projects/linux·blob v4.4 · refs

5969 lines · 161.1 KiB

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 <linux/cgroup.h>
32#include <linux/cred.h>
33#include <linux/ctype.h>
34#include <linux/errno.h>
35#include <linux/init_task.h>
36#include <linux/kernel.h>
37#include <linux/list.h>
38#include <linux/magic.h>
39#include <linux/mm.h>
40#include <linux/mutex.h>
41#include <linux/mount.h>
42#include <linux/pagemap.h>
43#include <linux/proc_fs.h>
44#include <linux/rcupdate.h>
45#include <linux/sched.h>
46#include <linux/slab.h>
47#include <linux/spinlock.h>
48#include <linux/percpu-rwsem.h>
49#include <linux/string.h>
50#include <linux/sort.h>
51#include <linux/kmod.h>
52#include <linux/delayacct.h>
53#include <linux/cgroupstats.h>
54#include <linux/hashtable.h>
55#include <linux/pid_namespace.h>
56#include <linux/idr.h>
57#include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58#include <linux/kthread.h>
59#include <linux/delay.h>
60
61#include <linux/atomic.h>
62
63/*
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
68 */
69#define CGROUP_PIDLIST_DESTROY_DELAY HZ
70
71#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
72 MAX_CFTYPE_NAME + 2)
73
74/*
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
77 *
78 * css_set_lock protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
80 *
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
83 */
84#ifdef CONFIG_PROVE_RCU
85DEFINE_MUTEX(cgroup_mutex);
86DEFINE_SPINLOCK(css_set_lock);
87EXPORT_SYMBOL_GPL(cgroup_mutex);
88EXPORT_SYMBOL_GPL(css_set_lock);
89#else
90static DEFINE_MUTEX(cgroup_mutex);
91static DEFINE_SPINLOCK(css_set_lock);
92#endif
93
94/*
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
97 */
98static DEFINE_SPINLOCK(cgroup_idr_lock);
99
100/*
101 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
102 * against file removal/re-creation across css hiding.
103 */
104static DEFINE_SPINLOCK(cgroup_file_kn_lock);
105
106/*
107 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
108 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
109 */
110static DEFINE_SPINLOCK(release_agent_path_lock);
111
112struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
113
114#define cgroup_assert_mutex_or_rcu_locked() \
115 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
116 !lockdep_is_held(&cgroup_mutex), \
117 "cgroup_mutex or RCU read lock required");
118
119/*
120 * cgroup destruction makes heavy use of work items and there can be a lot
121 * of concurrent destructions. Use a separate workqueue so that cgroup
122 * destruction work items don't end up filling up max_active of system_wq
123 * which may lead to deadlock.
124 */
125static struct workqueue_struct *cgroup_destroy_wq;
126
127/*
128 * pidlist destructions need to be flushed on cgroup destruction. Use a
129 * separate workqueue as flush domain.
130 */
131static struct workqueue_struct *cgroup_pidlist_destroy_wq;
132
133/* generate an array of cgroup subsystem pointers */
134#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
135static struct cgroup_subsys *cgroup_subsys[] = {
136#include <linux/cgroup_subsys.h>
137};
138#undef SUBSYS
139
140/* array of cgroup subsystem names */
141#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
142static const char *cgroup_subsys_name[] = {
143#include <linux/cgroup_subsys.h>
144};
145#undef SUBSYS
146
147/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
148#define SUBSYS(_x) \
149 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
150 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
151 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
152 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
153#include <linux/cgroup_subsys.h>
154#undef SUBSYS
155
156#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
157static struct static_key_true *cgroup_subsys_enabled_key[] = {
158#include <linux/cgroup_subsys.h>
159};
160#undef SUBSYS
161
162#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
163static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
164#include <linux/cgroup_subsys.h>
165};
166#undef SUBSYS
167
168/*
169 * The default hierarchy, reserved for the subsystems that are otherwise
170 * unattached - it never has more than a single cgroup, and all tasks are
171 * part of that cgroup.
172 */
173struct cgroup_root cgrp_dfl_root;
174EXPORT_SYMBOL_GPL(cgrp_dfl_root);
175
176/*
177 * The default hierarchy always exists but is hidden until mounted for the
178 * first time. This is for backward compatibility.
179 */
180static bool cgrp_dfl_root_visible;
181
182/* some controllers are not supported in the default hierarchy */
183static unsigned long cgrp_dfl_root_inhibit_ss_mask;
184
185/* The list of hierarchy roots */
186
187static LIST_HEAD(cgroup_roots);
188static int cgroup_root_count;
189
190/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
191static DEFINE_IDR(cgroup_hierarchy_idr);
192
193/*
194 * Assign a monotonically increasing serial number to csses. It guarantees
195 * cgroups with bigger numbers are newer than those with smaller numbers.
196 * Also, as csses are always appended to the parent's ->children list, it
197 * guarantees that sibling csses are always sorted in the ascending serial
198 * number order on the list. Protected by cgroup_mutex.
199 */
200static u64 css_serial_nr_next = 1;
201
202/*
203 * These bitmask flags indicate whether tasks in the fork and exit paths have
204 * fork/exit handlers to call. This avoids us having to do extra work in the
205 * fork/exit path to check which subsystems have fork/exit callbacks.
206 */
207static unsigned long have_fork_callback __read_mostly;
208static unsigned long have_exit_callback __read_mostly;
209static unsigned long have_free_callback __read_mostly;
210
211/* Ditto for the can_fork callback. */
212static unsigned long have_canfork_callback __read_mostly;
213
214static struct cftype cgroup_dfl_base_files[];
215static struct cftype cgroup_legacy_base_files[];
216
217static int rebind_subsystems(struct cgroup_root *dst_root,
218 unsigned long ss_mask);
219static void css_task_iter_advance(struct css_task_iter *it);
220static int cgroup_destroy_locked(struct cgroup *cgrp);
221static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
222 bool visible);
223static void css_release(struct percpu_ref *ref);
224static void kill_css(struct cgroup_subsys_state *css);
225static int cgroup_addrm_files(struct cgroup_subsys_state *css,
226 struct cgroup *cgrp, struct cftype cfts[],
227 bool is_add);
228
229/**
230 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
231 * @ssid: subsys ID of interest
232 *
233 * cgroup_subsys_enabled() can only be used with literal subsys names which
234 * is fine for individual subsystems but unsuitable for cgroup core. This
235 * is slower static_key_enabled() based test indexed by @ssid.
236 */
237static bool cgroup_ssid_enabled(int ssid)
238{
239 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
240}
241
242/**
243 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
244 * @cgrp: the cgroup of interest
245 *
246 * The default hierarchy is the v2 interface of cgroup and this function
247 * can be used to test whether a cgroup is on the default hierarchy for
248 * cases where a subsystem should behave differnetly depending on the
249 * interface version.
250 *
251 * The set of behaviors which change on the default hierarchy are still
252 * being determined and the mount option is prefixed with __DEVEL__.
253 *
254 * List of changed behaviors:
255 *
256 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
257 * and "name" are disallowed.
258 *
259 * - When mounting an existing superblock, mount options should match.
260 *
261 * - Remount is disallowed.
262 *
263 * - rename(2) is disallowed.
264 *
265 * - "tasks" is removed. Everything should be at process granularity. Use
266 * "cgroup.procs" instead.
267 *
268 * - "cgroup.procs" is not sorted. pids will be unique unless they got
269 * recycled inbetween reads.
270 *
271 * - "release_agent" and "notify_on_release" are removed. Replacement
272 * notification mechanism will be implemented.
273 *
274 * - "cgroup.clone_children" is removed.
275 *
276 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
277 * and its descendants contain no task; otherwise, 1. The file also
278 * generates kernfs notification which can be monitored through poll and
279 * [di]notify when the value of the file changes.
280 *
281 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
282 * take masks of ancestors with non-empty cpus/mems, instead of being
283 * moved to an ancestor.
284 *
285 * - cpuset: a task can be moved into an empty cpuset, and again it takes
286 * masks of ancestors.
287 *
288 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
289 * is not created.
290 *
291 * - blkcg: blk-throttle becomes properly hierarchical.
292 *
293 * - debug: disallowed on the default hierarchy.
294 */
295static bool cgroup_on_dfl(const struct cgroup *cgrp)
296{
297 return cgrp->root == &cgrp_dfl_root;
298}
299
300/* IDR wrappers which synchronize using cgroup_idr_lock */
301static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
302 gfp_t gfp_mask)
303{
304 int ret;
305
306 idr_preload(gfp_mask);
307 spin_lock_bh(&cgroup_idr_lock);
308 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
309 spin_unlock_bh(&cgroup_idr_lock);
310 idr_preload_end();
311 return ret;
312}
313
314static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
315{
316 void *ret;
317
318 spin_lock_bh(&cgroup_idr_lock);
319 ret = idr_replace(idr, ptr, id);
320 spin_unlock_bh(&cgroup_idr_lock);
321 return ret;
322}
323
324static void cgroup_idr_remove(struct idr *idr, int id)
325{
326 spin_lock_bh(&cgroup_idr_lock);
327 idr_remove(idr, id);
328 spin_unlock_bh(&cgroup_idr_lock);
329}
330
331static struct cgroup *cgroup_parent(struct cgroup *cgrp)
332{
333 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
334
335 if (parent_css)
336 return container_of(parent_css, struct cgroup, self);
337 return NULL;
338}
339
340/**
341 * cgroup_css - obtain a cgroup's css for the specified subsystem
342 * @cgrp: the cgroup of interest
343 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
344 *
345 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
346 * function must be called either under cgroup_mutex or rcu_read_lock() and
347 * the caller is responsible for pinning the returned css if it wants to
348 * keep accessing it outside the said locks. This function may return
349 * %NULL if @cgrp doesn't have @subsys_id enabled.
350 */
351static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
352 struct cgroup_subsys *ss)
353{
354 if (ss)
355 return rcu_dereference_check(cgrp->subsys[ss->id],
356 lockdep_is_held(&cgroup_mutex));
357 else
358 return &cgrp->self;
359}
360
361/**
362 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
363 * @cgrp: the cgroup of interest
364 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
365 *
366 * Similar to cgroup_css() but returns the effective css, which is defined
367 * as the matching css of the nearest ancestor including self which has @ss
368 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
369 * function is guaranteed to return non-NULL css.
370 */
371static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
372 struct cgroup_subsys *ss)
373{
374 lockdep_assert_held(&cgroup_mutex);
375
376 if (!ss)
377 return &cgrp->self;
378
379 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
380 return NULL;
381
382 /*
383 * This function is used while updating css associations and thus
384 * can't test the csses directly. Use ->child_subsys_mask.
385 */
386 while (cgroup_parent(cgrp) &&
387 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
388 cgrp = cgroup_parent(cgrp);
389
390 return cgroup_css(cgrp, ss);
391}
392
393/**
394 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
395 * @cgrp: the cgroup of interest
396 * @ss: the subsystem of interest
397 *
398 * Find and get the effective css of @cgrp for @ss. The effective css is
399 * defined as the matching css of the nearest ancestor including self which
400 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
401 * the root css is returned, so this function always returns a valid css.
402 * The returned css must be put using css_put().
403 */
404struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
405 struct cgroup_subsys *ss)
406{
407 struct cgroup_subsys_state *css;
408
409 rcu_read_lock();
410
411 do {
412 css = cgroup_css(cgrp, ss);
413
414 if (css && css_tryget_online(css))
415 goto out_unlock;
416 cgrp = cgroup_parent(cgrp);
417 } while (cgrp);
418
419 css = init_css_set.subsys[ss->id];
420 css_get(css);
421out_unlock:
422 rcu_read_unlock();
423 return css;
424}
425
426/* convenient tests for these bits */
427static inline bool cgroup_is_dead(const struct cgroup *cgrp)
428{
429 return !(cgrp->self.flags & CSS_ONLINE);
430}
431
432static void cgroup_get(struct cgroup *cgrp)
433{
434 WARN_ON_ONCE(cgroup_is_dead(cgrp));
435 css_get(&cgrp->self);
436}
437
438static bool cgroup_tryget(struct cgroup *cgrp)
439{
440 return css_tryget(&cgrp->self);
441}
442
443static void cgroup_put(struct cgroup *cgrp)
444{
445 css_put(&cgrp->self);
446}
447
448struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
449{
450 struct cgroup *cgrp = of->kn->parent->priv;
451 struct cftype *cft = of_cft(of);
452
453 /*
454 * This is open and unprotected implementation of cgroup_css().
455 * seq_css() is only called from a kernfs file operation which has
456 * an active reference on the file. Because all the subsystem
457 * files are drained before a css is disassociated with a cgroup,
458 * the matching css from the cgroup's subsys table is guaranteed to
459 * be and stay valid until the enclosing operation is complete.
460 */
461 if (cft->ss)
462 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
463 else
464 return &cgrp->self;
465}
466EXPORT_SYMBOL_GPL(of_css);
467
468/**
469 * cgroup_is_descendant - test ancestry
470 * @cgrp: the cgroup to be tested
471 * @ancestor: possible ancestor of @cgrp
472 *
473 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
474 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
475 * and @ancestor are accessible.
476 */
477bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
478{
479 while (cgrp) {
480 if (cgrp == ancestor)
481 return true;
482 cgrp = cgroup_parent(cgrp);
483 }
484 return false;
485}
486
487static int notify_on_release(const struct cgroup *cgrp)
488{
489 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
490}
491
492/**
493 * for_each_css - iterate all css's of a cgroup
494 * @css: the iteration cursor
495 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
496 * @cgrp: the target cgroup to iterate css's of
497 *
498 * Should be called under cgroup_[tree_]mutex.
499 */
500#define for_each_css(css, ssid, cgrp) \
501 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
502 if (!((css) = rcu_dereference_check( \
503 (cgrp)->subsys[(ssid)], \
504 lockdep_is_held(&cgroup_mutex)))) { } \
505 else
506
507/**
508 * for_each_e_css - iterate all effective css's of a cgroup
509 * @css: the iteration cursor
510 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
511 * @cgrp: the target cgroup to iterate css's of
512 *
513 * Should be called under cgroup_[tree_]mutex.
514 */
515#define for_each_e_css(css, ssid, cgrp) \
516 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
517 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
518 ; \
519 else
520
521/**
522 * for_each_subsys - iterate all enabled cgroup subsystems
523 * @ss: the iteration cursor
524 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
525 */
526#define for_each_subsys(ss, ssid) \
527 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
528 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
529
530/**
531 * for_each_subsys_which - filter for_each_subsys with a bitmask
532 * @ss: the iteration cursor
533 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
534 * @ss_maskp: a pointer to the bitmask
535 *
536 * The block will only run for cases where the ssid-th bit (1 << ssid) of
537 * mask is set to 1.
538 */
539#define for_each_subsys_which(ss, ssid, ss_maskp) \
540 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
541 (ssid) = 0; \
542 else \
543 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
544 if (((ss) = cgroup_subsys[ssid]) && false) \
545 break; \
546 else
547
548/* iterate across the hierarchies */
549#define for_each_root(root) \
550 list_for_each_entry((root), &cgroup_roots, root_list)
551
552/* iterate over child cgrps, lock should be held throughout iteration */
553#define cgroup_for_each_live_child(child, cgrp) \
554 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
555 if (({ lockdep_assert_held(&cgroup_mutex); \
556 cgroup_is_dead(child); })) \
557 ; \
558 else
559
560static void cgroup_release_agent(struct work_struct *work);
561static void check_for_release(struct cgroup *cgrp);
562
563/*
564 * A cgroup can be associated with multiple css_sets as different tasks may
565 * belong to different cgroups on different hierarchies. In the other
566 * direction, a css_set is naturally associated with multiple cgroups.
567 * This M:N relationship is represented by the following link structure
568 * which exists for each association and allows traversing the associations
569 * from both sides.
570 */
571struct cgrp_cset_link {
572 /* the cgroup and css_set this link associates */
573 struct cgroup *cgrp;
574 struct css_set *cset;
575
576 /* list of cgrp_cset_links anchored at cgrp->cset_links */
577 struct list_head cset_link;
578
579 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
580 struct list_head cgrp_link;
581};
582
583/*
584 * The default css_set - used by init and its children prior to any
585 * hierarchies being mounted. It contains a pointer to the root state
586 * for each subsystem. Also used to anchor the list of css_sets. Not
587 * reference-counted, to improve performance when child cgroups
588 * haven't been created.
589 */
590struct css_set init_css_set = {
591 .refcount = ATOMIC_INIT(1),
592 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
593 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
594 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
595 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
596 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
597 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
598};
599
600static int css_set_count = 1; /* 1 for init_css_set */
601
602/**
603 * css_set_populated - does a css_set contain any tasks?
604 * @cset: target css_set
605 */
606static bool css_set_populated(struct css_set *cset)
607{
608 lockdep_assert_held(&css_set_lock);
609
610 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
611}
612
613/**
614 * cgroup_update_populated - updated populated count of a cgroup
615 * @cgrp: the target cgroup
616 * @populated: inc or dec populated count
617 *
618 * One of the css_sets associated with @cgrp is either getting its first
619 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
620 * count is propagated towards root so that a given cgroup's populated_cnt
621 * is zero iff the cgroup and all its descendants don't contain any tasks.
622 *
623 * @cgrp's interface file "cgroup.populated" is zero if
624 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
625 * changes from or to zero, userland is notified that the content of the
626 * interface file has changed. This can be used to detect when @cgrp and
627 * its descendants become populated or empty.
628 */
629static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
630{
631 lockdep_assert_held(&css_set_lock);
632
633 do {
634 bool trigger;
635
636 if (populated)
637 trigger = !cgrp->populated_cnt++;
638 else
639 trigger = !--cgrp->populated_cnt;
640
641 if (!trigger)
642 break;
643
644 check_for_release(cgrp);
645 cgroup_file_notify(&cgrp->events_file);
646
647 cgrp = cgroup_parent(cgrp);
648 } while (cgrp);
649}
650
651/**
652 * css_set_update_populated - update populated state of a css_set
653 * @cset: target css_set
654 * @populated: whether @cset is populated or depopulated
655 *
656 * @cset is either getting the first task or losing the last. Update the
657 * ->populated_cnt of all associated cgroups accordingly.
658 */
659static void css_set_update_populated(struct css_set *cset, bool populated)
660{
661 struct cgrp_cset_link *link;
662
663 lockdep_assert_held(&css_set_lock);
664
665 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
666 cgroup_update_populated(link->cgrp, populated);
667}
668
669/**
670 * css_set_move_task - move a task from one css_set to another
671 * @task: task being moved
672 * @from_cset: css_set @task currently belongs to (may be NULL)
673 * @to_cset: new css_set @task is being moved to (may be NULL)
674 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
675 *
676 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
677 * css_set, @from_cset can be NULL. If @task is being disassociated
678 * instead of moved, @to_cset can be NULL.
679 *
680 * This function automatically handles populated_cnt updates and
681 * css_task_iter adjustments but the caller is responsible for managing
682 * @from_cset and @to_cset's reference counts.
683 */
684static void css_set_move_task(struct task_struct *task,
685 struct css_set *from_cset, struct css_set *to_cset,
686 bool use_mg_tasks)
687{
688 lockdep_assert_held(&css_set_lock);
689
690 if (from_cset) {
691 struct css_task_iter *it, *pos;
692
693 WARN_ON_ONCE(list_empty(&task->cg_list));
694
695 /*
696 * @task is leaving, advance task iterators which are
697 * pointing to it so that they can resume at the next
698 * position. Advancing an iterator might remove it from
699 * the list, use safe walk. See css_task_iter_advance*()
700 * for details.
701 */
702 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
703 iters_node)
704 if (it->task_pos == &task->cg_list)
705 css_task_iter_advance(it);
706
707 list_del_init(&task->cg_list);
708 if (!css_set_populated(from_cset))
709 css_set_update_populated(from_cset, false);
710 } else {
711 WARN_ON_ONCE(!list_empty(&task->cg_list));
712 }
713
714 if (to_cset) {
715 /*
716 * We are synchronized through cgroup_threadgroup_rwsem
717 * against PF_EXITING setting such that we can't race
718 * against cgroup_exit() changing the css_set to
719 * init_css_set and dropping the old one.
720 */
721 WARN_ON_ONCE(task->flags & PF_EXITING);
722
723 if (!css_set_populated(to_cset))
724 css_set_update_populated(to_cset, true);
725 rcu_assign_pointer(task->cgroups, to_cset);
726 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
727 &to_cset->tasks);
728 }
729}
730
731/*
732 * hash table for cgroup groups. This improves the performance to find
733 * an existing css_set. This hash doesn't (currently) take into
734 * account cgroups in empty hierarchies.
735 */
736#define CSS_SET_HASH_BITS 7
737static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
738
739static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
740{
741 unsigned long key = 0UL;
742 struct cgroup_subsys *ss;
743 int i;
744
745 for_each_subsys(ss, i)
746 key += (unsigned long)css[i];
747 key = (key >> 16) ^ key;
748
749 return key;
750}
751
752static void put_css_set_locked(struct css_set *cset)
753{
754 struct cgrp_cset_link *link, *tmp_link;
755 struct cgroup_subsys *ss;
756 int ssid;
757
758 lockdep_assert_held(&css_set_lock);
759
760 if (!atomic_dec_and_test(&cset->refcount))
761 return;
762
763 /* This css_set is dead. unlink it and release cgroup and css refs */
764 for_each_subsys(ss, ssid) {
765 list_del(&cset->e_cset_node[ssid]);
766 css_put(cset->subsys[ssid]);
767 }
768 hash_del(&cset->hlist);
769 css_set_count--;
770
771 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
772 list_del(&link->cset_link);
773 list_del(&link->cgrp_link);
774 if (cgroup_parent(link->cgrp))
775 cgroup_put(link->cgrp);
776 kfree(link);
777 }
778
779 kfree_rcu(cset, rcu_head);
780}
781
782static void put_css_set(struct css_set *cset)
783{
784 /*
785 * Ensure that the refcount doesn't hit zero while any readers
786 * can see it. Similar to atomic_dec_and_lock(), but for an
787 * rwlock
788 */
789 if (atomic_add_unless(&cset->refcount, -1, 1))
790 return;
791
792 spin_lock_bh(&css_set_lock);
793 put_css_set_locked(cset);
794 spin_unlock_bh(&css_set_lock);
795}
796
797/*
798 * refcounted get/put for css_set objects
799 */
800static inline void get_css_set(struct css_set *cset)
801{
802 atomic_inc(&cset->refcount);
803}
804
805/**
806 * compare_css_sets - helper function for find_existing_css_set().
807 * @cset: candidate css_set being tested
808 * @old_cset: existing css_set for a task
809 * @new_cgrp: cgroup that's being entered by the task
810 * @template: desired set of css pointers in css_set (pre-calculated)
811 *
812 * Returns true if "cset" matches "old_cset" except for the hierarchy
813 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
814 */
815static bool compare_css_sets(struct css_set *cset,
816 struct css_set *old_cset,
817 struct cgroup *new_cgrp,
818 struct cgroup_subsys_state *template[])
819{
820 struct list_head *l1, *l2;
821
822 /*
823 * On the default hierarchy, there can be csets which are
824 * associated with the same set of cgroups but different csses.
825 * Let's first ensure that csses match.
826 */
827 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
828 return false;
829
830 /*
831 * Compare cgroup pointers in order to distinguish between
832 * different cgroups in hierarchies. As different cgroups may
833 * share the same effective css, this comparison is always
834 * necessary.
835 */
836 l1 = &cset->cgrp_links;
837 l2 = &old_cset->cgrp_links;
838 while (1) {
839 struct cgrp_cset_link *link1, *link2;
840 struct cgroup *cgrp1, *cgrp2;
841
842 l1 = l1->next;
843 l2 = l2->next;
844 /* See if we reached the end - both lists are equal length. */
845 if (l1 == &cset->cgrp_links) {
846 BUG_ON(l2 != &old_cset->cgrp_links);
847 break;
848 } else {
849 BUG_ON(l2 == &old_cset->cgrp_links);
850 }
851 /* Locate the cgroups associated with these links. */
852 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
853 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
854 cgrp1 = link1->cgrp;
855 cgrp2 = link2->cgrp;
856 /* Hierarchies should be linked in the same order. */
857 BUG_ON(cgrp1->root != cgrp2->root);
858
859 /*
860 * If this hierarchy is the hierarchy of the cgroup
861 * that's changing, then we need to check that this
862 * css_set points to the new cgroup; if it's any other
863 * hierarchy, then this css_set should point to the
864 * same cgroup as the old css_set.
865 */
866 if (cgrp1->root == new_cgrp->root) {
867 if (cgrp1 != new_cgrp)
868 return false;
869 } else {
870 if (cgrp1 != cgrp2)
871 return false;
872 }
873 }
874 return true;
875}
876
877/**
878 * find_existing_css_set - init css array and find the matching css_set
879 * @old_cset: the css_set that we're using before the cgroup transition
880 * @cgrp: the cgroup that we're moving into
881 * @template: out param for the new set of csses, should be clear on entry
882 */
883static struct css_set *find_existing_css_set(struct css_set *old_cset,
884 struct cgroup *cgrp,
885 struct cgroup_subsys_state *template[])
886{
887 struct cgroup_root *root = cgrp->root;
888 struct cgroup_subsys *ss;
889 struct css_set *cset;
890 unsigned long key;
891 int i;
892
893 /*
894 * Build the set of subsystem state objects that we want to see in the
895 * new css_set. while subsystems can change globally, the entries here
896 * won't change, so no need for locking.
897 */
898 for_each_subsys(ss, i) {
899 if (root->subsys_mask & (1UL << i)) {
900 /*
901 * @ss is in this hierarchy, so we want the
902 * effective css from @cgrp.
903 */
904 template[i] = cgroup_e_css(cgrp, ss);
905 } else {
906 /*
907 * @ss is not in this hierarchy, so we don't want
908 * to change the css.
909 */
910 template[i] = old_cset->subsys[i];
911 }
912 }
913
914 key = css_set_hash(template);
915 hash_for_each_possible(css_set_table, cset, hlist, key) {
916 if (!compare_css_sets(cset, old_cset, cgrp, template))
917 continue;
918
919 /* This css_set matches what we need */
920 return cset;
921 }
922
923 /* No existing cgroup group matched */
924 return NULL;
925}
926
927static void free_cgrp_cset_links(struct list_head *links_to_free)
928{
929 struct cgrp_cset_link *link, *tmp_link;
930
931 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
932 list_del(&link->cset_link);
933 kfree(link);
934 }
935}
936
937/**
938 * allocate_cgrp_cset_links - allocate cgrp_cset_links
939 * @count: the number of links to allocate
940 * @tmp_links: list_head the allocated links are put on
941 *
942 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
943 * through ->cset_link. Returns 0 on success or -errno.
944 */
945static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
946{
947 struct cgrp_cset_link *link;
948 int i;
949
950 INIT_LIST_HEAD(tmp_links);
951
952 for (i = 0; i < count; i++) {
953 link = kzalloc(sizeof(*link), GFP_KERNEL);
954 if (!link) {
955 free_cgrp_cset_links(tmp_links);
956 return -ENOMEM;
957 }
958 list_add(&link->cset_link, tmp_links);
959 }
960 return 0;
961}
962
963/**
964 * link_css_set - a helper function to link a css_set to a cgroup
965 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
966 * @cset: the css_set to be linked
967 * @cgrp: the destination cgroup
968 */
969static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
970 struct cgroup *cgrp)
971{
972 struct cgrp_cset_link *link;
973
974 BUG_ON(list_empty(tmp_links));
975
976 if (cgroup_on_dfl(cgrp))
977 cset->dfl_cgrp = cgrp;
978
979 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
980 link->cset = cset;
981 link->cgrp = cgrp;
982
983 /*
984 * Always add links to the tail of the lists so that the lists are
985 * in choronological order.
986 */
987 list_move_tail(&link->cset_link, &cgrp->cset_links);
988 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
989
990 if (cgroup_parent(cgrp))
991 cgroup_get(cgrp);
992}
993
994/**
995 * find_css_set - return a new css_set with one cgroup updated
996 * @old_cset: the baseline css_set
997 * @cgrp: the cgroup to be updated
998 *
999 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1000 * substituted into the appropriate hierarchy.
1001 */
1002static struct css_set *find_css_set(struct css_set *old_cset,
1003 struct cgroup *cgrp)
1004{
1005 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1006 struct css_set *cset;
1007 struct list_head tmp_links;
1008 struct cgrp_cset_link *link;
1009 struct cgroup_subsys *ss;
1010 unsigned long key;
1011 int ssid;
1012
1013 lockdep_assert_held(&cgroup_mutex);
1014
1015 /* First see if we already have a cgroup group that matches
1016 * the desired set */
1017 spin_lock_bh(&css_set_lock);
1018 cset = find_existing_css_set(old_cset, cgrp, template);
1019 if (cset)
1020 get_css_set(cset);
1021 spin_unlock_bh(&css_set_lock);
1022
1023 if (cset)
1024 return cset;
1025
1026 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1027 if (!cset)
1028 return NULL;
1029
1030 /* Allocate all the cgrp_cset_link objects that we'll need */
1031 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1032 kfree(cset);
1033 return NULL;
1034 }
1035
1036 atomic_set(&cset->refcount, 1);
1037 INIT_LIST_HEAD(&cset->cgrp_links);
1038 INIT_LIST_HEAD(&cset->tasks);
1039 INIT_LIST_HEAD(&cset->mg_tasks);
1040 INIT_LIST_HEAD(&cset->mg_preload_node);
1041 INIT_LIST_HEAD(&cset->mg_node);
1042 INIT_LIST_HEAD(&cset->task_iters);
1043 INIT_HLIST_NODE(&cset->hlist);
1044
1045 /* Copy the set of subsystem state objects generated in
1046 * find_existing_css_set() */
1047 memcpy(cset->subsys, template, sizeof(cset->subsys));
1048
1049 spin_lock_bh(&css_set_lock);
1050 /* Add reference counts and links from the new css_set. */
1051 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1052 struct cgroup *c = link->cgrp;
1053
1054 if (c->root == cgrp->root)
1055 c = cgrp;
1056 link_css_set(&tmp_links, cset, c);
1057 }
1058
1059 BUG_ON(!list_empty(&tmp_links));
1060
1061 css_set_count++;
1062
1063 /* Add @cset to the hash table */
1064 key = css_set_hash(cset->subsys);
1065 hash_add(css_set_table, &cset->hlist, key);
1066
1067 for_each_subsys(ss, ssid) {
1068 struct cgroup_subsys_state *css = cset->subsys[ssid];
1069
1070 list_add_tail(&cset->e_cset_node[ssid],
1071 &css->cgroup->e_csets[ssid]);
1072 css_get(css);
1073 }
1074
1075 spin_unlock_bh(&css_set_lock);
1076
1077 return cset;
1078}
1079
1080static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1081{
1082 struct cgroup *root_cgrp = kf_root->kn->priv;
1083
1084 return root_cgrp->root;
1085}
1086
1087static int cgroup_init_root_id(struct cgroup_root *root)
1088{
1089 int id;
1090
1091 lockdep_assert_held(&cgroup_mutex);
1092
1093 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1094 if (id < 0)
1095 return id;
1096
1097 root->hierarchy_id = id;
1098 return 0;
1099}
1100
1101static void cgroup_exit_root_id(struct cgroup_root *root)
1102{
1103 lockdep_assert_held(&cgroup_mutex);
1104
1105 if (root->hierarchy_id) {
1106 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1107 root->hierarchy_id = 0;
1108 }
1109}
1110
1111static void cgroup_free_root(struct cgroup_root *root)
1112{
1113 if (root) {
1114 /* hierarchy ID should already have been released */
1115 WARN_ON_ONCE(root->hierarchy_id);
1116
1117 idr_destroy(&root->cgroup_idr);
1118 kfree(root);
1119 }
1120}
1121
1122static void cgroup_destroy_root(struct cgroup_root *root)
1123{
1124 struct cgroup *cgrp = &root->cgrp;
1125 struct cgrp_cset_link *link, *tmp_link;
1126
1127 mutex_lock(&cgroup_mutex);
1128
1129 BUG_ON(atomic_read(&root->nr_cgrps));
1130 BUG_ON(!list_empty(&cgrp->self.children));
1131
1132 /* Rebind all subsystems back to the default hierarchy */
1133 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1134
1135 /*
1136 * Release all the links from cset_links to this hierarchy's
1137 * root cgroup
1138 */
1139 spin_lock_bh(&css_set_lock);
1140
1141 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1142 list_del(&link->cset_link);
1143 list_del(&link->cgrp_link);
1144 kfree(link);
1145 }
1146
1147 spin_unlock_bh(&css_set_lock);
1148
1149 if (!list_empty(&root->root_list)) {
1150 list_del(&root->root_list);
1151 cgroup_root_count--;
1152 }
1153
1154 cgroup_exit_root_id(root);
1155
1156 mutex_unlock(&cgroup_mutex);
1157
1158 kernfs_destroy_root(root->kf_root);
1159 cgroup_free_root(root);
1160}
1161
1162/* look up cgroup associated with given css_set on the specified hierarchy */
1163static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1164 struct cgroup_root *root)
1165{
1166 struct cgroup *res = NULL;
1167
1168 lockdep_assert_held(&cgroup_mutex);
1169 lockdep_assert_held(&css_set_lock);
1170
1171 if (cset == &init_css_set) {
1172 res = &root->cgrp;
1173 } else {
1174 struct cgrp_cset_link *link;
1175
1176 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1177 struct cgroup *c = link->cgrp;
1178
1179 if (c->root == root) {
1180 res = c;
1181 break;
1182 }
1183 }
1184 }
1185
1186 BUG_ON(!res);
1187 return res;
1188}
1189
1190/*
1191 * Return the cgroup for "task" from the given hierarchy. Must be
1192 * called with cgroup_mutex and css_set_lock held.
1193 */
1194static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1195 struct cgroup_root *root)
1196{
1197 /*
1198 * No need to lock the task - since we hold cgroup_mutex the
1199 * task can't change groups, so the only thing that can happen
1200 * is that it exits and its css is set back to init_css_set.
1201 */
1202 return cset_cgroup_from_root(task_css_set(task), root);
1203}
1204
1205/*
1206 * A task must hold cgroup_mutex to modify cgroups.
1207 *
1208 * Any task can increment and decrement the count field without lock.
1209 * So in general, code holding cgroup_mutex can't rely on the count
1210 * field not changing. However, if the count goes to zero, then only
1211 * cgroup_attach_task() can increment it again. Because a count of zero
1212 * means that no tasks are currently attached, therefore there is no
1213 * way a task attached to that cgroup can fork (the other way to
1214 * increment the count). So code holding cgroup_mutex can safely
1215 * assume that if the count is zero, it will stay zero. Similarly, if
1216 * a task holds cgroup_mutex on a cgroup with zero count, it
1217 * knows that the cgroup won't be removed, as cgroup_rmdir()
1218 * needs that mutex.
1219 *
1220 * A cgroup can only be deleted if both its 'count' of using tasks
1221 * is zero, and its list of 'children' cgroups is empty. Since all
1222 * tasks in the system use _some_ cgroup, and since there is always at
1223 * least one task in the system (init, pid == 1), therefore, root cgroup
1224 * always has either children cgroups and/or using tasks. So we don't
1225 * need a special hack to ensure that root cgroup cannot be deleted.
1226 *
1227 * P.S. One more locking exception. RCU is used to guard the
1228 * update of a tasks cgroup pointer by cgroup_attach_task()
1229 */
1230
1231static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1232static const struct file_operations proc_cgroupstats_operations;
1233
1234static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1235 char *buf)
1236{
1237 struct cgroup_subsys *ss = cft->ss;
1238
1239 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1240 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1241 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1242 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1243 cft->name);
1244 else
1245 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1246 return buf;
1247}
1248
1249/**
1250 * cgroup_file_mode - deduce file mode of a control file
1251 * @cft: the control file in question
1252 *
1253 * S_IRUGO for read, S_IWUSR for write.
1254 */
1255static umode_t cgroup_file_mode(const struct cftype *cft)
1256{
1257 umode_t mode = 0;
1258
1259 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1260 mode |= S_IRUGO;
1261
1262 if (cft->write_u64 || cft->write_s64 || cft->write) {
1263 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1264 mode |= S_IWUGO;
1265 else
1266 mode |= S_IWUSR;
1267 }
1268
1269 return mode;
1270}
1271
1272/**
1273 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1274 * @cgrp: the target cgroup
1275 * @subtree_control: the new subtree_control mask to consider
1276 *
1277 * On the default hierarchy, a subsystem may request other subsystems to be
1278 * enabled together through its ->depends_on mask. In such cases, more
1279 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1280 *
1281 * This function calculates which subsystems need to be enabled if
1282 * @subtree_control is to be applied to @cgrp. The returned mask is always
1283 * a superset of @subtree_control and follows the usual hierarchy rules.
1284 */
1285static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1286 unsigned long subtree_control)
1287{
1288 struct cgroup *parent = cgroup_parent(cgrp);
1289 unsigned long cur_ss_mask = subtree_control;
1290 struct cgroup_subsys *ss;
1291 int ssid;
1292
1293 lockdep_assert_held(&cgroup_mutex);
1294
1295 if (!cgroup_on_dfl(cgrp))
1296 return cur_ss_mask;
1297
1298 while (true) {
1299 unsigned long new_ss_mask = cur_ss_mask;
1300
1301 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1302 new_ss_mask |= ss->depends_on;
1303
1304 /*
1305 * Mask out subsystems which aren't available. This can
1306 * happen only if some depended-upon subsystems were bound
1307 * to non-default hierarchies.
1308 */
1309 if (parent)
1310 new_ss_mask &= parent->child_subsys_mask;
1311 else
1312 new_ss_mask &= cgrp->root->subsys_mask;
1313
1314 if (new_ss_mask == cur_ss_mask)
1315 break;
1316 cur_ss_mask = new_ss_mask;
1317 }
1318
1319 return cur_ss_mask;
1320}
1321
1322/**
1323 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1324 * @cgrp: the target cgroup
1325 *
1326 * Update @cgrp->child_subsys_mask according to the current
1327 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1328 */
1329static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1330{
1331 cgrp->child_subsys_mask =
1332 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1333}
1334
1335/**
1336 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1337 * @kn: the kernfs_node being serviced
1338 *
1339 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1340 * the method finishes if locking succeeded. Note that once this function
1341 * returns the cgroup returned by cgroup_kn_lock_live() may become
1342 * inaccessible any time. If the caller intends to continue to access the
1343 * cgroup, it should pin it before invoking this function.
1344 */
1345static void cgroup_kn_unlock(struct kernfs_node *kn)
1346{
1347 struct cgroup *cgrp;
1348
1349 if (kernfs_type(kn) == KERNFS_DIR)
1350 cgrp = kn->priv;
1351 else
1352 cgrp = kn->parent->priv;
1353
1354 mutex_unlock(&cgroup_mutex);
1355
1356 kernfs_unbreak_active_protection(kn);
1357 cgroup_put(cgrp);
1358}
1359
1360/**
1361 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1362 * @kn: the kernfs_node being serviced
1363 *
1364 * This helper is to be used by a cgroup kernfs method currently servicing
1365 * @kn. It breaks the active protection, performs cgroup locking and
1366 * verifies that the associated cgroup is alive. Returns the cgroup if
1367 * alive; otherwise, %NULL. A successful return should be undone by a
1368 * matching cgroup_kn_unlock() invocation.
1369 *
1370 * Any cgroup kernfs method implementation which requires locking the
1371 * associated cgroup should use this helper. It avoids nesting cgroup
1372 * locking under kernfs active protection and allows all kernfs operations
1373 * including self-removal.
1374 */
1375static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1376{
1377 struct cgroup *cgrp;
1378
1379 if (kernfs_type(kn) == KERNFS_DIR)
1380 cgrp = kn->priv;
1381 else
1382 cgrp = kn->parent->priv;
1383
1384 /*
1385 * We're gonna grab cgroup_mutex which nests outside kernfs
1386 * active_ref. cgroup liveliness check alone provides enough
1387 * protection against removal. Ensure @cgrp stays accessible and
1388 * break the active_ref protection.
1389 */
1390 if (!cgroup_tryget(cgrp))
1391 return NULL;
1392 kernfs_break_active_protection(kn);
1393
1394 mutex_lock(&cgroup_mutex);
1395
1396 if (!cgroup_is_dead(cgrp))
1397 return cgrp;
1398
1399 cgroup_kn_unlock(kn);
1400 return NULL;
1401}
1402
1403static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1404{
1405 char name[CGROUP_FILE_NAME_MAX];
1406
1407 lockdep_assert_held(&cgroup_mutex);
1408
1409 if (cft->file_offset) {
1410 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1411 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1412
1413 spin_lock_irq(&cgroup_file_kn_lock);
1414 cfile->kn = NULL;
1415 spin_unlock_irq(&cgroup_file_kn_lock);
1416 }
1417
1418 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1419}
1420
1421/**
1422 * css_clear_dir - remove subsys files in a cgroup directory
1423 * @css: taget css
1424 * @cgrp_override: specify if target cgroup is different from css->cgroup
1425 */
1426static void css_clear_dir(struct cgroup_subsys_state *css,
1427 struct cgroup *cgrp_override)
1428{
1429 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1430 struct cftype *cfts;
1431
1432 list_for_each_entry(cfts, &css->ss->cfts, node)
1433 cgroup_addrm_files(css, cgrp, cfts, false);
1434}
1435
1436/**
1437 * css_populate_dir - create subsys files in a cgroup directory
1438 * @css: target css
1439 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1440 *
1441 * On failure, no file is added.
1442 */
1443static int css_populate_dir(struct cgroup_subsys_state *css,
1444 struct cgroup *cgrp_override)
1445{
1446 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1447 struct cftype *cfts, *failed_cfts;
1448 int ret;
1449
1450 if (!css->ss) {
1451 if (cgroup_on_dfl(cgrp))
1452 cfts = cgroup_dfl_base_files;
1453 else
1454 cfts = cgroup_legacy_base_files;
1455
1456 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1457 }
1458
1459 list_for_each_entry(cfts, &css->ss->cfts, node) {
1460 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1461 if (ret < 0) {
1462 failed_cfts = cfts;
1463 goto err;
1464 }
1465 }
1466 return 0;
1467err:
1468 list_for_each_entry(cfts, &css->ss->cfts, node) {
1469 if (cfts == failed_cfts)
1470 break;
1471 cgroup_addrm_files(css, cgrp, cfts, false);
1472 }
1473 return ret;
1474}
1475
1476static int rebind_subsystems(struct cgroup_root *dst_root,
1477 unsigned long ss_mask)
1478{
1479 struct cgroup *dcgrp = &dst_root->cgrp;
1480 struct cgroup_subsys *ss;
1481 unsigned long tmp_ss_mask;
1482 int ssid, i, ret;
1483
1484 lockdep_assert_held(&cgroup_mutex);
1485
1486 for_each_subsys_which(ss, ssid, &ss_mask) {
1487 /* if @ss has non-root csses attached to it, can't move */
1488 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1489 return -EBUSY;
1490
1491 /* can't move between two non-dummy roots either */
1492 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1493 return -EBUSY;
1494 }
1495
1496 /* skip creating root files on dfl_root for inhibited subsystems */
1497 tmp_ss_mask = ss_mask;
1498 if (dst_root == &cgrp_dfl_root)
1499 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1500
1501 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1502 struct cgroup *scgrp = &ss->root->cgrp;
1503 int tssid;
1504
1505 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1506 if (!ret)
1507 continue;
1508
1509 /*
1510 * Rebinding back to the default root is not allowed to
1511 * fail. Using both default and non-default roots should
1512 * be rare. Moving subsystems back and forth even more so.
1513 * Just warn about it and continue.
1514 */
1515 if (dst_root == &cgrp_dfl_root) {
1516 if (cgrp_dfl_root_visible) {
1517 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1518 ret, ss_mask);
1519 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1520 }
1521 continue;
1522 }
1523
1524 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1525 if (tssid == ssid)
1526 break;
1527 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1528 }
1529 return ret;
1530 }
1531
1532 /*
1533 * Nothing can fail from this point on. Remove files for the
1534 * removed subsystems and rebind each subsystem.
1535 */
1536 for_each_subsys_which(ss, ssid, &ss_mask) {
1537 struct cgroup_root *src_root = ss->root;
1538 struct cgroup *scgrp = &src_root->cgrp;
1539 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1540 struct css_set *cset;
1541
1542 WARN_ON(!css || cgroup_css(dcgrp, ss));
1543
1544 css_clear_dir(css, NULL);
1545
1546 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1547 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1548 ss->root = dst_root;
1549 css->cgroup = dcgrp;
1550
1551 spin_lock_bh(&css_set_lock);
1552 hash_for_each(css_set_table, i, cset, hlist)
1553 list_move_tail(&cset->e_cset_node[ss->id],
1554 &dcgrp->e_csets[ss->id]);
1555 spin_unlock_bh(&css_set_lock);
1556
1557 src_root->subsys_mask &= ~(1 << ssid);
1558 scgrp->subtree_control &= ~(1 << ssid);
1559 cgroup_refresh_child_subsys_mask(scgrp);
1560
1561 /* default hierarchy doesn't enable controllers by default */
1562 dst_root->subsys_mask |= 1 << ssid;
1563 if (dst_root == &cgrp_dfl_root) {
1564 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1565 } else {
1566 dcgrp->subtree_control |= 1 << ssid;
1567 cgroup_refresh_child_subsys_mask(dcgrp);
1568 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1569 }
1570
1571 if (ss->bind)
1572 ss->bind(css);
1573 }
1574
1575 kernfs_activate(dcgrp->kn);
1576 return 0;
1577}
1578
1579static int cgroup_show_options(struct seq_file *seq,
1580 struct kernfs_root *kf_root)
1581{
1582 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1583 struct cgroup_subsys *ss;
1584 int ssid;
1585
1586 if (root != &cgrp_dfl_root)
1587 for_each_subsys(ss, ssid)
1588 if (root->subsys_mask & (1 << ssid))
1589 seq_show_option(seq, ss->legacy_name, NULL);
1590 if (root->flags & CGRP_ROOT_NOPREFIX)
1591 seq_puts(seq, ",noprefix");
1592 if (root->flags & CGRP_ROOT_XATTR)
1593 seq_puts(seq, ",xattr");
1594
1595 spin_lock(&release_agent_path_lock);
1596 if (strlen(root->release_agent_path))
1597 seq_show_option(seq, "release_agent",
1598 root->release_agent_path);
1599 spin_unlock(&release_agent_path_lock);
1600
1601 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1602 seq_puts(seq, ",clone_children");
1603 if (strlen(root->name))
1604 seq_show_option(seq, "name", root->name);
1605 return 0;
1606}
1607
1608struct cgroup_sb_opts {
1609 unsigned long subsys_mask;
1610 unsigned int flags;
1611 char *release_agent;
1612 bool cpuset_clone_children;
1613 char *name;
1614 /* User explicitly requested empty subsystem */
1615 bool none;
1616};
1617
1618static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1619{
1620 char *token, *o = data;
1621 bool all_ss = false, one_ss = false;
1622 unsigned long mask = -1UL;
1623 struct cgroup_subsys *ss;
1624 int nr_opts = 0;
1625 int i;
1626
1627#ifdef CONFIG_CPUSETS
1628 mask = ~(1U << cpuset_cgrp_id);
1629#endif
1630
1631 memset(opts, 0, sizeof(*opts));
1632
1633 while ((token = strsep(&o, ",")) != NULL) {
1634 nr_opts++;
1635
1636 if (!*token)
1637 return -EINVAL;
1638 if (!strcmp(token, "none")) {
1639 /* Explicitly have no subsystems */
1640 opts->none = true;
1641 continue;
1642 }
1643 if (!strcmp(token, "all")) {
1644 /* Mutually exclusive option 'all' + subsystem name */
1645 if (one_ss)
1646 return -EINVAL;
1647 all_ss = true;
1648 continue;
1649 }
1650 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1651 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1652 continue;
1653 }
1654 if (!strcmp(token, "noprefix")) {
1655 opts->flags |= CGRP_ROOT_NOPREFIX;
1656 continue;
1657 }
1658 if (!strcmp(token, "clone_children")) {
1659 opts->cpuset_clone_children = true;
1660 continue;
1661 }
1662 if (!strcmp(token, "xattr")) {
1663 opts->flags |= CGRP_ROOT_XATTR;
1664 continue;
1665 }
1666 if (!strncmp(token, "release_agent=", 14)) {
1667 /* Specifying two release agents is forbidden */
1668 if (opts->release_agent)
1669 return -EINVAL;
1670 opts->release_agent =
1671 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1672 if (!opts->release_agent)
1673 return -ENOMEM;
1674 continue;
1675 }
1676 if (!strncmp(token, "name=", 5)) {
1677 const char *name = token + 5;
1678 /* Can't specify an empty name */
1679 if (!strlen(name))
1680 return -EINVAL;
1681 /* Must match [\w.-]+ */
1682 for (i = 0; i < strlen(name); i++) {
1683 char c = name[i];
1684 if (isalnum(c))
1685 continue;
1686 if ((c == '.') || (c == '-') || (c == '_'))
1687 continue;
1688 return -EINVAL;
1689 }
1690 /* Specifying two names is forbidden */
1691 if (opts->name)
1692 return -EINVAL;
1693 opts->name = kstrndup(name,
1694 MAX_CGROUP_ROOT_NAMELEN - 1,
1695 GFP_KERNEL);
1696 if (!opts->name)
1697 return -ENOMEM;
1698
1699 continue;
1700 }
1701
1702 for_each_subsys(ss, i) {
1703 if (strcmp(token, ss->legacy_name))
1704 continue;
1705 if (!cgroup_ssid_enabled(i))
1706 continue;
1707
1708 /* Mutually exclusive option 'all' + subsystem name */
1709 if (all_ss)
1710 return -EINVAL;
1711 opts->subsys_mask |= (1 << i);
1712 one_ss = true;
1713
1714 break;
1715 }
1716 if (i == CGROUP_SUBSYS_COUNT)
1717 return -ENOENT;
1718 }
1719
1720 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1721 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1722 if (nr_opts != 1) {
1723 pr_err("sane_behavior: no other mount options allowed\n");
1724 return -EINVAL;
1725 }
1726 return 0;
1727 }
1728
1729 /*
1730 * If the 'all' option was specified select all the subsystems,
1731 * otherwise if 'none', 'name=' and a subsystem name options were
1732 * not specified, let's default to 'all'
1733 */
1734 if (all_ss || (!one_ss && !opts->none && !opts->name))
1735 for_each_subsys(ss, i)
1736 if (cgroup_ssid_enabled(i))
1737 opts->subsys_mask |= (1 << i);
1738
1739 /*
1740 * We either have to specify by name or by subsystems. (So all
1741 * empty hierarchies must have a name).
1742 */
1743 if (!opts->subsys_mask && !opts->name)
1744 return -EINVAL;
1745
1746 /*
1747 * Option noprefix was introduced just for backward compatibility
1748 * with the old cpuset, so we allow noprefix only if mounting just
1749 * the cpuset subsystem.
1750 */
1751 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1752 return -EINVAL;
1753
1754 /* Can't specify "none" and some subsystems */
1755 if (opts->subsys_mask && opts->none)
1756 return -EINVAL;
1757
1758 return 0;
1759}
1760
1761static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1762{
1763 int ret = 0;
1764 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1765 struct cgroup_sb_opts opts;
1766 unsigned long added_mask, removed_mask;
1767
1768 if (root == &cgrp_dfl_root) {
1769 pr_err("remount is not allowed\n");
1770 return -EINVAL;
1771 }
1772
1773 mutex_lock(&cgroup_mutex);
1774
1775 /* See what subsystems are wanted */
1776 ret = parse_cgroupfs_options(data, &opts);
1777 if (ret)
1778 goto out_unlock;
1779
1780 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1781 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1782 task_tgid_nr(current), current->comm);
1783
1784 added_mask = opts.subsys_mask & ~root->subsys_mask;
1785 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1786
1787 /* Don't allow flags or name to change at remount */
1788 if ((opts.flags ^ root->flags) ||
1789 (opts.name && strcmp(opts.name, root->name))) {
1790 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1791 opts.flags, opts.name ?: "", root->flags, root->name);
1792 ret = -EINVAL;
1793 goto out_unlock;
1794 }
1795
1796 /* remounting is not allowed for populated hierarchies */
1797 if (!list_empty(&root->cgrp.self.children)) {
1798 ret = -EBUSY;
1799 goto out_unlock;
1800 }
1801
1802 ret = rebind_subsystems(root, added_mask);
1803 if (ret)
1804 goto out_unlock;
1805
1806 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1807
1808 if (opts.release_agent) {
1809 spin_lock(&release_agent_path_lock);
1810 strcpy(root->release_agent_path, opts.release_agent);
1811 spin_unlock(&release_agent_path_lock);
1812 }
1813 out_unlock:
1814 kfree(opts.release_agent);
1815 kfree(opts.name);
1816 mutex_unlock(&cgroup_mutex);
1817 return ret;
1818}
1819
1820/*
1821 * To reduce the fork() overhead for systems that are not actually using
1822 * their cgroups capability, we don't maintain the lists running through
1823 * each css_set to its tasks until we see the list actually used - in other
1824 * words after the first mount.
1825 */
1826static bool use_task_css_set_links __read_mostly;
1827
1828static void cgroup_enable_task_cg_lists(void)
1829{
1830 struct task_struct *p, *g;
1831
1832 spin_lock_bh(&css_set_lock);
1833
1834 if (use_task_css_set_links)
1835 goto out_unlock;
1836
1837 use_task_css_set_links = true;
1838
1839 /*
1840 * We need tasklist_lock because RCU is not safe against
1841 * while_each_thread(). Besides, a forking task that has passed
1842 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1843 * is not guaranteed to have its child immediately visible in the
1844 * tasklist if we walk through it with RCU.
1845 */
1846 read_lock(&tasklist_lock);
1847 do_each_thread(g, p) {
1848 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1849 task_css_set(p) != &init_css_set);
1850
1851 /*
1852 * We should check if the process is exiting, otherwise
1853 * it will race with cgroup_exit() in that the list
1854 * entry won't be deleted though the process has exited.
1855 * Do it while holding siglock so that we don't end up
1856 * racing against cgroup_exit().
1857 */
1858 spin_lock_irq(&p->sighand->siglock);
1859 if (!(p->flags & PF_EXITING)) {
1860 struct css_set *cset = task_css_set(p);
1861
1862 if (!css_set_populated(cset))
1863 css_set_update_populated(cset, true);
1864 list_add_tail(&p->cg_list, &cset->tasks);
1865 get_css_set(cset);
1866 }
1867 spin_unlock_irq(&p->sighand->siglock);
1868 } while_each_thread(g, p);
1869 read_unlock(&tasklist_lock);
1870out_unlock:
1871 spin_unlock_bh(&css_set_lock);
1872}
1873
1874static void init_cgroup_housekeeping(struct cgroup *cgrp)
1875{
1876 struct cgroup_subsys *ss;
1877 int ssid;
1878
1879 INIT_LIST_HEAD(&cgrp->self.sibling);
1880 INIT_LIST_HEAD(&cgrp->self.children);
1881 INIT_LIST_HEAD(&cgrp->cset_links);
1882 INIT_LIST_HEAD(&cgrp->pidlists);
1883 mutex_init(&cgrp->pidlist_mutex);
1884 cgrp->self.cgroup = cgrp;
1885 cgrp->self.flags |= CSS_ONLINE;
1886
1887 for_each_subsys(ss, ssid)
1888 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1889
1890 init_waitqueue_head(&cgrp->offline_waitq);
1891 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1892}
1893
1894static void init_cgroup_root(struct cgroup_root *root,
1895 struct cgroup_sb_opts *opts)
1896{
1897 struct cgroup *cgrp = &root->cgrp;
1898
1899 INIT_LIST_HEAD(&root->root_list);
1900 atomic_set(&root->nr_cgrps, 1);
1901 cgrp->root = root;
1902 init_cgroup_housekeeping(cgrp);
1903 idr_init(&root->cgroup_idr);
1904
1905 root->flags = opts->flags;
1906 if (opts->release_agent)
1907 strcpy(root->release_agent_path, opts->release_agent);
1908 if (opts->name)
1909 strcpy(root->name, opts->name);
1910 if (opts->cpuset_clone_children)
1911 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1912}
1913
1914static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1915{
1916 LIST_HEAD(tmp_links);
1917 struct cgroup *root_cgrp = &root->cgrp;
1918 struct css_set *cset;
1919 int i, ret;
1920
1921 lockdep_assert_held(&cgroup_mutex);
1922
1923 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1924 if (ret < 0)
1925 goto out;
1926 root_cgrp->id = ret;
1927
1928 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1929 GFP_KERNEL);
1930 if (ret)
1931 goto out;
1932
1933 /*
1934 * We're accessing css_set_count without locking css_set_lock here,
1935 * but that's OK - it can only be increased by someone holding
1936 * cgroup_lock, and that's us. The worst that can happen is that we
1937 * have some link structures left over
1938 */
1939 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1940 if (ret)
1941 goto cancel_ref;
1942
1943 ret = cgroup_init_root_id(root);
1944 if (ret)
1945 goto cancel_ref;
1946
1947 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1948 KERNFS_ROOT_CREATE_DEACTIVATED,
1949 root_cgrp);
1950 if (IS_ERR(root->kf_root)) {
1951 ret = PTR_ERR(root->kf_root);
1952 goto exit_root_id;
1953 }
1954 root_cgrp->kn = root->kf_root->kn;
1955
1956 ret = css_populate_dir(&root_cgrp->self, NULL);
1957 if (ret)
1958 goto destroy_root;
1959
1960 ret = rebind_subsystems(root, ss_mask);
1961 if (ret)
1962 goto destroy_root;
1963
1964 /*
1965 * There must be no failure case after here, since rebinding takes
1966 * care of subsystems' refcounts, which are explicitly dropped in
1967 * the failure exit path.
1968 */
1969 list_add(&root->root_list, &cgroup_roots);
1970 cgroup_root_count++;
1971
1972 /*
1973 * Link the root cgroup in this hierarchy into all the css_set
1974 * objects.
1975 */
1976 spin_lock_bh(&css_set_lock);
1977 hash_for_each(css_set_table, i, cset, hlist) {
1978 link_css_set(&tmp_links, cset, root_cgrp);
1979 if (css_set_populated(cset))
1980 cgroup_update_populated(root_cgrp, true);
1981 }
1982 spin_unlock_bh(&css_set_lock);
1983
1984 BUG_ON(!list_empty(&root_cgrp->self.children));
1985 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1986
1987 kernfs_activate(root_cgrp->kn);
1988 ret = 0;
1989 goto out;
1990
1991destroy_root:
1992 kernfs_destroy_root(root->kf_root);
1993 root->kf_root = NULL;
1994exit_root_id:
1995 cgroup_exit_root_id(root);
1996cancel_ref:
1997 percpu_ref_exit(&root_cgrp->self.refcnt);
1998out:
1999 free_cgrp_cset_links(&tmp_links);
2000 return ret;
2001}
2002
2003static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2004 int flags, const char *unused_dev_name,
2005 void *data)
2006{
2007 struct super_block *pinned_sb = NULL;
2008 struct cgroup_subsys *ss;
2009 struct cgroup_root *root;
2010 struct cgroup_sb_opts opts;
2011 struct dentry *dentry;
2012 int ret;
2013 int i;
2014 bool new_sb;
2015
2016 /*
2017 * The first time anyone tries to mount a cgroup, enable the list
2018 * linking each css_set to its tasks and fix up all existing tasks.
2019 */
2020 if (!use_task_css_set_links)
2021 cgroup_enable_task_cg_lists();
2022
2023 mutex_lock(&cgroup_mutex);
2024
2025 /* First find the desired set of subsystems */
2026 ret = parse_cgroupfs_options(data, &opts);
2027 if (ret)
2028 goto out_unlock;
2029
2030 /* look for a matching existing root */
2031 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
2032 cgrp_dfl_root_visible = true;
2033 root = &cgrp_dfl_root;
2034 cgroup_get(&root->cgrp);
2035 ret = 0;
2036 goto out_unlock;
2037 }
2038
2039 /*
2040 * Destruction of cgroup root is asynchronous, so subsystems may
2041 * still be dying after the previous unmount. Let's drain the
2042 * dying subsystems. We just need to ensure that the ones
2043 * unmounted previously finish dying and don't care about new ones
2044 * starting. Testing ref liveliness is good enough.
2045 */
2046 for_each_subsys(ss, i) {
2047 if (!(opts.subsys_mask & (1 << i)) ||
2048 ss->root == &cgrp_dfl_root)
2049 continue;
2050
2051 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2052 mutex_unlock(&cgroup_mutex);
2053 msleep(10);
2054 ret = restart_syscall();
2055 goto out_free;
2056 }
2057 cgroup_put(&ss->root->cgrp);
2058 }
2059
2060 for_each_root(root) {
2061 bool name_match = false;
2062
2063 if (root == &cgrp_dfl_root)
2064 continue;
2065
2066 /*
2067 * If we asked for a name then it must match. Also, if
2068 * name matches but sybsys_mask doesn't, we should fail.
2069 * Remember whether name matched.
2070 */
2071 if (opts.name) {
2072 if (strcmp(opts.name, root->name))
2073 continue;
2074 name_match = true;
2075 }
2076
2077 /*
2078 * If we asked for subsystems (or explicitly for no
2079 * subsystems) then they must match.
2080 */
2081 if ((opts.subsys_mask || opts.none) &&
2082 (opts.subsys_mask != root->subsys_mask)) {
2083 if (!name_match)
2084 continue;
2085 ret = -EBUSY;
2086 goto out_unlock;
2087 }
2088
2089 if (root->flags ^ opts.flags)
2090 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2091
2092 /*
2093 * We want to reuse @root whose lifetime is governed by its
2094 * ->cgrp. Let's check whether @root is alive and keep it
2095 * that way. As cgroup_kill_sb() can happen anytime, we
2096 * want to block it by pinning the sb so that @root doesn't
2097 * get killed before mount is complete.
2098 *
2099 * With the sb pinned, tryget_live can reliably indicate
2100 * whether @root can be reused. If it's being killed,
2101 * drain it. We can use wait_queue for the wait but this
2102 * path is super cold. Let's just sleep a bit and retry.
2103 */
2104 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2105 if (IS_ERR(pinned_sb) ||
2106 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2107 mutex_unlock(&cgroup_mutex);
2108 if (!IS_ERR_OR_NULL(pinned_sb))
2109 deactivate_super(pinned_sb);
2110 msleep(10);
2111 ret = restart_syscall();
2112 goto out_free;
2113 }
2114
2115 ret = 0;
2116 goto out_unlock;
2117 }
2118
2119 /*
2120 * No such thing, create a new one. name= matching without subsys
2121 * specification is allowed for already existing hierarchies but we
2122 * can't create new one without subsys specification.
2123 */
2124 if (!opts.subsys_mask && !opts.none) {
2125 ret = -EINVAL;
2126 goto out_unlock;
2127 }
2128
2129 root = kzalloc(sizeof(*root), GFP_KERNEL);
2130 if (!root) {
2131 ret = -ENOMEM;
2132 goto out_unlock;
2133 }
2134
2135 init_cgroup_root(root, &opts);
2136
2137 ret = cgroup_setup_root(root, opts.subsys_mask);
2138 if (ret)
2139 cgroup_free_root(root);
2140
2141out_unlock:
2142 mutex_unlock(&cgroup_mutex);
2143out_free:
2144 kfree(opts.release_agent);
2145 kfree(opts.name);
2146
2147 if (ret)
2148 return ERR_PTR(ret);
2149
2150 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2151 CGROUP_SUPER_MAGIC, &new_sb);
2152 if (IS_ERR(dentry) || !new_sb)
2153 cgroup_put(&root->cgrp);
2154
2155 /*
2156 * If @pinned_sb, we're reusing an existing root and holding an
2157 * extra ref on its sb. Mount is complete. Put the extra ref.
2158 */
2159 if (pinned_sb) {
2160 WARN_ON(new_sb);
2161 deactivate_super(pinned_sb);
2162 }
2163
2164 return dentry;
2165}
2166
2167static void cgroup_kill_sb(struct super_block *sb)
2168{
2169 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2170 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2171
2172 /*
2173 * If @root doesn't have any mounts or children, start killing it.
2174 * This prevents new mounts by disabling percpu_ref_tryget_live().
2175 * cgroup_mount() may wait for @root's release.
2176 *
2177 * And don't kill the default root.
2178 */
2179 if (!list_empty(&root->cgrp.self.children) ||
2180 root == &cgrp_dfl_root)
2181 cgroup_put(&root->cgrp);
2182 else
2183 percpu_ref_kill(&root->cgrp.self.refcnt);
2184
2185 kernfs_kill_sb(sb);
2186}
2187
2188static struct file_system_type cgroup_fs_type = {
2189 .name = "cgroup",
2190 .mount = cgroup_mount,
2191 .kill_sb = cgroup_kill_sb,
2192};
2193
2194/**
2195 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2196 * @task: target task
2197 * @buf: the buffer to write the path into
2198 * @buflen: the length of the buffer
2199 *
2200 * Determine @task's cgroup on the first (the one with the lowest non-zero
2201 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2202 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2203 * cgroup controller callbacks.
2204 *
2205 * Return value is the same as kernfs_path().
2206 */
2207char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2208{
2209 struct cgroup_root *root;
2210 struct cgroup *cgrp;
2211 int hierarchy_id = 1;
2212 char *path = NULL;
2213
2214 mutex_lock(&cgroup_mutex);
2215 spin_lock_bh(&css_set_lock);
2216
2217 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2218
2219 if (root) {
2220 cgrp = task_cgroup_from_root(task, root);
2221 path = cgroup_path(cgrp, buf, buflen);
2222 } else {
2223 /* if no hierarchy exists, everyone is in "/" */
2224 if (strlcpy(buf, "/", buflen) < buflen)
2225 path = buf;
2226 }
2227
2228 spin_unlock_bh(&css_set_lock);
2229 mutex_unlock(&cgroup_mutex);
2230 return path;
2231}
2232EXPORT_SYMBOL_GPL(task_cgroup_path);
2233
2234/* used to track tasks and other necessary states during migration */
2235struct cgroup_taskset {
2236 /* the src and dst cset list running through cset->mg_node */
2237 struct list_head src_csets;
2238 struct list_head dst_csets;
2239
2240 /* the subsys currently being processed */
2241 int ssid;
2242
2243 /*
2244 * Fields for cgroup_taskset_*() iteration.
2245 *
2246 * Before migration is committed, the target migration tasks are on
2247 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2248 * the csets on ->dst_csets. ->csets point to either ->src_csets
2249 * or ->dst_csets depending on whether migration is committed.
2250 *
2251 * ->cur_csets and ->cur_task point to the current task position
2252 * during iteration.
2253 */
2254 struct list_head *csets;
2255 struct css_set *cur_cset;
2256 struct task_struct *cur_task;
2257};
2258
2259#define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2260 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2261 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2262 .csets = &tset.src_csets, \
2263}
2264
2265/**
2266 * cgroup_taskset_add - try to add a migration target task to a taskset
2267 * @task: target task
2268 * @tset: target taskset
2269 *
2270 * Add @task, which is a migration target, to @tset. This function becomes
2271 * noop if @task doesn't need to be migrated. @task's css_set should have
2272 * been added as a migration source and @task->cg_list will be moved from
2273 * the css_set's tasks list to mg_tasks one.
2274 */
2275static void cgroup_taskset_add(struct task_struct *task,
2276 struct cgroup_taskset *tset)
2277{
2278 struct css_set *cset;
2279
2280 lockdep_assert_held(&css_set_lock);
2281
2282 /* @task either already exited or can't exit until the end */
2283 if (task->flags & PF_EXITING)
2284 return;
2285
2286 /* leave @task alone if post_fork() hasn't linked it yet */
2287 if (list_empty(&task->cg_list))
2288 return;
2289
2290 cset = task_css_set(task);
2291 if (!cset->mg_src_cgrp)
2292 return;
2293
2294 list_move_tail(&task->cg_list, &cset->mg_tasks);
2295 if (list_empty(&cset->mg_node))
2296 list_add_tail(&cset->mg_node, &tset->src_csets);
2297 if (list_empty(&cset->mg_dst_cset->mg_node))
2298 list_move_tail(&cset->mg_dst_cset->mg_node,
2299 &tset->dst_csets);
2300}
2301
2302/**
2303 * cgroup_taskset_first - reset taskset and return the first task
2304 * @tset: taskset of interest
2305 * @dst_cssp: output variable for the destination css
2306 *
2307 * @tset iteration is initialized and the first task is returned.
2308 */
2309struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2310 struct cgroup_subsys_state **dst_cssp)
2311{
2312 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2313 tset->cur_task = NULL;
2314
2315 return cgroup_taskset_next(tset, dst_cssp);
2316}
2317
2318/**
2319 * cgroup_taskset_next - iterate to the next task in taskset
2320 * @tset: taskset of interest
2321 * @dst_cssp: output variable for the destination css
2322 *
2323 * Return the next task in @tset. Iteration must have been initialized
2324 * with cgroup_taskset_first().
2325 */
2326struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2327 struct cgroup_subsys_state **dst_cssp)
2328{
2329 struct css_set *cset = tset->cur_cset;
2330 struct task_struct *task = tset->cur_task;
2331
2332 while (&cset->mg_node != tset->csets) {
2333 if (!task)
2334 task = list_first_entry(&cset->mg_tasks,
2335 struct task_struct, cg_list);
2336 else
2337 task = list_next_entry(task, cg_list);
2338
2339 if (&task->cg_list != &cset->mg_tasks) {
2340 tset->cur_cset = cset;
2341 tset->cur_task = task;
2342
2343 /*
2344 * This function may be called both before and
2345 * after cgroup_taskset_migrate(). The two cases
2346 * can be distinguished by looking at whether @cset
2347 * has its ->mg_dst_cset set.
2348 */
2349 if (cset->mg_dst_cset)
2350 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2351 else
2352 *dst_cssp = cset->subsys[tset->ssid];
2353
2354 return task;
2355 }
2356
2357 cset = list_next_entry(cset, mg_node);
2358 task = NULL;
2359 }
2360
2361 return NULL;
2362}
2363
2364/**
2365 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2366 * @tset: taget taskset
2367 * @dst_cgrp: destination cgroup
2368 *
2369 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2370 * ->can_attach callbacks fails and guarantees that either all or none of
2371 * the tasks in @tset are migrated. @tset is consumed regardless of
2372 * success.
2373 */
2374static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2375 struct cgroup *dst_cgrp)
2376{
2377 struct cgroup_subsys_state *css, *failed_css = NULL;
2378 struct task_struct *task, *tmp_task;
2379 struct css_set *cset, *tmp_cset;
2380 int i, ret;
2381
2382 /* methods shouldn't be called if no task is actually migrating */
2383 if (list_empty(&tset->src_csets))
2384 return 0;
2385
2386 /* check that we can legitimately attach to the cgroup */
2387 for_each_e_css(css, i, dst_cgrp) {
2388 if (css->ss->can_attach) {
2389 tset->ssid = i;
2390 ret = css->ss->can_attach(tset);
2391 if (ret) {
2392 failed_css = css;
2393 goto out_cancel_attach;
2394 }
2395 }
2396 }
2397
2398 /*
2399 * Now that we're guaranteed success, proceed to move all tasks to
2400 * the new cgroup. There are no failure cases after here, so this
2401 * is the commit point.
2402 */
2403 spin_lock_bh(&css_set_lock);
2404 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2405 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2406 struct css_set *from_cset = task_css_set(task);
2407 struct css_set *to_cset = cset->mg_dst_cset;
2408
2409 get_css_set(to_cset);
2410 css_set_move_task(task, from_cset, to_cset, true);
2411 put_css_set_locked(from_cset);
2412 }
2413 }
2414 spin_unlock_bh(&css_set_lock);
2415
2416 /*
2417 * Migration is committed, all target tasks are now on dst_csets.
2418 * Nothing is sensitive to fork() after this point. Notify
2419 * controllers that migration is complete.
2420 */
2421 tset->csets = &tset->dst_csets;
2422
2423 for_each_e_css(css, i, dst_cgrp) {
2424 if (css->ss->attach) {
2425 tset->ssid = i;
2426 css->ss->attach(tset);
2427 }
2428 }
2429
2430 ret = 0;
2431 goto out_release_tset;
2432
2433out_cancel_attach:
2434 for_each_e_css(css, i, dst_cgrp) {
2435 if (css == failed_css)
2436 break;
2437 if (css->ss->cancel_attach) {
2438 tset->ssid = i;
2439 css->ss->cancel_attach(tset);
2440 }
2441 }
2442out_release_tset:
2443 spin_lock_bh(&css_set_lock);
2444 list_splice_init(&tset->dst_csets, &tset->src_csets);
2445 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2446 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2447 list_del_init(&cset->mg_node);
2448 }
2449 spin_unlock_bh(&css_set_lock);
2450 return ret;
2451}
2452
2453/**
2454 * cgroup_migrate_finish - cleanup after attach
2455 * @preloaded_csets: list of preloaded css_sets
2456 *
2457 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2458 * those functions for details.
2459 */
2460static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2461{
2462 struct css_set *cset, *tmp_cset;
2463
2464 lockdep_assert_held(&cgroup_mutex);
2465
2466 spin_lock_bh(&css_set_lock);
2467 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2468 cset->mg_src_cgrp = NULL;
2469 cset->mg_dst_cset = NULL;
2470 list_del_init(&cset->mg_preload_node);
2471 put_css_set_locked(cset);
2472 }
2473 spin_unlock_bh(&css_set_lock);
2474}
2475
2476/**
2477 * cgroup_migrate_add_src - add a migration source css_set
2478 * @src_cset: the source css_set to add
2479 * @dst_cgrp: the destination cgroup
2480 * @preloaded_csets: list of preloaded css_sets
2481 *
2482 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2483 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2484 * up by cgroup_migrate_finish().
2485 *
2486 * This function may be called without holding cgroup_threadgroup_rwsem
2487 * even if the target is a process. Threads may be created and destroyed
2488 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2489 * into play and the preloaded css_sets are guaranteed to cover all
2490 * migrations.
2491 */
2492static void cgroup_migrate_add_src(struct css_set *src_cset,
2493 struct cgroup *dst_cgrp,
2494 struct list_head *preloaded_csets)
2495{
2496 struct cgroup *src_cgrp;
2497
2498 lockdep_assert_held(&cgroup_mutex);
2499 lockdep_assert_held(&css_set_lock);
2500
2501 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2502
2503 if (!list_empty(&src_cset->mg_preload_node))
2504 return;
2505
2506 WARN_ON(src_cset->mg_src_cgrp);
2507 WARN_ON(!list_empty(&src_cset->mg_tasks));
2508 WARN_ON(!list_empty(&src_cset->mg_node));
2509
2510 src_cset->mg_src_cgrp = src_cgrp;
2511 get_css_set(src_cset);
2512 list_add(&src_cset->mg_preload_node, preloaded_csets);
2513}
2514
2515/**
2516 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2517 * @dst_cgrp: the destination cgroup (may be %NULL)
2518 * @preloaded_csets: list of preloaded source css_sets
2519 *
2520 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2521 * have been preloaded to @preloaded_csets. This function looks up and
2522 * pins all destination css_sets, links each to its source, and append them
2523 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2524 * source css_set is assumed to be its cgroup on the default hierarchy.
2525 *
2526 * This function must be called after cgroup_migrate_add_src() has been
2527 * called on each migration source css_set. After migration is performed
2528 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2529 * @preloaded_csets.
2530 */
2531static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2532 struct list_head *preloaded_csets)
2533{
2534 LIST_HEAD(csets);
2535 struct css_set *src_cset, *tmp_cset;
2536
2537 lockdep_assert_held(&cgroup_mutex);
2538
2539 /*
2540 * Except for the root, child_subsys_mask must be zero for a cgroup
2541 * with tasks so that child cgroups don't compete against tasks.
2542 */
2543 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2544 dst_cgrp->child_subsys_mask)
2545 return -EBUSY;
2546
2547 /* look up the dst cset for each src cset and link it to src */
2548 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2549 struct css_set *dst_cset;
2550
2551 dst_cset = find_css_set(src_cset,
2552 dst_cgrp ?: src_cset->dfl_cgrp);
2553 if (!dst_cset)
2554 goto err;
2555
2556 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2557
2558 /*
2559 * If src cset equals dst, it's noop. Drop the src.
2560 * cgroup_migrate() will skip the cset too. Note that we
2561 * can't handle src == dst as some nodes are used by both.
2562 */
2563 if (src_cset == dst_cset) {
2564 src_cset->mg_src_cgrp = NULL;
2565 list_del_init(&src_cset->mg_preload_node);
2566 put_css_set(src_cset);
2567 put_css_set(dst_cset);
2568 continue;
2569 }
2570
2571 src_cset->mg_dst_cset = dst_cset;
2572
2573 if (list_empty(&dst_cset->mg_preload_node))
2574 list_add(&dst_cset->mg_preload_node, &csets);
2575 else
2576 put_css_set(dst_cset);
2577 }
2578
2579 list_splice_tail(&csets, preloaded_csets);
2580 return 0;
2581err:
2582 cgroup_migrate_finish(&csets);
2583 return -ENOMEM;
2584}
2585
2586/**
2587 * cgroup_migrate - migrate a process or task to a cgroup
2588 * @leader: the leader of the process or the task to migrate
2589 * @threadgroup: whether @leader points to the whole process or a single task
2590 * @cgrp: the destination cgroup
2591 *
2592 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2593 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2594 * caller is also responsible for invoking cgroup_migrate_add_src() and
2595 * cgroup_migrate_prepare_dst() on the targets before invoking this
2596 * function and following up with cgroup_migrate_finish().
2597 *
2598 * As long as a controller's ->can_attach() doesn't fail, this function is
2599 * guaranteed to succeed. This means that, excluding ->can_attach()
2600 * failure, when migrating multiple targets, the success or failure can be
2601 * decided for all targets by invoking group_migrate_prepare_dst() before
2602 * actually starting migrating.
2603 */
2604static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2605 struct cgroup *cgrp)
2606{
2607 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2608 struct task_struct *task;
2609
2610 /*
2611 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2612 * already PF_EXITING could be freed from underneath us unless we
2613 * take an rcu_read_lock.
2614 */
2615 spin_lock_bh(&css_set_lock);
2616 rcu_read_lock();
2617 task = leader;
2618 do {
2619 cgroup_taskset_add(task, &tset);
2620 if (!threadgroup)
2621 break;
2622 } while_each_thread(leader, task);
2623 rcu_read_unlock();
2624 spin_unlock_bh(&css_set_lock);
2625
2626 return cgroup_taskset_migrate(&tset, cgrp);
2627}
2628
2629/**
2630 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2631 * @dst_cgrp: the cgroup to attach to
2632 * @leader: the task or the leader of the threadgroup to be attached
2633 * @threadgroup: attach the whole threadgroup?
2634 *
2635 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2636 */
2637static int cgroup_attach_task(struct cgroup *dst_cgrp,
2638 struct task_struct *leader, bool threadgroup)
2639{
2640 LIST_HEAD(preloaded_csets);
2641 struct task_struct *task;
2642 int ret;
2643
2644 /* look up all src csets */
2645 spin_lock_bh(&css_set_lock);
2646 rcu_read_lock();
2647 task = leader;
2648 do {
2649 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2650 &preloaded_csets);
2651 if (!threadgroup)
2652 break;
2653 } while_each_thread(leader, task);
2654 rcu_read_unlock();
2655 spin_unlock_bh(&css_set_lock);
2656
2657 /* prepare dst csets and commit */
2658 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2659 if (!ret)
2660 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2661
2662 cgroup_migrate_finish(&preloaded_csets);
2663 return ret;
2664}
2665
2666static int cgroup_procs_write_permission(struct task_struct *task,
2667 struct cgroup *dst_cgrp,
2668 struct kernfs_open_file *of)
2669{
2670 const struct cred *cred = current_cred();
2671 const struct cred *tcred = get_task_cred(task);
2672 int ret = 0;
2673
2674 /*
2675 * even if we're attaching all tasks in the thread group, we only
2676 * need to check permissions on one of them.
2677 */
2678 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2679 !uid_eq(cred->euid, tcred->uid) &&
2680 !uid_eq(cred->euid, tcred->suid))
2681 ret = -EACCES;
2682
2683 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2684 struct super_block *sb = of->file->f_path.dentry->d_sb;
2685 struct cgroup *cgrp;
2686 struct inode *inode;
2687
2688 spin_lock_bh(&css_set_lock);
2689 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2690 spin_unlock_bh(&css_set_lock);
2691
2692 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2693 cgrp = cgroup_parent(cgrp);
2694
2695 ret = -ENOMEM;
2696 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2697 if (inode) {
2698 ret = inode_permission(inode, MAY_WRITE);
2699 iput(inode);
2700 }
2701 }
2702
2703 put_cred(tcred);
2704 return ret;
2705}
2706
2707/*
2708 * Find the task_struct of the task to attach by vpid and pass it along to the
2709 * function to attach either it or all tasks in its threadgroup. Will lock
2710 * cgroup_mutex and threadgroup.
2711 */
2712static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2713 size_t nbytes, loff_t off, bool threadgroup)
2714{
2715 struct task_struct *tsk;
2716 struct cgroup *cgrp;
2717 pid_t pid;
2718 int ret;
2719
2720 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2721 return -EINVAL;
2722
2723 cgrp = cgroup_kn_lock_live(of->kn);
2724 if (!cgrp)
2725 return -ENODEV;
2726
2727 percpu_down_write(&cgroup_threadgroup_rwsem);
2728 rcu_read_lock();
2729 if (pid) {
2730 tsk = find_task_by_vpid(pid);
2731 if (!tsk) {
2732 ret = -ESRCH;
2733 goto out_unlock_rcu;
2734 }
2735 } else {
2736 tsk = current;
2737 }
2738
2739 if (threadgroup)
2740 tsk = tsk->group_leader;
2741
2742 /*
2743 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2744 * trapped in a cpuset, or RT worker may be born in a cgroup
2745 * with no rt_runtime allocated. Just say no.
2746 */
2747 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2748 ret = -EINVAL;
2749 goto out_unlock_rcu;
2750 }
2751
2752 get_task_struct(tsk);
2753 rcu_read_unlock();
2754
2755 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2756 if (!ret)
2757 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2758
2759 put_task_struct(tsk);
2760 goto out_unlock_threadgroup;
2761
2762out_unlock_rcu:
2763 rcu_read_unlock();
2764out_unlock_threadgroup:
2765 percpu_up_write(&cgroup_threadgroup_rwsem);
2766 cgroup_kn_unlock(of->kn);
2767 return ret ?: nbytes;
2768}
2769
2770/**
2771 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2772 * @from: attach to all cgroups of a given task
2773 * @tsk: the task to be attached
2774 */
2775int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2776{
2777 struct cgroup_root *root;
2778 int retval = 0;
2779
2780 mutex_lock(&cgroup_mutex);
2781 for_each_root(root) {
2782 struct cgroup *from_cgrp;
2783
2784 if (root == &cgrp_dfl_root)
2785 continue;
2786
2787 spin_lock_bh(&css_set_lock);
2788 from_cgrp = task_cgroup_from_root(from, root);
2789 spin_unlock_bh(&css_set_lock);
2790
2791 retval = cgroup_attach_task(from_cgrp, tsk, false);
2792 if (retval)
2793 break;
2794 }
2795 mutex_unlock(&cgroup_mutex);
2796
2797 return retval;
2798}
2799EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2800
2801static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2802 char *buf, size_t nbytes, loff_t off)
2803{
2804 return __cgroup_procs_write(of, buf, nbytes, off, false);
2805}
2806
2807static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2808 char *buf, size_t nbytes, loff_t off)
2809{
2810 return __cgroup_procs_write(of, buf, nbytes, off, true);
2811}
2812
2813static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2814 char *buf, size_t nbytes, loff_t off)
2815{
2816 struct cgroup *cgrp;
2817
2818 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2819
2820 cgrp = cgroup_kn_lock_live(of->kn);
2821 if (!cgrp)
2822 return -ENODEV;
2823 spin_lock(&release_agent_path_lock);
2824 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2825 sizeof(cgrp->root->release_agent_path));
2826 spin_unlock(&release_agent_path_lock);
2827 cgroup_kn_unlock(of->kn);
2828 return nbytes;
2829}
2830
2831static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2832{
2833 struct cgroup *cgrp = seq_css(seq)->cgroup;
2834
2835 spin_lock(&release_agent_path_lock);
2836 seq_puts(seq, cgrp->root->release_agent_path);
2837 spin_unlock(&release_agent_path_lock);
2838 seq_putc(seq, '\n');
2839 return 0;
2840}
2841
2842static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2843{
2844 seq_puts(seq, "0\n");
2845 return 0;
2846}
2847
2848static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2849{
2850 struct cgroup_subsys *ss;
2851 bool printed = false;
2852 int ssid;
2853
2854 for_each_subsys_which(ss, ssid, &ss_mask) {
2855 if (printed)
2856 seq_putc(seq, ' ');
2857 seq_printf(seq, "%s", ss->name);
2858 printed = true;
2859 }
2860 if (printed)
2861 seq_putc(seq, '\n');
2862}
2863
2864/* show controllers which are currently attached to the default hierarchy */
2865static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2866{
2867 struct cgroup *cgrp = seq_css(seq)->cgroup;
2868
2869 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2870 ~cgrp_dfl_root_inhibit_ss_mask);
2871 return 0;
2872}
2873
2874/* show controllers which are enabled from the parent */
2875static int cgroup_controllers_show(struct seq_file *seq, void *v)
2876{
2877 struct cgroup *cgrp = seq_css(seq)->cgroup;
2878
2879 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2880 return 0;
2881}
2882
2883/* show controllers which are enabled for a given cgroup's children */
2884static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2885{
2886 struct cgroup *cgrp = seq_css(seq)->cgroup;
2887
2888 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2889 return 0;
2890}
2891
2892/**
2893 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2894 * @cgrp: root of the subtree to update csses for
2895 *
2896 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2897 * css associations need to be updated accordingly. This function looks up
2898 * all css_sets which are attached to the subtree, creates the matching
2899 * updated css_sets and migrates the tasks to the new ones.
2900 */
2901static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2902{
2903 LIST_HEAD(preloaded_csets);
2904 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2905 struct cgroup_subsys_state *css;
2906 struct css_set *src_cset;
2907 int ret;
2908
2909 lockdep_assert_held(&cgroup_mutex);
2910
2911 percpu_down_write(&cgroup_threadgroup_rwsem);
2912
2913 /* look up all csses currently attached to @cgrp's subtree */
2914 spin_lock_bh(&css_set_lock);
2915 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2916 struct cgrp_cset_link *link;
2917
2918 /* self is not affected by child_subsys_mask change */
2919 if (css->cgroup == cgrp)
2920 continue;
2921
2922 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2923 cgroup_migrate_add_src(link->cset, cgrp,
2924 &preloaded_csets);
2925 }
2926 spin_unlock_bh(&css_set_lock);
2927
2928 /* NULL dst indicates self on default hierarchy */
2929 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2930 if (ret)
2931 goto out_finish;
2932
2933 spin_lock_bh(&css_set_lock);
2934 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2935 struct task_struct *task, *ntask;
2936
2937 /* src_csets precede dst_csets, break on the first dst_cset */
2938 if (!src_cset->mg_src_cgrp)
2939 break;
2940
2941 /* all tasks in src_csets need to be migrated */
2942 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2943 cgroup_taskset_add(task, &tset);
2944 }
2945 spin_unlock_bh(&css_set_lock);
2946
2947 ret = cgroup_taskset_migrate(&tset, cgrp);
2948out_finish:
2949 cgroup_migrate_finish(&preloaded_csets);
2950 percpu_up_write(&cgroup_threadgroup_rwsem);
2951 return ret;
2952}
2953
2954/* change the enabled child controllers for a cgroup in the default hierarchy */
2955static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2956 char *buf, size_t nbytes,
2957 loff_t off)
2958{
2959 unsigned long enable = 0, disable = 0;
2960 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2961 struct cgroup *cgrp, *child;
2962 struct cgroup_subsys *ss;
2963 char *tok;
2964 int ssid, ret;
2965
2966 /*
2967 * Parse input - space separated list of subsystem names prefixed
2968 * with either + or -.
2969 */
2970 buf = strstrip(buf);
2971 while ((tok = strsep(&buf, " "))) {
2972 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2973
2974 if (tok[0] == '\0')
2975 continue;
2976 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2977 if (!cgroup_ssid_enabled(ssid) ||
2978 strcmp(tok + 1, ss->name))
2979 continue;
2980
2981 if (*tok == '+') {
2982 enable |= 1 << ssid;
2983 disable &= ~(1 << ssid);
2984 } else if (*tok == '-') {
2985 disable |= 1 << ssid;
2986 enable &= ~(1 << ssid);
2987 } else {
2988 return -EINVAL;
2989 }
2990 break;
2991 }
2992 if (ssid == CGROUP_SUBSYS_COUNT)
2993 return -EINVAL;
2994 }
2995
2996 cgrp = cgroup_kn_lock_live(of->kn);
2997 if (!cgrp)
2998 return -ENODEV;
2999
3000 for_each_subsys(ss, ssid) {
3001 if (enable & (1 << ssid)) {
3002 if (cgrp->subtree_control & (1 << ssid)) {
3003 enable &= ~(1 << ssid);
3004 continue;
3005 }
3006
3007 /* unavailable or not enabled on the parent? */
3008 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
3009 (cgroup_parent(cgrp) &&
3010 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
3011 ret = -ENOENT;
3012 goto out_unlock;
3013 }
3014 } else if (disable & (1 << ssid)) {
3015 if (!(cgrp->subtree_control & (1 << ssid))) {
3016 disable &= ~(1 << ssid);
3017 continue;
3018 }
3019
3020 /* a child has it enabled? */
3021 cgroup_for_each_live_child(child, cgrp) {
3022 if (child->subtree_control & (1 << ssid)) {
3023 ret = -EBUSY;
3024 goto out_unlock;
3025 }
3026 }
3027 }
3028 }
3029
3030 if (!enable && !disable) {
3031 ret = 0;
3032 goto out_unlock;
3033 }
3034
3035 /*
3036 * Except for the root, subtree_control must be zero for a cgroup
3037 * with tasks so that child cgroups don't compete against tasks.
3038 */
3039 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3040 ret = -EBUSY;
3041 goto out_unlock;
3042 }
3043
3044 /*
3045 * Update subsys masks and calculate what needs to be done. More
3046 * subsystems than specified may need to be enabled or disabled
3047 * depending on subsystem dependencies.
3048 */
3049 old_sc = cgrp->subtree_control;
3050 old_ss = cgrp->child_subsys_mask;
3051 new_sc = (old_sc | enable) & ~disable;
3052 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
3053
3054 css_enable = ~old_ss & new_ss;
3055 css_disable = old_ss & ~new_ss;
3056 enable |= css_enable;
3057 disable |= css_disable;
3058
3059 /*
3060 * Because css offlining is asynchronous, userland might try to
3061 * re-enable the same controller while the previous instance is
3062 * still around. In such cases, wait till it's gone using
3063 * offline_waitq.
3064 */
3065 for_each_subsys_which(ss, ssid, &css_enable) {
3066 cgroup_for_each_live_child(child, cgrp) {
3067 DEFINE_WAIT(wait);
3068
3069 if (!cgroup_css(child, ss))
3070 continue;
3071
3072 cgroup_get(child);
3073 prepare_to_wait(&child->offline_waitq, &wait,
3074 TASK_UNINTERRUPTIBLE);
3075 cgroup_kn_unlock(of->kn);
3076 schedule();
3077 finish_wait(&child->offline_waitq, &wait);
3078 cgroup_put(child);
3079
3080 return restart_syscall();
3081 }
3082 }
3083
3084 cgrp->subtree_control = new_sc;
3085 cgrp->child_subsys_mask = new_ss;
3086
3087 /*
3088 * Create new csses or make the existing ones visible. A css is
3089 * created invisible if it's being implicitly enabled through
3090 * dependency. An invisible css is made visible when the userland
3091 * explicitly enables it.
3092 */
3093 for_each_subsys(ss, ssid) {
3094 if (!(enable & (1 << ssid)))
3095 continue;
3096
3097 cgroup_for_each_live_child(child, cgrp) {
3098 if (css_enable & (1 << ssid))
3099 ret = create_css(child, ss,
3100 cgrp->subtree_control & (1 << ssid));
3101 else
3102 ret = css_populate_dir(cgroup_css(child, ss),
3103 NULL);
3104 if (ret)
3105 goto err_undo_css;
3106 }
3107 }
3108
3109 /*
3110 * At this point, cgroup_e_css() results reflect the new csses
3111 * making the following cgroup_update_dfl_csses() properly update
3112 * css associations of all tasks in the subtree.
3113 */
3114 ret = cgroup_update_dfl_csses(cgrp);
3115 if (ret)
3116 goto err_undo_css;
3117
3118 /*
3119 * All tasks are migrated out of disabled csses. Kill or hide
3120 * them. A css is hidden when the userland requests it to be
3121 * disabled while other subsystems are still depending on it. The
3122 * css must not actively control resources and be in the vanilla
3123 * state if it's made visible again later. Controllers which may
3124 * be depended upon should provide ->css_reset() for this purpose.
3125 */
3126 for_each_subsys(ss, ssid) {
3127 if (!(disable & (1 << ssid)))
3128 continue;
3129
3130 cgroup_for_each_live_child(child, cgrp) {
3131 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3132
3133 if (css_disable & (1 << ssid)) {
3134 kill_css(css);
3135 } else {
3136 css_clear_dir(css, NULL);
3137 if (ss->css_reset)
3138 ss->css_reset(css);
3139 }
3140 }
3141 }
3142
3143 /*
3144 * The effective csses of all the descendants (excluding @cgrp) may
3145 * have changed. Subsystems can optionally subscribe to this event
3146 * by implementing ->css_e_css_changed() which is invoked if any of
3147 * the effective csses seen from the css's cgroup may have changed.
3148 */
3149 for_each_subsys(ss, ssid) {
3150 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
3151 struct cgroup_subsys_state *css;
3152
3153 if (!ss->css_e_css_changed || !this_css)
3154 continue;
3155
3156 css_for_each_descendant_pre(css, this_css)
3157 if (css != this_css)
3158 ss->css_e_css_changed(css);
3159 }
3160
3161 kernfs_activate(cgrp->kn);
3162 ret = 0;
3163out_unlock:
3164 cgroup_kn_unlock(of->kn);
3165 return ret ?: nbytes;
3166
3167err_undo_css:
3168 cgrp->subtree_control = old_sc;
3169 cgrp->child_subsys_mask = old_ss;
3170
3171 for_each_subsys(ss, ssid) {
3172 if (!(enable & (1 << ssid)))
3173 continue;
3174
3175 cgroup_for_each_live_child(child, cgrp) {
3176 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3177
3178 if (!css)
3179 continue;
3180
3181 if (css_enable & (1 << ssid))
3182 kill_css(css);
3183 else
3184 css_clear_dir(css, NULL);
3185 }
3186 }
3187 goto out_unlock;
3188}
3189
3190static int cgroup_events_show(struct seq_file *seq, void *v)
3191{
3192 seq_printf(seq, "populated %d\n",
3193 cgroup_is_populated(seq_css(seq)->cgroup));
3194 return 0;
3195}
3196
3197static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3198 size_t nbytes, loff_t off)
3199{
3200 struct cgroup *cgrp = of->kn->parent->priv;
3201 struct cftype *cft = of->kn->priv;
3202 struct cgroup_subsys_state *css;
3203 int ret;
3204
3205 if (cft->write)
3206 return cft->write(of, buf, nbytes, off);
3207
3208 /*
3209 * kernfs guarantees that a file isn't deleted with operations in
3210 * flight, which means that the matching css is and stays alive and
3211 * doesn't need to be pinned. The RCU locking is not necessary
3212 * either. It's just for the convenience of using cgroup_css().
3213 */
3214 rcu_read_lock();
3215 css = cgroup_css(cgrp, cft->ss);
3216 rcu_read_unlock();
3217
3218 if (cft->write_u64) {
3219 unsigned long long v;
3220 ret = kstrtoull(buf, 0, &v);
3221 if (!ret)
3222 ret = cft->write_u64(css, cft, v);
3223 } else if (cft->write_s64) {
3224 long long v;
3225 ret = kstrtoll(buf, 0, &v);
3226 if (!ret)
3227 ret = cft->write_s64(css, cft, v);
3228 } else {
3229 ret = -EINVAL;
3230 }
3231
3232 return ret ?: nbytes;
3233}
3234
3235static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3236{
3237 return seq_cft(seq)->seq_start(seq, ppos);
3238}
3239
3240static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3241{
3242 return seq_cft(seq)->seq_next(seq, v, ppos);
3243}
3244
3245static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3246{
3247 seq_cft(seq)->seq_stop(seq, v);
3248}
3249
3250static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3251{
3252 struct cftype *cft = seq_cft(m);
3253 struct cgroup_subsys_state *css = seq_css(m);
3254
3255 if (cft->seq_show)
3256 return cft->seq_show(m, arg);
3257
3258 if (cft->read_u64)
3259 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3260 else if (cft->read_s64)
3261 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3262 else
3263 return -EINVAL;
3264 return 0;
3265}
3266
3267static struct kernfs_ops cgroup_kf_single_ops = {
3268 .atomic_write_len = PAGE_SIZE,
3269 .write = cgroup_file_write,
3270 .seq_show = cgroup_seqfile_show,
3271};
3272
3273static struct kernfs_ops cgroup_kf_ops = {
3274 .atomic_write_len = PAGE_SIZE,
3275 .write = cgroup_file_write,
3276 .seq_start = cgroup_seqfile_start,
3277 .seq_next = cgroup_seqfile_next,
3278 .seq_stop = cgroup_seqfile_stop,
3279 .seq_show = cgroup_seqfile_show,
3280};
3281
3282/*
3283 * cgroup_rename - Only allow simple rename of directories in place.
3284 */
3285static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3286 const char *new_name_str)
3287{
3288 struct cgroup *cgrp = kn->priv;
3289 int ret;
3290
3291 if (kernfs_type(kn) != KERNFS_DIR)
3292 return -ENOTDIR;
3293 if (kn->parent != new_parent)
3294 return -EIO;
3295
3296 /*
3297 * This isn't a proper migration and its usefulness is very
3298 * limited. Disallow on the default hierarchy.
3299 */
3300 if (cgroup_on_dfl(cgrp))
3301 return -EPERM;
3302
3303 /*
3304 * We're gonna grab cgroup_mutex which nests outside kernfs
3305 * active_ref. kernfs_rename() doesn't require active_ref
3306 * protection. Break them before grabbing cgroup_mutex.
3307 */
3308 kernfs_break_active_protection(new_parent);
3309 kernfs_break_active_protection(kn);
3310
3311 mutex_lock(&cgroup_mutex);
3312
3313 ret = kernfs_rename(kn, new_parent, new_name_str);
3314
3315 mutex_unlock(&cgroup_mutex);
3316
3317 kernfs_unbreak_active_protection(kn);
3318 kernfs_unbreak_active_protection(new_parent);
3319 return ret;
3320}
3321
3322/* set uid and gid of cgroup dirs and files to that of the creator */
3323static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3324{
3325 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3326 .ia_uid = current_fsuid(),
3327 .ia_gid = current_fsgid(), };
3328
3329 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3330 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3331 return 0;
3332
3333 return kernfs_setattr(kn, &iattr);
3334}
3335
3336static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3337 struct cftype *cft)
3338{
3339 char name[CGROUP_FILE_NAME_MAX];
3340 struct kernfs_node *kn;
3341 struct lock_class_key *key = NULL;
3342 int ret;
3343
3344#ifdef CONFIG_DEBUG_LOCK_ALLOC
3345 key = &cft->lockdep_key;
3346#endif
3347 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3348 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3349 NULL, key);
3350 if (IS_ERR(kn))
3351 return PTR_ERR(kn);
3352
3353 ret = cgroup_kn_set_ugid(kn);
3354 if (ret) {
3355 kernfs_remove(kn);
3356 return ret;
3357 }
3358
3359 if (cft->file_offset) {
3360 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3361
3362 spin_lock_irq(&cgroup_file_kn_lock);
3363 cfile->kn = kn;
3364 spin_unlock_irq(&cgroup_file_kn_lock);
3365 }
3366
3367 return 0;
3368}
3369
3370/**
3371 * cgroup_addrm_files - add or remove files to a cgroup directory
3372 * @css: the target css
3373 * @cgrp: the target cgroup (usually css->cgroup)
3374 * @cfts: array of cftypes to be added
3375 * @is_add: whether to add or remove
3376 *
3377 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3378 * For removals, this function never fails.
3379 */
3380static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3381 struct cgroup *cgrp, struct cftype cfts[],
3382 bool is_add)
3383{
3384 struct cftype *cft, *cft_end = NULL;
3385 int ret;
3386
3387 lockdep_assert_held(&cgroup_mutex);
3388
3389restart:
3390 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3391 /* does cft->flags tell us to skip this file on @cgrp? */
3392 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3393 continue;
3394 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3395 continue;
3396 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3397 continue;
3398 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3399 continue;
3400
3401 if (is_add) {
3402 ret = cgroup_add_file(css, cgrp, cft);
3403 if (ret) {
3404 pr_warn("%s: failed to add %s, err=%d\n",
3405 __func__, cft->name, ret);
3406 cft_end = cft;
3407 is_add = false;
3408 goto restart;
3409 }
3410 } else {
3411 cgroup_rm_file(cgrp, cft);
3412 }
3413 }
3414 return 0;
3415}
3416
3417static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3418{
3419 LIST_HEAD(pending);
3420 struct cgroup_subsys *ss = cfts[0].ss;
3421 struct cgroup *root = &ss->root->cgrp;
3422 struct cgroup_subsys_state *css;
3423 int ret = 0;
3424
3425 lockdep_assert_held(&cgroup_mutex);
3426
3427 /* add/rm files for all cgroups created before */
3428 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3429 struct cgroup *cgrp = css->cgroup;
3430
3431 if (cgroup_is_dead(cgrp))
3432 continue;
3433
3434 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3435 if (ret)
3436 break;
3437 }
3438
3439 if (is_add && !ret)
3440 kernfs_activate(root->kn);
3441 return ret;
3442}
3443
3444static void cgroup_exit_cftypes(struct cftype *cfts)
3445{
3446 struct cftype *cft;
3447
3448 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3449 /* free copy for custom atomic_write_len, see init_cftypes() */
3450 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3451 kfree(cft->kf_ops);
3452 cft->kf_ops = NULL;
3453 cft->ss = NULL;
3454
3455 /* revert flags set by cgroup core while adding @cfts */
3456 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3457 }
3458}
3459
3460static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3461{
3462 struct cftype *cft;
3463
3464 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3465 struct kernfs_ops *kf_ops;
3466
3467 WARN_ON(cft->ss || cft->kf_ops);
3468
3469 if (cft->seq_start)
3470 kf_ops = &cgroup_kf_ops;
3471 else
3472 kf_ops = &cgroup_kf_single_ops;
3473
3474 /*
3475 * Ugh... if @cft wants a custom max_write_len, we need to
3476 * make a copy of kf_ops to set its atomic_write_len.
3477 */
3478 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3479 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3480 if (!kf_ops) {
3481 cgroup_exit_cftypes(cfts);
3482 return -ENOMEM;
3483 }
3484 kf_ops->atomic_write_len = cft->max_write_len;
3485 }
3486
3487 cft->kf_ops = kf_ops;
3488 cft->ss = ss;
3489 }
3490
3491 return 0;
3492}
3493
3494static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3495{
3496 lockdep_assert_held(&cgroup_mutex);
3497
3498 if (!cfts || !cfts[0].ss)
3499 return -ENOENT;
3500
3501 list_del(&cfts->node);
3502 cgroup_apply_cftypes(cfts, false);
3503 cgroup_exit_cftypes(cfts);
3504 return 0;
3505}
3506
3507/**
3508 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3509 * @cfts: zero-length name terminated array of cftypes
3510 *
3511 * Unregister @cfts. Files described by @cfts are removed from all
3512 * existing cgroups and all future cgroups won't have them either. This
3513 * function can be called anytime whether @cfts' subsys is attached or not.
3514 *
3515 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3516 * registered.
3517 */
3518int cgroup_rm_cftypes(struct cftype *cfts)
3519{
3520 int ret;
3521
3522 mutex_lock(&cgroup_mutex);
3523 ret = cgroup_rm_cftypes_locked(cfts);
3524 mutex_unlock(&cgroup_mutex);
3525 return ret;
3526}
3527
3528/**
3529 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3530 * @ss: target cgroup subsystem
3531 * @cfts: zero-length name terminated array of cftypes
3532 *
3533 * Register @cfts to @ss. Files described by @cfts are created for all
3534 * existing cgroups to which @ss is attached and all future cgroups will
3535 * have them too. This function can be called anytime whether @ss is
3536 * attached or not.
3537 *
3538 * Returns 0 on successful registration, -errno on failure. Note that this
3539 * function currently returns 0 as long as @cfts registration is successful
3540 * even if some file creation attempts on existing cgroups fail.
3541 */
3542static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3543{
3544 int ret;
3545
3546 if (!cgroup_ssid_enabled(ss->id))
3547 return 0;
3548
3549 if (!cfts || cfts[0].name[0] == '\0')
3550 return 0;
3551
3552 ret = cgroup_init_cftypes(ss, cfts);
3553 if (ret)
3554 return ret;
3555
3556 mutex_lock(&cgroup_mutex);
3557
3558 list_add_tail(&cfts->node, &ss->cfts);
3559 ret = cgroup_apply_cftypes(cfts, true);
3560 if (ret)
3561 cgroup_rm_cftypes_locked(cfts);
3562
3563 mutex_unlock(&cgroup_mutex);
3564 return ret;
3565}
3566
3567/**
3568 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3569 * @ss: target cgroup subsystem
3570 * @cfts: zero-length name terminated array of cftypes
3571 *
3572 * Similar to cgroup_add_cftypes() but the added files are only used for
3573 * the default hierarchy.
3574 */
3575int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3576{
3577 struct cftype *cft;
3578
3579 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3580 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3581 return cgroup_add_cftypes(ss, cfts);
3582}
3583
3584/**
3585 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3586 * @ss: target cgroup subsystem
3587 * @cfts: zero-length name terminated array of cftypes
3588 *
3589 * Similar to cgroup_add_cftypes() but the added files are only used for
3590 * the legacy hierarchies.
3591 */
3592int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3593{
3594 struct cftype *cft;
3595
3596 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3597 cft->flags |= __CFTYPE_NOT_ON_DFL;
3598 return cgroup_add_cftypes(ss, cfts);
3599}
3600
3601/**
3602 * cgroup_file_notify - generate a file modified event for a cgroup_file
3603 * @cfile: target cgroup_file
3604 *
3605 * @cfile must have been obtained by setting cftype->file_offset.
3606 */
3607void cgroup_file_notify(struct cgroup_file *cfile)
3608{
3609 unsigned long flags;
3610
3611 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3612 if (cfile->kn)
3613 kernfs_notify(cfile->kn);
3614 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3615}
3616
3617/**
3618 * cgroup_task_count - count the number of tasks in a cgroup.
3619 * @cgrp: the cgroup in question
3620 *
3621 * Return the number of tasks in the cgroup.
3622 */
3623static int cgroup_task_count(const struct cgroup *cgrp)
3624{
3625 int count = 0;
3626 struct cgrp_cset_link *link;
3627
3628 spin_lock_bh(&css_set_lock);
3629 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3630 count += atomic_read(&link->cset->refcount);
3631 spin_unlock_bh(&css_set_lock);
3632 return count;
3633}
3634
3635/**
3636 * css_next_child - find the next child of a given css
3637 * @pos: the current position (%NULL to initiate traversal)
3638 * @parent: css whose children to walk
3639 *
3640 * This function returns the next child of @parent and should be called
3641 * under either cgroup_mutex or RCU read lock. The only requirement is
3642 * that @parent and @pos are accessible. The next sibling is guaranteed to
3643 * be returned regardless of their states.
3644 *
3645 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3646 * css which finished ->css_online() is guaranteed to be visible in the
3647 * future iterations and will stay visible until the last reference is put.
3648 * A css which hasn't finished ->css_online() or already finished
3649 * ->css_offline() may show up during traversal. It's each subsystem's
3650 * responsibility to synchronize against on/offlining.
3651 */
3652struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3653 struct cgroup_subsys_state *parent)
3654{
3655 struct cgroup_subsys_state *next;
3656
3657 cgroup_assert_mutex_or_rcu_locked();
3658
3659 /*
3660 * @pos could already have been unlinked from the sibling list.
3661 * Once a cgroup is removed, its ->sibling.next is no longer
3662 * updated when its next sibling changes. CSS_RELEASED is set when
3663 * @pos is taken off list, at which time its next pointer is valid,
3664 * and, as releases are serialized, the one pointed to by the next
3665 * pointer is guaranteed to not have started release yet. This
3666 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3667 * critical section, the one pointed to by its next pointer is
3668 * guaranteed to not have finished its RCU grace period even if we
3669 * have dropped rcu_read_lock() inbetween iterations.
3670 *
3671 * If @pos has CSS_RELEASED set, its next pointer can't be
3672 * dereferenced; however, as each css is given a monotonically
3673 * increasing unique serial number and always appended to the
3674 * sibling list, the next one can be found by walking the parent's
3675 * children until the first css with higher serial number than
3676 * @pos's. While this path can be slower, it happens iff iteration
3677 * races against release and the race window is very small.
3678 */
3679 if (!pos) {
3680 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3681 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3682 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3683 } else {
3684 list_for_each_entry_rcu(next, &parent->children, sibling)
3685 if (next->serial_nr > pos->serial_nr)
3686 break;
3687 }
3688
3689 /*
3690 * @next, if not pointing to the head, can be dereferenced and is
3691 * the next sibling.
3692 */
3693 if (&next->sibling != &parent->children)
3694 return next;
3695 return NULL;
3696}
3697
3698/**
3699 * css_next_descendant_pre - find the next descendant for pre-order walk
3700 * @pos: the current position (%NULL to initiate traversal)
3701 * @root: css whose descendants to walk
3702 *
3703 * To be used by css_for_each_descendant_pre(). Find the next descendant
3704 * to visit for pre-order traversal of @root's descendants. @root is
3705 * included in the iteration and the first node to be visited.
3706 *
3707 * While this function requires cgroup_mutex or RCU read locking, it
3708 * doesn't require the whole traversal to be contained in a single critical
3709 * section. This function will return the correct next descendant as long
3710 * as both @pos and @root are accessible and @pos is a descendant of @root.
3711 *
3712 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3713 * css which finished ->css_online() is guaranteed to be visible in the
3714 * future iterations and will stay visible until the last reference is put.
3715 * A css which hasn't finished ->css_online() or already finished
3716 * ->css_offline() may show up during traversal. It's each subsystem's
3717 * responsibility to synchronize against on/offlining.
3718 */
3719struct cgroup_subsys_state *
3720css_next_descendant_pre(struct cgroup_subsys_state *pos,
3721 struct cgroup_subsys_state *root)
3722{
3723 struct cgroup_subsys_state *next;
3724
3725 cgroup_assert_mutex_or_rcu_locked();
3726
3727 /* if first iteration, visit @root */
3728 if (!pos)
3729 return root;
3730
3731 /* visit the first child if exists */
3732 next = css_next_child(NULL, pos);
3733 if (next)
3734 return next;
3735
3736 /* no child, visit my or the closest ancestor's next sibling */
3737 while (pos != root) {
3738 next = css_next_child(pos, pos->parent);
3739 if (next)
3740 return next;
3741 pos = pos->parent;
3742 }
3743
3744 return NULL;
3745}
3746
3747/**
3748 * css_rightmost_descendant - return the rightmost descendant of a css
3749 * @pos: css of interest
3750 *
3751 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3752 * is returned. This can be used during pre-order traversal to skip
3753 * subtree of @pos.
3754 *
3755 * While this function requires cgroup_mutex or RCU read locking, it
3756 * doesn't require the whole traversal to be contained in a single critical
3757 * section. This function will return the correct rightmost descendant as
3758 * long as @pos is accessible.
3759 */
3760struct cgroup_subsys_state *
3761css_rightmost_descendant(struct cgroup_subsys_state *pos)
3762{
3763 struct cgroup_subsys_state *last, *tmp;
3764
3765 cgroup_assert_mutex_or_rcu_locked();
3766
3767 do {
3768 last = pos;
3769 /* ->prev isn't RCU safe, walk ->next till the end */
3770 pos = NULL;
3771 css_for_each_child(tmp, last)
3772 pos = tmp;
3773 } while (pos);
3774
3775 return last;
3776}
3777
3778static struct cgroup_subsys_state *
3779css_leftmost_descendant(struct cgroup_subsys_state *pos)
3780{
3781 struct cgroup_subsys_state *last;
3782
3783 do {
3784 last = pos;
3785 pos = css_next_child(NULL, pos);
3786 } while (pos);
3787
3788 return last;
3789}
3790
3791/**
3792 * css_next_descendant_post - find the next descendant for post-order walk
3793 * @pos: the current position (%NULL to initiate traversal)
3794 * @root: css whose descendants to walk
3795 *
3796 * To be used by css_for_each_descendant_post(). Find the next descendant
3797 * to visit for post-order traversal of @root's descendants. @root is
3798 * included in the iteration and the last node to be visited.
3799 *
3800 * While this function requires cgroup_mutex or RCU read locking, it
3801 * doesn't require the whole traversal to be contained in a single critical
3802 * section. This function will return the correct next descendant as long
3803 * as both @pos and @cgroup are accessible and @pos is a descendant of
3804 * @cgroup.
3805 *
3806 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3807 * css which finished ->css_online() is guaranteed to be visible in the
3808 * future iterations and will stay visible until the last reference is put.
3809 * A css which hasn't finished ->css_online() or already finished
3810 * ->css_offline() may show up during traversal. It's each subsystem's
3811 * responsibility to synchronize against on/offlining.
3812 */
3813struct cgroup_subsys_state *
3814css_next_descendant_post(struct cgroup_subsys_state *pos,
3815 struct cgroup_subsys_state *root)
3816{
3817 struct cgroup_subsys_state *next;
3818
3819 cgroup_assert_mutex_or_rcu_locked();
3820
3821 /* if first iteration, visit leftmost descendant which may be @root */
3822 if (!pos)
3823 return css_leftmost_descendant(root);
3824
3825 /* if we visited @root, we're done */
3826 if (pos == root)
3827 return NULL;
3828
3829 /* if there's an unvisited sibling, visit its leftmost descendant */
3830 next = css_next_child(pos, pos->parent);
3831 if (next)
3832 return css_leftmost_descendant(next);
3833
3834 /* no sibling left, visit parent */
3835 return pos->parent;
3836}
3837
3838/**
3839 * css_has_online_children - does a css have online children
3840 * @css: the target css
3841 *
3842 * Returns %true if @css has any online children; otherwise, %false. This
3843 * function can be called from any context but the caller is responsible
3844 * for synchronizing against on/offlining as necessary.
3845 */
3846bool css_has_online_children(struct cgroup_subsys_state *css)
3847{
3848 struct cgroup_subsys_state *child;
3849 bool ret = false;
3850
3851 rcu_read_lock();
3852 css_for_each_child(child, css) {
3853 if (child->flags & CSS_ONLINE) {
3854 ret = true;
3855 break;
3856 }
3857 }
3858 rcu_read_unlock();
3859 return ret;
3860}
3861
3862/**
3863 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3864 * @it: the iterator to advance
3865 *
3866 * Advance @it to the next css_set to walk.
3867 */
3868static void css_task_iter_advance_css_set(struct css_task_iter *it)
3869{
3870 struct list_head *l = it->cset_pos;
3871 struct cgrp_cset_link *link;
3872 struct css_set *cset;
3873
3874 lockdep_assert_held(&css_set_lock);
3875
3876 /* Advance to the next non-empty css_set */
3877 do {
3878 l = l->next;
3879 if (l == it->cset_head) {
3880 it->cset_pos = NULL;
3881 it->task_pos = NULL;
3882 return;
3883 }
3884
3885 if (it->ss) {
3886 cset = container_of(l, struct css_set,
3887 e_cset_node[it->ss->id]);
3888 } else {
3889 link = list_entry(l, struct cgrp_cset_link, cset_link);
3890 cset = link->cset;
3891 }
3892 } while (!css_set_populated(cset));
3893
3894 it->cset_pos = l;
3895
3896 if (!list_empty(&cset->tasks))
3897 it->task_pos = cset->tasks.next;
3898 else
3899 it->task_pos = cset->mg_tasks.next;
3900
3901 it->tasks_head = &cset->tasks;
3902 it->mg_tasks_head = &cset->mg_tasks;
3903
3904 /*
3905 * We don't keep css_sets locked across iteration steps and thus
3906 * need to take steps to ensure that iteration can be resumed after
3907 * the lock is re-acquired. Iteration is performed at two levels -
3908 * css_sets and tasks in them.
3909 *
3910 * Once created, a css_set never leaves its cgroup lists, so a
3911 * pinned css_set is guaranteed to stay put and we can resume
3912 * iteration afterwards.
3913 *
3914 * Tasks may leave @cset across iteration steps. This is resolved
3915 * by registering each iterator with the css_set currently being
3916 * walked and making css_set_move_task() advance iterators whose
3917 * next task is leaving.
3918 */
3919 if (it->cur_cset) {
3920 list_del(&it->iters_node);
3921 put_css_set_locked(it->cur_cset);
3922 }
3923 get_css_set(cset);
3924 it->cur_cset = cset;
3925 list_add(&it->iters_node, &cset->task_iters);
3926}
3927
3928static void css_task_iter_advance(struct css_task_iter *it)
3929{
3930 struct list_head *l = it->task_pos;
3931
3932 lockdep_assert_held(&css_set_lock);
3933 WARN_ON_ONCE(!l);
3934
3935 /*
3936 * Advance iterator to find next entry. cset->tasks is consumed
3937 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3938 * next cset.
3939 */
3940 l = l->next;
3941
3942 if (l == it->tasks_head)
3943 l = it->mg_tasks_head->next;
3944
3945 if (l == it->mg_tasks_head)
3946 css_task_iter_advance_css_set(it);
3947 else
3948 it->task_pos = l;
3949}
3950
3951/**
3952 * css_task_iter_start - initiate task iteration
3953 * @css: the css to walk tasks of
3954 * @it: the task iterator to use
3955 *
3956 * Initiate iteration through the tasks of @css. The caller can call
3957 * css_task_iter_next() to walk through the tasks until the function
3958 * returns NULL. On completion of iteration, css_task_iter_end() must be
3959 * called.
3960 */
3961void css_task_iter_start(struct cgroup_subsys_state *css,
3962 struct css_task_iter *it)
3963{
3964 /* no one should try to iterate before mounting cgroups */
3965 WARN_ON_ONCE(!use_task_css_set_links);
3966
3967 memset(it, 0, sizeof(*it));
3968
3969 spin_lock_bh(&css_set_lock);
3970
3971 it->ss = css->ss;
3972
3973 if (it->ss)
3974 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3975 else
3976 it->cset_pos = &css->cgroup->cset_links;
3977
3978 it->cset_head = it->cset_pos;
3979
3980 css_task_iter_advance_css_set(it);
3981
3982 spin_unlock_bh(&css_set_lock);
3983}
3984
3985/**
3986 * css_task_iter_next - return the next task for the iterator
3987 * @it: the task iterator being iterated
3988 *
3989 * The "next" function for task iteration. @it should have been
3990 * initialized via css_task_iter_start(). Returns NULL when the iteration
3991 * reaches the end.
3992 */
3993struct task_struct *css_task_iter_next(struct css_task_iter *it)
3994{
3995 if (it->cur_task) {
3996 put_task_struct(it->cur_task);
3997 it->cur_task = NULL;
3998 }
3999
4000 spin_lock_bh(&css_set_lock);
4001
4002 if (it->task_pos) {
4003 it->cur_task = list_entry(it->task_pos, struct task_struct,
4004 cg_list);
4005 get_task_struct(it->cur_task);
4006 css_task_iter_advance(it);
4007 }
4008
4009 spin_unlock_bh(&css_set_lock);
4010
4011 return it->cur_task;
4012}
4013
4014/**
4015 * css_task_iter_end - finish task iteration
4016 * @it: the task iterator to finish
4017 *
4018 * Finish task iteration started by css_task_iter_start().
4019 */
4020void css_task_iter_end(struct css_task_iter *it)
4021{
4022 if (it->cur_cset) {
4023 spin_lock_bh(&css_set_lock);
4024 list_del(&it->iters_node);
4025 put_css_set_locked(it->cur_cset);
4026 spin_unlock_bh(&css_set_lock);
4027 }
4028
4029 if (it->cur_task)
4030 put_task_struct(it->cur_task);
4031}
4032
4033/**
4034 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4035 * @to: cgroup to which the tasks will be moved
4036 * @from: cgroup in which the tasks currently reside
4037 *
4038 * Locking rules between cgroup_post_fork() and the migration path
4039 * guarantee that, if a task is forking while being migrated, the new child
4040 * is guaranteed to be either visible in the source cgroup after the
4041 * parent's migration is complete or put into the target cgroup. No task
4042 * can slip out of migration through forking.
4043 */
4044int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4045{
4046 LIST_HEAD(preloaded_csets);
4047 struct cgrp_cset_link *link;
4048 struct css_task_iter it;
4049 struct task_struct *task;
4050 int ret;
4051
4052 mutex_lock(&cgroup_mutex);
4053
4054 /* all tasks in @from are being moved, all csets are source */
4055 spin_lock_bh(&css_set_lock);
4056 list_for_each_entry(link, &from->cset_links, cset_link)
4057 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4058 spin_unlock_bh(&css_set_lock);
4059
4060 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
4061 if (ret)
4062 goto out_err;
4063
4064 /*
4065 * Migrate tasks one-by-one until @form is empty. This fails iff
4066 * ->can_attach() fails.
4067 */
4068 do {
4069 css_task_iter_start(&from->self, &it);
4070 task = css_task_iter_next(&it);
4071 if (task)
4072 get_task_struct(task);
4073 css_task_iter_end(&it);
4074
4075 if (task) {
4076 ret = cgroup_migrate(task, false, to);
4077 put_task_struct(task);
4078 }
4079 } while (task && !ret);
4080out_err:
4081 cgroup_migrate_finish(&preloaded_csets);
4082 mutex_unlock(&cgroup_mutex);
4083 return ret;
4084}
4085
4086/*
4087 * Stuff for reading the 'tasks'/'procs' files.
4088 *
4089 * Reading this file can return large amounts of data if a cgroup has
4090 * *lots* of attached tasks. So it may need several calls to read(),
4091 * but we cannot guarantee that the information we produce is correct
4092 * unless we produce it entirely atomically.
4093 *
4094 */
4095
4096/* which pidlist file are we talking about? */
4097enum cgroup_filetype {
4098 CGROUP_FILE_PROCS,
4099 CGROUP_FILE_TASKS,
4100};
4101
4102/*
4103 * A pidlist is a list of pids that virtually represents the contents of one
4104 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4105 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4106 * to the cgroup.
4107 */
4108struct cgroup_pidlist {
4109 /*
4110 * used to find which pidlist is wanted. doesn't change as long as
4111 * this particular list stays in the list.
4112 */
4113 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4114 /* array of xids */
4115 pid_t *list;
4116 /* how many elements the above list has */
4117 int length;
4118 /* each of these stored in a list by its cgroup */
4119 struct list_head links;
4120 /* pointer to the cgroup we belong to, for list removal purposes */
4121 struct cgroup *owner;
4122 /* for delayed destruction */
4123 struct delayed_work destroy_dwork;
4124};
4125
4126/*
4127 * The following two functions "fix" the issue where there are more pids
4128 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4129 * TODO: replace with a kernel-wide solution to this problem
4130 */
4131#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4132static void *pidlist_allocate(int count)
4133{
4134 if (PIDLIST_TOO_LARGE(count))
4135 return vmalloc(count * sizeof(pid_t));
4136 else
4137 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4138}
4139
4140static void pidlist_free(void *p)
4141{
4142 kvfree(p);
4143}
4144
4145/*
4146 * Used to destroy all pidlists lingering waiting for destroy timer. None
4147 * should be left afterwards.
4148 */
4149static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4150{
4151 struct cgroup_pidlist *l, *tmp_l;
4152
4153 mutex_lock(&cgrp->pidlist_mutex);
4154 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4155 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4156 mutex_unlock(&cgrp->pidlist_mutex);
4157
4158 flush_workqueue(cgroup_pidlist_destroy_wq);
4159 BUG_ON(!list_empty(&cgrp->pidlists));
4160}
4161
4162static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4163{
4164 struct delayed_work *dwork = to_delayed_work(work);
4165 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4166 destroy_dwork);
4167 struct cgroup_pidlist *tofree = NULL;
4168
4169 mutex_lock(&l->owner->pidlist_mutex);
4170
4171 /*
4172 * Destroy iff we didn't get queued again. The state won't change
4173 * as destroy_dwork can only be queued while locked.
4174 */
4175 if (!delayed_work_pending(dwork)) {
4176 list_del(&l->links);
4177 pidlist_free(l->list);
4178 put_pid_ns(l->key.ns);
4179 tofree = l;
4180 }
4181
4182 mutex_unlock(&l->owner->pidlist_mutex);
4183 kfree(tofree);
4184}
4185
4186/*
4187 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4188 * Returns the number of unique elements.
4189 */
4190static int pidlist_uniq(pid_t *list, int length)
4191{
4192 int src, dest = 1;
4193
4194 /*
4195 * we presume the 0th element is unique, so i starts at 1. trivial
4196 * edge cases first; no work needs to be done for either
4197 */
4198 if (length == 0 || length == 1)
4199 return length;
4200 /* src and dest walk down the list; dest counts unique elements */
4201 for (src = 1; src < length; src++) {
4202 /* find next unique element */
4203 while (list[src] == list[src-1]) {
4204 src++;
4205 if (src == length)
4206 goto after;
4207 }
4208 /* dest always points to where the next unique element goes */
4209 list[dest] = list[src];
4210 dest++;
4211 }
4212after:
4213 return dest;
4214}
4215
4216/*
4217 * The two pid files - task and cgroup.procs - guaranteed that the result
4218 * is sorted, which forced this whole pidlist fiasco. As pid order is
4219 * different per namespace, each namespace needs differently sorted list,
4220 * making it impossible to use, for example, single rbtree of member tasks
4221 * sorted by task pointer. As pidlists can be fairly large, allocating one
4222 * per open file is dangerous, so cgroup had to implement shared pool of
4223 * pidlists keyed by cgroup and namespace.
4224 *
4225 * All this extra complexity was caused by the original implementation
4226 * committing to an entirely unnecessary property. In the long term, we
4227 * want to do away with it. Explicitly scramble sort order if on the
4228 * default hierarchy so that no such expectation exists in the new
4229 * interface.
4230 *
4231 * Scrambling is done by swapping every two consecutive bits, which is
4232 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4233 */
4234static pid_t pid_fry(pid_t pid)
4235{
4236 unsigned a = pid & 0x55555555;
4237 unsigned b = pid & 0xAAAAAAAA;
4238
4239 return (a << 1) | (b >> 1);
4240}
4241
4242static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4243{
4244 if (cgroup_on_dfl(cgrp))
4245 return pid_fry(pid);
4246 else
4247 return pid;
4248}
4249
4250static int cmppid(const void *a, const void *b)
4251{
4252 return *(pid_t *)a - *(pid_t *)b;
4253}
4254
4255static int fried_cmppid(const void *a, const void *b)
4256{
4257 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4258}
4259
4260static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4261 enum cgroup_filetype type)
4262{
4263 struct cgroup_pidlist *l;
4264 /* don't need task_nsproxy() if we're looking at ourself */
4265 struct pid_namespace *ns = task_active_pid_ns(current);
4266
4267 lockdep_assert_held(&cgrp->pidlist_mutex);
4268
4269 list_for_each_entry(l, &cgrp->pidlists, links)
4270 if (l->key.type == type && l->key.ns == ns)
4271 return l;
4272 return NULL;
4273}
4274
4275/*
4276 * find the appropriate pidlist for our purpose (given procs vs tasks)
4277 * returns with the lock on that pidlist already held, and takes care
4278 * of the use count, or returns NULL with no locks held if we're out of
4279 * memory.
4280 */
4281static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4282 enum cgroup_filetype type)
4283{
4284 struct cgroup_pidlist *l;
4285
4286 lockdep_assert_held(&cgrp->pidlist_mutex);
4287
4288 l = cgroup_pidlist_find(cgrp, type);
4289 if (l)
4290 return l;
4291
4292 /* entry not found; create a new one */
4293 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4294 if (!l)
4295 return l;
4296
4297 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4298 l->key.type = type;
4299 /* don't need task_nsproxy() if we're looking at ourself */
4300 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4301 l->owner = cgrp;
4302 list_add(&l->links, &cgrp->pidlists);
4303 return l;
4304}
4305
4306/*
4307 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4308 */
4309static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4310 struct cgroup_pidlist **lp)
4311{
4312 pid_t *array;
4313 int length;
4314 int pid, n = 0; /* used for populating the array */
4315 struct css_task_iter it;
4316 struct task_struct *tsk;
4317 struct cgroup_pidlist *l;
4318
4319 lockdep_assert_held(&cgrp->pidlist_mutex);
4320
4321 /*
4322 * If cgroup gets more users after we read count, we won't have
4323 * enough space - tough. This race is indistinguishable to the
4324 * caller from the case that the additional cgroup users didn't
4325 * show up until sometime later on.
4326 */
4327 length = cgroup_task_count(cgrp);
4328 array = pidlist_allocate(length);
4329 if (!array)
4330 return -ENOMEM;
4331 /* now, populate the array */
4332 css_task_iter_start(&cgrp->self, &it);
4333 while ((tsk = css_task_iter_next(&it))) {
4334 if (unlikely(n == length))
4335 break;
4336 /* get tgid or pid for procs or tasks file respectively */
4337 if (type == CGROUP_FILE_PROCS)
4338 pid = task_tgid_vnr(tsk);
4339 else
4340 pid = task_pid_vnr(tsk);
4341 if (pid > 0) /* make sure to only use valid results */
4342 array[n++] = pid;
4343 }
4344 css_task_iter_end(&it);
4345 length = n;
4346 /* now sort & (if procs) strip out duplicates */
4347 if (cgroup_on_dfl(cgrp))
4348 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4349 else
4350 sort(array, length, sizeof(pid_t), cmppid, NULL);
4351 if (type == CGROUP_FILE_PROCS)
4352 length = pidlist_uniq(array, length);
4353
4354 l = cgroup_pidlist_find_create(cgrp, type);
4355 if (!l) {
4356 pidlist_free(array);
4357 return -ENOMEM;
4358 }
4359
4360 /* store array, freeing old if necessary */
4361 pidlist_free(l->list);
4362 l->list = array;
4363 l->length = length;
4364 *lp = l;
4365 return 0;
4366}
4367
4368/**
4369 * cgroupstats_build - build and fill cgroupstats
4370 * @stats: cgroupstats to fill information into
4371 * @dentry: A dentry entry belonging to the cgroup for which stats have
4372 * been requested.
4373 *
4374 * Build and fill cgroupstats so that taskstats can export it to user
4375 * space.
4376 */
4377int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4378{
4379 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4380 struct cgroup *cgrp;
4381 struct css_task_iter it;
4382 struct task_struct *tsk;
4383
4384 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4385 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4386 kernfs_type(kn) != KERNFS_DIR)
4387 return -EINVAL;
4388
4389 mutex_lock(&cgroup_mutex);
4390
4391 /*
4392 * We aren't being called from kernfs and there's no guarantee on
4393 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4394 * @kn->priv is RCU safe. Let's do the RCU dancing.
4395 */
4396 rcu_read_lock();
4397 cgrp = rcu_dereference(kn->priv);
4398 if (!cgrp || cgroup_is_dead(cgrp)) {
4399 rcu_read_unlock();
4400 mutex_unlock(&cgroup_mutex);
4401 return -ENOENT;
4402 }
4403 rcu_read_unlock();
4404
4405 css_task_iter_start(&cgrp->self, &it);
4406 while ((tsk = css_task_iter_next(&it))) {
4407 switch (tsk->state) {
4408 case TASK_RUNNING:
4409 stats->nr_running++;
4410 break;
4411 case TASK_INTERRUPTIBLE:
4412 stats->nr_sleeping++;
4413 break;
4414 case TASK_UNINTERRUPTIBLE:
4415 stats->nr_uninterruptible++;
4416 break;
4417 case TASK_STOPPED:
4418 stats->nr_stopped++;
4419 break;
4420 default:
4421 if (delayacct_is_task_waiting_on_io(tsk))
4422 stats->nr_io_wait++;
4423 break;
4424 }
4425 }
4426 css_task_iter_end(&it);
4427
4428 mutex_unlock(&cgroup_mutex);
4429 return 0;
4430}
4431
4432
4433/*
4434 * seq_file methods for the tasks/procs files. The seq_file position is the
4435 * next pid to display; the seq_file iterator is a pointer to the pid
4436 * in the cgroup->l->list array.
4437 */
4438
4439static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4440{
4441 /*
4442 * Initially we receive a position value that corresponds to
4443 * one more than the last pid shown (or 0 on the first call or
4444 * after a seek to the start). Use a binary-search to find the
4445 * next pid to display, if any
4446 */
4447 struct kernfs_open_file *of = s->private;
4448 struct cgroup *cgrp = seq_css(s)->cgroup;
4449 struct cgroup_pidlist *l;
4450 enum cgroup_filetype type = seq_cft(s)->private;
4451 int index = 0, pid = *pos;
4452 int *iter, ret;
4453
4454 mutex_lock(&cgrp->pidlist_mutex);
4455
4456 /*
4457 * !NULL @of->priv indicates that this isn't the first start()
4458 * after open. If the matching pidlist is around, we can use that.
4459 * Look for it. Note that @of->priv can't be used directly. It
4460 * could already have been destroyed.
4461 */
4462 if (of->priv)
4463 of->priv = cgroup_pidlist_find(cgrp, type);
4464
4465 /*
4466 * Either this is the first start() after open or the matching
4467 * pidlist has been destroyed inbetween. Create a new one.
4468 */
4469 if (!of->priv) {
4470 ret = pidlist_array_load(cgrp, type,
4471 (struct cgroup_pidlist **)&of->priv);
4472 if (ret)
4473 return ERR_PTR(ret);
4474 }
4475 l = of->priv;
4476
4477 if (pid) {
4478 int end = l->length;
4479
4480 while (index < end) {
4481 int mid = (index + end) / 2;
4482 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4483 index = mid;
4484 break;
4485 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4486 index = mid + 1;
4487 else
4488 end = mid;
4489 }
4490 }
4491 /* If we're off the end of the array, we're done */
4492 if (index >= l->length)
4493 return NULL;
4494 /* Update the abstract position to be the actual pid that we found */
4495 iter = l->list + index;
4496 *pos = cgroup_pid_fry(cgrp, *iter);
4497 return iter;
4498}
4499
4500static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4501{
4502 struct kernfs_open_file *of = s->private;
4503 struct cgroup_pidlist *l = of->priv;
4504
4505 if (l)
4506 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4507 CGROUP_PIDLIST_DESTROY_DELAY);
4508 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4509}
4510
4511static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4512{
4513 struct kernfs_open_file *of = s->private;
4514 struct cgroup_pidlist *l = of->priv;
4515 pid_t *p = v;
4516 pid_t *end = l->list + l->length;
4517 /*
4518 * Advance to the next pid in the array. If this goes off the
4519 * end, we're done
4520 */
4521 p++;
4522 if (p >= end) {
4523 return NULL;
4524 } else {
4525 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4526 return p;
4527 }
4528}
4529
4530static int cgroup_pidlist_show(struct seq_file *s, void *v)
4531{
4532 seq_printf(s, "%d\n", *(int *)v);
4533
4534 return 0;
4535}
4536
4537static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4538 struct cftype *cft)
4539{
4540 return notify_on_release(css->cgroup);
4541}
4542
4543static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4544 struct cftype *cft, u64 val)
4545{
4546 if (val)
4547 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4548 else
4549 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4550 return 0;
4551}
4552
4553static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4554 struct cftype *cft)
4555{
4556 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4557}
4558
4559static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4560 struct cftype *cft, u64 val)
4561{
4562 if (val)
4563 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4564 else
4565 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4566 return 0;
4567}
4568
4569/* cgroup core interface files for the default hierarchy */
4570static struct cftype cgroup_dfl_base_files[] = {
4571 {
4572 .name = "cgroup.procs",
4573 .file_offset = offsetof(struct cgroup, procs_file),
4574 .seq_start = cgroup_pidlist_start,
4575 .seq_next = cgroup_pidlist_next,
4576 .seq_stop = cgroup_pidlist_stop,
4577 .seq_show = cgroup_pidlist_show,
4578 .private = CGROUP_FILE_PROCS,
4579 .write = cgroup_procs_write,
4580 },
4581 {
4582 .name = "cgroup.controllers",
4583 .flags = CFTYPE_ONLY_ON_ROOT,
4584 .seq_show = cgroup_root_controllers_show,
4585 },
4586 {
4587 .name = "cgroup.controllers",
4588 .flags = CFTYPE_NOT_ON_ROOT,
4589 .seq_show = cgroup_controllers_show,
4590 },
4591 {
4592 .name = "cgroup.subtree_control",
4593 .seq_show = cgroup_subtree_control_show,
4594 .write = cgroup_subtree_control_write,
4595 },
4596 {
4597 .name = "cgroup.events",
4598 .flags = CFTYPE_NOT_ON_ROOT,
4599 .file_offset = offsetof(struct cgroup, events_file),
4600 .seq_show = cgroup_events_show,
4601 },
4602 { } /* terminate */
4603};
4604
4605/* cgroup core interface files for the legacy hierarchies */
4606static struct cftype cgroup_legacy_base_files[] = {
4607 {
4608 .name = "cgroup.procs",
4609 .seq_start = cgroup_pidlist_start,
4610 .seq_next = cgroup_pidlist_next,
4611 .seq_stop = cgroup_pidlist_stop,
4612 .seq_show = cgroup_pidlist_show,
4613 .private = CGROUP_FILE_PROCS,
4614 .write = cgroup_procs_write,
4615 },
4616 {
4617 .name = "cgroup.clone_children",
4618 .read_u64 = cgroup_clone_children_read,
4619 .write_u64 = cgroup_clone_children_write,
4620 },
4621 {
4622 .name = "cgroup.sane_behavior",
4623 .flags = CFTYPE_ONLY_ON_ROOT,
4624 .seq_show = cgroup_sane_behavior_show,
4625 },
4626 {
4627 .name = "tasks",
4628 .seq_start = cgroup_pidlist_start,
4629 .seq_next = cgroup_pidlist_next,
4630 .seq_stop = cgroup_pidlist_stop,
4631 .seq_show = cgroup_pidlist_show,
4632 .private = CGROUP_FILE_TASKS,
4633 .write = cgroup_tasks_write,
4634 },
4635 {
4636 .name = "notify_on_release",
4637 .read_u64 = cgroup_read_notify_on_release,
4638 .write_u64 = cgroup_write_notify_on_release,
4639 },
4640 {
4641 .name = "release_agent",
4642 .flags = CFTYPE_ONLY_ON_ROOT,
4643 .seq_show = cgroup_release_agent_show,
4644 .write = cgroup_release_agent_write,
4645 .max_write_len = PATH_MAX - 1,
4646 },
4647 { } /* terminate */
4648};
4649
4650/*
4651 * css destruction is four-stage process.
4652 *
4653 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4654 * Implemented in kill_css().
4655 *
4656 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4657 * and thus css_tryget_online() is guaranteed to fail, the css can be
4658 * offlined by invoking offline_css(). After offlining, the base ref is
4659 * put. Implemented in css_killed_work_fn().
4660 *
4661 * 3. When the percpu_ref reaches zero, the only possible remaining
4662 * accessors are inside RCU read sections. css_release() schedules the
4663 * RCU callback.
4664 *
4665 * 4. After the grace period, the css can be freed. Implemented in
4666 * css_free_work_fn().
4667 *
4668 * It is actually hairier because both step 2 and 4 require process context
4669 * and thus involve punting to css->destroy_work adding two additional
4670 * steps to the already complex sequence.
4671 */
4672static void css_free_work_fn(struct work_struct *work)
4673{
4674 struct cgroup_subsys_state *css =
4675 container_of(work, struct cgroup_subsys_state, destroy_work);
4676 struct cgroup_subsys *ss = css->ss;
4677 struct cgroup *cgrp = css->cgroup;
4678
4679 percpu_ref_exit(&css->refcnt);
4680
4681 if (ss) {
4682 /* css free path */
4683 int id = css->id;
4684
4685 if (css->parent)
4686 css_put(css->parent);
4687
4688 ss->css_free(css);
4689 cgroup_idr_remove(&ss->css_idr, id);
4690 cgroup_put(cgrp);
4691 } else {
4692 /* cgroup free path */
4693 atomic_dec(&cgrp->root->nr_cgrps);
4694 cgroup_pidlist_destroy_all(cgrp);
4695 cancel_work_sync(&cgrp->release_agent_work);
4696
4697 if (cgroup_parent(cgrp)) {
4698 /*
4699 * We get a ref to the parent, and put the ref when
4700 * this cgroup is being freed, so it's guaranteed
4701 * that the parent won't be destroyed before its
4702 * children.
4703 */
4704 cgroup_put(cgroup_parent(cgrp));
4705 kernfs_put(cgrp->kn);
4706 kfree(cgrp);
4707 } else {
4708 /*
4709 * This is root cgroup's refcnt reaching zero,
4710 * which indicates that the root should be
4711 * released.
4712 */
4713 cgroup_destroy_root(cgrp->root);
4714 }
4715 }
4716}
4717
4718static void css_free_rcu_fn(struct rcu_head *rcu_head)
4719{
4720 struct cgroup_subsys_state *css =
4721 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4722
4723 INIT_WORK(&css->destroy_work, css_free_work_fn);
4724 queue_work(cgroup_destroy_wq, &css->destroy_work);
4725}
4726
4727static void css_release_work_fn(struct work_struct *work)
4728{
4729 struct cgroup_subsys_state *css =
4730 container_of(work, struct cgroup_subsys_state, destroy_work);
4731 struct cgroup_subsys *ss = css->ss;
4732 struct cgroup *cgrp = css->cgroup;
4733
4734 mutex_lock(&cgroup_mutex);
4735
4736 css->flags |= CSS_RELEASED;
4737 list_del_rcu(&css->sibling);
4738
4739 if (ss) {
4740 /* css release path */
4741 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4742 if (ss->css_released)
4743 ss->css_released(css);
4744 } else {
4745 /* cgroup release path */
4746 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4747 cgrp->id = -1;
4748
4749 /*
4750 * There are two control paths which try to determine
4751 * cgroup from dentry without going through kernfs -
4752 * cgroupstats_build() and css_tryget_online_from_dir().
4753 * Those are supported by RCU protecting clearing of
4754 * cgrp->kn->priv backpointer.
4755 */
4756 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4757 }
4758
4759 mutex_unlock(&cgroup_mutex);
4760
4761 call_rcu(&css->rcu_head, css_free_rcu_fn);
4762}
4763
4764static void css_release(struct percpu_ref *ref)
4765{
4766 struct cgroup_subsys_state *css =
4767 container_of(ref, struct cgroup_subsys_state, refcnt);
4768
4769 INIT_WORK(&css->destroy_work, css_release_work_fn);
4770 queue_work(cgroup_destroy_wq, &css->destroy_work);
4771}
4772
4773static void init_and_link_css(struct cgroup_subsys_state *css,
4774 struct cgroup_subsys *ss, struct cgroup *cgrp)
4775{
4776 lockdep_assert_held(&cgroup_mutex);
4777
4778 cgroup_get(cgrp);
4779
4780 memset(css, 0, sizeof(*css));
4781 css->cgroup = cgrp;
4782 css->ss = ss;
4783 INIT_LIST_HEAD(&css->sibling);
4784 INIT_LIST_HEAD(&css->children);
4785 css->serial_nr = css_serial_nr_next++;
4786
4787 if (cgroup_parent(cgrp)) {
4788 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4789 css_get(css->parent);
4790 }
4791
4792 BUG_ON(cgroup_css(cgrp, ss));
4793}
4794
4795/* invoke ->css_online() on a new CSS and mark it online if successful */
4796static int online_css(struct cgroup_subsys_state *css)
4797{
4798 struct cgroup_subsys *ss = css->ss;
4799 int ret = 0;
4800
4801 lockdep_assert_held(&cgroup_mutex);
4802
4803 if (ss->css_online)
4804 ret = ss->css_online(css);
4805 if (!ret) {
4806 css->flags |= CSS_ONLINE;
4807 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4808 }
4809 return ret;
4810}
4811
4812/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4813static void offline_css(struct cgroup_subsys_state *css)
4814{
4815 struct cgroup_subsys *ss = css->ss;
4816
4817 lockdep_assert_held(&cgroup_mutex);
4818
4819 if (!(css->flags & CSS_ONLINE))
4820 return;
4821
4822 if (ss->css_offline)
4823 ss->css_offline(css);
4824
4825 css->flags &= ~CSS_ONLINE;
4826 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4827
4828 wake_up_all(&css->cgroup->offline_waitq);
4829}
4830
4831/**
4832 * create_css - create a cgroup_subsys_state
4833 * @cgrp: the cgroup new css will be associated with
4834 * @ss: the subsys of new css
4835 * @visible: whether to create control knobs for the new css or not
4836 *
4837 * Create a new css associated with @cgrp - @ss pair. On success, the new
4838 * css is online and installed in @cgrp with all interface files created if
4839 * @visible. Returns 0 on success, -errno on failure.
4840 */
4841static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4842 bool visible)
4843{
4844 struct cgroup *parent = cgroup_parent(cgrp);
4845 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4846 struct cgroup_subsys_state *css;
4847 int err;
4848
4849 lockdep_assert_held(&cgroup_mutex);
4850
4851 css = ss->css_alloc(parent_css);
4852 if (IS_ERR(css))
4853 return PTR_ERR(css);
4854
4855 init_and_link_css(css, ss, cgrp);
4856
4857 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4858 if (err)
4859 goto err_free_css;
4860
4861 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4862 if (err < 0)
4863 goto err_free_percpu_ref;
4864 css->id = err;
4865
4866 if (visible) {
4867 err = css_populate_dir(css, NULL);
4868 if (err)
4869 goto err_free_id;
4870 }
4871
4872 /* @css is ready to be brought online now, make it visible */
4873 list_add_tail_rcu(&css->sibling, &parent_css->children);
4874 cgroup_idr_replace(&ss->css_idr, css, css->id);
4875
4876 err = online_css(css);
4877 if (err)
4878 goto err_list_del;
4879
4880 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4881 cgroup_parent(parent)) {
4882 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4883 current->comm, current->pid, ss->name);
4884 if (!strcmp(ss->name, "memory"))
4885 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4886 ss->warned_broken_hierarchy = true;
4887 }
4888
4889 return 0;
4890
4891err_list_del:
4892 list_del_rcu(&css->sibling);
4893 css_clear_dir(css, NULL);
4894err_free_id:
4895 cgroup_idr_remove(&ss->css_idr, css->id);
4896err_free_percpu_ref:
4897 percpu_ref_exit(&css->refcnt);
4898err_free_css:
4899 call_rcu(&css->rcu_head, css_free_rcu_fn);
4900 return err;
4901}
4902
4903static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4904 umode_t mode)
4905{
4906 struct cgroup *parent, *cgrp;
4907 struct cgroup_root *root;
4908 struct cgroup_subsys *ss;
4909 struct kernfs_node *kn;
4910 int ssid, ret;
4911
4912 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4913 */
4914 if (strchr(name, '\n'))
4915 return -EINVAL;
4916
4917 parent = cgroup_kn_lock_live(parent_kn);
4918 if (!parent)
4919 return -ENODEV;
4920 root = parent->root;
4921
4922 /* allocate the cgroup and its ID, 0 is reserved for the root */
4923 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4924 if (!cgrp) {
4925 ret = -ENOMEM;
4926 goto out_unlock;
4927 }
4928
4929 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4930 if (ret)
4931 goto out_free_cgrp;
4932
4933 /*
4934 * Temporarily set the pointer to NULL, so idr_find() won't return
4935 * a half-baked cgroup.
4936 */
4937 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4938 if (cgrp->id < 0) {
4939 ret = -ENOMEM;
4940 goto out_cancel_ref;
4941 }
4942
4943 init_cgroup_housekeeping(cgrp);
4944
4945 cgrp->self.parent = &parent->self;
4946 cgrp->root = root;
4947
4948 if (notify_on_release(parent))
4949 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4950
4951 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4952 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4953
4954 /* create the directory */
4955 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4956 if (IS_ERR(kn)) {
4957 ret = PTR_ERR(kn);
4958 goto out_free_id;
4959 }
4960 cgrp->kn = kn;
4961
4962 /*
4963 * This extra ref will be put in cgroup_free_fn() and guarantees
4964 * that @cgrp->kn is always accessible.
4965 */
4966 kernfs_get(kn);
4967
4968 cgrp->self.serial_nr = css_serial_nr_next++;
4969
4970 /* allocation complete, commit to creation */
4971 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4972 atomic_inc(&root->nr_cgrps);
4973 cgroup_get(parent);
4974
4975 /*
4976 * @cgrp is now fully operational. If something fails after this
4977 * point, it'll be released via the normal destruction path.
4978 */
4979 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4980
4981 ret = cgroup_kn_set_ugid(kn);
4982 if (ret)
4983 goto out_destroy;
4984
4985 ret = css_populate_dir(&cgrp->self, NULL);
4986 if (ret)
4987 goto out_destroy;
4988
4989 /* let's create and online css's */
4990 for_each_subsys(ss, ssid) {
4991 if (parent->child_subsys_mask & (1 << ssid)) {
4992 ret = create_css(cgrp, ss,
4993 parent->subtree_control & (1 << ssid));
4994 if (ret)
4995 goto out_destroy;
4996 }
4997 }
4998
4999 /*
5000 * On the default hierarchy, a child doesn't automatically inherit
5001 * subtree_control from the parent. Each is configured manually.
5002 */
5003 if (!cgroup_on_dfl(cgrp)) {
5004 cgrp->subtree_control = parent->subtree_control;
5005 cgroup_refresh_child_subsys_mask(cgrp);
5006 }
5007
5008 kernfs_activate(kn);
5009
5010 ret = 0;
5011 goto out_unlock;
5012
5013out_free_id:
5014 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
5015out_cancel_ref:
5016 percpu_ref_exit(&cgrp->self.refcnt);
5017out_free_cgrp:
5018 kfree(cgrp);
5019out_unlock:
5020 cgroup_kn_unlock(parent_kn);
5021 return ret;
5022
5023out_destroy:
5024 cgroup_destroy_locked(cgrp);
5025 goto out_unlock;
5026}
5027
5028/*
5029 * This is called when the refcnt of a css is confirmed to be killed.
5030 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5031 * initate destruction and put the css ref from kill_css().
5032 */
5033static void css_killed_work_fn(struct work_struct *work)
5034{
5035 struct cgroup_subsys_state *css =
5036 container_of(work, struct cgroup_subsys_state, destroy_work);
5037
5038 mutex_lock(&cgroup_mutex);
5039 offline_css(css);
5040 mutex_unlock(&cgroup_mutex);
5041
5042 css_put(css);
5043}
5044
5045/* css kill confirmation processing requires process context, bounce */
5046static void css_killed_ref_fn(struct percpu_ref *ref)
5047{
5048 struct cgroup_subsys_state *css =
5049 container_of(ref, struct cgroup_subsys_state, refcnt);
5050
5051 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5052 queue_work(cgroup_destroy_wq, &css->destroy_work);
5053}
5054
5055/**
5056 * kill_css - destroy a css
5057 * @css: css to destroy
5058 *
5059 * This function initiates destruction of @css by removing cgroup interface
5060 * files and putting its base reference. ->css_offline() will be invoked
5061 * asynchronously once css_tryget_online() is guaranteed to fail and when
5062 * the reference count reaches zero, @css will be released.
5063 */
5064static void kill_css(struct cgroup_subsys_state *css)
5065{
5066 lockdep_assert_held(&cgroup_mutex);
5067
5068 /*
5069 * This must happen before css is disassociated with its cgroup.
5070 * See seq_css() for details.
5071 */
5072 css_clear_dir(css, NULL);
5073
5074 /*
5075 * Killing would put the base ref, but we need to keep it alive
5076 * until after ->css_offline().
5077 */
5078 css_get(css);
5079
5080 /*
5081 * cgroup core guarantees that, by the time ->css_offline() is
5082 * invoked, no new css reference will be given out via
5083 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5084 * proceed to offlining css's because percpu_ref_kill() doesn't
5085 * guarantee that the ref is seen as killed on all CPUs on return.
5086 *
5087 * Use percpu_ref_kill_and_confirm() to get notifications as each
5088 * css is confirmed to be seen as killed on all CPUs.
5089 */
5090 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5091}
5092
5093/**
5094 * cgroup_destroy_locked - the first stage of cgroup destruction
5095 * @cgrp: cgroup to be destroyed
5096 *
5097 * css's make use of percpu refcnts whose killing latency shouldn't be
5098 * exposed to userland and are RCU protected. Also, cgroup core needs to
5099 * guarantee that css_tryget_online() won't succeed by the time
5100 * ->css_offline() is invoked. To satisfy all the requirements,
5101 * destruction is implemented in the following two steps.
5102 *
5103 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5104 * userland visible parts and start killing the percpu refcnts of
5105 * css's. Set up so that the next stage will be kicked off once all
5106 * the percpu refcnts are confirmed to be killed.
5107 *
5108 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5109 * rest of destruction. Once all cgroup references are gone, the
5110 * cgroup is RCU-freed.
5111 *
5112 * This function implements s1. After this step, @cgrp is gone as far as
5113 * the userland is concerned and a new cgroup with the same name may be
5114 * created. As cgroup doesn't care about the names internally, this
5115 * doesn't cause any problem.
5116 */
5117static int cgroup_destroy_locked(struct cgroup *cgrp)
5118 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5119{
5120 struct cgroup_subsys_state *css;
5121 int ssid;
5122
5123 lockdep_assert_held(&cgroup_mutex);
5124
5125 /*
5126 * Only migration can raise populated from zero and we're already
5127 * holding cgroup_mutex.
5128 */
5129 if (cgroup_is_populated(cgrp))
5130 return -EBUSY;
5131
5132 /*
5133 * Make sure there's no live children. We can't test emptiness of
5134 * ->self.children as dead children linger on it while being
5135 * drained; otherwise, "rmdir parent/child parent" may fail.
5136 */
5137 if (css_has_online_children(&cgrp->self))
5138 return -EBUSY;
5139
5140 /*
5141 * Mark @cgrp dead. This prevents further task migration and child
5142 * creation by disabling cgroup_lock_live_group().
5143 */
5144 cgrp->self.flags &= ~CSS_ONLINE;
5145
5146 /* initiate massacre of all css's */
5147 for_each_css(css, ssid, cgrp)
5148 kill_css(css);
5149
5150 /*
5151 * Remove @cgrp directory along with the base files. @cgrp has an
5152 * extra ref on its kn.
5153 */
5154 kernfs_remove(cgrp->kn);
5155
5156 check_for_release(cgroup_parent(cgrp));
5157
5158 /* put the base reference */
5159 percpu_ref_kill(&cgrp->self.refcnt);
5160
5161 return 0;
5162};
5163
5164static int cgroup_rmdir(struct kernfs_node *kn)
5165{
5166 struct cgroup *cgrp;
5167 int ret = 0;
5168
5169 cgrp = cgroup_kn_lock_live(kn);
5170 if (!cgrp)
5171 return 0;
5172
5173 ret = cgroup_destroy_locked(cgrp);
5174
5175 cgroup_kn_unlock(kn);
5176 return ret;
5177}
5178
5179static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5180 .remount_fs = cgroup_remount,
5181 .show_options = cgroup_show_options,
5182 .mkdir = cgroup_mkdir,
5183 .rmdir = cgroup_rmdir,
5184 .rename = cgroup_rename,
5185};
5186
5187static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5188{
5189 struct cgroup_subsys_state *css;
5190
5191 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
5192
5193 mutex_lock(&cgroup_mutex);
5194
5195 idr_init(&ss->css_idr);
5196 INIT_LIST_HEAD(&ss->cfts);
5197
5198 /* Create the root cgroup state for this subsystem */
5199 ss->root = &cgrp_dfl_root;
5200 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5201 /* We don't handle early failures gracefully */
5202 BUG_ON(IS_ERR(css));
5203 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5204
5205 /*
5206 * Root csses are never destroyed and we can't initialize
5207 * percpu_ref during early init. Disable refcnting.
5208 */
5209 css->flags |= CSS_NO_REF;
5210
5211 if (early) {
5212 /* allocation can't be done safely during early init */
5213 css->id = 1;
5214 } else {
5215 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5216 BUG_ON(css->id < 0);
5217 }
5218
5219 /* Update the init_css_set to contain a subsys
5220 * pointer to this state - since the subsystem is
5221 * newly registered, all tasks and hence the
5222 * init_css_set is in the subsystem's root cgroup. */
5223 init_css_set.subsys[ss->id] = css;
5224
5225 have_fork_callback |= (bool)ss->fork << ss->id;
5226 have_exit_callback |= (bool)ss->exit << ss->id;
5227 have_free_callback |= (bool)ss->free << ss->id;
5228 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5229
5230 /* At system boot, before all subsystems have been
5231 * registered, no tasks have been forked, so we don't
5232 * need to invoke fork callbacks here. */
5233 BUG_ON(!list_empty(&init_task.tasks));
5234
5235 BUG_ON(online_css(css));
5236
5237 mutex_unlock(&cgroup_mutex);
5238}
5239
5240/**
5241 * cgroup_init_early - cgroup initialization at system boot
5242 *
5243 * Initialize cgroups at system boot, and initialize any
5244 * subsystems that request early init.
5245 */
5246int __init cgroup_init_early(void)
5247{
5248 static struct cgroup_sb_opts __initdata opts;
5249 struct cgroup_subsys *ss;
5250 int i;
5251
5252 init_cgroup_root(&cgrp_dfl_root, &opts);
5253 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5254
5255 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5256
5257 for_each_subsys(ss, i) {
5258 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5259 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5260 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5261 ss->id, ss->name);
5262 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5263 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5264
5265 ss->id = i;
5266 ss->name = cgroup_subsys_name[i];
5267 if (!ss->legacy_name)
5268 ss->legacy_name = cgroup_subsys_name[i];
5269
5270 if (ss->early_init)
5271 cgroup_init_subsys(ss, true);
5272 }
5273 return 0;
5274}
5275
5276static unsigned long cgroup_disable_mask __initdata;
5277
5278/**
5279 * cgroup_init - cgroup initialization
5280 *
5281 * Register cgroup filesystem and /proc file, and initialize
5282 * any subsystems that didn't request early init.
5283 */
5284int __init cgroup_init(void)
5285{
5286 struct cgroup_subsys *ss;
5287 unsigned long key;
5288 int ssid;
5289
5290 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5291 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5292 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5293
5294 mutex_lock(&cgroup_mutex);
5295
5296 /* Add init_css_set to the hash table */
5297 key = css_set_hash(init_css_set.subsys);
5298 hash_add(css_set_table, &init_css_set.hlist, key);
5299
5300 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5301
5302 mutex_unlock(&cgroup_mutex);
5303
5304 for_each_subsys(ss, ssid) {
5305 if (ss->early_init) {
5306 struct cgroup_subsys_state *css =
5307 init_css_set.subsys[ss->id];
5308
5309 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5310 GFP_KERNEL);
5311 BUG_ON(css->id < 0);
5312 } else {
5313 cgroup_init_subsys(ss, false);
5314 }
5315
5316 list_add_tail(&init_css_set.e_cset_node[ssid],
5317 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5318
5319 /*
5320 * Setting dfl_root subsys_mask needs to consider the
5321 * disabled flag and cftype registration needs kmalloc,
5322 * both of which aren't available during early_init.
5323 */
5324 if (cgroup_disable_mask & (1 << ssid)) {
5325 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5326 printk(KERN_INFO "Disabling %s control group subsystem\n",
5327 ss->name);
5328 continue;
5329 }
5330
5331 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5332
5333 if (!ss->dfl_cftypes)
5334 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5335
5336 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5337 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5338 } else {
5339 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5340 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5341 }
5342
5343 if (ss->bind)
5344 ss->bind(init_css_set.subsys[ssid]);
5345 }
5346
5347 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5348 WARN_ON(register_filesystem(&cgroup_fs_type));
5349 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5350
5351 return 0;
5352}
5353
5354static int __init cgroup_wq_init(void)
5355{
5356 /*
5357 * There isn't much point in executing destruction path in
5358 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5359 * Use 1 for @max_active.
5360 *
5361 * We would prefer to do this in cgroup_init() above, but that
5362 * is called before init_workqueues(): so leave this until after.
5363 */
5364 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5365 BUG_ON(!cgroup_destroy_wq);
5366
5367 /*
5368 * Used to destroy pidlists and separate to serve as flush domain.
5369 * Cap @max_active to 1 too.
5370 */
5371 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5372 0, 1);
5373 BUG_ON(!cgroup_pidlist_destroy_wq);
5374
5375 return 0;
5376}
5377core_initcall(cgroup_wq_init);
5378
5379/*
5380 * proc_cgroup_show()
5381 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5382 * - Used for /proc/<pid>/cgroup.
5383 */
5384int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5385 struct pid *pid, struct task_struct *tsk)
5386{
5387 char *buf, *path;
5388 int retval;
5389 struct cgroup_root *root;
5390
5391 retval = -ENOMEM;
5392 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5393 if (!buf)
5394 goto out;
5395
5396 mutex_lock(&cgroup_mutex);
5397 spin_lock_bh(&css_set_lock);
5398
5399 for_each_root(root) {
5400 struct cgroup_subsys *ss;
5401 struct cgroup *cgrp;
5402 int ssid, count = 0;
5403
5404 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5405 continue;
5406
5407 seq_printf(m, "%d:", root->hierarchy_id);
5408 if (root != &cgrp_dfl_root)
5409 for_each_subsys(ss, ssid)
5410 if (root->subsys_mask & (1 << ssid))
5411 seq_printf(m, "%s%s", count++ ? "," : "",
5412 ss->legacy_name);
5413 if (strlen(root->name))
5414 seq_printf(m, "%sname=%s", count ? "," : "",
5415 root->name);
5416 seq_putc(m, ':');
5417
5418 cgrp = task_cgroup_from_root(tsk, root);
5419
5420 /*
5421 * On traditional hierarchies, all zombie tasks show up as
5422 * belonging to the root cgroup. On the default hierarchy,
5423 * while a zombie doesn't show up in "cgroup.procs" and
5424 * thus can't be migrated, its /proc/PID/cgroup keeps
5425 * reporting the cgroup it belonged to before exiting. If
5426 * the cgroup is removed before the zombie is reaped,
5427 * " (deleted)" is appended to the cgroup path.
5428 */
5429 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5430 path = cgroup_path(cgrp, buf, PATH_MAX);
5431 if (!path) {
5432 retval = -ENAMETOOLONG;
5433 goto out_unlock;
5434 }
5435 } else {
5436 path = "/";
5437 }
5438
5439 seq_puts(m, path);
5440
5441 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5442 seq_puts(m, " (deleted)\n");
5443 else
5444 seq_putc(m, '\n');
5445 }
5446
5447 retval = 0;
5448out_unlock:
5449 spin_unlock_bh(&css_set_lock);
5450 mutex_unlock(&cgroup_mutex);
5451 kfree(buf);
5452out:
5453 return retval;
5454}
5455
5456/* Display information about each subsystem and each hierarchy */
5457static int proc_cgroupstats_show(struct seq_file *m, void *v)
5458{
5459 struct cgroup_subsys *ss;
5460 int i;
5461
5462 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5463 /*
5464 * ideally we don't want subsystems moving around while we do this.
5465 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5466 * subsys/hierarchy state.
5467 */
5468 mutex_lock(&cgroup_mutex);
5469
5470 for_each_subsys(ss, i)
5471 seq_printf(m, "%s\t%d\t%d\t%d\n",
5472 ss->legacy_name, ss->root->hierarchy_id,
5473 atomic_read(&ss->root->nr_cgrps),
5474 cgroup_ssid_enabled(i));
5475
5476 mutex_unlock(&cgroup_mutex);
5477 return 0;
5478}
5479
5480static int cgroupstats_open(struct inode *inode, struct file *file)
5481{
5482 return single_open(file, proc_cgroupstats_show, NULL);
5483}
5484
5485static const struct file_operations proc_cgroupstats_operations = {
5486 .open = cgroupstats_open,
5487 .read = seq_read,
5488 .llseek = seq_lseek,
5489 .release = single_release,
5490};
5491
5492static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5493{
5494 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5495 return &ss_priv[i - CGROUP_CANFORK_START];
5496 return NULL;
5497}
5498
5499static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5500{
5501 void **private = subsys_canfork_priv_p(ss_priv, i);
5502 return private ? *private : NULL;
5503}
5504
5505/**
5506 * cgroup_fork - initialize cgroup related fields during copy_process()
5507 * @child: pointer to task_struct of forking parent process.
5508 *
5509 * A task is associated with the init_css_set until cgroup_post_fork()
5510 * attaches it to the parent's css_set. Empty cg_list indicates that
5511 * @child isn't holding reference to its css_set.
5512 */
5513void cgroup_fork(struct task_struct *child)
5514{
5515 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5516 INIT_LIST_HEAD(&child->cg_list);
5517}
5518
5519/**
5520 * cgroup_can_fork - called on a new task before the process is exposed
5521 * @child: the task in question.
5522 *
5523 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5524 * returns an error, the fork aborts with that error code. This allows for
5525 * a cgroup subsystem to conditionally allow or deny new forks.
5526 */
5527int cgroup_can_fork(struct task_struct *child,
5528 void *ss_priv[CGROUP_CANFORK_COUNT])
5529{
5530 struct cgroup_subsys *ss;
5531 int i, j, ret;
5532
5533 for_each_subsys_which(ss, i, &have_canfork_callback) {
5534 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5535 if (ret)
5536 goto out_revert;
5537 }
5538
5539 return 0;
5540
5541out_revert:
5542 for_each_subsys(ss, j) {
5543 if (j >= i)
5544 break;
5545 if (ss->cancel_fork)
5546 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5547 }
5548
5549 return ret;
5550}
5551
5552/**
5553 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5554 * @child: the task in question
5555 *
5556 * This calls the cancel_fork() callbacks if a fork failed *after*
5557 * cgroup_can_fork() succeded.
5558 */
5559void cgroup_cancel_fork(struct task_struct *child,
5560 void *ss_priv[CGROUP_CANFORK_COUNT])
5561{
5562 struct cgroup_subsys *ss;
5563 int i;
5564
5565 for_each_subsys(ss, i)
5566 if (ss->cancel_fork)
5567 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5568}
5569
5570/**
5571 * cgroup_post_fork - called on a new task after adding it to the task list
5572 * @child: the task in question
5573 *
5574 * Adds the task to the list running through its css_set if necessary and
5575 * call the subsystem fork() callbacks. Has to be after the task is
5576 * visible on the task list in case we race with the first call to
5577 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5578 * list.
5579 */
5580void cgroup_post_fork(struct task_struct *child,
5581 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5582{
5583 struct cgroup_subsys *ss;
5584 int i;
5585
5586 /*
5587 * This may race against cgroup_enable_task_cg_lists(). As that
5588 * function sets use_task_css_set_links before grabbing
5589 * tasklist_lock and we just went through tasklist_lock to add
5590 * @child, it's guaranteed that either we see the set
5591 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5592 * @child during its iteration.
5593 *
5594 * If we won the race, @child is associated with %current's
5595 * css_set. Grabbing css_set_lock guarantees both that the
5596 * association is stable, and, on completion of the parent's
5597 * migration, @child is visible in the source of migration or
5598 * already in the destination cgroup. This guarantee is necessary
5599 * when implementing operations which need to migrate all tasks of
5600 * a cgroup to another.
5601 *
5602 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5603 * will remain in init_css_set. This is safe because all tasks are
5604 * in the init_css_set before cg_links is enabled and there's no
5605 * operation which transfers all tasks out of init_css_set.
5606 */
5607 if (use_task_css_set_links) {
5608 struct css_set *cset;
5609
5610 spin_lock_bh(&css_set_lock);
5611 cset = task_css_set(current);
5612 if (list_empty(&child->cg_list)) {
5613 get_css_set(cset);
5614 css_set_move_task(child, NULL, cset, false);
5615 }
5616 spin_unlock_bh(&css_set_lock);
5617 }
5618
5619 /*
5620 * Call ss->fork(). This must happen after @child is linked on
5621 * css_set; otherwise, @child might change state between ->fork()
5622 * and addition to css_set.
5623 */
5624 for_each_subsys_which(ss, i, &have_fork_callback)
5625 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5626}
5627
5628/**
5629 * cgroup_exit - detach cgroup from exiting task
5630 * @tsk: pointer to task_struct of exiting process
5631 *
5632 * Description: Detach cgroup from @tsk and release it.
5633 *
5634 * Note that cgroups marked notify_on_release force every task in
5635 * them to take the global cgroup_mutex mutex when exiting.
5636 * This could impact scaling on very large systems. Be reluctant to
5637 * use notify_on_release cgroups where very high task exit scaling
5638 * is required on large systems.
5639 *
5640 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5641 * call cgroup_exit() while the task is still competent to handle
5642 * notify_on_release(), then leave the task attached to the root cgroup in
5643 * each hierarchy for the remainder of its exit. No need to bother with
5644 * init_css_set refcnting. init_css_set never goes away and we can't race
5645 * with migration path - PF_EXITING is visible to migration path.
5646 */
5647void cgroup_exit(struct task_struct *tsk)
5648{
5649 struct cgroup_subsys *ss;
5650 struct css_set *cset;
5651 int i;
5652
5653 /*
5654 * Unlink from @tsk from its css_set. As migration path can't race
5655 * with us, we can check css_set and cg_list without synchronization.
5656 */
5657 cset = task_css_set(tsk);
5658
5659 if (!list_empty(&tsk->cg_list)) {
5660 spin_lock_bh(&css_set_lock);
5661 css_set_move_task(tsk, cset, NULL, false);
5662 spin_unlock_bh(&css_set_lock);
5663 } else {
5664 get_css_set(cset);
5665 }
5666
5667 /* see cgroup_post_fork() for details */
5668 for_each_subsys_which(ss, i, &have_exit_callback)
5669 ss->exit(tsk);
5670}
5671
5672void cgroup_free(struct task_struct *task)
5673{
5674 struct css_set *cset = task_css_set(task);
5675 struct cgroup_subsys *ss;
5676 int ssid;
5677
5678 for_each_subsys_which(ss, ssid, &have_free_callback)
5679 ss->free(task);
5680
5681 put_css_set(cset);
5682}
5683
5684static void check_for_release(struct cgroup *cgrp)
5685{
5686 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5687 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5688 schedule_work(&cgrp->release_agent_work);
5689}
5690
5691/*
5692 * Notify userspace when a cgroup is released, by running the
5693 * configured release agent with the name of the cgroup (path
5694 * relative to the root of cgroup file system) as the argument.
5695 *
5696 * Most likely, this user command will try to rmdir this cgroup.
5697 *
5698 * This races with the possibility that some other task will be
5699 * attached to this cgroup before it is removed, or that some other
5700 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5701 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5702 * unused, and this cgroup will be reprieved from its death sentence,
5703 * to continue to serve a useful existence. Next time it's released,
5704 * we will get notified again, if it still has 'notify_on_release' set.
5705 *
5706 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5707 * means only wait until the task is successfully execve()'d. The
5708 * separate release agent task is forked by call_usermodehelper(),
5709 * then control in this thread returns here, without waiting for the
5710 * release agent task. We don't bother to wait because the caller of
5711 * this routine has no use for the exit status of the release agent
5712 * task, so no sense holding our caller up for that.
5713 */
5714static void cgroup_release_agent(struct work_struct *work)
5715{
5716 struct cgroup *cgrp =
5717 container_of(work, struct cgroup, release_agent_work);
5718 char *pathbuf = NULL, *agentbuf = NULL, *path;
5719 char *argv[3], *envp[3];
5720
5721 mutex_lock(&cgroup_mutex);
5722
5723 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5724 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5725 if (!pathbuf || !agentbuf)
5726 goto out;
5727
5728 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5729 if (!path)
5730 goto out;
5731
5732 argv[0] = agentbuf;
5733 argv[1] = path;
5734 argv[2] = NULL;
5735
5736 /* minimal command environment */
5737 envp[0] = "HOME=/";
5738 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5739 envp[2] = NULL;
5740
5741 mutex_unlock(&cgroup_mutex);
5742 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5743 goto out_free;
5744out:
5745 mutex_unlock(&cgroup_mutex);
5746out_free:
5747 kfree(agentbuf);
5748 kfree(pathbuf);
5749}
5750
5751static int __init cgroup_disable(char *str)
5752{
5753 struct cgroup_subsys *ss;
5754 char *token;
5755 int i;
5756
5757 while ((token = strsep(&str, ",")) != NULL) {
5758 if (!*token)
5759 continue;
5760
5761 for_each_subsys(ss, i) {
5762 if (strcmp(token, ss->name) &&
5763 strcmp(token, ss->legacy_name))
5764 continue;
5765 cgroup_disable_mask |= 1 << i;
5766 }
5767 }
5768 return 1;
5769}
5770__setup("cgroup_disable=", cgroup_disable);
5771
5772/**
5773 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5774 * @dentry: directory dentry of interest
5775 * @ss: subsystem of interest
5776 *
5777 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5778 * to get the corresponding css and return it. If such css doesn't exist
5779 * or can't be pinned, an ERR_PTR value is returned.
5780 */
5781struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5782 struct cgroup_subsys *ss)
5783{
5784 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5785 struct cgroup_subsys_state *css = NULL;
5786 struct cgroup *cgrp;
5787
5788 /* is @dentry a cgroup dir? */
5789 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5790 kernfs_type(kn) != KERNFS_DIR)
5791 return ERR_PTR(-EBADF);
5792
5793 rcu_read_lock();
5794
5795 /*
5796 * This path doesn't originate from kernfs and @kn could already
5797 * have been or be removed at any point. @kn->priv is RCU
5798 * protected for this access. See css_release_work_fn() for details.
5799 */
5800 cgrp = rcu_dereference(kn->priv);
5801 if (cgrp)
5802 css = cgroup_css(cgrp, ss);
5803
5804 if (!css || !css_tryget_online(css))
5805 css = ERR_PTR(-ENOENT);
5806
5807 rcu_read_unlock();
5808 return css;
5809}
5810
5811/**
5812 * css_from_id - lookup css by id
5813 * @id: the cgroup id
5814 * @ss: cgroup subsys to be looked into
5815 *
5816 * Returns the css if there's valid one with @id, otherwise returns NULL.
5817 * Should be called under rcu_read_lock().
5818 */
5819struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5820{
5821 WARN_ON_ONCE(!rcu_read_lock_held());
5822 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5823}
5824
5825#ifdef CONFIG_CGROUP_DEBUG
5826static struct cgroup_subsys_state *
5827debug_css_alloc(struct cgroup_subsys_state *parent_css)
5828{
5829 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5830
5831 if (!css)
5832 return ERR_PTR(-ENOMEM);
5833
5834 return css;
5835}
5836
5837static void debug_css_free(struct cgroup_subsys_state *css)
5838{
5839 kfree(css);
5840}
5841
5842static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5843 struct cftype *cft)
5844{
5845 return cgroup_task_count(css->cgroup);
5846}
5847
5848static u64 current_css_set_read(struct cgroup_subsys_state *css,
5849 struct cftype *cft)
5850{
5851 return (u64)(unsigned long)current->cgroups;
5852}
5853
5854static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5855 struct cftype *cft)
5856{
5857 u64 count;
5858
5859 rcu_read_lock();
5860 count = atomic_read(&task_css_set(current)->refcount);
5861 rcu_read_unlock();
5862 return count;
5863}
5864
5865static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5866{
5867 struct cgrp_cset_link *link;
5868 struct css_set *cset;
5869 char *name_buf;
5870
5871 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5872 if (!name_buf)
5873 return -ENOMEM;
5874
5875 spin_lock_bh(&css_set_lock);
5876 rcu_read_lock();
5877 cset = rcu_dereference(current->cgroups);
5878 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5879 struct cgroup *c = link->cgrp;
5880
5881 cgroup_name(c, name_buf, NAME_MAX + 1);
5882 seq_printf(seq, "Root %d group %s\n",
5883 c->root->hierarchy_id, name_buf);
5884 }
5885 rcu_read_unlock();
5886 spin_unlock_bh(&css_set_lock);
5887 kfree(name_buf);
5888 return 0;
5889}
5890
5891#define MAX_TASKS_SHOWN_PER_CSS 25
5892static int cgroup_css_links_read(struct seq_file *seq, void *v)
5893{
5894 struct cgroup_subsys_state *css = seq_css(seq);
5895 struct cgrp_cset_link *link;
5896
5897 spin_lock_bh(&css_set_lock);
5898 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5899 struct css_set *cset = link->cset;
5900 struct task_struct *task;
5901 int count = 0;
5902
5903 seq_printf(seq, "css_set %p\n", cset);
5904
5905 list_for_each_entry(task, &cset->tasks, cg_list) {
5906 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5907 goto overflow;
5908 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5909 }
5910
5911 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5912 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5913 goto overflow;
5914 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5915 }
5916 continue;
5917 overflow:
5918 seq_puts(seq, " ...\n");
5919 }
5920 spin_unlock_bh(&css_set_lock);
5921 return 0;
5922}
5923
5924static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5925{
5926 return (!cgroup_is_populated(css->cgroup) &&
5927 !css_has_online_children(&css->cgroup->self));
5928}
5929
5930static struct cftype debug_files[] = {
5931 {
5932 .name = "taskcount",
5933 .read_u64 = debug_taskcount_read,
5934 },
5935
5936 {
5937 .name = "current_css_set",
5938 .read_u64 = current_css_set_read,
5939 },
5940
5941 {
5942 .name = "current_css_set_refcount",
5943 .read_u64 = current_css_set_refcount_read,
5944 },
5945
5946 {
5947 .name = "current_css_set_cg_links",
5948 .seq_show = current_css_set_cg_links_read,
5949 },
5950
5951 {
5952 .name = "cgroup_css_links",
5953 .seq_show = cgroup_css_links_read,
5954 },
5955
5956 {
5957 .name = "releasable",
5958 .read_u64 = releasable_read,
5959 },
5960
5961 { } /* terminate */
5962};
5963
5964struct cgroup_subsys debug_cgrp_subsys = {
5965 .css_alloc = debug_css_alloc,
5966 .css_free = debug_css_free,
5967 .legacy_cftypes = debug_files,
5968};
5969#endif /* CONFIG_CGROUP_DEBUG */