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