Loading...
Note: File does not exist in v3.15.
1/*
2 * Generic process-grouping system.
3 *
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
6 *
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
10 *
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
15 *
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
18 *
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
23 *
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
27 */
28
29#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
31#include "cgroup-internal.h"
32
33#include <linux/cred.h>
34#include <linux/errno.h>
35#include <linux/init_task.h>
36#include <linux/kernel.h>
37#include <linux/magic.h>
38#include <linux/mutex.h>
39#include <linux/mount.h>
40#include <linux/pagemap.h>
41#include <linux/proc_fs.h>
42#include <linux/rcupdate.h>
43#include <linux/sched.h>
44#include <linux/sched/task.h>
45#include <linux/slab.h>
46#include <linux/spinlock.h>
47#include <linux/percpu-rwsem.h>
48#include <linux/string.h>
49#include <linux/hashtable.h>
50#include <linux/idr.h>
51#include <linux/kthread.h>
52#include <linux/atomic.h>
53#include <linux/cpuset.h>
54#include <linux/proc_ns.h>
55#include <linux/nsproxy.h>
56#include <linux/file.h>
57#include <linux/fs_parser.h>
58#include <linux/sched/cputime.h>
59#include <linux/psi.h>
60#include <net/sock.h>
61
62#define CREATE_TRACE_POINTS
63#include <trace/events/cgroup.h>
64
65#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
66 MAX_CFTYPE_NAME + 2)
67/* let's not notify more than 100 times per second */
68#define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
69
70/*
71 * cgroup_mutex is the master lock. Any modification to cgroup or its
72 * hierarchy must be performed while holding it.
73 *
74 * css_set_lock protects task->cgroups pointer, the list of css_set
75 * objects, and the chain of tasks off each css_set.
76 *
77 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
78 * cgroup.h can use them for lockdep annotations.
79 */
80DEFINE_MUTEX(cgroup_mutex);
81DEFINE_SPINLOCK(css_set_lock);
82
83#ifdef CONFIG_PROVE_RCU
84EXPORT_SYMBOL_GPL(cgroup_mutex);
85EXPORT_SYMBOL_GPL(css_set_lock);
86#endif
87
88DEFINE_SPINLOCK(trace_cgroup_path_lock);
89char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
90bool cgroup_debug __read_mostly;
91
92/*
93 * Protects cgroup_idr and css_idr so that IDs can be released without
94 * grabbing cgroup_mutex.
95 */
96static DEFINE_SPINLOCK(cgroup_idr_lock);
97
98/*
99 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
100 * against file removal/re-creation across css hiding.
101 */
102static DEFINE_SPINLOCK(cgroup_file_kn_lock);
103
104DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
105
106#define cgroup_assert_mutex_or_rcu_locked() \
107 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
108 !lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
110
111/*
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
116 */
117static struct workqueue_struct *cgroup_destroy_wq;
118
119/* generate an array of cgroup subsystem pointers */
120#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
121struct cgroup_subsys *cgroup_subsys[] = {
122#include <linux/cgroup_subsys.h>
123};
124#undef SUBSYS
125
126/* array of cgroup subsystem names */
127#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
128static const char *cgroup_subsys_name[] = {
129#include <linux/cgroup_subsys.h>
130};
131#undef SUBSYS
132
133/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
134#define SUBSYS(_x) \
135 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
136 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
137 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
138 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
139#include <linux/cgroup_subsys.h>
140#undef SUBSYS
141
142#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
143static struct static_key_true *cgroup_subsys_enabled_key[] = {
144#include <linux/cgroup_subsys.h>
145};
146#undef SUBSYS
147
148#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
149static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
150#include <linux/cgroup_subsys.h>
151};
152#undef SUBSYS
153
154static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
155
156/* the default hierarchy */
157struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
158EXPORT_SYMBOL_GPL(cgrp_dfl_root);
159
160/*
161 * The default hierarchy always exists but is hidden until mounted for the
162 * first time. This is for backward compatibility.
163 */
164static bool cgrp_dfl_visible;
165
166/* some controllers are not supported in the default hierarchy */
167static u16 cgrp_dfl_inhibit_ss_mask;
168
169/* some controllers are implicitly enabled on the default hierarchy */
170static u16 cgrp_dfl_implicit_ss_mask;
171
172/* some controllers can be threaded on the default hierarchy */
173static u16 cgrp_dfl_threaded_ss_mask;
174
175/* The list of hierarchy roots */
176LIST_HEAD(cgroup_roots);
177static int cgroup_root_count;
178
179/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
180static DEFINE_IDR(cgroup_hierarchy_idr);
181
182/*
183 * Assign a monotonically increasing serial number to csses. It guarantees
184 * cgroups with bigger numbers are newer than those with smaller numbers.
185 * Also, as csses are always appended to the parent's ->children list, it
186 * guarantees that sibling csses are always sorted in the ascending serial
187 * number order on the list. Protected by cgroup_mutex.
188 */
189static u64 css_serial_nr_next = 1;
190
191/*
192 * These bitmasks identify subsystems with specific features to avoid
193 * having to do iterative checks repeatedly.
194 */
195static u16 have_fork_callback __read_mostly;
196static u16 have_exit_callback __read_mostly;
197static u16 have_release_callback __read_mostly;
198static u16 have_canfork_callback __read_mostly;
199
200/* cgroup namespace for init task */
201struct cgroup_namespace init_cgroup_ns = {
202 .count = REFCOUNT_INIT(2),
203 .user_ns = &init_user_ns,
204 .ns.ops = &cgroupns_operations,
205 .ns.inum = PROC_CGROUP_INIT_INO,
206 .root_cset = &init_css_set,
207};
208
209static struct file_system_type cgroup2_fs_type;
210static struct cftype cgroup_base_files[];
211
212static int cgroup_apply_control(struct cgroup *cgrp);
213static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
214static void css_task_iter_skip(struct css_task_iter *it,
215 struct task_struct *task);
216static int cgroup_destroy_locked(struct cgroup *cgrp);
217static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
218 struct cgroup_subsys *ss);
219static void css_release(struct percpu_ref *ref);
220static void kill_css(struct cgroup_subsys_state *css);
221static int cgroup_addrm_files(struct cgroup_subsys_state *css,
222 struct cgroup *cgrp, struct cftype cfts[],
223 bool is_add);
224
225/**
226 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
227 * @ssid: subsys ID of interest
228 *
229 * cgroup_subsys_enabled() can only be used with literal subsys names which
230 * is fine for individual subsystems but unsuitable for cgroup core. This
231 * is slower static_key_enabled() based test indexed by @ssid.
232 */
233bool cgroup_ssid_enabled(int ssid)
234{
235 if (CGROUP_SUBSYS_COUNT == 0)
236 return false;
237
238 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
239}
240
241/**
242 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
243 * @cgrp: the cgroup of interest
244 *
245 * The default hierarchy is the v2 interface of cgroup and this function
246 * can be used to test whether a cgroup is on the default hierarchy for
247 * cases where a subsystem should behave differnetly depending on the
248 * interface version.
249 *
250 * List of changed behaviors:
251 *
252 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
253 * and "name" are disallowed.
254 *
255 * - When mounting an existing superblock, mount options should match.
256 *
257 * - Remount is disallowed.
258 *
259 * - rename(2) is disallowed.
260 *
261 * - "tasks" is removed. Everything should be at process granularity. Use
262 * "cgroup.procs" instead.
263 *
264 * - "cgroup.procs" is not sorted. pids will be unique unless they got
265 * recycled inbetween reads.
266 *
267 * - "release_agent" and "notify_on_release" are removed. Replacement
268 * notification mechanism will be implemented.
269 *
270 * - "cgroup.clone_children" is removed.
271 *
272 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
273 * and its descendants contain no task; otherwise, 1. The file also
274 * generates kernfs notification which can be monitored through poll and
275 * [di]notify when the value of the file changes.
276 *
277 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
278 * take masks of ancestors with non-empty cpus/mems, instead of being
279 * moved to an ancestor.
280 *
281 * - cpuset: a task can be moved into an empty cpuset, and again it takes
282 * masks of ancestors.
283 *
284 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
285 * is not created.
286 *
287 * - blkcg: blk-throttle becomes properly hierarchical.
288 *
289 * - debug: disallowed on the default hierarchy.
290 */
291bool cgroup_on_dfl(const struct cgroup *cgrp)
292{
293 return cgrp->root == &cgrp_dfl_root;
294}
295
296/* IDR wrappers which synchronize using cgroup_idr_lock */
297static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
298 gfp_t gfp_mask)
299{
300 int ret;
301
302 idr_preload(gfp_mask);
303 spin_lock_bh(&cgroup_idr_lock);
304 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
305 spin_unlock_bh(&cgroup_idr_lock);
306 idr_preload_end();
307 return ret;
308}
309
310static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
311{
312 void *ret;
313
314 spin_lock_bh(&cgroup_idr_lock);
315 ret = idr_replace(idr, ptr, id);
316 spin_unlock_bh(&cgroup_idr_lock);
317 return ret;
318}
319
320static void cgroup_idr_remove(struct idr *idr, int id)
321{
322 spin_lock_bh(&cgroup_idr_lock);
323 idr_remove(idr, id);
324 spin_unlock_bh(&cgroup_idr_lock);
325}
326
327static bool cgroup_has_tasks(struct cgroup *cgrp)
328{
329 return cgrp->nr_populated_csets;
330}
331
332bool cgroup_is_threaded(struct cgroup *cgrp)
333{
334 return cgrp->dom_cgrp != cgrp;
335}
336
337/* can @cgrp host both domain and threaded children? */
338static bool cgroup_is_mixable(struct cgroup *cgrp)
339{
340 /*
341 * Root isn't under domain level resource control exempting it from
342 * the no-internal-process constraint, so it can serve as a thread
343 * root and a parent of resource domains at the same time.
344 */
345 return !cgroup_parent(cgrp);
346}
347
348/* can @cgrp become a thread root? should always be true for a thread root */
349static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
350{
351 /* mixables don't care */
352 if (cgroup_is_mixable(cgrp))
353 return true;
354
355 /* domain roots can't be nested under threaded */
356 if (cgroup_is_threaded(cgrp))
357 return false;
358
359 /* can only have either domain or threaded children */
360 if (cgrp->nr_populated_domain_children)
361 return false;
362
363 /* and no domain controllers can be enabled */
364 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
365 return false;
366
367 return true;
368}
369
370/* is @cgrp root of a threaded subtree? */
371bool cgroup_is_thread_root(struct cgroup *cgrp)
372{
373 /* thread root should be a domain */
374 if (cgroup_is_threaded(cgrp))
375 return false;
376
377 /* a domain w/ threaded children is a thread root */
378 if (cgrp->nr_threaded_children)
379 return true;
380
381 /*
382 * A domain which has tasks and explicit threaded controllers
383 * enabled is a thread root.
384 */
385 if (cgroup_has_tasks(cgrp) &&
386 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
387 return true;
388
389 return false;
390}
391
392/* a domain which isn't connected to the root w/o brekage can't be used */
393static bool cgroup_is_valid_domain(struct cgroup *cgrp)
394{
395 /* the cgroup itself can be a thread root */
396 if (cgroup_is_threaded(cgrp))
397 return false;
398
399 /* but the ancestors can't be unless mixable */
400 while ((cgrp = cgroup_parent(cgrp))) {
401 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
402 return false;
403 if (cgroup_is_threaded(cgrp))
404 return false;
405 }
406
407 return true;
408}
409
410/* subsystems visibly enabled on a cgroup */
411static u16 cgroup_control(struct cgroup *cgrp)
412{
413 struct cgroup *parent = cgroup_parent(cgrp);
414 u16 root_ss_mask = cgrp->root->subsys_mask;
415
416 if (parent) {
417 u16 ss_mask = parent->subtree_control;
418
419 /* threaded cgroups can only have threaded controllers */
420 if (cgroup_is_threaded(cgrp))
421 ss_mask &= cgrp_dfl_threaded_ss_mask;
422 return ss_mask;
423 }
424
425 if (cgroup_on_dfl(cgrp))
426 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
427 cgrp_dfl_implicit_ss_mask);
428 return root_ss_mask;
429}
430
431/* subsystems enabled on a cgroup */
432static u16 cgroup_ss_mask(struct cgroup *cgrp)
433{
434 struct cgroup *parent = cgroup_parent(cgrp);
435
436 if (parent) {
437 u16 ss_mask = parent->subtree_ss_mask;
438
439 /* threaded cgroups can only have threaded controllers */
440 if (cgroup_is_threaded(cgrp))
441 ss_mask &= cgrp_dfl_threaded_ss_mask;
442 return ss_mask;
443 }
444
445 return cgrp->root->subsys_mask;
446}
447
448/**
449 * cgroup_css - obtain a cgroup's css for the specified subsystem
450 * @cgrp: the cgroup of interest
451 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
452 *
453 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
454 * function must be called either under cgroup_mutex or rcu_read_lock() and
455 * the caller is responsible for pinning the returned css if it wants to
456 * keep accessing it outside the said locks. This function may return
457 * %NULL if @cgrp doesn't have @subsys_id enabled.
458 */
459static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
460 struct cgroup_subsys *ss)
461{
462 if (ss)
463 return rcu_dereference_check(cgrp->subsys[ss->id],
464 lockdep_is_held(&cgroup_mutex));
465 else
466 return &cgrp->self;
467}
468
469/**
470 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
471 * @cgrp: the cgroup of interest
472 * @ss: the subsystem of interest
473 *
474 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
475 * or is offline, %NULL is returned.
476 */
477static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
478 struct cgroup_subsys *ss)
479{
480 struct cgroup_subsys_state *css;
481
482 rcu_read_lock();
483 css = cgroup_css(cgrp, ss);
484 if (css && !css_tryget_online(css))
485 css = NULL;
486 rcu_read_unlock();
487
488 return css;
489}
490
491/**
492 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
493 * @cgrp: the cgroup of interest
494 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
495 *
496 * Similar to cgroup_css() but returns the effective css, which is defined
497 * as the matching css of the nearest ancestor including self which has @ss
498 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
499 * function is guaranteed to return non-NULL css.
500 */
501static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
502 struct cgroup_subsys *ss)
503{
504 lockdep_assert_held(&cgroup_mutex);
505
506 if (!ss)
507 return &cgrp->self;
508
509 /*
510 * This function is used while updating css associations and thus
511 * can't test the csses directly. Test ss_mask.
512 */
513 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
514 cgrp = cgroup_parent(cgrp);
515 if (!cgrp)
516 return NULL;
517 }
518
519 return cgroup_css(cgrp, ss);
520}
521
522/**
523 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
524 * @cgrp: the cgroup of interest
525 * @ss: the subsystem of interest
526 *
527 * Find and get the effective css of @cgrp for @ss. The effective css is
528 * defined as the matching css of the nearest ancestor including self which
529 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
530 * the root css is returned, so this function always returns a valid css.
531 *
532 * The returned css is not guaranteed to be online, and therefore it is the
533 * callers responsiblity to tryget a reference for it.
534 */
535struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
536 struct cgroup_subsys *ss)
537{
538 struct cgroup_subsys_state *css;
539
540 do {
541 css = cgroup_css(cgrp, ss);
542
543 if (css)
544 return css;
545 cgrp = cgroup_parent(cgrp);
546 } while (cgrp);
547
548 return init_css_set.subsys[ss->id];
549}
550
551/**
552 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
553 * @cgrp: the cgroup of interest
554 * @ss: the subsystem of interest
555 *
556 * Find and get the effective css of @cgrp for @ss. The effective css is
557 * defined as the matching css of the nearest ancestor including self which
558 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
559 * the root css is returned, so this function always returns a valid css.
560 * The returned css must be put using css_put().
561 */
562struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
563 struct cgroup_subsys *ss)
564{
565 struct cgroup_subsys_state *css;
566
567 rcu_read_lock();
568
569 do {
570 css = cgroup_css(cgrp, ss);
571
572 if (css && css_tryget_online(css))
573 goto out_unlock;
574 cgrp = cgroup_parent(cgrp);
575 } while (cgrp);
576
577 css = init_css_set.subsys[ss->id];
578 css_get(css);
579out_unlock:
580 rcu_read_unlock();
581 return css;
582}
583
584static void cgroup_get_live(struct cgroup *cgrp)
585{
586 WARN_ON_ONCE(cgroup_is_dead(cgrp));
587 css_get(&cgrp->self);
588}
589
590/**
591 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
592 * is responsible for taking the css_set_lock.
593 * @cgrp: the cgroup in question
594 */
595int __cgroup_task_count(const struct cgroup *cgrp)
596{
597 int count = 0;
598 struct cgrp_cset_link *link;
599
600 lockdep_assert_held(&css_set_lock);
601
602 list_for_each_entry(link, &cgrp->cset_links, cset_link)
603 count += link->cset->nr_tasks;
604
605 return count;
606}
607
608/**
609 * cgroup_task_count - count the number of tasks in a cgroup.
610 * @cgrp: the cgroup in question
611 */
612int cgroup_task_count(const struct cgroup *cgrp)
613{
614 int count;
615
616 spin_lock_irq(&css_set_lock);
617 count = __cgroup_task_count(cgrp);
618 spin_unlock_irq(&css_set_lock);
619
620 return count;
621}
622
623struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
624{
625 struct cgroup *cgrp = of->kn->parent->priv;
626 struct cftype *cft = of_cft(of);
627
628 /*
629 * This is open and unprotected implementation of cgroup_css().
630 * seq_css() is only called from a kernfs file operation which has
631 * an active reference on the file. Because all the subsystem
632 * files are drained before a css is disassociated with a cgroup,
633 * the matching css from the cgroup's subsys table is guaranteed to
634 * be and stay valid until the enclosing operation is complete.
635 */
636 if (cft->ss)
637 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
638 else
639 return &cgrp->self;
640}
641EXPORT_SYMBOL_GPL(of_css);
642
643/**
644 * for_each_css - iterate all css's of a cgroup
645 * @css: the iteration cursor
646 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
647 * @cgrp: the target cgroup to iterate css's of
648 *
649 * Should be called under cgroup_[tree_]mutex.
650 */
651#define for_each_css(css, ssid, cgrp) \
652 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
653 if (!((css) = rcu_dereference_check( \
654 (cgrp)->subsys[(ssid)], \
655 lockdep_is_held(&cgroup_mutex)))) { } \
656 else
657
658/**
659 * for_each_e_css - iterate all effective css's of a cgroup
660 * @css: the iteration cursor
661 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
662 * @cgrp: the target cgroup to iterate css's of
663 *
664 * Should be called under cgroup_[tree_]mutex.
665 */
666#define for_each_e_css(css, ssid, cgrp) \
667 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
668 if (!((css) = cgroup_e_css_by_mask(cgrp, \
669 cgroup_subsys[(ssid)]))) \
670 ; \
671 else
672
673/**
674 * do_each_subsys_mask - filter for_each_subsys with a bitmask
675 * @ss: the iteration cursor
676 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
677 * @ss_mask: the bitmask
678 *
679 * The block will only run for cases where the ssid-th bit (1 << ssid) of
680 * @ss_mask is set.
681 */
682#define do_each_subsys_mask(ss, ssid, ss_mask) do { \
683 unsigned long __ss_mask = (ss_mask); \
684 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
685 (ssid) = 0; \
686 break; \
687 } \
688 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
689 (ss) = cgroup_subsys[ssid]; \
690 {
691
692#define while_each_subsys_mask() \
693 } \
694 } \
695} while (false)
696
697/* iterate over child cgrps, lock should be held throughout iteration */
698#define cgroup_for_each_live_child(child, cgrp) \
699 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
700 if (({ lockdep_assert_held(&cgroup_mutex); \
701 cgroup_is_dead(child); })) \
702 ; \
703 else
704
705/* walk live descendants in preorder */
706#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
707 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
708 if (({ lockdep_assert_held(&cgroup_mutex); \
709 (dsct) = (d_css)->cgroup; \
710 cgroup_is_dead(dsct); })) \
711 ; \
712 else
713
714/* walk live descendants in postorder */
715#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
716 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
717 if (({ lockdep_assert_held(&cgroup_mutex); \
718 (dsct) = (d_css)->cgroup; \
719 cgroup_is_dead(dsct); })) \
720 ; \
721 else
722
723/*
724 * The default css_set - used by init and its children prior to any
725 * hierarchies being mounted. It contains a pointer to the root state
726 * for each subsystem. Also used to anchor the list of css_sets. Not
727 * reference-counted, to improve performance when child cgroups
728 * haven't been created.
729 */
730struct css_set init_css_set = {
731 .refcount = REFCOUNT_INIT(1),
732 .dom_cset = &init_css_set,
733 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
734 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
735 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
736 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
737 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
738 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
739 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
740 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
741
742 /*
743 * The following field is re-initialized when this cset gets linked
744 * in cgroup_init(). However, let's initialize the field
745 * statically too so that the default cgroup can be accessed safely
746 * early during boot.
747 */
748 .dfl_cgrp = &cgrp_dfl_root.cgrp,
749};
750
751static int css_set_count = 1; /* 1 for init_css_set */
752
753static bool css_set_threaded(struct css_set *cset)
754{
755 return cset->dom_cset != cset;
756}
757
758/**
759 * css_set_populated - does a css_set contain any tasks?
760 * @cset: target css_set
761 *
762 * css_set_populated() should be the same as !!cset->nr_tasks at steady
763 * state. However, css_set_populated() can be called while a task is being
764 * added to or removed from the linked list before the nr_tasks is
765 * properly updated. Hence, we can't just look at ->nr_tasks here.
766 */
767static bool css_set_populated(struct css_set *cset)
768{
769 lockdep_assert_held(&css_set_lock);
770
771 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
772}
773
774/**
775 * cgroup_update_populated - update the populated count of a cgroup
776 * @cgrp: the target cgroup
777 * @populated: inc or dec populated count
778 *
779 * One of the css_sets associated with @cgrp is either getting its first
780 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
781 * count is propagated towards root so that a given cgroup's
782 * nr_populated_children is zero iff none of its descendants contain any
783 * tasks.
784 *
785 * @cgrp's interface file "cgroup.populated" is zero if both
786 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
787 * 1 otherwise. When the sum changes from or to zero, userland is notified
788 * that the content of the interface file has changed. This can be used to
789 * detect when @cgrp and its descendants become populated or empty.
790 */
791static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
792{
793 struct cgroup *child = NULL;
794 int adj = populated ? 1 : -1;
795
796 lockdep_assert_held(&css_set_lock);
797
798 do {
799 bool was_populated = cgroup_is_populated(cgrp);
800
801 if (!child) {
802 cgrp->nr_populated_csets += adj;
803 } else {
804 if (cgroup_is_threaded(child))
805 cgrp->nr_populated_threaded_children += adj;
806 else
807 cgrp->nr_populated_domain_children += adj;
808 }
809
810 if (was_populated == cgroup_is_populated(cgrp))
811 break;
812
813 cgroup1_check_for_release(cgrp);
814 TRACE_CGROUP_PATH(notify_populated, cgrp,
815 cgroup_is_populated(cgrp));
816 cgroup_file_notify(&cgrp->events_file);
817
818 child = cgrp;
819 cgrp = cgroup_parent(cgrp);
820 } while (cgrp);
821}
822
823/**
824 * css_set_update_populated - update populated state of a css_set
825 * @cset: target css_set
826 * @populated: whether @cset is populated or depopulated
827 *
828 * @cset is either getting the first task or losing the last. Update the
829 * populated counters of all associated cgroups accordingly.
830 */
831static void css_set_update_populated(struct css_set *cset, bool populated)
832{
833 struct cgrp_cset_link *link;
834
835 lockdep_assert_held(&css_set_lock);
836
837 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
838 cgroup_update_populated(link->cgrp, populated);
839}
840
841/*
842 * @task is leaving, advance task iterators which are pointing to it so
843 * that they can resume at the next position. Advancing an iterator might
844 * remove it from the list, use safe walk. See css_task_iter_skip() for
845 * details.
846 */
847static void css_set_skip_task_iters(struct css_set *cset,
848 struct task_struct *task)
849{
850 struct css_task_iter *it, *pos;
851
852 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
853 css_task_iter_skip(it, task);
854}
855
856/**
857 * css_set_move_task - move a task from one css_set to another
858 * @task: task being moved
859 * @from_cset: css_set @task currently belongs to (may be NULL)
860 * @to_cset: new css_set @task is being moved to (may be NULL)
861 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
862 *
863 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
864 * css_set, @from_cset can be NULL. If @task is being disassociated
865 * instead of moved, @to_cset can be NULL.
866 *
867 * This function automatically handles populated counter updates and
868 * css_task_iter adjustments but the caller is responsible for managing
869 * @from_cset and @to_cset's reference counts.
870 */
871static void css_set_move_task(struct task_struct *task,
872 struct css_set *from_cset, struct css_set *to_cset,
873 bool use_mg_tasks)
874{
875 lockdep_assert_held(&css_set_lock);
876
877 if (to_cset && !css_set_populated(to_cset))
878 css_set_update_populated(to_cset, true);
879
880 if (from_cset) {
881 WARN_ON_ONCE(list_empty(&task->cg_list));
882
883 css_set_skip_task_iters(from_cset, task);
884 list_del_init(&task->cg_list);
885 if (!css_set_populated(from_cset))
886 css_set_update_populated(from_cset, false);
887 } else {
888 WARN_ON_ONCE(!list_empty(&task->cg_list));
889 }
890
891 if (to_cset) {
892 /*
893 * We are synchronized through cgroup_threadgroup_rwsem
894 * against PF_EXITING setting such that we can't race
895 * against cgroup_exit()/cgroup_free() dropping the css_set.
896 */
897 WARN_ON_ONCE(task->flags & PF_EXITING);
898
899 cgroup_move_task(task, to_cset);
900 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
901 &to_cset->tasks);
902 }
903}
904
905/*
906 * hash table for cgroup groups. This improves the performance to find
907 * an existing css_set. This hash doesn't (currently) take into
908 * account cgroups in empty hierarchies.
909 */
910#define CSS_SET_HASH_BITS 7
911static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
912
913static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
914{
915 unsigned long key = 0UL;
916 struct cgroup_subsys *ss;
917 int i;
918
919 for_each_subsys(ss, i)
920 key += (unsigned long)css[i];
921 key = (key >> 16) ^ key;
922
923 return key;
924}
925
926void put_css_set_locked(struct css_set *cset)
927{
928 struct cgrp_cset_link *link, *tmp_link;
929 struct cgroup_subsys *ss;
930 int ssid;
931
932 lockdep_assert_held(&css_set_lock);
933
934 if (!refcount_dec_and_test(&cset->refcount))
935 return;
936
937 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
938
939 /* This css_set is dead. unlink it and release cgroup and css refs */
940 for_each_subsys(ss, ssid) {
941 list_del(&cset->e_cset_node[ssid]);
942 css_put(cset->subsys[ssid]);
943 }
944 hash_del(&cset->hlist);
945 css_set_count--;
946
947 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
948 list_del(&link->cset_link);
949 list_del(&link->cgrp_link);
950 if (cgroup_parent(link->cgrp))
951 cgroup_put(link->cgrp);
952 kfree(link);
953 }
954
955 if (css_set_threaded(cset)) {
956 list_del(&cset->threaded_csets_node);
957 put_css_set_locked(cset->dom_cset);
958 }
959
960 kfree_rcu(cset, rcu_head);
961}
962
963/**
964 * compare_css_sets - helper function for find_existing_css_set().
965 * @cset: candidate css_set being tested
966 * @old_cset: existing css_set for a task
967 * @new_cgrp: cgroup that's being entered by the task
968 * @template: desired set of css pointers in css_set (pre-calculated)
969 *
970 * Returns true if "cset" matches "old_cset" except for the hierarchy
971 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
972 */
973static bool compare_css_sets(struct css_set *cset,
974 struct css_set *old_cset,
975 struct cgroup *new_cgrp,
976 struct cgroup_subsys_state *template[])
977{
978 struct cgroup *new_dfl_cgrp;
979 struct list_head *l1, *l2;
980
981 /*
982 * On the default hierarchy, there can be csets which are
983 * associated with the same set of cgroups but different csses.
984 * Let's first ensure that csses match.
985 */
986 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
987 return false;
988
989
990 /* @cset's domain should match the default cgroup's */
991 if (cgroup_on_dfl(new_cgrp))
992 new_dfl_cgrp = new_cgrp;
993 else
994 new_dfl_cgrp = old_cset->dfl_cgrp;
995
996 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
997 return false;
998
999 /*
1000 * Compare cgroup pointers in order to distinguish between
1001 * different cgroups in hierarchies. As different cgroups may
1002 * share the same effective css, this comparison is always
1003 * necessary.
1004 */
1005 l1 = &cset->cgrp_links;
1006 l2 = &old_cset->cgrp_links;
1007 while (1) {
1008 struct cgrp_cset_link *link1, *link2;
1009 struct cgroup *cgrp1, *cgrp2;
1010
1011 l1 = l1->next;
1012 l2 = l2->next;
1013 /* See if we reached the end - both lists are equal length. */
1014 if (l1 == &cset->cgrp_links) {
1015 BUG_ON(l2 != &old_cset->cgrp_links);
1016 break;
1017 } else {
1018 BUG_ON(l2 == &old_cset->cgrp_links);
1019 }
1020 /* Locate the cgroups associated with these links. */
1021 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1022 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1023 cgrp1 = link1->cgrp;
1024 cgrp2 = link2->cgrp;
1025 /* Hierarchies should be linked in the same order. */
1026 BUG_ON(cgrp1->root != cgrp2->root);
1027
1028 /*
1029 * If this hierarchy is the hierarchy of the cgroup
1030 * that's changing, then we need to check that this
1031 * css_set points to the new cgroup; if it's any other
1032 * hierarchy, then this css_set should point to the
1033 * same cgroup as the old css_set.
1034 */
1035 if (cgrp1->root == new_cgrp->root) {
1036 if (cgrp1 != new_cgrp)
1037 return false;
1038 } else {
1039 if (cgrp1 != cgrp2)
1040 return false;
1041 }
1042 }
1043 return true;
1044}
1045
1046/**
1047 * find_existing_css_set - init css array and find the matching css_set
1048 * @old_cset: the css_set that we're using before the cgroup transition
1049 * @cgrp: the cgroup that we're moving into
1050 * @template: out param for the new set of csses, should be clear on entry
1051 */
1052static struct css_set *find_existing_css_set(struct css_set *old_cset,
1053 struct cgroup *cgrp,
1054 struct cgroup_subsys_state *template[])
1055{
1056 struct cgroup_root *root = cgrp->root;
1057 struct cgroup_subsys *ss;
1058 struct css_set *cset;
1059 unsigned long key;
1060 int i;
1061
1062 /*
1063 * Build the set of subsystem state objects that we want to see in the
1064 * new css_set. while subsystems can change globally, the entries here
1065 * won't change, so no need for locking.
1066 */
1067 for_each_subsys(ss, i) {
1068 if (root->subsys_mask & (1UL << i)) {
1069 /*
1070 * @ss is in this hierarchy, so we want the
1071 * effective css from @cgrp.
1072 */
1073 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1074 } else {
1075 /*
1076 * @ss is not in this hierarchy, so we don't want
1077 * to change the css.
1078 */
1079 template[i] = old_cset->subsys[i];
1080 }
1081 }
1082
1083 key = css_set_hash(template);
1084 hash_for_each_possible(css_set_table, cset, hlist, key) {
1085 if (!compare_css_sets(cset, old_cset, cgrp, template))
1086 continue;
1087
1088 /* This css_set matches what we need */
1089 return cset;
1090 }
1091
1092 /* No existing cgroup group matched */
1093 return NULL;
1094}
1095
1096static void free_cgrp_cset_links(struct list_head *links_to_free)
1097{
1098 struct cgrp_cset_link *link, *tmp_link;
1099
1100 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1101 list_del(&link->cset_link);
1102 kfree(link);
1103 }
1104}
1105
1106/**
1107 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1108 * @count: the number of links to allocate
1109 * @tmp_links: list_head the allocated links are put on
1110 *
1111 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1112 * through ->cset_link. Returns 0 on success or -errno.
1113 */
1114static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1115{
1116 struct cgrp_cset_link *link;
1117 int i;
1118
1119 INIT_LIST_HEAD(tmp_links);
1120
1121 for (i = 0; i < count; i++) {
1122 link = kzalloc(sizeof(*link), GFP_KERNEL);
1123 if (!link) {
1124 free_cgrp_cset_links(tmp_links);
1125 return -ENOMEM;
1126 }
1127 list_add(&link->cset_link, tmp_links);
1128 }
1129 return 0;
1130}
1131
1132/**
1133 * link_css_set - a helper function to link a css_set to a cgroup
1134 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1135 * @cset: the css_set to be linked
1136 * @cgrp: the destination cgroup
1137 */
1138static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1139 struct cgroup *cgrp)
1140{
1141 struct cgrp_cset_link *link;
1142
1143 BUG_ON(list_empty(tmp_links));
1144
1145 if (cgroup_on_dfl(cgrp))
1146 cset->dfl_cgrp = cgrp;
1147
1148 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1149 link->cset = cset;
1150 link->cgrp = cgrp;
1151
1152 /*
1153 * Always add links to the tail of the lists so that the lists are
1154 * in choronological order.
1155 */
1156 list_move_tail(&link->cset_link, &cgrp->cset_links);
1157 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1158
1159 if (cgroup_parent(cgrp))
1160 cgroup_get_live(cgrp);
1161}
1162
1163/**
1164 * find_css_set - return a new css_set with one cgroup updated
1165 * @old_cset: the baseline css_set
1166 * @cgrp: the cgroup to be updated
1167 *
1168 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1169 * substituted into the appropriate hierarchy.
1170 */
1171static struct css_set *find_css_set(struct css_set *old_cset,
1172 struct cgroup *cgrp)
1173{
1174 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1175 struct css_set *cset;
1176 struct list_head tmp_links;
1177 struct cgrp_cset_link *link;
1178 struct cgroup_subsys *ss;
1179 unsigned long key;
1180 int ssid;
1181
1182 lockdep_assert_held(&cgroup_mutex);
1183
1184 /* First see if we already have a cgroup group that matches
1185 * the desired set */
1186 spin_lock_irq(&css_set_lock);
1187 cset = find_existing_css_set(old_cset, cgrp, template);
1188 if (cset)
1189 get_css_set(cset);
1190 spin_unlock_irq(&css_set_lock);
1191
1192 if (cset)
1193 return cset;
1194
1195 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1196 if (!cset)
1197 return NULL;
1198
1199 /* Allocate all the cgrp_cset_link objects that we'll need */
1200 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1201 kfree(cset);
1202 return NULL;
1203 }
1204
1205 refcount_set(&cset->refcount, 1);
1206 cset->dom_cset = cset;
1207 INIT_LIST_HEAD(&cset->tasks);
1208 INIT_LIST_HEAD(&cset->mg_tasks);
1209 INIT_LIST_HEAD(&cset->dying_tasks);
1210 INIT_LIST_HEAD(&cset->task_iters);
1211 INIT_LIST_HEAD(&cset->threaded_csets);
1212 INIT_HLIST_NODE(&cset->hlist);
1213 INIT_LIST_HEAD(&cset->cgrp_links);
1214 INIT_LIST_HEAD(&cset->mg_preload_node);
1215 INIT_LIST_HEAD(&cset->mg_node);
1216
1217 /* Copy the set of subsystem state objects generated in
1218 * find_existing_css_set() */
1219 memcpy(cset->subsys, template, sizeof(cset->subsys));
1220
1221 spin_lock_irq(&css_set_lock);
1222 /* Add reference counts and links from the new css_set. */
1223 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1224 struct cgroup *c = link->cgrp;
1225
1226 if (c->root == cgrp->root)
1227 c = cgrp;
1228 link_css_set(&tmp_links, cset, c);
1229 }
1230
1231 BUG_ON(!list_empty(&tmp_links));
1232
1233 css_set_count++;
1234
1235 /* Add @cset to the hash table */
1236 key = css_set_hash(cset->subsys);
1237 hash_add(css_set_table, &cset->hlist, key);
1238
1239 for_each_subsys(ss, ssid) {
1240 struct cgroup_subsys_state *css = cset->subsys[ssid];
1241
1242 list_add_tail(&cset->e_cset_node[ssid],
1243 &css->cgroup->e_csets[ssid]);
1244 css_get(css);
1245 }
1246
1247 spin_unlock_irq(&css_set_lock);
1248
1249 /*
1250 * If @cset should be threaded, look up the matching dom_cset and
1251 * link them up. We first fully initialize @cset then look for the
1252 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1253 * to stay empty until we return.
1254 */
1255 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1256 struct css_set *dcset;
1257
1258 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1259 if (!dcset) {
1260 put_css_set(cset);
1261 return NULL;
1262 }
1263
1264 spin_lock_irq(&css_set_lock);
1265 cset->dom_cset = dcset;
1266 list_add_tail(&cset->threaded_csets_node,
1267 &dcset->threaded_csets);
1268 spin_unlock_irq(&css_set_lock);
1269 }
1270
1271 return cset;
1272}
1273
1274struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1275{
1276 struct cgroup *root_cgrp = kf_root->kn->priv;
1277
1278 return root_cgrp->root;
1279}
1280
1281static int cgroup_init_root_id(struct cgroup_root *root)
1282{
1283 int id;
1284
1285 lockdep_assert_held(&cgroup_mutex);
1286
1287 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1288 if (id < 0)
1289 return id;
1290
1291 root->hierarchy_id = id;
1292 return 0;
1293}
1294
1295static void cgroup_exit_root_id(struct cgroup_root *root)
1296{
1297 lockdep_assert_held(&cgroup_mutex);
1298
1299 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1300}
1301
1302void cgroup_free_root(struct cgroup_root *root)
1303{
1304 kfree(root);
1305}
1306
1307static void cgroup_destroy_root(struct cgroup_root *root)
1308{
1309 struct cgroup *cgrp = &root->cgrp;
1310 struct cgrp_cset_link *link, *tmp_link;
1311
1312 trace_cgroup_destroy_root(root);
1313
1314 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1315
1316 BUG_ON(atomic_read(&root->nr_cgrps));
1317 BUG_ON(!list_empty(&cgrp->self.children));
1318
1319 /* Rebind all subsystems back to the default hierarchy */
1320 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1321
1322 /*
1323 * Release all the links from cset_links to this hierarchy's
1324 * root cgroup
1325 */
1326 spin_lock_irq(&css_set_lock);
1327
1328 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1329 list_del(&link->cset_link);
1330 list_del(&link->cgrp_link);
1331 kfree(link);
1332 }
1333
1334 spin_unlock_irq(&css_set_lock);
1335
1336 if (!list_empty(&root->root_list)) {
1337 list_del(&root->root_list);
1338 cgroup_root_count--;
1339 }
1340
1341 cgroup_exit_root_id(root);
1342
1343 mutex_unlock(&cgroup_mutex);
1344
1345 kernfs_destroy_root(root->kf_root);
1346 cgroup_free_root(root);
1347}
1348
1349/*
1350 * look up cgroup associated with current task's cgroup namespace on the
1351 * specified hierarchy
1352 */
1353static struct cgroup *
1354current_cgns_cgroup_from_root(struct cgroup_root *root)
1355{
1356 struct cgroup *res = NULL;
1357 struct css_set *cset;
1358
1359 lockdep_assert_held(&css_set_lock);
1360
1361 rcu_read_lock();
1362
1363 cset = current->nsproxy->cgroup_ns->root_cset;
1364 if (cset == &init_css_set) {
1365 res = &root->cgrp;
1366 } else if (root == &cgrp_dfl_root) {
1367 res = cset->dfl_cgrp;
1368 } else {
1369 struct cgrp_cset_link *link;
1370
1371 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1372 struct cgroup *c = link->cgrp;
1373
1374 if (c->root == root) {
1375 res = c;
1376 break;
1377 }
1378 }
1379 }
1380 rcu_read_unlock();
1381
1382 BUG_ON(!res);
1383 return res;
1384}
1385
1386/* look up cgroup associated with given css_set on the specified hierarchy */
1387static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1388 struct cgroup_root *root)
1389{
1390 struct cgroup *res = NULL;
1391
1392 lockdep_assert_held(&cgroup_mutex);
1393 lockdep_assert_held(&css_set_lock);
1394
1395 if (cset == &init_css_set) {
1396 res = &root->cgrp;
1397 } else if (root == &cgrp_dfl_root) {
1398 res = cset->dfl_cgrp;
1399 } else {
1400 struct cgrp_cset_link *link;
1401
1402 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1403 struct cgroup *c = link->cgrp;
1404
1405 if (c->root == root) {
1406 res = c;
1407 break;
1408 }
1409 }
1410 }
1411
1412 BUG_ON(!res);
1413 return res;
1414}
1415
1416/*
1417 * Return the cgroup for "task" from the given hierarchy. Must be
1418 * called with cgroup_mutex and css_set_lock held.
1419 */
1420struct cgroup *task_cgroup_from_root(struct task_struct *task,
1421 struct cgroup_root *root)
1422{
1423 /*
1424 * No need to lock the task - since we hold css_set_lock the
1425 * task can't change groups.
1426 */
1427 return cset_cgroup_from_root(task_css_set(task), root);
1428}
1429
1430/*
1431 * A task must hold cgroup_mutex to modify cgroups.
1432 *
1433 * Any task can increment and decrement the count field without lock.
1434 * So in general, code holding cgroup_mutex can't rely on the count
1435 * field not changing. However, if the count goes to zero, then only
1436 * cgroup_attach_task() can increment it again. Because a count of zero
1437 * means that no tasks are currently attached, therefore there is no
1438 * way a task attached to that cgroup can fork (the other way to
1439 * increment the count). So code holding cgroup_mutex can safely
1440 * assume that if the count is zero, it will stay zero. Similarly, if
1441 * a task holds cgroup_mutex on a cgroup with zero count, it
1442 * knows that the cgroup won't be removed, as cgroup_rmdir()
1443 * needs that mutex.
1444 *
1445 * A cgroup can only be deleted if both its 'count' of using tasks
1446 * is zero, and its list of 'children' cgroups is empty. Since all
1447 * tasks in the system use _some_ cgroup, and since there is always at
1448 * least one task in the system (init, pid == 1), therefore, root cgroup
1449 * always has either children cgroups and/or using tasks. So we don't
1450 * need a special hack to ensure that root cgroup cannot be deleted.
1451 *
1452 * P.S. One more locking exception. RCU is used to guard the
1453 * update of a tasks cgroup pointer by cgroup_attach_task()
1454 */
1455
1456static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1457
1458static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1459 char *buf)
1460{
1461 struct cgroup_subsys *ss = cft->ss;
1462
1463 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1464 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1465 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1466
1467 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1468 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1469 cft->name);
1470 } else {
1471 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1472 }
1473 return buf;
1474}
1475
1476/**
1477 * cgroup_file_mode - deduce file mode of a control file
1478 * @cft: the control file in question
1479 *
1480 * S_IRUGO for read, S_IWUSR for write.
1481 */
1482static umode_t cgroup_file_mode(const struct cftype *cft)
1483{
1484 umode_t mode = 0;
1485
1486 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1487 mode |= S_IRUGO;
1488
1489 if (cft->write_u64 || cft->write_s64 || cft->write) {
1490 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1491 mode |= S_IWUGO;
1492 else
1493 mode |= S_IWUSR;
1494 }
1495
1496 return mode;
1497}
1498
1499/**
1500 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1501 * @subtree_control: the new subtree_control mask to consider
1502 * @this_ss_mask: available subsystems
1503 *
1504 * On the default hierarchy, a subsystem may request other subsystems to be
1505 * enabled together through its ->depends_on mask. In such cases, more
1506 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1507 *
1508 * This function calculates which subsystems need to be enabled if
1509 * @subtree_control is to be applied while restricted to @this_ss_mask.
1510 */
1511static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1512{
1513 u16 cur_ss_mask = subtree_control;
1514 struct cgroup_subsys *ss;
1515 int ssid;
1516
1517 lockdep_assert_held(&cgroup_mutex);
1518
1519 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1520
1521 while (true) {
1522 u16 new_ss_mask = cur_ss_mask;
1523
1524 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1525 new_ss_mask |= ss->depends_on;
1526 } while_each_subsys_mask();
1527
1528 /*
1529 * Mask out subsystems which aren't available. This can
1530 * happen only if some depended-upon subsystems were bound
1531 * to non-default hierarchies.
1532 */
1533 new_ss_mask &= this_ss_mask;
1534
1535 if (new_ss_mask == cur_ss_mask)
1536 break;
1537 cur_ss_mask = new_ss_mask;
1538 }
1539
1540 return cur_ss_mask;
1541}
1542
1543/**
1544 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1545 * @kn: the kernfs_node being serviced
1546 *
1547 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1548 * the method finishes if locking succeeded. Note that once this function
1549 * returns the cgroup returned by cgroup_kn_lock_live() may become
1550 * inaccessible any time. If the caller intends to continue to access the
1551 * cgroup, it should pin it before invoking this function.
1552 */
1553void cgroup_kn_unlock(struct kernfs_node *kn)
1554{
1555 struct cgroup *cgrp;
1556
1557 if (kernfs_type(kn) == KERNFS_DIR)
1558 cgrp = kn->priv;
1559 else
1560 cgrp = kn->parent->priv;
1561
1562 mutex_unlock(&cgroup_mutex);
1563
1564 kernfs_unbreak_active_protection(kn);
1565 cgroup_put(cgrp);
1566}
1567
1568/**
1569 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1570 * @kn: the kernfs_node being serviced
1571 * @drain_offline: perform offline draining on the cgroup
1572 *
1573 * This helper is to be used by a cgroup kernfs method currently servicing
1574 * @kn. It breaks the active protection, performs cgroup locking and
1575 * verifies that the associated cgroup is alive. Returns the cgroup if
1576 * alive; otherwise, %NULL. A successful return should be undone by a
1577 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1578 * cgroup is drained of offlining csses before return.
1579 *
1580 * Any cgroup kernfs method implementation which requires locking the
1581 * associated cgroup should use this helper. It avoids nesting cgroup
1582 * locking under kernfs active protection and allows all kernfs operations
1583 * including self-removal.
1584 */
1585struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1586{
1587 struct cgroup *cgrp;
1588
1589 if (kernfs_type(kn) == KERNFS_DIR)
1590 cgrp = kn->priv;
1591 else
1592 cgrp = kn->parent->priv;
1593
1594 /*
1595 * We're gonna grab cgroup_mutex which nests outside kernfs
1596 * active_ref. cgroup liveliness check alone provides enough
1597 * protection against removal. Ensure @cgrp stays accessible and
1598 * break the active_ref protection.
1599 */
1600 if (!cgroup_tryget(cgrp))
1601 return NULL;
1602 kernfs_break_active_protection(kn);
1603
1604 if (drain_offline)
1605 cgroup_lock_and_drain_offline(cgrp);
1606 else
1607 mutex_lock(&cgroup_mutex);
1608
1609 if (!cgroup_is_dead(cgrp))
1610 return cgrp;
1611
1612 cgroup_kn_unlock(kn);
1613 return NULL;
1614}
1615
1616static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1617{
1618 char name[CGROUP_FILE_NAME_MAX];
1619
1620 lockdep_assert_held(&cgroup_mutex);
1621
1622 if (cft->file_offset) {
1623 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1624 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1625
1626 spin_lock_irq(&cgroup_file_kn_lock);
1627 cfile->kn = NULL;
1628 spin_unlock_irq(&cgroup_file_kn_lock);
1629
1630 del_timer_sync(&cfile->notify_timer);
1631 }
1632
1633 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1634}
1635
1636/**
1637 * css_clear_dir - remove subsys files in a cgroup directory
1638 * @css: taget css
1639 */
1640static void css_clear_dir(struct cgroup_subsys_state *css)
1641{
1642 struct cgroup *cgrp = css->cgroup;
1643 struct cftype *cfts;
1644
1645 if (!(css->flags & CSS_VISIBLE))
1646 return;
1647
1648 css->flags &= ~CSS_VISIBLE;
1649
1650 if (!css->ss) {
1651 if (cgroup_on_dfl(cgrp))
1652 cfts = cgroup_base_files;
1653 else
1654 cfts = cgroup1_base_files;
1655
1656 cgroup_addrm_files(css, cgrp, cfts, false);
1657 } else {
1658 list_for_each_entry(cfts, &css->ss->cfts, node)
1659 cgroup_addrm_files(css, cgrp, cfts, false);
1660 }
1661}
1662
1663/**
1664 * css_populate_dir - create subsys files in a cgroup directory
1665 * @css: target css
1666 *
1667 * On failure, no file is added.
1668 */
1669static int css_populate_dir(struct cgroup_subsys_state *css)
1670{
1671 struct cgroup *cgrp = css->cgroup;
1672 struct cftype *cfts, *failed_cfts;
1673 int ret;
1674
1675 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1676 return 0;
1677
1678 if (!css->ss) {
1679 if (cgroup_on_dfl(cgrp))
1680 cfts = cgroup_base_files;
1681 else
1682 cfts = cgroup1_base_files;
1683
1684 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1685 if (ret < 0)
1686 return ret;
1687 } else {
1688 list_for_each_entry(cfts, &css->ss->cfts, node) {
1689 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1690 if (ret < 0) {
1691 failed_cfts = cfts;
1692 goto err;
1693 }
1694 }
1695 }
1696
1697 css->flags |= CSS_VISIBLE;
1698
1699 return 0;
1700err:
1701 list_for_each_entry(cfts, &css->ss->cfts, node) {
1702 if (cfts == failed_cfts)
1703 break;
1704 cgroup_addrm_files(css, cgrp, cfts, false);
1705 }
1706 return ret;
1707}
1708
1709int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1710{
1711 struct cgroup *dcgrp = &dst_root->cgrp;
1712 struct cgroup_subsys *ss;
1713 int ssid, i, ret;
1714
1715 lockdep_assert_held(&cgroup_mutex);
1716
1717 do_each_subsys_mask(ss, ssid, ss_mask) {
1718 /*
1719 * If @ss has non-root csses attached to it, can't move.
1720 * If @ss is an implicit controller, it is exempt from this
1721 * rule and can be stolen.
1722 */
1723 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1724 !ss->implicit_on_dfl)
1725 return -EBUSY;
1726
1727 /* can't move between two non-dummy roots either */
1728 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1729 return -EBUSY;
1730 } while_each_subsys_mask();
1731
1732 do_each_subsys_mask(ss, ssid, ss_mask) {
1733 struct cgroup_root *src_root = ss->root;
1734 struct cgroup *scgrp = &src_root->cgrp;
1735 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1736 struct css_set *cset;
1737
1738 WARN_ON(!css || cgroup_css(dcgrp, ss));
1739
1740 /* disable from the source */
1741 src_root->subsys_mask &= ~(1 << ssid);
1742 WARN_ON(cgroup_apply_control(scgrp));
1743 cgroup_finalize_control(scgrp, 0);
1744
1745 /* rebind */
1746 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1747 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1748 ss->root = dst_root;
1749 css->cgroup = dcgrp;
1750
1751 spin_lock_irq(&css_set_lock);
1752 hash_for_each(css_set_table, i, cset, hlist)
1753 list_move_tail(&cset->e_cset_node[ss->id],
1754 &dcgrp->e_csets[ss->id]);
1755 spin_unlock_irq(&css_set_lock);
1756
1757 /* default hierarchy doesn't enable controllers by default */
1758 dst_root->subsys_mask |= 1 << ssid;
1759 if (dst_root == &cgrp_dfl_root) {
1760 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1761 } else {
1762 dcgrp->subtree_control |= 1 << ssid;
1763 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1764 }
1765
1766 ret = cgroup_apply_control(dcgrp);
1767 if (ret)
1768 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1769 ss->name, ret);
1770
1771 if (ss->bind)
1772 ss->bind(css);
1773 } while_each_subsys_mask();
1774
1775 kernfs_activate(dcgrp->kn);
1776 return 0;
1777}
1778
1779int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1780 struct kernfs_root *kf_root)
1781{
1782 int len = 0;
1783 char *buf = NULL;
1784 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1785 struct cgroup *ns_cgroup;
1786
1787 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1788 if (!buf)
1789 return -ENOMEM;
1790
1791 spin_lock_irq(&css_set_lock);
1792 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1793 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1794 spin_unlock_irq(&css_set_lock);
1795
1796 if (len >= PATH_MAX)
1797 len = -ERANGE;
1798 else if (len > 0) {
1799 seq_escape(sf, buf, " \t\n\\");
1800 len = 0;
1801 }
1802 kfree(buf);
1803 return len;
1804}
1805
1806enum cgroup2_param {
1807 Opt_nsdelegate,
1808 Opt_memory_localevents,
1809 Opt_memory_recursiveprot,
1810 nr__cgroup2_params
1811};
1812
1813static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1814 fsparam_flag("nsdelegate", Opt_nsdelegate),
1815 fsparam_flag("memory_localevents", Opt_memory_localevents),
1816 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1817 {}
1818};
1819
1820static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1821{
1822 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1823 struct fs_parse_result result;
1824 int opt;
1825
1826 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1827 if (opt < 0)
1828 return opt;
1829
1830 switch (opt) {
1831 case Opt_nsdelegate:
1832 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1833 return 0;
1834 case Opt_memory_localevents:
1835 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1836 return 0;
1837 case Opt_memory_recursiveprot:
1838 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1839 return 0;
1840 }
1841 return -EINVAL;
1842}
1843
1844static void apply_cgroup_root_flags(unsigned int root_flags)
1845{
1846 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1847 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1848 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1849 else
1850 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1851
1852 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1853 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1854 else
1855 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1856
1857 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1858 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1859 else
1860 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1861 }
1862}
1863
1864static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1865{
1866 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1867 seq_puts(seq, ",nsdelegate");
1868 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1869 seq_puts(seq, ",memory_localevents");
1870 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1871 seq_puts(seq, ",memory_recursiveprot");
1872 return 0;
1873}
1874
1875static int cgroup_reconfigure(struct fs_context *fc)
1876{
1877 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1878
1879 apply_cgroup_root_flags(ctx->flags);
1880 return 0;
1881}
1882
1883static void init_cgroup_housekeeping(struct cgroup *cgrp)
1884{
1885 struct cgroup_subsys *ss;
1886 int ssid;
1887
1888 INIT_LIST_HEAD(&cgrp->self.sibling);
1889 INIT_LIST_HEAD(&cgrp->self.children);
1890 INIT_LIST_HEAD(&cgrp->cset_links);
1891 INIT_LIST_HEAD(&cgrp->pidlists);
1892 mutex_init(&cgrp->pidlist_mutex);
1893 cgrp->self.cgroup = cgrp;
1894 cgrp->self.flags |= CSS_ONLINE;
1895 cgrp->dom_cgrp = cgrp;
1896 cgrp->max_descendants = INT_MAX;
1897 cgrp->max_depth = INT_MAX;
1898 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1899 prev_cputime_init(&cgrp->prev_cputime);
1900
1901 for_each_subsys(ss, ssid)
1902 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1903
1904 init_waitqueue_head(&cgrp->offline_waitq);
1905 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1906}
1907
1908void init_cgroup_root(struct cgroup_fs_context *ctx)
1909{
1910 struct cgroup_root *root = ctx->root;
1911 struct cgroup *cgrp = &root->cgrp;
1912
1913 INIT_LIST_HEAD(&root->root_list);
1914 atomic_set(&root->nr_cgrps, 1);
1915 cgrp->root = root;
1916 init_cgroup_housekeeping(cgrp);
1917
1918 root->flags = ctx->flags;
1919 if (ctx->release_agent)
1920 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1921 if (ctx->name)
1922 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1923 if (ctx->cpuset_clone_children)
1924 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1925}
1926
1927int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1928{
1929 LIST_HEAD(tmp_links);
1930 struct cgroup *root_cgrp = &root->cgrp;
1931 struct kernfs_syscall_ops *kf_sops;
1932 struct css_set *cset;
1933 int i, ret;
1934
1935 lockdep_assert_held(&cgroup_mutex);
1936
1937 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1938 0, GFP_KERNEL);
1939 if (ret)
1940 goto out;
1941
1942 /*
1943 * We're accessing css_set_count without locking css_set_lock here,
1944 * but that's OK - it can only be increased by someone holding
1945 * cgroup_lock, and that's us. Later rebinding may disable
1946 * controllers on the default hierarchy and thus create new csets,
1947 * which can't be more than the existing ones. Allocate 2x.
1948 */
1949 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1950 if (ret)
1951 goto cancel_ref;
1952
1953 ret = cgroup_init_root_id(root);
1954 if (ret)
1955 goto cancel_ref;
1956
1957 kf_sops = root == &cgrp_dfl_root ?
1958 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1959
1960 root->kf_root = kernfs_create_root(kf_sops,
1961 KERNFS_ROOT_CREATE_DEACTIVATED |
1962 KERNFS_ROOT_SUPPORT_EXPORTOP |
1963 KERNFS_ROOT_SUPPORT_USER_XATTR,
1964 root_cgrp);
1965 if (IS_ERR(root->kf_root)) {
1966 ret = PTR_ERR(root->kf_root);
1967 goto exit_root_id;
1968 }
1969 root_cgrp->kn = root->kf_root->kn;
1970 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
1971 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
1972
1973 ret = css_populate_dir(&root_cgrp->self);
1974 if (ret)
1975 goto destroy_root;
1976
1977 ret = rebind_subsystems(root, ss_mask);
1978 if (ret)
1979 goto destroy_root;
1980
1981 ret = cgroup_bpf_inherit(root_cgrp);
1982 WARN_ON_ONCE(ret);
1983
1984 trace_cgroup_setup_root(root);
1985
1986 /*
1987 * There must be no failure case after here, since rebinding takes
1988 * care of subsystems' refcounts, which are explicitly dropped in
1989 * the failure exit path.
1990 */
1991 list_add(&root->root_list, &cgroup_roots);
1992 cgroup_root_count++;
1993
1994 /*
1995 * Link the root cgroup in this hierarchy into all the css_set
1996 * objects.
1997 */
1998 spin_lock_irq(&css_set_lock);
1999 hash_for_each(css_set_table, i, cset, hlist) {
2000 link_css_set(&tmp_links, cset, root_cgrp);
2001 if (css_set_populated(cset))
2002 cgroup_update_populated(root_cgrp, true);
2003 }
2004 spin_unlock_irq(&css_set_lock);
2005
2006 BUG_ON(!list_empty(&root_cgrp->self.children));
2007 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2008
2009 kernfs_activate(root_cgrp->kn);
2010 ret = 0;
2011 goto out;
2012
2013destroy_root:
2014 kernfs_destroy_root(root->kf_root);
2015 root->kf_root = NULL;
2016exit_root_id:
2017 cgroup_exit_root_id(root);
2018cancel_ref:
2019 percpu_ref_exit(&root_cgrp->self.refcnt);
2020out:
2021 free_cgrp_cset_links(&tmp_links);
2022 return ret;
2023}
2024
2025int cgroup_do_get_tree(struct fs_context *fc)
2026{
2027 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2028 int ret;
2029
2030 ctx->kfc.root = ctx->root->kf_root;
2031 if (fc->fs_type == &cgroup2_fs_type)
2032 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2033 else
2034 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2035 ret = kernfs_get_tree(fc);
2036
2037 /*
2038 * In non-init cgroup namespace, instead of root cgroup's dentry,
2039 * we return the dentry corresponding to the cgroupns->root_cgrp.
2040 */
2041 if (!ret && ctx->ns != &init_cgroup_ns) {
2042 struct dentry *nsdentry;
2043 struct super_block *sb = fc->root->d_sb;
2044 struct cgroup *cgrp;
2045
2046 mutex_lock(&cgroup_mutex);
2047 spin_lock_irq(&css_set_lock);
2048
2049 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2050
2051 spin_unlock_irq(&css_set_lock);
2052 mutex_unlock(&cgroup_mutex);
2053
2054 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2055 dput(fc->root);
2056 if (IS_ERR(nsdentry)) {
2057 deactivate_locked_super(sb);
2058 ret = PTR_ERR(nsdentry);
2059 nsdentry = NULL;
2060 }
2061 fc->root = nsdentry;
2062 }
2063
2064 if (!ctx->kfc.new_sb_created)
2065 cgroup_put(&ctx->root->cgrp);
2066
2067 return ret;
2068}
2069
2070/*
2071 * Destroy a cgroup filesystem context.
2072 */
2073static void cgroup_fs_context_free(struct fs_context *fc)
2074{
2075 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2076
2077 kfree(ctx->name);
2078 kfree(ctx->release_agent);
2079 put_cgroup_ns(ctx->ns);
2080 kernfs_free_fs_context(fc);
2081 kfree(ctx);
2082}
2083
2084static int cgroup_get_tree(struct fs_context *fc)
2085{
2086 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2087 int ret;
2088
2089 cgrp_dfl_visible = true;
2090 cgroup_get_live(&cgrp_dfl_root.cgrp);
2091 ctx->root = &cgrp_dfl_root;
2092
2093 ret = cgroup_do_get_tree(fc);
2094 if (!ret)
2095 apply_cgroup_root_flags(ctx->flags);
2096 return ret;
2097}
2098
2099static const struct fs_context_operations cgroup_fs_context_ops = {
2100 .free = cgroup_fs_context_free,
2101 .parse_param = cgroup2_parse_param,
2102 .get_tree = cgroup_get_tree,
2103 .reconfigure = cgroup_reconfigure,
2104};
2105
2106static const struct fs_context_operations cgroup1_fs_context_ops = {
2107 .free = cgroup_fs_context_free,
2108 .parse_param = cgroup1_parse_param,
2109 .get_tree = cgroup1_get_tree,
2110 .reconfigure = cgroup1_reconfigure,
2111};
2112
2113/*
2114 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2115 * we select the namespace we're going to use.
2116 */
2117static int cgroup_init_fs_context(struct fs_context *fc)
2118{
2119 struct cgroup_fs_context *ctx;
2120
2121 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2122 if (!ctx)
2123 return -ENOMEM;
2124
2125 ctx->ns = current->nsproxy->cgroup_ns;
2126 get_cgroup_ns(ctx->ns);
2127 fc->fs_private = &ctx->kfc;
2128 if (fc->fs_type == &cgroup2_fs_type)
2129 fc->ops = &cgroup_fs_context_ops;
2130 else
2131 fc->ops = &cgroup1_fs_context_ops;
2132 put_user_ns(fc->user_ns);
2133 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2134 fc->global = true;
2135 return 0;
2136}
2137
2138static void cgroup_kill_sb(struct super_block *sb)
2139{
2140 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2141 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2142
2143 /*
2144 * If @root doesn't have any children, start killing it.
2145 * This prevents new mounts by disabling percpu_ref_tryget_live().
2146 * cgroup_mount() may wait for @root's release.
2147 *
2148 * And don't kill the default root.
2149 */
2150 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2151 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2152 percpu_ref_kill(&root->cgrp.self.refcnt);
2153 cgroup_put(&root->cgrp);
2154 kernfs_kill_sb(sb);
2155}
2156
2157struct file_system_type cgroup_fs_type = {
2158 .name = "cgroup",
2159 .init_fs_context = cgroup_init_fs_context,
2160 .parameters = cgroup1_fs_parameters,
2161 .kill_sb = cgroup_kill_sb,
2162 .fs_flags = FS_USERNS_MOUNT,
2163};
2164
2165static struct file_system_type cgroup2_fs_type = {
2166 .name = "cgroup2",
2167 .init_fs_context = cgroup_init_fs_context,
2168 .parameters = cgroup2_fs_parameters,
2169 .kill_sb = cgroup_kill_sb,
2170 .fs_flags = FS_USERNS_MOUNT,
2171};
2172
2173#ifdef CONFIG_CPUSETS
2174static const struct fs_context_operations cpuset_fs_context_ops = {
2175 .get_tree = cgroup1_get_tree,
2176 .free = cgroup_fs_context_free,
2177};
2178
2179/*
2180 * This is ugly, but preserves the userspace API for existing cpuset
2181 * users. If someone tries to mount the "cpuset" filesystem, we
2182 * silently switch it to mount "cgroup" instead
2183 */
2184static int cpuset_init_fs_context(struct fs_context *fc)
2185{
2186 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2187 struct cgroup_fs_context *ctx;
2188 int err;
2189
2190 err = cgroup_init_fs_context(fc);
2191 if (err) {
2192 kfree(agent);
2193 return err;
2194 }
2195
2196 fc->ops = &cpuset_fs_context_ops;
2197
2198 ctx = cgroup_fc2context(fc);
2199 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2200 ctx->flags |= CGRP_ROOT_NOPREFIX;
2201 ctx->release_agent = agent;
2202
2203 get_filesystem(&cgroup_fs_type);
2204 put_filesystem(fc->fs_type);
2205 fc->fs_type = &cgroup_fs_type;
2206
2207 return 0;
2208}
2209
2210static struct file_system_type cpuset_fs_type = {
2211 .name = "cpuset",
2212 .init_fs_context = cpuset_init_fs_context,
2213 .fs_flags = FS_USERNS_MOUNT,
2214};
2215#endif
2216
2217int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2218 struct cgroup_namespace *ns)
2219{
2220 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2221
2222 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2223}
2224
2225int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2226 struct cgroup_namespace *ns)
2227{
2228 int ret;
2229
2230 mutex_lock(&cgroup_mutex);
2231 spin_lock_irq(&css_set_lock);
2232
2233 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2234
2235 spin_unlock_irq(&css_set_lock);
2236 mutex_unlock(&cgroup_mutex);
2237
2238 return ret;
2239}
2240EXPORT_SYMBOL_GPL(cgroup_path_ns);
2241
2242/**
2243 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2244 * @task: target task
2245 * @buf: the buffer to write the path into
2246 * @buflen: the length of the buffer
2247 *
2248 * Determine @task's cgroup on the first (the one with the lowest non-zero
2249 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2250 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2251 * cgroup controller callbacks.
2252 *
2253 * Return value is the same as kernfs_path().
2254 */
2255int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2256{
2257 struct cgroup_root *root;
2258 struct cgroup *cgrp;
2259 int hierarchy_id = 1;
2260 int ret;
2261
2262 mutex_lock(&cgroup_mutex);
2263 spin_lock_irq(&css_set_lock);
2264
2265 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2266
2267 if (root) {
2268 cgrp = task_cgroup_from_root(task, root);
2269 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2270 } else {
2271 /* if no hierarchy exists, everyone is in "/" */
2272 ret = strlcpy(buf, "/", buflen);
2273 }
2274
2275 spin_unlock_irq(&css_set_lock);
2276 mutex_unlock(&cgroup_mutex);
2277 return ret;
2278}
2279EXPORT_SYMBOL_GPL(task_cgroup_path);
2280
2281/**
2282 * cgroup_migrate_add_task - add a migration target task to a migration context
2283 * @task: target task
2284 * @mgctx: target migration context
2285 *
2286 * Add @task, which is a migration target, to @mgctx->tset. This function
2287 * becomes noop if @task doesn't need to be migrated. @task's css_set
2288 * should have been added as a migration source and @task->cg_list will be
2289 * moved from the css_set's tasks list to mg_tasks one.
2290 */
2291static void cgroup_migrate_add_task(struct task_struct *task,
2292 struct cgroup_mgctx *mgctx)
2293{
2294 struct css_set *cset;
2295
2296 lockdep_assert_held(&css_set_lock);
2297
2298 /* @task either already exited or can't exit until the end */
2299 if (task->flags & PF_EXITING)
2300 return;
2301
2302 /* cgroup_threadgroup_rwsem protects racing against forks */
2303 WARN_ON_ONCE(list_empty(&task->cg_list));
2304
2305 cset = task_css_set(task);
2306 if (!cset->mg_src_cgrp)
2307 return;
2308
2309 mgctx->tset.nr_tasks++;
2310
2311 list_move_tail(&task->cg_list, &cset->mg_tasks);
2312 if (list_empty(&cset->mg_node))
2313 list_add_tail(&cset->mg_node,
2314 &mgctx->tset.src_csets);
2315 if (list_empty(&cset->mg_dst_cset->mg_node))
2316 list_add_tail(&cset->mg_dst_cset->mg_node,
2317 &mgctx->tset.dst_csets);
2318}
2319
2320/**
2321 * cgroup_taskset_first - reset taskset and return the first task
2322 * @tset: taskset of interest
2323 * @dst_cssp: output variable for the destination css
2324 *
2325 * @tset iteration is initialized and the first task is returned.
2326 */
2327struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2328 struct cgroup_subsys_state **dst_cssp)
2329{
2330 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2331 tset->cur_task = NULL;
2332
2333 return cgroup_taskset_next(tset, dst_cssp);
2334}
2335
2336/**
2337 * cgroup_taskset_next - iterate to the next task in taskset
2338 * @tset: taskset of interest
2339 * @dst_cssp: output variable for the destination css
2340 *
2341 * Return the next task in @tset. Iteration must have been initialized
2342 * with cgroup_taskset_first().
2343 */
2344struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2345 struct cgroup_subsys_state **dst_cssp)
2346{
2347 struct css_set *cset = tset->cur_cset;
2348 struct task_struct *task = tset->cur_task;
2349
2350 while (&cset->mg_node != tset->csets) {
2351 if (!task)
2352 task = list_first_entry(&cset->mg_tasks,
2353 struct task_struct, cg_list);
2354 else
2355 task = list_next_entry(task, cg_list);
2356
2357 if (&task->cg_list != &cset->mg_tasks) {
2358 tset->cur_cset = cset;
2359 tset->cur_task = task;
2360
2361 /*
2362 * This function may be called both before and
2363 * after cgroup_taskset_migrate(). The two cases
2364 * can be distinguished by looking at whether @cset
2365 * has its ->mg_dst_cset set.
2366 */
2367 if (cset->mg_dst_cset)
2368 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2369 else
2370 *dst_cssp = cset->subsys[tset->ssid];
2371
2372 return task;
2373 }
2374
2375 cset = list_next_entry(cset, mg_node);
2376 task = NULL;
2377 }
2378
2379 return NULL;
2380}
2381
2382/**
2383 * cgroup_taskset_migrate - migrate a taskset
2384 * @mgctx: migration context
2385 *
2386 * Migrate tasks in @mgctx as setup by migration preparation functions.
2387 * This function fails iff one of the ->can_attach callbacks fails and
2388 * guarantees that either all or none of the tasks in @mgctx are migrated.
2389 * @mgctx is consumed regardless of success.
2390 */
2391static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2392{
2393 struct cgroup_taskset *tset = &mgctx->tset;
2394 struct cgroup_subsys *ss;
2395 struct task_struct *task, *tmp_task;
2396 struct css_set *cset, *tmp_cset;
2397 int ssid, failed_ssid, ret;
2398
2399 /* check that we can legitimately attach to the cgroup */
2400 if (tset->nr_tasks) {
2401 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2402 if (ss->can_attach) {
2403 tset->ssid = ssid;
2404 ret = ss->can_attach(tset);
2405 if (ret) {
2406 failed_ssid = ssid;
2407 goto out_cancel_attach;
2408 }
2409 }
2410 } while_each_subsys_mask();
2411 }
2412
2413 /*
2414 * Now that we're guaranteed success, proceed to move all tasks to
2415 * the new cgroup. There are no failure cases after here, so this
2416 * is the commit point.
2417 */
2418 spin_lock_irq(&css_set_lock);
2419 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2420 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2421 struct css_set *from_cset = task_css_set(task);
2422 struct css_set *to_cset = cset->mg_dst_cset;
2423
2424 get_css_set(to_cset);
2425 to_cset->nr_tasks++;
2426 css_set_move_task(task, from_cset, to_cset, true);
2427 from_cset->nr_tasks--;
2428 /*
2429 * If the source or destination cgroup is frozen,
2430 * the task might require to change its state.
2431 */
2432 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2433 to_cset->dfl_cgrp);
2434 put_css_set_locked(from_cset);
2435
2436 }
2437 }
2438 spin_unlock_irq(&css_set_lock);
2439
2440 /*
2441 * Migration is committed, all target tasks are now on dst_csets.
2442 * Nothing is sensitive to fork() after this point. Notify
2443 * controllers that migration is complete.
2444 */
2445 tset->csets = &tset->dst_csets;
2446
2447 if (tset->nr_tasks) {
2448 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2449 if (ss->attach) {
2450 tset->ssid = ssid;
2451 ss->attach(tset);
2452 }
2453 } while_each_subsys_mask();
2454 }
2455
2456 ret = 0;
2457 goto out_release_tset;
2458
2459out_cancel_attach:
2460 if (tset->nr_tasks) {
2461 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2462 if (ssid == failed_ssid)
2463 break;
2464 if (ss->cancel_attach) {
2465 tset->ssid = ssid;
2466 ss->cancel_attach(tset);
2467 }
2468 } while_each_subsys_mask();
2469 }
2470out_release_tset:
2471 spin_lock_irq(&css_set_lock);
2472 list_splice_init(&tset->dst_csets, &tset->src_csets);
2473 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2474 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2475 list_del_init(&cset->mg_node);
2476 }
2477 spin_unlock_irq(&css_set_lock);
2478
2479 /*
2480 * Re-initialize the cgroup_taskset structure in case it is reused
2481 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2482 * iteration.
2483 */
2484 tset->nr_tasks = 0;
2485 tset->csets = &tset->src_csets;
2486 return ret;
2487}
2488
2489/**
2490 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2491 * @dst_cgrp: destination cgroup to test
2492 *
2493 * On the default hierarchy, except for the mixable, (possible) thread root
2494 * and threaded cgroups, subtree_control must be zero for migration
2495 * destination cgroups with tasks so that child cgroups don't compete
2496 * against tasks.
2497 */
2498int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2499{
2500 /* v1 doesn't have any restriction */
2501 if (!cgroup_on_dfl(dst_cgrp))
2502 return 0;
2503
2504 /* verify @dst_cgrp can host resources */
2505 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2506 return -EOPNOTSUPP;
2507
2508 /* mixables don't care */
2509 if (cgroup_is_mixable(dst_cgrp))
2510 return 0;
2511
2512 /*
2513 * If @dst_cgrp is already or can become a thread root or is
2514 * threaded, it doesn't matter.
2515 */
2516 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2517 return 0;
2518
2519 /* apply no-internal-process constraint */
2520 if (dst_cgrp->subtree_control)
2521 return -EBUSY;
2522
2523 return 0;
2524}
2525
2526/**
2527 * cgroup_migrate_finish - cleanup after attach
2528 * @mgctx: migration context
2529 *
2530 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2531 * those functions for details.
2532 */
2533void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2534{
2535 LIST_HEAD(preloaded);
2536 struct css_set *cset, *tmp_cset;
2537
2538 lockdep_assert_held(&cgroup_mutex);
2539
2540 spin_lock_irq(&css_set_lock);
2541
2542 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2543 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2544
2545 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2546 cset->mg_src_cgrp = NULL;
2547 cset->mg_dst_cgrp = NULL;
2548 cset->mg_dst_cset = NULL;
2549 list_del_init(&cset->mg_preload_node);
2550 put_css_set_locked(cset);
2551 }
2552
2553 spin_unlock_irq(&css_set_lock);
2554}
2555
2556/**
2557 * cgroup_migrate_add_src - add a migration source css_set
2558 * @src_cset: the source css_set to add
2559 * @dst_cgrp: the destination cgroup
2560 * @mgctx: migration context
2561 *
2562 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2563 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2564 * up by cgroup_migrate_finish().
2565 *
2566 * This function may be called without holding cgroup_threadgroup_rwsem
2567 * even if the target is a process. Threads may be created and destroyed
2568 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2569 * into play and the preloaded css_sets are guaranteed to cover all
2570 * migrations.
2571 */
2572void cgroup_migrate_add_src(struct css_set *src_cset,
2573 struct cgroup *dst_cgrp,
2574 struct cgroup_mgctx *mgctx)
2575{
2576 struct cgroup *src_cgrp;
2577
2578 lockdep_assert_held(&cgroup_mutex);
2579 lockdep_assert_held(&css_set_lock);
2580
2581 /*
2582 * If ->dead, @src_set is associated with one or more dead cgroups
2583 * and doesn't contain any migratable tasks. Ignore it early so
2584 * that the rest of migration path doesn't get confused by it.
2585 */
2586 if (src_cset->dead)
2587 return;
2588
2589 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2590
2591 if (!list_empty(&src_cset->mg_preload_node))
2592 return;
2593
2594 WARN_ON(src_cset->mg_src_cgrp);
2595 WARN_ON(src_cset->mg_dst_cgrp);
2596 WARN_ON(!list_empty(&src_cset->mg_tasks));
2597 WARN_ON(!list_empty(&src_cset->mg_node));
2598
2599 src_cset->mg_src_cgrp = src_cgrp;
2600 src_cset->mg_dst_cgrp = dst_cgrp;
2601 get_css_set(src_cset);
2602 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2603}
2604
2605/**
2606 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2607 * @mgctx: migration context
2608 *
2609 * Tasks are about to be moved and all the source css_sets have been
2610 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2611 * pins all destination css_sets, links each to its source, and append them
2612 * to @mgctx->preloaded_dst_csets.
2613 *
2614 * This function must be called after cgroup_migrate_add_src() has been
2615 * called on each migration source css_set. After migration is performed
2616 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2617 * @mgctx.
2618 */
2619int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2620{
2621 struct css_set *src_cset, *tmp_cset;
2622
2623 lockdep_assert_held(&cgroup_mutex);
2624
2625 /* look up the dst cset for each src cset and link it to src */
2626 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2627 mg_preload_node) {
2628 struct css_set *dst_cset;
2629 struct cgroup_subsys *ss;
2630 int ssid;
2631
2632 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2633 if (!dst_cset)
2634 return -ENOMEM;
2635
2636 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2637
2638 /*
2639 * If src cset equals dst, it's noop. Drop the src.
2640 * cgroup_migrate() will skip the cset too. Note that we
2641 * can't handle src == dst as some nodes are used by both.
2642 */
2643 if (src_cset == dst_cset) {
2644 src_cset->mg_src_cgrp = NULL;
2645 src_cset->mg_dst_cgrp = NULL;
2646 list_del_init(&src_cset->mg_preload_node);
2647 put_css_set(src_cset);
2648 put_css_set(dst_cset);
2649 continue;
2650 }
2651
2652 src_cset->mg_dst_cset = dst_cset;
2653
2654 if (list_empty(&dst_cset->mg_preload_node))
2655 list_add_tail(&dst_cset->mg_preload_node,
2656 &mgctx->preloaded_dst_csets);
2657 else
2658 put_css_set(dst_cset);
2659
2660 for_each_subsys(ss, ssid)
2661 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2662 mgctx->ss_mask |= 1 << ssid;
2663 }
2664
2665 return 0;
2666}
2667
2668/**
2669 * cgroup_migrate - migrate a process or task to a cgroup
2670 * @leader: the leader of the process or the task to migrate
2671 * @threadgroup: whether @leader points to the whole process or a single task
2672 * @mgctx: migration context
2673 *
2674 * Migrate a process or task denoted by @leader. If migrating a process,
2675 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2676 * responsible for invoking cgroup_migrate_add_src() and
2677 * cgroup_migrate_prepare_dst() on the targets before invoking this
2678 * function and following up with cgroup_migrate_finish().
2679 *
2680 * As long as a controller's ->can_attach() doesn't fail, this function is
2681 * guaranteed to succeed. This means that, excluding ->can_attach()
2682 * failure, when migrating multiple targets, the success or failure can be
2683 * decided for all targets by invoking group_migrate_prepare_dst() before
2684 * actually starting migrating.
2685 */
2686int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2687 struct cgroup_mgctx *mgctx)
2688{
2689 struct task_struct *task;
2690
2691 /*
2692 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2693 * already PF_EXITING could be freed from underneath us unless we
2694 * take an rcu_read_lock.
2695 */
2696 spin_lock_irq(&css_set_lock);
2697 rcu_read_lock();
2698 task = leader;
2699 do {
2700 cgroup_migrate_add_task(task, mgctx);
2701 if (!threadgroup)
2702 break;
2703 } while_each_thread(leader, task);
2704 rcu_read_unlock();
2705 spin_unlock_irq(&css_set_lock);
2706
2707 return cgroup_migrate_execute(mgctx);
2708}
2709
2710/**
2711 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2712 * @dst_cgrp: the cgroup to attach to
2713 * @leader: the task or the leader of the threadgroup to be attached
2714 * @threadgroup: attach the whole threadgroup?
2715 *
2716 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2717 */
2718int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2719 bool threadgroup)
2720{
2721 DEFINE_CGROUP_MGCTX(mgctx);
2722 struct task_struct *task;
2723 int ret = 0;
2724
2725 /* look up all src csets */
2726 spin_lock_irq(&css_set_lock);
2727 rcu_read_lock();
2728 task = leader;
2729 do {
2730 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2731 if (!threadgroup)
2732 break;
2733 } while_each_thread(leader, task);
2734 rcu_read_unlock();
2735 spin_unlock_irq(&css_set_lock);
2736
2737 /* prepare dst csets and commit */
2738 ret = cgroup_migrate_prepare_dst(&mgctx);
2739 if (!ret)
2740 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2741
2742 cgroup_migrate_finish(&mgctx);
2743
2744 if (!ret)
2745 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2746
2747 return ret;
2748}
2749
2750struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2751 bool *locked)
2752 __acquires(&cgroup_threadgroup_rwsem)
2753{
2754 struct task_struct *tsk;
2755 pid_t pid;
2756
2757 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2758 return ERR_PTR(-EINVAL);
2759
2760 /*
2761 * If we migrate a single thread, we don't care about threadgroup
2762 * stability. If the thread is `current`, it won't exit(2) under our
2763 * hands or change PID through exec(2). We exclude
2764 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2765 * callers by cgroup_mutex.
2766 * Therefore, we can skip the global lock.
2767 */
2768 lockdep_assert_held(&cgroup_mutex);
2769 if (pid || threadgroup) {
2770 percpu_down_write(&cgroup_threadgroup_rwsem);
2771 *locked = true;
2772 } else {
2773 *locked = false;
2774 }
2775
2776 rcu_read_lock();
2777 if (pid) {
2778 tsk = find_task_by_vpid(pid);
2779 if (!tsk) {
2780 tsk = ERR_PTR(-ESRCH);
2781 goto out_unlock_threadgroup;
2782 }
2783 } else {
2784 tsk = current;
2785 }
2786
2787 if (threadgroup)
2788 tsk = tsk->group_leader;
2789
2790 /*
2791 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2792 * If userland migrates such a kthread to a non-root cgroup, it can
2793 * become trapped in a cpuset, or RT kthread may be born in a
2794 * cgroup with no rt_runtime allocated. Just say no.
2795 */
2796 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2797 tsk = ERR_PTR(-EINVAL);
2798 goto out_unlock_threadgroup;
2799 }
2800
2801 get_task_struct(tsk);
2802 goto out_unlock_rcu;
2803
2804out_unlock_threadgroup:
2805 if (*locked) {
2806 percpu_up_write(&cgroup_threadgroup_rwsem);
2807 *locked = false;
2808 }
2809out_unlock_rcu:
2810 rcu_read_unlock();
2811 return tsk;
2812}
2813
2814void cgroup_procs_write_finish(struct task_struct *task, bool locked)
2815 __releases(&cgroup_threadgroup_rwsem)
2816{
2817 struct cgroup_subsys *ss;
2818 int ssid;
2819
2820 /* release reference from cgroup_procs_write_start() */
2821 put_task_struct(task);
2822
2823 if (locked)
2824 percpu_up_write(&cgroup_threadgroup_rwsem);
2825 for_each_subsys(ss, ssid)
2826 if (ss->post_attach)
2827 ss->post_attach();
2828}
2829
2830static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2831{
2832 struct cgroup_subsys *ss;
2833 bool printed = false;
2834 int ssid;
2835
2836 do_each_subsys_mask(ss, ssid, ss_mask) {
2837 if (printed)
2838 seq_putc(seq, ' ');
2839 seq_puts(seq, ss->name);
2840 printed = true;
2841 } while_each_subsys_mask();
2842 if (printed)
2843 seq_putc(seq, '\n');
2844}
2845
2846/* show controllers which are enabled from the parent */
2847static int cgroup_controllers_show(struct seq_file *seq, void *v)
2848{
2849 struct cgroup *cgrp = seq_css(seq)->cgroup;
2850
2851 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2852 return 0;
2853}
2854
2855/* show controllers which are enabled for a given cgroup's children */
2856static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2857{
2858 struct cgroup *cgrp = seq_css(seq)->cgroup;
2859
2860 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2861 return 0;
2862}
2863
2864/**
2865 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2866 * @cgrp: root of the subtree to update csses for
2867 *
2868 * @cgrp's control masks have changed and its subtree's css associations
2869 * need to be updated accordingly. This function looks up all css_sets
2870 * which are attached to the subtree, creates the matching updated css_sets
2871 * and migrates the tasks to the new ones.
2872 */
2873static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2874{
2875 DEFINE_CGROUP_MGCTX(mgctx);
2876 struct cgroup_subsys_state *d_css;
2877 struct cgroup *dsct;
2878 struct css_set *src_cset;
2879 int ret;
2880
2881 lockdep_assert_held(&cgroup_mutex);
2882
2883 percpu_down_write(&cgroup_threadgroup_rwsem);
2884
2885 /* look up all csses currently attached to @cgrp's subtree */
2886 spin_lock_irq(&css_set_lock);
2887 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2888 struct cgrp_cset_link *link;
2889
2890 list_for_each_entry(link, &dsct->cset_links, cset_link)
2891 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2892 }
2893 spin_unlock_irq(&css_set_lock);
2894
2895 /* NULL dst indicates self on default hierarchy */
2896 ret = cgroup_migrate_prepare_dst(&mgctx);
2897 if (ret)
2898 goto out_finish;
2899
2900 spin_lock_irq(&css_set_lock);
2901 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2902 struct task_struct *task, *ntask;
2903
2904 /* all tasks in src_csets need to be migrated */
2905 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2906 cgroup_migrate_add_task(task, &mgctx);
2907 }
2908 spin_unlock_irq(&css_set_lock);
2909
2910 ret = cgroup_migrate_execute(&mgctx);
2911out_finish:
2912 cgroup_migrate_finish(&mgctx);
2913 percpu_up_write(&cgroup_threadgroup_rwsem);
2914 return ret;
2915}
2916
2917/**
2918 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2919 * @cgrp: root of the target subtree
2920 *
2921 * Because css offlining is asynchronous, userland may try to re-enable a
2922 * controller while the previous css is still around. This function grabs
2923 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2924 */
2925void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2926 __acquires(&cgroup_mutex)
2927{
2928 struct cgroup *dsct;
2929 struct cgroup_subsys_state *d_css;
2930 struct cgroup_subsys *ss;
2931 int ssid;
2932
2933restart:
2934 mutex_lock(&cgroup_mutex);
2935
2936 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2937 for_each_subsys(ss, ssid) {
2938 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2939 DEFINE_WAIT(wait);
2940
2941 if (!css || !percpu_ref_is_dying(&css->refcnt))
2942 continue;
2943
2944 cgroup_get_live(dsct);
2945 prepare_to_wait(&dsct->offline_waitq, &wait,
2946 TASK_UNINTERRUPTIBLE);
2947
2948 mutex_unlock(&cgroup_mutex);
2949 schedule();
2950 finish_wait(&dsct->offline_waitq, &wait);
2951
2952 cgroup_put(dsct);
2953 goto restart;
2954 }
2955 }
2956}
2957
2958/**
2959 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2960 * @cgrp: root of the target subtree
2961 *
2962 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2963 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2964 * itself.
2965 */
2966static void cgroup_save_control(struct cgroup *cgrp)
2967{
2968 struct cgroup *dsct;
2969 struct cgroup_subsys_state *d_css;
2970
2971 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2972 dsct->old_subtree_control = dsct->subtree_control;
2973 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2974 dsct->old_dom_cgrp = dsct->dom_cgrp;
2975 }
2976}
2977
2978/**
2979 * cgroup_propagate_control - refresh control masks of a subtree
2980 * @cgrp: root of the target subtree
2981 *
2982 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2983 * ->subtree_control and propagate controller availability through the
2984 * subtree so that descendants don't have unavailable controllers enabled.
2985 */
2986static void cgroup_propagate_control(struct cgroup *cgrp)
2987{
2988 struct cgroup *dsct;
2989 struct cgroup_subsys_state *d_css;
2990
2991 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2992 dsct->subtree_control &= cgroup_control(dsct);
2993 dsct->subtree_ss_mask =
2994 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2995 cgroup_ss_mask(dsct));
2996 }
2997}
2998
2999/**
3000 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3001 * @cgrp: root of the target subtree
3002 *
3003 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3004 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3005 * itself.
3006 */
3007static void cgroup_restore_control(struct cgroup *cgrp)
3008{
3009 struct cgroup *dsct;
3010 struct cgroup_subsys_state *d_css;
3011
3012 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3013 dsct->subtree_control = dsct->old_subtree_control;
3014 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3015 dsct->dom_cgrp = dsct->old_dom_cgrp;
3016 }
3017}
3018
3019static bool css_visible(struct cgroup_subsys_state *css)
3020{
3021 struct cgroup_subsys *ss = css->ss;
3022 struct cgroup *cgrp = css->cgroup;
3023
3024 if (cgroup_control(cgrp) & (1 << ss->id))
3025 return true;
3026 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3027 return false;
3028 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3029}
3030
3031/**
3032 * cgroup_apply_control_enable - enable or show csses according to control
3033 * @cgrp: root of the target subtree
3034 *
3035 * Walk @cgrp's subtree and create new csses or make the existing ones
3036 * visible. A css is created invisible if it's being implicitly enabled
3037 * through dependency. An invisible css is made visible when the userland
3038 * explicitly enables it.
3039 *
3040 * Returns 0 on success, -errno on failure. On failure, csses which have
3041 * been processed already aren't cleaned up. The caller is responsible for
3042 * cleaning up with cgroup_apply_control_disable().
3043 */
3044static int cgroup_apply_control_enable(struct cgroup *cgrp)
3045{
3046 struct cgroup *dsct;
3047 struct cgroup_subsys_state *d_css;
3048 struct cgroup_subsys *ss;
3049 int ssid, ret;
3050
3051 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3052 for_each_subsys(ss, ssid) {
3053 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3054
3055 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3056 continue;
3057
3058 if (!css) {
3059 css = css_create(dsct, ss);
3060 if (IS_ERR(css))
3061 return PTR_ERR(css);
3062 }
3063
3064 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3065
3066 if (css_visible(css)) {
3067 ret = css_populate_dir(css);
3068 if (ret)
3069 return ret;
3070 }
3071 }
3072 }
3073
3074 return 0;
3075}
3076
3077/**
3078 * cgroup_apply_control_disable - kill or hide csses according to control
3079 * @cgrp: root of the target subtree
3080 *
3081 * Walk @cgrp's subtree and kill and hide csses so that they match
3082 * cgroup_ss_mask() and cgroup_visible_mask().
3083 *
3084 * A css is hidden when the userland requests it to be disabled while other
3085 * subsystems are still depending on it. The css must not actively control
3086 * resources and be in the vanilla state if it's made visible again later.
3087 * Controllers which may be depended upon should provide ->css_reset() for
3088 * this purpose.
3089 */
3090static void cgroup_apply_control_disable(struct cgroup *cgrp)
3091{
3092 struct cgroup *dsct;
3093 struct cgroup_subsys_state *d_css;
3094 struct cgroup_subsys *ss;
3095 int ssid;
3096
3097 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3098 for_each_subsys(ss, ssid) {
3099 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3100
3101 if (!css)
3102 continue;
3103
3104 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3105
3106 if (css->parent &&
3107 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3108 kill_css(css);
3109 } else if (!css_visible(css)) {
3110 css_clear_dir(css);
3111 if (ss->css_reset)
3112 ss->css_reset(css);
3113 }
3114 }
3115 }
3116}
3117
3118/**
3119 * cgroup_apply_control - apply control mask updates to the subtree
3120 * @cgrp: root of the target subtree
3121 *
3122 * subsystems can be enabled and disabled in a subtree using the following
3123 * steps.
3124 *
3125 * 1. Call cgroup_save_control() to stash the current state.
3126 * 2. Update ->subtree_control masks in the subtree as desired.
3127 * 3. Call cgroup_apply_control() to apply the changes.
3128 * 4. Optionally perform other related operations.
3129 * 5. Call cgroup_finalize_control() to finish up.
3130 *
3131 * This function implements step 3 and propagates the mask changes
3132 * throughout @cgrp's subtree, updates csses accordingly and perform
3133 * process migrations.
3134 */
3135static int cgroup_apply_control(struct cgroup *cgrp)
3136{
3137 int ret;
3138
3139 cgroup_propagate_control(cgrp);
3140
3141 ret = cgroup_apply_control_enable(cgrp);
3142 if (ret)
3143 return ret;
3144
3145 /*
3146 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3147 * making the following cgroup_update_dfl_csses() properly update
3148 * css associations of all tasks in the subtree.
3149 */
3150 ret = cgroup_update_dfl_csses(cgrp);
3151 if (ret)
3152 return ret;
3153
3154 return 0;
3155}
3156
3157/**
3158 * cgroup_finalize_control - finalize control mask update
3159 * @cgrp: root of the target subtree
3160 * @ret: the result of the update
3161 *
3162 * Finalize control mask update. See cgroup_apply_control() for more info.
3163 */
3164static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3165{
3166 if (ret) {
3167 cgroup_restore_control(cgrp);
3168 cgroup_propagate_control(cgrp);
3169 }
3170
3171 cgroup_apply_control_disable(cgrp);
3172}
3173
3174static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3175{
3176 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3177
3178 /* if nothing is getting enabled, nothing to worry about */
3179 if (!enable)
3180 return 0;
3181
3182 /* can @cgrp host any resources? */
3183 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3184 return -EOPNOTSUPP;
3185
3186 /* mixables don't care */
3187 if (cgroup_is_mixable(cgrp))
3188 return 0;
3189
3190 if (domain_enable) {
3191 /* can't enable domain controllers inside a thread subtree */
3192 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3193 return -EOPNOTSUPP;
3194 } else {
3195 /*
3196 * Threaded controllers can handle internal competitions
3197 * and are always allowed inside a (prospective) thread
3198 * subtree.
3199 */
3200 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3201 return 0;
3202 }
3203
3204 /*
3205 * Controllers can't be enabled for a cgroup with tasks to avoid
3206 * child cgroups competing against tasks.
3207 */
3208 if (cgroup_has_tasks(cgrp))
3209 return -EBUSY;
3210
3211 return 0;
3212}
3213
3214/* change the enabled child controllers for a cgroup in the default hierarchy */
3215static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3216 char *buf, size_t nbytes,
3217 loff_t off)
3218{
3219 u16 enable = 0, disable = 0;
3220 struct cgroup *cgrp, *child;
3221 struct cgroup_subsys *ss;
3222 char *tok;
3223 int ssid, ret;
3224
3225 /*
3226 * Parse input - space separated list of subsystem names prefixed
3227 * with either + or -.
3228 */
3229 buf = strstrip(buf);
3230 while ((tok = strsep(&buf, " "))) {
3231 if (tok[0] == '\0')
3232 continue;
3233 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3234 if (!cgroup_ssid_enabled(ssid) ||
3235 strcmp(tok + 1, ss->name))
3236 continue;
3237
3238 if (*tok == '+') {
3239 enable |= 1 << ssid;
3240 disable &= ~(1 << ssid);
3241 } else if (*tok == '-') {
3242 disable |= 1 << ssid;
3243 enable &= ~(1 << ssid);
3244 } else {
3245 return -EINVAL;
3246 }
3247 break;
3248 } while_each_subsys_mask();
3249 if (ssid == CGROUP_SUBSYS_COUNT)
3250 return -EINVAL;
3251 }
3252
3253 cgrp = cgroup_kn_lock_live(of->kn, true);
3254 if (!cgrp)
3255 return -ENODEV;
3256
3257 for_each_subsys(ss, ssid) {
3258 if (enable & (1 << ssid)) {
3259 if (cgrp->subtree_control & (1 << ssid)) {
3260 enable &= ~(1 << ssid);
3261 continue;
3262 }
3263
3264 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3265 ret = -ENOENT;
3266 goto out_unlock;
3267 }
3268 } else if (disable & (1 << ssid)) {
3269 if (!(cgrp->subtree_control & (1 << ssid))) {
3270 disable &= ~(1 << ssid);
3271 continue;
3272 }
3273
3274 /* a child has it enabled? */
3275 cgroup_for_each_live_child(child, cgrp) {
3276 if (child->subtree_control & (1 << ssid)) {
3277 ret = -EBUSY;
3278 goto out_unlock;
3279 }
3280 }
3281 }
3282 }
3283
3284 if (!enable && !disable) {
3285 ret = 0;
3286 goto out_unlock;
3287 }
3288
3289 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3290 if (ret)
3291 goto out_unlock;
3292
3293 /* save and update control masks and prepare csses */
3294 cgroup_save_control(cgrp);
3295
3296 cgrp->subtree_control |= enable;
3297 cgrp->subtree_control &= ~disable;
3298
3299 ret = cgroup_apply_control(cgrp);
3300 cgroup_finalize_control(cgrp, ret);
3301 if (ret)
3302 goto out_unlock;
3303
3304 kernfs_activate(cgrp->kn);
3305out_unlock:
3306 cgroup_kn_unlock(of->kn);
3307 return ret ?: nbytes;
3308}
3309
3310/**
3311 * cgroup_enable_threaded - make @cgrp threaded
3312 * @cgrp: the target cgroup
3313 *
3314 * Called when "threaded" is written to the cgroup.type interface file and
3315 * tries to make @cgrp threaded and join the parent's resource domain.
3316 * This function is never called on the root cgroup as cgroup.type doesn't
3317 * exist on it.
3318 */
3319static int cgroup_enable_threaded(struct cgroup *cgrp)
3320{
3321 struct cgroup *parent = cgroup_parent(cgrp);
3322 struct cgroup *dom_cgrp = parent->dom_cgrp;
3323 struct cgroup *dsct;
3324 struct cgroup_subsys_state *d_css;
3325 int ret;
3326
3327 lockdep_assert_held(&cgroup_mutex);
3328
3329 /* noop if already threaded */
3330 if (cgroup_is_threaded(cgrp))
3331 return 0;
3332
3333 /*
3334 * If @cgroup is populated or has domain controllers enabled, it
3335 * can't be switched. While the below cgroup_can_be_thread_root()
3336 * test can catch the same conditions, that's only when @parent is
3337 * not mixable, so let's check it explicitly.
3338 */
3339 if (cgroup_is_populated(cgrp) ||
3340 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3341 return -EOPNOTSUPP;
3342
3343 /* we're joining the parent's domain, ensure its validity */
3344 if (!cgroup_is_valid_domain(dom_cgrp) ||
3345 !cgroup_can_be_thread_root(dom_cgrp))
3346 return -EOPNOTSUPP;
3347
3348 /*
3349 * The following shouldn't cause actual migrations and should
3350 * always succeed.
3351 */
3352 cgroup_save_control(cgrp);
3353
3354 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3355 if (dsct == cgrp || cgroup_is_threaded(dsct))
3356 dsct->dom_cgrp = dom_cgrp;
3357
3358 ret = cgroup_apply_control(cgrp);
3359 if (!ret)
3360 parent->nr_threaded_children++;
3361
3362 cgroup_finalize_control(cgrp, ret);
3363 return ret;
3364}
3365
3366static int cgroup_type_show(struct seq_file *seq, void *v)
3367{
3368 struct cgroup *cgrp = seq_css(seq)->cgroup;
3369
3370 if (cgroup_is_threaded(cgrp))
3371 seq_puts(seq, "threaded\n");
3372 else if (!cgroup_is_valid_domain(cgrp))
3373 seq_puts(seq, "domain invalid\n");
3374 else if (cgroup_is_thread_root(cgrp))
3375 seq_puts(seq, "domain threaded\n");
3376 else
3377 seq_puts(seq, "domain\n");
3378
3379 return 0;
3380}
3381
3382static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3383 size_t nbytes, loff_t off)
3384{
3385 struct cgroup *cgrp;
3386 int ret;
3387
3388 /* only switching to threaded mode is supported */
3389 if (strcmp(strstrip(buf), "threaded"))
3390 return -EINVAL;
3391
3392 /* drain dying csses before we re-apply (threaded) subtree control */
3393 cgrp = cgroup_kn_lock_live(of->kn, true);
3394 if (!cgrp)
3395 return -ENOENT;
3396
3397 /* threaded can only be enabled */
3398 ret = cgroup_enable_threaded(cgrp);
3399
3400 cgroup_kn_unlock(of->kn);
3401 return ret ?: nbytes;
3402}
3403
3404static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3405{
3406 struct cgroup *cgrp = seq_css(seq)->cgroup;
3407 int descendants = READ_ONCE(cgrp->max_descendants);
3408
3409 if (descendants == INT_MAX)
3410 seq_puts(seq, "max\n");
3411 else
3412 seq_printf(seq, "%d\n", descendants);
3413
3414 return 0;
3415}
3416
3417static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3418 char *buf, size_t nbytes, loff_t off)
3419{
3420 struct cgroup *cgrp;
3421 int descendants;
3422 ssize_t ret;
3423
3424 buf = strstrip(buf);
3425 if (!strcmp(buf, "max")) {
3426 descendants = INT_MAX;
3427 } else {
3428 ret = kstrtoint(buf, 0, &descendants);
3429 if (ret)
3430 return ret;
3431 }
3432
3433 if (descendants < 0)
3434 return -ERANGE;
3435
3436 cgrp = cgroup_kn_lock_live(of->kn, false);
3437 if (!cgrp)
3438 return -ENOENT;
3439
3440 cgrp->max_descendants = descendants;
3441
3442 cgroup_kn_unlock(of->kn);
3443
3444 return nbytes;
3445}
3446
3447static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3448{
3449 struct cgroup *cgrp = seq_css(seq)->cgroup;
3450 int depth = READ_ONCE(cgrp->max_depth);
3451
3452 if (depth == INT_MAX)
3453 seq_puts(seq, "max\n");
3454 else
3455 seq_printf(seq, "%d\n", depth);
3456
3457 return 0;
3458}
3459
3460static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3461 char *buf, size_t nbytes, loff_t off)
3462{
3463 struct cgroup *cgrp;
3464 ssize_t ret;
3465 int depth;
3466
3467 buf = strstrip(buf);
3468 if (!strcmp(buf, "max")) {
3469 depth = INT_MAX;
3470 } else {
3471 ret = kstrtoint(buf, 0, &depth);
3472 if (ret)
3473 return ret;
3474 }
3475
3476 if (depth < 0)
3477 return -ERANGE;
3478
3479 cgrp = cgroup_kn_lock_live(of->kn, false);
3480 if (!cgrp)
3481 return -ENOENT;
3482
3483 cgrp->max_depth = depth;
3484
3485 cgroup_kn_unlock(of->kn);
3486
3487 return nbytes;
3488}
3489
3490static int cgroup_events_show(struct seq_file *seq, void *v)
3491{
3492 struct cgroup *cgrp = seq_css(seq)->cgroup;
3493
3494 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3495 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3496
3497 return 0;
3498}
3499
3500static int cgroup_stat_show(struct seq_file *seq, void *v)
3501{
3502 struct cgroup *cgroup = seq_css(seq)->cgroup;
3503
3504 seq_printf(seq, "nr_descendants %d\n",
3505 cgroup->nr_descendants);
3506 seq_printf(seq, "nr_dying_descendants %d\n",
3507 cgroup->nr_dying_descendants);
3508
3509 return 0;
3510}
3511
3512static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3513 struct cgroup *cgrp, int ssid)
3514{
3515 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3516 struct cgroup_subsys_state *css;
3517 int ret;
3518
3519 if (!ss->css_extra_stat_show)
3520 return 0;
3521
3522 css = cgroup_tryget_css(cgrp, ss);
3523 if (!css)
3524 return 0;
3525
3526 ret = ss->css_extra_stat_show(seq, css);
3527 css_put(css);
3528 return ret;
3529}
3530
3531static int cpu_stat_show(struct seq_file *seq, void *v)
3532{
3533 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3534 int ret = 0;
3535
3536 cgroup_base_stat_cputime_show(seq);
3537#ifdef CONFIG_CGROUP_SCHED
3538 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3539#endif
3540 return ret;
3541}
3542
3543#ifdef CONFIG_PSI
3544static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3545{
3546 struct cgroup *cgrp = seq_css(seq)->cgroup;
3547 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3548
3549 return psi_show(seq, psi, PSI_IO);
3550}
3551static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3552{
3553 struct cgroup *cgrp = seq_css(seq)->cgroup;
3554 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3555
3556 return psi_show(seq, psi, PSI_MEM);
3557}
3558static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3559{
3560 struct cgroup *cgrp = seq_css(seq)->cgroup;
3561 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3562
3563 return psi_show(seq, psi, PSI_CPU);
3564}
3565
3566static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3567 size_t nbytes, enum psi_res res)
3568{
3569 struct psi_trigger *new;
3570 struct cgroup *cgrp;
3571
3572 cgrp = cgroup_kn_lock_live(of->kn, false);
3573 if (!cgrp)
3574 return -ENODEV;
3575
3576 cgroup_get(cgrp);
3577 cgroup_kn_unlock(of->kn);
3578
3579 new = psi_trigger_create(&cgrp->psi, buf, nbytes, res);
3580 if (IS_ERR(new)) {
3581 cgroup_put(cgrp);
3582 return PTR_ERR(new);
3583 }
3584
3585 psi_trigger_replace(&of->priv, new);
3586
3587 cgroup_put(cgrp);
3588
3589 return nbytes;
3590}
3591
3592static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3593 char *buf, size_t nbytes,
3594 loff_t off)
3595{
3596 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3597}
3598
3599static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3600 char *buf, size_t nbytes,
3601 loff_t off)
3602{
3603 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3604}
3605
3606static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3607 char *buf, size_t nbytes,
3608 loff_t off)
3609{
3610 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3611}
3612
3613static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3614 poll_table *pt)
3615{
3616 return psi_trigger_poll(&of->priv, of->file, pt);
3617}
3618
3619static void cgroup_pressure_release(struct kernfs_open_file *of)
3620{
3621 psi_trigger_replace(&of->priv, NULL);
3622}
3623#endif /* CONFIG_PSI */
3624
3625static int cgroup_freeze_show(struct seq_file *seq, void *v)
3626{
3627 struct cgroup *cgrp = seq_css(seq)->cgroup;
3628
3629 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3630
3631 return 0;
3632}
3633
3634static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3635 char *buf, size_t nbytes, loff_t off)
3636{
3637 struct cgroup *cgrp;
3638 ssize_t ret;
3639 int freeze;
3640
3641 ret = kstrtoint(strstrip(buf), 0, &freeze);
3642 if (ret)
3643 return ret;
3644
3645 if (freeze < 0 || freeze > 1)
3646 return -ERANGE;
3647
3648 cgrp = cgroup_kn_lock_live(of->kn, false);
3649 if (!cgrp)
3650 return -ENOENT;
3651
3652 cgroup_freeze(cgrp, freeze);
3653
3654 cgroup_kn_unlock(of->kn);
3655
3656 return nbytes;
3657}
3658
3659static int cgroup_file_open(struct kernfs_open_file *of)
3660{
3661 struct cftype *cft = of->kn->priv;
3662
3663 if (cft->open)
3664 return cft->open(of);
3665 return 0;
3666}
3667
3668static void cgroup_file_release(struct kernfs_open_file *of)
3669{
3670 struct cftype *cft = of->kn->priv;
3671
3672 if (cft->release)
3673 cft->release(of);
3674}
3675
3676static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3677 size_t nbytes, loff_t off)
3678{
3679 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3680 struct cgroup *cgrp = of->kn->parent->priv;
3681 struct cftype *cft = of->kn->priv;
3682 struct cgroup_subsys_state *css;
3683 int ret;
3684
3685 /*
3686 * If namespaces are delegation boundaries, disallow writes to
3687 * files in an non-init namespace root from inside the namespace
3688 * except for the files explicitly marked delegatable -
3689 * cgroup.procs and cgroup.subtree_control.
3690 */
3691 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3692 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3693 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3694 return -EPERM;
3695
3696 if (cft->write)
3697 return cft->write(of, buf, nbytes, off);
3698
3699 /*
3700 * kernfs guarantees that a file isn't deleted with operations in
3701 * flight, which means that the matching css is and stays alive and
3702 * doesn't need to be pinned. The RCU locking is not necessary
3703 * either. It's just for the convenience of using cgroup_css().
3704 */
3705 rcu_read_lock();
3706 css = cgroup_css(cgrp, cft->ss);
3707 rcu_read_unlock();
3708
3709 if (cft->write_u64) {
3710 unsigned long long v;
3711 ret = kstrtoull(buf, 0, &v);
3712 if (!ret)
3713 ret = cft->write_u64(css, cft, v);
3714 } else if (cft->write_s64) {
3715 long long v;
3716 ret = kstrtoll(buf, 0, &v);
3717 if (!ret)
3718 ret = cft->write_s64(css, cft, v);
3719 } else {
3720 ret = -EINVAL;
3721 }
3722
3723 return ret ?: nbytes;
3724}
3725
3726static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3727{
3728 struct cftype *cft = of->kn->priv;
3729
3730 if (cft->poll)
3731 return cft->poll(of, pt);
3732
3733 return kernfs_generic_poll(of, pt);
3734}
3735
3736static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3737{
3738 return seq_cft(seq)->seq_start(seq, ppos);
3739}
3740
3741static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3742{
3743 return seq_cft(seq)->seq_next(seq, v, ppos);
3744}
3745
3746static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3747{
3748 if (seq_cft(seq)->seq_stop)
3749 seq_cft(seq)->seq_stop(seq, v);
3750}
3751
3752static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3753{
3754 struct cftype *cft = seq_cft(m);
3755 struct cgroup_subsys_state *css = seq_css(m);
3756
3757 if (cft->seq_show)
3758 return cft->seq_show(m, arg);
3759
3760 if (cft->read_u64)
3761 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3762 else if (cft->read_s64)
3763 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3764 else
3765 return -EINVAL;
3766 return 0;
3767}
3768
3769static struct kernfs_ops cgroup_kf_single_ops = {
3770 .atomic_write_len = PAGE_SIZE,
3771 .open = cgroup_file_open,
3772 .release = cgroup_file_release,
3773 .write = cgroup_file_write,
3774 .poll = cgroup_file_poll,
3775 .seq_show = cgroup_seqfile_show,
3776};
3777
3778static struct kernfs_ops cgroup_kf_ops = {
3779 .atomic_write_len = PAGE_SIZE,
3780 .open = cgroup_file_open,
3781 .release = cgroup_file_release,
3782 .write = cgroup_file_write,
3783 .poll = cgroup_file_poll,
3784 .seq_start = cgroup_seqfile_start,
3785 .seq_next = cgroup_seqfile_next,
3786 .seq_stop = cgroup_seqfile_stop,
3787 .seq_show = cgroup_seqfile_show,
3788};
3789
3790/* set uid and gid of cgroup dirs and files to that of the creator */
3791static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3792{
3793 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3794 .ia_uid = current_fsuid(),
3795 .ia_gid = current_fsgid(), };
3796
3797 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3798 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3799 return 0;
3800
3801 return kernfs_setattr(kn, &iattr);
3802}
3803
3804static void cgroup_file_notify_timer(struct timer_list *timer)
3805{
3806 cgroup_file_notify(container_of(timer, struct cgroup_file,
3807 notify_timer));
3808}
3809
3810static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3811 struct cftype *cft)
3812{
3813 char name[CGROUP_FILE_NAME_MAX];
3814 struct kernfs_node *kn;
3815 struct lock_class_key *key = NULL;
3816 int ret;
3817
3818#ifdef CONFIG_DEBUG_LOCK_ALLOC
3819 key = &cft->lockdep_key;
3820#endif
3821 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3822 cgroup_file_mode(cft),
3823 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3824 0, cft->kf_ops, cft,
3825 NULL, key);
3826 if (IS_ERR(kn))
3827 return PTR_ERR(kn);
3828
3829 ret = cgroup_kn_set_ugid(kn);
3830 if (ret) {
3831 kernfs_remove(kn);
3832 return ret;
3833 }
3834
3835 if (cft->file_offset) {
3836 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3837
3838 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3839
3840 spin_lock_irq(&cgroup_file_kn_lock);
3841 cfile->kn = kn;
3842 spin_unlock_irq(&cgroup_file_kn_lock);
3843 }
3844
3845 return 0;
3846}
3847
3848/**
3849 * cgroup_addrm_files - add or remove files to a cgroup directory
3850 * @css: the target css
3851 * @cgrp: the target cgroup (usually css->cgroup)
3852 * @cfts: array of cftypes to be added
3853 * @is_add: whether to add or remove
3854 *
3855 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3856 * For removals, this function never fails.
3857 */
3858static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3859 struct cgroup *cgrp, struct cftype cfts[],
3860 bool is_add)
3861{
3862 struct cftype *cft, *cft_end = NULL;
3863 int ret = 0;
3864
3865 lockdep_assert_held(&cgroup_mutex);
3866
3867restart:
3868 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3869 /* does cft->flags tell us to skip this file on @cgrp? */
3870 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3871 continue;
3872 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3873 continue;
3874 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3875 continue;
3876 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3877 continue;
3878 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3879 continue;
3880 if (is_add) {
3881 ret = cgroup_add_file(css, cgrp, cft);
3882 if (ret) {
3883 pr_warn("%s: failed to add %s, err=%d\n",
3884 __func__, cft->name, ret);
3885 cft_end = cft;
3886 is_add = false;
3887 goto restart;
3888 }
3889 } else {
3890 cgroup_rm_file(cgrp, cft);
3891 }
3892 }
3893 return ret;
3894}
3895
3896static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3897{
3898 struct cgroup_subsys *ss = cfts[0].ss;
3899 struct cgroup *root = &ss->root->cgrp;
3900 struct cgroup_subsys_state *css;
3901 int ret = 0;
3902
3903 lockdep_assert_held(&cgroup_mutex);
3904
3905 /* add/rm files for all cgroups created before */
3906 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3907 struct cgroup *cgrp = css->cgroup;
3908
3909 if (!(css->flags & CSS_VISIBLE))
3910 continue;
3911
3912 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3913 if (ret)
3914 break;
3915 }
3916
3917 if (is_add && !ret)
3918 kernfs_activate(root->kn);
3919 return ret;
3920}
3921
3922static void cgroup_exit_cftypes(struct cftype *cfts)
3923{
3924 struct cftype *cft;
3925
3926 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3927 /* free copy for custom atomic_write_len, see init_cftypes() */
3928 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3929 kfree(cft->kf_ops);
3930 cft->kf_ops = NULL;
3931 cft->ss = NULL;
3932
3933 /* revert flags set by cgroup core while adding @cfts */
3934 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3935 }
3936}
3937
3938static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3939{
3940 struct cftype *cft;
3941
3942 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3943 struct kernfs_ops *kf_ops;
3944
3945 WARN_ON(cft->ss || cft->kf_ops);
3946
3947 if (cft->seq_start)
3948 kf_ops = &cgroup_kf_ops;
3949 else
3950 kf_ops = &cgroup_kf_single_ops;
3951
3952 /*
3953 * Ugh... if @cft wants a custom max_write_len, we need to
3954 * make a copy of kf_ops to set its atomic_write_len.
3955 */
3956 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3957 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3958 if (!kf_ops) {
3959 cgroup_exit_cftypes(cfts);
3960 return -ENOMEM;
3961 }
3962 kf_ops->atomic_write_len = cft->max_write_len;
3963 }
3964
3965 cft->kf_ops = kf_ops;
3966 cft->ss = ss;
3967 }
3968
3969 return 0;
3970}
3971
3972static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3973{
3974 lockdep_assert_held(&cgroup_mutex);
3975
3976 if (!cfts || !cfts[0].ss)
3977 return -ENOENT;
3978
3979 list_del(&cfts->node);
3980 cgroup_apply_cftypes(cfts, false);
3981 cgroup_exit_cftypes(cfts);
3982 return 0;
3983}
3984
3985/**
3986 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3987 * @cfts: zero-length name terminated array of cftypes
3988 *
3989 * Unregister @cfts. Files described by @cfts are removed from all
3990 * existing cgroups and all future cgroups won't have them either. This
3991 * function can be called anytime whether @cfts' subsys is attached or not.
3992 *
3993 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3994 * registered.
3995 */
3996int cgroup_rm_cftypes(struct cftype *cfts)
3997{
3998 int ret;
3999
4000 mutex_lock(&cgroup_mutex);
4001 ret = cgroup_rm_cftypes_locked(cfts);
4002 mutex_unlock(&cgroup_mutex);
4003 return ret;
4004}
4005
4006/**
4007 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4008 * @ss: target cgroup subsystem
4009 * @cfts: zero-length name terminated array of cftypes
4010 *
4011 * Register @cfts to @ss. Files described by @cfts are created for all
4012 * existing cgroups to which @ss is attached and all future cgroups will
4013 * have them too. This function can be called anytime whether @ss is
4014 * attached or not.
4015 *
4016 * Returns 0 on successful registration, -errno on failure. Note that this
4017 * function currently returns 0 as long as @cfts registration is successful
4018 * even if some file creation attempts on existing cgroups fail.
4019 */
4020static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4021{
4022 int ret;
4023
4024 if (!cgroup_ssid_enabled(ss->id))
4025 return 0;
4026
4027 if (!cfts || cfts[0].name[0] == '\0')
4028 return 0;
4029
4030 ret = cgroup_init_cftypes(ss, cfts);
4031 if (ret)
4032 return ret;
4033
4034 mutex_lock(&cgroup_mutex);
4035
4036 list_add_tail(&cfts->node, &ss->cfts);
4037 ret = cgroup_apply_cftypes(cfts, true);
4038 if (ret)
4039 cgroup_rm_cftypes_locked(cfts);
4040
4041 mutex_unlock(&cgroup_mutex);
4042 return ret;
4043}
4044
4045/**
4046 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4047 * @ss: target cgroup subsystem
4048 * @cfts: zero-length name terminated array of cftypes
4049 *
4050 * Similar to cgroup_add_cftypes() but the added files are only used for
4051 * the default hierarchy.
4052 */
4053int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4054{
4055 struct cftype *cft;
4056
4057 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4058 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4059 return cgroup_add_cftypes(ss, cfts);
4060}
4061
4062/**
4063 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4064 * @ss: target cgroup subsystem
4065 * @cfts: zero-length name terminated array of cftypes
4066 *
4067 * Similar to cgroup_add_cftypes() but the added files are only used for
4068 * the legacy hierarchies.
4069 */
4070int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4071{
4072 struct cftype *cft;
4073
4074 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4075 cft->flags |= __CFTYPE_NOT_ON_DFL;
4076 return cgroup_add_cftypes(ss, cfts);
4077}
4078
4079/**
4080 * cgroup_file_notify - generate a file modified event for a cgroup_file
4081 * @cfile: target cgroup_file
4082 *
4083 * @cfile must have been obtained by setting cftype->file_offset.
4084 */
4085void cgroup_file_notify(struct cgroup_file *cfile)
4086{
4087 unsigned long flags;
4088
4089 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4090 if (cfile->kn) {
4091 unsigned long last = cfile->notified_at;
4092 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4093
4094 if (time_in_range(jiffies, last, next)) {
4095 timer_reduce(&cfile->notify_timer, next);
4096 } else {
4097 kernfs_notify(cfile->kn);
4098 cfile->notified_at = jiffies;
4099 }
4100 }
4101 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4102}
4103
4104/**
4105 * css_next_child - find the next child of a given css
4106 * @pos: the current position (%NULL to initiate traversal)
4107 * @parent: css whose children to walk
4108 *
4109 * This function returns the next child of @parent and should be called
4110 * under either cgroup_mutex or RCU read lock. The only requirement is
4111 * that @parent and @pos are accessible. The next sibling is guaranteed to
4112 * be returned regardless of their states.
4113 *
4114 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4115 * css which finished ->css_online() is guaranteed to be visible in the
4116 * future iterations and will stay visible until the last reference is put.
4117 * A css which hasn't finished ->css_online() or already finished
4118 * ->css_offline() may show up during traversal. It's each subsystem's
4119 * responsibility to synchronize against on/offlining.
4120 */
4121struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4122 struct cgroup_subsys_state *parent)
4123{
4124 struct cgroup_subsys_state *next;
4125
4126 cgroup_assert_mutex_or_rcu_locked();
4127
4128 /*
4129 * @pos could already have been unlinked from the sibling list.
4130 * Once a cgroup is removed, its ->sibling.next is no longer
4131 * updated when its next sibling changes. CSS_RELEASED is set when
4132 * @pos is taken off list, at which time its next pointer is valid,
4133 * and, as releases are serialized, the one pointed to by the next
4134 * pointer is guaranteed to not have started release yet. This
4135 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4136 * critical section, the one pointed to by its next pointer is
4137 * guaranteed to not have finished its RCU grace period even if we
4138 * have dropped rcu_read_lock() inbetween iterations.
4139 *
4140 * If @pos has CSS_RELEASED set, its next pointer can't be
4141 * dereferenced; however, as each css is given a monotonically
4142 * increasing unique serial number and always appended to the
4143 * sibling list, the next one can be found by walking the parent's
4144 * children until the first css with higher serial number than
4145 * @pos's. While this path can be slower, it happens iff iteration
4146 * races against release and the race window is very small.
4147 */
4148 if (!pos) {
4149 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4150 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4151 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4152 } else {
4153 list_for_each_entry_rcu(next, &parent->children, sibling,
4154 lockdep_is_held(&cgroup_mutex))
4155 if (next->serial_nr > pos->serial_nr)
4156 break;
4157 }
4158
4159 /*
4160 * @next, if not pointing to the head, can be dereferenced and is
4161 * the next sibling.
4162 */
4163 if (&next->sibling != &parent->children)
4164 return next;
4165 return NULL;
4166}
4167
4168/**
4169 * css_next_descendant_pre - find the next descendant for pre-order walk
4170 * @pos: the current position (%NULL to initiate traversal)
4171 * @root: css whose descendants to walk
4172 *
4173 * To be used by css_for_each_descendant_pre(). Find the next descendant
4174 * to visit for pre-order traversal of @root's descendants. @root is
4175 * included in the iteration and the first node to be visited.
4176 *
4177 * While this function requires cgroup_mutex or RCU read locking, it
4178 * doesn't require the whole traversal to be contained in a single critical
4179 * section. This function will return the correct next descendant as long
4180 * as both @pos and @root are accessible and @pos is a descendant of @root.
4181 *
4182 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4183 * css which finished ->css_online() is guaranteed to be visible in the
4184 * future iterations and will stay visible until the last reference is put.
4185 * A css which hasn't finished ->css_online() or already finished
4186 * ->css_offline() may show up during traversal. It's each subsystem's
4187 * responsibility to synchronize against on/offlining.
4188 */
4189struct cgroup_subsys_state *
4190css_next_descendant_pre(struct cgroup_subsys_state *pos,
4191 struct cgroup_subsys_state *root)
4192{
4193 struct cgroup_subsys_state *next;
4194
4195 cgroup_assert_mutex_or_rcu_locked();
4196
4197 /* if first iteration, visit @root */
4198 if (!pos)
4199 return root;
4200
4201 /* visit the first child if exists */
4202 next = css_next_child(NULL, pos);
4203 if (next)
4204 return next;
4205
4206 /* no child, visit my or the closest ancestor's next sibling */
4207 while (pos != root) {
4208 next = css_next_child(pos, pos->parent);
4209 if (next)
4210 return next;
4211 pos = pos->parent;
4212 }
4213
4214 return NULL;
4215}
4216EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4217
4218/**
4219 * css_rightmost_descendant - return the rightmost descendant of a css
4220 * @pos: css of interest
4221 *
4222 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4223 * is returned. This can be used during pre-order traversal to skip
4224 * subtree of @pos.
4225 *
4226 * While this function requires cgroup_mutex or RCU read locking, it
4227 * doesn't require the whole traversal to be contained in a single critical
4228 * section. This function will return the correct rightmost descendant as
4229 * long as @pos is accessible.
4230 */
4231struct cgroup_subsys_state *
4232css_rightmost_descendant(struct cgroup_subsys_state *pos)
4233{
4234 struct cgroup_subsys_state *last, *tmp;
4235
4236 cgroup_assert_mutex_or_rcu_locked();
4237
4238 do {
4239 last = pos;
4240 /* ->prev isn't RCU safe, walk ->next till the end */
4241 pos = NULL;
4242 css_for_each_child(tmp, last)
4243 pos = tmp;
4244 } while (pos);
4245
4246 return last;
4247}
4248
4249static struct cgroup_subsys_state *
4250css_leftmost_descendant(struct cgroup_subsys_state *pos)
4251{
4252 struct cgroup_subsys_state *last;
4253
4254 do {
4255 last = pos;
4256 pos = css_next_child(NULL, pos);
4257 } while (pos);
4258
4259 return last;
4260}
4261
4262/**
4263 * css_next_descendant_post - find the next descendant for post-order walk
4264 * @pos: the current position (%NULL to initiate traversal)
4265 * @root: css whose descendants to walk
4266 *
4267 * To be used by css_for_each_descendant_post(). Find the next descendant
4268 * to visit for post-order traversal of @root's descendants. @root is
4269 * included in the iteration and the last node to be visited.
4270 *
4271 * While this function requires cgroup_mutex or RCU read locking, it
4272 * doesn't require the whole traversal to be contained in a single critical
4273 * section. This function will return the correct next descendant as long
4274 * as both @pos and @cgroup are accessible and @pos is a descendant of
4275 * @cgroup.
4276 *
4277 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4278 * css which finished ->css_online() is guaranteed to be visible in the
4279 * future iterations and will stay visible until the last reference is put.
4280 * A css which hasn't finished ->css_online() or already finished
4281 * ->css_offline() may show up during traversal. It's each subsystem's
4282 * responsibility to synchronize against on/offlining.
4283 */
4284struct cgroup_subsys_state *
4285css_next_descendant_post(struct cgroup_subsys_state *pos,
4286 struct cgroup_subsys_state *root)
4287{
4288 struct cgroup_subsys_state *next;
4289
4290 cgroup_assert_mutex_or_rcu_locked();
4291
4292 /* if first iteration, visit leftmost descendant which may be @root */
4293 if (!pos)
4294 return css_leftmost_descendant(root);
4295
4296 /* if we visited @root, we're done */
4297 if (pos == root)
4298 return NULL;
4299
4300 /* if there's an unvisited sibling, visit its leftmost descendant */
4301 next = css_next_child(pos, pos->parent);
4302 if (next)
4303 return css_leftmost_descendant(next);
4304
4305 /* no sibling left, visit parent */
4306 return pos->parent;
4307}
4308
4309/**
4310 * css_has_online_children - does a css have online children
4311 * @css: the target css
4312 *
4313 * Returns %true if @css has any online children; otherwise, %false. This
4314 * function can be called from any context but the caller is responsible
4315 * for synchronizing against on/offlining as necessary.
4316 */
4317bool css_has_online_children(struct cgroup_subsys_state *css)
4318{
4319 struct cgroup_subsys_state *child;
4320 bool ret = false;
4321
4322 rcu_read_lock();
4323 css_for_each_child(child, css) {
4324 if (child->flags & CSS_ONLINE) {
4325 ret = true;
4326 break;
4327 }
4328 }
4329 rcu_read_unlock();
4330 return ret;
4331}
4332
4333static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4334{
4335 struct list_head *l;
4336 struct cgrp_cset_link *link;
4337 struct css_set *cset;
4338
4339 lockdep_assert_held(&css_set_lock);
4340
4341 /* find the next threaded cset */
4342 if (it->tcset_pos) {
4343 l = it->tcset_pos->next;
4344
4345 if (l != it->tcset_head) {
4346 it->tcset_pos = l;
4347 return container_of(l, struct css_set,
4348 threaded_csets_node);
4349 }
4350
4351 it->tcset_pos = NULL;
4352 }
4353
4354 /* find the next cset */
4355 l = it->cset_pos;
4356 l = l->next;
4357 if (l == it->cset_head) {
4358 it->cset_pos = NULL;
4359 return NULL;
4360 }
4361
4362 if (it->ss) {
4363 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4364 } else {
4365 link = list_entry(l, struct cgrp_cset_link, cset_link);
4366 cset = link->cset;
4367 }
4368
4369 it->cset_pos = l;
4370
4371 /* initialize threaded css_set walking */
4372 if (it->flags & CSS_TASK_ITER_THREADED) {
4373 if (it->cur_dcset)
4374 put_css_set_locked(it->cur_dcset);
4375 it->cur_dcset = cset;
4376 get_css_set(cset);
4377
4378 it->tcset_head = &cset->threaded_csets;
4379 it->tcset_pos = &cset->threaded_csets;
4380 }
4381
4382 return cset;
4383}
4384
4385/**
4386 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4387 * @it: the iterator to advance
4388 *
4389 * Advance @it to the next css_set to walk.
4390 */
4391static void css_task_iter_advance_css_set(struct css_task_iter *it)
4392{
4393 struct css_set *cset;
4394
4395 lockdep_assert_held(&css_set_lock);
4396
4397 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4398 while ((cset = css_task_iter_next_css_set(it))) {
4399 if (!list_empty(&cset->tasks)) {
4400 it->cur_tasks_head = &cset->tasks;
4401 break;
4402 } else if (!list_empty(&cset->mg_tasks)) {
4403 it->cur_tasks_head = &cset->mg_tasks;
4404 break;
4405 } else if (!list_empty(&cset->dying_tasks)) {
4406 it->cur_tasks_head = &cset->dying_tasks;
4407 break;
4408 }
4409 }
4410 if (!cset) {
4411 it->task_pos = NULL;
4412 return;
4413 }
4414 it->task_pos = it->cur_tasks_head->next;
4415
4416 /*
4417 * We don't keep css_sets locked across iteration steps and thus
4418 * need to take steps to ensure that iteration can be resumed after
4419 * the lock is re-acquired. Iteration is performed at two levels -
4420 * css_sets and tasks in them.
4421 *
4422 * Once created, a css_set never leaves its cgroup lists, so a
4423 * pinned css_set is guaranteed to stay put and we can resume
4424 * iteration afterwards.
4425 *
4426 * Tasks may leave @cset across iteration steps. This is resolved
4427 * by registering each iterator with the css_set currently being
4428 * walked and making css_set_move_task() advance iterators whose
4429 * next task is leaving.
4430 */
4431 if (it->cur_cset) {
4432 list_del(&it->iters_node);
4433 put_css_set_locked(it->cur_cset);
4434 }
4435 get_css_set(cset);
4436 it->cur_cset = cset;
4437 list_add(&it->iters_node, &cset->task_iters);
4438}
4439
4440static void css_task_iter_skip(struct css_task_iter *it,
4441 struct task_struct *task)
4442{
4443 lockdep_assert_held(&css_set_lock);
4444
4445 if (it->task_pos == &task->cg_list) {
4446 it->task_pos = it->task_pos->next;
4447 it->flags |= CSS_TASK_ITER_SKIPPED;
4448 }
4449}
4450
4451static void css_task_iter_advance(struct css_task_iter *it)
4452{
4453 struct task_struct *task;
4454
4455 lockdep_assert_held(&css_set_lock);
4456repeat:
4457 if (it->task_pos) {
4458 /*
4459 * Advance iterator to find next entry. We go through cset
4460 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4461 * the next cset.
4462 */
4463 if (it->flags & CSS_TASK_ITER_SKIPPED)
4464 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4465 else
4466 it->task_pos = it->task_pos->next;
4467
4468 if (it->task_pos == &it->cur_cset->tasks) {
4469 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4470 it->task_pos = it->cur_tasks_head->next;
4471 }
4472 if (it->task_pos == &it->cur_cset->mg_tasks) {
4473 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4474 it->task_pos = it->cur_tasks_head->next;
4475 }
4476 if (it->task_pos == &it->cur_cset->dying_tasks)
4477 css_task_iter_advance_css_set(it);
4478 } else {
4479 /* called from start, proceed to the first cset */
4480 css_task_iter_advance_css_set(it);
4481 }
4482
4483 if (!it->task_pos)
4484 return;
4485
4486 task = list_entry(it->task_pos, struct task_struct, cg_list);
4487
4488 if (it->flags & CSS_TASK_ITER_PROCS) {
4489 /* if PROCS, skip over tasks which aren't group leaders */
4490 if (!thread_group_leader(task))
4491 goto repeat;
4492
4493 /* and dying leaders w/o live member threads */
4494 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4495 !atomic_read(&task->signal->live))
4496 goto repeat;
4497 } else {
4498 /* skip all dying ones */
4499 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4500 goto repeat;
4501 }
4502}
4503
4504/**
4505 * css_task_iter_start - initiate task iteration
4506 * @css: the css to walk tasks of
4507 * @flags: CSS_TASK_ITER_* flags
4508 * @it: the task iterator to use
4509 *
4510 * Initiate iteration through the tasks of @css. The caller can call
4511 * css_task_iter_next() to walk through the tasks until the function
4512 * returns NULL. On completion of iteration, css_task_iter_end() must be
4513 * called.
4514 */
4515void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4516 struct css_task_iter *it)
4517{
4518 memset(it, 0, sizeof(*it));
4519
4520 spin_lock_irq(&css_set_lock);
4521
4522 it->ss = css->ss;
4523 it->flags = flags;
4524
4525 if (it->ss)
4526 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4527 else
4528 it->cset_pos = &css->cgroup->cset_links;
4529
4530 it->cset_head = it->cset_pos;
4531
4532 css_task_iter_advance(it);
4533
4534 spin_unlock_irq(&css_set_lock);
4535}
4536
4537/**
4538 * css_task_iter_next - return the next task for the iterator
4539 * @it: the task iterator being iterated
4540 *
4541 * The "next" function for task iteration. @it should have been
4542 * initialized via css_task_iter_start(). Returns NULL when the iteration
4543 * reaches the end.
4544 */
4545struct task_struct *css_task_iter_next(struct css_task_iter *it)
4546{
4547 if (it->cur_task) {
4548 put_task_struct(it->cur_task);
4549 it->cur_task = NULL;
4550 }
4551
4552 spin_lock_irq(&css_set_lock);
4553
4554 /* @it may be half-advanced by skips, finish advancing */
4555 if (it->flags & CSS_TASK_ITER_SKIPPED)
4556 css_task_iter_advance(it);
4557
4558 if (it->task_pos) {
4559 it->cur_task = list_entry(it->task_pos, struct task_struct,
4560 cg_list);
4561 get_task_struct(it->cur_task);
4562 css_task_iter_advance(it);
4563 }
4564
4565 spin_unlock_irq(&css_set_lock);
4566
4567 return it->cur_task;
4568}
4569
4570/**
4571 * css_task_iter_end - finish task iteration
4572 * @it: the task iterator to finish
4573 *
4574 * Finish task iteration started by css_task_iter_start().
4575 */
4576void css_task_iter_end(struct css_task_iter *it)
4577{
4578 if (it->cur_cset) {
4579 spin_lock_irq(&css_set_lock);
4580 list_del(&it->iters_node);
4581 put_css_set_locked(it->cur_cset);
4582 spin_unlock_irq(&css_set_lock);
4583 }
4584
4585 if (it->cur_dcset)
4586 put_css_set(it->cur_dcset);
4587
4588 if (it->cur_task)
4589 put_task_struct(it->cur_task);
4590}
4591
4592static void cgroup_procs_release(struct kernfs_open_file *of)
4593{
4594 if (of->priv) {
4595 css_task_iter_end(of->priv);
4596 kfree(of->priv);
4597 }
4598}
4599
4600static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4601{
4602 struct kernfs_open_file *of = s->private;
4603 struct css_task_iter *it = of->priv;
4604
4605 if (pos)
4606 (*pos)++;
4607
4608 return css_task_iter_next(it);
4609}
4610
4611static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4612 unsigned int iter_flags)
4613{
4614 struct kernfs_open_file *of = s->private;
4615 struct cgroup *cgrp = seq_css(s)->cgroup;
4616 struct css_task_iter *it = of->priv;
4617
4618 /*
4619 * When a seq_file is seeked, it's always traversed sequentially
4620 * from position 0, so we can simply keep iterating on !0 *pos.
4621 */
4622 if (!it) {
4623 if (WARN_ON_ONCE((*pos)))
4624 return ERR_PTR(-EINVAL);
4625
4626 it = kzalloc(sizeof(*it), GFP_KERNEL);
4627 if (!it)
4628 return ERR_PTR(-ENOMEM);
4629 of->priv = it;
4630 css_task_iter_start(&cgrp->self, iter_flags, it);
4631 } else if (!(*pos)) {
4632 css_task_iter_end(it);
4633 css_task_iter_start(&cgrp->self, iter_flags, it);
4634 } else
4635 return it->cur_task;
4636
4637 return cgroup_procs_next(s, NULL, NULL);
4638}
4639
4640static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4641{
4642 struct cgroup *cgrp = seq_css(s)->cgroup;
4643
4644 /*
4645 * All processes of a threaded subtree belong to the domain cgroup
4646 * of the subtree. Only threads can be distributed across the
4647 * subtree. Reject reads on cgroup.procs in the subtree proper.
4648 * They're always empty anyway.
4649 */
4650 if (cgroup_is_threaded(cgrp))
4651 return ERR_PTR(-EOPNOTSUPP);
4652
4653 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4654 CSS_TASK_ITER_THREADED);
4655}
4656
4657static int cgroup_procs_show(struct seq_file *s, void *v)
4658{
4659 seq_printf(s, "%d\n", task_pid_vnr(v));
4660 return 0;
4661}
4662
4663static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
4664{
4665 int ret;
4666 struct inode *inode;
4667
4668 lockdep_assert_held(&cgroup_mutex);
4669
4670 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
4671 if (!inode)
4672 return -ENOMEM;
4673
4674 ret = inode_permission(inode, MAY_WRITE);
4675 iput(inode);
4676 return ret;
4677}
4678
4679static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4680 struct cgroup *dst_cgrp,
4681 struct super_block *sb)
4682{
4683 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4684 struct cgroup *com_cgrp = src_cgrp;
4685 int ret;
4686
4687 lockdep_assert_held(&cgroup_mutex);
4688
4689 /* find the common ancestor */
4690 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4691 com_cgrp = cgroup_parent(com_cgrp);
4692
4693 /* %current should be authorized to migrate to the common ancestor */
4694 ret = cgroup_may_write(com_cgrp, sb);
4695 if (ret)
4696 return ret;
4697
4698 /*
4699 * If namespaces are delegation boundaries, %current must be able
4700 * to see both source and destination cgroups from its namespace.
4701 */
4702 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4703 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4704 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4705 return -ENOENT;
4706
4707 return 0;
4708}
4709
4710static int cgroup_attach_permissions(struct cgroup *src_cgrp,
4711 struct cgroup *dst_cgrp,
4712 struct super_block *sb, bool threadgroup)
4713{
4714 int ret = 0;
4715
4716 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb);
4717 if (ret)
4718 return ret;
4719
4720 ret = cgroup_migrate_vet_dst(dst_cgrp);
4721 if (ret)
4722 return ret;
4723
4724 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
4725 ret = -EOPNOTSUPP;
4726
4727 return ret;
4728}
4729
4730static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4731 char *buf, size_t nbytes, loff_t off)
4732{
4733 struct cgroup *src_cgrp, *dst_cgrp;
4734 struct task_struct *task;
4735 ssize_t ret;
4736 bool locked;
4737
4738 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4739 if (!dst_cgrp)
4740 return -ENODEV;
4741
4742 task = cgroup_procs_write_start(buf, true, &locked);
4743 ret = PTR_ERR_OR_ZERO(task);
4744 if (ret)
4745 goto out_unlock;
4746
4747 /* find the source cgroup */
4748 spin_lock_irq(&css_set_lock);
4749 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4750 spin_unlock_irq(&css_set_lock);
4751
4752 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
4753 of->file->f_path.dentry->d_sb, true);
4754 if (ret)
4755 goto out_finish;
4756
4757 ret = cgroup_attach_task(dst_cgrp, task, true);
4758
4759out_finish:
4760 cgroup_procs_write_finish(task, locked);
4761out_unlock:
4762 cgroup_kn_unlock(of->kn);
4763
4764 return ret ?: nbytes;
4765}
4766
4767static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4768{
4769 return __cgroup_procs_start(s, pos, 0);
4770}
4771
4772static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4773 char *buf, size_t nbytes, loff_t off)
4774{
4775 struct cgroup *src_cgrp, *dst_cgrp;
4776 struct task_struct *task;
4777 ssize_t ret;
4778 bool locked;
4779
4780 buf = strstrip(buf);
4781
4782 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4783 if (!dst_cgrp)
4784 return -ENODEV;
4785
4786 task = cgroup_procs_write_start(buf, false, &locked);
4787 ret = PTR_ERR_OR_ZERO(task);
4788 if (ret)
4789 goto out_unlock;
4790
4791 /* find the source cgroup */
4792 spin_lock_irq(&css_set_lock);
4793 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4794 spin_unlock_irq(&css_set_lock);
4795
4796 /* thread migrations follow the cgroup.procs delegation rule */
4797 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
4798 of->file->f_path.dentry->d_sb, false);
4799 if (ret)
4800 goto out_finish;
4801
4802 ret = cgroup_attach_task(dst_cgrp, task, false);
4803
4804out_finish:
4805 cgroup_procs_write_finish(task, locked);
4806out_unlock:
4807 cgroup_kn_unlock(of->kn);
4808
4809 return ret ?: nbytes;
4810}
4811
4812/* cgroup core interface files for the default hierarchy */
4813static struct cftype cgroup_base_files[] = {
4814 {
4815 .name = "cgroup.type",
4816 .flags = CFTYPE_NOT_ON_ROOT,
4817 .seq_show = cgroup_type_show,
4818 .write = cgroup_type_write,
4819 },
4820 {
4821 .name = "cgroup.procs",
4822 .flags = CFTYPE_NS_DELEGATABLE,
4823 .file_offset = offsetof(struct cgroup, procs_file),
4824 .release = cgroup_procs_release,
4825 .seq_start = cgroup_procs_start,
4826 .seq_next = cgroup_procs_next,
4827 .seq_show = cgroup_procs_show,
4828 .write = cgroup_procs_write,
4829 },
4830 {
4831 .name = "cgroup.threads",
4832 .flags = CFTYPE_NS_DELEGATABLE,
4833 .release = cgroup_procs_release,
4834 .seq_start = cgroup_threads_start,
4835 .seq_next = cgroup_procs_next,
4836 .seq_show = cgroup_procs_show,
4837 .write = cgroup_threads_write,
4838 },
4839 {
4840 .name = "cgroup.controllers",
4841 .seq_show = cgroup_controllers_show,
4842 },
4843 {
4844 .name = "cgroup.subtree_control",
4845 .flags = CFTYPE_NS_DELEGATABLE,
4846 .seq_show = cgroup_subtree_control_show,
4847 .write = cgroup_subtree_control_write,
4848 },
4849 {
4850 .name = "cgroup.events",
4851 .flags = CFTYPE_NOT_ON_ROOT,
4852 .file_offset = offsetof(struct cgroup, events_file),
4853 .seq_show = cgroup_events_show,
4854 },
4855 {
4856 .name = "cgroup.max.descendants",
4857 .seq_show = cgroup_max_descendants_show,
4858 .write = cgroup_max_descendants_write,
4859 },
4860 {
4861 .name = "cgroup.max.depth",
4862 .seq_show = cgroup_max_depth_show,
4863 .write = cgroup_max_depth_write,
4864 },
4865 {
4866 .name = "cgroup.stat",
4867 .seq_show = cgroup_stat_show,
4868 },
4869 {
4870 .name = "cgroup.freeze",
4871 .flags = CFTYPE_NOT_ON_ROOT,
4872 .seq_show = cgroup_freeze_show,
4873 .write = cgroup_freeze_write,
4874 },
4875 {
4876 .name = "cpu.stat",
4877 .seq_show = cpu_stat_show,
4878 },
4879#ifdef CONFIG_PSI
4880 {
4881 .name = "io.pressure",
4882 .seq_show = cgroup_io_pressure_show,
4883 .write = cgroup_io_pressure_write,
4884 .poll = cgroup_pressure_poll,
4885 .release = cgroup_pressure_release,
4886 },
4887 {
4888 .name = "memory.pressure",
4889 .seq_show = cgroup_memory_pressure_show,
4890 .write = cgroup_memory_pressure_write,
4891 .poll = cgroup_pressure_poll,
4892 .release = cgroup_pressure_release,
4893 },
4894 {
4895 .name = "cpu.pressure",
4896 .seq_show = cgroup_cpu_pressure_show,
4897 .write = cgroup_cpu_pressure_write,
4898 .poll = cgroup_pressure_poll,
4899 .release = cgroup_pressure_release,
4900 },
4901#endif /* CONFIG_PSI */
4902 { } /* terminate */
4903};
4904
4905/*
4906 * css destruction is four-stage process.
4907 *
4908 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4909 * Implemented in kill_css().
4910 *
4911 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4912 * and thus css_tryget_online() is guaranteed to fail, the css can be
4913 * offlined by invoking offline_css(). After offlining, the base ref is
4914 * put. Implemented in css_killed_work_fn().
4915 *
4916 * 3. When the percpu_ref reaches zero, the only possible remaining
4917 * accessors are inside RCU read sections. css_release() schedules the
4918 * RCU callback.
4919 *
4920 * 4. After the grace period, the css can be freed. Implemented in
4921 * css_free_work_fn().
4922 *
4923 * It is actually hairier because both step 2 and 4 require process context
4924 * and thus involve punting to css->destroy_work adding two additional
4925 * steps to the already complex sequence.
4926 */
4927static void css_free_rwork_fn(struct work_struct *work)
4928{
4929 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4930 struct cgroup_subsys_state, destroy_rwork);
4931 struct cgroup_subsys *ss = css->ss;
4932 struct cgroup *cgrp = css->cgroup;
4933
4934 percpu_ref_exit(&css->refcnt);
4935
4936 if (ss) {
4937 /* css free path */
4938 struct cgroup_subsys_state *parent = css->parent;
4939 int id = css->id;
4940
4941 ss->css_free(css);
4942 cgroup_idr_remove(&ss->css_idr, id);
4943 cgroup_put(cgrp);
4944
4945 if (parent)
4946 css_put(parent);
4947 } else {
4948 /* cgroup free path */
4949 atomic_dec(&cgrp->root->nr_cgrps);
4950 cgroup1_pidlist_destroy_all(cgrp);
4951 cancel_work_sync(&cgrp->release_agent_work);
4952
4953 if (cgroup_parent(cgrp)) {
4954 /*
4955 * We get a ref to the parent, and put the ref when
4956 * this cgroup is being freed, so it's guaranteed
4957 * that the parent won't be destroyed before its
4958 * children.
4959 */
4960 cgroup_put(cgroup_parent(cgrp));
4961 kernfs_put(cgrp->kn);
4962 psi_cgroup_free(cgrp);
4963 if (cgroup_on_dfl(cgrp))
4964 cgroup_rstat_exit(cgrp);
4965 kfree(cgrp);
4966 } else {
4967 /*
4968 * This is root cgroup's refcnt reaching zero,
4969 * which indicates that the root should be
4970 * released.
4971 */
4972 cgroup_destroy_root(cgrp->root);
4973 }
4974 }
4975}
4976
4977static void css_release_work_fn(struct work_struct *work)
4978{
4979 struct cgroup_subsys_state *css =
4980 container_of(work, struct cgroup_subsys_state, destroy_work);
4981 struct cgroup_subsys *ss = css->ss;
4982 struct cgroup *cgrp = css->cgroup;
4983
4984 mutex_lock(&cgroup_mutex);
4985
4986 css->flags |= CSS_RELEASED;
4987 list_del_rcu(&css->sibling);
4988
4989 if (ss) {
4990 /* css release path */
4991 if (!list_empty(&css->rstat_css_node)) {
4992 cgroup_rstat_flush(cgrp);
4993 list_del_rcu(&css->rstat_css_node);
4994 }
4995
4996 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4997 if (ss->css_released)
4998 ss->css_released(css);
4999 } else {
5000 struct cgroup *tcgrp;
5001
5002 /* cgroup release path */
5003 TRACE_CGROUP_PATH(release, cgrp);
5004
5005 if (cgroup_on_dfl(cgrp))
5006 cgroup_rstat_flush(cgrp);
5007
5008 spin_lock_irq(&css_set_lock);
5009 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5010 tcgrp = cgroup_parent(tcgrp))
5011 tcgrp->nr_dying_descendants--;
5012 spin_unlock_irq(&css_set_lock);
5013
5014 /*
5015 * There are two control paths which try to determine
5016 * cgroup from dentry without going through kernfs -
5017 * cgroupstats_build() and css_tryget_online_from_dir().
5018 * Those are supported by RCU protecting clearing of
5019 * cgrp->kn->priv backpointer.
5020 */
5021 if (cgrp->kn)
5022 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5023 NULL);
5024 }
5025
5026 mutex_unlock(&cgroup_mutex);
5027
5028 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5029 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5030}
5031
5032static void css_release(struct percpu_ref *ref)
5033{
5034 struct cgroup_subsys_state *css =
5035 container_of(ref, struct cgroup_subsys_state, refcnt);
5036
5037 INIT_WORK(&css->destroy_work, css_release_work_fn);
5038 queue_work(cgroup_destroy_wq, &css->destroy_work);
5039}
5040
5041static void init_and_link_css(struct cgroup_subsys_state *css,
5042 struct cgroup_subsys *ss, struct cgroup *cgrp)
5043{
5044 lockdep_assert_held(&cgroup_mutex);
5045
5046 cgroup_get_live(cgrp);
5047
5048 memset(css, 0, sizeof(*css));
5049 css->cgroup = cgrp;
5050 css->ss = ss;
5051 css->id = -1;
5052 INIT_LIST_HEAD(&css->sibling);
5053 INIT_LIST_HEAD(&css->children);
5054 INIT_LIST_HEAD(&css->rstat_css_node);
5055 css->serial_nr = css_serial_nr_next++;
5056 atomic_set(&css->online_cnt, 0);
5057
5058 if (cgroup_parent(cgrp)) {
5059 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5060 css_get(css->parent);
5061 }
5062
5063 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush)
5064 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5065
5066 BUG_ON(cgroup_css(cgrp, ss));
5067}
5068
5069/* invoke ->css_online() on a new CSS and mark it online if successful */
5070static int online_css(struct cgroup_subsys_state *css)
5071{
5072 struct cgroup_subsys *ss = css->ss;
5073 int ret = 0;
5074
5075 lockdep_assert_held(&cgroup_mutex);
5076
5077 if (ss->css_online)
5078 ret = ss->css_online(css);
5079 if (!ret) {
5080 css->flags |= CSS_ONLINE;
5081 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5082
5083 atomic_inc(&css->online_cnt);
5084 if (css->parent)
5085 atomic_inc(&css->parent->online_cnt);
5086 }
5087 return ret;
5088}
5089
5090/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5091static void offline_css(struct cgroup_subsys_state *css)
5092{
5093 struct cgroup_subsys *ss = css->ss;
5094
5095 lockdep_assert_held(&cgroup_mutex);
5096
5097 if (!(css->flags & CSS_ONLINE))
5098 return;
5099
5100 if (ss->css_offline)
5101 ss->css_offline(css);
5102
5103 css->flags &= ~CSS_ONLINE;
5104 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5105
5106 wake_up_all(&css->cgroup->offline_waitq);
5107}
5108
5109/**
5110 * css_create - create a cgroup_subsys_state
5111 * @cgrp: the cgroup new css will be associated with
5112 * @ss: the subsys of new css
5113 *
5114 * Create a new css associated with @cgrp - @ss pair. On success, the new
5115 * css is online and installed in @cgrp. This function doesn't create the
5116 * interface files. Returns 0 on success, -errno on failure.
5117 */
5118static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5119 struct cgroup_subsys *ss)
5120{
5121 struct cgroup *parent = cgroup_parent(cgrp);
5122 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5123 struct cgroup_subsys_state *css;
5124 int err;
5125
5126 lockdep_assert_held(&cgroup_mutex);
5127
5128 css = ss->css_alloc(parent_css);
5129 if (!css)
5130 css = ERR_PTR(-ENOMEM);
5131 if (IS_ERR(css))
5132 return css;
5133
5134 init_and_link_css(css, ss, cgrp);
5135
5136 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5137 if (err)
5138 goto err_free_css;
5139
5140 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5141 if (err < 0)
5142 goto err_free_css;
5143 css->id = err;
5144
5145 /* @css is ready to be brought online now, make it visible */
5146 list_add_tail_rcu(&css->sibling, &parent_css->children);
5147 cgroup_idr_replace(&ss->css_idr, css, css->id);
5148
5149 err = online_css(css);
5150 if (err)
5151 goto err_list_del;
5152
5153 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5154 cgroup_parent(parent)) {
5155 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5156 current->comm, current->pid, ss->name);
5157 if (!strcmp(ss->name, "memory"))
5158 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5159 ss->warned_broken_hierarchy = true;
5160 }
5161
5162 return css;
5163
5164err_list_del:
5165 list_del_rcu(&css->sibling);
5166err_free_css:
5167 list_del_rcu(&css->rstat_css_node);
5168 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5169 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5170 return ERR_PTR(err);
5171}
5172
5173/*
5174 * The returned cgroup is fully initialized including its control mask, but
5175 * it isn't associated with its kernfs_node and doesn't have the control
5176 * mask applied.
5177 */
5178static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5179 umode_t mode)
5180{
5181 struct cgroup_root *root = parent->root;
5182 struct cgroup *cgrp, *tcgrp;
5183 struct kernfs_node *kn;
5184 int level = parent->level + 1;
5185 int ret;
5186
5187 /* allocate the cgroup and its ID, 0 is reserved for the root */
5188 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5189 GFP_KERNEL);
5190 if (!cgrp)
5191 return ERR_PTR(-ENOMEM);
5192
5193 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5194 if (ret)
5195 goto out_free_cgrp;
5196
5197 if (cgroup_on_dfl(parent)) {
5198 ret = cgroup_rstat_init(cgrp);
5199 if (ret)
5200 goto out_cancel_ref;
5201 }
5202
5203 /* create the directory */
5204 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5205 if (IS_ERR(kn)) {
5206 ret = PTR_ERR(kn);
5207 goto out_stat_exit;
5208 }
5209 cgrp->kn = kn;
5210
5211 init_cgroup_housekeeping(cgrp);
5212
5213 cgrp->self.parent = &parent->self;
5214 cgrp->root = root;
5215 cgrp->level = level;
5216
5217 ret = psi_cgroup_alloc(cgrp);
5218 if (ret)
5219 goto out_kernfs_remove;
5220
5221 ret = cgroup_bpf_inherit(cgrp);
5222 if (ret)
5223 goto out_psi_free;
5224
5225 /*
5226 * New cgroup inherits effective freeze counter, and
5227 * if the parent has to be frozen, the child has too.
5228 */
5229 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5230 if (cgrp->freezer.e_freeze) {
5231 /*
5232 * Set the CGRP_FREEZE flag, so when a process will be
5233 * attached to the child cgroup, it will become frozen.
5234 * At this point the new cgroup is unpopulated, so we can
5235 * consider it frozen immediately.
5236 */
5237 set_bit(CGRP_FREEZE, &cgrp->flags);
5238 set_bit(CGRP_FROZEN, &cgrp->flags);
5239 }
5240
5241 spin_lock_irq(&css_set_lock);
5242 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5243 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp);
5244
5245 if (tcgrp != cgrp) {
5246 tcgrp->nr_descendants++;
5247
5248 /*
5249 * If the new cgroup is frozen, all ancestor cgroups
5250 * get a new frozen descendant, but their state can't
5251 * change because of this.
5252 */
5253 if (cgrp->freezer.e_freeze)
5254 tcgrp->freezer.nr_frozen_descendants++;
5255 }
5256 }
5257 spin_unlock_irq(&css_set_lock);
5258
5259 if (notify_on_release(parent))
5260 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5261
5262 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5263 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5264
5265 cgrp->self.serial_nr = css_serial_nr_next++;
5266
5267 /* allocation complete, commit to creation */
5268 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5269 atomic_inc(&root->nr_cgrps);
5270 cgroup_get_live(parent);
5271
5272 /*
5273 * On the default hierarchy, a child doesn't automatically inherit
5274 * subtree_control from the parent. Each is configured manually.
5275 */
5276 if (!cgroup_on_dfl(cgrp))
5277 cgrp->subtree_control = cgroup_control(cgrp);
5278
5279 cgroup_propagate_control(cgrp);
5280
5281 return cgrp;
5282
5283out_psi_free:
5284 psi_cgroup_free(cgrp);
5285out_kernfs_remove:
5286 kernfs_remove(cgrp->kn);
5287out_stat_exit:
5288 if (cgroup_on_dfl(parent))
5289 cgroup_rstat_exit(cgrp);
5290out_cancel_ref:
5291 percpu_ref_exit(&cgrp->self.refcnt);
5292out_free_cgrp:
5293 kfree(cgrp);
5294 return ERR_PTR(ret);
5295}
5296
5297static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5298{
5299 struct cgroup *cgroup;
5300 int ret = false;
5301 int level = 1;
5302
5303 lockdep_assert_held(&cgroup_mutex);
5304
5305 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5306 if (cgroup->nr_descendants >= cgroup->max_descendants)
5307 goto fail;
5308
5309 if (level > cgroup->max_depth)
5310 goto fail;
5311
5312 level++;
5313 }
5314
5315 ret = true;
5316fail:
5317 return ret;
5318}
5319
5320int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5321{
5322 struct cgroup *parent, *cgrp;
5323 int ret;
5324
5325 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5326 if (strchr(name, '\n'))
5327 return -EINVAL;
5328
5329 parent = cgroup_kn_lock_live(parent_kn, false);
5330 if (!parent)
5331 return -ENODEV;
5332
5333 if (!cgroup_check_hierarchy_limits(parent)) {
5334 ret = -EAGAIN;
5335 goto out_unlock;
5336 }
5337
5338 cgrp = cgroup_create(parent, name, mode);
5339 if (IS_ERR(cgrp)) {
5340 ret = PTR_ERR(cgrp);
5341 goto out_unlock;
5342 }
5343
5344 /*
5345 * This extra ref will be put in cgroup_free_fn() and guarantees
5346 * that @cgrp->kn is always accessible.
5347 */
5348 kernfs_get(cgrp->kn);
5349
5350 ret = cgroup_kn_set_ugid(cgrp->kn);
5351 if (ret)
5352 goto out_destroy;
5353
5354 ret = css_populate_dir(&cgrp->self);
5355 if (ret)
5356 goto out_destroy;
5357
5358 ret = cgroup_apply_control_enable(cgrp);
5359 if (ret)
5360 goto out_destroy;
5361
5362 TRACE_CGROUP_PATH(mkdir, cgrp);
5363
5364 /* let's create and online css's */
5365 kernfs_activate(cgrp->kn);
5366
5367 ret = 0;
5368 goto out_unlock;
5369
5370out_destroy:
5371 cgroup_destroy_locked(cgrp);
5372out_unlock:
5373 cgroup_kn_unlock(parent_kn);
5374 return ret;
5375}
5376
5377/*
5378 * This is called when the refcnt of a css is confirmed to be killed.
5379 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5380 * initate destruction and put the css ref from kill_css().
5381 */
5382static void css_killed_work_fn(struct work_struct *work)
5383{
5384 struct cgroup_subsys_state *css =
5385 container_of(work, struct cgroup_subsys_state, destroy_work);
5386
5387 mutex_lock(&cgroup_mutex);
5388
5389 do {
5390 offline_css(css);
5391 css_put(css);
5392 /* @css can't go away while we're holding cgroup_mutex */
5393 css = css->parent;
5394 } while (css && atomic_dec_and_test(&css->online_cnt));
5395
5396 mutex_unlock(&cgroup_mutex);
5397}
5398
5399/* css kill confirmation processing requires process context, bounce */
5400static void css_killed_ref_fn(struct percpu_ref *ref)
5401{
5402 struct cgroup_subsys_state *css =
5403 container_of(ref, struct cgroup_subsys_state, refcnt);
5404
5405 if (atomic_dec_and_test(&css->online_cnt)) {
5406 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5407 queue_work(cgroup_destroy_wq, &css->destroy_work);
5408 }
5409}
5410
5411/**
5412 * kill_css - destroy a css
5413 * @css: css to destroy
5414 *
5415 * This function initiates destruction of @css by removing cgroup interface
5416 * files and putting its base reference. ->css_offline() will be invoked
5417 * asynchronously once css_tryget_online() is guaranteed to fail and when
5418 * the reference count reaches zero, @css will be released.
5419 */
5420static void kill_css(struct cgroup_subsys_state *css)
5421{
5422 lockdep_assert_held(&cgroup_mutex);
5423
5424 if (css->flags & CSS_DYING)
5425 return;
5426
5427 css->flags |= CSS_DYING;
5428
5429 /*
5430 * This must happen before css is disassociated with its cgroup.
5431 * See seq_css() for details.
5432 */
5433 css_clear_dir(css);
5434
5435 /*
5436 * Killing would put the base ref, but we need to keep it alive
5437 * until after ->css_offline().
5438 */
5439 css_get(css);
5440
5441 /*
5442 * cgroup core guarantees that, by the time ->css_offline() is
5443 * invoked, no new css reference will be given out via
5444 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5445 * proceed to offlining css's because percpu_ref_kill() doesn't
5446 * guarantee that the ref is seen as killed on all CPUs on return.
5447 *
5448 * Use percpu_ref_kill_and_confirm() to get notifications as each
5449 * css is confirmed to be seen as killed on all CPUs.
5450 */
5451 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5452}
5453
5454/**
5455 * cgroup_destroy_locked - the first stage of cgroup destruction
5456 * @cgrp: cgroup to be destroyed
5457 *
5458 * css's make use of percpu refcnts whose killing latency shouldn't be
5459 * exposed to userland and are RCU protected. Also, cgroup core needs to
5460 * guarantee that css_tryget_online() won't succeed by the time
5461 * ->css_offline() is invoked. To satisfy all the requirements,
5462 * destruction is implemented in the following two steps.
5463 *
5464 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5465 * userland visible parts and start killing the percpu refcnts of
5466 * css's. Set up so that the next stage will be kicked off once all
5467 * the percpu refcnts are confirmed to be killed.
5468 *
5469 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5470 * rest of destruction. Once all cgroup references are gone, the
5471 * cgroup is RCU-freed.
5472 *
5473 * This function implements s1. After this step, @cgrp is gone as far as
5474 * the userland is concerned and a new cgroup with the same name may be
5475 * created. As cgroup doesn't care about the names internally, this
5476 * doesn't cause any problem.
5477 */
5478static int cgroup_destroy_locked(struct cgroup *cgrp)
5479 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5480{
5481 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5482 struct cgroup_subsys_state *css;
5483 struct cgrp_cset_link *link;
5484 int ssid;
5485
5486 lockdep_assert_held(&cgroup_mutex);
5487
5488 /*
5489 * Only migration can raise populated from zero and we're already
5490 * holding cgroup_mutex.
5491 */
5492 if (cgroup_is_populated(cgrp))
5493 return -EBUSY;
5494
5495 /*
5496 * Make sure there's no live children. We can't test emptiness of
5497 * ->self.children as dead children linger on it while being
5498 * drained; otherwise, "rmdir parent/child parent" may fail.
5499 */
5500 if (css_has_online_children(&cgrp->self))
5501 return -EBUSY;
5502
5503 /*
5504 * Mark @cgrp and the associated csets dead. The former prevents
5505 * further task migration and child creation by disabling
5506 * cgroup_lock_live_group(). The latter makes the csets ignored by
5507 * the migration path.
5508 */
5509 cgrp->self.flags &= ~CSS_ONLINE;
5510
5511 spin_lock_irq(&css_set_lock);
5512 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5513 link->cset->dead = true;
5514 spin_unlock_irq(&css_set_lock);
5515
5516 /* initiate massacre of all css's */
5517 for_each_css(css, ssid, cgrp)
5518 kill_css(css);
5519
5520 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5521 css_clear_dir(&cgrp->self);
5522 kernfs_remove(cgrp->kn);
5523
5524 if (parent && cgroup_is_threaded(cgrp))
5525 parent->nr_threaded_children--;
5526
5527 spin_lock_irq(&css_set_lock);
5528 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5529 tcgrp->nr_descendants--;
5530 tcgrp->nr_dying_descendants++;
5531 /*
5532 * If the dying cgroup is frozen, decrease frozen descendants
5533 * counters of ancestor cgroups.
5534 */
5535 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5536 tcgrp->freezer.nr_frozen_descendants--;
5537 }
5538 spin_unlock_irq(&css_set_lock);
5539
5540 cgroup1_check_for_release(parent);
5541
5542 cgroup_bpf_offline(cgrp);
5543
5544 /* put the base reference */
5545 percpu_ref_kill(&cgrp->self.refcnt);
5546
5547 return 0;
5548};
5549
5550int cgroup_rmdir(struct kernfs_node *kn)
5551{
5552 struct cgroup *cgrp;
5553 int ret = 0;
5554
5555 cgrp = cgroup_kn_lock_live(kn, false);
5556 if (!cgrp)
5557 return 0;
5558
5559 ret = cgroup_destroy_locked(cgrp);
5560 if (!ret)
5561 TRACE_CGROUP_PATH(rmdir, cgrp);
5562
5563 cgroup_kn_unlock(kn);
5564 return ret;
5565}
5566
5567static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5568 .show_options = cgroup_show_options,
5569 .mkdir = cgroup_mkdir,
5570 .rmdir = cgroup_rmdir,
5571 .show_path = cgroup_show_path,
5572};
5573
5574static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5575{
5576 struct cgroup_subsys_state *css;
5577
5578 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5579
5580 mutex_lock(&cgroup_mutex);
5581
5582 idr_init(&ss->css_idr);
5583 INIT_LIST_HEAD(&ss->cfts);
5584
5585 /* Create the root cgroup state for this subsystem */
5586 ss->root = &cgrp_dfl_root;
5587 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5588 /* We don't handle early failures gracefully */
5589 BUG_ON(IS_ERR(css));
5590 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5591
5592 /*
5593 * Root csses are never destroyed and we can't initialize
5594 * percpu_ref during early init. Disable refcnting.
5595 */
5596 css->flags |= CSS_NO_REF;
5597
5598 if (early) {
5599 /* allocation can't be done safely during early init */
5600 css->id = 1;
5601 } else {
5602 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5603 BUG_ON(css->id < 0);
5604 }
5605
5606 /* Update the init_css_set to contain a subsys
5607 * pointer to this state - since the subsystem is
5608 * newly registered, all tasks and hence the
5609 * init_css_set is in the subsystem's root cgroup. */
5610 init_css_set.subsys[ss->id] = css;
5611
5612 have_fork_callback |= (bool)ss->fork << ss->id;
5613 have_exit_callback |= (bool)ss->exit << ss->id;
5614 have_release_callback |= (bool)ss->release << ss->id;
5615 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5616
5617 /* At system boot, before all subsystems have been
5618 * registered, no tasks have been forked, so we don't
5619 * need to invoke fork callbacks here. */
5620 BUG_ON(!list_empty(&init_task.tasks));
5621
5622 BUG_ON(online_css(css));
5623
5624 mutex_unlock(&cgroup_mutex);
5625}
5626
5627/**
5628 * cgroup_init_early - cgroup initialization at system boot
5629 *
5630 * Initialize cgroups at system boot, and initialize any
5631 * subsystems that request early init.
5632 */
5633int __init cgroup_init_early(void)
5634{
5635 static struct cgroup_fs_context __initdata ctx;
5636 struct cgroup_subsys *ss;
5637 int i;
5638
5639 ctx.root = &cgrp_dfl_root;
5640 init_cgroup_root(&ctx);
5641 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5642
5643 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5644
5645 for_each_subsys(ss, i) {
5646 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5647 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5648 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5649 ss->id, ss->name);
5650 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5651 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5652
5653 ss->id = i;
5654 ss->name = cgroup_subsys_name[i];
5655 if (!ss->legacy_name)
5656 ss->legacy_name = cgroup_subsys_name[i];
5657
5658 if (ss->early_init)
5659 cgroup_init_subsys(ss, true);
5660 }
5661 return 0;
5662}
5663
5664static u16 cgroup_disable_mask __initdata;
5665
5666/**
5667 * cgroup_init - cgroup initialization
5668 *
5669 * Register cgroup filesystem and /proc file, and initialize
5670 * any subsystems that didn't request early init.
5671 */
5672int __init cgroup_init(void)
5673{
5674 struct cgroup_subsys *ss;
5675 int ssid;
5676
5677 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5678 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5679 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5680
5681 cgroup_rstat_boot();
5682
5683 /*
5684 * The latency of the synchronize_rcu() is too high for cgroups,
5685 * avoid it at the cost of forcing all readers into the slow path.
5686 */
5687 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5688
5689 get_user_ns(init_cgroup_ns.user_ns);
5690
5691 mutex_lock(&cgroup_mutex);
5692
5693 /*
5694 * Add init_css_set to the hash table so that dfl_root can link to
5695 * it during init.
5696 */
5697 hash_add(css_set_table, &init_css_set.hlist,
5698 css_set_hash(init_css_set.subsys));
5699
5700 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5701
5702 mutex_unlock(&cgroup_mutex);
5703
5704 for_each_subsys(ss, ssid) {
5705 if (ss->early_init) {
5706 struct cgroup_subsys_state *css =
5707 init_css_set.subsys[ss->id];
5708
5709 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5710 GFP_KERNEL);
5711 BUG_ON(css->id < 0);
5712 } else {
5713 cgroup_init_subsys(ss, false);
5714 }
5715
5716 list_add_tail(&init_css_set.e_cset_node[ssid],
5717 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5718
5719 /*
5720 * Setting dfl_root subsys_mask needs to consider the
5721 * disabled flag and cftype registration needs kmalloc,
5722 * both of which aren't available during early_init.
5723 */
5724 if (cgroup_disable_mask & (1 << ssid)) {
5725 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5726 printk(KERN_INFO "Disabling %s control group subsystem\n",
5727 ss->name);
5728 continue;
5729 }
5730
5731 if (cgroup1_ssid_disabled(ssid))
5732 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5733 ss->name);
5734
5735 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5736
5737 /* implicit controllers must be threaded too */
5738 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5739
5740 if (ss->implicit_on_dfl)
5741 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5742 else if (!ss->dfl_cftypes)
5743 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5744
5745 if (ss->threaded)
5746 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5747
5748 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5749 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5750 } else {
5751 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5752 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5753 }
5754
5755 if (ss->bind)
5756 ss->bind(init_css_set.subsys[ssid]);
5757
5758 mutex_lock(&cgroup_mutex);
5759 css_populate_dir(init_css_set.subsys[ssid]);
5760 mutex_unlock(&cgroup_mutex);
5761 }
5762
5763 /* init_css_set.subsys[] has been updated, re-hash */
5764 hash_del(&init_css_set.hlist);
5765 hash_add(css_set_table, &init_css_set.hlist,
5766 css_set_hash(init_css_set.subsys));
5767
5768 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5769 WARN_ON(register_filesystem(&cgroup_fs_type));
5770 WARN_ON(register_filesystem(&cgroup2_fs_type));
5771 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5772#ifdef CONFIG_CPUSETS
5773 WARN_ON(register_filesystem(&cpuset_fs_type));
5774#endif
5775
5776 return 0;
5777}
5778
5779static int __init cgroup_wq_init(void)
5780{
5781 /*
5782 * There isn't much point in executing destruction path in
5783 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5784 * Use 1 for @max_active.
5785 *
5786 * We would prefer to do this in cgroup_init() above, but that
5787 * is called before init_workqueues(): so leave this until after.
5788 */
5789 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5790 BUG_ON(!cgroup_destroy_wq);
5791 return 0;
5792}
5793core_initcall(cgroup_wq_init);
5794
5795void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
5796{
5797 struct kernfs_node *kn;
5798
5799 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
5800 if (!kn)
5801 return;
5802 kernfs_path(kn, buf, buflen);
5803 kernfs_put(kn);
5804}
5805
5806/*
5807 * proc_cgroup_show()
5808 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5809 * - Used for /proc/<pid>/cgroup.
5810 */
5811int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5812 struct pid *pid, struct task_struct *tsk)
5813{
5814 char *buf;
5815 int retval;
5816 struct cgroup_root *root;
5817
5818 retval = -ENOMEM;
5819 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5820 if (!buf)
5821 goto out;
5822
5823 mutex_lock(&cgroup_mutex);
5824 spin_lock_irq(&css_set_lock);
5825
5826 for_each_root(root) {
5827 struct cgroup_subsys *ss;
5828 struct cgroup *cgrp;
5829 int ssid, count = 0;
5830
5831 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5832 continue;
5833
5834 seq_printf(m, "%d:", root->hierarchy_id);
5835 if (root != &cgrp_dfl_root)
5836 for_each_subsys(ss, ssid)
5837 if (root->subsys_mask & (1 << ssid))
5838 seq_printf(m, "%s%s", count++ ? "," : "",
5839 ss->legacy_name);
5840 if (strlen(root->name))
5841 seq_printf(m, "%sname=%s", count ? "," : "",
5842 root->name);
5843 seq_putc(m, ':');
5844
5845 cgrp = task_cgroup_from_root(tsk, root);
5846
5847 /*
5848 * On traditional hierarchies, all zombie tasks show up as
5849 * belonging to the root cgroup. On the default hierarchy,
5850 * while a zombie doesn't show up in "cgroup.procs" and
5851 * thus can't be migrated, its /proc/PID/cgroup keeps
5852 * reporting the cgroup it belonged to before exiting. If
5853 * the cgroup is removed before the zombie is reaped,
5854 * " (deleted)" is appended to the cgroup path.
5855 */
5856 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5857 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5858 current->nsproxy->cgroup_ns);
5859 if (retval >= PATH_MAX)
5860 retval = -ENAMETOOLONG;
5861 if (retval < 0)
5862 goto out_unlock;
5863
5864 seq_puts(m, buf);
5865 } else {
5866 seq_puts(m, "/");
5867 }
5868
5869 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5870 seq_puts(m, " (deleted)\n");
5871 else
5872 seq_putc(m, '\n');
5873 }
5874
5875 retval = 0;
5876out_unlock:
5877 spin_unlock_irq(&css_set_lock);
5878 mutex_unlock(&cgroup_mutex);
5879 kfree(buf);
5880out:
5881 return retval;
5882}
5883
5884/**
5885 * cgroup_fork - initialize cgroup related fields during copy_process()
5886 * @child: pointer to task_struct of forking parent process.
5887 *
5888 * A task is associated with the init_css_set until cgroup_post_fork()
5889 * attaches it to the target css_set.
5890 */
5891void cgroup_fork(struct task_struct *child)
5892{
5893 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5894 INIT_LIST_HEAD(&child->cg_list);
5895}
5896
5897static struct cgroup *cgroup_get_from_file(struct file *f)
5898{
5899 struct cgroup_subsys_state *css;
5900 struct cgroup *cgrp;
5901
5902 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5903 if (IS_ERR(css))
5904 return ERR_CAST(css);
5905
5906 cgrp = css->cgroup;
5907 if (!cgroup_on_dfl(cgrp)) {
5908 cgroup_put(cgrp);
5909 return ERR_PTR(-EBADF);
5910 }
5911
5912 return cgrp;
5913}
5914
5915/**
5916 * cgroup_css_set_fork - find or create a css_set for a child process
5917 * @kargs: the arguments passed to create the child process
5918 *
5919 * This functions finds or creates a new css_set which the child
5920 * process will be attached to in cgroup_post_fork(). By default,
5921 * the child process will be given the same css_set as its parent.
5922 *
5923 * If CLONE_INTO_CGROUP is specified this function will try to find an
5924 * existing css_set which includes the requested cgroup and if not create
5925 * a new css_set that the child will be attached to later. If this function
5926 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
5927 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
5928 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
5929 * to the target cgroup.
5930 */
5931static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
5932 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
5933{
5934 int ret;
5935 struct cgroup *dst_cgrp = NULL;
5936 struct css_set *cset;
5937 struct super_block *sb;
5938 struct file *f;
5939
5940 if (kargs->flags & CLONE_INTO_CGROUP)
5941 mutex_lock(&cgroup_mutex);
5942
5943 cgroup_threadgroup_change_begin(current);
5944
5945 spin_lock_irq(&css_set_lock);
5946 cset = task_css_set(current);
5947 get_css_set(cset);
5948 spin_unlock_irq(&css_set_lock);
5949
5950 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
5951 kargs->cset = cset;
5952 return 0;
5953 }
5954
5955 f = fget_raw(kargs->cgroup);
5956 if (!f) {
5957 ret = -EBADF;
5958 goto err;
5959 }
5960 sb = f->f_path.dentry->d_sb;
5961
5962 dst_cgrp = cgroup_get_from_file(f);
5963 if (IS_ERR(dst_cgrp)) {
5964 ret = PTR_ERR(dst_cgrp);
5965 dst_cgrp = NULL;
5966 goto err;
5967 }
5968
5969 if (cgroup_is_dead(dst_cgrp)) {
5970 ret = -ENODEV;
5971 goto err;
5972 }
5973
5974 /*
5975 * Verify that we the target cgroup is writable for us. This is
5976 * usually done by the vfs layer but since we're not going through
5977 * the vfs layer here we need to do it "manually".
5978 */
5979 ret = cgroup_may_write(dst_cgrp, sb);
5980 if (ret)
5981 goto err;
5982
5983 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
5984 !(kargs->flags & CLONE_THREAD));
5985 if (ret)
5986 goto err;
5987
5988 kargs->cset = find_css_set(cset, dst_cgrp);
5989 if (!kargs->cset) {
5990 ret = -ENOMEM;
5991 goto err;
5992 }
5993
5994 put_css_set(cset);
5995 fput(f);
5996 kargs->cgrp = dst_cgrp;
5997 return ret;
5998
5999err:
6000 cgroup_threadgroup_change_end(current);
6001 mutex_unlock(&cgroup_mutex);
6002 if (f)
6003 fput(f);
6004 if (dst_cgrp)
6005 cgroup_put(dst_cgrp);
6006 put_css_set(cset);
6007 if (kargs->cset)
6008 put_css_set(kargs->cset);
6009 return ret;
6010}
6011
6012/**
6013 * cgroup_css_set_put_fork - drop references we took during fork
6014 * @kargs: the arguments passed to create the child process
6015 *
6016 * Drop references to the prepared css_set and target cgroup if
6017 * CLONE_INTO_CGROUP was requested.
6018 */
6019static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6020 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6021{
6022 cgroup_threadgroup_change_end(current);
6023
6024 if (kargs->flags & CLONE_INTO_CGROUP) {
6025 struct cgroup *cgrp = kargs->cgrp;
6026 struct css_set *cset = kargs->cset;
6027
6028 mutex_unlock(&cgroup_mutex);
6029
6030 if (cset) {
6031 put_css_set(cset);
6032 kargs->cset = NULL;
6033 }
6034
6035 if (cgrp) {
6036 cgroup_put(cgrp);
6037 kargs->cgrp = NULL;
6038 }
6039 }
6040}
6041
6042/**
6043 * cgroup_can_fork - called on a new task before the process is exposed
6044 * @child: the child process
6045 *
6046 * This prepares a new css_set for the child process which the child will
6047 * be attached to in cgroup_post_fork().
6048 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6049 * callback returns an error, the fork aborts with that error code. This
6050 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6051 */
6052int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6053{
6054 struct cgroup_subsys *ss;
6055 int i, j, ret;
6056
6057 ret = cgroup_css_set_fork(kargs);
6058 if (ret)
6059 return ret;
6060
6061 do_each_subsys_mask(ss, i, have_canfork_callback) {
6062 ret = ss->can_fork(child, kargs->cset);
6063 if (ret)
6064 goto out_revert;
6065 } while_each_subsys_mask();
6066
6067 return 0;
6068
6069out_revert:
6070 for_each_subsys(ss, j) {
6071 if (j >= i)
6072 break;
6073 if (ss->cancel_fork)
6074 ss->cancel_fork(child, kargs->cset);
6075 }
6076
6077 cgroup_css_set_put_fork(kargs);
6078
6079 return ret;
6080}
6081
6082/**
6083 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6084 * @child: the child process
6085 * @kargs: the arguments passed to create the child process
6086 *
6087 * This calls the cancel_fork() callbacks if a fork failed *after*
6088 * cgroup_can_fork() succeded and cleans up references we took to
6089 * prepare a new css_set for the child process in cgroup_can_fork().
6090 */
6091void cgroup_cancel_fork(struct task_struct *child,
6092 struct kernel_clone_args *kargs)
6093{
6094 struct cgroup_subsys *ss;
6095 int i;
6096
6097 for_each_subsys(ss, i)
6098 if (ss->cancel_fork)
6099 ss->cancel_fork(child, kargs->cset);
6100
6101 cgroup_css_set_put_fork(kargs);
6102}
6103
6104/**
6105 * cgroup_post_fork - finalize cgroup setup for the child process
6106 * @child: the child process
6107 *
6108 * Attach the child process to its css_set calling the subsystem fork()
6109 * callbacks.
6110 */
6111void cgroup_post_fork(struct task_struct *child,
6112 struct kernel_clone_args *kargs)
6113 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6114{
6115 struct cgroup_subsys *ss;
6116 struct css_set *cset;
6117 int i;
6118
6119 cset = kargs->cset;
6120 kargs->cset = NULL;
6121
6122 spin_lock_irq(&css_set_lock);
6123
6124 /* init tasks are special, only link regular threads */
6125 if (likely(child->pid)) {
6126 WARN_ON_ONCE(!list_empty(&child->cg_list));
6127 cset->nr_tasks++;
6128 css_set_move_task(child, NULL, cset, false);
6129 } else {
6130 put_css_set(cset);
6131 cset = NULL;
6132 }
6133
6134 /*
6135 * If the cgroup has to be frozen, the new task has too. Let's set
6136 * the JOBCTL_TRAP_FREEZE jobctl bit to get the task into the
6137 * frozen state.
6138 */
6139 if (unlikely(cgroup_task_freeze(child))) {
6140 spin_lock(&child->sighand->siglock);
6141 WARN_ON_ONCE(child->frozen);
6142 child->jobctl |= JOBCTL_TRAP_FREEZE;
6143 spin_unlock(&child->sighand->siglock);
6144
6145 /*
6146 * Calling cgroup_update_frozen() isn't required here,
6147 * because it will be called anyway a bit later from
6148 * do_freezer_trap(). So we avoid cgroup's transient switch
6149 * from the frozen state and back.
6150 */
6151 }
6152
6153 spin_unlock_irq(&css_set_lock);
6154
6155 /*
6156 * Call ss->fork(). This must happen after @child is linked on
6157 * css_set; otherwise, @child might change state between ->fork()
6158 * and addition to css_set.
6159 */
6160 do_each_subsys_mask(ss, i, have_fork_callback) {
6161 ss->fork(child);
6162 } while_each_subsys_mask();
6163
6164 /* Make the new cset the root_cset of the new cgroup namespace. */
6165 if (kargs->flags & CLONE_NEWCGROUP) {
6166 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6167
6168 get_css_set(cset);
6169 child->nsproxy->cgroup_ns->root_cset = cset;
6170 put_css_set(rcset);
6171 }
6172
6173 cgroup_css_set_put_fork(kargs);
6174}
6175
6176/**
6177 * cgroup_exit - detach cgroup from exiting task
6178 * @tsk: pointer to task_struct of exiting process
6179 *
6180 * Description: Detach cgroup from @tsk.
6181 *
6182 */
6183void cgroup_exit(struct task_struct *tsk)
6184{
6185 struct cgroup_subsys *ss;
6186 struct css_set *cset;
6187 int i;
6188
6189 spin_lock_irq(&css_set_lock);
6190
6191 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6192 cset = task_css_set(tsk);
6193 css_set_move_task(tsk, cset, NULL, false);
6194 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6195 cset->nr_tasks--;
6196
6197 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6198 if (unlikely(cgroup_task_freeze(tsk)))
6199 cgroup_update_frozen(task_dfl_cgroup(tsk));
6200
6201 spin_unlock_irq(&css_set_lock);
6202
6203 /* see cgroup_post_fork() for details */
6204 do_each_subsys_mask(ss, i, have_exit_callback) {
6205 ss->exit(tsk);
6206 } while_each_subsys_mask();
6207}
6208
6209void cgroup_release(struct task_struct *task)
6210{
6211 struct cgroup_subsys *ss;
6212 int ssid;
6213
6214 do_each_subsys_mask(ss, ssid, have_release_callback) {
6215 ss->release(task);
6216 } while_each_subsys_mask();
6217
6218 spin_lock_irq(&css_set_lock);
6219 css_set_skip_task_iters(task_css_set(task), task);
6220 list_del_init(&task->cg_list);
6221 spin_unlock_irq(&css_set_lock);
6222}
6223
6224void cgroup_free(struct task_struct *task)
6225{
6226 struct css_set *cset = task_css_set(task);
6227 put_css_set(cset);
6228}
6229
6230static int __init cgroup_disable(char *str)
6231{
6232 struct cgroup_subsys *ss;
6233 char *token;
6234 int i;
6235
6236 while ((token = strsep(&str, ",")) != NULL) {
6237 if (!*token)
6238 continue;
6239
6240 for_each_subsys(ss, i) {
6241 if (strcmp(token, ss->name) &&
6242 strcmp(token, ss->legacy_name))
6243 continue;
6244 cgroup_disable_mask |= 1 << i;
6245 }
6246 }
6247 return 1;
6248}
6249__setup("cgroup_disable=", cgroup_disable);
6250
6251void __init __weak enable_debug_cgroup(void) { }
6252
6253static int __init enable_cgroup_debug(char *str)
6254{
6255 cgroup_debug = true;
6256 enable_debug_cgroup();
6257 return 1;
6258}
6259__setup("cgroup_debug", enable_cgroup_debug);
6260
6261/**
6262 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6263 * @dentry: directory dentry of interest
6264 * @ss: subsystem of interest
6265 *
6266 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6267 * to get the corresponding css and return it. If such css doesn't exist
6268 * or can't be pinned, an ERR_PTR value is returned.
6269 */
6270struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6271 struct cgroup_subsys *ss)
6272{
6273 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6274 struct file_system_type *s_type = dentry->d_sb->s_type;
6275 struct cgroup_subsys_state *css = NULL;
6276 struct cgroup *cgrp;
6277
6278 /* is @dentry a cgroup dir? */
6279 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6280 !kn || kernfs_type(kn) != KERNFS_DIR)
6281 return ERR_PTR(-EBADF);
6282
6283 rcu_read_lock();
6284
6285 /*
6286 * This path doesn't originate from kernfs and @kn could already
6287 * have been or be removed at any point. @kn->priv is RCU
6288 * protected for this access. See css_release_work_fn() for details.
6289 */
6290 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6291 if (cgrp)
6292 css = cgroup_css(cgrp, ss);
6293
6294 if (!css || !css_tryget_online(css))
6295 css = ERR_PTR(-ENOENT);
6296
6297 rcu_read_unlock();
6298 return css;
6299}
6300
6301/**
6302 * css_from_id - lookup css by id
6303 * @id: the cgroup id
6304 * @ss: cgroup subsys to be looked into
6305 *
6306 * Returns the css if there's valid one with @id, otherwise returns NULL.
6307 * Should be called under rcu_read_lock().
6308 */
6309struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6310{
6311 WARN_ON_ONCE(!rcu_read_lock_held());
6312 return idr_find(&ss->css_idr, id);
6313}
6314
6315/**
6316 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6317 * @path: path on the default hierarchy
6318 *
6319 * Find the cgroup at @path on the default hierarchy, increment its
6320 * reference count and return it. Returns pointer to the found cgroup on
6321 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6322 * if @path points to a non-directory.
6323 */
6324struct cgroup *cgroup_get_from_path(const char *path)
6325{
6326 struct kernfs_node *kn;
6327 struct cgroup *cgrp;
6328
6329 mutex_lock(&cgroup_mutex);
6330
6331 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6332 if (kn) {
6333 if (kernfs_type(kn) == KERNFS_DIR) {
6334 cgrp = kn->priv;
6335 cgroup_get_live(cgrp);
6336 } else {
6337 cgrp = ERR_PTR(-ENOTDIR);
6338 }
6339 kernfs_put(kn);
6340 } else {
6341 cgrp = ERR_PTR(-ENOENT);
6342 }
6343
6344 mutex_unlock(&cgroup_mutex);
6345 return cgrp;
6346}
6347EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6348
6349/**
6350 * cgroup_get_from_fd - get a cgroup pointer from a fd
6351 * @fd: fd obtained by open(cgroup2_dir)
6352 *
6353 * Find the cgroup from a fd which should be obtained
6354 * by opening a cgroup directory. Returns a pointer to the
6355 * cgroup on success. ERR_PTR is returned if the cgroup
6356 * cannot be found.
6357 */
6358struct cgroup *cgroup_get_from_fd(int fd)
6359{
6360 struct cgroup *cgrp;
6361 struct file *f;
6362
6363 f = fget_raw(fd);
6364 if (!f)
6365 return ERR_PTR(-EBADF);
6366
6367 cgrp = cgroup_get_from_file(f);
6368 fput(f);
6369 return cgrp;
6370}
6371EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6372
6373static u64 power_of_ten(int power)
6374{
6375 u64 v = 1;
6376 while (power--)
6377 v *= 10;
6378 return v;
6379}
6380
6381/**
6382 * cgroup_parse_float - parse a floating number
6383 * @input: input string
6384 * @dec_shift: number of decimal digits to shift
6385 * @v: output
6386 *
6387 * Parse a decimal floating point number in @input and store the result in
6388 * @v with decimal point right shifted @dec_shift times. For example, if
6389 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6390 * Returns 0 on success, -errno otherwise.
6391 *
6392 * There's nothing cgroup specific about this function except that it's
6393 * currently the only user.
6394 */
6395int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6396{
6397 s64 whole, frac = 0;
6398 int fstart = 0, fend = 0, flen;
6399
6400 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6401 return -EINVAL;
6402 if (frac < 0)
6403 return -EINVAL;
6404
6405 flen = fend > fstart ? fend - fstart : 0;
6406 if (flen < dec_shift)
6407 frac *= power_of_ten(dec_shift - flen);
6408 else
6409 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6410
6411 *v = whole * power_of_ten(dec_shift) + frac;
6412 return 0;
6413}
6414
6415/*
6416 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6417 * definition in cgroup-defs.h.
6418 */
6419#ifdef CONFIG_SOCK_CGROUP_DATA
6420
6421#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6422
6423DEFINE_SPINLOCK(cgroup_sk_update_lock);
6424static bool cgroup_sk_alloc_disabled __read_mostly;
6425
6426void cgroup_sk_alloc_disable(void)
6427{
6428 if (cgroup_sk_alloc_disabled)
6429 return;
6430 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6431 cgroup_sk_alloc_disabled = true;
6432}
6433
6434#else
6435
6436#define cgroup_sk_alloc_disabled false
6437
6438#endif
6439
6440void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6441{
6442 if (cgroup_sk_alloc_disabled) {
6443 skcd->no_refcnt = 1;
6444 return;
6445 }
6446
6447 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6448 if (in_interrupt())
6449 return;
6450
6451 rcu_read_lock();
6452
6453 while (true) {
6454 struct css_set *cset;
6455
6456 cset = task_css_set(current);
6457 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6458 skcd->val = (unsigned long)cset->dfl_cgrp;
6459 cgroup_bpf_get(cset->dfl_cgrp);
6460 break;
6461 }
6462 cpu_relax();
6463 }
6464
6465 rcu_read_unlock();
6466}
6467
6468void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6469{
6470 if (skcd->val) {
6471 if (skcd->no_refcnt)
6472 return;
6473 /*
6474 * We might be cloning a socket which is left in an empty
6475 * cgroup and the cgroup might have already been rmdir'd.
6476 * Don't use cgroup_get_live().
6477 */
6478 cgroup_get(sock_cgroup_ptr(skcd));
6479 cgroup_bpf_get(sock_cgroup_ptr(skcd));
6480 }
6481}
6482
6483void cgroup_sk_free(struct sock_cgroup_data *skcd)
6484{
6485 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6486
6487 if (skcd->no_refcnt)
6488 return;
6489 cgroup_bpf_put(cgrp);
6490 cgroup_put(cgrp);
6491}
6492
6493#endif /* CONFIG_SOCK_CGROUP_DATA */
6494
6495#ifdef CONFIG_CGROUP_BPF
6496int cgroup_bpf_attach(struct cgroup *cgrp,
6497 struct bpf_prog *prog, struct bpf_prog *replace_prog,
6498 struct bpf_cgroup_link *link,
6499 enum bpf_attach_type type,
6500 u32 flags)
6501{
6502 int ret;
6503
6504 mutex_lock(&cgroup_mutex);
6505 ret = __cgroup_bpf_attach(cgrp, prog, replace_prog, link, type, flags);
6506 mutex_unlock(&cgroup_mutex);
6507 return ret;
6508}
6509
6510int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6511 enum bpf_attach_type type)
6512{
6513 int ret;
6514
6515 mutex_lock(&cgroup_mutex);
6516 ret = __cgroup_bpf_detach(cgrp, prog, NULL, type);
6517 mutex_unlock(&cgroup_mutex);
6518 return ret;
6519}
6520
6521int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6522 union bpf_attr __user *uattr)
6523{
6524 int ret;
6525
6526 mutex_lock(&cgroup_mutex);
6527 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6528 mutex_unlock(&cgroup_mutex);
6529 return ret;
6530}
6531#endif /* CONFIG_CGROUP_BPF */
6532
6533#ifdef CONFIG_SYSFS
6534static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6535 ssize_t size, const char *prefix)
6536{
6537 struct cftype *cft;
6538 ssize_t ret = 0;
6539
6540 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6541 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6542 continue;
6543
6544 if (prefix)
6545 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6546
6547 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6548
6549 if (WARN_ON(ret >= size))
6550 break;
6551 }
6552
6553 return ret;
6554}
6555
6556static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6557 char *buf)
6558{
6559 struct cgroup_subsys *ss;
6560 int ssid;
6561 ssize_t ret = 0;
6562
6563 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6564 NULL);
6565
6566 for_each_subsys(ss, ssid)
6567 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6568 PAGE_SIZE - ret,
6569 cgroup_subsys_name[ssid]);
6570
6571 return ret;
6572}
6573static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6574
6575static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6576 char *buf)
6577{
6578 return snprintf(buf, PAGE_SIZE,
6579 "nsdelegate\n"
6580 "memory_localevents\n"
6581 "memory_recursiveprot\n");
6582}
6583static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6584
6585static struct attribute *cgroup_sysfs_attrs[] = {
6586 &cgroup_delegate_attr.attr,
6587 &cgroup_features_attr.attr,
6588 NULL,
6589};
6590
6591static const struct attribute_group cgroup_sysfs_attr_group = {
6592 .attrs = cgroup_sysfs_attrs,
6593 .name = "cgroup",
6594};
6595
6596static int __init cgroup_sysfs_init(void)
6597{
6598 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6599}
6600subsys_initcall(cgroup_sysfs_init);
6601
6602#endif /* CONFIG_SYSFS */