
3809
3810static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3811{
3812 struct cftype *cft = seq_cft(m);
3813 struct cgroup_subsys_state *css = seq_css(m);
3814
3815 if (cft->seq_show)
3816 return cft->seq_show(m, arg);
3817
3818 if (cft->read_u64)
3819 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3820 else if (cft->read_s64)
3821 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3822 else
3823 return -EINVAL;
3824 return 0;
3825}
3826
3827static struct kernfs_ops cgroup_kf_single_ops = {
3828 .atomic_write_len = PAGE_SIZE,
3829 .open = cgroup_file_open,
3830 .release = cgroup_file_release,
3831 .write = cgroup_file_write,
3832 .poll = cgroup_file_poll,
3833 .seq_show = cgroup_seqfile_show,
3834};
3835
3836static struct kernfs_ops cgroup_kf_ops = {
3837 .atomic_write_len = PAGE_SIZE,
3838 .open = cgroup_file_open,
3839 .release = cgroup_file_release,
3840 .write = cgroup_file_write,
3841 .poll = cgroup_file_poll,
3842 .seq_start = cgroup_seqfile_start,
3843 .seq_next = cgroup_seqfile_next,
3844 .seq_stop = cgroup_seqfile_stop,
3845 .seq_show = cgroup_seqfile_show,
3846};
3847
3848/* set uid and gid of cgroup dirs and files to that of the creator */
3849static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3850{
3851 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3852 .ia_uid = current_fsuid(),
3853 .ia_gid = current_fsgid(), };
3854
3855 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3856 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3857 return 0;
3858
3859 return kernfs_setattr(kn, &iattr);
3860}
3861
3862static void cgroup_file_notify_timer(struct timer_list *timer)
3863{
3864 cgroup_file_notify(container_of(timer, struct cgroup_file,
3865 notify_timer));
3866}
3867
3868static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3869 struct cftype *cft)
3870{
3871 char name[CGROUP_FILE_NAME_MAX];
3872 struct kernfs_node *kn;
3873 struct lock_class_key *key = NULL;
3874 int ret;
3875
3876#ifdef CONFIG_DEBUG_LOCK_ALLOC
3877 key = &cft->lockdep_key;
3878#endif
3879 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3880 cgroup_file_mode(cft),
3881 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3882 0, cft->kf_ops, cft,
3883 NULL, key);
3884 if (IS_ERR(kn))
3885 return PTR_ERR(kn);
3886
3887 ret = cgroup_kn_set_ugid(kn);
3888 if (ret) {
3889 kernfs_remove(kn);
3890 return ret;
3891 }
3892
3893 if (cft->file_offset) {
3894 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3895
3896 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3897
3898 spin_lock_irq(&cgroup_file_kn_lock);
3899 cfile->kn = kn;
3900 spin_unlock_irq(&cgroup_file_kn_lock);
3901 }
3902
3903 return 0;
3904}
3905
3906/**
3907 * cgroup_addrm_files - add or remove files to a cgroup directory
3908 * @css: the target css
3909 * @cgrp: the target cgroup (usually css->cgroup)
3910 * @cfts: array of cftypes to be added
3911 * @is_add: whether to add or remove
3912 *
3913 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3914 * For removals, this function never fails.
3915 */
3916static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3917 struct cgroup *cgrp, struct cftype cfts[],
3918 bool is_add)
3919{
3920 struct cftype *cft, *cft_end = NULL;
3921 int ret = 0;
3922
3923 lockdep_assert_held(&cgroup_mutex);
3924
3925restart:
3926 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3927 /* does cft->flags tell us to skip this file on @cgrp? */
3928 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3929 continue;
3930 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3931 continue;
3932 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3933 continue;
3934 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3935 continue;
3936 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3937 continue;
3938 if (is_add) {
3939 ret = cgroup_add_file(css, cgrp, cft);
3940 if (ret) {
3941 pr_warn("%s: failed to add %s, err=%d\n",
3942 __func__, cft->name, ret);
3943 cft_end = cft;
3944 is_add = false;
3945 goto restart;
3946 }
3947 } else {
3948 cgroup_rm_file(cgrp, cft);
3949 }
3950 }
3951 return ret;
3952}
3953
3954static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3955{
3956 struct cgroup_subsys *ss = cfts[0].ss;
3957 struct cgroup *root = &ss->root->cgrp;
3958 struct cgroup_subsys_state *css;
3959 int ret = 0;
3960
3961 lockdep_assert_held(&cgroup_mutex);
3962
3963 /* add/rm files for all cgroups created before */
3964 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3965 struct cgroup *cgrp = css->cgroup;
3966
3967 if (!(css->flags & CSS_VISIBLE))
3968 continue;
3969
3970 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3971 if (ret)
3972 break;
3973 }
3974
3975 if (is_add && !ret)
3976 kernfs_activate(root->kn);
3977 return ret;
3978}
3979
3980static void cgroup_exit_cftypes(struct cftype *cfts)
3981{
3982 struct cftype *cft;
3983
3984 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3985 /* free copy for custom atomic_write_len, see init_cftypes() */
3986 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3987 kfree(cft->kf_ops);
3988 cft->kf_ops = NULL;
3989 cft->ss = NULL;
3990
3991 /* revert flags set by cgroup core while adding @cfts */
3992 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3993 }
3994}
3995
3996static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3997{
3998 struct cftype *cft;
3999
4000 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4001 struct kernfs_ops *kf_ops;
4002
4003 WARN_ON(cft->ss || cft->kf_ops);
4004
4005 if (cft->seq_start)
4006 kf_ops = &cgroup_kf_ops;
4007 else
4008 kf_ops = &cgroup_kf_single_ops;
4009
4010 /*
4011 * Ugh... if @cft wants a custom max_write_len, we need to
4012 * make a copy of kf_ops to set its atomic_write_len.
4013 */
4014 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4015 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4016 if (!kf_ops) {
4017 cgroup_exit_cftypes(cfts);
4018 return -ENOMEM;
4019 }
4020 kf_ops->atomic_write_len = cft->max_write_len;
4021 }
4022
4023 cft->kf_ops = kf_ops;
4024 cft->ss = ss;
4025 }
4026
4027 return 0;
4028}
4029
4030static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4031{
4032 lockdep_assert_held(&cgroup_mutex);
4033
4034 if (!cfts || !cfts[0].ss)
4035 return -ENOENT;
4036
4037 list_del(&cfts->node);
4038 cgroup_apply_cftypes(cfts, false);
4039 cgroup_exit_cftypes(cfts);
4040 return 0;
4041}
4042
4043/**
4044 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4045 * @cfts: zero-length name terminated array of cftypes
4046 *
4047 * Unregister @cfts. Files described by @cfts are removed from all
4048 * existing cgroups and all future cgroups won't have them either. This
4049 * function can be called anytime whether @cfts' subsys is attached or not.
4050 *
4051 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4052 * registered.
4053 */
4054int cgroup_rm_cftypes(struct cftype *cfts)
4055{
4056 int ret;
4057
4058 mutex_lock(&cgroup_mutex);
4059 ret = cgroup_rm_cftypes_locked(cfts);
4060 mutex_unlock(&cgroup_mutex);
4061 return ret;
4062}
4063
4064/**
4065 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4066 * @ss: target cgroup subsystem
4067 * @cfts: zero-length name terminated array of cftypes
4068 *
4069 * Register @cfts to @ss. Files described by @cfts are created for all
4070 * existing cgroups to which @ss is attached and all future cgroups will
4071 * have them too. This function can be called anytime whether @ss is
4072 * attached or not.
4073 *
4074 * Returns 0 on successful registration, -errno on failure. Note that this
4075 * function currently returns 0 as long as @cfts registration is successful
4076 * even if some file creation attempts on existing cgroups fail.
4077 */
4078static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4079{
4080 int ret;
4081
4082 if (!cgroup_ssid_enabled(ss->id))
4083 return 0;
4084
4085 if (!cfts || cfts[0].name[0] == '\0')
4086 return 0;
4087
4088 ret = cgroup_init_cftypes(ss, cfts);
4089 if (ret)
4090 return ret;
4091
4092 mutex_lock(&cgroup_mutex);
4093
4094 list_add_tail(&cfts->node, &ss->cfts);
4095 ret = cgroup_apply_cftypes(cfts, true);
4096 if (ret)
4097 cgroup_rm_cftypes_locked(cfts);
4098
4099 mutex_unlock(&cgroup_mutex);
4100 return ret;
4101}
4102
4103/**
4104 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4105 * @ss: target cgroup subsystem
4106 * @cfts: zero-length name terminated array of cftypes
4107 *
4108 * Similar to cgroup_add_cftypes() but the added files are only used for
4109 * the default hierarchy.
4110 */
4111int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4112{
4113 struct cftype *cft;
4114
4115 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4116 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4117 return cgroup_add_cftypes(ss, cfts);
4118}
4119
4120/**
4121 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4122 * @ss: target cgroup subsystem
4123 * @cfts: zero-length name terminated array of cftypes
4124 *
4125 * Similar to cgroup_add_cftypes() but the added files are only used for
4126 * the legacy hierarchies.
4127 */
4128int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4129{
4130 struct cftype *cft;
4131
4132 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4133 cft->flags |= __CFTYPE_NOT_ON_DFL;
4134 return cgroup_add_cftypes(ss, cfts);
4135}
4136
4137/**
4138 * cgroup_file_notify - generate a file modified event for a cgroup_file
4139 * @cfile: target cgroup_file
4140 *
4141 * @cfile must have been obtained by setting cftype->file_offset.
4142 */
4143void cgroup_file_notify(struct cgroup_file *cfile)
4144{
4145 unsigned long flags;
4146
4147 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4148 if (cfile->kn) {
4149 unsigned long last = cfile->notified_at;
4150 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4151
4152 if (time_in_range(jiffies, last, next)) {
4153 timer_reduce(&cfile->notify_timer, next);
4154 } else {
4155 kernfs_notify(cfile->kn);
4156 cfile->notified_at = jiffies;
4157 }
4158 }
4159 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4160}
4161
4162/**
4163 * css_next_child - find the next child of a given css
4164 * @pos: the current position (%NULL to initiate traversal)
4165 * @parent: css whose children to walk
4166 *
4167 * This function returns the next child of @parent and should be called
4168 * under either cgroup_mutex or RCU read lock. The only requirement is
4169 * that @parent and @pos are accessible. The next sibling is guaranteed to
4170 * be returned regardless of their states.
4171 *
4172 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4173 * css which finished ->css_online() is guaranteed to be visible in the
4174 * future iterations and will stay visible until the last reference is put.
4175 * A css which hasn't finished ->css_online() or already finished
4176 * ->css_offline() may show up during traversal. It's each subsystem's
4177 * responsibility to synchronize against on/offlining.
4178 */
4179struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4180 struct cgroup_subsys_state *parent)
4181{
4182 struct cgroup_subsys_state *next;
4183
4184 cgroup_assert_mutex_or_rcu_locked();
4185
4186 /*
4187 * @pos could already have been unlinked from the sibling list.
4188 * Once a cgroup is removed, its ->sibling.next is no longer
4189 * updated when its next sibling changes. CSS_RELEASED is set when
4190 * @pos is taken off list, at which time its next pointer is valid,
4191 * and, as releases are serialized, the one pointed to by the next
4192 * pointer is guaranteed to not have started release yet. This
4193 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4194 * critical section, the one pointed to by its next pointer is
4195 * guaranteed to not have finished its RCU grace period even if we
4196 * have dropped rcu_read_lock() inbetween iterations.
4197 *
4198 * If @pos has CSS_RELEASED set, its next pointer can't be
4199 * dereferenced; however, as each css is given a monotonically
4200 * increasing unique serial number and always appended to the
4201 * sibling list, the next one can be found by walking the parent's
4202 * children until the first css with higher serial number than
4203 * @pos's. While this path can be slower, it happens iff iteration
4204 * races against release and the race window is very small.
4205 */
4206 if (!pos) {
4207 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4208 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4209 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4210 } else {
4211 list_for_each_entry_rcu(next, &parent->children, sibling)
4212 if (next->serial_nr > pos->serial_nr)
4213 break;
4214 }
4215
4216 /*
4217 * @next, if not pointing to the head, can be dereferenced and is
4218 * the next sibling.
4219 */
4220 if (&next->sibling != &parent->children)
4221 return next;
4222 return NULL;
4223}
4224
4225/**
4226 * css_next_descendant_pre - find the next descendant for pre-order walk
4227 * @pos: the current position (%NULL to initiate traversal)
4228 * @root: css whose descendants to walk
4229 *
4230 * To be used by css_for_each_descendant_pre(). Find the next descendant
4231 * to visit for pre-order traversal of @root's descendants. @root is
4232 * included in the iteration and the first node to be visited.
4233 *
4234 * While this function requires cgroup_mutex or RCU read locking, it
4235 * doesn't require the whole traversal to be contained in a single critical
4236 * section. This function will return the correct next descendant as long
4237 * as both @pos and @root are accessible and @pos is a descendant of @root.
4238 *
4239 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4240 * css which finished ->css_online() is guaranteed to be visible in the
4241 * future iterations and will stay visible until the last reference is put.
4242 * A css which hasn't finished ->css_online() or already finished
4243 * ->css_offline() may show up during traversal. It's each subsystem's
4244 * responsibility to synchronize against on/offlining.
4245 */
4246struct cgroup_subsys_state *
4247css_next_descendant_pre(struct cgroup_subsys_state *pos,
4248 struct cgroup_subsys_state *root)
4249{
4250 struct cgroup_subsys_state *next;
4251
4252 cgroup_assert_mutex_or_rcu_locked();
4253
4254 /* if first iteration, visit @root */
4255 if (!pos)
4256 return root;
4257
4258 /* visit the first child if exists */
4259 next = css_next_child(NULL, pos);
4260 if (next)
4261 return next;
4262
4263 /* no child, visit my or the closest ancestor's next sibling */
4264 while (pos != root) {
4265 next = css_next_child(pos, pos->parent);
4266 if (next)
4267 return next;
4268 pos = pos->parent;
4269 }
4270
4271 return NULL;
4272}
4273EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4274
4275/**
4276 * css_rightmost_descendant - return the rightmost descendant of a css
4277 * @pos: css of interest
4278 *
4279 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4280 * is returned. This can be used during pre-order traversal to skip
4281 * subtree of @pos.
4282 *
4283 * While this function requires cgroup_mutex or RCU read locking, it
4284 * doesn't require the whole traversal to be contained in a single critical
4285 * section. This function will return the correct rightmost descendant as
4286 * long as @pos is accessible.
4287 */
4288struct cgroup_subsys_state *
4289css_rightmost_descendant(struct cgroup_subsys_state *pos)
4290{
4291 struct cgroup_subsys_state *last, *tmp;
4292
4293 cgroup_assert_mutex_or_rcu_locked();
4294
4295 do {
4296 last = pos;
4297 /* ->prev isn't RCU safe, walk ->next till the end */
4298 pos = NULL;
4299 css_for_each_child(tmp, last)
4300 pos = tmp;
4301 } while (pos);
4302
4303 return last;
4304}
4305
4306static struct cgroup_subsys_state *
4307css_leftmost_descendant(struct cgroup_subsys_state *pos)
4308{
4309 struct cgroup_subsys_state *last;
4310
4311 do {
4312 last = pos;
4313 pos = css_next_child(NULL, pos);
4314 } while (pos);
4315
4316 return last;
4317}
4318
4319/**
4320 * css_next_descendant_post - find the next descendant for post-order walk
4321 * @pos: the current position (%NULL to initiate traversal)
4322 * @root: css whose descendants to walk
4323 *
4324 * To be used by css_for_each_descendant_post(). Find the next descendant
4325 * to visit for post-order traversal of @root's descendants. @root is
4326 * included in the iteration and the last node to be visited.
4327 *
4328 * While this function requires cgroup_mutex or RCU read locking, it
4329 * doesn't require the whole traversal to be contained in a single critical
4330 * section. This function will return the correct next descendant as long
4331 * as both @pos and @cgroup are accessible and @pos is a descendant of
4332 * @cgroup.
4333 *
4334 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4335 * css which finished ->css_online() is guaranteed to be visible in the
4336 * future iterations and will stay visible until the last reference is put.
4337 * A css which hasn't finished ->css_online() or already finished
4338 * ->css_offline() may show up during traversal. It's each subsystem's
4339 * responsibility to synchronize against on/offlining.
4340 */
4341struct cgroup_subsys_state *
4342css_next_descendant_post(struct cgroup_subsys_state *pos,
4343 struct cgroup_subsys_state *root)
4344{
4345 struct cgroup_subsys_state *next;
4346
4347 cgroup_assert_mutex_or_rcu_locked();
4348
4349 /* if first iteration, visit leftmost descendant which may be @root */
4350 if (!pos)
4351 return css_leftmost_descendant(root);
4352
4353 /* if we visited @root, we're done */
4354 if (pos == root)
4355 return NULL;
4356
4357 /* if there's an unvisited sibling, visit its leftmost descendant */
4358 next = css_next_child(pos, pos->parent);
4359 if (next)
4360 return css_leftmost_descendant(next);
4361
4362 /* no sibling left, visit parent */
4363 return pos->parent;
4364}
4365
4366/**
4367 * css_has_online_children - does a css have online children
4368 * @css: the target css
4369 *
4370 * Returns %true if @css has any online children; otherwise, %false. This
4371 * function can be called from any context but the caller is responsible
4372 * for synchronizing against on/offlining as necessary.
4373 */
4374bool css_has_online_children(struct cgroup_subsys_state *css)
4375{
4376 struct cgroup_subsys_state *child;
4377 bool ret = false;
4378
4379 rcu_read_lock();
4380 css_for_each_child(child, css) {
4381 if (child->flags & CSS_ONLINE) {
4382 ret = true;
4383 break;
4384 }
4385 }
4386 rcu_read_unlock();
4387 return ret;
4388}
4389
4390static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4391{
4392 struct list_head *l;
4393 struct cgrp_cset_link *link;
4394 struct css_set *cset;
4395
4396 lockdep_assert_held(&css_set_lock);
4397
4398 /* find the next threaded cset */
4399 if (it->tcset_pos) {
4400 l = it->tcset_pos->next;
4401
4402 if (l != it->tcset_head) {
4403 it->tcset_pos = l;
4404 return container_of(l, struct css_set,
4405 threaded_csets_node);
4406 }
4407
4408 it->tcset_pos = NULL;
4409 }
4410
4411 /* find the next cset */
4412 l = it->cset_pos;
4413 l = l->next;
4414 if (l == it->cset_head) {
4415 it->cset_pos = NULL;
4416 return NULL;
4417 }
4418
4419 if (it->ss) {
4420 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4421 } else {
4422 link = list_entry(l, struct cgrp_cset_link, cset_link);
4423 cset = link->cset;
4424 }
4425
4426 it->cset_pos = l;
4427
4428 /* initialize threaded css_set walking */
4429 if (it->flags & CSS_TASK_ITER_THREADED) {
4430 if (it->cur_dcset)
4431 put_css_set_locked(it->cur_dcset);
4432 it->cur_dcset = cset;
4433 get_css_set(cset);
4434
4435 it->tcset_head = &cset->threaded_csets;
4436 it->tcset_pos = &cset->threaded_csets;
4437 }
4438
4439 return cset;
4440}
4441
4442/**
4443 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4444 * @it: the iterator to advance
4445 *
4446 * Advance @it to the next css_set to walk.
4447 */
4448static void css_task_iter_advance_css_set(struct css_task_iter *it)
4449{
4450 struct css_set *cset;
4451
4452 lockdep_assert_held(&css_set_lock);
4453
4454 /* Advance to the next non-empty css_set */
4455 do {
4456 cset = css_task_iter_next_css_set(it);
4457 if (!cset) {
4458 it->task_pos = NULL;
4459 return;
4460 }
4461 } while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
4462
4463 if (!list_empty(&cset->tasks))
4464 it->task_pos = cset->tasks.next;
4465 else if (!list_empty(&cset->mg_tasks))
4466 it->task_pos = cset->mg_tasks.next;
4467 else
4468 it->task_pos = cset->dying_tasks.next;
4469
4470 it->tasks_head = &cset->tasks;
4471 it->mg_tasks_head = &cset->mg_tasks;
4472 it->dying_tasks_head = &cset->dying_tasks;
4473
4474 /*
4475 * We don't keep css_sets locked across iteration steps and thus
4476 * need to take steps to ensure that iteration can be resumed after
4477 * the lock is re-acquired. Iteration is performed at two levels -
4478 * css_sets and tasks in them.
4479 *
4480 * Once created, a css_set never leaves its cgroup lists, so a
4481 * pinned css_set is guaranteed to stay put and we can resume
4482 * iteration afterwards.
4483 *
4484 * Tasks may leave @cset across iteration steps. This is resolved
4485 * by registering each iterator with the css_set currently being
4486 * walked and making css_set_move_task() advance iterators whose
4487 * next task is leaving.
4488 */
4489 if (it->cur_cset) {
4490 list_del(&it->iters_node);
4491 put_css_set_locked(it->cur_cset);
4492 }
4493 get_css_set(cset);
4494 it->cur_cset = cset;
4495 list_add(&it->iters_node, &cset->task_iters);
4496}
4497
4498static void css_task_iter_skip(struct css_task_iter *it,
4499 struct task_struct *task)
4500{
4501 lockdep_assert_held(&css_set_lock);
4502
4503 if (it->task_pos == &task->cg_list) {
4504 it->task_pos = it->task_pos->next;
4505 it->flags |= CSS_TASK_ITER_SKIPPED;
4506 }
4507}
4508
4509static void css_task_iter_advance(struct css_task_iter *it)
4510{
4511 struct task_struct *task;
4512
4513 lockdep_assert_held(&css_set_lock);
4514repeat:
4515 if (it->task_pos) {
4516 /*
4517 * Advance iterator to find next entry. cset->tasks is
4518 * consumed first and then ->mg_tasks. After ->mg_tasks,
4519 * we move onto the next cset.
4520 */
4521 if (it->flags & CSS_TASK_ITER_SKIPPED)
4522 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4523 else
4524 it->task_pos = it->task_pos->next;
4525
4526 if (it->task_pos == it->tasks_head)
4527 it->task_pos = it->mg_tasks_head->next;
4528 if (it->task_pos == it->mg_tasks_head)
4529 it->task_pos = it->dying_tasks_head->next;
4530 if (it->task_pos == it->dying_tasks_head)
4531 css_task_iter_advance_css_set(it);
4532 } else {
4533 /* called from start, proceed to the first cset */
4534 css_task_iter_advance_css_set(it);
4535 }
4536
4537 if (!it->task_pos)
4538 return;
4539
4540 task = list_entry(it->task_pos, struct task_struct, cg_list);
4541
4542 if (it->flags & CSS_TASK_ITER_PROCS) {
4543 /* if PROCS, skip over tasks which aren't group leaders */
4544 if (!thread_group_leader(task))
4545 goto repeat;
4546
4547 /* and dying leaders w/o live member threads */
4548 if (!atomic_read(&task->signal->live))
4549 goto repeat;
4550 } else {
4551 /* skip all dying ones */
4552 if (task->flags & PF_EXITING)
4553 goto repeat;
4554 }
4555}
4556
4557/**
4558 * css_task_iter_start - initiate task iteration
4559 * @css: the css to walk tasks of
4560 * @flags: CSS_TASK_ITER_* flags
4561 * @it: the task iterator to use
4562 *
4563 * Initiate iteration through the tasks of @css. The caller can call
4564 * css_task_iter_next() to walk through the tasks until the function
4565 * returns NULL. On completion of iteration, css_task_iter_end() must be
4566 * called.
4567 */
4568void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4569 struct css_task_iter *it)
4570{
4571 /* no one should try to iterate before mounting cgroups */
4572 WARN_ON_ONCE(!use_task_css_set_links);
4573
4574 memset(it, 0, sizeof(*it));
4575
4576 spin_lock_irq(&css_set_lock);
4577
4578 it->ss = css->ss;
4579 it->flags = flags;
4580
4581 if (it->ss)
4582 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4583 else
4584 it->cset_pos = &css->cgroup->cset_links;
4585
4586 it->cset_head = it->cset_pos;
4587
4588 css_task_iter_advance(it);
4589
4590 spin_unlock_irq(&css_set_lock);
4591}
4592
4593/**
4594 * css_task_iter_next - return the next task for the iterator
4595 * @it: the task iterator being iterated
4596 *
4597 * The "next" function for task iteration. @it should have been
4598 * initialized via css_task_iter_start(). Returns NULL when the iteration
4599 * reaches the end.
4600 */
4601struct task_struct *css_task_iter_next(struct css_task_iter *it)
4602{
4603 if (it->cur_task) {
4604 put_task_struct(it->cur_task);
4605 it->cur_task = NULL;
4606 }
4607
4608 spin_lock_irq(&css_set_lock);
4609
4610 /* @it may be half-advanced by skips, finish advancing */
4611 if (it->flags & CSS_TASK_ITER_SKIPPED)
4612 css_task_iter_advance(it);
4613
4614 if (it->task_pos) {
4615 it->cur_task = list_entry(it->task_pos, struct task_struct,
4616 cg_list);
4617 get_task_struct(it->cur_task);
4618 css_task_iter_advance(it);
4619 }
4620
4621 spin_unlock_irq(&css_set_lock);
4622
4623 return it->cur_task;
4624}
4625
4626/**
4627 * css_task_iter_end - finish task iteration
4628 * @it: the task iterator to finish
4629 *
4630 * Finish task iteration started by css_task_iter_start().
4631 */
4632void css_task_iter_end(struct css_task_iter *it)
4633{
4634 if (it->cur_cset) {
4635 spin_lock_irq(&css_set_lock);
4636 list_del(&it->iters_node);
4637 put_css_set_locked(it->cur_cset);
4638 spin_unlock_irq(&css_set_lock);
4639 }
4640
4641 if (it->cur_dcset)
4642 put_css_set(it->cur_dcset);
4643
4644 if (it->cur_task)
4645 put_task_struct(it->cur_task);
4646}
4647
4648static void cgroup_procs_release(struct kernfs_open_file *of)
4649{
4650 if (of->priv) {
4651 css_task_iter_end(of->priv);
4652 kfree(of->priv);
4653 }
4654}
4655
4656static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4657{
4658 struct kernfs_open_file *of = s->private;
4659 struct css_task_iter *it = of->priv;
4660
4661 return css_task_iter_next(it);
4662}
4663
4664static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4665 unsigned int iter_flags)
4666{
4667 struct kernfs_open_file *of = s->private;
4668 struct cgroup *cgrp = seq_css(s)->cgroup;
4669 struct css_task_iter *it = of->priv;
4670
4671 /*
4672 * When a seq_file is seeked, it's always traversed sequentially
4673 * from position 0, so we can simply keep iterating on !0 *pos.
4674 */
4675 if (!it) {
4676 if (WARN_ON_ONCE((*pos)++))
4677 return ERR_PTR(-EINVAL);
4678
4679 it = kzalloc(sizeof(*it), GFP_KERNEL);
4680 if (!it)
4681 return ERR_PTR(-ENOMEM);
4682 of->priv = it;
4683 css_task_iter_start(&cgrp->self, iter_flags, it);
4684 } else if (!(*pos)++) {
4685 css_task_iter_end(it);
4686 css_task_iter_start(&cgrp->self, iter_flags, it);
4687 }
4688
4689 return cgroup_procs_next(s, NULL, NULL);
4690}
4691
4692static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4693{
4694 struct cgroup *cgrp = seq_css(s)->cgroup;
4695
4696 /*
4697 * All processes of a threaded subtree belong to the domain cgroup
4698 * of the subtree. Only threads can be distributed across the
4699 * subtree. Reject reads on cgroup.procs in the subtree proper.
4700 * They're always empty anyway.
4701 */
4702 if (cgroup_is_threaded(cgrp))
4703 return ERR_PTR(-EOPNOTSUPP);
4704
4705 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4706 CSS_TASK_ITER_THREADED);
4707}
4708
4709static int cgroup_procs_show(struct seq_file *s, void *v)
4710{
4711 seq_printf(s, "%d\n", task_pid_vnr(v));
4712 return 0;
4713}
4714
4715static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4716 struct cgroup *dst_cgrp,
4717 struct super_block *sb)
4718{
4719 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4720 struct cgroup *com_cgrp = src_cgrp;
4721 struct inode *inode;
4722 int ret;
4723
4724 lockdep_assert_held(&cgroup_mutex);
4725
4726 /* find the common ancestor */
4727 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4728 com_cgrp = cgroup_parent(com_cgrp);
4729
4730 /* %current should be authorized to migrate to the common ancestor */
4731 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4732 if (!inode)
4733 return -ENOMEM;
4734
4735 ret = inode_permission(inode, MAY_WRITE);
4736 iput(inode);
4737 if (ret)
4738 return ret;
4739
4740 /*
4741 * If namespaces are delegation boundaries, %current must be able
4742 * to see both source and destination cgroups from its namespace.
4743 */
4744 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4745 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4746 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4747 return -ENOENT;
4748
4749 return 0;
4750}
4751
4752static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4753 char *buf, size_t nbytes, loff_t off)
4754{
4755 struct cgroup *src_cgrp, *dst_cgrp;
4756 struct task_struct *task;
4757 ssize_t ret;
4758
4759 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4760 if (!dst_cgrp)
4761 return -ENODEV;
4762
4763 task = cgroup_procs_write_start(buf, true);
4764 ret = PTR_ERR_OR_ZERO(task);
4765 if (ret)
4766 goto out_unlock;
4767
4768 /* find the source cgroup */
4769 spin_lock_irq(&css_set_lock);
4770 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4771 spin_unlock_irq(&css_set_lock);
4772
4773 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4774 of->file->f_path.dentry->d_sb);
4775 if (ret)
4776 goto out_finish;
4777
4778 ret = cgroup_attach_task(dst_cgrp, task, true);
4779
4780out_finish:
4781 cgroup_procs_write_finish(task);
4782out_unlock:
4783 cgroup_kn_unlock(of->kn);
4784
4785 return ret ?: nbytes;
4786}
4787
4788static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4789{
4790 return __cgroup_procs_start(s, pos, 0);
4791}
4792
4793static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4794 char *buf, size_t nbytes, loff_t off)
4795{
4796 struct cgroup *src_cgrp, *dst_cgrp;
4797 struct task_struct *task;
4798 ssize_t ret;
4799
4800 buf = strstrip(buf);
4801
4802 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4803 if (!dst_cgrp)
4804 return -ENODEV;
4805
4806 task = cgroup_procs_write_start(buf, false);
4807 ret = PTR_ERR_OR_ZERO(task);
4808 if (ret)
4809 goto out_unlock;
4810
4811 /* find the source cgroup */
4812 spin_lock_irq(&css_set_lock);
4813 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4814 spin_unlock_irq(&css_set_lock);
4815
4816 /* thread migrations follow the cgroup.procs delegation rule */
4817 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4818 of->file->f_path.dentry->d_sb);
4819 if (ret)
4820 goto out_finish;
4821
4822 /* and must be contained in the same domain */
4823 ret = -EOPNOTSUPP;
4824 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4825 goto out_finish;
4826
4827 ret = cgroup_attach_task(dst_cgrp, task, false);
4828
4829out_finish:
4830 cgroup_procs_write_finish(task);
4831out_unlock:
4832 cgroup_kn_unlock(of->kn);
4833
4834 return ret ?: nbytes;
4835}
4836
4837/* cgroup core interface files for the default hierarchy */
4838static struct cftype cgroup_base_files[] = {
4839 {
4840 .name = "cgroup.type",
4841 .flags = CFTYPE_NOT_ON_ROOT,
4842 .seq_show = cgroup_type_show,
4843 .write = cgroup_type_write,
4844 },
4845 {
4846 .name = "cgroup.procs",
4847 .flags = CFTYPE_NS_DELEGATABLE,
4848 .file_offset = offsetof(struct cgroup, procs_file),
4849 .release = cgroup_procs_release,
4850 .seq_start = cgroup_procs_start,
4851 .seq_next = cgroup_procs_next,
4852 .seq_show = cgroup_procs_show,
4853 .write = cgroup_procs_write,
4854 },
4855 {
4856 .name = "cgroup.threads",
4857 .flags = CFTYPE_NS_DELEGATABLE,
4858 .release = cgroup_procs_release,
4859 .seq_start = cgroup_threads_start,
4860 .seq_next = cgroup_procs_next,
4861 .seq_show = cgroup_procs_show,
4862 .write = cgroup_threads_write,
4863 },
4864 {
4865 .name = "cgroup.controllers",
4866 .seq_show = cgroup_controllers_show,
4867 },
4868 {
4869 .name = "cgroup.subtree_control",
4870 .flags = CFTYPE_NS_DELEGATABLE,
4871 .seq_show = cgroup_subtree_control_show,
4872 .write = cgroup_subtree_control_write,
4873 },
4874 {
4875 .name = "cgroup.events",
4876 .flags = CFTYPE_NOT_ON_ROOT,
4877 .file_offset = offsetof(struct cgroup, events_file),
4878 .seq_show = cgroup_events_show,
4879 },
4880 {
4881 .name = "cgroup.max.descendants",
4882 .seq_show = cgroup_max_descendants_show,
4883 .write = cgroup_max_descendants_write,
4884 },
4885 {
4886 .name = "cgroup.max.depth",
4887 .seq_show = cgroup_max_depth_show,
4888 .write = cgroup_max_depth_write,
4889 },
4890 {
4891 .name = "cgroup.stat",
4892 .seq_show = cgroup_stat_show,
4893 },
4894 {
4895 .name = "cgroup.freeze",
4896 .flags = CFTYPE_NOT_ON_ROOT,
4897 .seq_show = cgroup_freeze_show,
4898 .write = cgroup_freeze_write,
4899 },
4900 {
4901 .name = "cpu.stat",
4902 .flags = CFTYPE_NOT_ON_ROOT,
4903 .seq_show = cpu_stat_show,
4904 },
4905#ifdef CONFIG_PSI
4906 {
4907 .name = "io.pressure",
4908 .seq_show = cgroup_io_pressure_show,
4909 .write = cgroup_io_pressure_write,
4910 .poll = cgroup_pressure_poll,
4911 .release = cgroup_pressure_release,
4912 },
4913 {
4914 .name = "memory.pressure",
4915 .seq_show = cgroup_memory_pressure_show,
4916 .write = cgroup_memory_pressure_write,
4917 .poll = cgroup_pressure_poll,
4918 .release = cgroup_pressure_release,
4919 },
4920 {
4921 .name = "cpu.pressure",
4922 .seq_show = cgroup_cpu_pressure_show,
4923 .write = cgroup_cpu_pressure_write,
4924 .poll = cgroup_pressure_poll,
4925 .release = cgroup_pressure_release,
4926 },
4927#endif /* CONFIG_PSI */
4928 { } /* terminate */
4929};
4930
4931/*
4932 * css destruction is four-stage process.
4933 *
4934 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4935 * Implemented in kill_css().
4936 *
4937 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4938 * and thus css_tryget_online() is guaranteed to fail, the css can be
4939 * offlined by invoking offline_css(). After offlining, the base ref is
4940 * put. Implemented in css_killed_work_fn().
4941 *
4942 * 3. When the percpu_ref reaches zero, the only possible remaining
4943 * accessors are inside RCU read sections. css_release() schedules the
4944 * RCU callback.
4945 *
4946 * 4. After the grace period, the css can be freed. Implemented in
4947 * css_free_work_fn().
4948 *
4949 * It is actually hairier because both step 2 and 4 require process context
4950 * and thus involve punting to css->destroy_work adding two additional
4951 * steps to the already complex sequence.
4952 */
4953static void css_free_rwork_fn(struct work_struct *work)
4954{
4955 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4956 struct cgroup_subsys_state, destroy_rwork);
4957 struct cgroup_subsys *ss = css->ss;
4958 struct cgroup *cgrp = css->cgroup;
4959
4960 percpu_ref_exit(&css->refcnt);
4961
4962 if (ss) {
4963 /* css free path */
4964 struct cgroup_subsys_state *parent = css->parent;
4965 int id = css->id;
4966
4967 ss->css_free(css);
4968 cgroup_idr_remove(&ss->css_idr, id);
4969 cgroup_put(cgrp);
4970
4971 if (parent)
4972 css_put(parent);
4973 } else {
4974 /* cgroup free path */
4975 atomic_dec(&cgrp->root->nr_cgrps);
4976 cgroup1_pidlist_destroy_all(cgrp);
4977 cancel_work_sync(&cgrp->release_agent_work);
4978
4979 if (cgroup_parent(cgrp)) {
4980 /*
4981 * We get a ref to the parent, and put the ref when
4982 * this cgroup is being freed, so it's guaranteed
4983 * that the parent won't be destroyed before its
4984 * children.
4985 */
4986 cgroup_put(cgroup_parent(cgrp));
4987 kernfs_put(cgrp->kn);
4988 psi_cgroup_free(cgrp);
4989 if (cgroup_on_dfl(cgrp))
4990 cgroup_rstat_exit(cgrp);
4991 kfree(cgrp);
4992 } else {
4993 /*
4994 * This is root cgroup's refcnt reaching zero,
4995 * which indicates that the root should be
4996 * released.
4997 */
4998 cgroup_destroy_root(cgrp->root);
4999 }
5000 }
5001}
5002
5003static void css_release_work_fn(struct work_struct *work)
5004{
5005 struct cgroup_subsys_state *css =
5006 container_of(work, struct cgroup_subsys_state, destroy_work);
5007 struct cgroup_subsys *ss = css->ss;
5008 struct cgroup *cgrp = css->cgroup;
5009
5010 mutex_lock(&cgroup_mutex);
5011
5012 css->flags |= CSS_RELEASED;
5013 list_del_rcu(&css->sibling);
5014
5015 if (ss) {
5016 /* css release path */
5017 if (!list_empty(&css->rstat_css_node)) {
5018 cgroup_rstat_flush(cgrp);
5019 list_del_rcu(&css->rstat_css_node);
5020 }
5021
5022 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5023 if (ss->css_released)
5024 ss->css_released(css);
5025 } else {
5026 struct cgroup *tcgrp;
5027
5028 /* cgroup release path */
5029 TRACE_CGROUP_PATH(release, cgrp);
5030
5031 if (cgroup_on_dfl(cgrp))
5032 cgroup_rstat_flush(cgrp);
5033
5034 spin_lock_irq(&css_set_lock);
5035 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5036 tcgrp = cgroup_parent(tcgrp))
5037 tcgrp->nr_dying_descendants--;
5038 spin_unlock_irq(&css_set_lock);
5039
5040 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
5041 cgrp->id = -1;
5042
5043 /*
5044 * There are two control paths which try to determine
5045 * cgroup from dentry without going through kernfs -
5046 * cgroupstats_build() and css_tryget_online_from_dir().
5047 * Those are supported by RCU protecting clearing of
5048 * cgrp->kn->priv backpointer.
5049 */
5050 if (cgrp->kn)
5051 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5052 NULL);
5053 }
5054
5055 mutex_unlock(&cgroup_mutex);
5056
5057 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5058 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5059}
5060
5061static void css_release(struct percpu_ref *ref)
5062{
5063 struct cgroup_subsys_state *css =
5064 container_of(ref, struct cgroup_subsys_state, refcnt);
5065
5066 INIT_WORK(&css->destroy_work, css_release_work_fn);
5067 queue_work(cgroup_destroy_wq, &css->destroy_work);
5068}
5069
5070static void init_and_link_css(struct cgroup_subsys_state *css,
5071 struct cgroup_subsys *ss, struct cgroup *cgrp)
5072{
5073 lockdep_assert_held(&cgroup_mutex);
5074
5075 cgroup_get_live(cgrp);
5076
5077 memset(css, 0, sizeof(*css));
5078 css->cgroup = cgrp;
5079 css->ss = ss;
5080 css->id = -1;
5081 INIT_LIST_HEAD(&css->sibling);
5082 INIT_LIST_HEAD(&css->children);
5083 INIT_LIST_HEAD(&css->rstat_css_node);
5084 css->serial_nr = css_serial_nr_next++;
5085 atomic_set(&css->online_cnt, 0);
5086
5087 if (cgroup_parent(cgrp)) {
5088 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5089 css_get(css->parent);
5090 }
5091
5092 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush)
5093 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5094
5095 BUG_ON(cgroup_css(cgrp, ss));
5096}
5097
5098/* invoke ->css_online() on a new CSS and mark it online if successful */
5099static int online_css(struct cgroup_subsys_state *css)
5100{
5101 struct cgroup_subsys *ss = css->ss;
5102 int ret = 0;
5103
5104 lockdep_assert_held(&cgroup_mutex);
5105
5106 if (ss->css_online)
5107 ret = ss->css_online(css);
5108 if (!ret) {
5109 css->flags |= CSS_ONLINE;
5110 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5111
5112 atomic_inc(&css->online_cnt);
5113 if (css->parent)
5114 atomic_inc(&css->parent->online_cnt);
5115 }
5116 return ret;
5117}
5118
5119/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5120static void offline_css(struct cgroup_subsys_state *css)
5121{
5122 struct cgroup_subsys *ss = css->ss;
5123
5124 lockdep_assert_held(&cgroup_mutex);
5125
5126 if (!(css->flags & CSS_ONLINE))
5127 return;
5128
5129 if (ss->css_offline)
5130 ss->css_offline(css);
5131
5132 css->flags &= ~CSS_ONLINE;
5133 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5134
5135 wake_up_all(&css->cgroup->offline_waitq);
5136}
5137
5138/**
5139 * css_create - create a cgroup_subsys_state
5140 * @cgrp: the cgroup new css will be associated with
5141 * @ss: the subsys of new css
5142 *
5143 * Create a new css associated with @cgrp - @ss pair. On success, the new
5144 * css is online and installed in @cgrp. This function doesn't create the
5145 * interface files. Returns 0 on success, -errno on failure.
5146 */
5147static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5148 struct cgroup_subsys *ss)
5149{
5150 struct cgroup *parent = cgroup_parent(cgrp);
5151 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5152 struct cgroup_subsys_state *css;
5153 int err;
5154
5155 lockdep_assert_held(&cgroup_mutex);
5156
5157 css = ss->css_alloc(parent_css);
5158 if (!css)
5159 css = ERR_PTR(-ENOMEM);
5160 if (IS_ERR(css))
5161 return css;
5162
5163 init_and_link_css(css, ss, cgrp);
5164
5165 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5166 if (err)
5167 goto err_free_css;
5168
5169 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5170 if (err < 0)
5171 goto err_free_css;
5172 css->id = err;
5173
5174 /* @css is ready to be brought online now, make it visible */
5175 list_add_tail_rcu(&css->sibling, &parent_css->children);
5176 cgroup_idr_replace(&ss->css_idr, css, css->id);
5177
5178 err = online_css(css);
5179 if (err)
5180 goto err_list_del;
5181
5182 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5183 cgroup_parent(parent)) {
5184 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5185 current->comm, current->pid, ss->name);
5186 if (!strcmp(ss->name, "memory"))
5187 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5188 ss->warned_broken_hierarchy = true;
5189 }
5190
5191 return css;
5192
5193err_list_del:
5194 list_del_rcu(&css->sibling);
5195err_free_css:
5196 list_del_rcu(&css->rstat_css_node);
5197 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5198 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5199 return ERR_PTR(err);
5200}
5201
5202/*
5203 * The returned cgroup is fully initialized including its control mask, but
5204 * it isn't associated with its kernfs_node and doesn't have the control
5205 * mask applied.
5206 */
5207static struct cgroup *cgroup_create(struct cgroup *parent)
5208{
5209 struct cgroup_root *root = parent->root;
5210 struct cgroup *cgrp, *tcgrp;
5211 int level = parent->level + 1;
5212 int ret;
5213
5214 /* allocate the cgroup and its ID, 0 is reserved for the root */
5215 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5216 GFP_KERNEL);
5217 if (!cgrp)
5218 return ERR_PTR(-ENOMEM);
5219
5220 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5221 if (ret)
5222 goto out_free_cgrp;
5223
5224 if (cgroup_on_dfl(parent)) {
5225 ret = cgroup_rstat_init(cgrp);
5226 if (ret)
5227 goto out_cancel_ref;
5228 }
5229
5230 /*
5231 * Temporarily set the pointer to NULL, so idr_find() won't return
5232 * a half-baked cgroup.
5233 */
5234 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5235 if (cgrp->id < 0) {
5236 ret = -ENOMEM;
5237 goto out_stat_exit;
5238 }
5239
5240 init_cgroup_housekeeping(cgrp);
5241
5242 cgrp->self.parent = &parent->self;
5243 cgrp->root = root;
5244 cgrp->level = level;
5245
5246 ret = psi_cgroup_alloc(cgrp);
5247 if (ret)
5248 goto out_idr_free;
5249
5250 ret = cgroup_bpf_inherit(cgrp);
5251 if (ret)
5252 goto out_psi_free;
5253
5254 /*
5255 * New cgroup inherits effective freeze counter, and
5256 * if the parent has to be frozen, the child has too.
5257 */
5258 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5259 if (cgrp->freezer.e_freeze) {
5260 /*
5261 * Set the CGRP_FREEZE flag, so when a process will be
5262 * attached to the child cgroup, it will become frozen.
5263 * At this point the new cgroup is unpopulated, so we can
5264 * consider it frozen immediately.
5265 */
5266 set_bit(CGRP_FREEZE, &cgrp->flags);
5267 set_bit(CGRP_FROZEN, &cgrp->flags);
5268 }
5269
5270 spin_lock_irq(&css_set_lock);
5271 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5272 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5273
5274 if (tcgrp != cgrp) {
5275 tcgrp->nr_descendants++;
5276
5277 /*
5278 * If the new cgroup is frozen, all ancestor cgroups
5279 * get a new frozen descendant, but their state can't
5280 * change because of this.
5281 */
5282 if (cgrp->freezer.e_freeze)
5283 tcgrp->freezer.nr_frozen_descendants++;
5284 }
5285 }
5286 spin_unlock_irq(&css_set_lock);
5287
5288 if (notify_on_release(parent))
5289 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5290
5291 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5292 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5293
5294 cgrp->self.serial_nr = css_serial_nr_next++;
5295
5296 /* allocation complete, commit to creation */
5297 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5298 atomic_inc(&root->nr_cgrps);
5299 cgroup_get_live(parent);
5300
5301 /*
5302 * @cgrp is now fully operational. If something fails after this
5303 * point, it'll be released via the normal destruction path.
5304 */
5305 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5306
5307 /*
5308 * On the default hierarchy, a child doesn't automatically inherit
5309 * subtree_control from the parent. Each is configured manually.
5310 */
5311 if (!cgroup_on_dfl(cgrp))
5312 cgrp->subtree_control = cgroup_control(cgrp);
5313
5314 cgroup_propagate_control(cgrp);
5315
5316 return cgrp;
5317
5318out_psi_free:
5319 psi_cgroup_free(cgrp);
5320out_idr_free:
5321 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
5322out_stat_exit:
5323 if (cgroup_on_dfl(parent))
5324 cgroup_rstat_exit(cgrp);
5325out_cancel_ref:
5326 percpu_ref_exit(&cgrp->self.refcnt);
5327out_free_cgrp:
5328 kfree(cgrp);
5329 return ERR_PTR(ret);
5330}
5331
5332static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5333{
5334 struct cgroup *cgroup;
5335 int ret = false;
5336 int level = 1;
5337
5338 lockdep_assert_held(&cgroup_mutex);
5339
5340 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5341 if (cgroup->nr_descendants >= cgroup->max_descendants)
5342 goto fail;
5343
5344 if (level > cgroup->max_depth)
5345 goto fail;
5346
5347 level++;
5348 }
5349
5350 ret = true;
5351fail:
5352 return ret;
5353}
5354
5355int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5356{
5357 struct cgroup *parent, *cgrp;
5358 struct kernfs_node *kn;
5359 int ret;
5360
5361 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5362 if (strchr(name, '\n'))
5363 return -EINVAL;
5364
5365 parent = cgroup_kn_lock_live(parent_kn, false);
5366 if (!parent)
5367 return -ENODEV;
5368
5369 if (!cgroup_check_hierarchy_limits(parent)) {
5370 ret = -EAGAIN;
5371 goto out_unlock;
5372 }
5373
5374 cgrp = cgroup_create(parent);
5375 if (IS_ERR(cgrp)) {
5376 ret = PTR_ERR(cgrp);
5377 goto out_unlock;
5378 }
5379
5380 /* create the directory */
5381 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5382 if (IS_ERR(kn)) {
5383 ret = PTR_ERR(kn);
5384 goto out_destroy;
5385 }
5386 cgrp->kn = kn;
5387
5388 /*
5389 * This extra ref will be put in cgroup_free_fn() and guarantees
5390 * that @cgrp->kn is always accessible.
5391 */
5392 kernfs_get(kn);
5393
5394 ret = cgroup_kn_set_ugid(kn);
5395 if (ret)
5396 goto out_destroy;
5397
5398 ret = css_populate_dir(&cgrp->self);
5399 if (ret)
5400 goto out_destroy;
5401
5402 ret = cgroup_apply_control_enable(cgrp);
5403 if (ret)
5404 goto out_destroy;
5405
5406 TRACE_CGROUP_PATH(mkdir, cgrp);
5407
5408 /* let's create and online css's */
5409 kernfs_activate(kn);
5410
5411 ret = 0;
5412 goto out_unlock;
5413
5414out_destroy:
5415 cgroup_destroy_locked(cgrp);
5416out_unlock:
5417 cgroup_kn_unlock(parent_kn);
5418 return ret;
5419}
5420
5421/*
5422 * This is called when the refcnt of a css is confirmed to be killed.
5423 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5424 * initate destruction and put the css ref from kill_css().
5425 */
5426static void css_killed_work_fn(struct work_struct *work)
5427{
5428 struct cgroup_subsys_state *css =
5429 container_of(work, struct cgroup_subsys_state, destroy_work);
5430
5431 mutex_lock(&cgroup_mutex);
5432
5433 do {
5434 offline_css(css);
5435 css_put(css);
5436 /* @css can't go away while we're holding cgroup_mutex */
5437 css = css->parent;
5438 } while (css && atomic_dec_and_test(&css->online_cnt));
5439
5440 mutex_unlock(&cgroup_mutex);
5441}
5442
5443/* css kill confirmation processing requires process context, bounce */
5444static void css_killed_ref_fn(struct percpu_ref *ref)
5445{
5446 struct cgroup_subsys_state *css =
5447 container_of(ref, struct cgroup_subsys_state, refcnt);
5448
5449 if (atomic_dec_and_test(&css->online_cnt)) {
5450 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5451 queue_work(cgroup_destroy_wq, &css->destroy_work);
5452 }
5453}
5454
5455/**
5456 * kill_css - destroy a css
5457 * @css: css to destroy
5458 *
5459 * This function initiates destruction of @css by removing cgroup interface
5460 * files and putting its base reference. ->css_offline() will be invoked
5461 * asynchronously once css_tryget_online() is guaranteed to fail and when
5462 * the reference count reaches zero, @css will be released.
5463 */
5464static void kill_css(struct cgroup_subsys_state *css)
5465{
5466 lockdep_assert_held(&cgroup_mutex);
5467
5468 if (css->flags & CSS_DYING)
5469 return;
5470
5471 css->flags |= CSS_DYING;
5472
5473 /*
5474 * This must happen before css is disassociated with its cgroup.
5475 * See seq_css() for details.
5476 */
5477 css_clear_dir(css);
5478
5479 /*
5480 * Killing would put the base ref, but we need to keep it alive
5481 * until after ->css_offline().
5482 */
5483 css_get(css);
5484
5485 /*
5486 * cgroup core guarantees that, by the time ->css_offline() is
5487 * invoked, no new css reference will be given out via
5488 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5489 * proceed to offlining css's because percpu_ref_kill() doesn't
5490 * guarantee that the ref is seen as killed on all CPUs on return.
5491 *
5492 * Use percpu_ref_kill_and_confirm() to get notifications as each
5493 * css is confirmed to be seen as killed on all CPUs.
5494 */
5495 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5496}
5497
5498/**
5499 * cgroup_destroy_locked - the first stage of cgroup destruction
5500 * @cgrp: cgroup to be destroyed
5501 *
5502 * css's make use of percpu refcnts whose killing latency shouldn't be
5503 * exposed to userland and are RCU protected. Also, cgroup core needs to
5504 * guarantee that css_tryget_online() won't succeed by the time
5505 * ->css_offline() is invoked. To satisfy all the requirements,
5506 * destruction is implemented in the following two steps.
5507 *
5508 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5509 * userland visible parts and start killing the percpu refcnts of
5510 * css's. Set up so that the next stage will be kicked off once all
5511 * the percpu refcnts are confirmed to be killed.
5512 *
5513 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5514 * rest of destruction. Once all cgroup references are gone, the
5515 * cgroup is RCU-freed.
5516 *
5517 * This function implements s1. After this step, @cgrp is gone as far as
5518 * the userland is concerned and a new cgroup with the same name may be
5519 * created. As cgroup doesn't care about the names internally, this
5520 * doesn't cause any problem.
5521 */
5522static int cgroup_destroy_locked(struct cgroup *cgrp)
5523 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5524{
5525 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5526 struct cgroup_subsys_state *css;
5527 struct cgrp_cset_link *link;
5528 int ssid;
5529
5530 lockdep_assert_held(&cgroup_mutex);
5531
5532 /*
5533 * Only migration can raise populated from zero and we're already
5534 * holding cgroup_mutex.
5535 */
5536 if (cgroup_is_populated(cgrp))
5537 return -EBUSY;
5538
5539 /*
5540 * Make sure there's no live children. We can't test emptiness of
5541 * ->self.children as dead children linger on it while being
5542 * drained; otherwise, "rmdir parent/child parent" may fail.
5543 */
5544 if (css_has_online_children(&cgrp->self))
5545 return -EBUSY;
5546
5547 /*
5548 * Mark @cgrp and the associated csets dead. The former prevents
5549 * further task migration and child creation by disabling
5550 * cgroup_lock_live_group(). The latter makes the csets ignored by
5551 * the migration path.
5552 */
5553 cgrp->self.flags &= ~CSS_ONLINE;
5554
5555 spin_lock_irq(&css_set_lock);
5556 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5557 link->cset->dead = true;
5558 spin_unlock_irq(&css_set_lock);
5559
5560 /* initiate massacre of all css's */
5561 for_each_css(css, ssid, cgrp)
5562 kill_css(css);
5563
5564 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5565 css_clear_dir(&cgrp->self);
5566 kernfs_remove(cgrp->kn);
5567
5568 if (parent && cgroup_is_threaded(cgrp))
5569 parent->nr_threaded_children--;
5570
5571 spin_lock_irq(&css_set_lock);
5572 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5573 tcgrp->nr_descendants--;
5574 tcgrp->nr_dying_descendants++;
5575 /*
5576 * If the dying cgroup is frozen, decrease frozen descendants
5577 * counters of ancestor cgroups.
5578 */
5579 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5580 tcgrp->freezer.nr_frozen_descendants--;
5581 }
5582 spin_unlock_irq(&css_set_lock);
5583
5584 cgroup1_check_for_release(parent);
5585
5586 cgroup_bpf_offline(cgrp);
5587
5588 /* put the base reference */
5589 percpu_ref_kill(&cgrp->self.refcnt);
5590
5591 return 0;
5592};
5593
5594int cgroup_rmdir(struct kernfs_node *kn)
5595{
5596 struct cgroup *cgrp;
5597 int ret = 0;
5598
5599 cgrp = cgroup_kn_lock_live(kn, false);
5600 if (!cgrp)
5601 return 0;
5602
5603 ret = cgroup_destroy_locked(cgrp);
5604 if (!ret)
5605 TRACE_CGROUP_PATH(rmdir, cgrp);
5606
5607 cgroup_kn_unlock(kn);
5608 return ret;
5609}
5610
5611static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5612 .show_options = cgroup_show_options,
5613 .mkdir = cgroup_mkdir,
5614 .rmdir = cgroup_rmdir,
5615 .show_path = cgroup_show_path,
5616};
5617
5618static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5619{
5620 struct cgroup_subsys_state *css;
5621
5622 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5623
5624 mutex_lock(&cgroup_mutex);
5625
5626 idr_init(&ss->css_idr);
5627 INIT_LIST_HEAD(&ss->cfts);
5628
5629 /* Create the root cgroup state for this subsystem */
5630 ss->root = &cgrp_dfl_root;
5631 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5632 /* We don't handle early failures gracefully */
5633 BUG_ON(IS_ERR(css));
5634 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5635
5636 /*
5637 * Root csses are never destroyed and we can't initialize
5638 * percpu_ref during early init. Disable refcnting.
5639 */
5640 css->flags |= CSS_NO_REF;
5641
5642 if (early) {
5643 /* allocation can't be done safely during early init */
5644 css->id = 1;
5645 } else {
5646 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5647 BUG_ON(css->id < 0);
5648 }
5649
5650 /* Update the init_css_set to contain a subsys
5651 * pointer to this state - since the subsystem is
5652 * newly registered, all tasks and hence the
5653 * init_css_set is in the subsystem's root cgroup. */
5654 init_css_set.subsys[ss->id] = css;
5655
5656 have_fork_callback |= (bool)ss->fork << ss->id;
5657 have_exit_callback |= (bool)ss->exit << ss->id;
5658 have_release_callback |= (bool)ss->release << ss->id;
5659 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5660
5661 /* At system boot, before all subsystems have been
5662 * registered, no tasks have been forked, so we don't
5663 * need to invoke fork callbacks here. */
5664 BUG_ON(!list_empty(&init_task.tasks));
5665
5666 BUG_ON(online_css(css));
5667
5668 mutex_unlock(&cgroup_mutex);
5669}
5670
5671/**
5672 * cgroup_init_early - cgroup initialization at system boot
5673 *
5674 * Initialize cgroups at system boot, and initialize any
5675 * subsystems that request early init.
5676 */
5677int __init cgroup_init_early(void)
5678{
5679 static struct cgroup_fs_context __initdata ctx;
5680 struct cgroup_subsys *ss;
5681 int i;
5682
5683 ctx.root = &cgrp_dfl_root;
5684 init_cgroup_root(&ctx);
5685 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5686
5687 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5688
5689 for_each_subsys(ss, i) {
5690 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5691 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5692 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5693 ss->id, ss->name);
5694 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5695 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5696
5697 ss->id = i;
5698 ss->name = cgroup_subsys_name[i];
5699 if (!ss->legacy_name)
5700 ss->legacy_name = cgroup_subsys_name[i];
5701
5702 if (ss->early_init)
5703 cgroup_init_subsys(ss, true);
5704 }
5705 return 0;
5706}
5707
5708static u16 cgroup_disable_mask __initdata;
5709
5710/**
5711 * cgroup_init - cgroup initialization
5712 *
5713 * Register cgroup filesystem and /proc file, and initialize
5714 * any subsystems that didn't request early init.
5715 */
5716int __init cgroup_init(void)
5717{
5718 struct cgroup_subsys *ss;
5719 int ssid;
5720
5721 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5722 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5723 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5724
5725 cgroup_rstat_boot();
5726
5727 /*
5728 * The latency of the synchronize_rcu() is too high for cgroups,
5729 * avoid it at the cost of forcing all readers into the slow path.
5730 */
5731 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5732
5733 get_user_ns(init_cgroup_ns.user_ns);
5734
5735 mutex_lock(&cgroup_mutex);
5736
5737 /*
5738 * Add init_css_set to the hash table so that dfl_root can link to
5739 * it during init.
5740 */
5741 hash_add(css_set_table, &init_css_set.hlist,
5742 css_set_hash(init_css_set.subsys));
5743
5744 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5745
5746 mutex_unlock(&cgroup_mutex);
5747
5748 for_each_subsys(ss, ssid) {
5749 if (ss->early_init) {
5750 struct cgroup_subsys_state *css =
5751 init_css_set.subsys[ss->id];
5752
5753 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5754 GFP_KERNEL);
5755 BUG_ON(css->id < 0);
5756 } else {
5757 cgroup_init_subsys(ss, false);
5758 }
5759
5760 list_add_tail(&init_css_set.e_cset_node[ssid],
5761 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5762
5763 /*
5764 * Setting dfl_root subsys_mask needs to consider the
5765 * disabled flag and cftype registration needs kmalloc,
5766 * both of which aren't available during early_init.
5767 */
5768 if (cgroup_disable_mask & (1 << ssid)) {
5769 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5770 printk(KERN_INFO "Disabling %s control group subsystem\n",
5771 ss->name);
5772 continue;
5773 }
5774
5775 if (cgroup1_ssid_disabled(ssid))
5776 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5777 ss->name);
5778
5779 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5780
5781 /* implicit controllers must be threaded too */
5782 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5783
5784 if (ss->implicit_on_dfl)
5785 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5786 else if (!ss->dfl_cftypes)
5787 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5788
5789 if (ss->threaded)
5790 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5791
5792 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5793 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5794 } else {
5795 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5796 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5797 }
5798
5799 if (ss->bind)
5800 ss->bind(init_css_set.subsys[ssid]);
5801
5802 mutex_lock(&cgroup_mutex);
5803 css_populate_dir(init_css_set.subsys[ssid]);
5804 mutex_unlock(&cgroup_mutex);
5805 }
5806
5807 /* init_css_set.subsys[] has been updated, re-hash */
5808 hash_del(&init_css_set.hlist);
5809 hash_add(css_set_table, &init_css_set.hlist,
5810 css_set_hash(init_css_set.subsys));
5811
5812 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5813 WARN_ON(register_filesystem(&cgroup_fs_type));
5814 WARN_ON(register_filesystem(&cgroup2_fs_type));
5815 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5816#ifdef CONFIG_CPUSETS
5817 WARN_ON(register_filesystem(&cpuset_fs_type));
5818#endif
5819
5820 return 0;
5821}
5822
5823static int __init cgroup_wq_init(void)
5824{
5825 /*
5826 * There isn't much point in executing destruction path in
5827 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5828 * Use 1 for @max_active.
5829 *
5830 * We would prefer to do this in cgroup_init() above, but that
5831 * is called before init_workqueues(): so leave this until after.
5832 */
5833 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5834 BUG_ON(!cgroup_destroy_wq);
5835 return 0;
5836}
5837core_initcall(cgroup_wq_init);
5838
5839void cgroup_path_from_kernfs_id(const union kernfs_node_id *id,
5840 char *buf, size_t buflen)
5841{
5842 struct kernfs_node *kn;
5843
5844 kn = kernfs_get_node_by_id(cgrp_dfl_root.kf_root, id);
5845 if (!kn)
5846 return;
5847 kernfs_path(kn, buf, buflen);
5848 kernfs_put(kn);
5849}
5850
5851/*
5852 * proc_cgroup_show()
5853 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5854 * - Used for /proc/<pid>/cgroup.
5855 */
5856int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5857 struct pid *pid, struct task_struct *tsk)
5858{
5859 char *buf;
5860 int retval;
5861 struct cgroup_root *root;
5862
5863 retval = -ENOMEM;
5864 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5865 if (!buf)
5866 goto out;
5867
5868 mutex_lock(&cgroup_mutex);
5869 spin_lock_irq(&css_set_lock);
5870
5871 for_each_root(root) {
5872 struct cgroup_subsys *ss;
5873 struct cgroup *cgrp;
5874 int ssid, count = 0;
5875
5876 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5877 continue;
5878
5879 seq_printf(m, "%d:", root->hierarchy_id);
5880 if (root != &cgrp_dfl_root)
5881 for_each_subsys(ss, ssid)
5882 if (root->subsys_mask & (1 << ssid))
5883 seq_printf(m, "%s%s", count++ ? "," : "",
5884 ss->legacy_name);
5885 if (strlen(root->name))
5886 seq_printf(m, "%sname=%s", count ? "," : "",
5887 root->name);
5888 seq_putc(m, ':');
5889
5890 cgrp = task_cgroup_from_root(tsk, root);
5891
5892 /*
5893 * On traditional hierarchies, all zombie tasks show up as
5894 * belonging to the root cgroup. On the default hierarchy,
5895 * while a zombie doesn't show up in "cgroup.procs" and
5896 * thus can't be migrated, its /proc/PID/cgroup keeps
5897 * reporting the cgroup it belonged to before exiting. If
5898 * the cgroup is removed before the zombie is reaped,
5899 * " (deleted)" is appended to the cgroup path.
5900 */
5901 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5902 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5903 current->nsproxy->cgroup_ns);
5904 if (retval >= PATH_MAX)
5905 retval = -ENAMETOOLONG;
5906 if (retval < 0)
5907 goto out_unlock;
5908
5909 seq_puts(m, buf);
5910 } else {
5911 seq_puts(m, "/");
5912 }
5913
5914 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5915 seq_puts(m, " (deleted)\n");
5916 else
5917 seq_putc(m, '\n');
5918 }
5919
5920 retval = 0;
5921out_unlock:
5922 spin_unlock_irq(&css_set_lock);
5923 mutex_unlock(&cgroup_mutex);
5924 kfree(buf);
5925out:
5926 return retval;
5927}
5928
5929/**
5930 * cgroup_fork - initialize cgroup related fields during copy_process()
5931 * @child: pointer to task_struct of forking parent process.
5932 *
5933 * A task is associated with the init_css_set until cgroup_post_fork()
5934 * attaches it to the parent's css_set. Empty cg_list indicates that
5935 * @child isn't holding reference to its css_set.
5936 */
5937void cgroup_fork(struct task_struct *child)
5938{
5939 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5940 INIT_LIST_HEAD(&child->cg_list);
5941}
5942
5943/**
5944 * cgroup_can_fork - called on a new task before the process is exposed
5945 * @child: the task in question.
5946 *
5947 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5948 * returns an error, the fork aborts with that error code. This allows for
5949 * a cgroup subsystem to conditionally allow or deny new forks.
5950 */
5951int cgroup_can_fork(struct task_struct *child)
5952{
5953 struct cgroup_subsys *ss;
5954 int i, j, ret;
5955
5956 do_each_subsys_mask(ss, i, have_canfork_callback) {
5957 ret = ss->can_fork(child);
5958 if (ret)
5959 goto out_revert;
5960 } while_each_subsys_mask();
5961
5962 return 0;
5963
5964out_revert:
5965 for_each_subsys(ss, j) {
5966 if (j >= i)
5967 break;
5968 if (ss->cancel_fork)
5969 ss->cancel_fork(child);
5970 }
5971
5972 return ret;
5973}
5974
5975/**
5976 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5977 * @child: the task in question
5978 *
5979 * This calls the cancel_fork() callbacks if a fork failed *after*
5980 * cgroup_can_fork() succeded.
5981 */
5982void cgroup_cancel_fork(struct task_struct *child)
5983{
5984 struct cgroup_subsys *ss;
5985 int i;
5986
5987 for_each_subsys(ss, i)
5988 if (ss->cancel_fork)
5989 ss->cancel_fork(child);
5990}
5991
5992/**
5993 * cgroup_post_fork - called on a new task after adding it to the task list
5994 * @child: the task in question
5995 *
5996 * Adds the task to the list running through its css_set if necessary and
5997 * call the subsystem fork() callbacks. Has to be after the task is
5998 * visible on the task list in case we race with the first call to
5999 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
6000 * list.
6001 */
6002void cgroup_post_fork(struct task_struct *child)
6003{
6004 struct cgroup_subsys *ss;
6005 int i;
6006
6007 /*
6008 * This may race against cgroup_enable_task_cg_lists(). As that
6009 * function sets use_task_css_set_links before grabbing
6010 * tasklist_lock and we just went through tasklist_lock to add
6011 * @child, it's guaranteed that either we see the set
6012 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
6013 * @child during its iteration.
6014 *
6015 * If we won the race, @child is associated with %current's
6016 * css_set. Grabbing css_set_lock guarantees both that the
6017 * association is stable, and, on completion of the parent's
6018 * migration, @child is visible in the source of migration or
6019 * already in the destination cgroup. This guarantee is necessary
6020 * when implementing operations which need to migrate all tasks of
6021 * a cgroup to another.
6022 *
6023 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
6024 * will remain in init_css_set. This is safe because all tasks are
6025 * in the init_css_set before cg_links is enabled and there's no
6026 * operation which transfers all tasks out of init_css_set.
6027 */
6028 if (use_task_css_set_links) {
6029 struct css_set *cset;
6030
6031 spin_lock_irq(&css_set_lock);
6032 cset = task_css_set(current);
6033 if (list_empty(&child->cg_list)) {
6034 get_css_set(cset);
6035 cset->nr_tasks++;
6036 css_set_move_task(child, NULL, cset, false);
6037 }
6038
6039 /*
6040 * If the cgroup has to be frozen, the new task has too.
6041 * Let's set the JOBCTL_TRAP_FREEZE jobctl bit to get
6042 * the task into the frozen state.
6043 */
6044 if (unlikely(cgroup_task_freeze(child))) {
6045 spin_lock(&child->sighand->siglock);
6046 WARN_ON_ONCE(child->frozen);
6047 child->jobctl |= JOBCTL_TRAP_FREEZE;
6048 spin_unlock(&child->sighand->siglock);
6049
6050 /*
6051 * Calling cgroup_update_frozen() isn't required here,
6052 * because it will be called anyway a bit later
6053 * from do_freezer_trap(). So we avoid cgroup's
6054 * transient switch from the frozen state and back.
6055 */
6056 }
6057
6058 spin_unlock_irq(&css_set_lock);
6059 }
6060
6061 /*
6062 * Call ss->fork(). This must happen after @child is linked on
6063 * css_set; otherwise, @child might change state between ->fork()
6064 * and addition to css_set.
6065 */
6066 do_each_subsys_mask(ss, i, have_fork_callback) {
6067 ss->fork(child);
6068 } while_each_subsys_mask();
6069}
6070
6071/**
6072 * cgroup_exit - detach cgroup from exiting task
6073 * @tsk: pointer to task_struct of exiting process
6074 *
6075 * Description: Detach cgroup from @tsk and release it.
6076 *
6077 * Note that cgroups marked notify_on_release force every task in
6078 * them to take the global cgroup_mutex mutex when exiting.
6079 * This could impact scaling on very large systems. Be reluctant to
6080 * use notify_on_release cgroups where very high task exit scaling
6081 * is required on large systems.
6082 *
6083 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
6084 * call cgroup_exit() while the task is still competent to handle
6085 * notify_on_release(), then leave the task attached to the root cgroup in
6086 * each hierarchy for the remainder of its exit. No need to bother with
6087 * init_css_set refcnting. init_css_set never goes away and we can't race
6088 * with migration path - PF_EXITING is visible to migration path.
6089 */
6090void cgroup_exit(struct task_struct *tsk)
6091{
6092 struct cgroup_subsys *ss;
6093 struct css_set *cset;
6094 int i;
6095
6096 /*
6097 * Unlink from @tsk from its css_set. As migration path can't race
6098 * with us, we can check css_set and cg_list without synchronization.
6099 */
6100 cset = task_css_set(tsk);
6101
6102 if (!list_empty(&tsk->cg_list)) {
6103 spin_lock_irq(&css_set_lock);
6104 css_set_move_task(tsk, cset, NULL, false);
6105 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6106 cset->nr_tasks--;
6107
6108 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6109 if (unlikely(cgroup_task_freeze(tsk)))
6110 cgroup_update_frozen(task_dfl_cgroup(tsk));
6111
6112 spin_unlock_irq(&css_set_lock);
6113 } else {
6114 get_css_set(cset);
6115 }
6116
6117 /* see cgroup_post_fork() for details */
6118 do_each_subsys_mask(ss, i, have_exit_callback) {
6119 ss->exit(tsk);
6120 } while_each_subsys_mask();
6121}
6122
6123void cgroup_release(struct task_struct *task)
6124{
6125 struct cgroup_subsys *ss;
6126 int ssid;
6127
6128 do_each_subsys_mask(ss, ssid, have_release_callback) {
6129 ss->release(task);
6130 } while_each_subsys_mask();
6131
6132 if (use_task_css_set_links) {
6133 spin_lock_irq(&css_set_lock);
6134 css_set_skip_task_iters(task_css_set(task), task);
6135 list_del_init(&task->cg_list);
6136 spin_unlock_irq(&css_set_lock);
6137 }
6138}
6139
6140void cgroup_free(struct task_struct *task)
6141{
6142 struct css_set *cset = task_css_set(task);
6143 put_css_set(cset);
6144}
6145
6146static int __init cgroup_disable(char *str)
6147{
6148 struct cgroup_subsys *ss;
6149 char *token;
6150 int i;
6151
6152 while ((token = strsep(&str, ",")) != NULL) {
6153 if (!*token)
6154 continue;
6155
6156 for_each_subsys(ss, i) {
6157 if (strcmp(token, ss->name) &&
6158 strcmp(token, ss->legacy_name))
6159 continue;
6160 cgroup_disable_mask |= 1 << i;
6161 }
6162 }
6163 return 1;
6164}
6165__setup("cgroup_disable=", cgroup_disable);
6166
6167void __init __weak enable_debug_cgroup(void) { }
6168
6169static int __init enable_cgroup_debug(char *str)
6170{
6171 cgroup_debug = true;
6172 enable_debug_cgroup();
6173 return 1;
6174}
6175__setup("cgroup_debug", enable_cgroup_debug);
6176
6177/**
6178 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6179 * @dentry: directory dentry of interest
6180 * @ss: subsystem of interest
6181 *
6182 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6183 * to get the corresponding css and return it. If such css doesn't exist
6184 * or can't be pinned, an ERR_PTR value is returned.
6185 */
6186struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6187 struct cgroup_subsys *ss)
6188{
6189 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6190 struct file_system_type *s_type = dentry->d_sb->s_type;
6191 struct cgroup_subsys_state *css = NULL;
6192 struct cgroup *cgrp;
6193
6194 /* is @dentry a cgroup dir? */
6195 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6196 !kn || kernfs_type(kn) != KERNFS_DIR)
6197 return ERR_PTR(-EBADF);
6198
6199 rcu_read_lock();
6200
6201 /*
6202 * This path doesn't originate from kernfs and @kn could already
6203 * have been or be removed at any point. @kn->priv is RCU
6204 * protected for this access. See css_release_work_fn() for details.
6205 */
6206 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6207 if (cgrp)
6208 css = cgroup_css(cgrp, ss);
6209
6210 if (!css || !css_tryget_online(css))
6211 css = ERR_PTR(-ENOENT);
6212
6213 rcu_read_unlock();
6214 return css;
6215}
6216
6217/**
6218 * css_from_id - lookup css by id
6219 * @id: the cgroup id
6220 * @ss: cgroup subsys to be looked into
6221 *
6222 * Returns the css if there's valid one with @id, otherwise returns NULL.
6223 * Should be called under rcu_read_lock().
6224 */
6225struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6226{
6227 WARN_ON_ONCE(!rcu_read_lock_held());
6228 return idr_find(&ss->css_idr, id);
6229}
6230
6231/**
6232 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6233 * @path: path on the default hierarchy
6234 *
6235 * Find the cgroup at @path on the default hierarchy, increment its
6236 * reference count and return it. Returns pointer to the found cgroup on
6237 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6238 * if @path points to a non-directory.
6239 */
6240struct cgroup *cgroup_get_from_path(const char *path)
6241{
6242 struct kernfs_node *kn;
6243 struct cgroup *cgrp;
6244
6245 mutex_lock(&cgroup_mutex);
6246
6247 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6248 if (kn) {
6249 if (kernfs_type(kn) == KERNFS_DIR) {
6250 cgrp = kn->priv;
6251 cgroup_get_live(cgrp);
6252 } else {
6253 cgrp = ERR_PTR(-ENOTDIR);
6254 }
6255 kernfs_put(kn);
6256 } else {
6257 cgrp = ERR_PTR(-ENOENT);
6258 }
6259
6260 mutex_unlock(&cgroup_mutex);
6261 return cgrp;
6262}
6263EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6264
6265/**
6266 * cgroup_get_from_fd - get a cgroup pointer from a fd
6267 * @fd: fd obtained by open(cgroup2_dir)
6268 *
6269 * Find the cgroup from a fd which should be obtained
6270 * by opening a cgroup directory. Returns a pointer to the
6271 * cgroup on success. ERR_PTR is returned if the cgroup
6272 * cannot be found.
6273 */
6274struct cgroup *cgroup_get_from_fd(int fd)
6275{
6276 struct cgroup_subsys_state *css;
6277 struct cgroup *cgrp;
6278 struct file *f;
6279
6280 f = fget_raw(fd);
6281 if (!f)
6282 return ERR_PTR(-EBADF);
6283
6284 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6285 fput(f);
6286 if (IS_ERR(css))
6287 return ERR_CAST(css);
6288
6289 cgrp = css->cgroup;
6290 if (!cgroup_on_dfl(cgrp)) {
6291 cgroup_put(cgrp);
6292 return ERR_PTR(-EBADF);
6293 }
6294
6295 return cgrp;
6296}
6297EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6298
6299static u64 power_of_ten(int power)
6300{
6301 u64 v = 1;
6302 while (power--)
6303 v *= 10;
6304 return v;
6305}
6306
6307/**
6308 * cgroup_parse_float - parse a floating number
6309 * @input: input string
6310 * @dec_shift: number of decimal digits to shift
6311 * @v: output
6312 *
6313 * Parse a decimal floating point number in @input and store the result in
6314 * @v with decimal point right shifted @dec_shift times. For example, if
6315 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6316 * Returns 0 on success, -errno otherwise.
6317 *
6318 * There's nothing cgroup specific about this function except that it's
6319 * currently the only user.
6320 */
6321int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6322{
6323 s64 whole, frac = 0;
6324 int fstart = 0, fend = 0, flen;
6325
6326 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6327 return -EINVAL;
6328 if (frac < 0)
6329 return -EINVAL;
6330
6331 flen = fend > fstart ? fend - fstart : 0;
6332 if (flen < dec_shift)
6333 frac *= power_of_ten(dec_shift - flen);
6334 else
6335 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6336
6337 *v = whole * power_of_ten(dec_shift) + frac;
6338 return 0;
6339}
6340
6341/*
6342 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6343 * definition in cgroup-defs.h.
6344 */
6345#ifdef CONFIG_SOCK_CGROUP_DATA
6346
6347#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6348
6349DEFINE_SPINLOCK(cgroup_sk_update_lock);
6350static bool cgroup_sk_alloc_disabled __read_mostly;
6351
6352void cgroup_sk_alloc_disable(void)
6353{
6354 if (cgroup_sk_alloc_disabled)
6355 return;
6356 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6357 cgroup_sk_alloc_disabled = true;
6358}
6359
6360#else
6361
6362#define cgroup_sk_alloc_disabled false
6363
6364#endif
6365
6366void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6367{
6368 if (cgroup_sk_alloc_disabled)
6369 return;
6370
6371 /* Socket clone path */
6372 if (skcd->val) {
6373 /*
6374 * We might be cloning a socket which is left in an empty
6375 * cgroup and the cgroup might have already been rmdir'd.
6376 * Don't use cgroup_get_live().
6377 */
6378 cgroup_get(sock_cgroup_ptr(skcd));
6379 cgroup_bpf_get(sock_cgroup_ptr(skcd));
6380 return;
6381 }
6382
6383 rcu_read_lock();
6384
6385 while (true) {
6386 struct css_set *cset;
6387
6388 cset = task_css_set(current);
6389 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6390 skcd->val = (unsigned long)cset->dfl_cgrp;
6391 cgroup_bpf_get(cset->dfl_cgrp);
6392 break;
6393 }
6394 cpu_relax();
6395 }
6396
6397 rcu_read_unlock();
6398}
6399
6400void cgroup_sk_free(struct sock_cgroup_data *skcd)
6401{
6402 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6403
6404 cgroup_bpf_put(cgrp);
6405 cgroup_put(cgrp);
6406}
6407
6408#endif /* CONFIG_SOCK_CGROUP_DATA */
6409
6410#ifdef CONFIG_CGROUP_BPF
6411int cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
6412 enum bpf_attach_type type, u32 flags)
6413{
6414 int ret;
6415
6416 mutex_lock(&cgroup_mutex);
6417 ret = __cgroup_bpf_attach(cgrp, prog, type, flags);
6418 mutex_unlock(&cgroup_mutex);
6419 return ret;
6420}
6421int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6422 enum bpf_attach_type type, u32 flags)
6423{
6424 int ret;
6425
6426 mutex_lock(&cgroup_mutex);
6427 ret = __cgroup_bpf_detach(cgrp, prog, type);
6428 mutex_unlock(&cgroup_mutex);
6429 return ret;
6430}
6431int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6432 union bpf_attr __user *uattr)
6433{
6434 int ret;
6435
6436 mutex_lock(&cgroup_mutex);
6437 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6438 mutex_unlock(&cgroup_mutex);
6439 return ret;
6440}
6441#endif /* CONFIG_CGROUP_BPF */
6442
6443#ifdef CONFIG_SYSFS
6444static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6445 ssize_t size, const char *prefix)
6446{
6447 struct cftype *cft;
6448 ssize_t ret = 0;
6449
6450 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6451 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6452 continue;
6453
6454 if (prefix)
6455 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6456
6457 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6458
6459 if (WARN_ON(ret >= size))
6460 break;
6461 }
6462
6463 return ret;
6464}
6465
6466static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6467 char *buf)
6468{
6469 struct cgroup_subsys *ss;
6470 int ssid;
6471 ssize_t ret = 0;
6472
6473 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6474 NULL);
6475
6476 for_each_subsys(ss, ssid)
6477 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6478 PAGE_SIZE - ret,
6479 cgroup_subsys_name[ssid]);
6480
6481 return ret;
6482}
6483static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6484
6485static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6486 char *buf)
6487{
6488 return snprintf(buf, PAGE_SIZE, "nsdelegate\nmemory_localevents\n");
6489}
6490static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6491
6492static struct attribute *cgroup_sysfs_attrs[] = {
6493 &cgroup_delegate_attr.attr,
6494 &cgroup_features_attr.attr,
6495 NULL,
6496};
6497
6498static const struct attribute_group cgroup_sysfs_attr_group = {
6499 .attrs = cgroup_sysfs_attrs,
6500 .name = "cgroup",
6501};
6502
6503static int __init cgroup_sysfs_init(void)
6504{
6505 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6506}
6507subsys_initcall(cgroup_sysfs_init);
6508
6509#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/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 return ret;
1749 }
1750 } else {
1751 ret = cgroup_addrm_files(css, cgrp,
1752 cgroup1_base_files, true);
1753 if (ret < 0)
1754 return ret;
1755 }
1756 } else {
1757 list_for_each_entry(cfts, &css->ss->cfts, node) {
1758 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1759 if (ret < 0) {
1760 failed_cfts = cfts;
1761 goto err;
1762 }
1763 }
1764 }
1765
1766 css->flags |= CSS_VISIBLE;
1767
1768 return 0;
1769err:
1770 list_for_each_entry(cfts, &css->ss->cfts, node) {
1771 if (cfts == failed_cfts)
1772 break;
1773 cgroup_addrm_files(css, cgrp, cfts, false);
1774 }
1775 return ret;
1776}
1777
1778int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1779{
1780 struct cgroup *dcgrp = &dst_root->cgrp;
1781 struct cgroup_subsys *ss;
1782 int ssid, ret;
1783 u16 dfl_disable_ss_mask = 0;
1784
1785 lockdep_assert_held(&cgroup_mutex);
1786
1787 do_each_subsys_mask(ss, ssid, ss_mask) {
1788 /*
1789 * If @ss has non-root csses attached to it, can't move.
1790 * If @ss is an implicit controller, it is exempt from this
1791 * rule and can be stolen.
1792 */
1793 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1794 !ss->implicit_on_dfl)
1795 return -EBUSY;
1796
1797 /* can't move between two non-dummy roots either */
1798 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1799 return -EBUSY;
1800
1801 /*
1802 * Collect ssid's that need to be disabled from default
1803 * hierarchy.
1804 */
1805 if (ss->root == &cgrp_dfl_root)
1806 dfl_disable_ss_mask |= 1 << ssid;
1807
1808 } while_each_subsys_mask();
1809
1810 if (dfl_disable_ss_mask) {
1811 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1812
1813 /*
1814 * Controllers from default hierarchy that need to be rebound
1815 * are all disabled together in one go.
1816 */
1817 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1818 WARN_ON(cgroup_apply_control(scgrp));
1819 cgroup_finalize_control(scgrp, 0);
1820 }
1821
1822 do_each_subsys_mask(ss, ssid, ss_mask) {
1823 struct cgroup_root *src_root = ss->root;
1824 struct cgroup *scgrp = &src_root->cgrp;
1825 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1826 struct css_set *cset, *cset_pos;
1827 struct css_task_iter *it;
1828
1829 WARN_ON(!css || cgroup_css(dcgrp, ss));
1830
1831 if (src_root != &cgrp_dfl_root) {
1832 /* disable from the source */
1833 src_root->subsys_mask &= ~(1 << ssid);
1834 WARN_ON(cgroup_apply_control(scgrp));
1835 cgroup_finalize_control(scgrp, 0);
1836 }
1837
1838 /* rebind */
1839 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1840 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1841 ss->root = dst_root;
1842 css->cgroup = dcgrp;
1843
1844 spin_lock_irq(&css_set_lock);
1845 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id]));
1846 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id],
1847 e_cset_node[ss->id]) {
1848 list_move_tail(&cset->e_cset_node[ss->id],
1849 &dcgrp->e_csets[ss->id]);
1850 /*
1851 * all css_sets of scgrp together in same order to dcgrp,
1852 * patch in-flight iterators to preserve correct iteration.
1853 * since the iterator is always advanced right away and
1854 * finished when it->cset_pos meets it->cset_head, so only
1855 * update it->cset_head is enough here.
1856 */
1857 list_for_each_entry(it, &cset->task_iters, iters_node)
1858 if (it->cset_head == &scgrp->e_csets[ss->id])
1859 it->cset_head = &dcgrp->e_csets[ss->id];
1860 }
1861 spin_unlock_irq(&css_set_lock);
1862
1863 if (ss->css_rstat_flush) {
1864 list_del_rcu(&css->rstat_css_node);
1865 synchronize_rcu();
1866 list_add_rcu(&css->rstat_css_node,
1867 &dcgrp->rstat_css_list);
1868 }
1869
1870 /* default hierarchy doesn't enable controllers by default */
1871 dst_root->subsys_mask |= 1 << ssid;
1872 if (dst_root == &cgrp_dfl_root) {
1873 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1874 } else {
1875 dcgrp->subtree_control |= 1 << ssid;
1876 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1877 }
1878
1879 ret = cgroup_apply_control(dcgrp);
1880 if (ret)
1881 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1882 ss->name, ret);
1883
1884 if (ss->bind)
1885 ss->bind(css);
1886 } while_each_subsys_mask();
1887
1888 kernfs_activate(dcgrp->kn);
1889 return 0;
1890}
1891
1892int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1893 struct kernfs_root *kf_root)
1894{
1895 int len = 0;
1896 char *buf = NULL;
1897 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1898 struct cgroup *ns_cgroup;
1899
1900 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1901 if (!buf)
1902 return -ENOMEM;
1903
1904 spin_lock_irq(&css_set_lock);
1905 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1906 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1907 spin_unlock_irq(&css_set_lock);
1908
1909 if (len == -E2BIG)
1910 len = -ERANGE;
1911 else if (len > 0) {
1912 seq_escape(sf, buf, " \t\n\\");
1913 len = 0;
1914 }
1915 kfree(buf);
1916 return len;
1917}
1918
1919enum cgroup2_param {
1920 Opt_nsdelegate,
1921 Opt_favordynmods,
1922 Opt_memory_localevents,
1923 Opt_memory_recursiveprot,
1924 Opt_memory_hugetlb_accounting,
1925 nr__cgroup2_params
1926};
1927
1928static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1929 fsparam_flag("nsdelegate", Opt_nsdelegate),
1930 fsparam_flag("favordynmods", Opt_favordynmods),
1931 fsparam_flag("memory_localevents", Opt_memory_localevents),
1932 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1933 fsparam_flag("memory_hugetlb_accounting", Opt_memory_hugetlb_accounting),
1934 {}
1935};
1936
1937static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1938{
1939 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1940 struct fs_parse_result result;
1941 int opt;
1942
1943 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1944 if (opt < 0)
1945 return opt;
1946
1947 switch (opt) {
1948 case Opt_nsdelegate:
1949 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1950 return 0;
1951 case Opt_favordynmods:
1952 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1953 return 0;
1954 case Opt_memory_localevents:
1955 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1956 return 0;
1957 case Opt_memory_recursiveprot:
1958 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1959 return 0;
1960 case Opt_memory_hugetlb_accounting:
1961 ctx->flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1962 return 0;
1963 }
1964 return -EINVAL;
1965}
1966
1967static void apply_cgroup_root_flags(unsigned int root_flags)
1968{
1969 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1970 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1971 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1972 else
1973 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1974
1975 cgroup_favor_dynmods(&cgrp_dfl_root,
1976 root_flags & CGRP_ROOT_FAVOR_DYNMODS);
1977
1978 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1979 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1980 else
1981 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1982
1983 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1984 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1985 else
1986 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1987
1988 if (root_flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
1989 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1990 else
1991 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1992 }
1993}
1994
1995static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1996{
1997 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1998 seq_puts(seq, ",nsdelegate");
1999 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
2000 seq_puts(seq, ",favordynmods");
2001 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
2002 seq_puts(seq, ",memory_localevents");
2003 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
2004 seq_puts(seq, ",memory_recursiveprot");
2005 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
2006 seq_puts(seq, ",memory_hugetlb_accounting");
2007 return 0;
2008}
2009
2010static int cgroup_reconfigure(struct fs_context *fc)
2011{
2012 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2013
2014 apply_cgroup_root_flags(ctx->flags);
2015 return 0;
2016}
2017
2018static void init_cgroup_housekeeping(struct cgroup *cgrp)
2019{
2020 struct cgroup_subsys *ss;
2021 int ssid;
2022
2023 INIT_LIST_HEAD(&cgrp->self.sibling);
2024 INIT_LIST_HEAD(&cgrp->self.children);
2025 INIT_LIST_HEAD(&cgrp->cset_links);
2026 INIT_LIST_HEAD(&cgrp->pidlists);
2027 mutex_init(&cgrp->pidlist_mutex);
2028 cgrp->self.cgroup = cgrp;
2029 cgrp->self.flags |= CSS_ONLINE;
2030 cgrp->dom_cgrp = cgrp;
2031 cgrp->max_descendants = INT_MAX;
2032 cgrp->max_depth = INT_MAX;
2033 INIT_LIST_HEAD(&cgrp->rstat_css_list);
2034 prev_cputime_init(&cgrp->prev_cputime);
2035
2036 for_each_subsys(ss, ssid)
2037 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2038
2039 init_waitqueue_head(&cgrp->offline_waitq);
2040 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2041}
2042
2043void init_cgroup_root(struct cgroup_fs_context *ctx)
2044{
2045 struct cgroup_root *root = ctx->root;
2046 struct cgroup *cgrp = &root->cgrp;
2047
2048 INIT_LIST_HEAD_RCU(&root->root_list);
2049 atomic_set(&root->nr_cgrps, 1);
2050 cgrp->root = root;
2051 init_cgroup_housekeeping(cgrp);
2052
2053 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2054 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2055 if (ctx->release_agent)
2056 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2057 if (ctx->name)
2058 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2059 if (ctx->cpuset_clone_children)
2060 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2061}
2062
2063int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2064{
2065 LIST_HEAD(tmp_links);
2066 struct cgroup *root_cgrp = &root->cgrp;
2067 struct kernfs_syscall_ops *kf_sops;
2068 struct css_set *cset;
2069 int i, ret;
2070
2071 lockdep_assert_held(&cgroup_mutex);
2072
2073 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2074 0, GFP_KERNEL);
2075 if (ret)
2076 goto out;
2077
2078 /*
2079 * We're accessing css_set_count without locking css_set_lock here,
2080 * but that's OK - it can only be increased by someone holding
2081 * cgroup_lock, and that's us. Later rebinding may disable
2082 * controllers on the default hierarchy and thus create new csets,
2083 * which can't be more than the existing ones. Allocate 2x.
2084 */
2085 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2086 if (ret)
2087 goto cancel_ref;
2088
2089 ret = cgroup_init_root_id(root);
2090 if (ret)
2091 goto cancel_ref;
2092
2093 kf_sops = root == &cgrp_dfl_root ?
2094 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2095
2096 root->kf_root = kernfs_create_root(kf_sops,
2097 KERNFS_ROOT_CREATE_DEACTIVATED |
2098 KERNFS_ROOT_SUPPORT_EXPORTOP |
2099 KERNFS_ROOT_SUPPORT_USER_XATTR,
2100 root_cgrp);
2101 if (IS_ERR(root->kf_root)) {
2102 ret = PTR_ERR(root->kf_root);
2103 goto exit_root_id;
2104 }
2105 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2106 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2107 root_cgrp->ancestors[0] = root_cgrp;
2108
2109 ret = css_populate_dir(&root_cgrp->self);
2110 if (ret)
2111 goto destroy_root;
2112
2113 ret = cgroup_rstat_init(root_cgrp);
2114 if (ret)
2115 goto destroy_root;
2116
2117 ret = rebind_subsystems(root, ss_mask);
2118 if (ret)
2119 goto exit_stats;
2120
2121 ret = cgroup_bpf_inherit(root_cgrp);
2122 WARN_ON_ONCE(ret);
2123
2124 trace_cgroup_setup_root(root);
2125
2126 /*
2127 * There must be no failure case after here, since rebinding takes
2128 * care of subsystems' refcounts, which are explicitly dropped in
2129 * the failure exit path.
2130 */
2131 list_add_rcu(&root->root_list, &cgroup_roots);
2132 cgroup_root_count++;
2133
2134 /*
2135 * Link the root cgroup in this hierarchy into all the css_set
2136 * objects.
2137 */
2138 spin_lock_irq(&css_set_lock);
2139 hash_for_each(css_set_table, i, cset, hlist) {
2140 link_css_set(&tmp_links, cset, root_cgrp);
2141 if (css_set_populated(cset))
2142 cgroup_update_populated(root_cgrp, true);
2143 }
2144 spin_unlock_irq(&css_set_lock);
2145
2146 BUG_ON(!list_empty(&root_cgrp->self.children));
2147 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2148
2149 ret = 0;
2150 goto out;
2151
2152exit_stats:
2153 cgroup_rstat_exit(root_cgrp);
2154destroy_root:
2155 kernfs_destroy_root(root->kf_root);
2156 root->kf_root = NULL;
2157exit_root_id:
2158 cgroup_exit_root_id(root);
2159cancel_ref:
2160 percpu_ref_exit(&root_cgrp->self.refcnt);
2161out:
2162 free_cgrp_cset_links(&tmp_links);
2163 return ret;
2164}
2165
2166int cgroup_do_get_tree(struct fs_context *fc)
2167{
2168 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2169 int ret;
2170
2171 ctx->kfc.root = ctx->root->kf_root;
2172 if (fc->fs_type == &cgroup2_fs_type)
2173 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2174 else
2175 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2176 ret = kernfs_get_tree(fc);
2177
2178 /*
2179 * In non-init cgroup namespace, instead of root cgroup's dentry,
2180 * we return the dentry corresponding to the cgroupns->root_cgrp.
2181 */
2182 if (!ret && ctx->ns != &init_cgroup_ns) {
2183 struct dentry *nsdentry;
2184 struct super_block *sb = fc->root->d_sb;
2185 struct cgroup *cgrp;
2186
2187 cgroup_lock();
2188 spin_lock_irq(&css_set_lock);
2189
2190 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2191
2192 spin_unlock_irq(&css_set_lock);
2193 cgroup_unlock();
2194
2195 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2196 dput(fc->root);
2197 if (IS_ERR(nsdentry)) {
2198 deactivate_locked_super(sb);
2199 ret = PTR_ERR(nsdentry);
2200 nsdentry = NULL;
2201 }
2202 fc->root = nsdentry;
2203 }
2204
2205 if (!ctx->kfc.new_sb_created)
2206 cgroup_put(&ctx->root->cgrp);
2207
2208 return ret;
2209}
2210
2211/*
2212 * Destroy a cgroup filesystem context.
2213 */
2214static void cgroup_fs_context_free(struct fs_context *fc)
2215{
2216 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2217
2218 kfree(ctx->name);
2219 kfree(ctx->release_agent);
2220 put_cgroup_ns(ctx->ns);
2221 kernfs_free_fs_context(fc);
2222 kfree(ctx);
2223}
2224
2225static int cgroup_get_tree(struct fs_context *fc)
2226{
2227 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2228 int ret;
2229
2230 WRITE_ONCE(cgrp_dfl_visible, true);
2231 cgroup_get_live(&cgrp_dfl_root.cgrp);
2232 ctx->root = &cgrp_dfl_root;
2233
2234 ret = cgroup_do_get_tree(fc);
2235 if (!ret)
2236 apply_cgroup_root_flags(ctx->flags);
2237 return ret;
2238}
2239
2240static const struct fs_context_operations cgroup_fs_context_ops = {
2241 .free = cgroup_fs_context_free,
2242 .parse_param = cgroup2_parse_param,
2243 .get_tree = cgroup_get_tree,
2244 .reconfigure = cgroup_reconfigure,
2245};
2246
2247static const struct fs_context_operations cgroup1_fs_context_ops = {
2248 .free = cgroup_fs_context_free,
2249 .parse_param = cgroup1_parse_param,
2250 .get_tree = cgroup1_get_tree,
2251 .reconfigure = cgroup1_reconfigure,
2252};
2253
2254/*
2255 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2256 * we select the namespace we're going to use.
2257 */
2258static int cgroup_init_fs_context(struct fs_context *fc)
2259{
2260 struct cgroup_fs_context *ctx;
2261
2262 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2263 if (!ctx)
2264 return -ENOMEM;
2265
2266 ctx->ns = current->nsproxy->cgroup_ns;
2267 get_cgroup_ns(ctx->ns);
2268 fc->fs_private = &ctx->kfc;
2269 if (fc->fs_type == &cgroup2_fs_type)
2270 fc->ops = &cgroup_fs_context_ops;
2271 else
2272 fc->ops = &cgroup1_fs_context_ops;
2273 put_user_ns(fc->user_ns);
2274 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2275 fc->global = true;
2276
2277 if (have_favordynmods)
2278 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2279
2280 return 0;
2281}
2282
2283static void cgroup_kill_sb(struct super_block *sb)
2284{
2285 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2286 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2287
2288 /*
2289 * If @root doesn't have any children, start killing it.
2290 * This prevents new mounts by disabling percpu_ref_tryget_live().
2291 *
2292 * And don't kill the default root.
2293 */
2294 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2295 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) {
2296 cgroup_bpf_offline(&root->cgrp);
2297 percpu_ref_kill(&root->cgrp.self.refcnt);
2298 }
2299 cgroup_put(&root->cgrp);
2300 kernfs_kill_sb(sb);
2301}
2302
2303struct file_system_type cgroup_fs_type = {
2304 .name = "cgroup",
2305 .init_fs_context = cgroup_init_fs_context,
2306 .parameters = cgroup1_fs_parameters,
2307 .kill_sb = cgroup_kill_sb,
2308 .fs_flags = FS_USERNS_MOUNT,
2309};
2310
2311static struct file_system_type cgroup2_fs_type = {
2312 .name = "cgroup2",
2313 .init_fs_context = cgroup_init_fs_context,
2314 .parameters = cgroup2_fs_parameters,
2315 .kill_sb = cgroup_kill_sb,
2316 .fs_flags = FS_USERNS_MOUNT,
2317};
2318
2319#ifdef CONFIG_CPUSETS
2320static const struct fs_context_operations cpuset_fs_context_ops = {
2321 .get_tree = cgroup1_get_tree,
2322 .free = cgroup_fs_context_free,
2323};
2324
2325/*
2326 * This is ugly, but preserves the userspace API for existing cpuset
2327 * users. If someone tries to mount the "cpuset" filesystem, we
2328 * silently switch it to mount "cgroup" instead
2329 */
2330static int cpuset_init_fs_context(struct fs_context *fc)
2331{
2332 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2333 struct cgroup_fs_context *ctx;
2334 int err;
2335
2336 err = cgroup_init_fs_context(fc);
2337 if (err) {
2338 kfree(agent);
2339 return err;
2340 }
2341
2342 fc->ops = &cpuset_fs_context_ops;
2343
2344 ctx = cgroup_fc2context(fc);
2345 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2346 ctx->flags |= CGRP_ROOT_NOPREFIX;
2347 ctx->release_agent = agent;
2348
2349 get_filesystem(&cgroup_fs_type);
2350 put_filesystem(fc->fs_type);
2351 fc->fs_type = &cgroup_fs_type;
2352
2353 return 0;
2354}
2355
2356static struct file_system_type cpuset_fs_type = {
2357 .name = "cpuset",
2358 .init_fs_context = cpuset_init_fs_context,
2359 .fs_flags = FS_USERNS_MOUNT,
2360};
2361#endif
2362
2363int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2364 struct cgroup_namespace *ns)
2365{
2366 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2367
2368 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2369}
2370
2371int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2372 struct cgroup_namespace *ns)
2373{
2374 int ret;
2375
2376 cgroup_lock();
2377 spin_lock_irq(&css_set_lock);
2378
2379 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2380
2381 spin_unlock_irq(&css_set_lock);
2382 cgroup_unlock();
2383
2384 return ret;
2385}
2386EXPORT_SYMBOL_GPL(cgroup_path_ns);
2387
2388/**
2389 * cgroup_attach_lock - Lock for ->attach()
2390 * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2391 *
2392 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2393 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2394 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2395 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2396 * lead to deadlocks.
2397 *
2398 * Bringing up a CPU may involve creating and destroying tasks which requires
2399 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2400 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2401 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2402 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2403 * the threadgroup_rwsem to be released to create new tasks. For more details:
2404 *
2405 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2406 *
2407 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2408 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2409 * CPU hotplug is disabled on entry.
2410 */
2411void cgroup_attach_lock(bool lock_threadgroup)
2412{
2413 cpus_read_lock();
2414 if (lock_threadgroup)
2415 percpu_down_write(&cgroup_threadgroup_rwsem);
2416}
2417
2418/**
2419 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2420 * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2421 */
2422void cgroup_attach_unlock(bool lock_threadgroup)
2423{
2424 if (lock_threadgroup)
2425 percpu_up_write(&cgroup_threadgroup_rwsem);
2426 cpus_read_unlock();
2427}
2428
2429/**
2430 * cgroup_migrate_add_task - add a migration target task to a migration context
2431 * @task: target task
2432 * @mgctx: target migration context
2433 *
2434 * Add @task, which is a migration target, to @mgctx->tset. This function
2435 * becomes noop if @task doesn't need to be migrated. @task's css_set
2436 * should have been added as a migration source and @task->cg_list will be
2437 * moved from the css_set's tasks list to mg_tasks one.
2438 */
2439static void cgroup_migrate_add_task(struct task_struct *task,
2440 struct cgroup_mgctx *mgctx)
2441{
2442 struct css_set *cset;
2443
2444 lockdep_assert_held(&css_set_lock);
2445
2446 /* @task either already exited or can't exit until the end */
2447 if (task->flags & PF_EXITING)
2448 return;
2449
2450 /* cgroup_threadgroup_rwsem protects racing against forks */
2451 WARN_ON_ONCE(list_empty(&task->cg_list));
2452
2453 cset = task_css_set(task);
2454 if (!cset->mg_src_cgrp)
2455 return;
2456
2457 mgctx->tset.nr_tasks++;
2458
2459 list_move_tail(&task->cg_list, &cset->mg_tasks);
2460 if (list_empty(&cset->mg_node))
2461 list_add_tail(&cset->mg_node,
2462 &mgctx->tset.src_csets);
2463 if (list_empty(&cset->mg_dst_cset->mg_node))
2464 list_add_tail(&cset->mg_dst_cset->mg_node,
2465 &mgctx->tset.dst_csets);
2466}
2467
2468/**
2469 * cgroup_taskset_first - reset taskset and return the first task
2470 * @tset: taskset of interest
2471 * @dst_cssp: output variable for the destination css
2472 *
2473 * @tset iteration is initialized and the first task is returned.
2474 */
2475struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2476 struct cgroup_subsys_state **dst_cssp)
2477{
2478 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2479 tset->cur_task = NULL;
2480
2481 return cgroup_taskset_next(tset, dst_cssp);
2482}
2483
2484/**
2485 * cgroup_taskset_next - iterate to the next task in taskset
2486 * @tset: taskset of interest
2487 * @dst_cssp: output variable for the destination css
2488 *
2489 * Return the next task in @tset. Iteration must have been initialized
2490 * with cgroup_taskset_first().
2491 */
2492struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2493 struct cgroup_subsys_state **dst_cssp)
2494{
2495 struct css_set *cset = tset->cur_cset;
2496 struct task_struct *task = tset->cur_task;
2497
2498 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2499 if (!task)
2500 task = list_first_entry(&cset->mg_tasks,
2501 struct task_struct, cg_list);
2502 else
2503 task = list_next_entry(task, cg_list);
2504
2505 if (&task->cg_list != &cset->mg_tasks) {
2506 tset->cur_cset = cset;
2507 tset->cur_task = task;
2508
2509 /*
2510 * This function may be called both before and
2511 * after cgroup_migrate_execute(). The two cases
2512 * can be distinguished by looking at whether @cset
2513 * has its ->mg_dst_cset set.
2514 */
2515 if (cset->mg_dst_cset)
2516 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2517 else
2518 *dst_cssp = cset->subsys[tset->ssid];
2519
2520 return task;
2521 }
2522
2523 cset = list_next_entry(cset, mg_node);
2524 task = NULL;
2525 }
2526
2527 return NULL;
2528}
2529
2530/**
2531 * cgroup_migrate_execute - migrate a taskset
2532 * @mgctx: migration context
2533 *
2534 * Migrate tasks in @mgctx as setup by migration preparation functions.
2535 * This function fails iff one of the ->can_attach callbacks fails and
2536 * guarantees that either all or none of the tasks in @mgctx are migrated.
2537 * @mgctx is consumed regardless of success.
2538 */
2539static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2540{
2541 struct cgroup_taskset *tset = &mgctx->tset;
2542 struct cgroup_subsys *ss;
2543 struct task_struct *task, *tmp_task;
2544 struct css_set *cset, *tmp_cset;
2545 int ssid, failed_ssid, ret;
2546
2547 /* check that we can legitimately attach to the cgroup */
2548 if (tset->nr_tasks) {
2549 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2550 if (ss->can_attach) {
2551 tset->ssid = ssid;
2552 ret = ss->can_attach(tset);
2553 if (ret) {
2554 failed_ssid = ssid;
2555 goto out_cancel_attach;
2556 }
2557 }
2558 } while_each_subsys_mask();
2559 }
2560
2561 /*
2562 * Now that we're guaranteed success, proceed to move all tasks to
2563 * the new cgroup. There are no failure cases after here, so this
2564 * is the commit point.
2565 */
2566 spin_lock_irq(&css_set_lock);
2567 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2568 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2569 struct css_set *from_cset = task_css_set(task);
2570 struct css_set *to_cset = cset->mg_dst_cset;
2571
2572 get_css_set(to_cset);
2573 to_cset->nr_tasks++;
2574 css_set_move_task(task, from_cset, to_cset, true);
2575 from_cset->nr_tasks--;
2576 /*
2577 * If the source or destination cgroup is frozen,
2578 * the task might require to change its state.
2579 */
2580 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2581 to_cset->dfl_cgrp);
2582 put_css_set_locked(from_cset);
2583
2584 }
2585 }
2586 spin_unlock_irq(&css_set_lock);
2587
2588 /*
2589 * Migration is committed, all target tasks are now on dst_csets.
2590 * Nothing is sensitive to fork() after this point. Notify
2591 * controllers that migration is complete.
2592 */
2593 tset->csets = &tset->dst_csets;
2594
2595 if (tset->nr_tasks) {
2596 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2597 if (ss->attach) {
2598 tset->ssid = ssid;
2599 ss->attach(tset);
2600 }
2601 } while_each_subsys_mask();
2602 }
2603
2604 ret = 0;
2605 goto out_release_tset;
2606
2607out_cancel_attach:
2608 if (tset->nr_tasks) {
2609 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2610 if (ssid == failed_ssid)
2611 break;
2612 if (ss->cancel_attach) {
2613 tset->ssid = ssid;
2614 ss->cancel_attach(tset);
2615 }
2616 } while_each_subsys_mask();
2617 }
2618out_release_tset:
2619 spin_lock_irq(&css_set_lock);
2620 list_splice_init(&tset->dst_csets, &tset->src_csets);
2621 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2622 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2623 list_del_init(&cset->mg_node);
2624 }
2625 spin_unlock_irq(&css_set_lock);
2626
2627 /*
2628 * Re-initialize the cgroup_taskset structure in case it is reused
2629 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2630 * iteration.
2631 */
2632 tset->nr_tasks = 0;
2633 tset->csets = &tset->src_csets;
2634 return ret;
2635}
2636
2637/**
2638 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2639 * @dst_cgrp: destination cgroup to test
2640 *
2641 * On the default hierarchy, except for the mixable, (possible) thread root
2642 * and threaded cgroups, subtree_control must be zero for migration
2643 * destination cgroups with tasks so that child cgroups don't compete
2644 * against tasks.
2645 */
2646int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2647{
2648 /* v1 doesn't have any restriction */
2649 if (!cgroup_on_dfl(dst_cgrp))
2650 return 0;
2651
2652 /* verify @dst_cgrp can host resources */
2653 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2654 return -EOPNOTSUPP;
2655
2656 /*
2657 * If @dst_cgrp is already or can become a thread root or is
2658 * threaded, it doesn't matter.
2659 */
2660 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2661 return 0;
2662
2663 /* apply no-internal-process constraint */
2664 if (dst_cgrp->subtree_control)
2665 return -EBUSY;
2666
2667 return 0;
2668}
2669
2670/**
2671 * cgroup_migrate_finish - cleanup after attach
2672 * @mgctx: migration context
2673 *
2674 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2675 * those functions for details.
2676 */
2677void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2678{
2679 struct css_set *cset, *tmp_cset;
2680
2681 lockdep_assert_held(&cgroup_mutex);
2682
2683 spin_lock_irq(&css_set_lock);
2684
2685 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2686 mg_src_preload_node) {
2687 cset->mg_src_cgrp = NULL;
2688 cset->mg_dst_cgrp = NULL;
2689 cset->mg_dst_cset = NULL;
2690 list_del_init(&cset->mg_src_preload_node);
2691 put_css_set_locked(cset);
2692 }
2693
2694 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2695 mg_dst_preload_node) {
2696 cset->mg_src_cgrp = NULL;
2697 cset->mg_dst_cgrp = NULL;
2698 cset->mg_dst_cset = NULL;
2699 list_del_init(&cset->mg_dst_preload_node);
2700 put_css_set_locked(cset);
2701 }
2702
2703 spin_unlock_irq(&css_set_lock);
2704}
2705
2706/**
2707 * cgroup_migrate_add_src - add a migration source css_set
2708 * @src_cset: the source css_set to add
2709 * @dst_cgrp: the destination cgroup
2710 * @mgctx: migration context
2711 *
2712 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2713 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2714 * up by cgroup_migrate_finish().
2715 *
2716 * This function may be called without holding cgroup_threadgroup_rwsem
2717 * even if the target is a process. Threads may be created and destroyed
2718 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2719 * into play and the preloaded css_sets are guaranteed to cover all
2720 * migrations.
2721 */
2722void cgroup_migrate_add_src(struct css_set *src_cset,
2723 struct cgroup *dst_cgrp,
2724 struct cgroup_mgctx *mgctx)
2725{
2726 struct cgroup *src_cgrp;
2727
2728 lockdep_assert_held(&cgroup_mutex);
2729 lockdep_assert_held(&css_set_lock);
2730
2731 /*
2732 * If ->dead, @src_set is associated with one or more dead cgroups
2733 * and doesn't contain any migratable tasks. Ignore it early so
2734 * that the rest of migration path doesn't get confused by it.
2735 */
2736 if (src_cset->dead)
2737 return;
2738
2739 if (!list_empty(&src_cset->mg_src_preload_node))
2740 return;
2741
2742 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2743
2744 WARN_ON(src_cset->mg_src_cgrp);
2745 WARN_ON(src_cset->mg_dst_cgrp);
2746 WARN_ON(!list_empty(&src_cset->mg_tasks));
2747 WARN_ON(!list_empty(&src_cset->mg_node));
2748
2749 src_cset->mg_src_cgrp = src_cgrp;
2750 src_cset->mg_dst_cgrp = dst_cgrp;
2751 get_css_set(src_cset);
2752 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2753}
2754
2755/**
2756 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2757 * @mgctx: migration context
2758 *
2759 * Tasks are about to be moved and all the source css_sets have been
2760 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2761 * pins all destination css_sets, links each to its source, and append them
2762 * to @mgctx->preloaded_dst_csets.
2763 *
2764 * This function must be called after cgroup_migrate_add_src() has been
2765 * called on each migration source css_set. After migration is performed
2766 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2767 * @mgctx.
2768 */
2769int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2770{
2771 struct css_set *src_cset, *tmp_cset;
2772
2773 lockdep_assert_held(&cgroup_mutex);
2774
2775 /* look up the dst cset for each src cset and link it to src */
2776 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2777 mg_src_preload_node) {
2778 struct css_set *dst_cset;
2779 struct cgroup_subsys *ss;
2780 int ssid;
2781
2782 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2783 if (!dst_cset)
2784 return -ENOMEM;
2785
2786 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2787
2788 /*
2789 * If src cset equals dst, it's noop. Drop the src.
2790 * cgroup_migrate() will skip the cset too. Note that we
2791 * can't handle src == dst as some nodes are used by both.
2792 */
2793 if (src_cset == dst_cset) {
2794 src_cset->mg_src_cgrp = NULL;
2795 src_cset->mg_dst_cgrp = NULL;
2796 list_del_init(&src_cset->mg_src_preload_node);
2797 put_css_set(src_cset);
2798 put_css_set(dst_cset);
2799 continue;
2800 }
2801
2802 src_cset->mg_dst_cset = dst_cset;
2803
2804 if (list_empty(&dst_cset->mg_dst_preload_node))
2805 list_add_tail(&dst_cset->mg_dst_preload_node,
2806 &mgctx->preloaded_dst_csets);
2807 else
2808 put_css_set(dst_cset);
2809
2810 for_each_subsys(ss, ssid)
2811 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2812 mgctx->ss_mask |= 1 << ssid;
2813 }
2814
2815 return 0;
2816}
2817
2818/**
2819 * cgroup_migrate - migrate a process or task to a cgroup
2820 * @leader: the leader of the process or the task to migrate
2821 * @threadgroup: whether @leader points to the whole process or a single task
2822 * @mgctx: migration context
2823 *
2824 * Migrate a process or task denoted by @leader. If migrating a process,
2825 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2826 * responsible for invoking cgroup_migrate_add_src() and
2827 * cgroup_migrate_prepare_dst() on the targets before invoking this
2828 * function and following up with cgroup_migrate_finish().
2829 *
2830 * As long as a controller's ->can_attach() doesn't fail, this function is
2831 * guaranteed to succeed. This means that, excluding ->can_attach()
2832 * failure, when migrating multiple targets, the success or failure can be
2833 * decided for all targets by invoking group_migrate_prepare_dst() before
2834 * actually starting migrating.
2835 */
2836int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2837 struct cgroup_mgctx *mgctx)
2838{
2839 struct task_struct *task;
2840
2841 /*
2842 * The following thread iteration should be inside an RCU critical
2843 * section to prevent tasks from being freed while taking the snapshot.
2844 * spin_lock_irq() implies RCU critical section here.
2845 */
2846 spin_lock_irq(&css_set_lock);
2847 task = leader;
2848 do {
2849 cgroup_migrate_add_task(task, mgctx);
2850 if (!threadgroup)
2851 break;
2852 } while_each_thread(leader, task);
2853 spin_unlock_irq(&css_set_lock);
2854
2855 return cgroup_migrate_execute(mgctx);
2856}
2857
2858/**
2859 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2860 * @dst_cgrp: the cgroup to attach to
2861 * @leader: the task or the leader of the threadgroup to be attached
2862 * @threadgroup: attach the whole threadgroup?
2863 *
2864 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2865 */
2866int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2867 bool threadgroup)
2868{
2869 DEFINE_CGROUP_MGCTX(mgctx);
2870 struct task_struct *task;
2871 int ret = 0;
2872
2873 /* look up all src csets */
2874 spin_lock_irq(&css_set_lock);
2875 rcu_read_lock();
2876 task = leader;
2877 do {
2878 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2879 if (!threadgroup)
2880 break;
2881 } while_each_thread(leader, task);
2882 rcu_read_unlock();
2883 spin_unlock_irq(&css_set_lock);
2884
2885 /* prepare dst csets and commit */
2886 ret = cgroup_migrate_prepare_dst(&mgctx);
2887 if (!ret)
2888 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2889
2890 cgroup_migrate_finish(&mgctx);
2891
2892 if (!ret)
2893 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2894
2895 return ret;
2896}
2897
2898struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2899 bool *threadgroup_locked)
2900{
2901 struct task_struct *tsk;
2902 pid_t pid;
2903
2904 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2905 return ERR_PTR(-EINVAL);
2906
2907 /*
2908 * If we migrate a single thread, we don't care about threadgroup
2909 * stability. If the thread is `current`, it won't exit(2) under our
2910 * hands or change PID through exec(2). We exclude
2911 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2912 * callers by cgroup_mutex.
2913 * Therefore, we can skip the global lock.
2914 */
2915 lockdep_assert_held(&cgroup_mutex);
2916 *threadgroup_locked = pid || threadgroup;
2917 cgroup_attach_lock(*threadgroup_locked);
2918
2919 rcu_read_lock();
2920 if (pid) {
2921 tsk = find_task_by_vpid(pid);
2922 if (!tsk) {
2923 tsk = ERR_PTR(-ESRCH);
2924 goto out_unlock_threadgroup;
2925 }
2926 } else {
2927 tsk = current;
2928 }
2929
2930 if (threadgroup)
2931 tsk = tsk->group_leader;
2932
2933 /*
2934 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2935 * If userland migrates such a kthread to a non-root cgroup, it can
2936 * become trapped in a cpuset, or RT kthread may be born in a
2937 * cgroup with no rt_runtime allocated. Just say no.
2938 */
2939 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2940 tsk = ERR_PTR(-EINVAL);
2941 goto out_unlock_threadgroup;
2942 }
2943
2944 get_task_struct(tsk);
2945 goto out_unlock_rcu;
2946
2947out_unlock_threadgroup:
2948 cgroup_attach_unlock(*threadgroup_locked);
2949 *threadgroup_locked = false;
2950out_unlock_rcu:
2951 rcu_read_unlock();
2952 return tsk;
2953}
2954
2955void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2956{
2957 struct cgroup_subsys *ss;
2958 int ssid;
2959
2960 /* release reference from cgroup_procs_write_start() */
2961 put_task_struct(task);
2962
2963 cgroup_attach_unlock(threadgroup_locked);
2964
2965 for_each_subsys(ss, ssid)
2966 if (ss->post_attach)
2967 ss->post_attach();
2968}
2969
2970static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2971{
2972 struct cgroup_subsys *ss;
2973 bool printed = false;
2974 int ssid;
2975
2976 do_each_subsys_mask(ss, ssid, ss_mask) {
2977 if (printed)
2978 seq_putc(seq, ' ');
2979 seq_puts(seq, ss->name);
2980 printed = true;
2981 } while_each_subsys_mask();
2982 if (printed)
2983 seq_putc(seq, '\n');
2984}
2985
2986/* show controllers which are enabled from the parent */
2987static int cgroup_controllers_show(struct seq_file *seq, void *v)
2988{
2989 struct cgroup *cgrp = seq_css(seq)->cgroup;
2990
2991 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2992 return 0;
2993}
2994
2995/* show controllers which are enabled for a given cgroup's children */
2996static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2997{
2998 struct cgroup *cgrp = seq_css(seq)->cgroup;
2999
3000 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3001 return 0;
3002}
3003
3004/**
3005 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3006 * @cgrp: root of the subtree to update csses for
3007 *
3008 * @cgrp's control masks have changed and its subtree's css associations
3009 * need to be updated accordingly. This function looks up all css_sets
3010 * which are attached to the subtree, creates the matching updated css_sets
3011 * and migrates the tasks to the new ones.
3012 */
3013static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3014{
3015 DEFINE_CGROUP_MGCTX(mgctx);
3016 struct cgroup_subsys_state *d_css;
3017 struct cgroup *dsct;
3018 struct css_set *src_cset;
3019 bool has_tasks;
3020 int ret;
3021
3022 lockdep_assert_held(&cgroup_mutex);
3023
3024 /* look up all csses currently attached to @cgrp's subtree */
3025 spin_lock_irq(&css_set_lock);
3026 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3027 struct cgrp_cset_link *link;
3028
3029 /*
3030 * As cgroup_update_dfl_csses() is only called by
3031 * cgroup_apply_control(). The csses associated with the
3032 * given cgrp will not be affected by changes made to
3033 * its subtree_control file. We can skip them.
3034 */
3035 if (dsct == cgrp)
3036 continue;
3037
3038 list_for_each_entry(link, &dsct->cset_links, cset_link)
3039 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3040 }
3041 spin_unlock_irq(&css_set_lock);
3042
3043 /*
3044 * We need to write-lock threadgroup_rwsem while migrating tasks.
3045 * However, if there are no source csets for @cgrp, changing its
3046 * controllers isn't gonna produce any task migrations and the
3047 * write-locking can be skipped safely.
3048 */
3049 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3050 cgroup_attach_lock(has_tasks);
3051
3052 /* NULL dst indicates self on default hierarchy */
3053 ret = cgroup_migrate_prepare_dst(&mgctx);
3054 if (ret)
3055 goto out_finish;
3056
3057 spin_lock_irq(&css_set_lock);
3058 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3059 mg_src_preload_node) {
3060 struct task_struct *task, *ntask;
3061
3062 /* all tasks in src_csets need to be migrated */
3063 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3064 cgroup_migrate_add_task(task, &mgctx);
3065 }
3066 spin_unlock_irq(&css_set_lock);
3067
3068 ret = cgroup_migrate_execute(&mgctx);
3069out_finish:
3070 cgroup_migrate_finish(&mgctx);
3071 cgroup_attach_unlock(has_tasks);
3072 return ret;
3073}
3074
3075/**
3076 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3077 * @cgrp: root of the target subtree
3078 *
3079 * Because css offlining is asynchronous, userland may try to re-enable a
3080 * controller while the previous css is still around. This function grabs
3081 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3082 */
3083void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3084 __acquires(&cgroup_mutex)
3085{
3086 struct cgroup *dsct;
3087 struct cgroup_subsys_state *d_css;
3088 struct cgroup_subsys *ss;
3089 int ssid;
3090
3091restart:
3092 cgroup_lock();
3093
3094 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3095 for_each_subsys(ss, ssid) {
3096 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3097 DEFINE_WAIT(wait);
3098
3099 if (!css || !percpu_ref_is_dying(&css->refcnt))
3100 continue;
3101
3102 cgroup_get_live(dsct);
3103 prepare_to_wait(&dsct->offline_waitq, &wait,
3104 TASK_UNINTERRUPTIBLE);
3105
3106 cgroup_unlock();
3107 schedule();
3108 finish_wait(&dsct->offline_waitq, &wait);
3109
3110 cgroup_put(dsct);
3111 goto restart;
3112 }
3113 }
3114}
3115
3116/**
3117 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3118 * @cgrp: root of the target subtree
3119 *
3120 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3121 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3122 * itself.
3123 */
3124static void cgroup_save_control(struct cgroup *cgrp)
3125{
3126 struct cgroup *dsct;
3127 struct cgroup_subsys_state *d_css;
3128
3129 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3130 dsct->old_subtree_control = dsct->subtree_control;
3131 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3132 dsct->old_dom_cgrp = dsct->dom_cgrp;
3133 }
3134}
3135
3136/**
3137 * cgroup_propagate_control - refresh control masks of a subtree
3138 * @cgrp: root of the target subtree
3139 *
3140 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3141 * ->subtree_control and propagate controller availability through the
3142 * subtree so that descendants don't have unavailable controllers enabled.
3143 */
3144static void cgroup_propagate_control(struct cgroup *cgrp)
3145{
3146 struct cgroup *dsct;
3147 struct cgroup_subsys_state *d_css;
3148
3149 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3150 dsct->subtree_control &= cgroup_control(dsct);
3151 dsct->subtree_ss_mask =
3152 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3153 cgroup_ss_mask(dsct));
3154 }
3155}
3156
3157/**
3158 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3159 * @cgrp: root of the target subtree
3160 *
3161 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3162 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3163 * itself.
3164 */
3165static void cgroup_restore_control(struct cgroup *cgrp)
3166{
3167 struct cgroup *dsct;
3168 struct cgroup_subsys_state *d_css;
3169
3170 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3171 dsct->subtree_control = dsct->old_subtree_control;
3172 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3173 dsct->dom_cgrp = dsct->old_dom_cgrp;
3174 }
3175}
3176
3177static bool css_visible(struct cgroup_subsys_state *css)
3178{
3179 struct cgroup_subsys *ss = css->ss;
3180 struct cgroup *cgrp = css->cgroup;
3181
3182 if (cgroup_control(cgrp) & (1 << ss->id))
3183 return true;
3184 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3185 return false;
3186 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3187}
3188
3189/**
3190 * cgroup_apply_control_enable - enable or show csses according to control
3191 * @cgrp: root of the target subtree
3192 *
3193 * Walk @cgrp's subtree and create new csses or make the existing ones
3194 * visible. A css is created invisible if it's being implicitly enabled
3195 * through dependency. An invisible css is made visible when the userland
3196 * explicitly enables it.
3197 *
3198 * Returns 0 on success, -errno on failure. On failure, csses which have
3199 * been processed already aren't cleaned up. The caller is responsible for
3200 * cleaning up with cgroup_apply_control_disable().
3201 */
3202static int cgroup_apply_control_enable(struct cgroup *cgrp)
3203{
3204 struct cgroup *dsct;
3205 struct cgroup_subsys_state *d_css;
3206 struct cgroup_subsys *ss;
3207 int ssid, ret;
3208
3209 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3210 for_each_subsys(ss, ssid) {
3211 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3212
3213 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3214 continue;
3215
3216 if (!css) {
3217 css = css_create(dsct, ss);
3218 if (IS_ERR(css))
3219 return PTR_ERR(css);
3220 }
3221
3222 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3223
3224 if (css_visible(css)) {
3225 ret = css_populate_dir(css);
3226 if (ret)
3227 return ret;
3228 }
3229 }
3230 }
3231
3232 return 0;
3233}
3234
3235/**
3236 * cgroup_apply_control_disable - kill or hide csses according to control
3237 * @cgrp: root of the target subtree
3238 *
3239 * Walk @cgrp's subtree and kill and hide csses so that they match
3240 * cgroup_ss_mask() and cgroup_visible_mask().
3241 *
3242 * A css is hidden when the userland requests it to be disabled while other
3243 * subsystems are still depending on it. The css must not actively control
3244 * resources and be in the vanilla state if it's made visible again later.
3245 * Controllers which may be depended upon should provide ->css_reset() for
3246 * this purpose.
3247 */
3248static void cgroup_apply_control_disable(struct cgroup *cgrp)
3249{
3250 struct cgroup *dsct;
3251 struct cgroup_subsys_state *d_css;
3252 struct cgroup_subsys *ss;
3253 int ssid;
3254
3255 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3256 for_each_subsys(ss, ssid) {
3257 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3258
3259 if (!css)
3260 continue;
3261
3262 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3263
3264 if (css->parent &&
3265 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3266 kill_css(css);
3267 } else if (!css_visible(css)) {
3268 css_clear_dir(css);
3269 if (ss->css_reset)
3270 ss->css_reset(css);
3271 }
3272 }
3273 }
3274}
3275
3276/**
3277 * cgroup_apply_control - apply control mask updates to the subtree
3278 * @cgrp: root of the target subtree
3279 *
3280 * subsystems can be enabled and disabled in a subtree using the following
3281 * steps.
3282 *
3283 * 1. Call cgroup_save_control() to stash the current state.
3284 * 2. Update ->subtree_control masks in the subtree as desired.
3285 * 3. Call cgroup_apply_control() to apply the changes.
3286 * 4. Optionally perform other related operations.
3287 * 5. Call cgroup_finalize_control() to finish up.
3288 *
3289 * This function implements step 3 and propagates the mask changes
3290 * throughout @cgrp's subtree, updates csses accordingly and perform
3291 * process migrations.
3292 */
3293static int cgroup_apply_control(struct cgroup *cgrp)
3294{
3295 int ret;
3296
3297 cgroup_propagate_control(cgrp);
3298
3299 ret = cgroup_apply_control_enable(cgrp);
3300 if (ret)
3301 return ret;
3302
3303 /*
3304 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3305 * making the following cgroup_update_dfl_csses() properly update
3306 * css associations of all tasks in the subtree.
3307 */
3308 return cgroup_update_dfl_csses(cgrp);
3309}
3310
3311/**
3312 * cgroup_finalize_control - finalize control mask update
3313 * @cgrp: root of the target subtree
3314 * @ret: the result of the update
3315 *
3316 * Finalize control mask update. See cgroup_apply_control() for more info.
3317 */
3318static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3319{
3320 if (ret) {
3321 cgroup_restore_control(cgrp);
3322 cgroup_propagate_control(cgrp);
3323 }
3324
3325 cgroup_apply_control_disable(cgrp);
3326}
3327
3328static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3329{
3330 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3331
3332 /* if nothing is getting enabled, nothing to worry about */
3333 if (!enable)
3334 return 0;
3335
3336 /* can @cgrp host any resources? */
3337 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3338 return -EOPNOTSUPP;
3339
3340 /* mixables don't care */
3341 if (cgroup_is_mixable(cgrp))
3342 return 0;
3343
3344 if (domain_enable) {
3345 /* can't enable domain controllers inside a thread subtree */
3346 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3347 return -EOPNOTSUPP;
3348 } else {
3349 /*
3350 * Threaded controllers can handle internal competitions
3351 * and are always allowed inside a (prospective) thread
3352 * subtree.
3353 */
3354 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3355 return 0;
3356 }
3357
3358 /*
3359 * Controllers can't be enabled for a cgroup with tasks to avoid
3360 * child cgroups competing against tasks.
3361 */
3362 if (cgroup_has_tasks(cgrp))
3363 return -EBUSY;
3364
3365 return 0;
3366}
3367
3368/* change the enabled child controllers for a cgroup in the default hierarchy */
3369static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3370 char *buf, size_t nbytes,
3371 loff_t off)
3372{
3373 u16 enable = 0, disable = 0;
3374 struct cgroup *cgrp, *child;
3375 struct cgroup_subsys *ss;
3376 char *tok;
3377 int ssid, ret;
3378
3379 /*
3380 * Parse input - space separated list of subsystem names prefixed
3381 * with either + or -.
3382 */
3383 buf = strstrip(buf);
3384 while ((tok = strsep(&buf, " "))) {
3385 if (tok[0] == '\0')
3386 continue;
3387 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3388 if (!cgroup_ssid_enabled(ssid) ||
3389 strcmp(tok + 1, ss->name))
3390 continue;
3391
3392 if (*tok == '+') {
3393 enable |= 1 << ssid;
3394 disable &= ~(1 << ssid);
3395 } else if (*tok == '-') {
3396 disable |= 1 << ssid;
3397 enable &= ~(1 << ssid);
3398 } else {
3399 return -EINVAL;
3400 }
3401 break;
3402 } while_each_subsys_mask();
3403 if (ssid == CGROUP_SUBSYS_COUNT)
3404 return -EINVAL;
3405 }
3406
3407 cgrp = cgroup_kn_lock_live(of->kn, true);
3408 if (!cgrp)
3409 return -ENODEV;
3410
3411 for_each_subsys(ss, ssid) {
3412 if (enable & (1 << ssid)) {
3413 if (cgrp->subtree_control & (1 << ssid)) {
3414 enable &= ~(1 << ssid);
3415 continue;
3416 }
3417
3418 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3419 ret = -ENOENT;
3420 goto out_unlock;
3421 }
3422 } else if (disable & (1 << ssid)) {
3423 if (!(cgrp->subtree_control & (1 << ssid))) {
3424 disable &= ~(1 << ssid);
3425 continue;
3426 }
3427
3428 /* a child has it enabled? */
3429 cgroup_for_each_live_child(child, cgrp) {
3430 if (child->subtree_control & (1 << ssid)) {
3431 ret = -EBUSY;
3432 goto out_unlock;
3433 }
3434 }
3435 }
3436 }
3437
3438 if (!enable && !disable) {
3439 ret = 0;
3440 goto out_unlock;
3441 }
3442
3443 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3444 if (ret)
3445 goto out_unlock;
3446
3447 /* save and update control masks and prepare csses */
3448 cgroup_save_control(cgrp);
3449
3450 cgrp->subtree_control |= enable;
3451 cgrp->subtree_control &= ~disable;
3452
3453 ret = cgroup_apply_control(cgrp);
3454 cgroup_finalize_control(cgrp, ret);
3455 if (ret)
3456 goto out_unlock;
3457
3458 kernfs_activate(cgrp->kn);
3459out_unlock:
3460 cgroup_kn_unlock(of->kn);
3461 return ret ?: nbytes;
3462}
3463
3464/**
3465 * cgroup_enable_threaded - make @cgrp threaded
3466 * @cgrp: the target cgroup
3467 *
3468 * Called when "threaded" is written to the cgroup.type interface file and
3469 * tries to make @cgrp threaded and join the parent's resource domain.
3470 * This function is never called on the root cgroup as cgroup.type doesn't
3471 * exist on it.
3472 */
3473static int cgroup_enable_threaded(struct cgroup *cgrp)
3474{
3475 struct cgroup *parent = cgroup_parent(cgrp);
3476 struct cgroup *dom_cgrp = parent->dom_cgrp;
3477 struct cgroup *dsct;
3478 struct cgroup_subsys_state *d_css;
3479 int ret;
3480
3481 lockdep_assert_held(&cgroup_mutex);
3482
3483 /* noop if already threaded */
3484 if (cgroup_is_threaded(cgrp))
3485 return 0;
3486
3487 /*
3488 * If @cgroup is populated or has domain controllers enabled, it
3489 * can't be switched. While the below cgroup_can_be_thread_root()
3490 * test can catch the same conditions, that's only when @parent is
3491 * not mixable, so let's check it explicitly.
3492 */
3493 if (cgroup_is_populated(cgrp) ||
3494 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3495 return -EOPNOTSUPP;
3496
3497 /* we're joining the parent's domain, ensure its validity */
3498 if (!cgroup_is_valid_domain(dom_cgrp) ||
3499 !cgroup_can_be_thread_root(dom_cgrp))
3500 return -EOPNOTSUPP;
3501
3502 /*
3503 * The following shouldn't cause actual migrations and should
3504 * always succeed.
3505 */
3506 cgroup_save_control(cgrp);
3507
3508 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3509 if (dsct == cgrp || cgroup_is_threaded(dsct))
3510 dsct->dom_cgrp = dom_cgrp;
3511
3512 ret = cgroup_apply_control(cgrp);
3513 if (!ret)
3514 parent->nr_threaded_children++;
3515
3516 cgroup_finalize_control(cgrp, ret);
3517 return ret;
3518}
3519
3520static int cgroup_type_show(struct seq_file *seq, void *v)
3521{
3522 struct cgroup *cgrp = seq_css(seq)->cgroup;
3523
3524 if (cgroup_is_threaded(cgrp))
3525 seq_puts(seq, "threaded\n");
3526 else if (!cgroup_is_valid_domain(cgrp))
3527 seq_puts(seq, "domain invalid\n");
3528 else if (cgroup_is_thread_root(cgrp))
3529 seq_puts(seq, "domain threaded\n");
3530 else
3531 seq_puts(seq, "domain\n");
3532
3533 return 0;
3534}
3535
3536static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3537 size_t nbytes, loff_t off)
3538{
3539 struct cgroup *cgrp;
3540 int ret;
3541
3542 /* only switching to threaded mode is supported */
3543 if (strcmp(strstrip(buf), "threaded"))
3544 return -EINVAL;
3545
3546 /* drain dying csses before we re-apply (threaded) subtree control */
3547 cgrp = cgroup_kn_lock_live(of->kn, true);
3548 if (!cgrp)
3549 return -ENOENT;
3550
3551 /* threaded can only be enabled */
3552 ret = cgroup_enable_threaded(cgrp);
3553
3554 cgroup_kn_unlock(of->kn);
3555 return ret ?: nbytes;
3556}
3557
3558static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3559{
3560 struct cgroup *cgrp = seq_css(seq)->cgroup;
3561 int descendants = READ_ONCE(cgrp->max_descendants);
3562
3563 if (descendants == INT_MAX)
3564 seq_puts(seq, "max\n");
3565 else
3566 seq_printf(seq, "%d\n", descendants);
3567
3568 return 0;
3569}
3570
3571static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3572 char *buf, size_t nbytes, loff_t off)
3573{
3574 struct cgroup *cgrp;
3575 int descendants;
3576 ssize_t ret;
3577
3578 buf = strstrip(buf);
3579 if (!strcmp(buf, "max")) {
3580 descendants = INT_MAX;
3581 } else {
3582 ret = kstrtoint(buf, 0, &descendants);
3583 if (ret)
3584 return ret;
3585 }
3586
3587 if (descendants < 0)
3588 return -ERANGE;
3589
3590 cgrp = cgroup_kn_lock_live(of->kn, false);
3591 if (!cgrp)
3592 return -ENOENT;
3593
3594 cgrp->max_descendants = descendants;
3595
3596 cgroup_kn_unlock(of->kn);
3597
3598 return nbytes;
3599}
3600
3601static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3602{
3603 struct cgroup *cgrp = seq_css(seq)->cgroup;
3604 int depth = READ_ONCE(cgrp->max_depth);
3605
3606 if (depth == INT_MAX)
3607 seq_puts(seq, "max\n");
3608 else
3609 seq_printf(seq, "%d\n", depth);
3610
3611 return 0;
3612}
3613
3614static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3615 char *buf, size_t nbytes, loff_t off)
3616{
3617 struct cgroup *cgrp;
3618 ssize_t ret;
3619 int depth;
3620
3621 buf = strstrip(buf);
3622 if (!strcmp(buf, "max")) {
3623 depth = INT_MAX;
3624 } else {
3625 ret = kstrtoint(buf, 0, &depth);
3626 if (ret)
3627 return ret;
3628 }
3629
3630 if (depth < 0)
3631 return -ERANGE;
3632
3633 cgrp = cgroup_kn_lock_live(of->kn, false);
3634 if (!cgrp)
3635 return -ENOENT;
3636
3637 cgrp->max_depth = depth;
3638
3639 cgroup_kn_unlock(of->kn);
3640
3641 return nbytes;
3642}
3643
3644static int cgroup_events_show(struct seq_file *seq, void *v)
3645{
3646 struct cgroup *cgrp = seq_css(seq)->cgroup;
3647
3648 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3649 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3650
3651 return 0;
3652}
3653
3654static int cgroup_stat_show(struct seq_file *seq, void *v)
3655{
3656 struct cgroup *cgroup = seq_css(seq)->cgroup;
3657
3658 seq_printf(seq, "nr_descendants %d\n",
3659 cgroup->nr_descendants);
3660 seq_printf(seq, "nr_dying_descendants %d\n",
3661 cgroup->nr_dying_descendants);
3662
3663 return 0;
3664}
3665
3666#ifdef CONFIG_CGROUP_SCHED
3667/**
3668 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
3669 * @cgrp: the cgroup of interest
3670 * @ss: the subsystem of interest
3671 *
3672 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
3673 * or is offline, %NULL is returned.
3674 */
3675static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
3676 struct cgroup_subsys *ss)
3677{
3678 struct cgroup_subsys_state *css;
3679
3680 rcu_read_lock();
3681 css = cgroup_css(cgrp, ss);
3682 if (css && !css_tryget_online(css))
3683 css = NULL;
3684 rcu_read_unlock();
3685
3686 return css;
3687}
3688
3689static int cgroup_extra_stat_show(struct seq_file *seq, int ssid)
3690{
3691 struct cgroup *cgrp = seq_css(seq)->cgroup;
3692 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3693 struct cgroup_subsys_state *css;
3694 int ret;
3695
3696 if (!ss->css_extra_stat_show)
3697 return 0;
3698
3699 css = cgroup_tryget_css(cgrp, ss);
3700 if (!css)
3701 return 0;
3702
3703 ret = ss->css_extra_stat_show(seq, css);
3704 css_put(css);
3705 return ret;
3706}
3707
3708static int cgroup_local_stat_show(struct seq_file *seq,
3709 struct cgroup *cgrp, int ssid)
3710{
3711 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3712 struct cgroup_subsys_state *css;
3713 int ret;
3714
3715 if (!ss->css_local_stat_show)
3716 return 0;
3717
3718 css = cgroup_tryget_css(cgrp, ss);
3719 if (!css)
3720 return 0;
3721
3722 ret = ss->css_local_stat_show(seq, css);
3723 css_put(css);
3724 return ret;
3725}
3726#endif
3727
3728static int cpu_stat_show(struct seq_file *seq, void *v)
3729{
3730 int ret = 0;
3731
3732 cgroup_base_stat_cputime_show(seq);
3733#ifdef CONFIG_CGROUP_SCHED
3734 ret = cgroup_extra_stat_show(seq, cpu_cgrp_id);
3735#endif
3736 return ret;
3737}
3738
3739static int cpu_local_stat_show(struct seq_file *seq, void *v)
3740{
3741 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3742 int ret = 0;
3743
3744#ifdef CONFIG_CGROUP_SCHED
3745 ret = cgroup_local_stat_show(seq, cgrp, cpu_cgrp_id);
3746#endif
3747 return ret;
3748}
3749
3750#ifdef CONFIG_PSI
3751static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3752{
3753 struct cgroup *cgrp = seq_css(seq)->cgroup;
3754 struct psi_group *psi = cgroup_psi(cgrp);
3755
3756 return psi_show(seq, psi, PSI_IO);
3757}
3758static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3759{
3760 struct cgroup *cgrp = seq_css(seq)->cgroup;
3761 struct psi_group *psi = cgroup_psi(cgrp);
3762
3763 return psi_show(seq, psi, PSI_MEM);
3764}
3765static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3766{
3767 struct cgroup *cgrp = seq_css(seq)->cgroup;
3768 struct psi_group *psi = cgroup_psi(cgrp);
3769
3770 return psi_show(seq, psi, PSI_CPU);
3771}
3772
3773static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3774 size_t nbytes, enum psi_res res)
3775{
3776 struct cgroup_file_ctx *ctx = of->priv;
3777 struct psi_trigger *new;
3778 struct cgroup *cgrp;
3779 struct psi_group *psi;
3780
3781 cgrp = cgroup_kn_lock_live(of->kn, false);
3782 if (!cgrp)
3783 return -ENODEV;
3784
3785 cgroup_get(cgrp);
3786 cgroup_kn_unlock(of->kn);
3787
3788 /* Allow only one trigger per file descriptor */
3789 if (ctx->psi.trigger) {
3790 cgroup_put(cgrp);
3791 return -EBUSY;
3792 }
3793
3794 psi = cgroup_psi(cgrp);
3795 new = psi_trigger_create(psi, buf, res, of->file, of);
3796 if (IS_ERR(new)) {
3797 cgroup_put(cgrp);
3798 return PTR_ERR(new);
3799 }
3800
3801 smp_store_release(&ctx->psi.trigger, new);
3802 cgroup_put(cgrp);
3803
3804 return nbytes;
3805}
3806
3807static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3808 char *buf, size_t nbytes,
3809 loff_t off)
3810{
3811 return pressure_write(of, buf, nbytes, PSI_IO);
3812}
3813
3814static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3815 char *buf, size_t nbytes,
3816 loff_t off)
3817{
3818 return pressure_write(of, buf, nbytes, PSI_MEM);
3819}
3820
3821static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3822 char *buf, size_t nbytes,
3823 loff_t off)
3824{
3825 return pressure_write(of, buf, nbytes, PSI_CPU);
3826}
3827
3828#ifdef CONFIG_IRQ_TIME_ACCOUNTING
3829static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
3830{
3831 struct cgroup *cgrp = seq_css(seq)->cgroup;
3832 struct psi_group *psi = cgroup_psi(cgrp);
3833
3834 return psi_show(seq, psi, PSI_IRQ);
3835}
3836
3837static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
3838 char *buf, size_t nbytes,
3839 loff_t off)
3840{
3841 return pressure_write(of, buf, nbytes, PSI_IRQ);
3842}
3843#endif
3844
3845static int cgroup_pressure_show(struct seq_file *seq, void *v)
3846{
3847 struct cgroup *cgrp = seq_css(seq)->cgroup;
3848 struct psi_group *psi = cgroup_psi(cgrp);
3849
3850 seq_printf(seq, "%d\n", psi->enabled);
3851
3852 return 0;
3853}
3854
3855static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
3856 char *buf, size_t nbytes,
3857 loff_t off)
3858{
3859 ssize_t ret;
3860 int enable;
3861 struct cgroup *cgrp;
3862 struct psi_group *psi;
3863
3864 ret = kstrtoint(strstrip(buf), 0, &enable);
3865 if (ret)
3866 return ret;
3867
3868 if (enable < 0 || enable > 1)
3869 return -ERANGE;
3870
3871 cgrp = cgroup_kn_lock_live(of->kn, false);
3872 if (!cgrp)
3873 return -ENOENT;
3874
3875 psi = cgroup_psi(cgrp);
3876 if (psi->enabled != enable) {
3877 int i;
3878
3879 /* show or hide {cpu,memory,io,irq}.pressure files */
3880 for (i = 0; i < NR_PSI_RESOURCES; i++)
3881 cgroup_file_show(&cgrp->psi_files[i], enable);
3882
3883 psi->enabled = enable;
3884 if (enable)
3885 psi_cgroup_restart(psi);
3886 }
3887
3888 cgroup_kn_unlock(of->kn);
3889
3890 return nbytes;
3891}
3892
3893static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3894 poll_table *pt)
3895{
3896 struct cgroup_file_ctx *ctx = of->priv;
3897
3898 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3899}
3900
3901static void cgroup_pressure_release(struct kernfs_open_file *of)
3902{
3903 struct cgroup_file_ctx *ctx = of->priv;
3904
3905 psi_trigger_destroy(ctx->psi.trigger);
3906}
3907
3908bool cgroup_psi_enabled(void)
3909{
3910 if (static_branch_likely(&psi_disabled))
3911 return false;
3912
3913 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3914}
3915
3916#else /* CONFIG_PSI */
3917bool cgroup_psi_enabled(void)
3918{
3919 return false;
3920}
3921
3922#endif /* CONFIG_PSI */
3923
3924static int cgroup_freeze_show(struct seq_file *seq, void *v)
3925{
3926 struct cgroup *cgrp = seq_css(seq)->cgroup;
3927
3928 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3929
3930 return 0;
3931}
3932
3933static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3934 char *buf, size_t nbytes, loff_t off)
3935{
3936 struct cgroup *cgrp;
3937 ssize_t ret;
3938 int freeze;
3939
3940 ret = kstrtoint(strstrip(buf), 0, &freeze);
3941 if (ret)
3942 return ret;
3943
3944 if (freeze < 0 || freeze > 1)
3945 return -ERANGE;
3946
3947 cgrp = cgroup_kn_lock_live(of->kn, false);
3948 if (!cgrp)
3949 return -ENOENT;
3950
3951 cgroup_freeze(cgrp, freeze);
3952
3953 cgroup_kn_unlock(of->kn);
3954
3955 return nbytes;
3956}
3957
3958static void __cgroup_kill(struct cgroup *cgrp)
3959{
3960 struct css_task_iter it;
3961 struct task_struct *task;
3962
3963 lockdep_assert_held(&cgroup_mutex);
3964
3965 spin_lock_irq(&css_set_lock);
3966 set_bit(CGRP_KILL, &cgrp->flags);
3967 spin_unlock_irq(&css_set_lock);
3968
3969 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3970 while ((task = css_task_iter_next(&it))) {
3971 /* Ignore kernel threads here. */
3972 if (task->flags & PF_KTHREAD)
3973 continue;
3974
3975 /* Skip tasks that are already dying. */
3976 if (__fatal_signal_pending(task))
3977 continue;
3978
3979 send_sig(SIGKILL, task, 0);
3980 }
3981 css_task_iter_end(&it);
3982
3983 spin_lock_irq(&css_set_lock);
3984 clear_bit(CGRP_KILL, &cgrp->flags);
3985 spin_unlock_irq(&css_set_lock);
3986}
3987
3988static void cgroup_kill(struct cgroup *cgrp)
3989{
3990 struct cgroup_subsys_state *css;
3991 struct cgroup *dsct;
3992
3993 lockdep_assert_held(&cgroup_mutex);
3994
3995 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3996 __cgroup_kill(dsct);
3997}
3998
3999static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
4000 size_t nbytes, loff_t off)
4001{
4002 ssize_t ret = 0;
4003 int kill;
4004 struct cgroup *cgrp;
4005
4006 ret = kstrtoint(strstrip(buf), 0, &kill);
4007 if (ret)
4008 return ret;
4009
4010 if (kill != 1)
4011 return -ERANGE;
4012
4013 cgrp = cgroup_kn_lock_live(of->kn, false);
4014 if (!cgrp)
4015 return -ENOENT;
4016
4017 /*
4018 * Killing is a process directed operation, i.e. the whole thread-group
4019 * is taken down so act like we do for cgroup.procs and only make this
4020 * writable in non-threaded cgroups.
4021 */
4022 if (cgroup_is_threaded(cgrp))
4023 ret = -EOPNOTSUPP;
4024 else
4025 cgroup_kill(cgrp);
4026
4027 cgroup_kn_unlock(of->kn);
4028
4029 return ret ?: nbytes;
4030}
4031
4032static int cgroup_file_open(struct kernfs_open_file *of)
4033{
4034 struct cftype *cft = of_cft(of);
4035 struct cgroup_file_ctx *ctx;
4036 int ret;
4037
4038 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
4039 if (!ctx)
4040 return -ENOMEM;
4041
4042 ctx->ns = current->nsproxy->cgroup_ns;
4043 get_cgroup_ns(ctx->ns);
4044 of->priv = ctx;
4045
4046 if (!cft->open)
4047 return 0;
4048
4049 ret = cft->open(of);
4050 if (ret) {
4051 put_cgroup_ns(ctx->ns);
4052 kfree(ctx);
4053 }
4054 return ret;
4055}
4056
4057static void cgroup_file_release(struct kernfs_open_file *of)
4058{
4059 struct cftype *cft = of_cft(of);
4060 struct cgroup_file_ctx *ctx = of->priv;
4061
4062 if (cft->release)
4063 cft->release(of);
4064 put_cgroup_ns(ctx->ns);
4065 kfree(ctx);
4066}
4067
4068static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4069 size_t nbytes, loff_t off)
4070{
4071 struct cgroup_file_ctx *ctx = of->priv;
4072 struct cgroup *cgrp = of->kn->parent->priv;
4073 struct cftype *cft = of_cft(of);
4074 struct cgroup_subsys_state *css;
4075 int ret;
4076
4077 if (!nbytes)
4078 return 0;
4079
4080 /*
4081 * If namespaces are delegation boundaries, disallow writes to
4082 * files in an non-init namespace root from inside the namespace
4083 * except for the files explicitly marked delegatable -
4084 * cgroup.procs and cgroup.subtree_control.
4085 */
4086 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4087 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4088 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4089 return -EPERM;
4090
4091 if (cft->write)
4092 return cft->write(of, buf, nbytes, off);
4093
4094 /*
4095 * kernfs guarantees that a file isn't deleted with operations in
4096 * flight, which means that the matching css is and stays alive and
4097 * doesn't need to be pinned. The RCU locking is not necessary
4098 * either. It's just for the convenience of using cgroup_css().
4099 */
4100 rcu_read_lock();
4101 css = cgroup_css(cgrp, cft->ss);
4102 rcu_read_unlock();
4103
4104 if (cft->write_u64) {
4105 unsigned long long v;
4106 ret = kstrtoull(buf, 0, &v);
4107 if (!ret)
4108 ret = cft->write_u64(css, cft, v);
4109 } else if (cft->write_s64) {
4110 long long v;
4111 ret = kstrtoll(buf, 0, &v);
4112 if (!ret)
4113 ret = cft->write_s64(css, cft, v);
4114 } else {
4115 ret = -EINVAL;
4116 }
4117
4118 return ret ?: nbytes;
4119}
4120
4121static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4122{
4123 struct cftype *cft = of_cft(of);
4124
4125 if (cft->poll)
4126 return cft->poll(of, pt);
4127
4128 return kernfs_generic_poll(of, pt);
4129}
4130
4131static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4132{
4133 return seq_cft(seq)->seq_start(seq, ppos);
4134}
4135
4136static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4137{
4138 return seq_cft(seq)->seq_next(seq, v, ppos);
4139}
4140
4141static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4142{
4143 if (seq_cft(seq)->seq_stop)
4144 seq_cft(seq)->seq_stop(seq, v);
4145}
4146
4147static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4148{
4149 struct cftype *cft = seq_cft(m);
4150 struct cgroup_subsys_state *css = seq_css(m);
4151
4152 if (cft->seq_show)
4153 return cft->seq_show(m, arg);
4154
4155 if (cft->read_u64)
4156 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4157 else if (cft->read_s64)
4158 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4159 else
4160 return -EINVAL;
4161 return 0;
4162}
4163
4164static struct kernfs_ops cgroup_kf_single_ops = {
4165 .atomic_write_len = PAGE_SIZE,
4166 .open = cgroup_file_open,
4167 .release = cgroup_file_release,
4168 .write = cgroup_file_write,
4169 .poll = cgroup_file_poll,
4170 .seq_show = cgroup_seqfile_show,
4171};
4172
4173static struct kernfs_ops cgroup_kf_ops = {
4174 .atomic_write_len = PAGE_SIZE,
4175 .open = cgroup_file_open,
4176 .release = cgroup_file_release,
4177 .write = cgroup_file_write,
4178 .poll = cgroup_file_poll,
4179 .seq_start = cgroup_seqfile_start,
4180 .seq_next = cgroup_seqfile_next,
4181 .seq_stop = cgroup_seqfile_stop,
4182 .seq_show = cgroup_seqfile_show,
4183};
4184
4185static void cgroup_file_notify_timer(struct timer_list *timer)
4186{
4187 cgroup_file_notify(container_of(timer, struct cgroup_file,
4188 notify_timer));
4189}
4190
4191static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4192 struct cftype *cft)
4193{
4194 char name[CGROUP_FILE_NAME_MAX];
4195 struct kernfs_node *kn;
4196 struct lock_class_key *key = NULL;
4197
4198#ifdef CONFIG_DEBUG_LOCK_ALLOC
4199 key = &cft->lockdep_key;
4200#endif
4201 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4202 cgroup_file_mode(cft),
4203 current_fsuid(), current_fsgid(),
4204 0, cft->kf_ops, cft,
4205 NULL, key);
4206 if (IS_ERR(kn))
4207 return PTR_ERR(kn);
4208
4209 if (cft->file_offset) {
4210 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4211
4212 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4213
4214 spin_lock_irq(&cgroup_file_kn_lock);
4215 cfile->kn = kn;
4216 spin_unlock_irq(&cgroup_file_kn_lock);
4217 }
4218
4219 return 0;
4220}
4221
4222/**
4223 * cgroup_addrm_files - add or remove files to a cgroup directory
4224 * @css: the target css
4225 * @cgrp: the target cgroup (usually css->cgroup)
4226 * @cfts: array of cftypes to be added
4227 * @is_add: whether to add or remove
4228 *
4229 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4230 * For removals, this function never fails.
4231 */
4232static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4233 struct cgroup *cgrp, struct cftype cfts[],
4234 bool is_add)
4235{
4236 struct cftype *cft, *cft_end = NULL;
4237 int ret = 0;
4238
4239 lockdep_assert_held(&cgroup_mutex);
4240
4241restart:
4242 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4243 /* does cft->flags tell us to skip this file on @cgrp? */
4244 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4245 continue;
4246 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4247 continue;
4248 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4249 continue;
4250 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4251 continue;
4252 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4253 continue;
4254 if (is_add) {
4255 ret = cgroup_add_file(css, cgrp, cft);
4256 if (ret) {
4257 pr_warn("%s: failed to add %s, err=%d\n",
4258 __func__, cft->name, ret);
4259 cft_end = cft;
4260 is_add = false;
4261 goto restart;
4262 }
4263 } else {
4264 cgroup_rm_file(cgrp, cft);
4265 }
4266 }
4267 return ret;
4268}
4269
4270static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4271{
4272 struct cgroup_subsys *ss = cfts[0].ss;
4273 struct cgroup *root = &ss->root->cgrp;
4274 struct cgroup_subsys_state *css;
4275 int ret = 0;
4276
4277 lockdep_assert_held(&cgroup_mutex);
4278
4279 /* add/rm files for all cgroups created before */
4280 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4281 struct cgroup *cgrp = css->cgroup;
4282
4283 if (!(css->flags & CSS_VISIBLE))
4284 continue;
4285
4286 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4287 if (ret)
4288 break;
4289 }
4290
4291 if (is_add && !ret)
4292 kernfs_activate(root->kn);
4293 return ret;
4294}
4295
4296static void cgroup_exit_cftypes(struct cftype *cfts)
4297{
4298 struct cftype *cft;
4299
4300 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4301 /* free copy for custom atomic_write_len, see init_cftypes() */
4302 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4303 kfree(cft->kf_ops);
4304 cft->kf_ops = NULL;
4305 cft->ss = NULL;
4306
4307 /* revert flags set by cgroup core while adding @cfts */
4308 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4309 __CFTYPE_ADDED);
4310 }
4311}
4312
4313static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4314{
4315 struct cftype *cft;
4316 int ret = 0;
4317
4318 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4319 struct kernfs_ops *kf_ops;
4320
4321 WARN_ON(cft->ss || cft->kf_ops);
4322
4323 if (cft->flags & __CFTYPE_ADDED) {
4324 ret = -EBUSY;
4325 break;
4326 }
4327
4328 if (cft->seq_start)
4329 kf_ops = &cgroup_kf_ops;
4330 else
4331 kf_ops = &cgroup_kf_single_ops;
4332
4333 /*
4334 * Ugh... if @cft wants a custom max_write_len, we need to
4335 * make a copy of kf_ops to set its atomic_write_len.
4336 */
4337 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4338 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4339 if (!kf_ops) {
4340 ret = -ENOMEM;
4341 break;
4342 }
4343 kf_ops->atomic_write_len = cft->max_write_len;
4344 }
4345
4346 cft->kf_ops = kf_ops;
4347 cft->ss = ss;
4348 cft->flags |= __CFTYPE_ADDED;
4349 }
4350
4351 if (ret)
4352 cgroup_exit_cftypes(cfts);
4353 return ret;
4354}
4355
4356static void cgroup_rm_cftypes_locked(struct cftype *cfts)
4357{
4358 lockdep_assert_held(&cgroup_mutex);
4359
4360 list_del(&cfts->node);
4361 cgroup_apply_cftypes(cfts, false);
4362 cgroup_exit_cftypes(cfts);
4363}
4364
4365/**
4366 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4367 * @cfts: zero-length name terminated array of cftypes
4368 *
4369 * Unregister @cfts. Files described by @cfts are removed from all
4370 * existing cgroups and all future cgroups won't have them either. This
4371 * function can be called anytime whether @cfts' subsys is attached or not.
4372 *
4373 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4374 * registered.
4375 */
4376int cgroup_rm_cftypes(struct cftype *cfts)
4377{
4378 if (!cfts || cfts[0].name[0] == '\0')
4379 return 0;
4380
4381 if (!(cfts[0].flags & __CFTYPE_ADDED))
4382 return -ENOENT;
4383
4384 cgroup_lock();
4385 cgroup_rm_cftypes_locked(cfts);
4386 cgroup_unlock();
4387 return 0;
4388}
4389
4390/**
4391 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4392 * @ss: target cgroup subsystem
4393 * @cfts: zero-length name terminated array of cftypes
4394 *
4395 * Register @cfts to @ss. Files described by @cfts are created for all
4396 * existing cgroups to which @ss is attached and all future cgroups will
4397 * have them too. This function can be called anytime whether @ss is
4398 * attached or not.
4399 *
4400 * Returns 0 on successful registration, -errno on failure. Note that this
4401 * function currently returns 0 as long as @cfts registration is successful
4402 * even if some file creation attempts on existing cgroups fail.
4403 */
4404static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4405{
4406 int ret;
4407
4408 if (!cgroup_ssid_enabled(ss->id))
4409 return 0;
4410
4411 if (!cfts || cfts[0].name[0] == '\0')
4412 return 0;
4413
4414 ret = cgroup_init_cftypes(ss, cfts);
4415 if (ret)
4416 return ret;
4417
4418 cgroup_lock();
4419
4420 list_add_tail(&cfts->node, &ss->cfts);
4421 ret = cgroup_apply_cftypes(cfts, true);
4422 if (ret)
4423 cgroup_rm_cftypes_locked(cfts);
4424
4425 cgroup_unlock();
4426 return ret;
4427}
4428
4429/**
4430 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4431 * @ss: target cgroup subsystem
4432 * @cfts: zero-length name terminated array of cftypes
4433 *
4434 * Similar to cgroup_add_cftypes() but the added files are only used for
4435 * the default hierarchy.
4436 */
4437int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4438{
4439 struct cftype *cft;
4440
4441 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4442 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4443 return cgroup_add_cftypes(ss, cfts);
4444}
4445
4446/**
4447 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4448 * @ss: target cgroup subsystem
4449 * @cfts: zero-length name terminated array of cftypes
4450 *
4451 * Similar to cgroup_add_cftypes() but the added files are only used for
4452 * the legacy hierarchies.
4453 */
4454int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4455{
4456 struct cftype *cft;
4457
4458 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4459 cft->flags |= __CFTYPE_NOT_ON_DFL;
4460 return cgroup_add_cftypes(ss, cfts);
4461}
4462
4463/**
4464 * cgroup_file_notify - generate a file modified event for a cgroup_file
4465 * @cfile: target cgroup_file
4466 *
4467 * @cfile must have been obtained by setting cftype->file_offset.
4468 */
4469void cgroup_file_notify(struct cgroup_file *cfile)
4470{
4471 unsigned long flags;
4472
4473 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4474 if (cfile->kn) {
4475 unsigned long last = cfile->notified_at;
4476 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4477
4478 if (time_in_range(jiffies, last, next)) {
4479 timer_reduce(&cfile->notify_timer, next);
4480 } else {
4481 kernfs_notify(cfile->kn);
4482 cfile->notified_at = jiffies;
4483 }
4484 }
4485 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4486}
4487
4488/**
4489 * cgroup_file_show - show or hide a hidden cgroup file
4490 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4491 * @show: whether to show or hide
4492 */
4493void cgroup_file_show(struct cgroup_file *cfile, bool show)
4494{
4495 struct kernfs_node *kn;
4496
4497 spin_lock_irq(&cgroup_file_kn_lock);
4498 kn = cfile->kn;
4499 kernfs_get(kn);
4500 spin_unlock_irq(&cgroup_file_kn_lock);
4501
4502 if (kn)
4503 kernfs_show(kn, show);
4504
4505 kernfs_put(kn);
4506}
4507
4508/**
4509 * css_next_child - find the next child of a given css
4510 * @pos: the current position (%NULL to initiate traversal)
4511 * @parent: css whose children to walk
4512 *
4513 * This function returns the next child of @parent and should be called
4514 * under either cgroup_mutex or RCU read lock. The only requirement is
4515 * that @parent and @pos are accessible. The next sibling is guaranteed to
4516 * be returned regardless of their states.
4517 *
4518 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4519 * css which finished ->css_online() is guaranteed to be visible in the
4520 * future iterations and will stay visible until the last reference is put.
4521 * A css which hasn't finished ->css_online() or already finished
4522 * ->css_offline() may show up during traversal. It's each subsystem's
4523 * responsibility to synchronize against on/offlining.
4524 */
4525struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4526 struct cgroup_subsys_state *parent)
4527{
4528 struct cgroup_subsys_state *next;
4529
4530 cgroup_assert_mutex_or_rcu_locked();
4531
4532 /*
4533 * @pos could already have been unlinked from the sibling list.
4534 * Once a cgroup is removed, its ->sibling.next is no longer
4535 * updated when its next sibling changes. CSS_RELEASED is set when
4536 * @pos is taken off list, at which time its next pointer is valid,
4537 * and, as releases are serialized, the one pointed to by the next
4538 * pointer is guaranteed to not have started release yet. This
4539 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4540 * critical section, the one pointed to by its next pointer is
4541 * guaranteed to not have finished its RCU grace period even if we
4542 * have dropped rcu_read_lock() in-between iterations.
4543 *
4544 * If @pos has CSS_RELEASED set, its next pointer can't be
4545 * dereferenced; however, as each css is given a monotonically
4546 * increasing unique serial number and always appended to the
4547 * sibling list, the next one can be found by walking the parent's
4548 * children until the first css with higher serial number than
4549 * @pos's. While this path can be slower, it happens iff iteration
4550 * races against release and the race window is very small.
4551 */
4552 if (!pos) {
4553 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4554 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4555 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4556 } else {
4557 list_for_each_entry_rcu(next, &parent->children, sibling,
4558 lockdep_is_held(&cgroup_mutex))
4559 if (next->serial_nr > pos->serial_nr)
4560 break;
4561 }
4562
4563 /*
4564 * @next, if not pointing to the head, can be dereferenced and is
4565 * the next sibling.
4566 */
4567 if (&next->sibling != &parent->children)
4568 return next;
4569 return NULL;
4570}
4571
4572/**
4573 * css_next_descendant_pre - find the next descendant for pre-order walk
4574 * @pos: the current position (%NULL to initiate traversal)
4575 * @root: css whose descendants to walk
4576 *
4577 * To be used by css_for_each_descendant_pre(). Find the next descendant
4578 * to visit for pre-order traversal of @root's descendants. @root is
4579 * included in the iteration and the first node to be visited.
4580 *
4581 * While this function requires cgroup_mutex or RCU read locking, it
4582 * doesn't require the whole traversal to be contained in a single critical
4583 * section. This function will return the correct next descendant as long
4584 * as both @pos and @root are accessible and @pos is a descendant of @root.
4585 *
4586 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4587 * css which finished ->css_online() is guaranteed to be visible in the
4588 * future iterations and will stay visible until the last reference is put.
4589 * A css which hasn't finished ->css_online() or already finished
4590 * ->css_offline() may show up during traversal. It's each subsystem's
4591 * responsibility to synchronize against on/offlining.
4592 */
4593struct cgroup_subsys_state *
4594css_next_descendant_pre(struct cgroup_subsys_state *pos,
4595 struct cgroup_subsys_state *root)
4596{
4597 struct cgroup_subsys_state *next;
4598
4599 cgroup_assert_mutex_or_rcu_locked();
4600
4601 /* if first iteration, visit @root */
4602 if (!pos)
4603 return root;
4604
4605 /* visit the first child if exists */
4606 next = css_next_child(NULL, pos);
4607 if (next)
4608 return next;
4609
4610 /* no child, visit my or the closest ancestor's next sibling */
4611 while (pos != root) {
4612 next = css_next_child(pos, pos->parent);
4613 if (next)
4614 return next;
4615 pos = pos->parent;
4616 }
4617
4618 return NULL;
4619}
4620EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4621
4622/**
4623 * css_rightmost_descendant - return the rightmost descendant of a css
4624 * @pos: css of interest
4625 *
4626 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4627 * is returned. This can be used during pre-order traversal to skip
4628 * subtree of @pos.
4629 *
4630 * While this function requires cgroup_mutex or RCU read locking, it
4631 * doesn't require the whole traversal to be contained in a single critical
4632 * section. This function will return the correct rightmost descendant as
4633 * long as @pos is accessible.
4634 */
4635struct cgroup_subsys_state *
4636css_rightmost_descendant(struct cgroup_subsys_state *pos)
4637{
4638 struct cgroup_subsys_state *last, *tmp;
4639
4640 cgroup_assert_mutex_or_rcu_locked();
4641
4642 do {
4643 last = pos;
4644 /* ->prev isn't RCU safe, walk ->next till the end */
4645 pos = NULL;
4646 css_for_each_child(tmp, last)
4647 pos = tmp;
4648 } while (pos);
4649
4650 return last;
4651}
4652
4653static struct cgroup_subsys_state *
4654css_leftmost_descendant(struct cgroup_subsys_state *pos)
4655{
4656 struct cgroup_subsys_state *last;
4657
4658 do {
4659 last = pos;
4660 pos = css_next_child(NULL, pos);
4661 } while (pos);
4662
4663 return last;
4664}
4665
4666/**
4667 * css_next_descendant_post - find the next descendant for post-order walk
4668 * @pos: the current position (%NULL to initiate traversal)
4669 * @root: css whose descendants to walk
4670 *
4671 * To be used by css_for_each_descendant_post(). Find the next descendant
4672 * to visit for post-order traversal of @root's descendants. @root is
4673 * included in the iteration and the last node to be visited.
4674 *
4675 * While this function requires cgroup_mutex or RCU read locking, it
4676 * doesn't require the whole traversal to be contained in a single critical
4677 * section. This function will return the correct next descendant as long
4678 * as both @pos and @cgroup are accessible and @pos is a descendant of
4679 * @cgroup.
4680 *
4681 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4682 * css which finished ->css_online() is guaranteed to be visible in the
4683 * future iterations and will stay visible until the last reference is put.
4684 * A css which hasn't finished ->css_online() or already finished
4685 * ->css_offline() may show up during traversal. It's each subsystem's
4686 * responsibility to synchronize against on/offlining.
4687 */
4688struct cgroup_subsys_state *
4689css_next_descendant_post(struct cgroup_subsys_state *pos,
4690 struct cgroup_subsys_state *root)
4691{
4692 struct cgroup_subsys_state *next;
4693
4694 cgroup_assert_mutex_or_rcu_locked();
4695
4696 /* if first iteration, visit leftmost descendant which may be @root */
4697 if (!pos)
4698 return css_leftmost_descendant(root);
4699
4700 /* if we visited @root, we're done */
4701 if (pos == root)
4702 return NULL;
4703
4704 /* if there's an unvisited sibling, visit its leftmost descendant */
4705 next = css_next_child(pos, pos->parent);
4706 if (next)
4707 return css_leftmost_descendant(next);
4708
4709 /* no sibling left, visit parent */
4710 return pos->parent;
4711}
4712
4713/**
4714 * css_has_online_children - does a css have online children
4715 * @css: the target css
4716 *
4717 * Returns %true if @css has any online children; otherwise, %false. This
4718 * function can be called from any context but the caller is responsible
4719 * for synchronizing against on/offlining as necessary.
4720 */
4721bool css_has_online_children(struct cgroup_subsys_state *css)
4722{
4723 struct cgroup_subsys_state *child;
4724 bool ret = false;
4725
4726 rcu_read_lock();
4727 css_for_each_child(child, css) {
4728 if (child->flags & CSS_ONLINE) {
4729 ret = true;
4730 break;
4731 }
4732 }
4733 rcu_read_unlock();
4734 return ret;
4735}
4736
4737static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4738{
4739 struct list_head *l;
4740 struct cgrp_cset_link *link;
4741 struct css_set *cset;
4742
4743 lockdep_assert_held(&css_set_lock);
4744
4745 /* find the next threaded cset */
4746 if (it->tcset_pos) {
4747 l = it->tcset_pos->next;
4748
4749 if (l != it->tcset_head) {
4750 it->tcset_pos = l;
4751 return container_of(l, struct css_set,
4752 threaded_csets_node);
4753 }
4754
4755 it->tcset_pos = NULL;
4756 }
4757
4758 /* find the next cset */
4759 l = it->cset_pos;
4760 l = l->next;
4761 if (l == it->cset_head) {
4762 it->cset_pos = NULL;
4763 return NULL;
4764 }
4765
4766 if (it->ss) {
4767 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4768 } else {
4769 link = list_entry(l, struct cgrp_cset_link, cset_link);
4770 cset = link->cset;
4771 }
4772
4773 it->cset_pos = l;
4774
4775 /* initialize threaded css_set walking */
4776 if (it->flags & CSS_TASK_ITER_THREADED) {
4777 if (it->cur_dcset)
4778 put_css_set_locked(it->cur_dcset);
4779 it->cur_dcset = cset;
4780 get_css_set(cset);
4781
4782 it->tcset_head = &cset->threaded_csets;
4783 it->tcset_pos = &cset->threaded_csets;
4784 }
4785
4786 return cset;
4787}
4788
4789/**
4790 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4791 * @it: the iterator to advance
4792 *
4793 * Advance @it to the next css_set to walk.
4794 */
4795static void css_task_iter_advance_css_set(struct css_task_iter *it)
4796{
4797 struct css_set *cset;
4798
4799 lockdep_assert_held(&css_set_lock);
4800
4801 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4802 while ((cset = css_task_iter_next_css_set(it))) {
4803 if (!list_empty(&cset->tasks)) {
4804 it->cur_tasks_head = &cset->tasks;
4805 break;
4806 } else if (!list_empty(&cset->mg_tasks)) {
4807 it->cur_tasks_head = &cset->mg_tasks;
4808 break;
4809 } else if (!list_empty(&cset->dying_tasks)) {
4810 it->cur_tasks_head = &cset->dying_tasks;
4811 break;
4812 }
4813 }
4814 if (!cset) {
4815 it->task_pos = NULL;
4816 return;
4817 }
4818 it->task_pos = it->cur_tasks_head->next;
4819
4820 /*
4821 * We don't keep css_sets locked across iteration steps and thus
4822 * need to take steps to ensure that iteration can be resumed after
4823 * the lock is re-acquired. Iteration is performed at two levels -
4824 * css_sets and tasks in them.
4825 *
4826 * Once created, a css_set never leaves its cgroup lists, so a
4827 * pinned css_set is guaranteed to stay put and we can resume
4828 * iteration afterwards.
4829 *
4830 * Tasks may leave @cset across iteration steps. This is resolved
4831 * by registering each iterator with the css_set currently being
4832 * walked and making css_set_move_task() advance iterators whose
4833 * next task is leaving.
4834 */
4835 if (it->cur_cset) {
4836 list_del(&it->iters_node);
4837 put_css_set_locked(it->cur_cset);
4838 }
4839 get_css_set(cset);
4840 it->cur_cset = cset;
4841 list_add(&it->iters_node, &cset->task_iters);
4842}
4843
4844static void css_task_iter_skip(struct css_task_iter *it,
4845 struct task_struct *task)
4846{
4847 lockdep_assert_held(&css_set_lock);
4848
4849 if (it->task_pos == &task->cg_list) {
4850 it->task_pos = it->task_pos->next;
4851 it->flags |= CSS_TASK_ITER_SKIPPED;
4852 }
4853}
4854
4855static void css_task_iter_advance(struct css_task_iter *it)
4856{
4857 struct task_struct *task;
4858
4859 lockdep_assert_held(&css_set_lock);
4860repeat:
4861 if (it->task_pos) {
4862 /*
4863 * Advance iterator to find next entry. We go through cset
4864 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4865 * the next cset.
4866 */
4867 if (it->flags & CSS_TASK_ITER_SKIPPED)
4868 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4869 else
4870 it->task_pos = it->task_pos->next;
4871
4872 if (it->task_pos == &it->cur_cset->tasks) {
4873 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4874 it->task_pos = it->cur_tasks_head->next;
4875 }
4876 if (it->task_pos == &it->cur_cset->mg_tasks) {
4877 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4878 it->task_pos = it->cur_tasks_head->next;
4879 }
4880 if (it->task_pos == &it->cur_cset->dying_tasks)
4881 css_task_iter_advance_css_set(it);
4882 } else {
4883 /* called from start, proceed to the first cset */
4884 css_task_iter_advance_css_set(it);
4885 }
4886
4887 if (!it->task_pos)
4888 return;
4889
4890 task = list_entry(it->task_pos, struct task_struct, cg_list);
4891
4892 if (it->flags & CSS_TASK_ITER_PROCS) {
4893 /* if PROCS, skip over tasks which aren't group leaders */
4894 if (!thread_group_leader(task))
4895 goto repeat;
4896
4897 /* and dying leaders w/o live member threads */
4898 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4899 !atomic_read(&task->signal->live))
4900 goto repeat;
4901 } else {
4902 /* skip all dying ones */
4903 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4904 goto repeat;
4905 }
4906}
4907
4908/**
4909 * css_task_iter_start - initiate task iteration
4910 * @css: the css to walk tasks of
4911 * @flags: CSS_TASK_ITER_* flags
4912 * @it: the task iterator to use
4913 *
4914 * Initiate iteration through the tasks of @css. The caller can call
4915 * css_task_iter_next() to walk through the tasks until the function
4916 * returns NULL. On completion of iteration, css_task_iter_end() must be
4917 * called.
4918 */
4919void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4920 struct css_task_iter *it)
4921{
4922 unsigned long irqflags;
4923
4924 memset(it, 0, sizeof(*it));
4925
4926 spin_lock_irqsave(&css_set_lock, irqflags);
4927
4928 it->ss = css->ss;
4929 it->flags = flags;
4930
4931 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4932 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4933 else
4934 it->cset_pos = &css->cgroup->cset_links;
4935
4936 it->cset_head = it->cset_pos;
4937
4938 css_task_iter_advance(it);
4939
4940 spin_unlock_irqrestore(&css_set_lock, irqflags);
4941}
4942
4943/**
4944 * css_task_iter_next - return the next task for the iterator
4945 * @it: the task iterator being iterated
4946 *
4947 * The "next" function for task iteration. @it should have been
4948 * initialized via css_task_iter_start(). Returns NULL when the iteration
4949 * reaches the end.
4950 */
4951struct task_struct *css_task_iter_next(struct css_task_iter *it)
4952{
4953 unsigned long irqflags;
4954
4955 if (it->cur_task) {
4956 put_task_struct(it->cur_task);
4957 it->cur_task = NULL;
4958 }
4959
4960 spin_lock_irqsave(&css_set_lock, irqflags);
4961
4962 /* @it may be half-advanced by skips, finish advancing */
4963 if (it->flags & CSS_TASK_ITER_SKIPPED)
4964 css_task_iter_advance(it);
4965
4966 if (it->task_pos) {
4967 it->cur_task = list_entry(it->task_pos, struct task_struct,
4968 cg_list);
4969 get_task_struct(it->cur_task);
4970 css_task_iter_advance(it);
4971 }
4972
4973 spin_unlock_irqrestore(&css_set_lock, irqflags);
4974
4975 return it->cur_task;
4976}
4977
4978/**
4979 * css_task_iter_end - finish task iteration
4980 * @it: the task iterator to finish
4981 *
4982 * Finish task iteration started by css_task_iter_start().
4983 */
4984void css_task_iter_end(struct css_task_iter *it)
4985{
4986 unsigned long irqflags;
4987
4988 if (it->cur_cset) {
4989 spin_lock_irqsave(&css_set_lock, irqflags);
4990 list_del(&it->iters_node);
4991 put_css_set_locked(it->cur_cset);
4992 spin_unlock_irqrestore(&css_set_lock, irqflags);
4993 }
4994
4995 if (it->cur_dcset)
4996 put_css_set(it->cur_dcset);
4997
4998 if (it->cur_task)
4999 put_task_struct(it->cur_task);
5000}
5001
5002static void cgroup_procs_release(struct kernfs_open_file *of)
5003{
5004 struct cgroup_file_ctx *ctx = of->priv;
5005
5006 if (ctx->procs.started)
5007 css_task_iter_end(&ctx->procs.iter);
5008}
5009
5010static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
5011{
5012 struct kernfs_open_file *of = s->private;
5013 struct cgroup_file_ctx *ctx = of->priv;
5014
5015 if (pos)
5016 (*pos)++;
5017
5018 return css_task_iter_next(&ctx->procs.iter);
5019}
5020
5021static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5022 unsigned int iter_flags)
5023{
5024 struct kernfs_open_file *of = s->private;
5025 struct cgroup *cgrp = seq_css(s)->cgroup;
5026 struct cgroup_file_ctx *ctx = of->priv;
5027 struct css_task_iter *it = &ctx->procs.iter;
5028
5029 /*
5030 * When a seq_file is seeked, it's always traversed sequentially
5031 * from position 0, so we can simply keep iterating on !0 *pos.
5032 */
5033 if (!ctx->procs.started) {
5034 if (WARN_ON_ONCE((*pos)))
5035 return ERR_PTR(-EINVAL);
5036 css_task_iter_start(&cgrp->self, iter_flags, it);
5037 ctx->procs.started = true;
5038 } else if (!(*pos)) {
5039 css_task_iter_end(it);
5040 css_task_iter_start(&cgrp->self, iter_flags, it);
5041 } else
5042 return it->cur_task;
5043
5044 return cgroup_procs_next(s, NULL, NULL);
5045}
5046
5047static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5048{
5049 struct cgroup *cgrp = seq_css(s)->cgroup;
5050
5051 /*
5052 * All processes of a threaded subtree belong to the domain cgroup
5053 * of the subtree. Only threads can be distributed across the
5054 * subtree. Reject reads on cgroup.procs in the subtree proper.
5055 * They're always empty anyway.
5056 */
5057 if (cgroup_is_threaded(cgrp))
5058 return ERR_PTR(-EOPNOTSUPP);
5059
5060 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5061 CSS_TASK_ITER_THREADED);
5062}
5063
5064static int cgroup_procs_show(struct seq_file *s, void *v)
5065{
5066 seq_printf(s, "%d\n", task_pid_vnr(v));
5067 return 0;
5068}
5069
5070static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5071{
5072 int ret;
5073 struct inode *inode;
5074
5075 lockdep_assert_held(&cgroup_mutex);
5076
5077 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5078 if (!inode)
5079 return -ENOMEM;
5080
5081 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5082 iput(inode);
5083 return ret;
5084}
5085
5086static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5087 struct cgroup *dst_cgrp,
5088 struct super_block *sb,
5089 struct cgroup_namespace *ns)
5090{
5091 struct cgroup *com_cgrp = src_cgrp;
5092 int ret;
5093
5094 lockdep_assert_held(&cgroup_mutex);
5095
5096 /* find the common ancestor */
5097 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5098 com_cgrp = cgroup_parent(com_cgrp);
5099
5100 /* %current should be authorized to migrate to the common ancestor */
5101 ret = cgroup_may_write(com_cgrp, sb);
5102 if (ret)
5103 return ret;
5104
5105 /*
5106 * If namespaces are delegation boundaries, %current must be able
5107 * to see both source and destination cgroups from its namespace.
5108 */
5109 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5110 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5111 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5112 return -ENOENT;
5113
5114 return 0;
5115}
5116
5117static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5118 struct cgroup *dst_cgrp,
5119 struct super_block *sb, bool threadgroup,
5120 struct cgroup_namespace *ns)
5121{
5122 int ret = 0;
5123
5124 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5125 if (ret)
5126 return ret;
5127
5128 ret = cgroup_migrate_vet_dst(dst_cgrp);
5129 if (ret)
5130 return ret;
5131
5132 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5133 ret = -EOPNOTSUPP;
5134
5135 return ret;
5136}
5137
5138static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5139 bool threadgroup)
5140{
5141 struct cgroup_file_ctx *ctx = of->priv;
5142 struct cgroup *src_cgrp, *dst_cgrp;
5143 struct task_struct *task;
5144 const struct cred *saved_cred;
5145 ssize_t ret;
5146 bool threadgroup_locked;
5147
5148 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5149 if (!dst_cgrp)
5150 return -ENODEV;
5151
5152 task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
5153 ret = PTR_ERR_OR_ZERO(task);
5154 if (ret)
5155 goto out_unlock;
5156
5157 /* find the source cgroup */
5158 spin_lock_irq(&css_set_lock);
5159 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5160 spin_unlock_irq(&css_set_lock);
5161
5162 /*
5163 * Process and thread migrations follow same delegation rule. Check
5164 * permissions using the credentials from file open to protect against
5165 * inherited fd attacks.
5166 */
5167 saved_cred = override_creds(of->file->f_cred);
5168 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5169 of->file->f_path.dentry->d_sb,
5170 threadgroup, ctx->ns);
5171 revert_creds(saved_cred);
5172 if (ret)
5173 goto out_finish;
5174
5175 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5176
5177out_finish:
5178 cgroup_procs_write_finish(task, threadgroup_locked);
5179out_unlock:
5180 cgroup_kn_unlock(of->kn);
5181
5182 return ret;
5183}
5184
5185static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5186 char *buf, size_t nbytes, loff_t off)
5187{
5188 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5189}
5190
5191static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5192{
5193 return __cgroup_procs_start(s, pos, 0);
5194}
5195
5196static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5197 char *buf, size_t nbytes, loff_t off)
5198{
5199 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5200}
5201
5202/* cgroup core interface files for the default hierarchy */
5203static struct cftype cgroup_base_files[] = {
5204 {
5205 .name = "cgroup.type",
5206 .flags = CFTYPE_NOT_ON_ROOT,
5207 .seq_show = cgroup_type_show,
5208 .write = cgroup_type_write,
5209 },
5210 {
5211 .name = "cgroup.procs",
5212 .flags = CFTYPE_NS_DELEGATABLE,
5213 .file_offset = offsetof(struct cgroup, procs_file),
5214 .release = cgroup_procs_release,
5215 .seq_start = cgroup_procs_start,
5216 .seq_next = cgroup_procs_next,
5217 .seq_show = cgroup_procs_show,
5218 .write = cgroup_procs_write,
5219 },
5220 {
5221 .name = "cgroup.threads",
5222 .flags = CFTYPE_NS_DELEGATABLE,
5223 .release = cgroup_procs_release,
5224 .seq_start = cgroup_threads_start,
5225 .seq_next = cgroup_procs_next,
5226 .seq_show = cgroup_procs_show,
5227 .write = cgroup_threads_write,
5228 },
5229 {
5230 .name = "cgroup.controllers",
5231 .seq_show = cgroup_controllers_show,
5232 },
5233 {
5234 .name = "cgroup.subtree_control",
5235 .flags = CFTYPE_NS_DELEGATABLE,
5236 .seq_show = cgroup_subtree_control_show,
5237 .write = cgroup_subtree_control_write,
5238 },
5239 {
5240 .name = "cgroup.events",
5241 .flags = CFTYPE_NOT_ON_ROOT,
5242 .file_offset = offsetof(struct cgroup, events_file),
5243 .seq_show = cgroup_events_show,
5244 },
5245 {
5246 .name = "cgroup.max.descendants",
5247 .seq_show = cgroup_max_descendants_show,
5248 .write = cgroup_max_descendants_write,
5249 },
5250 {
5251 .name = "cgroup.max.depth",
5252 .seq_show = cgroup_max_depth_show,
5253 .write = cgroup_max_depth_write,
5254 },
5255 {
5256 .name = "cgroup.stat",
5257 .seq_show = cgroup_stat_show,
5258 },
5259 {
5260 .name = "cgroup.freeze",
5261 .flags = CFTYPE_NOT_ON_ROOT,
5262 .seq_show = cgroup_freeze_show,
5263 .write = cgroup_freeze_write,
5264 },
5265 {
5266 .name = "cgroup.kill",
5267 .flags = CFTYPE_NOT_ON_ROOT,
5268 .write = cgroup_kill_write,
5269 },
5270 {
5271 .name = "cpu.stat",
5272 .seq_show = cpu_stat_show,
5273 },
5274 {
5275 .name = "cpu.stat.local",
5276 .seq_show = cpu_local_stat_show,
5277 },
5278 { } /* terminate */
5279};
5280
5281static struct cftype cgroup_psi_files[] = {
5282#ifdef CONFIG_PSI
5283 {
5284 .name = "io.pressure",
5285 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5286 .seq_show = cgroup_io_pressure_show,
5287 .write = cgroup_io_pressure_write,
5288 .poll = cgroup_pressure_poll,
5289 .release = cgroup_pressure_release,
5290 },
5291 {
5292 .name = "memory.pressure",
5293 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5294 .seq_show = cgroup_memory_pressure_show,
5295 .write = cgroup_memory_pressure_write,
5296 .poll = cgroup_pressure_poll,
5297 .release = cgroup_pressure_release,
5298 },
5299 {
5300 .name = "cpu.pressure",
5301 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5302 .seq_show = cgroup_cpu_pressure_show,
5303 .write = cgroup_cpu_pressure_write,
5304 .poll = cgroup_pressure_poll,
5305 .release = cgroup_pressure_release,
5306 },
5307#ifdef CONFIG_IRQ_TIME_ACCOUNTING
5308 {
5309 .name = "irq.pressure",
5310 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5311 .seq_show = cgroup_irq_pressure_show,
5312 .write = cgroup_irq_pressure_write,
5313 .poll = cgroup_pressure_poll,
5314 .release = cgroup_pressure_release,
5315 },
5316#endif
5317 {
5318 .name = "cgroup.pressure",
5319 .seq_show = cgroup_pressure_show,
5320 .write = cgroup_pressure_write,
5321 },
5322#endif /* CONFIG_PSI */
5323 { } /* terminate */
5324};
5325
5326/*
5327 * css destruction is four-stage process.
5328 *
5329 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5330 * Implemented in kill_css().
5331 *
5332 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5333 * and thus css_tryget_online() is guaranteed to fail, the css can be
5334 * offlined by invoking offline_css(). After offlining, the base ref is
5335 * put. Implemented in css_killed_work_fn().
5336 *
5337 * 3. When the percpu_ref reaches zero, the only possible remaining
5338 * accessors are inside RCU read sections. css_release() schedules the
5339 * RCU callback.
5340 *
5341 * 4. After the grace period, the css can be freed. Implemented in
5342 * css_free_rwork_fn().
5343 *
5344 * It is actually hairier because both step 2 and 4 require process context
5345 * and thus involve punting to css->destroy_work adding two additional
5346 * steps to the already complex sequence.
5347 */
5348static void css_free_rwork_fn(struct work_struct *work)
5349{
5350 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5351 struct cgroup_subsys_state, destroy_rwork);
5352 struct cgroup_subsys *ss = css->ss;
5353 struct cgroup *cgrp = css->cgroup;
5354
5355 percpu_ref_exit(&css->refcnt);
5356
5357 if (ss) {
5358 /* css free path */
5359 struct cgroup_subsys_state *parent = css->parent;
5360 int id = css->id;
5361
5362 ss->css_free(css);
5363 cgroup_idr_remove(&ss->css_idr, id);
5364 cgroup_put(cgrp);
5365
5366 if (parent)
5367 css_put(parent);
5368 } else {
5369 /* cgroup free path */
5370 atomic_dec(&cgrp->root->nr_cgrps);
5371 cgroup1_pidlist_destroy_all(cgrp);
5372 cancel_work_sync(&cgrp->release_agent_work);
5373 bpf_cgrp_storage_free(cgrp);
5374
5375 if (cgroup_parent(cgrp)) {
5376 /*
5377 * We get a ref to the parent, and put the ref when
5378 * this cgroup is being freed, so it's guaranteed
5379 * that the parent won't be destroyed before its
5380 * children.
5381 */
5382 cgroup_put(cgroup_parent(cgrp));
5383 kernfs_put(cgrp->kn);
5384 psi_cgroup_free(cgrp);
5385 cgroup_rstat_exit(cgrp);
5386 kfree(cgrp);
5387 } else {
5388 /*
5389 * This is root cgroup's refcnt reaching zero,
5390 * which indicates that the root should be
5391 * released.
5392 */
5393 cgroup_destroy_root(cgrp->root);
5394 }
5395 }
5396}
5397
5398static void css_release_work_fn(struct work_struct *work)
5399{
5400 struct cgroup_subsys_state *css =
5401 container_of(work, struct cgroup_subsys_state, destroy_work);
5402 struct cgroup_subsys *ss = css->ss;
5403 struct cgroup *cgrp = css->cgroup;
5404
5405 cgroup_lock();
5406
5407 css->flags |= CSS_RELEASED;
5408 list_del_rcu(&css->sibling);
5409
5410 if (ss) {
5411 /* css release path */
5412 if (!list_empty(&css->rstat_css_node)) {
5413 cgroup_rstat_flush(cgrp);
5414 list_del_rcu(&css->rstat_css_node);
5415 }
5416
5417 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5418 if (ss->css_released)
5419 ss->css_released(css);
5420 } else {
5421 struct cgroup *tcgrp;
5422
5423 /* cgroup release path */
5424 TRACE_CGROUP_PATH(release, cgrp);
5425
5426 cgroup_rstat_flush(cgrp);
5427
5428 spin_lock_irq(&css_set_lock);
5429 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5430 tcgrp = cgroup_parent(tcgrp))
5431 tcgrp->nr_dying_descendants--;
5432 spin_unlock_irq(&css_set_lock);
5433
5434 /*
5435 * There are two control paths which try to determine
5436 * cgroup from dentry without going through kernfs -
5437 * cgroupstats_build() and css_tryget_online_from_dir().
5438 * Those are supported by RCU protecting clearing of
5439 * cgrp->kn->priv backpointer.
5440 */
5441 if (cgrp->kn)
5442 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5443 NULL);
5444 }
5445
5446 cgroup_unlock();
5447
5448 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5449 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5450}
5451
5452static void css_release(struct percpu_ref *ref)
5453{
5454 struct cgroup_subsys_state *css =
5455 container_of(ref, struct cgroup_subsys_state, refcnt);
5456
5457 INIT_WORK(&css->destroy_work, css_release_work_fn);
5458 queue_work(cgroup_destroy_wq, &css->destroy_work);
5459}
5460
5461static void init_and_link_css(struct cgroup_subsys_state *css,
5462 struct cgroup_subsys *ss, struct cgroup *cgrp)
5463{
5464 lockdep_assert_held(&cgroup_mutex);
5465
5466 cgroup_get_live(cgrp);
5467
5468 memset(css, 0, sizeof(*css));
5469 css->cgroup = cgrp;
5470 css->ss = ss;
5471 css->id = -1;
5472 INIT_LIST_HEAD(&css->sibling);
5473 INIT_LIST_HEAD(&css->children);
5474 INIT_LIST_HEAD(&css->rstat_css_node);
5475 css->serial_nr = css_serial_nr_next++;
5476 atomic_set(&css->online_cnt, 0);
5477
5478 if (cgroup_parent(cgrp)) {
5479 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5480 css_get(css->parent);
5481 }
5482
5483 if (ss->css_rstat_flush)
5484 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5485
5486 BUG_ON(cgroup_css(cgrp, ss));
5487}
5488
5489/* invoke ->css_online() on a new CSS and mark it online if successful */
5490static int online_css(struct cgroup_subsys_state *css)
5491{
5492 struct cgroup_subsys *ss = css->ss;
5493 int ret = 0;
5494
5495 lockdep_assert_held(&cgroup_mutex);
5496
5497 if (ss->css_online)
5498 ret = ss->css_online(css);
5499 if (!ret) {
5500 css->flags |= CSS_ONLINE;
5501 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5502
5503 atomic_inc(&css->online_cnt);
5504 if (css->parent)
5505 atomic_inc(&css->parent->online_cnt);
5506 }
5507 return ret;
5508}
5509
5510/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5511static void offline_css(struct cgroup_subsys_state *css)
5512{
5513 struct cgroup_subsys *ss = css->ss;
5514
5515 lockdep_assert_held(&cgroup_mutex);
5516
5517 if (!(css->flags & CSS_ONLINE))
5518 return;
5519
5520 if (ss->css_offline)
5521 ss->css_offline(css);
5522
5523 css->flags &= ~CSS_ONLINE;
5524 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5525
5526 wake_up_all(&css->cgroup->offline_waitq);
5527}
5528
5529/**
5530 * css_create - create a cgroup_subsys_state
5531 * @cgrp: the cgroup new css will be associated with
5532 * @ss: the subsys of new css
5533 *
5534 * Create a new css associated with @cgrp - @ss pair. On success, the new
5535 * css is online and installed in @cgrp. This function doesn't create the
5536 * interface files. Returns 0 on success, -errno on failure.
5537 */
5538static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5539 struct cgroup_subsys *ss)
5540{
5541 struct cgroup *parent = cgroup_parent(cgrp);
5542 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5543 struct cgroup_subsys_state *css;
5544 int err;
5545
5546 lockdep_assert_held(&cgroup_mutex);
5547
5548 css = ss->css_alloc(parent_css);
5549 if (!css)
5550 css = ERR_PTR(-ENOMEM);
5551 if (IS_ERR(css))
5552 return css;
5553
5554 init_and_link_css(css, ss, cgrp);
5555
5556 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5557 if (err)
5558 goto err_free_css;
5559
5560 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5561 if (err < 0)
5562 goto err_free_css;
5563 css->id = err;
5564
5565 /* @css is ready to be brought online now, make it visible */
5566 list_add_tail_rcu(&css->sibling, &parent_css->children);
5567 cgroup_idr_replace(&ss->css_idr, css, css->id);
5568
5569 err = online_css(css);
5570 if (err)
5571 goto err_list_del;
5572
5573 return css;
5574
5575err_list_del:
5576 list_del_rcu(&css->sibling);
5577err_free_css:
5578 list_del_rcu(&css->rstat_css_node);
5579 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5580 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5581 return ERR_PTR(err);
5582}
5583
5584/*
5585 * The returned cgroup is fully initialized including its control mask, but
5586 * it doesn't have the control mask applied.
5587 */
5588static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5589 umode_t mode)
5590{
5591 struct cgroup_root *root = parent->root;
5592 struct cgroup *cgrp, *tcgrp;
5593 struct kernfs_node *kn;
5594 int level = parent->level + 1;
5595 int ret;
5596
5597 /* allocate the cgroup and its ID, 0 is reserved for the root */
5598 cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5599 if (!cgrp)
5600 return ERR_PTR(-ENOMEM);
5601
5602 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5603 if (ret)
5604 goto out_free_cgrp;
5605
5606 ret = cgroup_rstat_init(cgrp);
5607 if (ret)
5608 goto out_cancel_ref;
5609
5610 /* create the directory */
5611 kn = kernfs_create_dir_ns(parent->kn, name, mode,
5612 current_fsuid(), current_fsgid(),
5613 cgrp, NULL);
5614 if (IS_ERR(kn)) {
5615 ret = PTR_ERR(kn);
5616 goto out_stat_exit;
5617 }
5618 cgrp->kn = kn;
5619
5620 init_cgroup_housekeeping(cgrp);
5621
5622 cgrp->self.parent = &parent->self;
5623 cgrp->root = root;
5624 cgrp->level = level;
5625
5626 ret = psi_cgroup_alloc(cgrp);
5627 if (ret)
5628 goto out_kernfs_remove;
5629
5630 ret = cgroup_bpf_inherit(cgrp);
5631 if (ret)
5632 goto out_psi_free;
5633
5634 /*
5635 * New cgroup inherits effective freeze counter, and
5636 * if the parent has to be frozen, the child has too.
5637 */
5638 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5639 if (cgrp->freezer.e_freeze) {
5640 /*
5641 * Set the CGRP_FREEZE flag, so when a process will be
5642 * attached to the child cgroup, it will become frozen.
5643 * At this point the new cgroup is unpopulated, so we can
5644 * consider it frozen immediately.
5645 */
5646 set_bit(CGRP_FREEZE, &cgrp->flags);
5647 set_bit(CGRP_FROZEN, &cgrp->flags);
5648 }
5649
5650 spin_lock_irq(&css_set_lock);
5651 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5652 cgrp->ancestors[tcgrp->level] = tcgrp;
5653
5654 if (tcgrp != cgrp) {
5655 tcgrp->nr_descendants++;
5656
5657 /*
5658 * If the new cgroup is frozen, all ancestor cgroups
5659 * get a new frozen descendant, but their state can't
5660 * change because of this.
5661 */
5662 if (cgrp->freezer.e_freeze)
5663 tcgrp->freezer.nr_frozen_descendants++;
5664 }
5665 }
5666 spin_unlock_irq(&css_set_lock);
5667
5668 if (notify_on_release(parent))
5669 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5670
5671 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5672 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5673
5674 cgrp->self.serial_nr = css_serial_nr_next++;
5675
5676 /* allocation complete, commit to creation */
5677 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5678 atomic_inc(&root->nr_cgrps);
5679 cgroup_get_live(parent);
5680
5681 /*
5682 * On the default hierarchy, a child doesn't automatically inherit
5683 * subtree_control from the parent. Each is configured manually.
5684 */
5685 if (!cgroup_on_dfl(cgrp))
5686 cgrp->subtree_control = cgroup_control(cgrp);
5687
5688 cgroup_propagate_control(cgrp);
5689
5690 return cgrp;
5691
5692out_psi_free:
5693 psi_cgroup_free(cgrp);
5694out_kernfs_remove:
5695 kernfs_remove(cgrp->kn);
5696out_stat_exit:
5697 cgroup_rstat_exit(cgrp);
5698out_cancel_ref:
5699 percpu_ref_exit(&cgrp->self.refcnt);
5700out_free_cgrp:
5701 kfree(cgrp);
5702 return ERR_PTR(ret);
5703}
5704
5705static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5706{
5707 struct cgroup *cgroup;
5708 int ret = false;
5709 int level = 1;
5710
5711 lockdep_assert_held(&cgroup_mutex);
5712
5713 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5714 if (cgroup->nr_descendants >= cgroup->max_descendants)
5715 goto fail;
5716
5717 if (level > cgroup->max_depth)
5718 goto fail;
5719
5720 level++;
5721 }
5722
5723 ret = true;
5724fail:
5725 return ret;
5726}
5727
5728int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5729{
5730 struct cgroup *parent, *cgrp;
5731 int ret;
5732
5733 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5734 if (strchr(name, '\n'))
5735 return -EINVAL;
5736
5737 parent = cgroup_kn_lock_live(parent_kn, false);
5738 if (!parent)
5739 return -ENODEV;
5740
5741 if (!cgroup_check_hierarchy_limits(parent)) {
5742 ret = -EAGAIN;
5743 goto out_unlock;
5744 }
5745
5746 cgrp = cgroup_create(parent, name, mode);
5747 if (IS_ERR(cgrp)) {
5748 ret = PTR_ERR(cgrp);
5749 goto out_unlock;
5750 }
5751
5752 /*
5753 * This extra ref will be put in cgroup_free_fn() and guarantees
5754 * that @cgrp->kn is always accessible.
5755 */
5756 kernfs_get(cgrp->kn);
5757
5758 ret = css_populate_dir(&cgrp->self);
5759 if (ret)
5760 goto out_destroy;
5761
5762 ret = cgroup_apply_control_enable(cgrp);
5763 if (ret)
5764 goto out_destroy;
5765
5766 TRACE_CGROUP_PATH(mkdir, cgrp);
5767
5768 /* let's create and online css's */
5769 kernfs_activate(cgrp->kn);
5770
5771 ret = 0;
5772 goto out_unlock;
5773
5774out_destroy:
5775 cgroup_destroy_locked(cgrp);
5776out_unlock:
5777 cgroup_kn_unlock(parent_kn);
5778 return ret;
5779}
5780
5781/*
5782 * This is called when the refcnt of a css is confirmed to be killed.
5783 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5784 * initiate destruction and put the css ref from kill_css().
5785 */
5786static void css_killed_work_fn(struct work_struct *work)
5787{
5788 struct cgroup_subsys_state *css =
5789 container_of(work, struct cgroup_subsys_state, destroy_work);
5790
5791 cgroup_lock();
5792
5793 do {
5794 offline_css(css);
5795 css_put(css);
5796 /* @css can't go away while we're holding cgroup_mutex */
5797 css = css->parent;
5798 } while (css && atomic_dec_and_test(&css->online_cnt));
5799
5800 cgroup_unlock();
5801}
5802
5803/* css kill confirmation processing requires process context, bounce */
5804static void css_killed_ref_fn(struct percpu_ref *ref)
5805{
5806 struct cgroup_subsys_state *css =
5807 container_of(ref, struct cgroup_subsys_state, refcnt);
5808
5809 if (atomic_dec_and_test(&css->online_cnt)) {
5810 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5811 queue_work(cgroup_destroy_wq, &css->destroy_work);
5812 }
5813}
5814
5815/**
5816 * kill_css - destroy a css
5817 * @css: css to destroy
5818 *
5819 * This function initiates destruction of @css by removing cgroup interface
5820 * files and putting its base reference. ->css_offline() will be invoked
5821 * asynchronously once css_tryget_online() is guaranteed to fail and when
5822 * the reference count reaches zero, @css will be released.
5823 */
5824static void kill_css(struct cgroup_subsys_state *css)
5825{
5826 lockdep_assert_held(&cgroup_mutex);
5827
5828 if (css->flags & CSS_DYING)
5829 return;
5830
5831 css->flags |= CSS_DYING;
5832
5833 /*
5834 * This must happen before css is disassociated with its cgroup.
5835 * See seq_css() for details.
5836 */
5837 css_clear_dir(css);
5838
5839 /*
5840 * Killing would put the base ref, but we need to keep it alive
5841 * until after ->css_offline().
5842 */
5843 css_get(css);
5844
5845 /*
5846 * cgroup core guarantees that, by the time ->css_offline() is
5847 * invoked, no new css reference will be given out via
5848 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5849 * proceed to offlining css's because percpu_ref_kill() doesn't
5850 * guarantee that the ref is seen as killed on all CPUs on return.
5851 *
5852 * Use percpu_ref_kill_and_confirm() to get notifications as each
5853 * css is confirmed to be seen as killed on all CPUs.
5854 */
5855 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5856}
5857
5858/**
5859 * cgroup_destroy_locked - the first stage of cgroup destruction
5860 * @cgrp: cgroup to be destroyed
5861 *
5862 * css's make use of percpu refcnts whose killing latency shouldn't be
5863 * exposed to userland and are RCU protected. Also, cgroup core needs to
5864 * guarantee that css_tryget_online() won't succeed by the time
5865 * ->css_offline() is invoked. To satisfy all the requirements,
5866 * destruction is implemented in the following two steps.
5867 *
5868 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5869 * userland visible parts and start killing the percpu refcnts of
5870 * css's. Set up so that the next stage will be kicked off once all
5871 * the percpu refcnts are confirmed to be killed.
5872 *
5873 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5874 * rest of destruction. Once all cgroup references are gone, the
5875 * cgroup is RCU-freed.
5876 *
5877 * This function implements s1. After this step, @cgrp is gone as far as
5878 * the userland is concerned and a new cgroup with the same name may be
5879 * created. As cgroup doesn't care about the names internally, this
5880 * doesn't cause any problem.
5881 */
5882static int cgroup_destroy_locked(struct cgroup *cgrp)
5883 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5884{
5885 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5886 struct cgroup_subsys_state *css;
5887 struct cgrp_cset_link *link;
5888 int ssid;
5889
5890 lockdep_assert_held(&cgroup_mutex);
5891
5892 /*
5893 * Only migration can raise populated from zero and we're already
5894 * holding cgroup_mutex.
5895 */
5896 if (cgroup_is_populated(cgrp))
5897 return -EBUSY;
5898
5899 /*
5900 * Make sure there's no live children. We can't test emptiness of
5901 * ->self.children as dead children linger on it while being
5902 * drained; otherwise, "rmdir parent/child parent" may fail.
5903 */
5904 if (css_has_online_children(&cgrp->self))
5905 return -EBUSY;
5906
5907 /*
5908 * Mark @cgrp and the associated csets dead. The former prevents
5909 * further task migration and child creation by disabling
5910 * cgroup_kn_lock_live(). The latter makes the csets ignored by
5911 * the migration path.
5912 */
5913 cgrp->self.flags &= ~CSS_ONLINE;
5914
5915 spin_lock_irq(&css_set_lock);
5916 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5917 link->cset->dead = true;
5918 spin_unlock_irq(&css_set_lock);
5919
5920 /* initiate massacre of all css's */
5921 for_each_css(css, ssid, cgrp)
5922 kill_css(css);
5923
5924 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5925 css_clear_dir(&cgrp->self);
5926 kernfs_remove(cgrp->kn);
5927
5928 if (cgroup_is_threaded(cgrp))
5929 parent->nr_threaded_children--;
5930
5931 spin_lock_irq(&css_set_lock);
5932 for (tcgrp = parent; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5933 tcgrp->nr_descendants--;
5934 tcgrp->nr_dying_descendants++;
5935 /*
5936 * If the dying cgroup is frozen, decrease frozen descendants
5937 * counters of ancestor cgroups.
5938 */
5939 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5940 tcgrp->freezer.nr_frozen_descendants--;
5941 }
5942 spin_unlock_irq(&css_set_lock);
5943
5944 cgroup1_check_for_release(parent);
5945
5946 cgroup_bpf_offline(cgrp);
5947
5948 /* put the base reference */
5949 percpu_ref_kill(&cgrp->self.refcnt);
5950
5951 return 0;
5952};
5953
5954int cgroup_rmdir(struct kernfs_node *kn)
5955{
5956 struct cgroup *cgrp;
5957 int ret = 0;
5958
5959 cgrp = cgroup_kn_lock_live(kn, false);
5960 if (!cgrp)
5961 return 0;
5962
5963 ret = cgroup_destroy_locked(cgrp);
5964 if (!ret)
5965 TRACE_CGROUP_PATH(rmdir, cgrp);
5966
5967 cgroup_kn_unlock(kn);
5968 return ret;
5969}
5970
5971static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5972 .show_options = cgroup_show_options,
5973 .mkdir = cgroup_mkdir,
5974 .rmdir = cgroup_rmdir,
5975 .show_path = cgroup_show_path,
5976};
5977
5978static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5979{
5980 struct cgroup_subsys_state *css;
5981
5982 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5983
5984 cgroup_lock();
5985
5986 idr_init(&ss->css_idr);
5987 INIT_LIST_HEAD(&ss->cfts);
5988
5989 /* Create the root cgroup state for this subsystem */
5990 ss->root = &cgrp_dfl_root;
5991 css = ss->css_alloc(NULL);
5992 /* We don't handle early failures gracefully */
5993 BUG_ON(IS_ERR(css));
5994 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5995
5996 /*
5997 * Root csses are never destroyed and we can't initialize
5998 * percpu_ref during early init. Disable refcnting.
5999 */
6000 css->flags |= CSS_NO_REF;
6001
6002 if (early) {
6003 /* allocation can't be done safely during early init */
6004 css->id = 1;
6005 } else {
6006 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
6007 BUG_ON(css->id < 0);
6008 }
6009
6010 /* Update the init_css_set to contain a subsys
6011 * pointer to this state - since the subsystem is
6012 * newly registered, all tasks and hence the
6013 * init_css_set is in the subsystem's root cgroup. */
6014 init_css_set.subsys[ss->id] = css;
6015
6016 have_fork_callback |= (bool)ss->fork << ss->id;
6017 have_exit_callback |= (bool)ss->exit << ss->id;
6018 have_release_callback |= (bool)ss->release << ss->id;
6019 have_canfork_callback |= (bool)ss->can_fork << ss->id;
6020
6021 /* At system boot, before all subsystems have been
6022 * registered, no tasks have been forked, so we don't
6023 * need to invoke fork callbacks here. */
6024 BUG_ON(!list_empty(&init_task.tasks));
6025
6026 BUG_ON(online_css(css));
6027
6028 cgroup_unlock();
6029}
6030
6031/**
6032 * cgroup_init_early - cgroup initialization at system boot
6033 *
6034 * Initialize cgroups at system boot, and initialize any
6035 * subsystems that request early init.
6036 */
6037int __init cgroup_init_early(void)
6038{
6039 static struct cgroup_fs_context __initdata ctx;
6040 struct cgroup_subsys *ss;
6041 int i;
6042
6043 ctx.root = &cgrp_dfl_root;
6044 init_cgroup_root(&ctx);
6045 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6046
6047 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6048
6049 for_each_subsys(ss, i) {
6050 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6051 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6052 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6053 ss->id, ss->name);
6054 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6055 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6056
6057 ss->id = i;
6058 ss->name = cgroup_subsys_name[i];
6059 if (!ss->legacy_name)
6060 ss->legacy_name = cgroup_subsys_name[i];
6061
6062 if (ss->early_init)
6063 cgroup_init_subsys(ss, true);
6064 }
6065 return 0;
6066}
6067
6068/**
6069 * cgroup_init - cgroup initialization
6070 *
6071 * Register cgroup filesystem and /proc file, and initialize
6072 * any subsystems that didn't request early init.
6073 */
6074int __init cgroup_init(void)
6075{
6076 struct cgroup_subsys *ss;
6077 int ssid;
6078
6079 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6080 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6081 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6082 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6083
6084 cgroup_rstat_boot();
6085
6086 get_user_ns(init_cgroup_ns.user_ns);
6087
6088 cgroup_lock();
6089
6090 /*
6091 * Add init_css_set to the hash table so that dfl_root can link to
6092 * it during init.
6093 */
6094 hash_add(css_set_table, &init_css_set.hlist,
6095 css_set_hash(init_css_set.subsys));
6096
6097 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6098
6099 cgroup_unlock();
6100
6101 for_each_subsys(ss, ssid) {
6102 if (ss->early_init) {
6103 struct cgroup_subsys_state *css =
6104 init_css_set.subsys[ss->id];
6105
6106 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6107 GFP_KERNEL);
6108 BUG_ON(css->id < 0);
6109 } else {
6110 cgroup_init_subsys(ss, false);
6111 }
6112
6113 list_add_tail(&init_css_set.e_cset_node[ssid],
6114 &cgrp_dfl_root.cgrp.e_csets[ssid]);
6115
6116 /*
6117 * Setting dfl_root subsys_mask needs to consider the
6118 * disabled flag and cftype registration needs kmalloc,
6119 * both of which aren't available during early_init.
6120 */
6121 if (!cgroup_ssid_enabled(ssid))
6122 continue;
6123
6124 if (cgroup1_ssid_disabled(ssid))
6125 pr_info("Disabling %s control group subsystem in v1 mounts\n",
6126 ss->legacy_name);
6127
6128 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6129
6130 /* implicit controllers must be threaded too */
6131 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6132
6133 if (ss->implicit_on_dfl)
6134 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6135 else if (!ss->dfl_cftypes)
6136 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6137
6138 if (ss->threaded)
6139 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6140
6141 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6142 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6143 } else {
6144 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6145 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6146 }
6147
6148 if (ss->bind)
6149 ss->bind(init_css_set.subsys[ssid]);
6150
6151 cgroup_lock();
6152 css_populate_dir(init_css_set.subsys[ssid]);
6153 cgroup_unlock();
6154 }
6155
6156 /* init_css_set.subsys[] has been updated, re-hash */
6157 hash_del(&init_css_set.hlist);
6158 hash_add(css_set_table, &init_css_set.hlist,
6159 css_set_hash(init_css_set.subsys));
6160
6161 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6162 WARN_ON(register_filesystem(&cgroup_fs_type));
6163 WARN_ON(register_filesystem(&cgroup2_fs_type));
6164 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6165#ifdef CONFIG_CPUSETS
6166 WARN_ON(register_filesystem(&cpuset_fs_type));
6167#endif
6168
6169 return 0;
6170}
6171
6172static int __init cgroup_wq_init(void)
6173{
6174 /*
6175 * There isn't much point in executing destruction path in
6176 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6177 * Use 1 for @max_active.
6178 *
6179 * We would prefer to do this in cgroup_init() above, but that
6180 * is called before init_workqueues(): so leave this until after.
6181 */
6182 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
6183 BUG_ON(!cgroup_destroy_wq);
6184 return 0;
6185}
6186core_initcall(cgroup_wq_init);
6187
6188void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6189{
6190 struct kernfs_node *kn;
6191
6192 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6193 if (!kn)
6194 return;
6195 kernfs_path(kn, buf, buflen);
6196 kernfs_put(kn);
6197}
6198
6199/*
6200 * cgroup_get_from_id : get the cgroup associated with cgroup id
6201 * @id: cgroup id
6202 * On success return the cgrp or ERR_PTR on failure
6203 * Only cgroups within current task's cgroup NS are valid.
6204 */
6205struct cgroup *cgroup_get_from_id(u64 id)
6206{
6207 struct kernfs_node *kn;
6208 struct cgroup *cgrp, *root_cgrp;
6209
6210 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6211 if (!kn)
6212 return ERR_PTR(-ENOENT);
6213
6214 if (kernfs_type(kn) != KERNFS_DIR) {
6215 kernfs_put(kn);
6216 return ERR_PTR(-ENOENT);
6217 }
6218
6219 rcu_read_lock();
6220
6221 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6222 if (cgrp && !cgroup_tryget(cgrp))
6223 cgrp = NULL;
6224
6225 rcu_read_unlock();
6226 kernfs_put(kn);
6227
6228 if (!cgrp)
6229 return ERR_PTR(-ENOENT);
6230
6231 root_cgrp = current_cgns_cgroup_dfl();
6232 if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6233 cgroup_put(cgrp);
6234 return ERR_PTR(-ENOENT);
6235 }
6236
6237 return cgrp;
6238}
6239EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6240
6241/*
6242 * proc_cgroup_show()
6243 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6244 * - Used for /proc/<pid>/cgroup.
6245 */
6246int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6247 struct pid *pid, struct task_struct *tsk)
6248{
6249 char *buf;
6250 int retval;
6251 struct cgroup_root *root;
6252
6253 retval = -ENOMEM;
6254 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6255 if (!buf)
6256 goto out;
6257
6258 rcu_read_lock();
6259 spin_lock_irq(&css_set_lock);
6260
6261 for_each_root(root) {
6262 struct cgroup_subsys *ss;
6263 struct cgroup *cgrp;
6264 int ssid, count = 0;
6265
6266 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6267 continue;
6268
6269 cgrp = task_cgroup_from_root(tsk, root);
6270 /* The root has already been unmounted. */
6271 if (!cgrp)
6272 continue;
6273
6274 seq_printf(m, "%d:", root->hierarchy_id);
6275 if (root != &cgrp_dfl_root)
6276 for_each_subsys(ss, ssid)
6277 if (root->subsys_mask & (1 << ssid))
6278 seq_printf(m, "%s%s", count++ ? "," : "",
6279 ss->legacy_name);
6280 if (strlen(root->name))
6281 seq_printf(m, "%sname=%s", count ? "," : "",
6282 root->name);
6283 seq_putc(m, ':');
6284 /*
6285 * On traditional hierarchies, all zombie tasks show up as
6286 * belonging to the root cgroup. On the default hierarchy,
6287 * while a zombie doesn't show up in "cgroup.procs" and
6288 * thus can't be migrated, its /proc/PID/cgroup keeps
6289 * reporting the cgroup it belonged to before exiting. If
6290 * the cgroup is removed before the zombie is reaped,
6291 * " (deleted)" is appended to the cgroup path.
6292 */
6293 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6294 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6295 current->nsproxy->cgroup_ns);
6296 if (retval == -E2BIG)
6297 retval = -ENAMETOOLONG;
6298 if (retval < 0)
6299 goto out_unlock;
6300
6301 seq_puts(m, buf);
6302 } else {
6303 seq_puts(m, "/");
6304 }
6305
6306 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6307 seq_puts(m, " (deleted)\n");
6308 else
6309 seq_putc(m, '\n');
6310 }
6311
6312 retval = 0;
6313out_unlock:
6314 spin_unlock_irq(&css_set_lock);
6315 rcu_read_unlock();
6316 kfree(buf);
6317out:
6318 return retval;
6319}
6320
6321/**
6322 * cgroup_fork - initialize cgroup related fields during copy_process()
6323 * @child: pointer to task_struct of forking parent process.
6324 *
6325 * A task is associated with the init_css_set until cgroup_post_fork()
6326 * attaches it to the target css_set.
6327 */
6328void cgroup_fork(struct task_struct *child)
6329{
6330 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6331 INIT_LIST_HEAD(&child->cg_list);
6332}
6333
6334/**
6335 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6336 * @f: file corresponding to cgroup_dir
6337 *
6338 * Find the cgroup from a file pointer associated with a cgroup directory.
6339 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6340 * cgroup cannot be found.
6341 */
6342static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6343{
6344 struct cgroup_subsys_state *css;
6345
6346 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6347 if (IS_ERR(css))
6348 return ERR_CAST(css);
6349
6350 return css->cgroup;
6351}
6352
6353/**
6354 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6355 * cgroup2.
6356 * @f: file corresponding to cgroup2_dir
6357 */
6358static struct cgroup *cgroup_get_from_file(struct file *f)
6359{
6360 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6361
6362 if (IS_ERR(cgrp))
6363 return ERR_CAST(cgrp);
6364
6365 if (!cgroup_on_dfl(cgrp)) {
6366 cgroup_put(cgrp);
6367 return ERR_PTR(-EBADF);
6368 }
6369
6370 return cgrp;
6371}
6372
6373/**
6374 * cgroup_css_set_fork - find or create a css_set for a child process
6375 * @kargs: the arguments passed to create the child process
6376 *
6377 * This functions finds or creates a new css_set which the child
6378 * process will be attached to in cgroup_post_fork(). By default,
6379 * the child process will be given the same css_set as its parent.
6380 *
6381 * If CLONE_INTO_CGROUP is specified this function will try to find an
6382 * existing css_set which includes the requested cgroup and if not create
6383 * a new css_set that the child will be attached to later. If this function
6384 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6385 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6386 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6387 * to the target cgroup.
6388 */
6389static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6390 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6391{
6392 int ret;
6393 struct cgroup *dst_cgrp = NULL;
6394 struct css_set *cset;
6395 struct super_block *sb;
6396 struct file *f;
6397
6398 if (kargs->flags & CLONE_INTO_CGROUP)
6399 cgroup_lock();
6400
6401 cgroup_threadgroup_change_begin(current);
6402
6403 spin_lock_irq(&css_set_lock);
6404 cset = task_css_set(current);
6405 get_css_set(cset);
6406 spin_unlock_irq(&css_set_lock);
6407
6408 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6409 kargs->cset = cset;
6410 return 0;
6411 }
6412
6413 f = fget_raw(kargs->cgroup);
6414 if (!f) {
6415 ret = -EBADF;
6416 goto err;
6417 }
6418 sb = f->f_path.dentry->d_sb;
6419
6420 dst_cgrp = cgroup_get_from_file(f);
6421 if (IS_ERR(dst_cgrp)) {
6422 ret = PTR_ERR(dst_cgrp);
6423 dst_cgrp = NULL;
6424 goto err;
6425 }
6426
6427 if (cgroup_is_dead(dst_cgrp)) {
6428 ret = -ENODEV;
6429 goto err;
6430 }
6431
6432 /*
6433 * Verify that we the target cgroup is writable for us. This is
6434 * usually done by the vfs layer but since we're not going through
6435 * the vfs layer here we need to do it "manually".
6436 */
6437 ret = cgroup_may_write(dst_cgrp, sb);
6438 if (ret)
6439 goto err;
6440
6441 /*
6442 * Spawning a task directly into a cgroup works by passing a file
6443 * descriptor to the target cgroup directory. This can even be an O_PATH
6444 * file descriptor. But it can never be a cgroup.procs file descriptor.
6445 * This was done on purpose so spawning into a cgroup could be
6446 * conceptualized as an atomic
6447 *
6448 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6449 * write(fd, <child-pid>, ...);
6450 *
6451 * sequence, i.e. it's a shorthand for the caller opening and writing
6452 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6453 * to always use the caller's credentials.
6454 */
6455 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6456 !(kargs->flags & CLONE_THREAD),
6457 current->nsproxy->cgroup_ns);
6458 if (ret)
6459 goto err;
6460
6461 kargs->cset = find_css_set(cset, dst_cgrp);
6462 if (!kargs->cset) {
6463 ret = -ENOMEM;
6464 goto err;
6465 }
6466
6467 put_css_set(cset);
6468 fput(f);
6469 kargs->cgrp = dst_cgrp;
6470 return ret;
6471
6472err:
6473 cgroup_threadgroup_change_end(current);
6474 cgroup_unlock();
6475 if (f)
6476 fput(f);
6477 if (dst_cgrp)
6478 cgroup_put(dst_cgrp);
6479 put_css_set(cset);
6480 if (kargs->cset)
6481 put_css_set(kargs->cset);
6482 return ret;
6483}
6484
6485/**
6486 * cgroup_css_set_put_fork - drop references we took during fork
6487 * @kargs: the arguments passed to create the child process
6488 *
6489 * Drop references to the prepared css_set and target cgroup if
6490 * CLONE_INTO_CGROUP was requested.
6491 */
6492static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6493 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6494{
6495 struct cgroup *cgrp = kargs->cgrp;
6496 struct css_set *cset = kargs->cset;
6497
6498 cgroup_threadgroup_change_end(current);
6499
6500 if (cset) {
6501 put_css_set(cset);
6502 kargs->cset = NULL;
6503 }
6504
6505 if (kargs->flags & CLONE_INTO_CGROUP) {
6506 cgroup_unlock();
6507 if (cgrp) {
6508 cgroup_put(cgrp);
6509 kargs->cgrp = NULL;
6510 }
6511 }
6512}
6513
6514/**
6515 * cgroup_can_fork - called on a new task before the process is exposed
6516 * @child: the child process
6517 * @kargs: the arguments passed to create the child process
6518 *
6519 * This prepares a new css_set for the child process which the child will
6520 * be attached to in cgroup_post_fork().
6521 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6522 * callback returns an error, the fork aborts with that error code. This
6523 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6524 */
6525int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6526{
6527 struct cgroup_subsys *ss;
6528 int i, j, ret;
6529
6530 ret = cgroup_css_set_fork(kargs);
6531 if (ret)
6532 return ret;
6533
6534 do_each_subsys_mask(ss, i, have_canfork_callback) {
6535 ret = ss->can_fork(child, kargs->cset);
6536 if (ret)
6537 goto out_revert;
6538 } while_each_subsys_mask();
6539
6540 return 0;
6541
6542out_revert:
6543 for_each_subsys(ss, j) {
6544 if (j >= i)
6545 break;
6546 if (ss->cancel_fork)
6547 ss->cancel_fork(child, kargs->cset);
6548 }
6549
6550 cgroup_css_set_put_fork(kargs);
6551
6552 return ret;
6553}
6554
6555/**
6556 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6557 * @child: the child process
6558 * @kargs: the arguments passed to create the child process
6559 *
6560 * This calls the cancel_fork() callbacks if a fork failed *after*
6561 * cgroup_can_fork() succeeded and cleans up references we took to
6562 * prepare a new css_set for the child process in cgroup_can_fork().
6563 */
6564void cgroup_cancel_fork(struct task_struct *child,
6565 struct kernel_clone_args *kargs)
6566{
6567 struct cgroup_subsys *ss;
6568 int i;
6569
6570 for_each_subsys(ss, i)
6571 if (ss->cancel_fork)
6572 ss->cancel_fork(child, kargs->cset);
6573
6574 cgroup_css_set_put_fork(kargs);
6575}
6576
6577/**
6578 * cgroup_post_fork - finalize cgroup setup for the child process
6579 * @child: the child process
6580 * @kargs: the arguments passed to create the child process
6581 *
6582 * Attach the child process to its css_set calling the subsystem fork()
6583 * callbacks.
6584 */
6585void cgroup_post_fork(struct task_struct *child,
6586 struct kernel_clone_args *kargs)
6587 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6588{
6589 unsigned long cgrp_flags = 0;
6590 bool kill = false;
6591 struct cgroup_subsys *ss;
6592 struct css_set *cset;
6593 int i;
6594
6595 cset = kargs->cset;
6596 kargs->cset = NULL;
6597
6598 spin_lock_irq(&css_set_lock);
6599
6600 /* init tasks are special, only link regular threads */
6601 if (likely(child->pid)) {
6602 if (kargs->cgrp)
6603 cgrp_flags = kargs->cgrp->flags;
6604 else
6605 cgrp_flags = cset->dfl_cgrp->flags;
6606
6607 WARN_ON_ONCE(!list_empty(&child->cg_list));
6608 cset->nr_tasks++;
6609 css_set_move_task(child, NULL, cset, false);
6610 } else {
6611 put_css_set(cset);
6612 cset = NULL;
6613 }
6614
6615 if (!(child->flags & PF_KTHREAD)) {
6616 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6617 /*
6618 * If the cgroup has to be frozen, the new task has
6619 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6620 * get the task into the frozen state.
6621 */
6622 spin_lock(&child->sighand->siglock);
6623 WARN_ON_ONCE(child->frozen);
6624 child->jobctl |= JOBCTL_TRAP_FREEZE;
6625 spin_unlock(&child->sighand->siglock);
6626
6627 /*
6628 * Calling cgroup_update_frozen() isn't required here,
6629 * because it will be called anyway a bit later from
6630 * do_freezer_trap(). So we avoid cgroup's transient
6631 * switch from the frozen state and back.
6632 */
6633 }
6634
6635 /*
6636 * If the cgroup is to be killed notice it now and take the
6637 * child down right after we finished preparing it for
6638 * userspace.
6639 */
6640 kill = test_bit(CGRP_KILL, &cgrp_flags);
6641 }
6642
6643 spin_unlock_irq(&css_set_lock);
6644
6645 /*
6646 * Call ss->fork(). This must happen after @child is linked on
6647 * css_set; otherwise, @child might change state between ->fork()
6648 * and addition to css_set.
6649 */
6650 do_each_subsys_mask(ss, i, have_fork_callback) {
6651 ss->fork(child);
6652 } while_each_subsys_mask();
6653
6654 /* Make the new cset the root_cset of the new cgroup namespace. */
6655 if (kargs->flags & CLONE_NEWCGROUP) {
6656 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6657
6658 get_css_set(cset);
6659 child->nsproxy->cgroup_ns->root_cset = cset;
6660 put_css_set(rcset);
6661 }
6662
6663 /* Cgroup has to be killed so take down child immediately. */
6664 if (unlikely(kill))
6665 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6666
6667 cgroup_css_set_put_fork(kargs);
6668}
6669
6670/**
6671 * cgroup_exit - detach cgroup from exiting task
6672 * @tsk: pointer to task_struct of exiting process
6673 *
6674 * Description: Detach cgroup from @tsk.
6675 *
6676 */
6677void cgroup_exit(struct task_struct *tsk)
6678{
6679 struct cgroup_subsys *ss;
6680 struct css_set *cset;
6681 int i;
6682
6683 spin_lock_irq(&css_set_lock);
6684
6685 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6686 cset = task_css_set(tsk);
6687 css_set_move_task(tsk, cset, NULL, false);
6688 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6689 cset->nr_tasks--;
6690
6691 if (dl_task(tsk))
6692 dec_dl_tasks_cs(tsk);
6693
6694 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6695 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6696 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6697 cgroup_update_frozen(task_dfl_cgroup(tsk));
6698
6699 spin_unlock_irq(&css_set_lock);
6700
6701 /* see cgroup_post_fork() for details */
6702 do_each_subsys_mask(ss, i, have_exit_callback) {
6703 ss->exit(tsk);
6704 } while_each_subsys_mask();
6705}
6706
6707void cgroup_release(struct task_struct *task)
6708{
6709 struct cgroup_subsys *ss;
6710 int ssid;
6711
6712 do_each_subsys_mask(ss, ssid, have_release_callback) {
6713 ss->release(task);
6714 } while_each_subsys_mask();
6715
6716 spin_lock_irq(&css_set_lock);
6717 css_set_skip_task_iters(task_css_set(task), task);
6718 list_del_init(&task->cg_list);
6719 spin_unlock_irq(&css_set_lock);
6720}
6721
6722void cgroup_free(struct task_struct *task)
6723{
6724 struct css_set *cset = task_css_set(task);
6725 put_css_set(cset);
6726}
6727
6728static int __init cgroup_disable(char *str)
6729{
6730 struct cgroup_subsys *ss;
6731 char *token;
6732 int i;
6733
6734 while ((token = strsep(&str, ",")) != NULL) {
6735 if (!*token)
6736 continue;
6737
6738 for_each_subsys(ss, i) {
6739 if (strcmp(token, ss->name) &&
6740 strcmp(token, ss->legacy_name))
6741 continue;
6742
6743 static_branch_disable(cgroup_subsys_enabled_key[i]);
6744 pr_info("Disabling %s control group subsystem\n",
6745 ss->name);
6746 }
6747
6748 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6749 if (strcmp(token, cgroup_opt_feature_names[i]))
6750 continue;
6751 cgroup_feature_disable_mask |= 1 << i;
6752 pr_info("Disabling %s control group feature\n",
6753 cgroup_opt_feature_names[i]);
6754 break;
6755 }
6756 }
6757 return 1;
6758}
6759__setup("cgroup_disable=", cgroup_disable);
6760
6761void __init __weak enable_debug_cgroup(void) { }
6762
6763static int __init enable_cgroup_debug(char *str)
6764{
6765 cgroup_debug = true;
6766 enable_debug_cgroup();
6767 return 1;
6768}
6769__setup("cgroup_debug", enable_cgroup_debug);
6770
6771static int __init cgroup_favordynmods_setup(char *str)
6772{
6773 return (kstrtobool(str, &have_favordynmods) == 0);
6774}
6775__setup("cgroup_favordynmods=", cgroup_favordynmods_setup);
6776
6777/**
6778 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6779 * @dentry: directory dentry of interest
6780 * @ss: subsystem of interest
6781 *
6782 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6783 * to get the corresponding css and return it. If such css doesn't exist
6784 * or can't be pinned, an ERR_PTR value is returned.
6785 */
6786struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6787 struct cgroup_subsys *ss)
6788{
6789 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6790 struct file_system_type *s_type = dentry->d_sb->s_type;
6791 struct cgroup_subsys_state *css = NULL;
6792 struct cgroup *cgrp;
6793
6794 /* is @dentry a cgroup dir? */
6795 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6796 !kn || kernfs_type(kn) != KERNFS_DIR)
6797 return ERR_PTR(-EBADF);
6798
6799 rcu_read_lock();
6800
6801 /*
6802 * This path doesn't originate from kernfs and @kn could already
6803 * have been or be removed at any point. @kn->priv is RCU
6804 * protected for this access. See css_release_work_fn() for details.
6805 */
6806 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6807 if (cgrp)
6808 css = cgroup_css(cgrp, ss);
6809
6810 if (!css || !css_tryget_online(css))
6811 css = ERR_PTR(-ENOENT);
6812
6813 rcu_read_unlock();
6814 return css;
6815}
6816
6817/**
6818 * css_from_id - lookup css by id
6819 * @id: the cgroup id
6820 * @ss: cgroup subsys to be looked into
6821 *
6822 * Returns the css if there's valid one with @id, otherwise returns NULL.
6823 * Should be called under rcu_read_lock().
6824 */
6825struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6826{
6827 WARN_ON_ONCE(!rcu_read_lock_held());
6828 return idr_find(&ss->css_idr, id);
6829}
6830
6831/**
6832 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6833 * @path: path on the default hierarchy
6834 *
6835 * Find the cgroup at @path on the default hierarchy, increment its
6836 * reference count and return it. Returns pointer to the found cgroup on
6837 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6838 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6839 */
6840struct cgroup *cgroup_get_from_path(const char *path)
6841{
6842 struct kernfs_node *kn;
6843 struct cgroup *cgrp = ERR_PTR(-ENOENT);
6844 struct cgroup *root_cgrp;
6845
6846 root_cgrp = current_cgns_cgroup_dfl();
6847 kn = kernfs_walk_and_get(root_cgrp->kn, path);
6848 if (!kn)
6849 goto out;
6850
6851 if (kernfs_type(kn) != KERNFS_DIR) {
6852 cgrp = ERR_PTR(-ENOTDIR);
6853 goto out_kernfs;
6854 }
6855
6856 rcu_read_lock();
6857
6858 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6859 if (!cgrp || !cgroup_tryget(cgrp))
6860 cgrp = ERR_PTR(-ENOENT);
6861
6862 rcu_read_unlock();
6863
6864out_kernfs:
6865 kernfs_put(kn);
6866out:
6867 return cgrp;
6868}
6869EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6870
6871/**
6872 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
6873 * @fd: fd obtained by open(cgroup_dir)
6874 *
6875 * Find the cgroup from a fd which should be obtained
6876 * by opening a cgroup directory. Returns a pointer to the
6877 * cgroup on success. ERR_PTR is returned if the cgroup
6878 * cannot be found.
6879 */
6880struct cgroup *cgroup_v1v2_get_from_fd(int fd)
6881{
6882 struct cgroup *cgrp;
6883 struct fd f = fdget_raw(fd);
6884 if (!f.file)
6885 return ERR_PTR(-EBADF);
6886
6887 cgrp = cgroup_v1v2_get_from_file(f.file);
6888 fdput(f);
6889 return cgrp;
6890}
6891
6892/**
6893 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
6894 * cgroup2.
6895 * @fd: fd obtained by open(cgroup2_dir)
6896 */
6897struct cgroup *cgroup_get_from_fd(int fd)
6898{
6899 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
6900
6901 if (IS_ERR(cgrp))
6902 return ERR_CAST(cgrp);
6903
6904 if (!cgroup_on_dfl(cgrp)) {
6905 cgroup_put(cgrp);
6906 return ERR_PTR(-EBADF);
6907 }
6908 return cgrp;
6909}
6910EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6911
6912static u64 power_of_ten(int power)
6913{
6914 u64 v = 1;
6915 while (power--)
6916 v *= 10;
6917 return v;
6918}
6919
6920/**
6921 * cgroup_parse_float - parse a floating number
6922 * @input: input string
6923 * @dec_shift: number of decimal digits to shift
6924 * @v: output
6925 *
6926 * Parse a decimal floating point number in @input and store the result in
6927 * @v with decimal point right shifted @dec_shift times. For example, if
6928 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6929 * Returns 0 on success, -errno otherwise.
6930 *
6931 * There's nothing cgroup specific about this function except that it's
6932 * currently the only user.
6933 */
6934int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6935{
6936 s64 whole, frac = 0;
6937 int fstart = 0, fend = 0, flen;
6938
6939 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6940 return -EINVAL;
6941 if (frac < 0)
6942 return -EINVAL;
6943
6944 flen = fend > fstart ? fend - fstart : 0;
6945 if (flen < dec_shift)
6946 frac *= power_of_ten(dec_shift - flen);
6947 else
6948 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6949
6950 *v = whole * power_of_ten(dec_shift) + frac;
6951 return 0;
6952}
6953
6954/*
6955 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6956 * definition in cgroup-defs.h.
6957 */
6958#ifdef CONFIG_SOCK_CGROUP_DATA
6959
6960void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6961{
6962 struct cgroup *cgroup;
6963
6964 rcu_read_lock();
6965 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6966 if (in_interrupt()) {
6967 cgroup = &cgrp_dfl_root.cgrp;
6968 cgroup_get(cgroup);
6969 goto out;
6970 }
6971
6972 while (true) {
6973 struct css_set *cset;
6974
6975 cset = task_css_set(current);
6976 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6977 cgroup = cset->dfl_cgrp;
6978 break;
6979 }
6980 cpu_relax();
6981 }
6982out:
6983 skcd->cgroup = cgroup;
6984 cgroup_bpf_get(cgroup);
6985 rcu_read_unlock();
6986}
6987
6988void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6989{
6990 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6991
6992 /*
6993 * We might be cloning a socket which is left in an empty
6994 * cgroup and the cgroup might have already been rmdir'd.
6995 * Don't use cgroup_get_live().
6996 */
6997 cgroup_get(cgrp);
6998 cgroup_bpf_get(cgrp);
6999}
7000
7001void cgroup_sk_free(struct sock_cgroup_data *skcd)
7002{
7003 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7004
7005 cgroup_bpf_put(cgrp);
7006 cgroup_put(cgrp);
7007}
7008
7009#endif /* CONFIG_SOCK_CGROUP_DATA */
7010
7011#ifdef CONFIG_SYSFS
7012static ssize_t show_delegatable_files(struct cftype *files, char *buf,
7013 ssize_t size, const char *prefix)
7014{
7015 struct cftype *cft;
7016 ssize_t ret = 0;
7017
7018 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
7019 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
7020 continue;
7021
7022 if (prefix)
7023 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
7024
7025 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
7026
7027 if (WARN_ON(ret >= size))
7028 break;
7029 }
7030
7031 return ret;
7032}
7033
7034static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7035 char *buf)
7036{
7037 struct cgroup_subsys *ss;
7038 int ssid;
7039 ssize_t ret = 0;
7040
7041 ret = show_delegatable_files(cgroup_base_files, buf + ret,
7042 PAGE_SIZE - ret, NULL);
7043 if (cgroup_psi_enabled())
7044 ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7045 PAGE_SIZE - ret, NULL);
7046
7047 for_each_subsys(ss, ssid)
7048 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7049 PAGE_SIZE - ret,
7050 cgroup_subsys_name[ssid]);
7051
7052 return ret;
7053}
7054static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7055
7056static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7057 char *buf)
7058{
7059 return snprintf(buf, PAGE_SIZE,
7060 "nsdelegate\n"
7061 "favordynmods\n"
7062 "memory_localevents\n"
7063 "memory_recursiveprot\n"
7064 "memory_hugetlb_accounting\n");
7065}
7066static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7067
7068static struct attribute *cgroup_sysfs_attrs[] = {
7069 &cgroup_delegate_attr.attr,
7070 &cgroup_features_attr.attr,
7071 NULL,
7072};
7073
7074static const struct attribute_group cgroup_sysfs_attr_group = {
7075 .attrs = cgroup_sysfs_attrs,
7076 .name = "cgroup",
7077};
7078
7079static int __init cgroup_sysfs_init(void)
7080{
7081 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7082}
7083subsys_initcall(cgroup_sysfs_init);
7084
7085#endif /* CONFIG_SYSFS */