Loading...
Note: File does not exist in v3.5.6.
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 */