]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - kernel/cgroup.c
Merge tag 'docs-for-linus' of git://git.lwn.net/linux
[mirror_ubuntu-artful-kernel.git] / kernel / cgroup.c
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 #include <linux/atomic.h>
61 #include <linux/cpuset.h>
62 #include <linux/proc_ns.h>
63 #include <linux/nsproxy.h>
64 #include <linux/proc_ns.h>
65 #include <net/sock.h>
66
67 /*
68 * pidlists linger the following amount before being destroyed. The goal
69 * is avoiding frequent destruction in the middle of consecutive read calls
70 * Expiring in the middle is a performance problem not a correctness one.
71 * 1 sec should be enough.
72 */
73 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
74
75 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 MAX_CFTYPE_NAME + 2)
77
78 /*
79 * cgroup_mutex is the master lock. Any modification to cgroup or its
80 * hierarchy must be performed while holding it.
81 *
82 * css_set_lock protects task->cgroups pointer, the list of css_set
83 * objects, and the chain of tasks off each css_set.
84 *
85 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
86 * cgroup.h can use them for lockdep annotations.
87 */
88 #ifdef CONFIG_PROVE_RCU
89 DEFINE_MUTEX(cgroup_mutex);
90 DEFINE_SPINLOCK(css_set_lock);
91 EXPORT_SYMBOL_GPL(cgroup_mutex);
92 EXPORT_SYMBOL_GPL(css_set_lock);
93 #else
94 static DEFINE_MUTEX(cgroup_mutex);
95 static DEFINE_SPINLOCK(css_set_lock);
96 #endif
97
98 /*
99 * Protects cgroup_idr and css_idr so that IDs can be released without
100 * grabbing cgroup_mutex.
101 */
102 static DEFINE_SPINLOCK(cgroup_idr_lock);
103
104 /*
105 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
106 * against file removal/re-creation across css hiding.
107 */
108 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
109
110 /*
111 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
112 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
113 */
114 static DEFINE_SPINLOCK(release_agent_path_lock);
115
116 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
117
118 #define cgroup_assert_mutex_or_rcu_locked() \
119 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
120 !lockdep_is_held(&cgroup_mutex), \
121 "cgroup_mutex or RCU read lock required");
122
123 /*
124 * cgroup destruction makes heavy use of work items and there can be a lot
125 * of concurrent destructions. Use a separate workqueue so that cgroup
126 * destruction work items don't end up filling up max_active of system_wq
127 * which may lead to deadlock.
128 */
129 static struct workqueue_struct *cgroup_destroy_wq;
130
131 /*
132 * pidlist destructions need to be flushed on cgroup destruction. Use a
133 * separate workqueue as flush domain.
134 */
135 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
136
137 /* generate an array of cgroup subsystem pointers */
138 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
139 static struct cgroup_subsys *cgroup_subsys[] = {
140 #include <linux/cgroup_subsys.h>
141 };
142 #undef SUBSYS
143
144 /* array of cgroup subsystem names */
145 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
146 static const char *cgroup_subsys_name[] = {
147 #include <linux/cgroup_subsys.h>
148 };
149 #undef SUBSYS
150
151 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
152 #define SUBSYS(_x) \
153 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
154 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
155 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
156 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
157 #include <linux/cgroup_subsys.h>
158 #undef SUBSYS
159
160 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
161 static struct static_key_true *cgroup_subsys_enabled_key[] = {
162 #include <linux/cgroup_subsys.h>
163 };
164 #undef SUBSYS
165
166 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
167 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
168 #include <linux/cgroup_subsys.h>
169 };
170 #undef SUBSYS
171
172 /*
173 * The default hierarchy, reserved for the subsystems that are otherwise
174 * unattached - it never has more than a single cgroup, and all tasks are
175 * part of that cgroup.
176 */
177 struct cgroup_root cgrp_dfl_root;
178 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
179
180 /*
181 * The default hierarchy always exists but is hidden until mounted for the
182 * first time. This is for backward compatibility.
183 */
184 static bool cgrp_dfl_visible;
185
186 /* Controllers blocked by the commandline in v1 */
187 static u16 cgroup_no_v1_mask;
188
189 /* some controllers are not supported in the default hierarchy */
190 static u16 cgrp_dfl_inhibit_ss_mask;
191
192 /* some controllers are implicitly enabled on the default hierarchy */
193 static unsigned long cgrp_dfl_implicit_ss_mask;
194
195 /* The list of hierarchy roots */
196
197 static LIST_HEAD(cgroup_roots);
198 static int cgroup_root_count;
199
200 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
201 static DEFINE_IDR(cgroup_hierarchy_idr);
202
203 /*
204 * Assign a monotonically increasing serial number to csses. It guarantees
205 * cgroups with bigger numbers are newer than those with smaller numbers.
206 * Also, as csses are always appended to the parent's ->children list, it
207 * guarantees that sibling csses are always sorted in the ascending serial
208 * number order on the list. Protected by cgroup_mutex.
209 */
210 static u64 css_serial_nr_next = 1;
211
212 /*
213 * These bitmask flags indicate whether tasks in the fork and exit paths have
214 * fork/exit handlers to call. This avoids us having to do extra work in the
215 * fork/exit path to check which subsystems have fork/exit callbacks.
216 */
217 static u16 have_fork_callback __read_mostly;
218 static u16 have_exit_callback __read_mostly;
219 static u16 have_free_callback __read_mostly;
220
221 /* cgroup namespace for init task */
222 struct cgroup_namespace init_cgroup_ns = {
223 .count = { .counter = 2, },
224 .user_ns = &init_user_ns,
225 .ns.ops = &cgroupns_operations,
226 .ns.inum = PROC_CGROUP_INIT_INO,
227 .root_cset = &init_css_set,
228 };
229
230 /* Ditto for the can_fork callback. */
231 static u16 have_canfork_callback __read_mostly;
232
233 static struct file_system_type cgroup2_fs_type;
234 static struct cftype cgroup_dfl_base_files[];
235 static struct cftype cgroup_legacy_base_files[];
236
237 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
238 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
239 static int cgroup_apply_control(struct cgroup *cgrp);
240 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
241 static void css_task_iter_advance(struct css_task_iter *it);
242 static int cgroup_destroy_locked(struct cgroup *cgrp);
243 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
244 struct cgroup_subsys *ss);
245 static void css_release(struct percpu_ref *ref);
246 static void kill_css(struct cgroup_subsys_state *css);
247 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
248 struct cgroup *cgrp, struct cftype cfts[],
249 bool is_add);
250
251 /**
252 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
253 * @ssid: subsys ID of interest
254 *
255 * cgroup_subsys_enabled() can only be used with literal subsys names which
256 * is fine for individual subsystems but unsuitable for cgroup core. This
257 * is slower static_key_enabled() based test indexed by @ssid.
258 */
259 static bool cgroup_ssid_enabled(int ssid)
260 {
261 if (CGROUP_SUBSYS_COUNT == 0)
262 return false;
263
264 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
265 }
266
267 static bool cgroup_ssid_no_v1(int ssid)
268 {
269 return cgroup_no_v1_mask & (1 << ssid);
270 }
271
272 /**
273 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
274 * @cgrp: the cgroup of interest
275 *
276 * The default hierarchy is the v2 interface of cgroup and this function
277 * can be used to test whether a cgroup is on the default hierarchy for
278 * cases where a subsystem should behave differnetly depending on the
279 * interface version.
280 *
281 * The set of behaviors which change on the default hierarchy are still
282 * being determined and the mount option is prefixed with __DEVEL__.
283 *
284 * List of changed behaviors:
285 *
286 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
287 * and "name" are disallowed.
288 *
289 * - When mounting an existing superblock, mount options should match.
290 *
291 * - Remount is disallowed.
292 *
293 * - rename(2) is disallowed.
294 *
295 * - "tasks" is removed. Everything should be at process granularity. Use
296 * "cgroup.procs" instead.
297 *
298 * - "cgroup.procs" is not sorted. pids will be unique unless they got
299 * recycled inbetween reads.
300 *
301 * - "release_agent" and "notify_on_release" are removed. Replacement
302 * notification mechanism will be implemented.
303 *
304 * - "cgroup.clone_children" is removed.
305 *
306 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
307 * and its descendants contain no task; otherwise, 1. The file also
308 * generates kernfs notification which can be monitored through poll and
309 * [di]notify when the value of the file changes.
310 *
311 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
312 * take masks of ancestors with non-empty cpus/mems, instead of being
313 * moved to an ancestor.
314 *
315 * - cpuset: a task can be moved into an empty cpuset, and again it takes
316 * masks of ancestors.
317 *
318 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
319 * is not created.
320 *
321 * - blkcg: blk-throttle becomes properly hierarchical.
322 *
323 * - debug: disallowed on the default hierarchy.
324 */
325 static bool cgroup_on_dfl(const struct cgroup *cgrp)
326 {
327 return cgrp->root == &cgrp_dfl_root;
328 }
329
330 /* IDR wrappers which synchronize using cgroup_idr_lock */
331 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
332 gfp_t gfp_mask)
333 {
334 int ret;
335
336 idr_preload(gfp_mask);
337 spin_lock_bh(&cgroup_idr_lock);
338 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
339 spin_unlock_bh(&cgroup_idr_lock);
340 idr_preload_end();
341 return ret;
342 }
343
344 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
345 {
346 void *ret;
347
348 spin_lock_bh(&cgroup_idr_lock);
349 ret = idr_replace(idr, ptr, id);
350 spin_unlock_bh(&cgroup_idr_lock);
351 return ret;
352 }
353
354 static void cgroup_idr_remove(struct idr *idr, int id)
355 {
356 spin_lock_bh(&cgroup_idr_lock);
357 idr_remove(idr, id);
358 spin_unlock_bh(&cgroup_idr_lock);
359 }
360
361 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
362 {
363 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
364
365 if (parent_css)
366 return container_of(parent_css, struct cgroup, self);
367 return NULL;
368 }
369
370 /* subsystems visibly enabled on a cgroup */
371 static u16 cgroup_control(struct cgroup *cgrp)
372 {
373 struct cgroup *parent = cgroup_parent(cgrp);
374 u16 root_ss_mask = cgrp->root->subsys_mask;
375
376 if (parent)
377 return parent->subtree_control;
378
379 if (cgroup_on_dfl(cgrp))
380 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
381 cgrp_dfl_implicit_ss_mask);
382 return root_ss_mask;
383 }
384
385 /* subsystems enabled on a cgroup */
386 static u16 cgroup_ss_mask(struct cgroup *cgrp)
387 {
388 struct cgroup *parent = cgroup_parent(cgrp);
389
390 if (parent)
391 return parent->subtree_ss_mask;
392
393 return cgrp->root->subsys_mask;
394 }
395
396 /**
397 * cgroup_css - obtain a cgroup's css for the specified subsystem
398 * @cgrp: the cgroup of interest
399 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
400 *
401 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
402 * function must be called either under cgroup_mutex or rcu_read_lock() and
403 * the caller is responsible for pinning the returned css if it wants to
404 * keep accessing it outside the said locks. This function may return
405 * %NULL if @cgrp doesn't have @subsys_id enabled.
406 */
407 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
408 struct cgroup_subsys *ss)
409 {
410 if (ss)
411 return rcu_dereference_check(cgrp->subsys[ss->id],
412 lockdep_is_held(&cgroup_mutex));
413 else
414 return &cgrp->self;
415 }
416
417 /**
418 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
419 * @cgrp: the cgroup of interest
420 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
421 *
422 * Similar to cgroup_css() but returns the effective css, which is defined
423 * as the matching css of the nearest ancestor including self which has @ss
424 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
425 * function is guaranteed to return non-NULL css.
426 */
427 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
428 struct cgroup_subsys *ss)
429 {
430 lockdep_assert_held(&cgroup_mutex);
431
432 if (!ss)
433 return &cgrp->self;
434
435 /*
436 * This function is used while updating css associations and thus
437 * can't test the csses directly. Test ss_mask.
438 */
439 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
440 cgrp = cgroup_parent(cgrp);
441 if (!cgrp)
442 return NULL;
443 }
444
445 return cgroup_css(cgrp, ss);
446 }
447
448 /**
449 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
450 * @cgrp: the cgroup of interest
451 * @ss: the subsystem of interest
452 *
453 * Find and get the effective css of @cgrp for @ss. The effective css is
454 * defined as the matching css of the nearest ancestor including self which
455 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
456 * the root css is returned, so this function always returns a valid css.
457 * The returned css must be put using css_put().
458 */
459 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
460 struct cgroup_subsys *ss)
461 {
462 struct cgroup_subsys_state *css;
463
464 rcu_read_lock();
465
466 do {
467 css = cgroup_css(cgrp, ss);
468
469 if (css && css_tryget_online(css))
470 goto out_unlock;
471 cgrp = cgroup_parent(cgrp);
472 } while (cgrp);
473
474 css = init_css_set.subsys[ss->id];
475 css_get(css);
476 out_unlock:
477 rcu_read_unlock();
478 return css;
479 }
480
481 /* convenient tests for these bits */
482 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
483 {
484 return !(cgrp->self.flags & CSS_ONLINE);
485 }
486
487 static void cgroup_get(struct cgroup *cgrp)
488 {
489 WARN_ON_ONCE(cgroup_is_dead(cgrp));
490 css_get(&cgrp->self);
491 }
492
493 static bool cgroup_tryget(struct cgroup *cgrp)
494 {
495 return css_tryget(&cgrp->self);
496 }
497
498 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
499 {
500 struct cgroup *cgrp = of->kn->parent->priv;
501 struct cftype *cft = of_cft(of);
502
503 /*
504 * This is open and unprotected implementation of cgroup_css().
505 * seq_css() is only called from a kernfs file operation which has
506 * an active reference on the file. Because all the subsystem
507 * files are drained before a css is disassociated with a cgroup,
508 * the matching css from the cgroup's subsys table is guaranteed to
509 * be and stay valid until the enclosing operation is complete.
510 */
511 if (cft->ss)
512 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
513 else
514 return &cgrp->self;
515 }
516 EXPORT_SYMBOL_GPL(of_css);
517
518 static int notify_on_release(const struct cgroup *cgrp)
519 {
520 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
521 }
522
523 /**
524 * for_each_css - iterate all css's of a cgroup
525 * @css: the iteration cursor
526 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
527 * @cgrp: the target cgroup to iterate css's of
528 *
529 * Should be called under cgroup_[tree_]mutex.
530 */
531 #define for_each_css(css, ssid, cgrp) \
532 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
533 if (!((css) = rcu_dereference_check( \
534 (cgrp)->subsys[(ssid)], \
535 lockdep_is_held(&cgroup_mutex)))) { } \
536 else
537
538 /**
539 * for_each_e_css - iterate all effective css's of a cgroup
540 * @css: the iteration cursor
541 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
542 * @cgrp: the target cgroup to iterate css's of
543 *
544 * Should be called under cgroup_[tree_]mutex.
545 */
546 #define for_each_e_css(css, ssid, cgrp) \
547 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
548 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
549 ; \
550 else
551
552 /**
553 * for_each_subsys - iterate all enabled cgroup subsystems
554 * @ss: the iteration cursor
555 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
556 */
557 #define for_each_subsys(ss, ssid) \
558 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
559 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
560
561 /**
562 * do_each_subsys_mask - filter for_each_subsys with a bitmask
563 * @ss: the iteration cursor
564 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
565 * @ss_mask: the bitmask
566 *
567 * The block will only run for cases where the ssid-th bit (1 << ssid) of
568 * @ss_mask is set.
569 */
570 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
571 unsigned long __ss_mask = (ss_mask); \
572 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
573 (ssid) = 0; \
574 break; \
575 } \
576 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
577 (ss) = cgroup_subsys[ssid]; \
578 {
579
580 #define while_each_subsys_mask() \
581 } \
582 } \
583 } while (false)
584
585 /* iterate across the hierarchies */
586 #define for_each_root(root) \
587 list_for_each_entry((root), &cgroup_roots, root_list)
588
589 /* iterate over child cgrps, lock should be held throughout iteration */
590 #define cgroup_for_each_live_child(child, cgrp) \
591 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
592 if (({ lockdep_assert_held(&cgroup_mutex); \
593 cgroup_is_dead(child); })) \
594 ; \
595 else
596
597 /* walk live descendants in preorder */
598 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
599 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
600 if (({ lockdep_assert_held(&cgroup_mutex); \
601 (dsct) = (d_css)->cgroup; \
602 cgroup_is_dead(dsct); })) \
603 ; \
604 else
605
606 /* walk live descendants in postorder */
607 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
608 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
609 if (({ lockdep_assert_held(&cgroup_mutex); \
610 (dsct) = (d_css)->cgroup; \
611 cgroup_is_dead(dsct); })) \
612 ; \
613 else
614
615 static void cgroup_release_agent(struct work_struct *work);
616 static void check_for_release(struct cgroup *cgrp);
617
618 /*
619 * A cgroup can be associated with multiple css_sets as different tasks may
620 * belong to different cgroups on different hierarchies. In the other
621 * direction, a css_set is naturally associated with multiple cgroups.
622 * This M:N relationship is represented by the following link structure
623 * which exists for each association and allows traversing the associations
624 * from both sides.
625 */
626 struct cgrp_cset_link {
627 /* the cgroup and css_set this link associates */
628 struct cgroup *cgrp;
629 struct css_set *cset;
630
631 /* list of cgrp_cset_links anchored at cgrp->cset_links */
632 struct list_head cset_link;
633
634 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
635 struct list_head cgrp_link;
636 };
637
638 /*
639 * The default css_set - used by init and its children prior to any
640 * hierarchies being mounted. It contains a pointer to the root state
641 * for each subsystem. Also used to anchor the list of css_sets. Not
642 * reference-counted, to improve performance when child cgroups
643 * haven't been created.
644 */
645 struct css_set init_css_set = {
646 .refcount = ATOMIC_INIT(1),
647 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
648 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
649 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
650 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
651 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
652 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
653 };
654
655 static int css_set_count = 1; /* 1 for init_css_set */
656
657 /**
658 * css_set_populated - does a css_set contain any tasks?
659 * @cset: target css_set
660 */
661 static bool css_set_populated(struct css_set *cset)
662 {
663 lockdep_assert_held(&css_set_lock);
664
665 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
666 }
667
668 /**
669 * cgroup_update_populated - updated populated count of a cgroup
670 * @cgrp: the target cgroup
671 * @populated: inc or dec populated count
672 *
673 * One of the css_sets associated with @cgrp is either getting its first
674 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
675 * count is propagated towards root so that a given cgroup's populated_cnt
676 * is zero iff the cgroup and all its descendants don't contain any tasks.
677 *
678 * @cgrp's interface file "cgroup.populated" is zero if
679 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
680 * changes from or to zero, userland is notified that the content of the
681 * interface file has changed. This can be used to detect when @cgrp and
682 * its descendants become populated or empty.
683 */
684 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
685 {
686 lockdep_assert_held(&css_set_lock);
687
688 do {
689 bool trigger;
690
691 if (populated)
692 trigger = !cgrp->populated_cnt++;
693 else
694 trigger = !--cgrp->populated_cnt;
695
696 if (!trigger)
697 break;
698
699 check_for_release(cgrp);
700 cgroup_file_notify(&cgrp->events_file);
701
702 cgrp = cgroup_parent(cgrp);
703 } while (cgrp);
704 }
705
706 /**
707 * css_set_update_populated - update populated state of a css_set
708 * @cset: target css_set
709 * @populated: whether @cset is populated or depopulated
710 *
711 * @cset is either getting the first task or losing the last. Update the
712 * ->populated_cnt of all associated cgroups accordingly.
713 */
714 static void css_set_update_populated(struct css_set *cset, bool populated)
715 {
716 struct cgrp_cset_link *link;
717
718 lockdep_assert_held(&css_set_lock);
719
720 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
721 cgroup_update_populated(link->cgrp, populated);
722 }
723
724 /**
725 * css_set_move_task - move a task from one css_set to another
726 * @task: task being moved
727 * @from_cset: css_set @task currently belongs to (may be NULL)
728 * @to_cset: new css_set @task is being moved to (may be NULL)
729 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
730 *
731 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
732 * css_set, @from_cset can be NULL. If @task is being disassociated
733 * instead of moved, @to_cset can be NULL.
734 *
735 * This function automatically handles populated_cnt updates and
736 * css_task_iter adjustments but the caller is responsible for managing
737 * @from_cset and @to_cset's reference counts.
738 */
739 static void css_set_move_task(struct task_struct *task,
740 struct css_set *from_cset, struct css_set *to_cset,
741 bool use_mg_tasks)
742 {
743 lockdep_assert_held(&css_set_lock);
744
745 if (to_cset && !css_set_populated(to_cset))
746 css_set_update_populated(to_cset, true);
747
748 if (from_cset) {
749 struct css_task_iter *it, *pos;
750
751 WARN_ON_ONCE(list_empty(&task->cg_list));
752
753 /*
754 * @task is leaving, advance task iterators which are
755 * pointing to it so that they can resume at the next
756 * position. Advancing an iterator might remove it from
757 * the list, use safe walk. See css_task_iter_advance*()
758 * for details.
759 */
760 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
761 iters_node)
762 if (it->task_pos == &task->cg_list)
763 css_task_iter_advance(it);
764
765 list_del_init(&task->cg_list);
766 if (!css_set_populated(from_cset))
767 css_set_update_populated(from_cset, false);
768 } else {
769 WARN_ON_ONCE(!list_empty(&task->cg_list));
770 }
771
772 if (to_cset) {
773 /*
774 * We are synchronized through cgroup_threadgroup_rwsem
775 * against PF_EXITING setting such that we can't race
776 * against cgroup_exit() changing the css_set to
777 * init_css_set and dropping the old one.
778 */
779 WARN_ON_ONCE(task->flags & PF_EXITING);
780
781 rcu_assign_pointer(task->cgroups, to_cset);
782 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
783 &to_cset->tasks);
784 }
785 }
786
787 /*
788 * hash table for cgroup groups. This improves the performance to find
789 * an existing css_set. This hash doesn't (currently) take into
790 * account cgroups in empty hierarchies.
791 */
792 #define CSS_SET_HASH_BITS 7
793 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
794
795 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
796 {
797 unsigned long key = 0UL;
798 struct cgroup_subsys *ss;
799 int i;
800
801 for_each_subsys(ss, i)
802 key += (unsigned long)css[i];
803 key = (key >> 16) ^ key;
804
805 return key;
806 }
807
808 static void put_css_set_locked(struct css_set *cset)
809 {
810 struct cgrp_cset_link *link, *tmp_link;
811 struct cgroup_subsys *ss;
812 int ssid;
813
814 lockdep_assert_held(&css_set_lock);
815
816 if (!atomic_dec_and_test(&cset->refcount))
817 return;
818
819 /* This css_set is dead. unlink it and release cgroup and css refs */
820 for_each_subsys(ss, ssid) {
821 list_del(&cset->e_cset_node[ssid]);
822 css_put(cset->subsys[ssid]);
823 }
824 hash_del(&cset->hlist);
825 css_set_count--;
826
827 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
828 list_del(&link->cset_link);
829 list_del(&link->cgrp_link);
830 if (cgroup_parent(link->cgrp))
831 cgroup_put(link->cgrp);
832 kfree(link);
833 }
834
835 kfree_rcu(cset, rcu_head);
836 }
837
838 static void put_css_set(struct css_set *cset)
839 {
840 /*
841 * Ensure that the refcount doesn't hit zero while any readers
842 * can see it. Similar to atomic_dec_and_lock(), but for an
843 * rwlock
844 */
845 if (atomic_add_unless(&cset->refcount, -1, 1))
846 return;
847
848 spin_lock_bh(&css_set_lock);
849 put_css_set_locked(cset);
850 spin_unlock_bh(&css_set_lock);
851 }
852
853 /*
854 * refcounted get/put for css_set objects
855 */
856 static inline void get_css_set(struct css_set *cset)
857 {
858 atomic_inc(&cset->refcount);
859 }
860
861 /**
862 * compare_css_sets - helper function for find_existing_css_set().
863 * @cset: candidate css_set being tested
864 * @old_cset: existing css_set for a task
865 * @new_cgrp: cgroup that's being entered by the task
866 * @template: desired set of css pointers in css_set (pre-calculated)
867 *
868 * Returns true if "cset" matches "old_cset" except for the hierarchy
869 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
870 */
871 static bool compare_css_sets(struct css_set *cset,
872 struct css_set *old_cset,
873 struct cgroup *new_cgrp,
874 struct cgroup_subsys_state *template[])
875 {
876 struct list_head *l1, *l2;
877
878 /*
879 * On the default hierarchy, there can be csets which are
880 * associated with the same set of cgroups but different csses.
881 * Let's first ensure that csses match.
882 */
883 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
884 return false;
885
886 /*
887 * Compare cgroup pointers in order to distinguish between
888 * different cgroups in hierarchies. As different cgroups may
889 * share the same effective css, this comparison is always
890 * necessary.
891 */
892 l1 = &cset->cgrp_links;
893 l2 = &old_cset->cgrp_links;
894 while (1) {
895 struct cgrp_cset_link *link1, *link2;
896 struct cgroup *cgrp1, *cgrp2;
897
898 l1 = l1->next;
899 l2 = l2->next;
900 /* See if we reached the end - both lists are equal length. */
901 if (l1 == &cset->cgrp_links) {
902 BUG_ON(l2 != &old_cset->cgrp_links);
903 break;
904 } else {
905 BUG_ON(l2 == &old_cset->cgrp_links);
906 }
907 /* Locate the cgroups associated with these links. */
908 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
909 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
910 cgrp1 = link1->cgrp;
911 cgrp2 = link2->cgrp;
912 /* Hierarchies should be linked in the same order. */
913 BUG_ON(cgrp1->root != cgrp2->root);
914
915 /*
916 * If this hierarchy is the hierarchy of the cgroup
917 * that's changing, then we need to check that this
918 * css_set points to the new cgroup; if it's any other
919 * hierarchy, then this css_set should point to the
920 * same cgroup as the old css_set.
921 */
922 if (cgrp1->root == new_cgrp->root) {
923 if (cgrp1 != new_cgrp)
924 return false;
925 } else {
926 if (cgrp1 != cgrp2)
927 return false;
928 }
929 }
930 return true;
931 }
932
933 /**
934 * find_existing_css_set - init css array and find the matching css_set
935 * @old_cset: the css_set that we're using before the cgroup transition
936 * @cgrp: the cgroup that we're moving into
937 * @template: out param for the new set of csses, should be clear on entry
938 */
939 static struct css_set *find_existing_css_set(struct css_set *old_cset,
940 struct cgroup *cgrp,
941 struct cgroup_subsys_state *template[])
942 {
943 struct cgroup_root *root = cgrp->root;
944 struct cgroup_subsys *ss;
945 struct css_set *cset;
946 unsigned long key;
947 int i;
948
949 /*
950 * Build the set of subsystem state objects that we want to see in the
951 * new css_set. while subsystems can change globally, the entries here
952 * won't change, so no need for locking.
953 */
954 for_each_subsys(ss, i) {
955 if (root->subsys_mask & (1UL << i)) {
956 /*
957 * @ss is in this hierarchy, so we want the
958 * effective css from @cgrp.
959 */
960 template[i] = cgroup_e_css(cgrp, ss);
961 } else {
962 /*
963 * @ss is not in this hierarchy, so we don't want
964 * to change the css.
965 */
966 template[i] = old_cset->subsys[i];
967 }
968 }
969
970 key = css_set_hash(template);
971 hash_for_each_possible(css_set_table, cset, hlist, key) {
972 if (!compare_css_sets(cset, old_cset, cgrp, template))
973 continue;
974
975 /* This css_set matches what we need */
976 return cset;
977 }
978
979 /* No existing cgroup group matched */
980 return NULL;
981 }
982
983 static void free_cgrp_cset_links(struct list_head *links_to_free)
984 {
985 struct cgrp_cset_link *link, *tmp_link;
986
987 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
988 list_del(&link->cset_link);
989 kfree(link);
990 }
991 }
992
993 /**
994 * allocate_cgrp_cset_links - allocate cgrp_cset_links
995 * @count: the number of links to allocate
996 * @tmp_links: list_head the allocated links are put on
997 *
998 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
999 * through ->cset_link. Returns 0 on success or -errno.
1000 */
1001 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1002 {
1003 struct cgrp_cset_link *link;
1004 int i;
1005
1006 INIT_LIST_HEAD(tmp_links);
1007
1008 for (i = 0; i < count; i++) {
1009 link = kzalloc(sizeof(*link), GFP_KERNEL);
1010 if (!link) {
1011 free_cgrp_cset_links(tmp_links);
1012 return -ENOMEM;
1013 }
1014 list_add(&link->cset_link, tmp_links);
1015 }
1016 return 0;
1017 }
1018
1019 /**
1020 * link_css_set - a helper function to link a css_set to a cgroup
1021 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1022 * @cset: the css_set to be linked
1023 * @cgrp: the destination cgroup
1024 */
1025 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1026 struct cgroup *cgrp)
1027 {
1028 struct cgrp_cset_link *link;
1029
1030 BUG_ON(list_empty(tmp_links));
1031
1032 if (cgroup_on_dfl(cgrp))
1033 cset->dfl_cgrp = cgrp;
1034
1035 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1036 link->cset = cset;
1037 link->cgrp = cgrp;
1038
1039 /*
1040 * Always add links to the tail of the lists so that the lists are
1041 * in choronological order.
1042 */
1043 list_move_tail(&link->cset_link, &cgrp->cset_links);
1044 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1045
1046 if (cgroup_parent(cgrp))
1047 cgroup_get(cgrp);
1048 }
1049
1050 /**
1051 * find_css_set - return a new css_set with one cgroup updated
1052 * @old_cset: the baseline css_set
1053 * @cgrp: the cgroup to be updated
1054 *
1055 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1056 * substituted into the appropriate hierarchy.
1057 */
1058 static struct css_set *find_css_set(struct css_set *old_cset,
1059 struct cgroup *cgrp)
1060 {
1061 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1062 struct css_set *cset;
1063 struct list_head tmp_links;
1064 struct cgrp_cset_link *link;
1065 struct cgroup_subsys *ss;
1066 unsigned long key;
1067 int ssid;
1068
1069 lockdep_assert_held(&cgroup_mutex);
1070
1071 /* First see if we already have a cgroup group that matches
1072 * the desired set */
1073 spin_lock_bh(&css_set_lock);
1074 cset = find_existing_css_set(old_cset, cgrp, template);
1075 if (cset)
1076 get_css_set(cset);
1077 spin_unlock_bh(&css_set_lock);
1078
1079 if (cset)
1080 return cset;
1081
1082 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1083 if (!cset)
1084 return NULL;
1085
1086 /* Allocate all the cgrp_cset_link objects that we'll need */
1087 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1088 kfree(cset);
1089 return NULL;
1090 }
1091
1092 atomic_set(&cset->refcount, 1);
1093 INIT_LIST_HEAD(&cset->cgrp_links);
1094 INIT_LIST_HEAD(&cset->tasks);
1095 INIT_LIST_HEAD(&cset->mg_tasks);
1096 INIT_LIST_HEAD(&cset->mg_preload_node);
1097 INIT_LIST_HEAD(&cset->mg_node);
1098 INIT_LIST_HEAD(&cset->task_iters);
1099 INIT_HLIST_NODE(&cset->hlist);
1100
1101 /* Copy the set of subsystem state objects generated in
1102 * find_existing_css_set() */
1103 memcpy(cset->subsys, template, sizeof(cset->subsys));
1104
1105 spin_lock_bh(&css_set_lock);
1106 /* Add reference counts and links from the new css_set. */
1107 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1108 struct cgroup *c = link->cgrp;
1109
1110 if (c->root == cgrp->root)
1111 c = cgrp;
1112 link_css_set(&tmp_links, cset, c);
1113 }
1114
1115 BUG_ON(!list_empty(&tmp_links));
1116
1117 css_set_count++;
1118
1119 /* Add @cset to the hash table */
1120 key = css_set_hash(cset->subsys);
1121 hash_add(css_set_table, &cset->hlist, key);
1122
1123 for_each_subsys(ss, ssid) {
1124 struct cgroup_subsys_state *css = cset->subsys[ssid];
1125
1126 list_add_tail(&cset->e_cset_node[ssid],
1127 &css->cgroup->e_csets[ssid]);
1128 css_get(css);
1129 }
1130
1131 spin_unlock_bh(&css_set_lock);
1132
1133 return cset;
1134 }
1135
1136 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1137 {
1138 struct cgroup *root_cgrp = kf_root->kn->priv;
1139
1140 return root_cgrp->root;
1141 }
1142
1143 static int cgroup_init_root_id(struct cgroup_root *root)
1144 {
1145 int id;
1146
1147 lockdep_assert_held(&cgroup_mutex);
1148
1149 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1150 if (id < 0)
1151 return id;
1152
1153 root->hierarchy_id = id;
1154 return 0;
1155 }
1156
1157 static void cgroup_exit_root_id(struct cgroup_root *root)
1158 {
1159 lockdep_assert_held(&cgroup_mutex);
1160
1161 if (root->hierarchy_id) {
1162 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1163 root->hierarchy_id = 0;
1164 }
1165 }
1166
1167 static void cgroup_free_root(struct cgroup_root *root)
1168 {
1169 if (root) {
1170 /* hierarchy ID should already have been released */
1171 WARN_ON_ONCE(root->hierarchy_id);
1172
1173 idr_destroy(&root->cgroup_idr);
1174 kfree(root);
1175 }
1176 }
1177
1178 static void cgroup_destroy_root(struct cgroup_root *root)
1179 {
1180 struct cgroup *cgrp = &root->cgrp;
1181 struct cgrp_cset_link *link, *tmp_link;
1182
1183 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1184
1185 BUG_ON(atomic_read(&root->nr_cgrps));
1186 BUG_ON(!list_empty(&cgrp->self.children));
1187
1188 /* Rebind all subsystems back to the default hierarchy */
1189 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1190
1191 /*
1192 * Release all the links from cset_links to this hierarchy's
1193 * root cgroup
1194 */
1195 spin_lock_bh(&css_set_lock);
1196
1197 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1198 list_del(&link->cset_link);
1199 list_del(&link->cgrp_link);
1200 kfree(link);
1201 }
1202
1203 spin_unlock_bh(&css_set_lock);
1204
1205 if (!list_empty(&root->root_list)) {
1206 list_del(&root->root_list);
1207 cgroup_root_count--;
1208 }
1209
1210 cgroup_exit_root_id(root);
1211
1212 mutex_unlock(&cgroup_mutex);
1213
1214 kernfs_destroy_root(root->kf_root);
1215 cgroup_free_root(root);
1216 }
1217
1218 /*
1219 * look up cgroup associated with current task's cgroup namespace on the
1220 * specified hierarchy
1221 */
1222 static struct cgroup *
1223 current_cgns_cgroup_from_root(struct cgroup_root *root)
1224 {
1225 struct cgroup *res = NULL;
1226 struct css_set *cset;
1227
1228 lockdep_assert_held(&css_set_lock);
1229
1230 rcu_read_lock();
1231
1232 cset = current->nsproxy->cgroup_ns->root_cset;
1233 if (cset == &init_css_set) {
1234 res = &root->cgrp;
1235 } else {
1236 struct cgrp_cset_link *link;
1237
1238 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1239 struct cgroup *c = link->cgrp;
1240
1241 if (c->root == root) {
1242 res = c;
1243 break;
1244 }
1245 }
1246 }
1247 rcu_read_unlock();
1248
1249 BUG_ON(!res);
1250 return res;
1251 }
1252
1253 /* look up cgroup associated with given css_set on the specified hierarchy */
1254 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1255 struct cgroup_root *root)
1256 {
1257 struct cgroup *res = NULL;
1258
1259 lockdep_assert_held(&cgroup_mutex);
1260 lockdep_assert_held(&css_set_lock);
1261
1262 if (cset == &init_css_set) {
1263 res = &root->cgrp;
1264 } else {
1265 struct cgrp_cset_link *link;
1266
1267 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1268 struct cgroup *c = link->cgrp;
1269
1270 if (c->root == root) {
1271 res = c;
1272 break;
1273 }
1274 }
1275 }
1276
1277 BUG_ON(!res);
1278 return res;
1279 }
1280
1281 /*
1282 * Return the cgroup for "task" from the given hierarchy. Must be
1283 * called with cgroup_mutex and css_set_lock held.
1284 */
1285 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1286 struct cgroup_root *root)
1287 {
1288 /*
1289 * No need to lock the task - since we hold cgroup_mutex the
1290 * task can't change groups, so the only thing that can happen
1291 * is that it exits and its css is set back to init_css_set.
1292 */
1293 return cset_cgroup_from_root(task_css_set(task), root);
1294 }
1295
1296 /*
1297 * A task must hold cgroup_mutex to modify cgroups.
1298 *
1299 * Any task can increment and decrement the count field without lock.
1300 * So in general, code holding cgroup_mutex can't rely on the count
1301 * field not changing. However, if the count goes to zero, then only
1302 * cgroup_attach_task() can increment it again. Because a count of zero
1303 * means that no tasks are currently attached, therefore there is no
1304 * way a task attached to that cgroup can fork (the other way to
1305 * increment the count). So code holding cgroup_mutex can safely
1306 * assume that if the count is zero, it will stay zero. Similarly, if
1307 * a task holds cgroup_mutex on a cgroup with zero count, it
1308 * knows that the cgroup won't be removed, as cgroup_rmdir()
1309 * needs that mutex.
1310 *
1311 * A cgroup can only be deleted if both its 'count' of using tasks
1312 * is zero, and its list of 'children' cgroups is empty. Since all
1313 * tasks in the system use _some_ cgroup, and since there is always at
1314 * least one task in the system (init, pid == 1), therefore, root cgroup
1315 * always has either children cgroups and/or using tasks. So we don't
1316 * need a special hack to ensure that root cgroup cannot be deleted.
1317 *
1318 * P.S. One more locking exception. RCU is used to guard the
1319 * update of a tasks cgroup pointer by cgroup_attach_task()
1320 */
1321
1322 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1323 static const struct file_operations proc_cgroupstats_operations;
1324
1325 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1326 char *buf)
1327 {
1328 struct cgroup_subsys *ss = cft->ss;
1329
1330 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1331 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1332 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1333 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1334 cft->name);
1335 else
1336 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1337 return buf;
1338 }
1339
1340 /**
1341 * cgroup_file_mode - deduce file mode of a control file
1342 * @cft: the control file in question
1343 *
1344 * S_IRUGO for read, S_IWUSR for write.
1345 */
1346 static umode_t cgroup_file_mode(const struct cftype *cft)
1347 {
1348 umode_t mode = 0;
1349
1350 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1351 mode |= S_IRUGO;
1352
1353 if (cft->write_u64 || cft->write_s64 || cft->write) {
1354 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1355 mode |= S_IWUGO;
1356 else
1357 mode |= S_IWUSR;
1358 }
1359
1360 return mode;
1361 }
1362
1363 /**
1364 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1365 * @subtree_control: the new subtree_control mask to consider
1366 * @this_ss_mask: available subsystems
1367 *
1368 * On the default hierarchy, a subsystem may request other subsystems to be
1369 * enabled together through its ->depends_on mask. In such cases, more
1370 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1371 *
1372 * This function calculates which subsystems need to be enabled if
1373 * @subtree_control is to be applied while restricted to @this_ss_mask.
1374 */
1375 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1376 {
1377 u16 cur_ss_mask = subtree_control;
1378 struct cgroup_subsys *ss;
1379 int ssid;
1380
1381 lockdep_assert_held(&cgroup_mutex);
1382
1383 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1384
1385 while (true) {
1386 u16 new_ss_mask = cur_ss_mask;
1387
1388 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1389 new_ss_mask |= ss->depends_on;
1390 } while_each_subsys_mask();
1391
1392 /*
1393 * Mask out subsystems which aren't available. This can
1394 * happen only if some depended-upon subsystems were bound
1395 * to non-default hierarchies.
1396 */
1397 new_ss_mask &= this_ss_mask;
1398
1399 if (new_ss_mask == cur_ss_mask)
1400 break;
1401 cur_ss_mask = new_ss_mask;
1402 }
1403
1404 return cur_ss_mask;
1405 }
1406
1407 /**
1408 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1409 * @kn: the kernfs_node being serviced
1410 *
1411 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1412 * the method finishes if locking succeeded. Note that once this function
1413 * returns the cgroup returned by cgroup_kn_lock_live() may become
1414 * inaccessible any time. If the caller intends to continue to access the
1415 * cgroup, it should pin it before invoking this function.
1416 */
1417 static void cgroup_kn_unlock(struct kernfs_node *kn)
1418 {
1419 struct cgroup *cgrp;
1420
1421 if (kernfs_type(kn) == KERNFS_DIR)
1422 cgrp = kn->priv;
1423 else
1424 cgrp = kn->parent->priv;
1425
1426 mutex_unlock(&cgroup_mutex);
1427
1428 kernfs_unbreak_active_protection(kn);
1429 cgroup_put(cgrp);
1430 }
1431
1432 /**
1433 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1434 * @kn: the kernfs_node being serviced
1435 * @drain_offline: perform offline draining on the cgroup
1436 *
1437 * This helper is to be used by a cgroup kernfs method currently servicing
1438 * @kn. It breaks the active protection, performs cgroup locking and
1439 * verifies that the associated cgroup is alive. Returns the cgroup if
1440 * alive; otherwise, %NULL. A successful return should be undone by a
1441 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1442 * cgroup is drained of offlining csses before return.
1443 *
1444 * Any cgroup kernfs method implementation which requires locking the
1445 * associated cgroup should use this helper. It avoids nesting cgroup
1446 * locking under kernfs active protection and allows all kernfs operations
1447 * including self-removal.
1448 */
1449 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1450 bool drain_offline)
1451 {
1452 struct cgroup *cgrp;
1453
1454 if (kernfs_type(kn) == KERNFS_DIR)
1455 cgrp = kn->priv;
1456 else
1457 cgrp = kn->parent->priv;
1458
1459 /*
1460 * We're gonna grab cgroup_mutex which nests outside kernfs
1461 * active_ref. cgroup liveliness check alone provides enough
1462 * protection against removal. Ensure @cgrp stays accessible and
1463 * break the active_ref protection.
1464 */
1465 if (!cgroup_tryget(cgrp))
1466 return NULL;
1467 kernfs_break_active_protection(kn);
1468
1469 if (drain_offline)
1470 cgroup_lock_and_drain_offline(cgrp);
1471 else
1472 mutex_lock(&cgroup_mutex);
1473
1474 if (!cgroup_is_dead(cgrp))
1475 return cgrp;
1476
1477 cgroup_kn_unlock(kn);
1478 return NULL;
1479 }
1480
1481 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1482 {
1483 char name[CGROUP_FILE_NAME_MAX];
1484
1485 lockdep_assert_held(&cgroup_mutex);
1486
1487 if (cft->file_offset) {
1488 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1489 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1490
1491 spin_lock_irq(&cgroup_file_kn_lock);
1492 cfile->kn = NULL;
1493 spin_unlock_irq(&cgroup_file_kn_lock);
1494 }
1495
1496 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1497 }
1498
1499 /**
1500 * css_clear_dir - remove subsys files in a cgroup directory
1501 * @css: taget css
1502 */
1503 static void css_clear_dir(struct cgroup_subsys_state *css)
1504 {
1505 struct cgroup *cgrp = css->cgroup;
1506 struct cftype *cfts;
1507
1508 if (!(css->flags & CSS_VISIBLE))
1509 return;
1510
1511 css->flags &= ~CSS_VISIBLE;
1512
1513 list_for_each_entry(cfts, &css->ss->cfts, node)
1514 cgroup_addrm_files(css, cgrp, cfts, false);
1515 }
1516
1517 /**
1518 * css_populate_dir - create subsys files in a cgroup directory
1519 * @css: target css
1520 *
1521 * On failure, no file is added.
1522 */
1523 static int css_populate_dir(struct cgroup_subsys_state *css)
1524 {
1525 struct cgroup *cgrp = css->cgroup;
1526 struct cftype *cfts, *failed_cfts;
1527 int ret;
1528
1529 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1530 return 0;
1531
1532 if (!css->ss) {
1533 if (cgroup_on_dfl(cgrp))
1534 cfts = cgroup_dfl_base_files;
1535 else
1536 cfts = cgroup_legacy_base_files;
1537
1538 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1539 }
1540
1541 list_for_each_entry(cfts, &css->ss->cfts, node) {
1542 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1543 if (ret < 0) {
1544 failed_cfts = cfts;
1545 goto err;
1546 }
1547 }
1548
1549 css->flags |= CSS_VISIBLE;
1550
1551 return 0;
1552 err:
1553 list_for_each_entry(cfts, &css->ss->cfts, node) {
1554 if (cfts == failed_cfts)
1555 break;
1556 cgroup_addrm_files(css, cgrp, cfts, false);
1557 }
1558 return ret;
1559 }
1560
1561 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1562 {
1563 struct cgroup *dcgrp = &dst_root->cgrp;
1564 struct cgroup_subsys *ss;
1565 int ssid, i, ret;
1566
1567 lockdep_assert_held(&cgroup_mutex);
1568
1569 do_each_subsys_mask(ss, ssid, ss_mask) {
1570 /*
1571 * If @ss has non-root csses attached to it, can't move.
1572 * If @ss is an implicit controller, it is exempt from this
1573 * rule and can be stolen.
1574 */
1575 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1576 !ss->implicit_on_dfl)
1577 return -EBUSY;
1578
1579 /* can't move between two non-dummy roots either */
1580 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1581 return -EBUSY;
1582 } while_each_subsys_mask();
1583
1584 do_each_subsys_mask(ss, ssid, ss_mask) {
1585 struct cgroup_root *src_root = ss->root;
1586 struct cgroup *scgrp = &src_root->cgrp;
1587 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1588 struct css_set *cset;
1589
1590 WARN_ON(!css || cgroup_css(dcgrp, ss));
1591
1592 /* disable from the source */
1593 src_root->subsys_mask &= ~(1 << ssid);
1594 WARN_ON(cgroup_apply_control(scgrp));
1595 cgroup_finalize_control(scgrp, 0);
1596
1597 /* rebind */
1598 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1599 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1600 ss->root = dst_root;
1601 css->cgroup = dcgrp;
1602
1603 spin_lock_bh(&css_set_lock);
1604 hash_for_each(css_set_table, i, cset, hlist)
1605 list_move_tail(&cset->e_cset_node[ss->id],
1606 &dcgrp->e_csets[ss->id]);
1607 spin_unlock_bh(&css_set_lock);
1608
1609 /* default hierarchy doesn't enable controllers by default */
1610 dst_root->subsys_mask |= 1 << ssid;
1611 if (dst_root == &cgrp_dfl_root) {
1612 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1613 } else {
1614 dcgrp->subtree_control |= 1 << ssid;
1615 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1616 }
1617
1618 ret = cgroup_apply_control(dcgrp);
1619 if (ret)
1620 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1621 ss->name, ret);
1622
1623 if (ss->bind)
1624 ss->bind(css);
1625 } while_each_subsys_mask();
1626
1627 kernfs_activate(dcgrp->kn);
1628 return 0;
1629 }
1630
1631 static int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1632 struct kernfs_root *kf_root)
1633 {
1634 int len = 0;
1635 char *buf = NULL;
1636 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1637 struct cgroup *ns_cgroup;
1638
1639 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1640 if (!buf)
1641 return -ENOMEM;
1642
1643 spin_lock_bh(&css_set_lock);
1644 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1645 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1646 spin_unlock_bh(&css_set_lock);
1647
1648 if (len >= PATH_MAX)
1649 len = -ERANGE;
1650 else if (len > 0) {
1651 seq_escape(sf, buf, " \t\n\\");
1652 len = 0;
1653 }
1654 kfree(buf);
1655 return len;
1656 }
1657
1658 static int cgroup_show_options(struct seq_file *seq,
1659 struct kernfs_root *kf_root)
1660 {
1661 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1662 struct cgroup_subsys *ss;
1663 int ssid;
1664
1665 if (root != &cgrp_dfl_root)
1666 for_each_subsys(ss, ssid)
1667 if (root->subsys_mask & (1 << ssid))
1668 seq_show_option(seq, ss->legacy_name, NULL);
1669 if (root->flags & CGRP_ROOT_NOPREFIX)
1670 seq_puts(seq, ",noprefix");
1671 if (root->flags & CGRP_ROOT_XATTR)
1672 seq_puts(seq, ",xattr");
1673
1674 spin_lock(&release_agent_path_lock);
1675 if (strlen(root->release_agent_path))
1676 seq_show_option(seq, "release_agent",
1677 root->release_agent_path);
1678 spin_unlock(&release_agent_path_lock);
1679
1680 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1681 seq_puts(seq, ",clone_children");
1682 if (strlen(root->name))
1683 seq_show_option(seq, "name", root->name);
1684 return 0;
1685 }
1686
1687 struct cgroup_sb_opts {
1688 u16 subsys_mask;
1689 unsigned int flags;
1690 char *release_agent;
1691 bool cpuset_clone_children;
1692 char *name;
1693 /* User explicitly requested empty subsystem */
1694 bool none;
1695 };
1696
1697 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1698 {
1699 char *token, *o = data;
1700 bool all_ss = false, one_ss = false;
1701 u16 mask = U16_MAX;
1702 struct cgroup_subsys *ss;
1703 int nr_opts = 0;
1704 int i;
1705
1706 #ifdef CONFIG_CPUSETS
1707 mask = ~((u16)1 << cpuset_cgrp_id);
1708 #endif
1709
1710 memset(opts, 0, sizeof(*opts));
1711
1712 while ((token = strsep(&o, ",")) != NULL) {
1713 nr_opts++;
1714
1715 if (!*token)
1716 return -EINVAL;
1717 if (!strcmp(token, "none")) {
1718 /* Explicitly have no subsystems */
1719 opts->none = true;
1720 continue;
1721 }
1722 if (!strcmp(token, "all")) {
1723 /* Mutually exclusive option 'all' + subsystem name */
1724 if (one_ss)
1725 return -EINVAL;
1726 all_ss = true;
1727 continue;
1728 }
1729 if (!strcmp(token, "noprefix")) {
1730 opts->flags |= CGRP_ROOT_NOPREFIX;
1731 continue;
1732 }
1733 if (!strcmp(token, "clone_children")) {
1734 opts->cpuset_clone_children = true;
1735 continue;
1736 }
1737 if (!strcmp(token, "xattr")) {
1738 opts->flags |= CGRP_ROOT_XATTR;
1739 continue;
1740 }
1741 if (!strncmp(token, "release_agent=", 14)) {
1742 /* Specifying two release agents is forbidden */
1743 if (opts->release_agent)
1744 return -EINVAL;
1745 opts->release_agent =
1746 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1747 if (!opts->release_agent)
1748 return -ENOMEM;
1749 continue;
1750 }
1751 if (!strncmp(token, "name=", 5)) {
1752 const char *name = token + 5;
1753 /* Can't specify an empty name */
1754 if (!strlen(name))
1755 return -EINVAL;
1756 /* Must match [\w.-]+ */
1757 for (i = 0; i < strlen(name); i++) {
1758 char c = name[i];
1759 if (isalnum(c))
1760 continue;
1761 if ((c == '.') || (c == '-') || (c == '_'))
1762 continue;
1763 return -EINVAL;
1764 }
1765 /* Specifying two names is forbidden */
1766 if (opts->name)
1767 return -EINVAL;
1768 opts->name = kstrndup(name,
1769 MAX_CGROUP_ROOT_NAMELEN - 1,
1770 GFP_KERNEL);
1771 if (!opts->name)
1772 return -ENOMEM;
1773
1774 continue;
1775 }
1776
1777 for_each_subsys(ss, i) {
1778 if (strcmp(token, ss->legacy_name))
1779 continue;
1780 if (!cgroup_ssid_enabled(i))
1781 continue;
1782 if (cgroup_ssid_no_v1(i))
1783 continue;
1784
1785 /* Mutually exclusive option 'all' + subsystem name */
1786 if (all_ss)
1787 return -EINVAL;
1788 opts->subsys_mask |= (1 << i);
1789 one_ss = true;
1790
1791 break;
1792 }
1793 if (i == CGROUP_SUBSYS_COUNT)
1794 return -ENOENT;
1795 }
1796
1797 /*
1798 * If the 'all' option was specified select all the subsystems,
1799 * otherwise if 'none', 'name=' and a subsystem name options were
1800 * not specified, let's default to 'all'
1801 */
1802 if (all_ss || (!one_ss && !opts->none && !opts->name))
1803 for_each_subsys(ss, i)
1804 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1805 opts->subsys_mask |= (1 << i);
1806
1807 /*
1808 * We either have to specify by name or by subsystems. (So all
1809 * empty hierarchies must have a name).
1810 */
1811 if (!opts->subsys_mask && !opts->name)
1812 return -EINVAL;
1813
1814 /*
1815 * Option noprefix was introduced just for backward compatibility
1816 * with the old cpuset, so we allow noprefix only if mounting just
1817 * the cpuset subsystem.
1818 */
1819 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1820 return -EINVAL;
1821
1822 /* Can't specify "none" and some subsystems */
1823 if (opts->subsys_mask && opts->none)
1824 return -EINVAL;
1825
1826 return 0;
1827 }
1828
1829 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1830 {
1831 int ret = 0;
1832 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1833 struct cgroup_sb_opts opts;
1834 u16 added_mask, removed_mask;
1835
1836 if (root == &cgrp_dfl_root) {
1837 pr_err("remount is not allowed\n");
1838 return -EINVAL;
1839 }
1840
1841 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1842
1843 /* See what subsystems are wanted */
1844 ret = parse_cgroupfs_options(data, &opts);
1845 if (ret)
1846 goto out_unlock;
1847
1848 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1849 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1850 task_tgid_nr(current), current->comm);
1851
1852 added_mask = opts.subsys_mask & ~root->subsys_mask;
1853 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1854
1855 /* Don't allow flags or name to change at remount */
1856 if ((opts.flags ^ root->flags) ||
1857 (opts.name && strcmp(opts.name, root->name))) {
1858 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1859 opts.flags, opts.name ?: "", root->flags, root->name);
1860 ret = -EINVAL;
1861 goto out_unlock;
1862 }
1863
1864 /* remounting is not allowed for populated hierarchies */
1865 if (!list_empty(&root->cgrp.self.children)) {
1866 ret = -EBUSY;
1867 goto out_unlock;
1868 }
1869
1870 ret = rebind_subsystems(root, added_mask);
1871 if (ret)
1872 goto out_unlock;
1873
1874 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1875
1876 if (opts.release_agent) {
1877 spin_lock(&release_agent_path_lock);
1878 strcpy(root->release_agent_path, opts.release_agent);
1879 spin_unlock(&release_agent_path_lock);
1880 }
1881 out_unlock:
1882 kfree(opts.release_agent);
1883 kfree(opts.name);
1884 mutex_unlock(&cgroup_mutex);
1885 return ret;
1886 }
1887
1888 /*
1889 * To reduce the fork() overhead for systems that are not actually using
1890 * their cgroups capability, we don't maintain the lists running through
1891 * each css_set to its tasks until we see the list actually used - in other
1892 * words after the first mount.
1893 */
1894 static bool use_task_css_set_links __read_mostly;
1895
1896 static void cgroup_enable_task_cg_lists(void)
1897 {
1898 struct task_struct *p, *g;
1899
1900 spin_lock_bh(&css_set_lock);
1901
1902 if (use_task_css_set_links)
1903 goto out_unlock;
1904
1905 use_task_css_set_links = true;
1906
1907 /*
1908 * We need tasklist_lock because RCU is not safe against
1909 * while_each_thread(). Besides, a forking task that has passed
1910 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1911 * is not guaranteed to have its child immediately visible in the
1912 * tasklist if we walk through it with RCU.
1913 */
1914 read_lock(&tasklist_lock);
1915 do_each_thread(g, p) {
1916 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1917 task_css_set(p) != &init_css_set);
1918
1919 /*
1920 * We should check if the process is exiting, otherwise
1921 * it will race with cgroup_exit() in that the list
1922 * entry won't be deleted though the process has exited.
1923 * Do it while holding siglock so that we don't end up
1924 * racing against cgroup_exit().
1925 */
1926 spin_lock_irq(&p->sighand->siglock);
1927 if (!(p->flags & PF_EXITING)) {
1928 struct css_set *cset = task_css_set(p);
1929
1930 if (!css_set_populated(cset))
1931 css_set_update_populated(cset, true);
1932 list_add_tail(&p->cg_list, &cset->tasks);
1933 get_css_set(cset);
1934 }
1935 spin_unlock_irq(&p->sighand->siglock);
1936 } while_each_thread(g, p);
1937 read_unlock(&tasklist_lock);
1938 out_unlock:
1939 spin_unlock_bh(&css_set_lock);
1940 }
1941
1942 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1943 {
1944 struct cgroup_subsys *ss;
1945 int ssid;
1946
1947 INIT_LIST_HEAD(&cgrp->self.sibling);
1948 INIT_LIST_HEAD(&cgrp->self.children);
1949 INIT_LIST_HEAD(&cgrp->cset_links);
1950 INIT_LIST_HEAD(&cgrp->pidlists);
1951 mutex_init(&cgrp->pidlist_mutex);
1952 cgrp->self.cgroup = cgrp;
1953 cgrp->self.flags |= CSS_ONLINE;
1954
1955 for_each_subsys(ss, ssid)
1956 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1957
1958 init_waitqueue_head(&cgrp->offline_waitq);
1959 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1960 }
1961
1962 static void init_cgroup_root(struct cgroup_root *root,
1963 struct cgroup_sb_opts *opts)
1964 {
1965 struct cgroup *cgrp = &root->cgrp;
1966
1967 INIT_LIST_HEAD(&root->root_list);
1968 atomic_set(&root->nr_cgrps, 1);
1969 cgrp->root = root;
1970 init_cgroup_housekeeping(cgrp);
1971 idr_init(&root->cgroup_idr);
1972
1973 root->flags = opts->flags;
1974 if (opts->release_agent)
1975 strcpy(root->release_agent_path, opts->release_agent);
1976 if (opts->name)
1977 strcpy(root->name, opts->name);
1978 if (opts->cpuset_clone_children)
1979 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1980 }
1981
1982 static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1983 {
1984 LIST_HEAD(tmp_links);
1985 struct cgroup *root_cgrp = &root->cgrp;
1986 struct css_set *cset;
1987 int i, ret;
1988
1989 lockdep_assert_held(&cgroup_mutex);
1990
1991 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1992 if (ret < 0)
1993 goto out;
1994 root_cgrp->id = ret;
1995 root_cgrp->ancestor_ids[0] = ret;
1996
1997 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1998 GFP_KERNEL);
1999 if (ret)
2000 goto out;
2001
2002 /*
2003 * We're accessing css_set_count without locking css_set_lock here,
2004 * but that's OK - it can only be increased by someone holding
2005 * cgroup_lock, and that's us. Later rebinding may disable
2006 * controllers on the default hierarchy and thus create new csets,
2007 * which can't be more than the existing ones. Allocate 2x.
2008 */
2009 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2010 if (ret)
2011 goto cancel_ref;
2012
2013 ret = cgroup_init_root_id(root);
2014 if (ret)
2015 goto cancel_ref;
2016
2017 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
2018 KERNFS_ROOT_CREATE_DEACTIVATED,
2019 root_cgrp);
2020 if (IS_ERR(root->kf_root)) {
2021 ret = PTR_ERR(root->kf_root);
2022 goto exit_root_id;
2023 }
2024 root_cgrp->kn = root->kf_root->kn;
2025
2026 ret = css_populate_dir(&root_cgrp->self);
2027 if (ret)
2028 goto destroy_root;
2029
2030 ret = rebind_subsystems(root, ss_mask);
2031 if (ret)
2032 goto destroy_root;
2033
2034 /*
2035 * There must be no failure case after here, since rebinding takes
2036 * care of subsystems' refcounts, which are explicitly dropped in
2037 * the failure exit path.
2038 */
2039 list_add(&root->root_list, &cgroup_roots);
2040 cgroup_root_count++;
2041
2042 /*
2043 * Link the root cgroup in this hierarchy into all the css_set
2044 * objects.
2045 */
2046 spin_lock_bh(&css_set_lock);
2047 hash_for_each(css_set_table, i, cset, hlist) {
2048 link_css_set(&tmp_links, cset, root_cgrp);
2049 if (css_set_populated(cset))
2050 cgroup_update_populated(root_cgrp, true);
2051 }
2052 spin_unlock_bh(&css_set_lock);
2053
2054 BUG_ON(!list_empty(&root_cgrp->self.children));
2055 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2056
2057 kernfs_activate(root_cgrp->kn);
2058 ret = 0;
2059 goto out;
2060
2061 destroy_root:
2062 kernfs_destroy_root(root->kf_root);
2063 root->kf_root = NULL;
2064 exit_root_id:
2065 cgroup_exit_root_id(root);
2066 cancel_ref:
2067 percpu_ref_exit(&root_cgrp->self.refcnt);
2068 out:
2069 free_cgrp_cset_links(&tmp_links);
2070 return ret;
2071 }
2072
2073 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2074 int flags, const char *unused_dev_name,
2075 void *data)
2076 {
2077 bool is_v2 = fs_type == &cgroup2_fs_type;
2078 struct super_block *pinned_sb = NULL;
2079 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2080 struct cgroup_subsys *ss;
2081 struct cgroup_root *root;
2082 struct cgroup_sb_opts opts;
2083 struct dentry *dentry;
2084 int ret;
2085 int i;
2086 bool new_sb;
2087
2088 get_cgroup_ns(ns);
2089
2090 /* Check if the caller has permission to mount. */
2091 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2092 put_cgroup_ns(ns);
2093 return ERR_PTR(-EPERM);
2094 }
2095
2096 /*
2097 * The first time anyone tries to mount a cgroup, enable the list
2098 * linking each css_set to its tasks and fix up all existing tasks.
2099 */
2100 if (!use_task_css_set_links)
2101 cgroup_enable_task_cg_lists();
2102
2103 if (is_v2) {
2104 if (data) {
2105 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2106 put_cgroup_ns(ns);
2107 return ERR_PTR(-EINVAL);
2108 }
2109 cgrp_dfl_visible = true;
2110 root = &cgrp_dfl_root;
2111 cgroup_get(&root->cgrp);
2112 goto out_mount;
2113 }
2114
2115 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2116
2117 /* First find the desired set of subsystems */
2118 ret = parse_cgroupfs_options(data, &opts);
2119 if (ret)
2120 goto out_unlock;
2121
2122 /*
2123 * Destruction of cgroup root is asynchronous, so subsystems may
2124 * still be dying after the previous unmount. Let's drain the
2125 * dying subsystems. We just need to ensure that the ones
2126 * unmounted previously finish dying and don't care about new ones
2127 * starting. Testing ref liveliness is good enough.
2128 */
2129 for_each_subsys(ss, i) {
2130 if (!(opts.subsys_mask & (1 << i)) ||
2131 ss->root == &cgrp_dfl_root)
2132 continue;
2133
2134 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2135 mutex_unlock(&cgroup_mutex);
2136 msleep(10);
2137 ret = restart_syscall();
2138 goto out_free;
2139 }
2140 cgroup_put(&ss->root->cgrp);
2141 }
2142
2143 for_each_root(root) {
2144 bool name_match = false;
2145
2146 if (root == &cgrp_dfl_root)
2147 continue;
2148
2149 /*
2150 * If we asked for a name then it must match. Also, if
2151 * name matches but sybsys_mask doesn't, we should fail.
2152 * Remember whether name matched.
2153 */
2154 if (opts.name) {
2155 if (strcmp(opts.name, root->name))
2156 continue;
2157 name_match = true;
2158 }
2159
2160 /*
2161 * If we asked for subsystems (or explicitly for no
2162 * subsystems) then they must match.
2163 */
2164 if ((opts.subsys_mask || opts.none) &&
2165 (opts.subsys_mask != root->subsys_mask)) {
2166 if (!name_match)
2167 continue;
2168 ret = -EBUSY;
2169 goto out_unlock;
2170 }
2171
2172 if (root->flags ^ opts.flags)
2173 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2174
2175 /*
2176 * We want to reuse @root whose lifetime is governed by its
2177 * ->cgrp. Let's check whether @root is alive and keep it
2178 * that way. As cgroup_kill_sb() can happen anytime, we
2179 * want to block it by pinning the sb so that @root doesn't
2180 * get killed before mount is complete.
2181 *
2182 * With the sb pinned, tryget_live can reliably indicate
2183 * whether @root can be reused. If it's being killed,
2184 * drain it. We can use wait_queue for the wait but this
2185 * path is super cold. Let's just sleep a bit and retry.
2186 */
2187 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2188 if (IS_ERR(pinned_sb) ||
2189 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2190 mutex_unlock(&cgroup_mutex);
2191 if (!IS_ERR_OR_NULL(pinned_sb))
2192 deactivate_super(pinned_sb);
2193 msleep(10);
2194 ret = restart_syscall();
2195 goto out_free;
2196 }
2197
2198 ret = 0;
2199 goto out_unlock;
2200 }
2201
2202 /*
2203 * No such thing, create a new one. name= matching without subsys
2204 * specification is allowed for already existing hierarchies but we
2205 * can't create new one without subsys specification.
2206 */
2207 if (!opts.subsys_mask && !opts.none) {
2208 ret = -EINVAL;
2209 goto out_unlock;
2210 }
2211
2212 /*
2213 * We know this subsystem has not yet been bound. Users in a non-init
2214 * user namespace may only mount hierarchies with no bound subsystems,
2215 * i.e. 'none,name=user1'
2216 */
2217 if (!opts.none && !capable(CAP_SYS_ADMIN)) {
2218 ret = -EPERM;
2219 goto out_unlock;
2220 }
2221
2222 root = kzalloc(sizeof(*root), GFP_KERNEL);
2223 if (!root) {
2224 ret = -ENOMEM;
2225 goto out_unlock;
2226 }
2227
2228 init_cgroup_root(root, &opts);
2229
2230 ret = cgroup_setup_root(root, opts.subsys_mask);
2231 if (ret)
2232 cgroup_free_root(root);
2233
2234 out_unlock:
2235 mutex_unlock(&cgroup_mutex);
2236 out_free:
2237 kfree(opts.release_agent);
2238 kfree(opts.name);
2239
2240 if (ret) {
2241 put_cgroup_ns(ns);
2242 return ERR_PTR(ret);
2243 }
2244 out_mount:
2245 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2246 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2247 &new_sb);
2248
2249 /*
2250 * In non-init cgroup namespace, instead of root cgroup's
2251 * dentry, we return the dentry corresponding to the
2252 * cgroupns->root_cgrp.
2253 */
2254 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2255 struct dentry *nsdentry;
2256 struct cgroup *cgrp;
2257
2258 mutex_lock(&cgroup_mutex);
2259 spin_lock_bh(&css_set_lock);
2260
2261 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2262
2263 spin_unlock_bh(&css_set_lock);
2264 mutex_unlock(&cgroup_mutex);
2265
2266 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2267 dput(dentry);
2268 dentry = nsdentry;
2269 }
2270
2271 if (IS_ERR(dentry) || !new_sb)
2272 cgroup_put(&root->cgrp);
2273
2274 /*
2275 * If @pinned_sb, we're reusing an existing root and holding an
2276 * extra ref on its sb. Mount is complete. Put the extra ref.
2277 */
2278 if (pinned_sb) {
2279 WARN_ON(new_sb);
2280 deactivate_super(pinned_sb);
2281 }
2282
2283 put_cgroup_ns(ns);
2284 return dentry;
2285 }
2286
2287 static void cgroup_kill_sb(struct super_block *sb)
2288 {
2289 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2290 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2291
2292 /*
2293 * If @root doesn't have any mounts or children, start killing it.
2294 * This prevents new mounts by disabling percpu_ref_tryget_live().
2295 * cgroup_mount() may wait for @root's release.
2296 *
2297 * And don't kill the default root.
2298 */
2299 if (!list_empty(&root->cgrp.self.children) ||
2300 root == &cgrp_dfl_root)
2301 cgroup_put(&root->cgrp);
2302 else
2303 percpu_ref_kill(&root->cgrp.self.refcnt);
2304
2305 kernfs_kill_sb(sb);
2306 }
2307
2308 static struct file_system_type cgroup_fs_type = {
2309 .name = "cgroup",
2310 .mount = cgroup_mount,
2311 .kill_sb = cgroup_kill_sb,
2312 .fs_flags = FS_USERNS_MOUNT,
2313 };
2314
2315 static struct file_system_type cgroup2_fs_type = {
2316 .name = "cgroup2",
2317 .mount = cgroup_mount,
2318 .kill_sb = cgroup_kill_sb,
2319 .fs_flags = FS_USERNS_MOUNT,
2320 };
2321
2322 static char *cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2323 struct cgroup_namespace *ns)
2324 {
2325 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2326 int ret;
2327
2328 ret = kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2329 if (ret < 0 || ret >= buflen)
2330 return NULL;
2331 return buf;
2332 }
2333
2334 char *cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2335 struct cgroup_namespace *ns)
2336 {
2337 char *ret;
2338
2339 mutex_lock(&cgroup_mutex);
2340 spin_lock_bh(&css_set_lock);
2341
2342 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2343
2344 spin_unlock_bh(&css_set_lock);
2345 mutex_unlock(&cgroup_mutex);
2346
2347 return ret;
2348 }
2349 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2350
2351 /**
2352 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2353 * @task: target task
2354 * @buf: the buffer to write the path into
2355 * @buflen: the length of the buffer
2356 *
2357 * Determine @task's cgroup on the first (the one with the lowest non-zero
2358 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2359 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2360 * cgroup controller callbacks.
2361 *
2362 * Return value is the same as kernfs_path().
2363 */
2364 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2365 {
2366 struct cgroup_root *root;
2367 struct cgroup *cgrp;
2368 int hierarchy_id = 1;
2369 char *path = NULL;
2370
2371 mutex_lock(&cgroup_mutex);
2372 spin_lock_bh(&css_set_lock);
2373
2374 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2375
2376 if (root) {
2377 cgrp = task_cgroup_from_root(task, root);
2378 path = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2379 } else {
2380 /* if no hierarchy exists, everyone is in "/" */
2381 if (strlcpy(buf, "/", buflen) < buflen)
2382 path = buf;
2383 }
2384
2385 spin_unlock_bh(&css_set_lock);
2386 mutex_unlock(&cgroup_mutex);
2387 return path;
2388 }
2389 EXPORT_SYMBOL_GPL(task_cgroup_path);
2390
2391 /* used to track tasks and other necessary states during migration */
2392 struct cgroup_taskset {
2393 /* the src and dst cset list running through cset->mg_node */
2394 struct list_head src_csets;
2395 struct list_head dst_csets;
2396
2397 /* the subsys currently being processed */
2398 int ssid;
2399
2400 /*
2401 * Fields for cgroup_taskset_*() iteration.
2402 *
2403 * Before migration is committed, the target migration tasks are on
2404 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2405 * the csets on ->dst_csets. ->csets point to either ->src_csets
2406 * or ->dst_csets depending on whether migration is committed.
2407 *
2408 * ->cur_csets and ->cur_task point to the current task position
2409 * during iteration.
2410 */
2411 struct list_head *csets;
2412 struct css_set *cur_cset;
2413 struct task_struct *cur_task;
2414 };
2415
2416 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2417 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2418 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2419 .csets = &tset.src_csets, \
2420 }
2421
2422 /**
2423 * cgroup_taskset_add - try to add a migration target task to a taskset
2424 * @task: target task
2425 * @tset: target taskset
2426 *
2427 * Add @task, which is a migration target, to @tset. This function becomes
2428 * noop if @task doesn't need to be migrated. @task's css_set should have
2429 * been added as a migration source and @task->cg_list will be moved from
2430 * the css_set's tasks list to mg_tasks one.
2431 */
2432 static void cgroup_taskset_add(struct task_struct *task,
2433 struct cgroup_taskset *tset)
2434 {
2435 struct css_set *cset;
2436
2437 lockdep_assert_held(&css_set_lock);
2438
2439 /* @task either already exited or can't exit until the end */
2440 if (task->flags & PF_EXITING)
2441 return;
2442
2443 /* leave @task alone if post_fork() hasn't linked it yet */
2444 if (list_empty(&task->cg_list))
2445 return;
2446
2447 cset = task_css_set(task);
2448 if (!cset->mg_src_cgrp)
2449 return;
2450
2451 list_move_tail(&task->cg_list, &cset->mg_tasks);
2452 if (list_empty(&cset->mg_node))
2453 list_add_tail(&cset->mg_node, &tset->src_csets);
2454 if (list_empty(&cset->mg_dst_cset->mg_node))
2455 list_move_tail(&cset->mg_dst_cset->mg_node,
2456 &tset->dst_csets);
2457 }
2458
2459 /**
2460 * cgroup_taskset_first - reset taskset and return the first task
2461 * @tset: taskset of interest
2462 * @dst_cssp: output variable for the destination css
2463 *
2464 * @tset iteration is initialized and the first task is returned.
2465 */
2466 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2467 struct cgroup_subsys_state **dst_cssp)
2468 {
2469 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2470 tset->cur_task = NULL;
2471
2472 return cgroup_taskset_next(tset, dst_cssp);
2473 }
2474
2475 /**
2476 * cgroup_taskset_next - iterate to the next task in taskset
2477 * @tset: taskset of interest
2478 * @dst_cssp: output variable for the destination css
2479 *
2480 * Return the next task in @tset. Iteration must have been initialized
2481 * with cgroup_taskset_first().
2482 */
2483 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2484 struct cgroup_subsys_state **dst_cssp)
2485 {
2486 struct css_set *cset = tset->cur_cset;
2487 struct task_struct *task = tset->cur_task;
2488
2489 while (&cset->mg_node != tset->csets) {
2490 if (!task)
2491 task = list_first_entry(&cset->mg_tasks,
2492 struct task_struct, cg_list);
2493 else
2494 task = list_next_entry(task, cg_list);
2495
2496 if (&task->cg_list != &cset->mg_tasks) {
2497 tset->cur_cset = cset;
2498 tset->cur_task = task;
2499
2500 /*
2501 * This function may be called both before and
2502 * after cgroup_taskset_migrate(). The two cases
2503 * can be distinguished by looking at whether @cset
2504 * has its ->mg_dst_cset set.
2505 */
2506 if (cset->mg_dst_cset)
2507 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2508 else
2509 *dst_cssp = cset->subsys[tset->ssid];
2510
2511 return task;
2512 }
2513
2514 cset = list_next_entry(cset, mg_node);
2515 task = NULL;
2516 }
2517
2518 return NULL;
2519 }
2520
2521 /**
2522 * cgroup_taskset_migrate - migrate a taskset
2523 * @tset: taget taskset
2524 * @root: cgroup root the migration is taking place on
2525 *
2526 * Migrate tasks in @tset as setup by migration preparation functions.
2527 * This function fails iff one of the ->can_attach callbacks fails and
2528 * guarantees that either all or none of the tasks in @tset are migrated.
2529 * @tset is consumed regardless of success.
2530 */
2531 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2532 struct cgroup_root *root)
2533 {
2534 struct cgroup_subsys *ss;
2535 struct task_struct *task, *tmp_task;
2536 struct css_set *cset, *tmp_cset;
2537 int ssid, failed_ssid, ret;
2538
2539 /* methods shouldn't be called if no task is actually migrating */
2540 if (list_empty(&tset->src_csets))
2541 return 0;
2542
2543 /* check that we can legitimately attach to the cgroup */
2544 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2545 if (ss->can_attach) {
2546 tset->ssid = ssid;
2547 ret = ss->can_attach(tset);
2548 if (ret) {
2549 failed_ssid = ssid;
2550 goto out_cancel_attach;
2551 }
2552 }
2553 } while_each_subsys_mask();
2554
2555 /*
2556 * Now that we're guaranteed success, proceed to move all tasks to
2557 * the new cgroup. There are no failure cases after here, so this
2558 * is the commit point.
2559 */
2560 spin_lock_bh(&css_set_lock);
2561 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2562 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2563 struct css_set *from_cset = task_css_set(task);
2564 struct css_set *to_cset = cset->mg_dst_cset;
2565
2566 get_css_set(to_cset);
2567 css_set_move_task(task, from_cset, to_cset, true);
2568 put_css_set_locked(from_cset);
2569 }
2570 }
2571 spin_unlock_bh(&css_set_lock);
2572
2573 /*
2574 * Migration is committed, all target tasks are now on dst_csets.
2575 * Nothing is sensitive to fork() after this point. Notify
2576 * controllers that migration is complete.
2577 */
2578 tset->csets = &tset->dst_csets;
2579
2580 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2581 if (ss->attach) {
2582 tset->ssid = ssid;
2583 ss->attach(tset);
2584 }
2585 } while_each_subsys_mask();
2586
2587 ret = 0;
2588 goto out_release_tset;
2589
2590 out_cancel_attach:
2591 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2592 if (ssid == failed_ssid)
2593 break;
2594 if (ss->cancel_attach) {
2595 tset->ssid = ssid;
2596 ss->cancel_attach(tset);
2597 }
2598 } while_each_subsys_mask();
2599 out_release_tset:
2600 spin_lock_bh(&css_set_lock);
2601 list_splice_init(&tset->dst_csets, &tset->src_csets);
2602 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2603 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2604 list_del_init(&cset->mg_node);
2605 }
2606 spin_unlock_bh(&css_set_lock);
2607 return ret;
2608 }
2609
2610 /**
2611 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2612 * @dst_cgrp: destination cgroup to test
2613 *
2614 * On the default hierarchy, except for the root, subtree_control must be
2615 * zero for migration destination cgroups with tasks so that child cgroups
2616 * don't compete against tasks.
2617 */
2618 static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2619 {
2620 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2621 !dst_cgrp->subtree_control;
2622 }
2623
2624 /**
2625 * cgroup_migrate_finish - cleanup after attach
2626 * @preloaded_csets: list of preloaded css_sets
2627 *
2628 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2629 * those functions for details.
2630 */
2631 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2632 {
2633 struct css_set *cset, *tmp_cset;
2634
2635 lockdep_assert_held(&cgroup_mutex);
2636
2637 spin_lock_bh(&css_set_lock);
2638 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2639 cset->mg_src_cgrp = NULL;
2640 cset->mg_dst_cgrp = NULL;
2641 cset->mg_dst_cset = NULL;
2642 list_del_init(&cset->mg_preload_node);
2643 put_css_set_locked(cset);
2644 }
2645 spin_unlock_bh(&css_set_lock);
2646 }
2647
2648 /**
2649 * cgroup_migrate_add_src - add a migration source css_set
2650 * @src_cset: the source css_set to add
2651 * @dst_cgrp: the destination cgroup
2652 * @preloaded_csets: list of preloaded css_sets
2653 *
2654 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2655 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2656 * up by cgroup_migrate_finish().
2657 *
2658 * This function may be called without holding cgroup_threadgroup_rwsem
2659 * even if the target is a process. Threads may be created and destroyed
2660 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2661 * into play and the preloaded css_sets are guaranteed to cover all
2662 * migrations.
2663 */
2664 static void cgroup_migrate_add_src(struct css_set *src_cset,
2665 struct cgroup *dst_cgrp,
2666 struct list_head *preloaded_csets)
2667 {
2668 struct cgroup *src_cgrp;
2669
2670 lockdep_assert_held(&cgroup_mutex);
2671 lockdep_assert_held(&css_set_lock);
2672
2673 /*
2674 * If ->dead, @src_set is associated with one or more dead cgroups
2675 * and doesn't contain any migratable tasks. Ignore it early so
2676 * that the rest of migration path doesn't get confused by it.
2677 */
2678 if (src_cset->dead)
2679 return;
2680
2681 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2682
2683 if (!list_empty(&src_cset->mg_preload_node))
2684 return;
2685
2686 WARN_ON(src_cset->mg_src_cgrp);
2687 WARN_ON(src_cset->mg_dst_cgrp);
2688 WARN_ON(!list_empty(&src_cset->mg_tasks));
2689 WARN_ON(!list_empty(&src_cset->mg_node));
2690
2691 src_cset->mg_src_cgrp = src_cgrp;
2692 src_cset->mg_dst_cgrp = dst_cgrp;
2693 get_css_set(src_cset);
2694 list_add(&src_cset->mg_preload_node, preloaded_csets);
2695 }
2696
2697 /**
2698 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2699 * @preloaded_csets: list of preloaded source css_sets
2700 *
2701 * Tasks are about to be moved and all the source css_sets have been
2702 * preloaded to @preloaded_csets. This function looks up and pins all
2703 * destination css_sets, links each to its source, and append them to
2704 * @preloaded_csets.
2705 *
2706 * This function must be called after cgroup_migrate_add_src() has been
2707 * called on each migration source css_set. After migration is performed
2708 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2709 * @preloaded_csets.
2710 */
2711 static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2712 {
2713 LIST_HEAD(csets);
2714 struct css_set *src_cset, *tmp_cset;
2715
2716 lockdep_assert_held(&cgroup_mutex);
2717
2718 /* look up the dst cset for each src cset and link it to src */
2719 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2720 struct css_set *dst_cset;
2721
2722 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2723 if (!dst_cset)
2724 goto err;
2725
2726 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2727
2728 /*
2729 * If src cset equals dst, it's noop. Drop the src.
2730 * cgroup_migrate() will skip the cset too. Note that we
2731 * can't handle src == dst as some nodes are used by both.
2732 */
2733 if (src_cset == dst_cset) {
2734 src_cset->mg_src_cgrp = NULL;
2735 src_cset->mg_dst_cgrp = NULL;
2736 list_del_init(&src_cset->mg_preload_node);
2737 put_css_set(src_cset);
2738 put_css_set(dst_cset);
2739 continue;
2740 }
2741
2742 src_cset->mg_dst_cset = dst_cset;
2743
2744 if (list_empty(&dst_cset->mg_preload_node))
2745 list_add(&dst_cset->mg_preload_node, &csets);
2746 else
2747 put_css_set(dst_cset);
2748 }
2749
2750 list_splice_tail(&csets, preloaded_csets);
2751 return 0;
2752 err:
2753 cgroup_migrate_finish(&csets);
2754 return -ENOMEM;
2755 }
2756
2757 /**
2758 * cgroup_migrate - migrate a process or task to a cgroup
2759 * @leader: the leader of the process or the task to migrate
2760 * @threadgroup: whether @leader points to the whole process or a single task
2761 * @root: cgroup root migration is taking place on
2762 *
2763 * Migrate a process or task denoted by @leader. If migrating a process,
2764 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2765 * responsible for invoking cgroup_migrate_add_src() and
2766 * cgroup_migrate_prepare_dst() on the targets before invoking this
2767 * function and following up with cgroup_migrate_finish().
2768 *
2769 * As long as a controller's ->can_attach() doesn't fail, this function is
2770 * guaranteed to succeed. This means that, excluding ->can_attach()
2771 * failure, when migrating multiple targets, the success or failure can be
2772 * decided for all targets by invoking group_migrate_prepare_dst() before
2773 * actually starting migrating.
2774 */
2775 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2776 struct cgroup_root *root)
2777 {
2778 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2779 struct task_struct *task;
2780
2781 /*
2782 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2783 * already PF_EXITING could be freed from underneath us unless we
2784 * take an rcu_read_lock.
2785 */
2786 spin_lock_bh(&css_set_lock);
2787 rcu_read_lock();
2788 task = leader;
2789 do {
2790 cgroup_taskset_add(task, &tset);
2791 if (!threadgroup)
2792 break;
2793 } while_each_thread(leader, task);
2794 rcu_read_unlock();
2795 spin_unlock_bh(&css_set_lock);
2796
2797 return cgroup_taskset_migrate(&tset, root);
2798 }
2799
2800 /**
2801 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2802 * @dst_cgrp: the cgroup to attach to
2803 * @leader: the task or the leader of the threadgroup to be attached
2804 * @threadgroup: attach the whole threadgroup?
2805 *
2806 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2807 */
2808 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2809 struct task_struct *leader, bool threadgroup)
2810 {
2811 LIST_HEAD(preloaded_csets);
2812 struct task_struct *task;
2813 int ret;
2814
2815 if (!cgroup_may_migrate_to(dst_cgrp))
2816 return -EBUSY;
2817
2818 /* look up all src csets */
2819 spin_lock_bh(&css_set_lock);
2820 rcu_read_lock();
2821 task = leader;
2822 do {
2823 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2824 &preloaded_csets);
2825 if (!threadgroup)
2826 break;
2827 } while_each_thread(leader, task);
2828 rcu_read_unlock();
2829 spin_unlock_bh(&css_set_lock);
2830
2831 /* prepare dst csets and commit */
2832 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2833 if (!ret)
2834 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2835
2836 cgroup_migrate_finish(&preloaded_csets);
2837 return ret;
2838 }
2839
2840 static int cgroup_procs_write_permission(struct task_struct *task,
2841 struct cgroup *dst_cgrp,
2842 struct kernfs_open_file *of)
2843 {
2844 const struct cred *cred = current_cred();
2845 const struct cred *tcred = get_task_cred(task);
2846 int ret = 0;
2847
2848 /*
2849 * even if we're attaching all tasks in the thread group, we only
2850 * need to check permissions on one of them.
2851 */
2852 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2853 !uid_eq(cred->euid, tcred->uid) &&
2854 !uid_eq(cred->euid, tcred->suid))
2855 ret = -EACCES;
2856
2857 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2858 struct super_block *sb = of->file->f_path.dentry->d_sb;
2859 struct cgroup *cgrp;
2860 struct inode *inode;
2861
2862 spin_lock_bh(&css_set_lock);
2863 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2864 spin_unlock_bh(&css_set_lock);
2865
2866 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2867 cgrp = cgroup_parent(cgrp);
2868
2869 ret = -ENOMEM;
2870 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2871 if (inode) {
2872 ret = inode_permission(inode, MAY_WRITE);
2873 iput(inode);
2874 }
2875 }
2876
2877 put_cred(tcred);
2878 return ret;
2879 }
2880
2881 /*
2882 * Find the task_struct of the task to attach by vpid and pass it along to the
2883 * function to attach either it or all tasks in its threadgroup. Will lock
2884 * cgroup_mutex and threadgroup.
2885 */
2886 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2887 size_t nbytes, loff_t off, bool threadgroup)
2888 {
2889 struct task_struct *tsk;
2890 struct cgroup_subsys *ss;
2891 struct cgroup *cgrp;
2892 pid_t pid;
2893 int ssid, ret;
2894
2895 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2896 return -EINVAL;
2897
2898 cgrp = cgroup_kn_lock_live(of->kn, false);
2899 if (!cgrp)
2900 return -ENODEV;
2901
2902 percpu_down_write(&cgroup_threadgroup_rwsem);
2903 rcu_read_lock();
2904 if (pid) {
2905 tsk = find_task_by_vpid(pid);
2906 if (!tsk) {
2907 ret = -ESRCH;
2908 goto out_unlock_rcu;
2909 }
2910 } else {
2911 tsk = current;
2912 }
2913
2914 if (threadgroup)
2915 tsk = tsk->group_leader;
2916
2917 /*
2918 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2919 * trapped in a cpuset, or RT worker may be born in a cgroup
2920 * with no rt_runtime allocated. Just say no.
2921 */
2922 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2923 ret = -EINVAL;
2924 goto out_unlock_rcu;
2925 }
2926
2927 get_task_struct(tsk);
2928 rcu_read_unlock();
2929
2930 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2931 if (!ret)
2932 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2933
2934 put_task_struct(tsk);
2935 goto out_unlock_threadgroup;
2936
2937 out_unlock_rcu:
2938 rcu_read_unlock();
2939 out_unlock_threadgroup:
2940 percpu_up_write(&cgroup_threadgroup_rwsem);
2941 for_each_subsys(ss, ssid)
2942 if (ss->post_attach)
2943 ss->post_attach();
2944 cgroup_kn_unlock(of->kn);
2945 return ret ?: nbytes;
2946 }
2947
2948 /**
2949 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2950 * @from: attach to all cgroups of a given task
2951 * @tsk: the task to be attached
2952 */
2953 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2954 {
2955 struct cgroup_root *root;
2956 int retval = 0;
2957
2958 mutex_lock(&cgroup_mutex);
2959 for_each_root(root) {
2960 struct cgroup *from_cgrp;
2961
2962 if (root == &cgrp_dfl_root)
2963 continue;
2964
2965 spin_lock_bh(&css_set_lock);
2966 from_cgrp = task_cgroup_from_root(from, root);
2967 spin_unlock_bh(&css_set_lock);
2968
2969 retval = cgroup_attach_task(from_cgrp, tsk, false);
2970 if (retval)
2971 break;
2972 }
2973 mutex_unlock(&cgroup_mutex);
2974
2975 return retval;
2976 }
2977 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2978
2979 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2980 char *buf, size_t nbytes, loff_t off)
2981 {
2982 return __cgroup_procs_write(of, buf, nbytes, off, false);
2983 }
2984
2985 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2986 char *buf, size_t nbytes, loff_t off)
2987 {
2988 return __cgroup_procs_write(of, buf, nbytes, off, true);
2989 }
2990
2991 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2992 char *buf, size_t nbytes, loff_t off)
2993 {
2994 struct cgroup *cgrp;
2995
2996 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2997
2998 cgrp = cgroup_kn_lock_live(of->kn, false);
2999 if (!cgrp)
3000 return -ENODEV;
3001 spin_lock(&release_agent_path_lock);
3002 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
3003 sizeof(cgrp->root->release_agent_path));
3004 spin_unlock(&release_agent_path_lock);
3005 cgroup_kn_unlock(of->kn);
3006 return nbytes;
3007 }
3008
3009 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
3010 {
3011 struct cgroup *cgrp = seq_css(seq)->cgroup;
3012
3013 spin_lock(&release_agent_path_lock);
3014 seq_puts(seq, cgrp->root->release_agent_path);
3015 spin_unlock(&release_agent_path_lock);
3016 seq_putc(seq, '\n');
3017 return 0;
3018 }
3019
3020 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
3021 {
3022 seq_puts(seq, "0\n");
3023 return 0;
3024 }
3025
3026 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
3027 {
3028 struct cgroup_subsys *ss;
3029 bool printed = false;
3030 int ssid;
3031
3032 do_each_subsys_mask(ss, ssid, ss_mask) {
3033 if (printed)
3034 seq_putc(seq, ' ');
3035 seq_printf(seq, "%s", ss->name);
3036 printed = true;
3037 } while_each_subsys_mask();
3038 if (printed)
3039 seq_putc(seq, '\n');
3040 }
3041
3042 /* show controllers which are enabled from the parent */
3043 static int cgroup_controllers_show(struct seq_file *seq, void *v)
3044 {
3045 struct cgroup *cgrp = seq_css(seq)->cgroup;
3046
3047 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3048 return 0;
3049 }
3050
3051 /* show controllers which are enabled for a given cgroup's children */
3052 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3053 {
3054 struct cgroup *cgrp = seq_css(seq)->cgroup;
3055
3056 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3057 return 0;
3058 }
3059
3060 /**
3061 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3062 * @cgrp: root of the subtree to update csses for
3063 *
3064 * @cgrp's control masks have changed and its subtree's css associations
3065 * need to be updated accordingly. This function looks up all css_sets
3066 * which are attached to the subtree, creates the matching updated css_sets
3067 * and migrates the tasks to the new ones.
3068 */
3069 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3070 {
3071 LIST_HEAD(preloaded_csets);
3072 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
3073 struct cgroup_subsys_state *d_css;
3074 struct cgroup *dsct;
3075 struct css_set *src_cset;
3076 int ret;
3077
3078 lockdep_assert_held(&cgroup_mutex);
3079
3080 percpu_down_write(&cgroup_threadgroup_rwsem);
3081
3082 /* look up all csses currently attached to @cgrp's subtree */
3083 spin_lock_bh(&css_set_lock);
3084 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3085 struct cgrp_cset_link *link;
3086
3087 list_for_each_entry(link, &dsct->cset_links, cset_link)
3088 cgroup_migrate_add_src(link->cset, dsct,
3089 &preloaded_csets);
3090 }
3091 spin_unlock_bh(&css_set_lock);
3092
3093 /* NULL dst indicates self on default hierarchy */
3094 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
3095 if (ret)
3096 goto out_finish;
3097
3098 spin_lock_bh(&css_set_lock);
3099 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3100 struct task_struct *task, *ntask;
3101
3102 /* src_csets precede dst_csets, break on the first dst_cset */
3103 if (!src_cset->mg_src_cgrp)
3104 break;
3105
3106 /* all tasks in src_csets need to be migrated */
3107 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3108 cgroup_taskset_add(task, &tset);
3109 }
3110 spin_unlock_bh(&css_set_lock);
3111
3112 ret = cgroup_taskset_migrate(&tset, cgrp->root);
3113 out_finish:
3114 cgroup_migrate_finish(&preloaded_csets);
3115 percpu_up_write(&cgroup_threadgroup_rwsem);
3116 return ret;
3117 }
3118
3119 /**
3120 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3121 * @cgrp: root of the target subtree
3122 *
3123 * Because css offlining is asynchronous, userland may try to re-enable a
3124 * controller while the previous css is still around. This function grabs
3125 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3126 */
3127 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3128 __acquires(&cgroup_mutex)
3129 {
3130 struct cgroup *dsct;
3131 struct cgroup_subsys_state *d_css;
3132 struct cgroup_subsys *ss;
3133 int ssid;
3134
3135 restart:
3136 mutex_lock(&cgroup_mutex);
3137
3138 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3139 for_each_subsys(ss, ssid) {
3140 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3141 DEFINE_WAIT(wait);
3142
3143 if (!css || !percpu_ref_is_dying(&css->refcnt))
3144 continue;
3145
3146 cgroup_get(dsct);
3147 prepare_to_wait(&dsct->offline_waitq, &wait,
3148 TASK_UNINTERRUPTIBLE);
3149
3150 mutex_unlock(&cgroup_mutex);
3151 schedule();
3152 finish_wait(&dsct->offline_waitq, &wait);
3153
3154 cgroup_put(dsct);
3155 goto restart;
3156 }
3157 }
3158 }
3159
3160 /**
3161 * cgroup_save_control - save control masks of a subtree
3162 * @cgrp: root of the target subtree
3163 *
3164 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3165 * prefixed fields for @cgrp's subtree including @cgrp itself.
3166 */
3167 static void cgroup_save_control(struct cgroup *cgrp)
3168 {
3169 struct cgroup *dsct;
3170 struct cgroup_subsys_state *d_css;
3171
3172 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3173 dsct->old_subtree_control = dsct->subtree_control;
3174 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3175 }
3176 }
3177
3178 /**
3179 * cgroup_propagate_control - refresh control masks of a subtree
3180 * @cgrp: root of the target subtree
3181 *
3182 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3183 * ->subtree_control and propagate controller availability through the
3184 * subtree so that descendants don't have unavailable controllers enabled.
3185 */
3186 static void cgroup_propagate_control(struct cgroup *cgrp)
3187 {
3188 struct cgroup *dsct;
3189 struct cgroup_subsys_state *d_css;
3190
3191 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3192 dsct->subtree_control &= cgroup_control(dsct);
3193 dsct->subtree_ss_mask =
3194 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3195 cgroup_ss_mask(dsct));
3196 }
3197 }
3198
3199 /**
3200 * cgroup_restore_control - restore control masks of a subtree
3201 * @cgrp: root of the target subtree
3202 *
3203 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3204 * prefixed fields for @cgrp's subtree including @cgrp itself.
3205 */
3206 static void cgroup_restore_control(struct cgroup *cgrp)
3207 {
3208 struct cgroup *dsct;
3209 struct cgroup_subsys_state *d_css;
3210
3211 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3212 dsct->subtree_control = dsct->old_subtree_control;
3213 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3214 }
3215 }
3216
3217 static bool css_visible(struct cgroup_subsys_state *css)
3218 {
3219 struct cgroup_subsys *ss = css->ss;
3220 struct cgroup *cgrp = css->cgroup;
3221
3222 if (cgroup_control(cgrp) & (1 << ss->id))
3223 return true;
3224 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3225 return false;
3226 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3227 }
3228
3229 /**
3230 * cgroup_apply_control_enable - enable or show csses according to control
3231 * @cgrp: root of the target subtree
3232 *
3233 * Walk @cgrp's subtree and create new csses or make the existing ones
3234 * visible. A css is created invisible if it's being implicitly enabled
3235 * through dependency. An invisible css is made visible when the userland
3236 * explicitly enables it.
3237 *
3238 * Returns 0 on success, -errno on failure. On failure, csses which have
3239 * been processed already aren't cleaned up. The caller is responsible for
3240 * cleaning up with cgroup_apply_control_disble().
3241 */
3242 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3243 {
3244 struct cgroup *dsct;
3245 struct cgroup_subsys_state *d_css;
3246 struct cgroup_subsys *ss;
3247 int ssid, ret;
3248
3249 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3250 for_each_subsys(ss, ssid) {
3251 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3252
3253 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3254
3255 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3256 continue;
3257
3258 if (!css) {
3259 css = css_create(dsct, ss);
3260 if (IS_ERR(css))
3261 return PTR_ERR(css);
3262 }
3263
3264 if (css_visible(css)) {
3265 ret = css_populate_dir(css);
3266 if (ret)
3267 return ret;
3268 }
3269 }
3270 }
3271
3272 return 0;
3273 }
3274
3275 /**
3276 * cgroup_apply_control_disable - kill or hide csses according to control
3277 * @cgrp: root of the target subtree
3278 *
3279 * Walk @cgrp's subtree and kill and hide csses so that they match
3280 * cgroup_ss_mask() and cgroup_visible_mask().
3281 *
3282 * A css is hidden when the userland requests it to be disabled while other
3283 * subsystems are still depending on it. The css must not actively control
3284 * resources and be in the vanilla state if it's made visible again later.
3285 * Controllers which may be depended upon should provide ->css_reset() for
3286 * this purpose.
3287 */
3288 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3289 {
3290 struct cgroup *dsct;
3291 struct cgroup_subsys_state *d_css;
3292 struct cgroup_subsys *ss;
3293 int ssid;
3294
3295 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3296 for_each_subsys(ss, ssid) {
3297 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3298
3299 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3300
3301 if (!css)
3302 continue;
3303
3304 if (css->parent &&
3305 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3306 kill_css(css);
3307 } else if (!css_visible(css)) {
3308 css_clear_dir(css);
3309 if (ss->css_reset)
3310 ss->css_reset(css);
3311 }
3312 }
3313 }
3314 }
3315
3316 /**
3317 * cgroup_apply_control - apply control mask updates to the subtree
3318 * @cgrp: root of the target subtree
3319 *
3320 * subsystems can be enabled and disabled in a subtree using the following
3321 * steps.
3322 *
3323 * 1. Call cgroup_save_control() to stash the current state.
3324 * 2. Update ->subtree_control masks in the subtree as desired.
3325 * 3. Call cgroup_apply_control() to apply the changes.
3326 * 4. Optionally perform other related operations.
3327 * 5. Call cgroup_finalize_control() to finish up.
3328 *
3329 * This function implements step 3 and propagates the mask changes
3330 * throughout @cgrp's subtree, updates csses accordingly and perform
3331 * process migrations.
3332 */
3333 static int cgroup_apply_control(struct cgroup *cgrp)
3334 {
3335 int ret;
3336
3337 cgroup_propagate_control(cgrp);
3338
3339 ret = cgroup_apply_control_enable(cgrp);
3340 if (ret)
3341 return ret;
3342
3343 /*
3344 * At this point, cgroup_e_css() results reflect the new csses
3345 * making the following cgroup_update_dfl_csses() properly update
3346 * css associations of all tasks in the subtree.
3347 */
3348 ret = cgroup_update_dfl_csses(cgrp);
3349 if (ret)
3350 return ret;
3351
3352 return 0;
3353 }
3354
3355 /**
3356 * cgroup_finalize_control - finalize control mask update
3357 * @cgrp: root of the target subtree
3358 * @ret: the result of the update
3359 *
3360 * Finalize control mask update. See cgroup_apply_control() for more info.
3361 */
3362 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3363 {
3364 if (ret) {
3365 cgroup_restore_control(cgrp);
3366 cgroup_propagate_control(cgrp);
3367 }
3368
3369 cgroup_apply_control_disable(cgrp);
3370 }
3371
3372 /* change the enabled child controllers for a cgroup in the default hierarchy */
3373 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3374 char *buf, size_t nbytes,
3375 loff_t off)
3376 {
3377 u16 enable = 0, disable = 0;
3378 struct cgroup *cgrp, *child;
3379 struct cgroup_subsys *ss;
3380 char *tok;
3381 int ssid, ret;
3382
3383 /*
3384 * Parse input - space separated list of subsystem names prefixed
3385 * with either + or -.
3386 */
3387 buf = strstrip(buf);
3388 while ((tok = strsep(&buf, " "))) {
3389 if (tok[0] == '\0')
3390 continue;
3391 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3392 if (!cgroup_ssid_enabled(ssid) ||
3393 strcmp(tok + 1, ss->name))
3394 continue;
3395
3396 if (*tok == '+') {
3397 enable |= 1 << ssid;
3398 disable &= ~(1 << ssid);
3399 } else if (*tok == '-') {
3400 disable |= 1 << ssid;
3401 enable &= ~(1 << ssid);
3402 } else {
3403 return -EINVAL;
3404 }
3405 break;
3406 } while_each_subsys_mask();
3407 if (ssid == CGROUP_SUBSYS_COUNT)
3408 return -EINVAL;
3409 }
3410
3411 cgrp = cgroup_kn_lock_live(of->kn, true);
3412 if (!cgrp)
3413 return -ENODEV;
3414
3415 for_each_subsys(ss, ssid) {
3416 if (enable & (1 << ssid)) {
3417 if (cgrp->subtree_control & (1 << ssid)) {
3418 enable &= ~(1 << ssid);
3419 continue;
3420 }
3421
3422 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3423 ret = -ENOENT;
3424 goto out_unlock;
3425 }
3426 } else if (disable & (1 << ssid)) {
3427 if (!(cgrp->subtree_control & (1 << ssid))) {
3428 disable &= ~(1 << ssid);
3429 continue;
3430 }
3431
3432 /* a child has it enabled? */
3433 cgroup_for_each_live_child(child, cgrp) {
3434 if (child->subtree_control & (1 << ssid)) {
3435 ret = -EBUSY;
3436 goto out_unlock;
3437 }
3438 }
3439 }
3440 }
3441
3442 if (!enable && !disable) {
3443 ret = 0;
3444 goto out_unlock;
3445 }
3446
3447 /*
3448 * Except for the root, subtree_control must be zero for a cgroup
3449 * with tasks so that child cgroups don't compete against tasks.
3450 */
3451 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3452 ret = -EBUSY;
3453 goto out_unlock;
3454 }
3455
3456 /* save and update control masks and prepare csses */
3457 cgroup_save_control(cgrp);
3458
3459 cgrp->subtree_control |= enable;
3460 cgrp->subtree_control &= ~disable;
3461
3462 ret = cgroup_apply_control(cgrp);
3463
3464 cgroup_finalize_control(cgrp, ret);
3465
3466 kernfs_activate(cgrp->kn);
3467 ret = 0;
3468 out_unlock:
3469 cgroup_kn_unlock(of->kn);
3470 return ret ?: nbytes;
3471 }
3472
3473 static int cgroup_events_show(struct seq_file *seq, void *v)
3474 {
3475 seq_printf(seq, "populated %d\n",
3476 cgroup_is_populated(seq_css(seq)->cgroup));
3477 return 0;
3478 }
3479
3480 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3481 size_t nbytes, loff_t off)
3482 {
3483 struct cgroup *cgrp = of->kn->parent->priv;
3484 struct cftype *cft = of->kn->priv;
3485 struct cgroup_subsys_state *css;
3486 int ret;
3487
3488 if (cft->write)
3489 return cft->write(of, buf, nbytes, off);
3490
3491 /*
3492 * kernfs guarantees that a file isn't deleted with operations in
3493 * flight, which means that the matching css is and stays alive and
3494 * doesn't need to be pinned. The RCU locking is not necessary
3495 * either. It's just for the convenience of using cgroup_css().
3496 */
3497 rcu_read_lock();
3498 css = cgroup_css(cgrp, cft->ss);
3499 rcu_read_unlock();
3500
3501 if (cft->write_u64) {
3502 unsigned long long v;
3503 ret = kstrtoull(buf, 0, &v);
3504 if (!ret)
3505 ret = cft->write_u64(css, cft, v);
3506 } else if (cft->write_s64) {
3507 long long v;
3508 ret = kstrtoll(buf, 0, &v);
3509 if (!ret)
3510 ret = cft->write_s64(css, cft, v);
3511 } else {
3512 ret = -EINVAL;
3513 }
3514
3515 return ret ?: nbytes;
3516 }
3517
3518 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3519 {
3520 return seq_cft(seq)->seq_start(seq, ppos);
3521 }
3522
3523 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3524 {
3525 return seq_cft(seq)->seq_next(seq, v, ppos);
3526 }
3527
3528 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3529 {
3530 seq_cft(seq)->seq_stop(seq, v);
3531 }
3532
3533 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3534 {
3535 struct cftype *cft = seq_cft(m);
3536 struct cgroup_subsys_state *css = seq_css(m);
3537
3538 if (cft->seq_show)
3539 return cft->seq_show(m, arg);
3540
3541 if (cft->read_u64)
3542 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3543 else if (cft->read_s64)
3544 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3545 else
3546 return -EINVAL;
3547 return 0;
3548 }
3549
3550 static struct kernfs_ops cgroup_kf_single_ops = {
3551 .atomic_write_len = PAGE_SIZE,
3552 .write = cgroup_file_write,
3553 .seq_show = cgroup_seqfile_show,
3554 };
3555
3556 static struct kernfs_ops cgroup_kf_ops = {
3557 .atomic_write_len = PAGE_SIZE,
3558 .write = cgroup_file_write,
3559 .seq_start = cgroup_seqfile_start,
3560 .seq_next = cgroup_seqfile_next,
3561 .seq_stop = cgroup_seqfile_stop,
3562 .seq_show = cgroup_seqfile_show,
3563 };
3564
3565 /*
3566 * cgroup_rename - Only allow simple rename of directories in place.
3567 */
3568 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3569 const char *new_name_str)
3570 {
3571 struct cgroup *cgrp = kn->priv;
3572 int ret;
3573
3574 if (kernfs_type(kn) != KERNFS_DIR)
3575 return -ENOTDIR;
3576 if (kn->parent != new_parent)
3577 return -EIO;
3578
3579 /*
3580 * This isn't a proper migration and its usefulness is very
3581 * limited. Disallow on the default hierarchy.
3582 */
3583 if (cgroup_on_dfl(cgrp))
3584 return -EPERM;
3585
3586 /*
3587 * We're gonna grab cgroup_mutex which nests outside kernfs
3588 * active_ref. kernfs_rename() doesn't require active_ref
3589 * protection. Break them before grabbing cgroup_mutex.
3590 */
3591 kernfs_break_active_protection(new_parent);
3592 kernfs_break_active_protection(kn);
3593
3594 mutex_lock(&cgroup_mutex);
3595
3596 ret = kernfs_rename(kn, new_parent, new_name_str);
3597
3598 mutex_unlock(&cgroup_mutex);
3599
3600 kernfs_unbreak_active_protection(kn);
3601 kernfs_unbreak_active_protection(new_parent);
3602 return ret;
3603 }
3604
3605 /* set uid and gid of cgroup dirs and files to that of the creator */
3606 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3607 {
3608 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3609 .ia_uid = current_fsuid(),
3610 .ia_gid = current_fsgid(), };
3611
3612 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3613 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3614 return 0;
3615
3616 return kernfs_setattr(kn, &iattr);
3617 }
3618
3619 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3620 struct cftype *cft)
3621 {
3622 char name[CGROUP_FILE_NAME_MAX];
3623 struct kernfs_node *kn;
3624 struct lock_class_key *key = NULL;
3625 int ret;
3626
3627 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3628 key = &cft->lockdep_key;
3629 #endif
3630 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3631 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3632 NULL, key);
3633 if (IS_ERR(kn))
3634 return PTR_ERR(kn);
3635
3636 ret = cgroup_kn_set_ugid(kn);
3637 if (ret) {
3638 kernfs_remove(kn);
3639 return ret;
3640 }
3641
3642 if (cft->file_offset) {
3643 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3644
3645 spin_lock_irq(&cgroup_file_kn_lock);
3646 cfile->kn = kn;
3647 spin_unlock_irq(&cgroup_file_kn_lock);
3648 }
3649
3650 return 0;
3651 }
3652
3653 /**
3654 * cgroup_addrm_files - add or remove files to a cgroup directory
3655 * @css: the target css
3656 * @cgrp: the target cgroup (usually css->cgroup)
3657 * @cfts: array of cftypes to be added
3658 * @is_add: whether to add or remove
3659 *
3660 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3661 * For removals, this function never fails.
3662 */
3663 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3664 struct cgroup *cgrp, struct cftype cfts[],
3665 bool is_add)
3666 {
3667 struct cftype *cft, *cft_end = NULL;
3668 int ret = 0;
3669
3670 lockdep_assert_held(&cgroup_mutex);
3671
3672 restart:
3673 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3674 /* does cft->flags tell us to skip this file on @cgrp? */
3675 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3676 continue;
3677 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3678 continue;
3679 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3680 continue;
3681 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3682 continue;
3683
3684 if (is_add) {
3685 ret = cgroup_add_file(css, cgrp, cft);
3686 if (ret) {
3687 pr_warn("%s: failed to add %s, err=%d\n",
3688 __func__, cft->name, ret);
3689 cft_end = cft;
3690 is_add = false;
3691 goto restart;
3692 }
3693 } else {
3694 cgroup_rm_file(cgrp, cft);
3695 }
3696 }
3697 return ret;
3698 }
3699
3700 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3701 {
3702 LIST_HEAD(pending);
3703 struct cgroup_subsys *ss = cfts[0].ss;
3704 struct cgroup *root = &ss->root->cgrp;
3705 struct cgroup_subsys_state *css;
3706 int ret = 0;
3707
3708 lockdep_assert_held(&cgroup_mutex);
3709
3710 /* add/rm files for all cgroups created before */
3711 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3712 struct cgroup *cgrp = css->cgroup;
3713
3714 if (!(css->flags & CSS_VISIBLE))
3715 continue;
3716
3717 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3718 if (ret)
3719 break;
3720 }
3721
3722 if (is_add && !ret)
3723 kernfs_activate(root->kn);
3724 return ret;
3725 }
3726
3727 static void cgroup_exit_cftypes(struct cftype *cfts)
3728 {
3729 struct cftype *cft;
3730
3731 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3732 /* free copy for custom atomic_write_len, see init_cftypes() */
3733 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3734 kfree(cft->kf_ops);
3735 cft->kf_ops = NULL;
3736 cft->ss = NULL;
3737
3738 /* revert flags set by cgroup core while adding @cfts */
3739 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3740 }
3741 }
3742
3743 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3744 {
3745 struct cftype *cft;
3746
3747 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3748 struct kernfs_ops *kf_ops;
3749
3750 WARN_ON(cft->ss || cft->kf_ops);
3751
3752 if (cft->seq_start)
3753 kf_ops = &cgroup_kf_ops;
3754 else
3755 kf_ops = &cgroup_kf_single_ops;
3756
3757 /*
3758 * Ugh... if @cft wants a custom max_write_len, we need to
3759 * make a copy of kf_ops to set its atomic_write_len.
3760 */
3761 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3762 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3763 if (!kf_ops) {
3764 cgroup_exit_cftypes(cfts);
3765 return -ENOMEM;
3766 }
3767 kf_ops->atomic_write_len = cft->max_write_len;
3768 }
3769
3770 cft->kf_ops = kf_ops;
3771 cft->ss = ss;
3772 }
3773
3774 return 0;
3775 }
3776
3777 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3778 {
3779 lockdep_assert_held(&cgroup_mutex);
3780
3781 if (!cfts || !cfts[0].ss)
3782 return -ENOENT;
3783
3784 list_del(&cfts->node);
3785 cgroup_apply_cftypes(cfts, false);
3786 cgroup_exit_cftypes(cfts);
3787 return 0;
3788 }
3789
3790 /**
3791 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3792 * @cfts: zero-length name terminated array of cftypes
3793 *
3794 * Unregister @cfts. Files described by @cfts are removed from all
3795 * existing cgroups and all future cgroups won't have them either. This
3796 * function can be called anytime whether @cfts' subsys is attached or not.
3797 *
3798 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3799 * registered.
3800 */
3801 int cgroup_rm_cftypes(struct cftype *cfts)
3802 {
3803 int ret;
3804
3805 mutex_lock(&cgroup_mutex);
3806 ret = cgroup_rm_cftypes_locked(cfts);
3807 mutex_unlock(&cgroup_mutex);
3808 return ret;
3809 }
3810
3811 /**
3812 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3813 * @ss: target cgroup subsystem
3814 * @cfts: zero-length name terminated array of cftypes
3815 *
3816 * Register @cfts to @ss. Files described by @cfts are created for all
3817 * existing cgroups to which @ss is attached and all future cgroups will
3818 * have them too. This function can be called anytime whether @ss is
3819 * attached or not.
3820 *
3821 * Returns 0 on successful registration, -errno on failure. Note that this
3822 * function currently returns 0 as long as @cfts registration is successful
3823 * even if some file creation attempts on existing cgroups fail.
3824 */
3825 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3826 {
3827 int ret;
3828
3829 if (!cgroup_ssid_enabled(ss->id))
3830 return 0;
3831
3832 if (!cfts || cfts[0].name[0] == '\0')
3833 return 0;
3834
3835 ret = cgroup_init_cftypes(ss, cfts);
3836 if (ret)
3837 return ret;
3838
3839 mutex_lock(&cgroup_mutex);
3840
3841 list_add_tail(&cfts->node, &ss->cfts);
3842 ret = cgroup_apply_cftypes(cfts, true);
3843 if (ret)
3844 cgroup_rm_cftypes_locked(cfts);
3845
3846 mutex_unlock(&cgroup_mutex);
3847 return ret;
3848 }
3849
3850 /**
3851 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3852 * @ss: target cgroup subsystem
3853 * @cfts: zero-length name terminated array of cftypes
3854 *
3855 * Similar to cgroup_add_cftypes() but the added files are only used for
3856 * the default hierarchy.
3857 */
3858 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3859 {
3860 struct cftype *cft;
3861
3862 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3863 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3864 return cgroup_add_cftypes(ss, cfts);
3865 }
3866
3867 /**
3868 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3869 * @ss: target cgroup subsystem
3870 * @cfts: zero-length name terminated array of cftypes
3871 *
3872 * Similar to cgroup_add_cftypes() but the added files are only used for
3873 * the legacy hierarchies.
3874 */
3875 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3876 {
3877 struct cftype *cft;
3878
3879 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3880 cft->flags |= __CFTYPE_NOT_ON_DFL;
3881 return cgroup_add_cftypes(ss, cfts);
3882 }
3883
3884 /**
3885 * cgroup_file_notify - generate a file modified event for a cgroup_file
3886 * @cfile: target cgroup_file
3887 *
3888 * @cfile must have been obtained by setting cftype->file_offset.
3889 */
3890 void cgroup_file_notify(struct cgroup_file *cfile)
3891 {
3892 unsigned long flags;
3893
3894 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3895 if (cfile->kn)
3896 kernfs_notify(cfile->kn);
3897 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3898 }
3899
3900 /**
3901 * cgroup_task_count - count the number of tasks in a cgroup.
3902 * @cgrp: the cgroup in question
3903 *
3904 * Return the number of tasks in the cgroup.
3905 */
3906 static int cgroup_task_count(const struct cgroup *cgrp)
3907 {
3908 int count = 0;
3909 struct cgrp_cset_link *link;
3910
3911 spin_lock_bh(&css_set_lock);
3912 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3913 count += atomic_read(&link->cset->refcount);
3914 spin_unlock_bh(&css_set_lock);
3915 return count;
3916 }
3917
3918 /**
3919 * css_next_child - find the next child of a given css
3920 * @pos: the current position (%NULL to initiate traversal)
3921 * @parent: css whose children to walk
3922 *
3923 * This function returns the next child of @parent and should be called
3924 * under either cgroup_mutex or RCU read lock. The only requirement is
3925 * that @parent and @pos are accessible. The next sibling is guaranteed to
3926 * be returned regardless of their states.
3927 *
3928 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3929 * css which finished ->css_online() is guaranteed to be visible in the
3930 * future iterations and will stay visible until the last reference is put.
3931 * A css which hasn't finished ->css_online() or already finished
3932 * ->css_offline() may show up during traversal. It's each subsystem's
3933 * responsibility to synchronize against on/offlining.
3934 */
3935 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3936 struct cgroup_subsys_state *parent)
3937 {
3938 struct cgroup_subsys_state *next;
3939
3940 cgroup_assert_mutex_or_rcu_locked();
3941
3942 /*
3943 * @pos could already have been unlinked from the sibling list.
3944 * Once a cgroup is removed, its ->sibling.next is no longer
3945 * updated when its next sibling changes. CSS_RELEASED is set when
3946 * @pos is taken off list, at which time its next pointer is valid,
3947 * and, as releases are serialized, the one pointed to by the next
3948 * pointer is guaranteed to not have started release yet. This
3949 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3950 * critical section, the one pointed to by its next pointer is
3951 * guaranteed to not have finished its RCU grace period even if we
3952 * have dropped rcu_read_lock() inbetween iterations.
3953 *
3954 * If @pos has CSS_RELEASED set, its next pointer can't be
3955 * dereferenced; however, as each css is given a monotonically
3956 * increasing unique serial number and always appended to the
3957 * sibling list, the next one can be found by walking the parent's
3958 * children until the first css with higher serial number than
3959 * @pos's. While this path can be slower, it happens iff iteration
3960 * races against release and the race window is very small.
3961 */
3962 if (!pos) {
3963 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3964 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3965 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3966 } else {
3967 list_for_each_entry_rcu(next, &parent->children, sibling)
3968 if (next->serial_nr > pos->serial_nr)
3969 break;
3970 }
3971
3972 /*
3973 * @next, if not pointing to the head, can be dereferenced and is
3974 * the next sibling.
3975 */
3976 if (&next->sibling != &parent->children)
3977 return next;
3978 return NULL;
3979 }
3980
3981 /**
3982 * css_next_descendant_pre - find the next descendant for pre-order walk
3983 * @pos: the current position (%NULL to initiate traversal)
3984 * @root: css whose descendants to walk
3985 *
3986 * To be used by css_for_each_descendant_pre(). Find the next descendant
3987 * to visit for pre-order traversal of @root's descendants. @root is
3988 * included in the iteration and the first node to be visited.
3989 *
3990 * While this function requires cgroup_mutex or RCU read locking, it
3991 * doesn't require the whole traversal to be contained in a single critical
3992 * section. This function will return the correct next descendant as long
3993 * as both @pos and @root are accessible and @pos is a descendant of @root.
3994 *
3995 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3996 * css which finished ->css_online() is guaranteed to be visible in the
3997 * future iterations and will stay visible until the last reference is put.
3998 * A css which hasn't finished ->css_online() or already finished
3999 * ->css_offline() may show up during traversal. It's each subsystem's
4000 * responsibility to synchronize against on/offlining.
4001 */
4002 struct cgroup_subsys_state *
4003 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4004 struct cgroup_subsys_state *root)
4005 {
4006 struct cgroup_subsys_state *next;
4007
4008 cgroup_assert_mutex_or_rcu_locked();
4009
4010 /* if first iteration, visit @root */
4011 if (!pos)
4012 return root;
4013
4014 /* visit the first child if exists */
4015 next = css_next_child(NULL, pos);
4016 if (next)
4017 return next;
4018
4019 /* no child, visit my or the closest ancestor's next sibling */
4020 while (pos != root) {
4021 next = css_next_child(pos, pos->parent);
4022 if (next)
4023 return next;
4024 pos = pos->parent;
4025 }
4026
4027 return NULL;
4028 }
4029
4030 /**
4031 * css_rightmost_descendant - return the rightmost descendant of a css
4032 * @pos: css of interest
4033 *
4034 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4035 * is returned. This can be used during pre-order traversal to skip
4036 * subtree of @pos.
4037 *
4038 * While this function requires cgroup_mutex or RCU read locking, it
4039 * doesn't require the whole traversal to be contained in a single critical
4040 * section. This function will return the correct rightmost descendant as
4041 * long as @pos is accessible.
4042 */
4043 struct cgroup_subsys_state *
4044 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4045 {
4046 struct cgroup_subsys_state *last, *tmp;
4047
4048 cgroup_assert_mutex_or_rcu_locked();
4049
4050 do {
4051 last = pos;
4052 /* ->prev isn't RCU safe, walk ->next till the end */
4053 pos = NULL;
4054 css_for_each_child(tmp, last)
4055 pos = tmp;
4056 } while (pos);
4057
4058 return last;
4059 }
4060
4061 static struct cgroup_subsys_state *
4062 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4063 {
4064 struct cgroup_subsys_state *last;
4065
4066 do {
4067 last = pos;
4068 pos = css_next_child(NULL, pos);
4069 } while (pos);
4070
4071 return last;
4072 }
4073
4074 /**
4075 * css_next_descendant_post - find the next descendant for post-order walk
4076 * @pos: the current position (%NULL to initiate traversal)
4077 * @root: css whose descendants to walk
4078 *
4079 * To be used by css_for_each_descendant_post(). Find the next descendant
4080 * to visit for post-order traversal of @root's descendants. @root is
4081 * included in the iteration and the last node to be visited.
4082 *
4083 * While this function requires cgroup_mutex or RCU read locking, it
4084 * doesn't require the whole traversal to be contained in a single critical
4085 * section. This function will return the correct next descendant as long
4086 * as both @pos and @cgroup are accessible and @pos is a descendant of
4087 * @cgroup.
4088 *
4089 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4090 * css which finished ->css_online() is guaranteed to be visible in the
4091 * future iterations and will stay visible until the last reference is put.
4092 * A css which hasn't finished ->css_online() or already finished
4093 * ->css_offline() may show up during traversal. It's each subsystem's
4094 * responsibility to synchronize against on/offlining.
4095 */
4096 struct cgroup_subsys_state *
4097 css_next_descendant_post(struct cgroup_subsys_state *pos,
4098 struct cgroup_subsys_state *root)
4099 {
4100 struct cgroup_subsys_state *next;
4101
4102 cgroup_assert_mutex_or_rcu_locked();
4103
4104 /* if first iteration, visit leftmost descendant which may be @root */
4105 if (!pos)
4106 return css_leftmost_descendant(root);
4107
4108 /* if we visited @root, we're done */
4109 if (pos == root)
4110 return NULL;
4111
4112 /* if there's an unvisited sibling, visit its leftmost descendant */
4113 next = css_next_child(pos, pos->parent);
4114 if (next)
4115 return css_leftmost_descendant(next);
4116
4117 /* no sibling left, visit parent */
4118 return pos->parent;
4119 }
4120
4121 /**
4122 * css_has_online_children - does a css have online children
4123 * @css: the target css
4124 *
4125 * Returns %true if @css has any online children; otherwise, %false. This
4126 * function can be called from any context but the caller is responsible
4127 * for synchronizing against on/offlining as necessary.
4128 */
4129 bool css_has_online_children(struct cgroup_subsys_state *css)
4130 {
4131 struct cgroup_subsys_state *child;
4132 bool ret = false;
4133
4134 rcu_read_lock();
4135 css_for_each_child(child, css) {
4136 if (child->flags & CSS_ONLINE) {
4137 ret = true;
4138 break;
4139 }
4140 }
4141 rcu_read_unlock();
4142 return ret;
4143 }
4144
4145 /**
4146 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4147 * @it: the iterator to advance
4148 *
4149 * Advance @it to the next css_set to walk.
4150 */
4151 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4152 {
4153 struct list_head *l = it->cset_pos;
4154 struct cgrp_cset_link *link;
4155 struct css_set *cset;
4156
4157 lockdep_assert_held(&css_set_lock);
4158
4159 /* Advance to the next non-empty css_set */
4160 do {
4161 l = l->next;
4162 if (l == it->cset_head) {
4163 it->cset_pos = NULL;
4164 it->task_pos = NULL;
4165 return;
4166 }
4167
4168 if (it->ss) {
4169 cset = container_of(l, struct css_set,
4170 e_cset_node[it->ss->id]);
4171 } else {
4172 link = list_entry(l, struct cgrp_cset_link, cset_link);
4173 cset = link->cset;
4174 }
4175 } while (!css_set_populated(cset));
4176
4177 it->cset_pos = l;
4178
4179 if (!list_empty(&cset->tasks))
4180 it->task_pos = cset->tasks.next;
4181 else
4182 it->task_pos = cset->mg_tasks.next;
4183
4184 it->tasks_head = &cset->tasks;
4185 it->mg_tasks_head = &cset->mg_tasks;
4186
4187 /*
4188 * We don't keep css_sets locked across iteration steps and thus
4189 * need to take steps to ensure that iteration can be resumed after
4190 * the lock is re-acquired. Iteration is performed at two levels -
4191 * css_sets and tasks in them.
4192 *
4193 * Once created, a css_set never leaves its cgroup lists, so a
4194 * pinned css_set is guaranteed to stay put and we can resume
4195 * iteration afterwards.
4196 *
4197 * Tasks may leave @cset across iteration steps. This is resolved
4198 * by registering each iterator with the css_set currently being
4199 * walked and making css_set_move_task() advance iterators whose
4200 * next task is leaving.
4201 */
4202 if (it->cur_cset) {
4203 list_del(&it->iters_node);
4204 put_css_set_locked(it->cur_cset);
4205 }
4206 get_css_set(cset);
4207 it->cur_cset = cset;
4208 list_add(&it->iters_node, &cset->task_iters);
4209 }
4210
4211 static void css_task_iter_advance(struct css_task_iter *it)
4212 {
4213 struct list_head *l = it->task_pos;
4214
4215 lockdep_assert_held(&css_set_lock);
4216 WARN_ON_ONCE(!l);
4217
4218 /*
4219 * Advance iterator to find next entry. cset->tasks is consumed
4220 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4221 * next cset.
4222 */
4223 l = l->next;
4224
4225 if (l == it->tasks_head)
4226 l = it->mg_tasks_head->next;
4227
4228 if (l == it->mg_tasks_head)
4229 css_task_iter_advance_css_set(it);
4230 else
4231 it->task_pos = l;
4232 }
4233
4234 /**
4235 * css_task_iter_start - initiate task iteration
4236 * @css: the css to walk tasks of
4237 * @it: the task iterator to use
4238 *
4239 * Initiate iteration through the tasks of @css. The caller can call
4240 * css_task_iter_next() to walk through the tasks until the function
4241 * returns NULL. On completion of iteration, css_task_iter_end() must be
4242 * called.
4243 */
4244 void css_task_iter_start(struct cgroup_subsys_state *css,
4245 struct css_task_iter *it)
4246 {
4247 /* no one should try to iterate before mounting cgroups */
4248 WARN_ON_ONCE(!use_task_css_set_links);
4249
4250 memset(it, 0, sizeof(*it));
4251
4252 spin_lock_bh(&css_set_lock);
4253
4254 it->ss = css->ss;
4255
4256 if (it->ss)
4257 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4258 else
4259 it->cset_pos = &css->cgroup->cset_links;
4260
4261 it->cset_head = it->cset_pos;
4262
4263 css_task_iter_advance_css_set(it);
4264
4265 spin_unlock_bh(&css_set_lock);
4266 }
4267
4268 /**
4269 * css_task_iter_next - return the next task for the iterator
4270 * @it: the task iterator being iterated
4271 *
4272 * The "next" function for task iteration. @it should have been
4273 * initialized via css_task_iter_start(). Returns NULL when the iteration
4274 * reaches the end.
4275 */
4276 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4277 {
4278 if (it->cur_task) {
4279 put_task_struct(it->cur_task);
4280 it->cur_task = NULL;
4281 }
4282
4283 spin_lock_bh(&css_set_lock);
4284
4285 if (it->task_pos) {
4286 it->cur_task = list_entry(it->task_pos, struct task_struct,
4287 cg_list);
4288 get_task_struct(it->cur_task);
4289 css_task_iter_advance(it);
4290 }
4291
4292 spin_unlock_bh(&css_set_lock);
4293
4294 return it->cur_task;
4295 }
4296
4297 /**
4298 * css_task_iter_end - finish task iteration
4299 * @it: the task iterator to finish
4300 *
4301 * Finish task iteration started by css_task_iter_start().
4302 */
4303 void css_task_iter_end(struct css_task_iter *it)
4304 {
4305 if (it->cur_cset) {
4306 spin_lock_bh(&css_set_lock);
4307 list_del(&it->iters_node);
4308 put_css_set_locked(it->cur_cset);
4309 spin_unlock_bh(&css_set_lock);
4310 }
4311
4312 if (it->cur_task)
4313 put_task_struct(it->cur_task);
4314 }
4315
4316 /**
4317 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4318 * @to: cgroup to which the tasks will be moved
4319 * @from: cgroup in which the tasks currently reside
4320 *
4321 * Locking rules between cgroup_post_fork() and the migration path
4322 * guarantee that, if a task is forking while being migrated, the new child
4323 * is guaranteed to be either visible in the source cgroup after the
4324 * parent's migration is complete or put into the target cgroup. No task
4325 * can slip out of migration through forking.
4326 */
4327 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4328 {
4329 LIST_HEAD(preloaded_csets);
4330 struct cgrp_cset_link *link;
4331 struct css_task_iter it;
4332 struct task_struct *task;
4333 int ret;
4334
4335 if (!cgroup_may_migrate_to(to))
4336 return -EBUSY;
4337
4338 mutex_lock(&cgroup_mutex);
4339
4340 /* all tasks in @from are being moved, all csets are source */
4341 spin_lock_bh(&css_set_lock);
4342 list_for_each_entry(link, &from->cset_links, cset_link)
4343 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4344 spin_unlock_bh(&css_set_lock);
4345
4346 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4347 if (ret)
4348 goto out_err;
4349
4350 /*
4351 * Migrate tasks one-by-one until @from is empty. This fails iff
4352 * ->can_attach() fails.
4353 */
4354 do {
4355 css_task_iter_start(&from->self, &it);
4356 task = css_task_iter_next(&it);
4357 if (task)
4358 get_task_struct(task);
4359 css_task_iter_end(&it);
4360
4361 if (task) {
4362 ret = cgroup_migrate(task, false, to->root);
4363 put_task_struct(task);
4364 }
4365 } while (task && !ret);
4366 out_err:
4367 cgroup_migrate_finish(&preloaded_csets);
4368 mutex_unlock(&cgroup_mutex);
4369 return ret;
4370 }
4371
4372 /*
4373 * Stuff for reading the 'tasks'/'procs' files.
4374 *
4375 * Reading this file can return large amounts of data if a cgroup has
4376 * *lots* of attached tasks. So it may need several calls to read(),
4377 * but we cannot guarantee that the information we produce is correct
4378 * unless we produce it entirely atomically.
4379 *
4380 */
4381
4382 /* which pidlist file are we talking about? */
4383 enum cgroup_filetype {
4384 CGROUP_FILE_PROCS,
4385 CGROUP_FILE_TASKS,
4386 };
4387
4388 /*
4389 * A pidlist is a list of pids that virtually represents the contents of one
4390 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4391 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4392 * to the cgroup.
4393 */
4394 struct cgroup_pidlist {
4395 /*
4396 * used to find which pidlist is wanted. doesn't change as long as
4397 * this particular list stays in the list.
4398 */
4399 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4400 /* array of xids */
4401 pid_t *list;
4402 /* how many elements the above list has */
4403 int length;
4404 /* each of these stored in a list by its cgroup */
4405 struct list_head links;
4406 /* pointer to the cgroup we belong to, for list removal purposes */
4407 struct cgroup *owner;
4408 /* for delayed destruction */
4409 struct delayed_work destroy_dwork;
4410 };
4411
4412 /*
4413 * The following two functions "fix" the issue where there are more pids
4414 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4415 * TODO: replace with a kernel-wide solution to this problem
4416 */
4417 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4418 static void *pidlist_allocate(int count)
4419 {
4420 if (PIDLIST_TOO_LARGE(count))
4421 return vmalloc(count * sizeof(pid_t));
4422 else
4423 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4424 }
4425
4426 static void pidlist_free(void *p)
4427 {
4428 kvfree(p);
4429 }
4430
4431 /*
4432 * Used to destroy all pidlists lingering waiting for destroy timer. None
4433 * should be left afterwards.
4434 */
4435 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4436 {
4437 struct cgroup_pidlist *l, *tmp_l;
4438
4439 mutex_lock(&cgrp->pidlist_mutex);
4440 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4441 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4442 mutex_unlock(&cgrp->pidlist_mutex);
4443
4444 flush_workqueue(cgroup_pidlist_destroy_wq);
4445 BUG_ON(!list_empty(&cgrp->pidlists));
4446 }
4447
4448 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4449 {
4450 struct delayed_work *dwork = to_delayed_work(work);
4451 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4452 destroy_dwork);
4453 struct cgroup_pidlist *tofree = NULL;
4454
4455 mutex_lock(&l->owner->pidlist_mutex);
4456
4457 /*
4458 * Destroy iff we didn't get queued again. The state won't change
4459 * as destroy_dwork can only be queued while locked.
4460 */
4461 if (!delayed_work_pending(dwork)) {
4462 list_del(&l->links);
4463 pidlist_free(l->list);
4464 put_pid_ns(l->key.ns);
4465 tofree = l;
4466 }
4467
4468 mutex_unlock(&l->owner->pidlist_mutex);
4469 kfree(tofree);
4470 }
4471
4472 /*
4473 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4474 * Returns the number of unique elements.
4475 */
4476 static int pidlist_uniq(pid_t *list, int length)
4477 {
4478 int src, dest = 1;
4479
4480 /*
4481 * we presume the 0th element is unique, so i starts at 1. trivial
4482 * edge cases first; no work needs to be done for either
4483 */
4484 if (length == 0 || length == 1)
4485 return length;
4486 /* src and dest walk down the list; dest counts unique elements */
4487 for (src = 1; src < length; src++) {
4488 /* find next unique element */
4489 while (list[src] == list[src-1]) {
4490 src++;
4491 if (src == length)
4492 goto after;
4493 }
4494 /* dest always points to where the next unique element goes */
4495 list[dest] = list[src];
4496 dest++;
4497 }
4498 after:
4499 return dest;
4500 }
4501
4502 /*
4503 * The two pid files - task and cgroup.procs - guaranteed that the result
4504 * is sorted, which forced this whole pidlist fiasco. As pid order is
4505 * different per namespace, each namespace needs differently sorted list,
4506 * making it impossible to use, for example, single rbtree of member tasks
4507 * sorted by task pointer. As pidlists can be fairly large, allocating one
4508 * per open file is dangerous, so cgroup had to implement shared pool of
4509 * pidlists keyed by cgroup and namespace.
4510 *
4511 * All this extra complexity was caused by the original implementation
4512 * committing to an entirely unnecessary property. In the long term, we
4513 * want to do away with it. Explicitly scramble sort order if on the
4514 * default hierarchy so that no such expectation exists in the new
4515 * interface.
4516 *
4517 * Scrambling is done by swapping every two consecutive bits, which is
4518 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4519 */
4520 static pid_t pid_fry(pid_t pid)
4521 {
4522 unsigned a = pid & 0x55555555;
4523 unsigned b = pid & 0xAAAAAAAA;
4524
4525 return (a << 1) | (b >> 1);
4526 }
4527
4528 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4529 {
4530 if (cgroup_on_dfl(cgrp))
4531 return pid_fry(pid);
4532 else
4533 return pid;
4534 }
4535
4536 static int cmppid(const void *a, const void *b)
4537 {
4538 return *(pid_t *)a - *(pid_t *)b;
4539 }
4540
4541 static int fried_cmppid(const void *a, const void *b)
4542 {
4543 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4544 }
4545
4546 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4547 enum cgroup_filetype type)
4548 {
4549 struct cgroup_pidlist *l;
4550 /* don't need task_nsproxy() if we're looking at ourself */
4551 struct pid_namespace *ns = task_active_pid_ns(current);
4552
4553 lockdep_assert_held(&cgrp->pidlist_mutex);
4554
4555 list_for_each_entry(l, &cgrp->pidlists, links)
4556 if (l->key.type == type && l->key.ns == ns)
4557 return l;
4558 return NULL;
4559 }
4560
4561 /*
4562 * find the appropriate pidlist for our purpose (given procs vs tasks)
4563 * returns with the lock on that pidlist already held, and takes care
4564 * of the use count, or returns NULL with no locks held if we're out of
4565 * memory.
4566 */
4567 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4568 enum cgroup_filetype type)
4569 {
4570 struct cgroup_pidlist *l;
4571
4572 lockdep_assert_held(&cgrp->pidlist_mutex);
4573
4574 l = cgroup_pidlist_find(cgrp, type);
4575 if (l)
4576 return l;
4577
4578 /* entry not found; create a new one */
4579 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4580 if (!l)
4581 return l;
4582
4583 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4584 l->key.type = type;
4585 /* don't need task_nsproxy() if we're looking at ourself */
4586 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4587 l->owner = cgrp;
4588 list_add(&l->links, &cgrp->pidlists);
4589 return l;
4590 }
4591
4592 /*
4593 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4594 */
4595 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4596 struct cgroup_pidlist **lp)
4597 {
4598 pid_t *array;
4599 int length;
4600 int pid, n = 0; /* used for populating the array */
4601 struct css_task_iter it;
4602 struct task_struct *tsk;
4603 struct cgroup_pidlist *l;
4604
4605 lockdep_assert_held(&cgrp->pidlist_mutex);
4606
4607 /*
4608 * If cgroup gets more users after we read count, we won't have
4609 * enough space - tough. This race is indistinguishable to the
4610 * caller from the case that the additional cgroup users didn't
4611 * show up until sometime later on.
4612 */
4613 length = cgroup_task_count(cgrp);
4614 array = pidlist_allocate(length);
4615 if (!array)
4616 return -ENOMEM;
4617 /* now, populate the array */
4618 css_task_iter_start(&cgrp->self, &it);
4619 while ((tsk = css_task_iter_next(&it))) {
4620 if (unlikely(n == length))
4621 break;
4622 /* get tgid or pid for procs or tasks file respectively */
4623 if (type == CGROUP_FILE_PROCS)
4624 pid = task_tgid_vnr(tsk);
4625 else
4626 pid = task_pid_vnr(tsk);
4627 if (pid > 0) /* make sure to only use valid results */
4628 array[n++] = pid;
4629 }
4630 css_task_iter_end(&it);
4631 length = n;
4632 /* now sort & (if procs) strip out duplicates */
4633 if (cgroup_on_dfl(cgrp))
4634 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4635 else
4636 sort(array, length, sizeof(pid_t), cmppid, NULL);
4637 if (type == CGROUP_FILE_PROCS)
4638 length = pidlist_uniq(array, length);
4639
4640 l = cgroup_pidlist_find_create(cgrp, type);
4641 if (!l) {
4642 pidlist_free(array);
4643 return -ENOMEM;
4644 }
4645
4646 /* store array, freeing old if necessary */
4647 pidlist_free(l->list);
4648 l->list = array;
4649 l->length = length;
4650 *lp = l;
4651 return 0;
4652 }
4653
4654 /**
4655 * cgroupstats_build - build and fill cgroupstats
4656 * @stats: cgroupstats to fill information into
4657 * @dentry: A dentry entry belonging to the cgroup for which stats have
4658 * been requested.
4659 *
4660 * Build and fill cgroupstats so that taskstats can export it to user
4661 * space.
4662 */
4663 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4664 {
4665 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4666 struct cgroup *cgrp;
4667 struct css_task_iter it;
4668 struct task_struct *tsk;
4669
4670 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4671 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4672 kernfs_type(kn) != KERNFS_DIR)
4673 return -EINVAL;
4674
4675 mutex_lock(&cgroup_mutex);
4676
4677 /*
4678 * We aren't being called from kernfs and there's no guarantee on
4679 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4680 * @kn->priv is RCU safe. Let's do the RCU dancing.
4681 */
4682 rcu_read_lock();
4683 cgrp = rcu_dereference(kn->priv);
4684 if (!cgrp || cgroup_is_dead(cgrp)) {
4685 rcu_read_unlock();
4686 mutex_unlock(&cgroup_mutex);
4687 return -ENOENT;
4688 }
4689 rcu_read_unlock();
4690
4691 css_task_iter_start(&cgrp->self, &it);
4692 while ((tsk = css_task_iter_next(&it))) {
4693 switch (tsk->state) {
4694 case TASK_RUNNING:
4695 stats->nr_running++;
4696 break;
4697 case TASK_INTERRUPTIBLE:
4698 stats->nr_sleeping++;
4699 break;
4700 case TASK_UNINTERRUPTIBLE:
4701 stats->nr_uninterruptible++;
4702 break;
4703 case TASK_STOPPED:
4704 stats->nr_stopped++;
4705 break;
4706 default:
4707 if (delayacct_is_task_waiting_on_io(tsk))
4708 stats->nr_io_wait++;
4709 break;
4710 }
4711 }
4712 css_task_iter_end(&it);
4713
4714 mutex_unlock(&cgroup_mutex);
4715 return 0;
4716 }
4717
4718
4719 /*
4720 * seq_file methods for the tasks/procs files. The seq_file position is the
4721 * next pid to display; the seq_file iterator is a pointer to the pid
4722 * in the cgroup->l->list array.
4723 */
4724
4725 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4726 {
4727 /*
4728 * Initially we receive a position value that corresponds to
4729 * one more than the last pid shown (or 0 on the first call or
4730 * after a seek to the start). Use a binary-search to find the
4731 * next pid to display, if any
4732 */
4733 struct kernfs_open_file *of = s->private;
4734 struct cgroup *cgrp = seq_css(s)->cgroup;
4735 struct cgroup_pidlist *l;
4736 enum cgroup_filetype type = seq_cft(s)->private;
4737 int index = 0, pid = *pos;
4738 int *iter, ret;
4739
4740 mutex_lock(&cgrp->pidlist_mutex);
4741
4742 /*
4743 * !NULL @of->priv indicates that this isn't the first start()
4744 * after open. If the matching pidlist is around, we can use that.
4745 * Look for it. Note that @of->priv can't be used directly. It
4746 * could already have been destroyed.
4747 */
4748 if (of->priv)
4749 of->priv = cgroup_pidlist_find(cgrp, type);
4750
4751 /*
4752 * Either this is the first start() after open or the matching
4753 * pidlist has been destroyed inbetween. Create a new one.
4754 */
4755 if (!of->priv) {
4756 ret = pidlist_array_load(cgrp, type,
4757 (struct cgroup_pidlist **)&of->priv);
4758 if (ret)
4759 return ERR_PTR(ret);
4760 }
4761 l = of->priv;
4762
4763 if (pid) {
4764 int end = l->length;
4765
4766 while (index < end) {
4767 int mid = (index + end) / 2;
4768 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4769 index = mid;
4770 break;
4771 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4772 index = mid + 1;
4773 else
4774 end = mid;
4775 }
4776 }
4777 /* If we're off the end of the array, we're done */
4778 if (index >= l->length)
4779 return NULL;
4780 /* Update the abstract position to be the actual pid that we found */
4781 iter = l->list + index;
4782 *pos = cgroup_pid_fry(cgrp, *iter);
4783 return iter;
4784 }
4785
4786 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4787 {
4788 struct kernfs_open_file *of = s->private;
4789 struct cgroup_pidlist *l = of->priv;
4790
4791 if (l)
4792 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4793 CGROUP_PIDLIST_DESTROY_DELAY);
4794 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4795 }
4796
4797 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4798 {
4799 struct kernfs_open_file *of = s->private;
4800 struct cgroup_pidlist *l = of->priv;
4801 pid_t *p = v;
4802 pid_t *end = l->list + l->length;
4803 /*
4804 * Advance to the next pid in the array. If this goes off the
4805 * end, we're done
4806 */
4807 p++;
4808 if (p >= end) {
4809 return NULL;
4810 } else {
4811 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4812 return p;
4813 }
4814 }
4815
4816 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4817 {
4818 seq_printf(s, "%d\n", *(int *)v);
4819
4820 return 0;
4821 }
4822
4823 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4824 struct cftype *cft)
4825 {
4826 return notify_on_release(css->cgroup);
4827 }
4828
4829 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4830 struct cftype *cft, u64 val)
4831 {
4832 if (val)
4833 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4834 else
4835 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4836 return 0;
4837 }
4838
4839 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4840 struct cftype *cft)
4841 {
4842 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4843 }
4844
4845 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4846 struct cftype *cft, u64 val)
4847 {
4848 if (val)
4849 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4850 else
4851 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4852 return 0;
4853 }
4854
4855 /* cgroup core interface files for the default hierarchy */
4856 static struct cftype cgroup_dfl_base_files[] = {
4857 {
4858 .name = "cgroup.procs",
4859 .file_offset = offsetof(struct cgroup, procs_file),
4860 .seq_start = cgroup_pidlist_start,
4861 .seq_next = cgroup_pidlist_next,
4862 .seq_stop = cgroup_pidlist_stop,
4863 .seq_show = cgroup_pidlist_show,
4864 .private = CGROUP_FILE_PROCS,
4865 .write = cgroup_procs_write,
4866 },
4867 {
4868 .name = "cgroup.controllers",
4869 .seq_show = cgroup_controllers_show,
4870 },
4871 {
4872 .name = "cgroup.subtree_control",
4873 .seq_show = cgroup_subtree_control_show,
4874 .write = cgroup_subtree_control_write,
4875 },
4876 {
4877 .name = "cgroup.events",
4878 .flags = CFTYPE_NOT_ON_ROOT,
4879 .file_offset = offsetof(struct cgroup, events_file),
4880 .seq_show = cgroup_events_show,
4881 },
4882 { } /* terminate */
4883 };
4884
4885 /* cgroup core interface files for the legacy hierarchies */
4886 static struct cftype cgroup_legacy_base_files[] = {
4887 {
4888 .name = "cgroup.procs",
4889 .seq_start = cgroup_pidlist_start,
4890 .seq_next = cgroup_pidlist_next,
4891 .seq_stop = cgroup_pidlist_stop,
4892 .seq_show = cgroup_pidlist_show,
4893 .private = CGROUP_FILE_PROCS,
4894 .write = cgroup_procs_write,
4895 },
4896 {
4897 .name = "cgroup.clone_children",
4898 .read_u64 = cgroup_clone_children_read,
4899 .write_u64 = cgroup_clone_children_write,
4900 },
4901 {
4902 .name = "cgroup.sane_behavior",
4903 .flags = CFTYPE_ONLY_ON_ROOT,
4904 .seq_show = cgroup_sane_behavior_show,
4905 },
4906 {
4907 .name = "tasks",
4908 .seq_start = cgroup_pidlist_start,
4909 .seq_next = cgroup_pidlist_next,
4910 .seq_stop = cgroup_pidlist_stop,
4911 .seq_show = cgroup_pidlist_show,
4912 .private = CGROUP_FILE_TASKS,
4913 .write = cgroup_tasks_write,
4914 },
4915 {
4916 .name = "notify_on_release",
4917 .read_u64 = cgroup_read_notify_on_release,
4918 .write_u64 = cgroup_write_notify_on_release,
4919 },
4920 {
4921 .name = "release_agent",
4922 .flags = CFTYPE_ONLY_ON_ROOT,
4923 .seq_show = cgroup_release_agent_show,
4924 .write = cgroup_release_agent_write,
4925 .max_write_len = PATH_MAX - 1,
4926 },
4927 { } /* terminate */
4928 };
4929
4930 /*
4931 * css destruction is four-stage process.
4932 *
4933 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4934 * Implemented in kill_css().
4935 *
4936 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4937 * and thus css_tryget_online() is guaranteed to fail, the css can be
4938 * offlined by invoking offline_css(). After offlining, the base ref is
4939 * put. Implemented in css_killed_work_fn().
4940 *
4941 * 3. When the percpu_ref reaches zero, the only possible remaining
4942 * accessors are inside RCU read sections. css_release() schedules the
4943 * RCU callback.
4944 *
4945 * 4. After the grace period, the css can be freed. Implemented in
4946 * css_free_work_fn().
4947 *
4948 * It is actually hairier because both step 2 and 4 require process context
4949 * and thus involve punting to css->destroy_work adding two additional
4950 * steps to the already complex sequence.
4951 */
4952 static void css_free_work_fn(struct work_struct *work)
4953 {
4954 struct cgroup_subsys_state *css =
4955 container_of(work, struct cgroup_subsys_state, destroy_work);
4956 struct cgroup_subsys *ss = css->ss;
4957 struct cgroup *cgrp = css->cgroup;
4958
4959 percpu_ref_exit(&css->refcnt);
4960
4961 if (ss) {
4962 /* css free path */
4963 struct cgroup_subsys_state *parent = css->parent;
4964 int id = css->id;
4965
4966 ss->css_free(css);
4967 cgroup_idr_remove(&ss->css_idr, id);
4968 cgroup_put(cgrp);
4969
4970 if (parent)
4971 css_put(parent);
4972 } else {
4973 /* cgroup free path */
4974 atomic_dec(&cgrp->root->nr_cgrps);
4975 cgroup_pidlist_destroy_all(cgrp);
4976 cancel_work_sync(&cgrp->release_agent_work);
4977
4978 if (cgroup_parent(cgrp)) {
4979 /*
4980 * We get a ref to the parent, and put the ref when
4981 * this cgroup is being freed, so it's guaranteed
4982 * that the parent won't be destroyed before its
4983 * children.
4984 */
4985 cgroup_put(cgroup_parent(cgrp));
4986 kernfs_put(cgrp->kn);
4987 kfree(cgrp);
4988 } else {
4989 /*
4990 * This is root cgroup's refcnt reaching zero,
4991 * which indicates that the root should be
4992 * released.
4993 */
4994 cgroup_destroy_root(cgrp->root);
4995 }
4996 }
4997 }
4998
4999 static void css_free_rcu_fn(struct rcu_head *rcu_head)
5000 {
5001 struct cgroup_subsys_state *css =
5002 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
5003
5004 INIT_WORK(&css->destroy_work, css_free_work_fn);
5005 queue_work(cgroup_destroy_wq, &css->destroy_work);
5006 }
5007
5008 static void css_release_work_fn(struct work_struct *work)
5009 {
5010 struct cgroup_subsys_state *css =
5011 container_of(work, struct cgroup_subsys_state, destroy_work);
5012 struct cgroup_subsys *ss = css->ss;
5013 struct cgroup *cgrp = css->cgroup;
5014
5015 mutex_lock(&cgroup_mutex);
5016
5017 css->flags |= CSS_RELEASED;
5018 list_del_rcu(&css->sibling);
5019
5020 if (ss) {
5021 /* css release path */
5022 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5023 if (ss->css_released)
5024 ss->css_released(css);
5025 } else {
5026 /* cgroup release path */
5027 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
5028 cgrp->id = -1;
5029
5030 /*
5031 * There are two control paths which try to determine
5032 * cgroup from dentry without going through kernfs -
5033 * cgroupstats_build() and css_tryget_online_from_dir().
5034 * Those are supported by RCU protecting clearing of
5035 * cgrp->kn->priv backpointer.
5036 */
5037 if (cgrp->kn)
5038 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5039 NULL);
5040 }
5041
5042 mutex_unlock(&cgroup_mutex);
5043
5044 call_rcu(&css->rcu_head, css_free_rcu_fn);
5045 }
5046
5047 static void css_release(struct percpu_ref *ref)
5048 {
5049 struct cgroup_subsys_state *css =
5050 container_of(ref, struct cgroup_subsys_state, refcnt);
5051
5052 INIT_WORK(&css->destroy_work, css_release_work_fn);
5053 queue_work(cgroup_destroy_wq, &css->destroy_work);
5054 }
5055
5056 static void init_and_link_css(struct cgroup_subsys_state *css,
5057 struct cgroup_subsys *ss, struct cgroup *cgrp)
5058 {
5059 lockdep_assert_held(&cgroup_mutex);
5060
5061 cgroup_get(cgrp);
5062
5063 memset(css, 0, sizeof(*css));
5064 css->cgroup = cgrp;
5065 css->ss = ss;
5066 INIT_LIST_HEAD(&css->sibling);
5067 INIT_LIST_HEAD(&css->children);
5068 css->serial_nr = css_serial_nr_next++;
5069 atomic_set(&css->online_cnt, 0);
5070
5071 if (cgroup_parent(cgrp)) {
5072 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5073 css_get(css->parent);
5074 }
5075
5076 BUG_ON(cgroup_css(cgrp, ss));
5077 }
5078
5079 /* invoke ->css_online() on a new CSS and mark it online if successful */
5080 static int online_css(struct cgroup_subsys_state *css)
5081 {
5082 struct cgroup_subsys *ss = css->ss;
5083 int ret = 0;
5084
5085 lockdep_assert_held(&cgroup_mutex);
5086
5087 if (ss->css_online)
5088 ret = ss->css_online(css);
5089 if (!ret) {
5090 css->flags |= CSS_ONLINE;
5091 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5092
5093 atomic_inc(&css->online_cnt);
5094 if (css->parent)
5095 atomic_inc(&css->parent->online_cnt);
5096 }
5097 return ret;
5098 }
5099
5100 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5101 static void offline_css(struct cgroup_subsys_state *css)
5102 {
5103 struct cgroup_subsys *ss = css->ss;
5104
5105 lockdep_assert_held(&cgroup_mutex);
5106
5107 if (!(css->flags & CSS_ONLINE))
5108 return;
5109
5110 if (ss->css_reset)
5111 ss->css_reset(css);
5112
5113 if (ss->css_offline)
5114 ss->css_offline(css);
5115
5116 css->flags &= ~CSS_ONLINE;
5117 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5118
5119 wake_up_all(&css->cgroup->offline_waitq);
5120 }
5121
5122 /**
5123 * css_create - create a cgroup_subsys_state
5124 * @cgrp: the cgroup new css will be associated with
5125 * @ss: the subsys of new css
5126 *
5127 * Create a new css associated with @cgrp - @ss pair. On success, the new
5128 * css is online and installed in @cgrp. This function doesn't create the
5129 * interface files. Returns 0 on success, -errno on failure.
5130 */
5131 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5132 struct cgroup_subsys *ss)
5133 {
5134 struct cgroup *parent = cgroup_parent(cgrp);
5135 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5136 struct cgroup_subsys_state *css;
5137 int err;
5138
5139 lockdep_assert_held(&cgroup_mutex);
5140
5141 css = ss->css_alloc(parent_css);
5142 if (IS_ERR(css))
5143 return css;
5144
5145 init_and_link_css(css, ss, cgrp);
5146
5147 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5148 if (err)
5149 goto err_free_css;
5150
5151 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5152 if (err < 0)
5153 goto err_free_percpu_ref;
5154 css->id = err;
5155
5156 /* @css is ready to be brought online now, make it visible */
5157 list_add_tail_rcu(&css->sibling, &parent_css->children);
5158 cgroup_idr_replace(&ss->css_idr, css, css->id);
5159
5160 err = online_css(css);
5161 if (err)
5162 goto err_list_del;
5163
5164 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5165 cgroup_parent(parent)) {
5166 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5167 current->comm, current->pid, ss->name);
5168 if (!strcmp(ss->name, "memory"))
5169 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5170 ss->warned_broken_hierarchy = true;
5171 }
5172
5173 return css;
5174
5175 err_list_del:
5176 list_del_rcu(&css->sibling);
5177 cgroup_idr_remove(&ss->css_idr, css->id);
5178 err_free_percpu_ref:
5179 percpu_ref_exit(&css->refcnt);
5180 err_free_css:
5181 call_rcu(&css->rcu_head, css_free_rcu_fn);
5182 return ERR_PTR(err);
5183 }
5184
5185 static struct cgroup *cgroup_create(struct cgroup *parent)
5186 {
5187 struct cgroup_root *root = parent->root;
5188 struct cgroup *cgrp, *tcgrp;
5189 int level = parent->level + 1;
5190 int ret;
5191
5192 /* allocate the cgroup and its ID, 0 is reserved for the root */
5193 cgrp = kzalloc(sizeof(*cgrp) +
5194 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5195 if (!cgrp)
5196 return ERR_PTR(-ENOMEM);
5197
5198 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5199 if (ret)
5200 goto out_free_cgrp;
5201
5202 /*
5203 * Temporarily set the pointer to NULL, so idr_find() won't return
5204 * a half-baked cgroup.
5205 */
5206 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5207 if (cgrp->id < 0) {
5208 ret = -ENOMEM;
5209 goto out_cancel_ref;
5210 }
5211
5212 init_cgroup_housekeeping(cgrp);
5213
5214 cgrp->self.parent = &parent->self;
5215 cgrp->root = root;
5216 cgrp->level = level;
5217
5218 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5219 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5220
5221 if (notify_on_release(parent))
5222 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5223
5224 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5225 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5226
5227 cgrp->self.serial_nr = css_serial_nr_next++;
5228
5229 /* allocation complete, commit to creation */
5230 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5231 atomic_inc(&root->nr_cgrps);
5232 cgroup_get(parent);
5233
5234 /*
5235 * @cgrp is now fully operational. If something fails after this
5236 * point, it'll be released via the normal destruction path.
5237 */
5238 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5239
5240 /*
5241 * On the default hierarchy, a child doesn't automatically inherit
5242 * subtree_control from the parent. Each is configured manually.
5243 */
5244 if (!cgroup_on_dfl(cgrp))
5245 cgrp->subtree_control = cgroup_control(cgrp);
5246
5247 cgroup_propagate_control(cgrp);
5248
5249 /* @cgrp doesn't have dir yet so the following will only create csses */
5250 ret = cgroup_apply_control_enable(cgrp);
5251 if (ret)
5252 goto out_destroy;
5253
5254 return cgrp;
5255
5256 out_cancel_ref:
5257 percpu_ref_exit(&cgrp->self.refcnt);
5258 out_free_cgrp:
5259 kfree(cgrp);
5260 return ERR_PTR(ret);
5261 out_destroy:
5262 cgroup_destroy_locked(cgrp);
5263 return ERR_PTR(ret);
5264 }
5265
5266 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5267 umode_t mode)
5268 {
5269 struct cgroup *parent, *cgrp;
5270 struct kernfs_node *kn;
5271 int ret;
5272
5273 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5274 if (strchr(name, '\n'))
5275 return -EINVAL;
5276
5277 parent = cgroup_kn_lock_live(parent_kn, false);
5278 if (!parent)
5279 return -ENODEV;
5280
5281 cgrp = cgroup_create(parent);
5282 if (IS_ERR(cgrp)) {
5283 ret = PTR_ERR(cgrp);
5284 goto out_unlock;
5285 }
5286
5287 /* create the directory */
5288 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5289 if (IS_ERR(kn)) {
5290 ret = PTR_ERR(kn);
5291 goto out_destroy;
5292 }
5293 cgrp->kn = kn;
5294
5295 /*
5296 * This extra ref will be put in cgroup_free_fn() and guarantees
5297 * that @cgrp->kn is always accessible.
5298 */
5299 kernfs_get(kn);
5300
5301 ret = cgroup_kn_set_ugid(kn);
5302 if (ret)
5303 goto out_destroy;
5304
5305 ret = css_populate_dir(&cgrp->self);
5306 if (ret)
5307 goto out_destroy;
5308
5309 ret = cgroup_apply_control_enable(cgrp);
5310 if (ret)
5311 goto out_destroy;
5312
5313 /* let's create and online css's */
5314 kernfs_activate(kn);
5315
5316 ret = 0;
5317 goto out_unlock;
5318
5319 out_destroy:
5320 cgroup_destroy_locked(cgrp);
5321 out_unlock:
5322 cgroup_kn_unlock(parent_kn);
5323 return ret;
5324 }
5325
5326 /*
5327 * This is called when the refcnt of a css is confirmed to be killed.
5328 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5329 * initate destruction and put the css ref from kill_css().
5330 */
5331 static void css_killed_work_fn(struct work_struct *work)
5332 {
5333 struct cgroup_subsys_state *css =
5334 container_of(work, struct cgroup_subsys_state, destroy_work);
5335
5336 mutex_lock(&cgroup_mutex);
5337
5338 do {
5339 offline_css(css);
5340 css_put(css);
5341 /* @css can't go away while we're holding cgroup_mutex */
5342 css = css->parent;
5343 } while (css && atomic_dec_and_test(&css->online_cnt));
5344
5345 mutex_unlock(&cgroup_mutex);
5346 }
5347
5348 /* css kill confirmation processing requires process context, bounce */
5349 static void css_killed_ref_fn(struct percpu_ref *ref)
5350 {
5351 struct cgroup_subsys_state *css =
5352 container_of(ref, struct cgroup_subsys_state, refcnt);
5353
5354 if (atomic_dec_and_test(&css->online_cnt)) {
5355 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5356 queue_work(cgroup_destroy_wq, &css->destroy_work);
5357 }
5358 }
5359
5360 /**
5361 * kill_css - destroy a css
5362 * @css: css to destroy
5363 *
5364 * This function initiates destruction of @css by removing cgroup interface
5365 * files and putting its base reference. ->css_offline() will be invoked
5366 * asynchronously once css_tryget_online() is guaranteed to fail and when
5367 * the reference count reaches zero, @css will be released.
5368 */
5369 static void kill_css(struct cgroup_subsys_state *css)
5370 {
5371 lockdep_assert_held(&cgroup_mutex);
5372
5373 /*
5374 * This must happen before css is disassociated with its cgroup.
5375 * See seq_css() for details.
5376 */
5377 css_clear_dir(css);
5378
5379 /*
5380 * Killing would put the base ref, but we need to keep it alive
5381 * until after ->css_offline().
5382 */
5383 css_get(css);
5384
5385 /*
5386 * cgroup core guarantees that, by the time ->css_offline() is
5387 * invoked, no new css reference will be given out via
5388 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5389 * proceed to offlining css's because percpu_ref_kill() doesn't
5390 * guarantee that the ref is seen as killed on all CPUs on return.
5391 *
5392 * Use percpu_ref_kill_and_confirm() to get notifications as each
5393 * css is confirmed to be seen as killed on all CPUs.
5394 */
5395 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5396 }
5397
5398 /**
5399 * cgroup_destroy_locked - the first stage of cgroup destruction
5400 * @cgrp: cgroup to be destroyed
5401 *
5402 * css's make use of percpu refcnts whose killing latency shouldn't be
5403 * exposed to userland and are RCU protected. Also, cgroup core needs to
5404 * guarantee that css_tryget_online() won't succeed by the time
5405 * ->css_offline() is invoked. To satisfy all the requirements,
5406 * destruction is implemented in the following two steps.
5407 *
5408 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5409 * userland visible parts and start killing the percpu refcnts of
5410 * css's. Set up so that the next stage will be kicked off once all
5411 * the percpu refcnts are confirmed to be killed.
5412 *
5413 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5414 * rest of destruction. Once all cgroup references are gone, the
5415 * cgroup is RCU-freed.
5416 *
5417 * This function implements s1. After this step, @cgrp is gone as far as
5418 * the userland is concerned and a new cgroup with the same name may be
5419 * created. As cgroup doesn't care about the names internally, this
5420 * doesn't cause any problem.
5421 */
5422 static int cgroup_destroy_locked(struct cgroup *cgrp)
5423 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5424 {
5425 struct cgroup_subsys_state *css;
5426 struct cgrp_cset_link *link;
5427 int ssid;
5428
5429 lockdep_assert_held(&cgroup_mutex);
5430
5431 /*
5432 * Only migration can raise populated from zero and we're already
5433 * holding cgroup_mutex.
5434 */
5435 if (cgroup_is_populated(cgrp))
5436 return -EBUSY;
5437
5438 /*
5439 * Make sure there's no live children. We can't test emptiness of
5440 * ->self.children as dead children linger on it while being
5441 * drained; otherwise, "rmdir parent/child parent" may fail.
5442 */
5443 if (css_has_online_children(&cgrp->self))
5444 return -EBUSY;
5445
5446 /*
5447 * Mark @cgrp and the associated csets dead. The former prevents
5448 * further task migration and child creation by disabling
5449 * cgroup_lock_live_group(). The latter makes the csets ignored by
5450 * the migration path.
5451 */
5452 cgrp->self.flags &= ~CSS_ONLINE;
5453
5454 spin_lock_bh(&css_set_lock);
5455 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5456 link->cset->dead = true;
5457 spin_unlock_bh(&css_set_lock);
5458
5459 /* initiate massacre of all css's */
5460 for_each_css(css, ssid, cgrp)
5461 kill_css(css);
5462
5463 /*
5464 * Remove @cgrp directory along with the base files. @cgrp has an
5465 * extra ref on its kn.
5466 */
5467 kernfs_remove(cgrp->kn);
5468
5469 check_for_release(cgroup_parent(cgrp));
5470
5471 /* put the base reference */
5472 percpu_ref_kill(&cgrp->self.refcnt);
5473
5474 return 0;
5475 };
5476
5477 static int cgroup_rmdir(struct kernfs_node *kn)
5478 {
5479 struct cgroup *cgrp;
5480 int ret = 0;
5481
5482 cgrp = cgroup_kn_lock_live(kn, false);
5483 if (!cgrp)
5484 return 0;
5485
5486 ret = cgroup_destroy_locked(cgrp);
5487
5488 cgroup_kn_unlock(kn);
5489 return ret;
5490 }
5491
5492 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5493 .remount_fs = cgroup_remount,
5494 .show_options = cgroup_show_options,
5495 .mkdir = cgroup_mkdir,
5496 .rmdir = cgroup_rmdir,
5497 .rename = cgroup_rename,
5498 .show_path = cgroup_show_path,
5499 };
5500
5501 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5502 {
5503 struct cgroup_subsys_state *css;
5504
5505 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5506
5507 mutex_lock(&cgroup_mutex);
5508
5509 idr_init(&ss->css_idr);
5510 INIT_LIST_HEAD(&ss->cfts);
5511
5512 /* Create the root cgroup state for this subsystem */
5513 ss->root = &cgrp_dfl_root;
5514 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5515 /* We don't handle early failures gracefully */
5516 BUG_ON(IS_ERR(css));
5517 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5518
5519 /*
5520 * Root csses are never destroyed and we can't initialize
5521 * percpu_ref during early init. Disable refcnting.
5522 */
5523 css->flags |= CSS_NO_REF;
5524
5525 if (early) {
5526 /* allocation can't be done safely during early init */
5527 css->id = 1;
5528 } else {
5529 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5530 BUG_ON(css->id < 0);
5531 }
5532
5533 /* Update the init_css_set to contain a subsys
5534 * pointer to this state - since the subsystem is
5535 * newly registered, all tasks and hence the
5536 * init_css_set is in the subsystem's root cgroup. */
5537 init_css_set.subsys[ss->id] = css;
5538
5539 have_fork_callback |= (bool)ss->fork << ss->id;
5540 have_exit_callback |= (bool)ss->exit << ss->id;
5541 have_free_callback |= (bool)ss->free << ss->id;
5542 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5543
5544 /* At system boot, before all subsystems have been
5545 * registered, no tasks have been forked, so we don't
5546 * need to invoke fork callbacks here. */
5547 BUG_ON(!list_empty(&init_task.tasks));
5548
5549 BUG_ON(online_css(css));
5550
5551 mutex_unlock(&cgroup_mutex);
5552 }
5553
5554 /**
5555 * cgroup_init_early - cgroup initialization at system boot
5556 *
5557 * Initialize cgroups at system boot, and initialize any
5558 * subsystems that request early init.
5559 */
5560 int __init cgroup_init_early(void)
5561 {
5562 static struct cgroup_sb_opts __initdata opts;
5563 struct cgroup_subsys *ss;
5564 int i;
5565
5566 init_cgroup_root(&cgrp_dfl_root, &opts);
5567 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5568
5569 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5570
5571 for_each_subsys(ss, i) {
5572 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5573 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5574 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5575 ss->id, ss->name);
5576 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5577 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5578
5579 ss->id = i;
5580 ss->name = cgroup_subsys_name[i];
5581 if (!ss->legacy_name)
5582 ss->legacy_name = cgroup_subsys_name[i];
5583
5584 if (ss->early_init)
5585 cgroup_init_subsys(ss, true);
5586 }
5587 return 0;
5588 }
5589
5590 static u16 cgroup_disable_mask __initdata;
5591
5592 /**
5593 * cgroup_init - cgroup initialization
5594 *
5595 * Register cgroup filesystem and /proc file, and initialize
5596 * any subsystems that didn't request early init.
5597 */
5598 int __init cgroup_init(void)
5599 {
5600 struct cgroup_subsys *ss;
5601 int ssid;
5602
5603 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5604 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5605 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5606 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5607
5608 get_user_ns(init_cgroup_ns.user_ns);
5609
5610 mutex_lock(&cgroup_mutex);
5611
5612 /*
5613 * Add init_css_set to the hash table so that dfl_root can link to
5614 * it during init.
5615 */
5616 hash_add(css_set_table, &init_css_set.hlist,
5617 css_set_hash(init_css_set.subsys));
5618
5619 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5620
5621 mutex_unlock(&cgroup_mutex);
5622
5623 for_each_subsys(ss, ssid) {
5624 if (ss->early_init) {
5625 struct cgroup_subsys_state *css =
5626 init_css_set.subsys[ss->id];
5627
5628 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5629 GFP_KERNEL);
5630 BUG_ON(css->id < 0);
5631 } else {
5632 cgroup_init_subsys(ss, false);
5633 }
5634
5635 list_add_tail(&init_css_set.e_cset_node[ssid],
5636 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5637
5638 /*
5639 * Setting dfl_root subsys_mask needs to consider the
5640 * disabled flag and cftype registration needs kmalloc,
5641 * both of which aren't available during early_init.
5642 */
5643 if (cgroup_disable_mask & (1 << ssid)) {
5644 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5645 printk(KERN_INFO "Disabling %s control group subsystem\n",
5646 ss->name);
5647 continue;
5648 }
5649
5650 if (cgroup_ssid_no_v1(ssid))
5651 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5652 ss->name);
5653
5654 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5655
5656 if (ss->implicit_on_dfl)
5657 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5658 else if (!ss->dfl_cftypes)
5659 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5660
5661 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5662 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5663 } else {
5664 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5665 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5666 }
5667
5668 if (ss->bind)
5669 ss->bind(init_css_set.subsys[ssid]);
5670 }
5671
5672 /* init_css_set.subsys[] has been updated, re-hash */
5673 hash_del(&init_css_set.hlist);
5674 hash_add(css_set_table, &init_css_set.hlist,
5675 css_set_hash(init_css_set.subsys));
5676
5677 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5678 WARN_ON(register_filesystem(&cgroup_fs_type));
5679 WARN_ON(register_filesystem(&cgroup2_fs_type));
5680 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5681
5682 return 0;
5683 }
5684
5685 static int __init cgroup_wq_init(void)
5686 {
5687 /*
5688 * There isn't much point in executing destruction path in
5689 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5690 * Use 1 for @max_active.
5691 *
5692 * We would prefer to do this in cgroup_init() above, but that
5693 * is called before init_workqueues(): so leave this until after.
5694 */
5695 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5696 BUG_ON(!cgroup_destroy_wq);
5697
5698 /*
5699 * Used to destroy pidlists and separate to serve as flush domain.
5700 * Cap @max_active to 1 too.
5701 */
5702 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5703 0, 1);
5704 BUG_ON(!cgroup_pidlist_destroy_wq);
5705
5706 return 0;
5707 }
5708 core_initcall(cgroup_wq_init);
5709
5710 /*
5711 * proc_cgroup_show()
5712 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5713 * - Used for /proc/<pid>/cgroup.
5714 */
5715 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5716 struct pid *pid, struct task_struct *tsk)
5717 {
5718 char *buf, *path;
5719 int retval;
5720 struct cgroup_root *root;
5721
5722 retval = -ENOMEM;
5723 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5724 if (!buf)
5725 goto out;
5726
5727 mutex_lock(&cgroup_mutex);
5728 spin_lock_bh(&css_set_lock);
5729
5730 for_each_root(root) {
5731 struct cgroup_subsys *ss;
5732 struct cgroup *cgrp;
5733 int ssid, count = 0;
5734
5735 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5736 continue;
5737
5738 seq_printf(m, "%d:", root->hierarchy_id);
5739 if (root != &cgrp_dfl_root)
5740 for_each_subsys(ss, ssid)
5741 if (root->subsys_mask & (1 << ssid))
5742 seq_printf(m, "%s%s", count++ ? "," : "",
5743 ss->legacy_name);
5744 if (strlen(root->name))
5745 seq_printf(m, "%sname=%s", count ? "," : "",
5746 root->name);
5747 seq_putc(m, ':');
5748
5749 cgrp = task_cgroup_from_root(tsk, root);
5750
5751 /*
5752 * On traditional hierarchies, all zombie tasks show up as
5753 * belonging to the root cgroup. On the default hierarchy,
5754 * while a zombie doesn't show up in "cgroup.procs" and
5755 * thus can't be migrated, its /proc/PID/cgroup keeps
5756 * reporting the cgroup it belonged to before exiting. If
5757 * the cgroup is removed before the zombie is reaped,
5758 * " (deleted)" is appended to the cgroup path.
5759 */
5760 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5761 path = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5762 current->nsproxy->cgroup_ns);
5763 if (!path) {
5764 retval = -ENAMETOOLONG;
5765 goto out_unlock;
5766 }
5767 } else {
5768 path = "/";
5769 }
5770
5771 seq_puts(m, path);
5772
5773 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5774 seq_puts(m, " (deleted)\n");
5775 else
5776 seq_putc(m, '\n');
5777 }
5778
5779 retval = 0;
5780 out_unlock:
5781 spin_unlock_bh(&css_set_lock);
5782 mutex_unlock(&cgroup_mutex);
5783 kfree(buf);
5784 out:
5785 return retval;
5786 }
5787
5788 /* Display information about each subsystem and each hierarchy */
5789 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5790 {
5791 struct cgroup_subsys *ss;
5792 int i;
5793
5794 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5795 /*
5796 * ideally we don't want subsystems moving around while we do this.
5797 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5798 * subsys/hierarchy state.
5799 */
5800 mutex_lock(&cgroup_mutex);
5801
5802 for_each_subsys(ss, i)
5803 seq_printf(m, "%s\t%d\t%d\t%d\n",
5804 ss->legacy_name, ss->root->hierarchy_id,
5805 atomic_read(&ss->root->nr_cgrps),
5806 cgroup_ssid_enabled(i));
5807
5808 mutex_unlock(&cgroup_mutex);
5809 return 0;
5810 }
5811
5812 static int cgroupstats_open(struct inode *inode, struct file *file)
5813 {
5814 return single_open(file, proc_cgroupstats_show, NULL);
5815 }
5816
5817 static const struct file_operations proc_cgroupstats_operations = {
5818 .open = cgroupstats_open,
5819 .read = seq_read,
5820 .llseek = seq_lseek,
5821 .release = single_release,
5822 };
5823
5824 /**
5825 * cgroup_fork - initialize cgroup related fields during copy_process()
5826 * @child: pointer to task_struct of forking parent process.
5827 *
5828 * A task is associated with the init_css_set until cgroup_post_fork()
5829 * attaches it to the parent's css_set. Empty cg_list indicates that
5830 * @child isn't holding reference to its css_set.
5831 */
5832 void cgroup_fork(struct task_struct *child)
5833 {
5834 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5835 INIT_LIST_HEAD(&child->cg_list);
5836 }
5837
5838 /**
5839 * cgroup_can_fork - called on a new task before the process is exposed
5840 * @child: the task in question.
5841 *
5842 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5843 * returns an error, the fork aborts with that error code. This allows for
5844 * a cgroup subsystem to conditionally allow or deny new forks.
5845 */
5846 int cgroup_can_fork(struct task_struct *child)
5847 {
5848 struct cgroup_subsys *ss;
5849 int i, j, ret;
5850
5851 do_each_subsys_mask(ss, i, have_canfork_callback) {
5852 ret = ss->can_fork(child);
5853 if (ret)
5854 goto out_revert;
5855 } while_each_subsys_mask();
5856
5857 return 0;
5858
5859 out_revert:
5860 for_each_subsys(ss, j) {
5861 if (j >= i)
5862 break;
5863 if (ss->cancel_fork)
5864 ss->cancel_fork(child);
5865 }
5866
5867 return ret;
5868 }
5869
5870 /**
5871 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5872 * @child: the task in question
5873 *
5874 * This calls the cancel_fork() callbacks if a fork failed *after*
5875 * cgroup_can_fork() succeded.
5876 */
5877 void cgroup_cancel_fork(struct task_struct *child)
5878 {
5879 struct cgroup_subsys *ss;
5880 int i;
5881
5882 for_each_subsys(ss, i)
5883 if (ss->cancel_fork)
5884 ss->cancel_fork(child);
5885 }
5886
5887 /**
5888 * cgroup_post_fork - called on a new task after adding it to the task list
5889 * @child: the task in question
5890 *
5891 * Adds the task to the list running through its css_set if necessary and
5892 * call the subsystem fork() callbacks. Has to be after the task is
5893 * visible on the task list in case we race with the first call to
5894 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5895 * list.
5896 */
5897 void cgroup_post_fork(struct task_struct *child)
5898 {
5899 struct cgroup_subsys *ss;
5900 int i;
5901
5902 /*
5903 * This may race against cgroup_enable_task_cg_lists(). As that
5904 * function sets use_task_css_set_links before grabbing
5905 * tasklist_lock and we just went through tasklist_lock to add
5906 * @child, it's guaranteed that either we see the set
5907 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5908 * @child during its iteration.
5909 *
5910 * If we won the race, @child is associated with %current's
5911 * css_set. Grabbing css_set_lock guarantees both that the
5912 * association is stable, and, on completion of the parent's
5913 * migration, @child is visible in the source of migration or
5914 * already in the destination cgroup. This guarantee is necessary
5915 * when implementing operations which need to migrate all tasks of
5916 * a cgroup to another.
5917 *
5918 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5919 * will remain in init_css_set. This is safe because all tasks are
5920 * in the init_css_set before cg_links is enabled and there's no
5921 * operation which transfers all tasks out of init_css_set.
5922 */
5923 if (use_task_css_set_links) {
5924 struct css_set *cset;
5925
5926 spin_lock_bh(&css_set_lock);
5927 cset = task_css_set(current);
5928 if (list_empty(&child->cg_list)) {
5929 get_css_set(cset);
5930 css_set_move_task(child, NULL, cset, false);
5931 }
5932 spin_unlock_bh(&css_set_lock);
5933 }
5934
5935 /*
5936 * Call ss->fork(). This must happen after @child is linked on
5937 * css_set; otherwise, @child might change state between ->fork()
5938 * and addition to css_set.
5939 */
5940 do_each_subsys_mask(ss, i, have_fork_callback) {
5941 ss->fork(child);
5942 } while_each_subsys_mask();
5943 }
5944
5945 /**
5946 * cgroup_exit - detach cgroup from exiting task
5947 * @tsk: pointer to task_struct of exiting process
5948 *
5949 * Description: Detach cgroup from @tsk and release it.
5950 *
5951 * Note that cgroups marked notify_on_release force every task in
5952 * them to take the global cgroup_mutex mutex when exiting.
5953 * This could impact scaling on very large systems. Be reluctant to
5954 * use notify_on_release cgroups where very high task exit scaling
5955 * is required on large systems.
5956 *
5957 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5958 * call cgroup_exit() while the task is still competent to handle
5959 * notify_on_release(), then leave the task attached to the root cgroup in
5960 * each hierarchy for the remainder of its exit. No need to bother with
5961 * init_css_set refcnting. init_css_set never goes away and we can't race
5962 * with migration path - PF_EXITING is visible to migration path.
5963 */
5964 void cgroup_exit(struct task_struct *tsk)
5965 {
5966 struct cgroup_subsys *ss;
5967 struct css_set *cset;
5968 int i;
5969
5970 /*
5971 * Unlink from @tsk from its css_set. As migration path can't race
5972 * with us, we can check css_set and cg_list without synchronization.
5973 */
5974 cset = task_css_set(tsk);
5975
5976 if (!list_empty(&tsk->cg_list)) {
5977 spin_lock_bh(&css_set_lock);
5978 css_set_move_task(tsk, cset, NULL, false);
5979 spin_unlock_bh(&css_set_lock);
5980 } else {
5981 get_css_set(cset);
5982 }
5983
5984 /* see cgroup_post_fork() for details */
5985 do_each_subsys_mask(ss, i, have_exit_callback) {
5986 ss->exit(tsk);
5987 } while_each_subsys_mask();
5988 }
5989
5990 void cgroup_free(struct task_struct *task)
5991 {
5992 struct css_set *cset = task_css_set(task);
5993 struct cgroup_subsys *ss;
5994 int ssid;
5995
5996 do_each_subsys_mask(ss, ssid, have_free_callback) {
5997 ss->free(task);
5998 } while_each_subsys_mask();
5999
6000 put_css_set(cset);
6001 }
6002
6003 static void check_for_release(struct cgroup *cgrp)
6004 {
6005 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
6006 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
6007 schedule_work(&cgrp->release_agent_work);
6008 }
6009
6010 /*
6011 * Notify userspace when a cgroup is released, by running the
6012 * configured release agent with the name of the cgroup (path
6013 * relative to the root of cgroup file system) as the argument.
6014 *
6015 * Most likely, this user command will try to rmdir this cgroup.
6016 *
6017 * This races with the possibility that some other task will be
6018 * attached to this cgroup before it is removed, or that some other
6019 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
6020 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
6021 * unused, and this cgroup will be reprieved from its death sentence,
6022 * to continue to serve a useful existence. Next time it's released,
6023 * we will get notified again, if it still has 'notify_on_release' set.
6024 *
6025 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
6026 * means only wait until the task is successfully execve()'d. The
6027 * separate release agent task is forked by call_usermodehelper(),
6028 * then control in this thread returns here, without waiting for the
6029 * release agent task. We don't bother to wait because the caller of
6030 * this routine has no use for the exit status of the release agent
6031 * task, so no sense holding our caller up for that.
6032 */
6033 static void cgroup_release_agent(struct work_struct *work)
6034 {
6035 struct cgroup *cgrp =
6036 container_of(work, struct cgroup, release_agent_work);
6037 char *pathbuf = NULL, *agentbuf = NULL, *path;
6038 char *argv[3], *envp[3];
6039
6040 mutex_lock(&cgroup_mutex);
6041
6042 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
6043 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
6044 if (!pathbuf || !agentbuf)
6045 goto out;
6046
6047 spin_lock_bh(&css_set_lock);
6048 path = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
6049 spin_unlock_bh(&css_set_lock);
6050 if (!path)
6051 goto out;
6052
6053 argv[0] = agentbuf;
6054 argv[1] = path;
6055 argv[2] = NULL;
6056
6057 /* minimal command environment */
6058 envp[0] = "HOME=/";
6059 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
6060 envp[2] = NULL;
6061
6062 mutex_unlock(&cgroup_mutex);
6063 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
6064 goto out_free;
6065 out:
6066 mutex_unlock(&cgroup_mutex);
6067 out_free:
6068 kfree(agentbuf);
6069 kfree(pathbuf);
6070 }
6071
6072 static int __init cgroup_disable(char *str)
6073 {
6074 struct cgroup_subsys *ss;
6075 char *token;
6076 int i;
6077
6078 while ((token = strsep(&str, ",")) != NULL) {
6079 if (!*token)
6080 continue;
6081
6082 for_each_subsys(ss, i) {
6083 if (strcmp(token, ss->name) &&
6084 strcmp(token, ss->legacy_name))
6085 continue;
6086 cgroup_disable_mask |= 1 << i;
6087 }
6088 }
6089 return 1;
6090 }
6091 __setup("cgroup_disable=", cgroup_disable);
6092
6093 static int __init cgroup_no_v1(char *str)
6094 {
6095 struct cgroup_subsys *ss;
6096 char *token;
6097 int i;
6098
6099 while ((token = strsep(&str, ",")) != NULL) {
6100 if (!*token)
6101 continue;
6102
6103 if (!strcmp(token, "all")) {
6104 cgroup_no_v1_mask = U16_MAX;
6105 break;
6106 }
6107
6108 for_each_subsys(ss, i) {
6109 if (strcmp(token, ss->name) &&
6110 strcmp(token, ss->legacy_name))
6111 continue;
6112
6113 cgroup_no_v1_mask |= 1 << i;
6114 }
6115 }
6116 return 1;
6117 }
6118 __setup("cgroup_no_v1=", cgroup_no_v1);
6119
6120 /**
6121 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6122 * @dentry: directory dentry of interest
6123 * @ss: subsystem of interest
6124 *
6125 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6126 * to get the corresponding css and return it. If such css doesn't exist
6127 * or can't be pinned, an ERR_PTR value is returned.
6128 */
6129 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6130 struct cgroup_subsys *ss)
6131 {
6132 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6133 struct file_system_type *s_type = dentry->d_sb->s_type;
6134 struct cgroup_subsys_state *css = NULL;
6135 struct cgroup *cgrp;
6136
6137 /* is @dentry a cgroup dir? */
6138 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6139 !kn || kernfs_type(kn) != KERNFS_DIR)
6140 return ERR_PTR(-EBADF);
6141
6142 rcu_read_lock();
6143
6144 /*
6145 * This path doesn't originate from kernfs and @kn could already
6146 * have been or be removed at any point. @kn->priv is RCU
6147 * protected for this access. See css_release_work_fn() for details.
6148 */
6149 cgrp = rcu_dereference(kn->priv);
6150 if (cgrp)
6151 css = cgroup_css(cgrp, ss);
6152
6153 if (!css || !css_tryget_online(css))
6154 css = ERR_PTR(-ENOENT);
6155
6156 rcu_read_unlock();
6157 return css;
6158 }
6159
6160 /**
6161 * css_from_id - lookup css by id
6162 * @id: the cgroup id
6163 * @ss: cgroup subsys to be looked into
6164 *
6165 * Returns the css if there's valid one with @id, otherwise returns NULL.
6166 * Should be called under rcu_read_lock().
6167 */
6168 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6169 {
6170 WARN_ON_ONCE(!rcu_read_lock_held());
6171 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
6172 }
6173
6174 /**
6175 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6176 * @path: path on the default hierarchy
6177 *
6178 * Find the cgroup at @path on the default hierarchy, increment its
6179 * reference count and return it. Returns pointer to the found cgroup on
6180 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6181 * if @path points to a non-directory.
6182 */
6183 struct cgroup *cgroup_get_from_path(const char *path)
6184 {
6185 struct kernfs_node *kn;
6186 struct cgroup *cgrp;
6187
6188 mutex_lock(&cgroup_mutex);
6189
6190 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6191 if (kn) {
6192 if (kernfs_type(kn) == KERNFS_DIR) {
6193 cgrp = kn->priv;
6194 cgroup_get(cgrp);
6195 } else {
6196 cgrp = ERR_PTR(-ENOTDIR);
6197 }
6198 kernfs_put(kn);
6199 } else {
6200 cgrp = ERR_PTR(-ENOENT);
6201 }
6202
6203 mutex_unlock(&cgroup_mutex);
6204 return cgrp;
6205 }
6206 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6207
6208 /*
6209 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6210 * definition in cgroup-defs.h.
6211 */
6212 #ifdef CONFIG_SOCK_CGROUP_DATA
6213
6214 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6215
6216 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6217 static bool cgroup_sk_alloc_disabled __read_mostly;
6218
6219 void cgroup_sk_alloc_disable(void)
6220 {
6221 if (cgroup_sk_alloc_disabled)
6222 return;
6223 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6224 cgroup_sk_alloc_disabled = true;
6225 }
6226
6227 #else
6228
6229 #define cgroup_sk_alloc_disabled false
6230
6231 #endif
6232
6233 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6234 {
6235 if (cgroup_sk_alloc_disabled)
6236 return;
6237
6238 rcu_read_lock();
6239
6240 while (true) {
6241 struct css_set *cset;
6242
6243 cset = task_css_set(current);
6244 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6245 skcd->val = (unsigned long)cset->dfl_cgrp;
6246 break;
6247 }
6248 cpu_relax();
6249 }
6250
6251 rcu_read_unlock();
6252 }
6253
6254 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6255 {
6256 cgroup_put(sock_cgroup_ptr(skcd));
6257 }
6258
6259 #endif /* CONFIG_SOCK_CGROUP_DATA */
6260
6261 /* cgroup namespaces */
6262
6263 static struct cgroup_namespace *alloc_cgroup_ns(void)
6264 {
6265 struct cgroup_namespace *new_ns;
6266 int ret;
6267
6268 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6269 if (!new_ns)
6270 return ERR_PTR(-ENOMEM);
6271 ret = ns_alloc_inum(&new_ns->ns);
6272 if (ret) {
6273 kfree(new_ns);
6274 return ERR_PTR(ret);
6275 }
6276 atomic_set(&new_ns->count, 1);
6277 new_ns->ns.ops = &cgroupns_operations;
6278 return new_ns;
6279 }
6280
6281 void free_cgroup_ns(struct cgroup_namespace *ns)
6282 {
6283 put_css_set(ns->root_cset);
6284 put_user_ns(ns->user_ns);
6285 ns_free_inum(&ns->ns);
6286 kfree(ns);
6287 }
6288 EXPORT_SYMBOL(free_cgroup_ns);
6289
6290 struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6291 struct user_namespace *user_ns,
6292 struct cgroup_namespace *old_ns)
6293 {
6294 struct cgroup_namespace *new_ns;
6295 struct css_set *cset;
6296
6297 BUG_ON(!old_ns);
6298
6299 if (!(flags & CLONE_NEWCGROUP)) {
6300 get_cgroup_ns(old_ns);
6301 return old_ns;
6302 }
6303
6304 /* Allow only sysadmin to create cgroup namespace. */
6305 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6306 return ERR_PTR(-EPERM);
6307
6308 mutex_lock(&cgroup_mutex);
6309 spin_lock_bh(&css_set_lock);
6310
6311 cset = task_css_set(current);
6312 get_css_set(cset);
6313
6314 spin_unlock_bh(&css_set_lock);
6315 mutex_unlock(&cgroup_mutex);
6316
6317 new_ns = alloc_cgroup_ns();
6318 if (IS_ERR(new_ns)) {
6319 put_css_set(cset);
6320 return new_ns;
6321 }
6322
6323 new_ns->user_ns = get_user_ns(user_ns);
6324 new_ns->root_cset = cset;
6325
6326 return new_ns;
6327 }
6328
6329 static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6330 {
6331 return container_of(ns, struct cgroup_namespace, ns);
6332 }
6333
6334 static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6335 {
6336 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6337
6338 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6339 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6340 return -EPERM;
6341
6342 /* Don't need to do anything if we are attaching to our own cgroupns. */
6343 if (cgroup_ns == nsproxy->cgroup_ns)
6344 return 0;
6345
6346 get_cgroup_ns(cgroup_ns);
6347 put_cgroup_ns(nsproxy->cgroup_ns);
6348 nsproxy->cgroup_ns = cgroup_ns;
6349
6350 return 0;
6351 }
6352
6353 static struct ns_common *cgroupns_get(struct task_struct *task)
6354 {
6355 struct cgroup_namespace *ns = NULL;
6356 struct nsproxy *nsproxy;
6357
6358 task_lock(task);
6359 nsproxy = task->nsproxy;
6360 if (nsproxy) {
6361 ns = nsproxy->cgroup_ns;
6362 get_cgroup_ns(ns);
6363 }
6364 task_unlock(task);
6365
6366 return ns ? &ns->ns : NULL;
6367 }
6368
6369 static void cgroupns_put(struct ns_common *ns)
6370 {
6371 put_cgroup_ns(to_cg_ns(ns));
6372 }
6373
6374 const struct proc_ns_operations cgroupns_operations = {
6375 .name = "cgroup",
6376 .type = CLONE_NEWCGROUP,
6377 .get = cgroupns_get,
6378 .put = cgroupns_put,
6379 .install = cgroupns_install,
6380 };
6381
6382 static __init int cgroup_namespaces_init(void)
6383 {
6384 return 0;
6385 }
6386 subsys_initcall(cgroup_namespaces_init);
6387
6388 #ifdef CONFIG_CGROUP_DEBUG
6389 static struct cgroup_subsys_state *
6390 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6391 {
6392 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6393
6394 if (!css)
6395 return ERR_PTR(-ENOMEM);
6396
6397 return css;
6398 }
6399
6400 static void debug_css_free(struct cgroup_subsys_state *css)
6401 {
6402 kfree(css);
6403 }
6404
6405 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6406 struct cftype *cft)
6407 {
6408 return cgroup_task_count(css->cgroup);
6409 }
6410
6411 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6412 struct cftype *cft)
6413 {
6414 return (u64)(unsigned long)current->cgroups;
6415 }
6416
6417 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6418 struct cftype *cft)
6419 {
6420 u64 count;
6421
6422 rcu_read_lock();
6423 count = atomic_read(&task_css_set(current)->refcount);
6424 rcu_read_unlock();
6425 return count;
6426 }
6427
6428 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6429 {
6430 struct cgrp_cset_link *link;
6431 struct css_set *cset;
6432 char *name_buf;
6433
6434 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6435 if (!name_buf)
6436 return -ENOMEM;
6437
6438 spin_lock_bh(&css_set_lock);
6439 rcu_read_lock();
6440 cset = rcu_dereference(current->cgroups);
6441 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6442 struct cgroup *c = link->cgrp;
6443
6444 cgroup_name(c, name_buf, NAME_MAX + 1);
6445 seq_printf(seq, "Root %d group %s\n",
6446 c->root->hierarchy_id, name_buf);
6447 }
6448 rcu_read_unlock();
6449 spin_unlock_bh(&css_set_lock);
6450 kfree(name_buf);
6451 return 0;
6452 }
6453
6454 #define MAX_TASKS_SHOWN_PER_CSS 25
6455 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6456 {
6457 struct cgroup_subsys_state *css = seq_css(seq);
6458 struct cgrp_cset_link *link;
6459
6460 spin_lock_bh(&css_set_lock);
6461 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6462 struct css_set *cset = link->cset;
6463 struct task_struct *task;
6464 int count = 0;
6465
6466 seq_printf(seq, "css_set %p\n", cset);
6467
6468 list_for_each_entry(task, &cset->tasks, cg_list) {
6469 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6470 goto overflow;
6471 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6472 }
6473
6474 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6475 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6476 goto overflow;
6477 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6478 }
6479 continue;
6480 overflow:
6481 seq_puts(seq, " ...\n");
6482 }
6483 spin_unlock_bh(&css_set_lock);
6484 return 0;
6485 }
6486
6487 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6488 {
6489 return (!cgroup_is_populated(css->cgroup) &&
6490 !css_has_online_children(&css->cgroup->self));
6491 }
6492
6493 static struct cftype debug_files[] = {
6494 {
6495 .name = "taskcount",
6496 .read_u64 = debug_taskcount_read,
6497 },
6498
6499 {
6500 .name = "current_css_set",
6501 .read_u64 = current_css_set_read,
6502 },
6503
6504 {
6505 .name = "current_css_set_refcount",
6506 .read_u64 = current_css_set_refcount_read,
6507 },
6508
6509 {
6510 .name = "current_css_set_cg_links",
6511 .seq_show = current_css_set_cg_links_read,
6512 },
6513
6514 {
6515 .name = "cgroup_css_links",
6516 .seq_show = cgroup_css_links_read,
6517 },
6518
6519 {
6520 .name = "releasable",
6521 .read_u64 = releasable_read,
6522 },
6523
6524 { } /* terminate */
6525 };
6526
6527 struct cgroup_subsys debug_cgrp_subsys = {
6528 .css_alloc = debug_css_alloc,
6529 .css_free = debug_css_free,
6530 .legacy_cftypes = debug_files,
6531 };
6532 #endif /* CONFIG_CGROUP_DEBUG */