1 /*
2 * Generic process-grouping system.
3 *
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
6 *
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
10 *
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
15 *
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
18 *
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
23 *
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
27 */
28
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
31 #include "cgroup-internal.h"
32
33 #include <linux/bpf-cgroup.h>
34 #include <linux/cred.h>
35 #include <linux/errno.h>
36 #include <linux/init_task.h>
37 #include <linux/kernel.h>
38 #include <linux/magic.h>
39 #include <linux/mutex.h>
40 #include <linux/mount.h>
41 #include <linux/pagemap.h>
42 #include <linux/proc_fs.h>
43 #include <linux/rcupdate.h>
44 #include <linux/sched.h>
45 #include <linux/sched/task.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/hashtable.h>
51 #include <linux/idr.h>
52 #include <linux/kthread.h>
53 #include <linux/atomic.h>
54 #include <linux/cpuset.h>
55 #include <linux/proc_ns.h>
56 #include <linux/nsproxy.h>
57 #include <linux/file.h>
58 #include <linux/fs_parser.h>
59 #include <linux/sched/cputime.h>
60 #include <linux/psi.h>
61 #include <net/sock.h>
62
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/cgroup.h>
65
66 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
67 MAX_CFTYPE_NAME + 2)
68 /* let's not notify more than 100 times per second */
69 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
70
71 /*
72 * To avoid confusing the compiler (and generating warnings) with code
73 * that attempts to access what would be a 0-element array (i.e. sized
74 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this
75 * constant expression can be added.
76 */
77 #define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0)
78
79 /*
80 * cgroup_mutex is the master lock. Any modification to cgroup or its
81 * hierarchy must be performed while holding it.
82 *
83 * css_set_lock protects task->cgroups pointer, the list of css_set
84 * objects, and the chain of tasks off each css_set.
85 *
86 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
87 * cgroup.h can use them for lockdep annotations.
88 */
89 DEFINE_MUTEX(cgroup_mutex);
90 DEFINE_SPINLOCK(css_set_lock);
91
92 #ifdef CONFIG_PROVE_RCU
93 EXPORT_SYMBOL_GPL(cgroup_mutex);
94 EXPORT_SYMBOL_GPL(css_set_lock);
95 #endif
96
97 DEFINE_SPINLOCK(trace_cgroup_path_lock);
98 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
99 static bool cgroup_debug __read_mostly;
100
101 /*
102 * Protects cgroup_idr and css_idr so that IDs can be released without
103 * grabbing cgroup_mutex.
104 */
105 static DEFINE_SPINLOCK(cgroup_idr_lock);
106
107 /*
108 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
109 * against file removal/re-creation across css hiding.
110 */
111 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
112
113 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
114
115 #define cgroup_assert_mutex_or_rcu_locked() \
116 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
117 !lockdep_is_held(&cgroup_mutex), \
118 "cgroup_mutex or RCU read lock required");
119
120 /*
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
125 */
126 static struct workqueue_struct *cgroup_destroy_wq;
127
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 struct cgroup_subsys *cgroup_subsys[] = {
131 #include <linux/cgroup_subsys.h>
132 };
133 #undef SUBSYS
134
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name[] = {
138 #include <linux/cgroup_subsys.h>
139 };
140 #undef SUBSYS
141
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
143 #define SUBSYS(_x) \
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
149 #undef SUBSYS
150
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true *cgroup_subsys_enabled_key[] = {
153 #include <linux/cgroup_subsys.h>
154 };
155 #undef SUBSYS
156
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
159 #include <linux/cgroup_subsys.h>
160 };
161 #undef SUBSYS
162
163 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
164
165 /* the default hierarchy */
166 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
167 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
168
169 /*
170 * The default hierarchy always exists but is hidden until mounted for the
171 * first time. This is for backward compatibility.
172 */
173 static bool cgrp_dfl_visible;
174
175 /* some controllers are not supported in the default hierarchy */
176 static u16 cgrp_dfl_inhibit_ss_mask;
177
178 /* some controllers are implicitly enabled on the default hierarchy */
179 static u16 cgrp_dfl_implicit_ss_mask;
180
181 /* some controllers can be threaded on the default hierarchy */
182 static u16 cgrp_dfl_threaded_ss_mask;
183
184 /* The list of hierarchy roots */
185 LIST_HEAD(cgroup_roots);
186 static int cgroup_root_count;
187
188 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
189 static DEFINE_IDR(cgroup_hierarchy_idr);
190
191 /*
192 * Assign a monotonically increasing serial number to csses. It guarantees
193 * cgroups with bigger numbers are newer than those with smaller numbers.
194 * Also, as csses are always appended to the parent's ->children list, it
195 * guarantees that sibling csses are always sorted in the ascending serial
196 * number order on the list. Protected by cgroup_mutex.
197 */
198 static u64 css_serial_nr_next = 1;
199
200 /*
201 * These bitmasks identify subsystems with specific features to avoid
202 * having to do iterative checks repeatedly.
203 */
204 static u16 have_fork_callback __read_mostly;
205 static u16 have_exit_callback __read_mostly;
206 static u16 have_release_callback __read_mostly;
207 static u16 have_canfork_callback __read_mostly;
208
209 /* cgroup namespace for init task */
210 struct cgroup_namespace init_cgroup_ns = {
211 .ns.count = REFCOUNT_INIT(2),
212 .user_ns = &init_user_ns,
213 .ns.ops = &cgroupns_operations,
214 .ns.inum = PROC_CGROUP_INIT_INO,
215 .root_cset = &init_css_set,
216 };
217
218 static struct file_system_type cgroup2_fs_type;
219 static struct cftype cgroup_base_files[];
220
221 /* cgroup optional features */
222 enum cgroup_opt_features {
223 #ifdef CONFIG_PSI
224 OPT_FEATURE_PRESSURE,
225 #endif
226 OPT_FEATURE_COUNT
227 };
228
229 static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
230 #ifdef CONFIG_PSI
231 "pressure",
232 #endif
233 };
234
235 static u16 cgroup_feature_disable_mask __read_mostly;
236
237 static int cgroup_apply_control(struct cgroup *cgrp);
238 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
239 static void css_task_iter_skip(struct css_task_iter *it,
240 struct task_struct *task);
241 static int cgroup_destroy_locked(struct cgroup *cgrp);
242 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
243 struct cgroup_subsys *ss);
244 static void css_release(struct percpu_ref *ref);
245 static void kill_css(struct cgroup_subsys_state *css);
246 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
247 struct cgroup *cgrp, struct cftype cfts[],
248 bool is_add);
249
250 /**
251 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
252 * @ssid: subsys ID of interest
253 *
254 * cgroup_subsys_enabled() can only be used with literal subsys names which
255 * is fine for individual subsystems but unsuitable for cgroup core. This
256 * is slower static_key_enabled() based test indexed by @ssid.
257 */
cgroup_ssid_enabled(int ssid)258 bool cgroup_ssid_enabled(int ssid)
259 {
260 if (!CGROUP_HAS_SUBSYS_CONFIG)
261 return false;
262
263 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
264 }
265
266 /**
267 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
268 * @cgrp: the cgroup of interest
269 *
270 * The default hierarchy is the v2 interface of cgroup and this function
271 * can be used to test whether a cgroup is on the default hierarchy for
272 * cases where a subsystem should behave differently depending on the
273 * interface version.
274 *
275 * List of changed behaviors:
276 *
277 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
278 * and "name" are disallowed.
279 *
280 * - When mounting an existing superblock, mount options should match.
281 *
282 * - Remount is disallowed.
283 *
284 * - rename(2) is disallowed.
285 *
286 * - "tasks" is removed. Everything should be at process granularity. Use
287 * "cgroup.procs" instead.
288 *
289 * - "cgroup.procs" is not sorted. pids will be unique unless they got
290 * recycled in-between reads.
291 *
292 * - "release_agent" and "notify_on_release" are removed. Replacement
293 * notification mechanism will be implemented.
294 *
295 * - "cgroup.clone_children" is removed.
296 *
297 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
298 * and its descendants contain no task; otherwise, 1. The file also
299 * generates kernfs notification which can be monitored through poll and
300 * [di]notify when the value of the file changes.
301 *
302 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
303 * take masks of ancestors with non-empty cpus/mems, instead of being
304 * moved to an ancestor.
305 *
306 * - cpuset: a task can be moved into an empty cpuset, and again it takes
307 * masks of ancestors.
308 *
309 * - blkcg: blk-throttle becomes properly hierarchical.
310 *
311 * - debug: disallowed on the default hierarchy.
312 */
cgroup_on_dfl(const struct cgroup * cgrp)313 bool cgroup_on_dfl(const struct cgroup *cgrp)
314 {
315 return cgrp->root == &cgrp_dfl_root;
316 }
317
318 /* IDR wrappers which synchronize using cgroup_idr_lock */
cgroup_idr_alloc(struct idr * idr,void * ptr,int start,int end,gfp_t gfp_mask)319 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
320 gfp_t gfp_mask)
321 {
322 int ret;
323
324 idr_preload(gfp_mask);
325 spin_lock_bh(&cgroup_idr_lock);
326 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
327 spin_unlock_bh(&cgroup_idr_lock);
328 idr_preload_end();
329 return ret;
330 }
331
cgroup_idr_replace(struct idr * idr,void * ptr,int id)332 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
333 {
334 void *ret;
335
336 spin_lock_bh(&cgroup_idr_lock);
337 ret = idr_replace(idr, ptr, id);
338 spin_unlock_bh(&cgroup_idr_lock);
339 return ret;
340 }
341
cgroup_idr_remove(struct idr * idr,int id)342 static void cgroup_idr_remove(struct idr *idr, int id)
343 {
344 spin_lock_bh(&cgroup_idr_lock);
345 idr_remove(idr, id);
346 spin_unlock_bh(&cgroup_idr_lock);
347 }
348
cgroup_has_tasks(struct cgroup * cgrp)349 static bool cgroup_has_tasks(struct cgroup *cgrp)
350 {
351 return cgrp->nr_populated_csets;
352 }
353
cgroup_is_threaded(struct cgroup * cgrp)354 bool cgroup_is_threaded(struct cgroup *cgrp)
355 {
356 return cgrp->dom_cgrp != cgrp;
357 }
358
359 /* can @cgrp host both domain and threaded children? */
cgroup_is_mixable(struct cgroup * cgrp)360 static bool cgroup_is_mixable(struct cgroup *cgrp)
361 {
362 /*
363 * Root isn't under domain level resource control exempting it from
364 * the no-internal-process constraint, so it can serve as a thread
365 * root and a parent of resource domains at the same time.
366 */
367 return !cgroup_parent(cgrp);
368 }
369
370 /* can @cgrp become a thread root? Should always be true for a thread root */
cgroup_can_be_thread_root(struct cgroup * cgrp)371 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
372 {
373 /* mixables don't care */
374 if (cgroup_is_mixable(cgrp))
375 return true;
376
377 /* domain roots can't be nested under threaded */
378 if (cgroup_is_threaded(cgrp))
379 return false;
380
381 /* can only have either domain or threaded children */
382 if (cgrp->nr_populated_domain_children)
383 return false;
384
385 /* and no domain controllers can be enabled */
386 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
387 return false;
388
389 return true;
390 }
391
392 /* is @cgrp root of a threaded subtree? */
cgroup_is_thread_root(struct cgroup * cgrp)393 bool cgroup_is_thread_root(struct cgroup *cgrp)
394 {
395 /* thread root should be a domain */
396 if (cgroup_is_threaded(cgrp))
397 return false;
398
399 /* a domain w/ threaded children is a thread root */
400 if (cgrp->nr_threaded_children)
401 return true;
402
403 /*
404 * A domain which has tasks and explicit threaded controllers
405 * enabled is a thread root.
406 */
407 if (cgroup_has_tasks(cgrp) &&
408 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
409 return true;
410
411 return false;
412 }
413
414 /* a domain which isn't connected to the root w/o brekage can't be used */
cgroup_is_valid_domain(struct cgroup * cgrp)415 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
416 {
417 /* the cgroup itself can be a thread root */
418 if (cgroup_is_threaded(cgrp))
419 return false;
420
421 /* but the ancestors can't be unless mixable */
422 while ((cgrp = cgroup_parent(cgrp))) {
423 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
424 return false;
425 if (cgroup_is_threaded(cgrp))
426 return false;
427 }
428
429 return true;
430 }
431
432 /* subsystems visibly enabled on a cgroup */
cgroup_control(struct cgroup * cgrp)433 static u16 cgroup_control(struct cgroup *cgrp)
434 {
435 struct cgroup *parent = cgroup_parent(cgrp);
436 u16 root_ss_mask = cgrp->root->subsys_mask;
437
438 if (parent) {
439 u16 ss_mask = parent->subtree_control;
440
441 /* threaded cgroups can only have threaded controllers */
442 if (cgroup_is_threaded(cgrp))
443 ss_mask &= cgrp_dfl_threaded_ss_mask;
444 return ss_mask;
445 }
446
447 if (cgroup_on_dfl(cgrp))
448 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
449 cgrp_dfl_implicit_ss_mask);
450 return root_ss_mask;
451 }
452
453 /* subsystems enabled on a cgroup */
cgroup_ss_mask(struct cgroup * cgrp)454 static u16 cgroup_ss_mask(struct cgroup *cgrp)
455 {
456 struct cgroup *parent = cgroup_parent(cgrp);
457
458 if (parent) {
459 u16 ss_mask = parent->subtree_ss_mask;
460
461 /* threaded cgroups can only have threaded controllers */
462 if (cgroup_is_threaded(cgrp))
463 ss_mask &= cgrp_dfl_threaded_ss_mask;
464 return ss_mask;
465 }
466
467 return cgrp->root->subsys_mask;
468 }
469
470 /**
471 * cgroup_css - obtain a cgroup's css for the specified subsystem
472 * @cgrp: the cgroup of interest
473 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
474 *
475 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
476 * function must be called either under cgroup_mutex or rcu_read_lock() and
477 * the caller is responsible for pinning the returned css if it wants to
478 * keep accessing it outside the said locks. This function may return
479 * %NULL if @cgrp doesn't have @subsys_id enabled.
480 */
cgroup_css(struct cgroup * cgrp,struct cgroup_subsys * ss)481 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
482 struct cgroup_subsys *ss)
483 {
484 if (CGROUP_HAS_SUBSYS_CONFIG && ss)
485 return rcu_dereference_check(cgrp->subsys[ss->id],
486 lockdep_is_held(&cgroup_mutex));
487 else
488 return &cgrp->self;
489 }
490
491 /**
492 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
493 * @cgrp: the cgroup of interest
494 * @ss: the subsystem of interest
495 *
496 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
497 * or is offline, %NULL is returned.
498 */
cgroup_tryget_css(struct cgroup * cgrp,struct cgroup_subsys * ss)499 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
500 struct cgroup_subsys *ss)
501 {
502 struct cgroup_subsys_state *css;
503
504 rcu_read_lock();
505 css = cgroup_css(cgrp, ss);
506 if (css && !css_tryget_online(css))
507 css = NULL;
508 rcu_read_unlock();
509
510 return css;
511 }
512
513 /**
514 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
515 * @cgrp: the cgroup of interest
516 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
517 *
518 * Similar to cgroup_css() but returns the effective css, which is defined
519 * as the matching css of the nearest ancestor including self which has @ss
520 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
521 * function is guaranteed to return non-NULL css.
522 */
cgroup_e_css_by_mask(struct cgroup * cgrp,struct cgroup_subsys * ss)523 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
524 struct cgroup_subsys *ss)
525 {
526 lockdep_assert_held(&cgroup_mutex);
527
528 if (!ss)
529 return &cgrp->self;
530
531 /*
532 * This function is used while updating css associations and thus
533 * can't test the csses directly. Test ss_mask.
534 */
535 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
536 cgrp = cgroup_parent(cgrp);
537 if (!cgrp)
538 return NULL;
539 }
540
541 return cgroup_css(cgrp, ss);
542 }
543
544 /**
545 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
546 * @cgrp: the cgroup of interest
547 * @ss: the subsystem of interest
548 *
549 * Find and get the effective css of @cgrp for @ss. The effective css is
550 * defined as the matching css of the nearest ancestor including self which
551 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
552 * the root css is returned, so this function always returns a valid css.
553 *
554 * The returned css is not guaranteed to be online, and therefore it is the
555 * callers responsibility to try get a reference for it.
556 */
cgroup_e_css(struct cgroup * cgrp,struct cgroup_subsys * ss)557 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
558 struct cgroup_subsys *ss)
559 {
560 struct cgroup_subsys_state *css;
561
562 if (!CGROUP_HAS_SUBSYS_CONFIG)
563 return NULL;
564
565 do {
566 css = cgroup_css(cgrp, ss);
567
568 if (css)
569 return css;
570 cgrp = cgroup_parent(cgrp);
571 } while (cgrp);
572
573 return init_css_set.subsys[ss->id];
574 }
575
576 /**
577 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
578 * @cgrp: the cgroup of interest
579 * @ss: the subsystem of interest
580 *
581 * Find and get the effective css of @cgrp for @ss. The effective css is
582 * defined as the matching css of the nearest ancestor including self which
583 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
584 * the root css is returned, so this function always returns a valid css.
585 * The returned css must be put using css_put().
586 */
cgroup_get_e_css(struct cgroup * cgrp,struct cgroup_subsys * ss)587 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
588 struct cgroup_subsys *ss)
589 {
590 struct cgroup_subsys_state *css;
591
592 if (!CGROUP_HAS_SUBSYS_CONFIG)
593 return NULL;
594
595 rcu_read_lock();
596
597 do {
598 css = cgroup_css(cgrp, ss);
599
600 if (css && css_tryget_online(css))
601 goto out_unlock;
602 cgrp = cgroup_parent(cgrp);
603 } while (cgrp);
604
605 css = init_css_set.subsys[ss->id];
606 css_get(css);
607 out_unlock:
608 rcu_read_unlock();
609 return css;
610 }
611 EXPORT_SYMBOL_GPL(cgroup_get_e_css);
612
cgroup_get_live(struct cgroup * cgrp)613 static void cgroup_get_live(struct cgroup *cgrp)
614 {
615 WARN_ON_ONCE(cgroup_is_dead(cgrp));
616 css_get(&cgrp->self);
617 }
618
619 /**
620 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
621 * is responsible for taking the css_set_lock.
622 * @cgrp: the cgroup in question
623 */
__cgroup_task_count(const struct cgroup * cgrp)624 int __cgroup_task_count(const struct cgroup *cgrp)
625 {
626 int count = 0;
627 struct cgrp_cset_link *link;
628
629 lockdep_assert_held(&css_set_lock);
630
631 list_for_each_entry(link, &cgrp->cset_links, cset_link)
632 count += link->cset->nr_tasks;
633
634 return count;
635 }
636
637 /**
638 * cgroup_task_count - count the number of tasks in a cgroup.
639 * @cgrp: the cgroup in question
640 */
cgroup_task_count(const struct cgroup * cgrp)641 int cgroup_task_count(const struct cgroup *cgrp)
642 {
643 int count;
644
645 spin_lock_irq(&css_set_lock);
646 count = __cgroup_task_count(cgrp);
647 spin_unlock_irq(&css_set_lock);
648
649 return count;
650 }
651
of_css(struct kernfs_open_file * of)652 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
653 {
654 struct cgroup *cgrp = of->kn->parent->priv;
655 struct cftype *cft = of_cft(of);
656
657 /*
658 * This is open and unprotected implementation of cgroup_css().
659 * seq_css() is only called from a kernfs file operation which has
660 * an active reference on the file. Because all the subsystem
661 * files are drained before a css is disassociated with a cgroup,
662 * the matching css from the cgroup's subsys table is guaranteed to
663 * be and stay valid until the enclosing operation is complete.
664 */
665 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
666 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
667 else
668 return &cgrp->self;
669 }
670 EXPORT_SYMBOL_GPL(of_css);
671
672 /**
673 * for_each_css - iterate all css's of a cgroup
674 * @css: the iteration cursor
675 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
676 * @cgrp: the target cgroup to iterate css's of
677 *
678 * Should be called under cgroup_[tree_]mutex.
679 */
680 #define for_each_css(css, ssid, cgrp) \
681 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
682 if (!((css) = rcu_dereference_check( \
683 (cgrp)->subsys[(ssid)], \
684 lockdep_is_held(&cgroup_mutex)))) { } \
685 else
686
687 /**
688 * for_each_e_css - iterate all effective css's of a cgroup
689 * @css: the iteration cursor
690 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
691 * @cgrp: the target cgroup to iterate css's of
692 *
693 * Should be called under cgroup_[tree_]mutex.
694 */
695 #define for_each_e_css(css, ssid, cgrp) \
696 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
697 if (!((css) = cgroup_e_css_by_mask(cgrp, \
698 cgroup_subsys[(ssid)]))) \
699 ; \
700 else
701
702 /**
703 * do_each_subsys_mask - filter for_each_subsys with a bitmask
704 * @ss: the iteration cursor
705 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
706 * @ss_mask: the bitmask
707 *
708 * The block will only run for cases where the ssid-th bit (1 << ssid) of
709 * @ss_mask is set.
710 */
711 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
712 unsigned long __ss_mask = (ss_mask); \
713 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
714 (ssid) = 0; \
715 break; \
716 } \
717 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
718 (ss) = cgroup_subsys[ssid]; \
719 {
720
721 #define while_each_subsys_mask() \
722 } \
723 } \
724 } while (false)
725
726 /* iterate over child cgrps, lock should be held throughout iteration */
727 #define cgroup_for_each_live_child(child, cgrp) \
728 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
729 if (({ lockdep_assert_held(&cgroup_mutex); \
730 cgroup_is_dead(child); })) \
731 ; \
732 else
733
734 /* walk live descendants in pre order */
735 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
736 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
737 if (({ lockdep_assert_held(&cgroup_mutex); \
738 (dsct) = (d_css)->cgroup; \
739 cgroup_is_dead(dsct); })) \
740 ; \
741 else
742
743 /* walk live descendants in postorder */
744 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
745 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
746 if (({ lockdep_assert_held(&cgroup_mutex); \
747 (dsct) = (d_css)->cgroup; \
748 cgroup_is_dead(dsct); })) \
749 ; \
750 else
751
752 /*
753 * The default css_set - used by init and its children prior to any
754 * hierarchies being mounted. It contains a pointer to the root state
755 * for each subsystem. Also used to anchor the list of css_sets. Not
756 * reference-counted, to improve performance when child cgroups
757 * haven't been created.
758 */
759 struct css_set init_css_set = {
760 .refcount = REFCOUNT_INIT(1),
761 .dom_cset = &init_css_set,
762 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
763 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
764 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
765 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
766 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
767 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
768 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
769 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
770 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
771
772 /*
773 * The following field is re-initialized when this cset gets linked
774 * in cgroup_init(). However, let's initialize the field
775 * statically too so that the default cgroup can be accessed safely
776 * early during boot.
777 */
778 .dfl_cgrp = &cgrp_dfl_root.cgrp,
779 };
780
781 static int css_set_count = 1; /* 1 for init_css_set */
782
css_set_threaded(struct css_set * cset)783 static bool css_set_threaded(struct css_set *cset)
784 {
785 return cset->dom_cset != cset;
786 }
787
788 /**
789 * css_set_populated - does a css_set contain any tasks?
790 * @cset: target css_set
791 *
792 * css_set_populated() should be the same as !!cset->nr_tasks at steady
793 * state. However, css_set_populated() can be called while a task is being
794 * added to or removed from the linked list before the nr_tasks is
795 * properly updated. Hence, we can't just look at ->nr_tasks here.
796 */
css_set_populated(struct css_set * cset)797 static bool css_set_populated(struct css_set *cset)
798 {
799 lockdep_assert_held(&css_set_lock);
800
801 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
802 }
803
804 /**
805 * cgroup_update_populated - update the populated count of a cgroup
806 * @cgrp: the target cgroup
807 * @populated: inc or dec populated count
808 *
809 * One of the css_sets associated with @cgrp is either getting its first
810 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
811 * count is propagated towards root so that a given cgroup's
812 * nr_populated_children is zero iff none of its descendants contain any
813 * tasks.
814 *
815 * @cgrp's interface file "cgroup.populated" is zero if both
816 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
817 * 1 otherwise. When the sum changes from or to zero, userland is notified
818 * that the content of the interface file has changed. This can be used to
819 * detect when @cgrp and its descendants become populated or empty.
820 */
cgroup_update_populated(struct cgroup * cgrp,bool populated)821 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
822 {
823 struct cgroup *child = NULL;
824 int adj = populated ? 1 : -1;
825
826 lockdep_assert_held(&css_set_lock);
827
828 do {
829 bool was_populated = cgroup_is_populated(cgrp);
830
831 if (!child) {
832 cgrp->nr_populated_csets += adj;
833 } else {
834 if (cgroup_is_threaded(child))
835 cgrp->nr_populated_threaded_children += adj;
836 else
837 cgrp->nr_populated_domain_children += adj;
838 }
839
840 if (was_populated == cgroup_is_populated(cgrp))
841 break;
842
843 cgroup1_check_for_release(cgrp);
844 TRACE_CGROUP_PATH(notify_populated, cgrp,
845 cgroup_is_populated(cgrp));
846 cgroup_file_notify(&cgrp->events_file);
847
848 child = cgrp;
849 cgrp = cgroup_parent(cgrp);
850 } while (cgrp);
851 }
852
853 /**
854 * css_set_update_populated - update populated state of a css_set
855 * @cset: target css_set
856 * @populated: whether @cset is populated or depopulated
857 *
858 * @cset is either getting the first task or losing the last. Update the
859 * populated counters of all associated cgroups accordingly.
860 */
css_set_update_populated(struct css_set * cset,bool populated)861 static void css_set_update_populated(struct css_set *cset, bool populated)
862 {
863 struct cgrp_cset_link *link;
864
865 lockdep_assert_held(&css_set_lock);
866
867 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
868 cgroup_update_populated(link->cgrp, populated);
869 }
870
871 /*
872 * @task is leaving, advance task iterators which are pointing to it so
873 * that they can resume at the next position. Advancing an iterator might
874 * remove it from the list, use safe walk. See css_task_iter_skip() for
875 * details.
876 */
css_set_skip_task_iters(struct css_set * cset,struct task_struct * task)877 static void css_set_skip_task_iters(struct css_set *cset,
878 struct task_struct *task)
879 {
880 struct css_task_iter *it, *pos;
881
882 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
883 css_task_iter_skip(it, task);
884 }
885
886 /**
887 * css_set_move_task - move a task from one css_set to another
888 * @task: task being moved
889 * @from_cset: css_set @task currently belongs to (may be NULL)
890 * @to_cset: new css_set @task is being moved to (may be NULL)
891 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
892 *
893 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
894 * css_set, @from_cset can be NULL. If @task is being disassociated
895 * instead of moved, @to_cset can be NULL.
896 *
897 * This function automatically handles populated counter updates and
898 * css_task_iter adjustments but the caller is responsible for managing
899 * @from_cset and @to_cset's reference counts.
900 */
css_set_move_task(struct task_struct * task,struct css_set * from_cset,struct css_set * to_cset,bool use_mg_tasks)901 static void css_set_move_task(struct task_struct *task,
902 struct css_set *from_cset, struct css_set *to_cset,
903 bool use_mg_tasks)
904 {
905 lockdep_assert_held(&css_set_lock);
906
907 if (to_cset && !css_set_populated(to_cset))
908 css_set_update_populated(to_cset, true);
909
910 if (from_cset) {
911 WARN_ON_ONCE(list_empty(&task->cg_list));
912
913 css_set_skip_task_iters(from_cset, task);
914 list_del_init(&task->cg_list);
915 if (!css_set_populated(from_cset))
916 css_set_update_populated(from_cset, false);
917 } else {
918 WARN_ON_ONCE(!list_empty(&task->cg_list));
919 }
920
921 if (to_cset) {
922 /*
923 * We are synchronized through cgroup_threadgroup_rwsem
924 * against PF_EXITING setting such that we can't race
925 * against cgroup_exit()/cgroup_free() dropping the css_set.
926 */
927 WARN_ON_ONCE(task->flags & PF_EXITING);
928
929 cgroup_move_task(task, to_cset);
930 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
931 &to_cset->tasks);
932 }
933 }
934
935 /*
936 * hash table for cgroup groups. This improves the performance to find
937 * an existing css_set. This hash doesn't (currently) take into
938 * account cgroups in empty hierarchies.
939 */
940 #define CSS_SET_HASH_BITS 7
941 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
942
css_set_hash(struct cgroup_subsys_state * css[])943 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
944 {
945 unsigned long key = 0UL;
946 struct cgroup_subsys *ss;
947 int i;
948
949 for_each_subsys(ss, i)
950 key += (unsigned long)css[i];
951 key = (key >> 16) ^ key;
952
953 return key;
954 }
955
put_css_set_locked(struct css_set * cset)956 void put_css_set_locked(struct css_set *cset)
957 {
958 struct cgrp_cset_link *link, *tmp_link;
959 struct cgroup_subsys *ss;
960 int ssid;
961
962 lockdep_assert_held(&css_set_lock);
963
964 if (!refcount_dec_and_test(&cset->refcount))
965 return;
966
967 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
968
969 /* This css_set is dead. Unlink it and release cgroup and css refs */
970 for_each_subsys(ss, ssid) {
971 list_del(&cset->e_cset_node[ssid]);
972 css_put(cset->subsys[ssid]);
973 }
974 hash_del(&cset->hlist);
975 css_set_count--;
976
977 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
978 list_del(&link->cset_link);
979 list_del(&link->cgrp_link);
980 if (cgroup_parent(link->cgrp))
981 cgroup_put(link->cgrp);
982 kfree(link);
983 }
984
985 if (css_set_threaded(cset)) {
986 list_del(&cset->threaded_csets_node);
987 put_css_set_locked(cset->dom_cset);
988 }
989
990 kfree_rcu(cset, rcu_head);
991 }
992
993 /**
994 * compare_css_sets - helper function for find_existing_css_set().
995 * @cset: candidate css_set being tested
996 * @old_cset: existing css_set for a task
997 * @new_cgrp: cgroup that's being entered by the task
998 * @template: desired set of css pointers in css_set (pre-calculated)
999 *
1000 * Returns true if "cset" matches "old_cset" except for the hierarchy
1001 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
1002 */
compare_css_sets(struct css_set * cset,struct css_set * old_cset,struct cgroup * new_cgrp,struct cgroup_subsys_state * template[])1003 static bool compare_css_sets(struct css_set *cset,
1004 struct css_set *old_cset,
1005 struct cgroup *new_cgrp,
1006 struct cgroup_subsys_state *template[])
1007 {
1008 struct cgroup *new_dfl_cgrp;
1009 struct list_head *l1, *l2;
1010
1011 /*
1012 * On the default hierarchy, there can be csets which are
1013 * associated with the same set of cgroups but different csses.
1014 * Let's first ensure that csses match.
1015 */
1016 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1017 return false;
1018
1019
1020 /* @cset's domain should match the default cgroup's */
1021 if (cgroup_on_dfl(new_cgrp))
1022 new_dfl_cgrp = new_cgrp;
1023 else
1024 new_dfl_cgrp = old_cset->dfl_cgrp;
1025
1026 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1027 return false;
1028
1029 /*
1030 * Compare cgroup pointers in order to distinguish between
1031 * different cgroups in hierarchies. As different cgroups may
1032 * share the same effective css, this comparison is always
1033 * necessary.
1034 */
1035 l1 = &cset->cgrp_links;
1036 l2 = &old_cset->cgrp_links;
1037 while (1) {
1038 struct cgrp_cset_link *link1, *link2;
1039 struct cgroup *cgrp1, *cgrp2;
1040
1041 l1 = l1->next;
1042 l2 = l2->next;
1043 /* See if we reached the end - both lists are equal length. */
1044 if (l1 == &cset->cgrp_links) {
1045 BUG_ON(l2 != &old_cset->cgrp_links);
1046 break;
1047 } else {
1048 BUG_ON(l2 == &old_cset->cgrp_links);
1049 }
1050 /* Locate the cgroups associated with these links. */
1051 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1052 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1053 cgrp1 = link1->cgrp;
1054 cgrp2 = link2->cgrp;
1055 /* Hierarchies should be linked in the same order. */
1056 BUG_ON(cgrp1->root != cgrp2->root);
1057
1058 /*
1059 * If this hierarchy is the hierarchy of the cgroup
1060 * that's changing, then we need to check that this
1061 * css_set points to the new cgroup; if it's any other
1062 * hierarchy, then this css_set should point to the
1063 * same cgroup as the old css_set.
1064 */
1065 if (cgrp1->root == new_cgrp->root) {
1066 if (cgrp1 != new_cgrp)
1067 return false;
1068 } else {
1069 if (cgrp1 != cgrp2)
1070 return false;
1071 }
1072 }
1073 return true;
1074 }
1075
1076 /**
1077 * find_existing_css_set - init css array and find the matching css_set
1078 * @old_cset: the css_set that we're using before the cgroup transition
1079 * @cgrp: the cgroup that we're moving into
1080 * @template: out param for the new set of csses, should be clear on entry
1081 */
find_existing_css_set(struct css_set * old_cset,struct cgroup * cgrp,struct cgroup_subsys_state * template[])1082 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1083 struct cgroup *cgrp,
1084 struct cgroup_subsys_state *template[])
1085 {
1086 struct cgroup_root *root = cgrp->root;
1087 struct cgroup_subsys *ss;
1088 struct css_set *cset;
1089 unsigned long key;
1090 int i;
1091
1092 /*
1093 * Build the set of subsystem state objects that we want to see in the
1094 * new css_set. While subsystems can change globally, the entries here
1095 * won't change, so no need for locking.
1096 */
1097 for_each_subsys(ss, i) {
1098 if (root->subsys_mask & (1UL << i)) {
1099 /*
1100 * @ss is in this hierarchy, so we want the
1101 * effective css from @cgrp.
1102 */
1103 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1104 } else {
1105 /*
1106 * @ss is not in this hierarchy, so we don't want
1107 * to change the css.
1108 */
1109 template[i] = old_cset->subsys[i];
1110 }
1111 }
1112
1113 key = css_set_hash(template);
1114 hash_for_each_possible(css_set_table, cset, hlist, key) {
1115 if (!compare_css_sets(cset, old_cset, cgrp, template))
1116 continue;
1117
1118 /* This css_set matches what we need */
1119 return cset;
1120 }
1121
1122 /* No existing cgroup group matched */
1123 return NULL;
1124 }
1125
free_cgrp_cset_links(struct list_head * links_to_free)1126 static void free_cgrp_cset_links(struct list_head *links_to_free)
1127 {
1128 struct cgrp_cset_link *link, *tmp_link;
1129
1130 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1131 list_del(&link->cset_link);
1132 kfree(link);
1133 }
1134 }
1135
1136 /**
1137 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1138 * @count: the number of links to allocate
1139 * @tmp_links: list_head the allocated links are put on
1140 *
1141 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1142 * through ->cset_link. Returns 0 on success or -errno.
1143 */
allocate_cgrp_cset_links(int count,struct list_head * tmp_links)1144 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1145 {
1146 struct cgrp_cset_link *link;
1147 int i;
1148
1149 INIT_LIST_HEAD(tmp_links);
1150
1151 for (i = 0; i < count; i++) {
1152 link = kzalloc(sizeof(*link), GFP_KERNEL);
1153 if (!link) {
1154 free_cgrp_cset_links(tmp_links);
1155 return -ENOMEM;
1156 }
1157 list_add(&link->cset_link, tmp_links);
1158 }
1159 return 0;
1160 }
1161
1162 /**
1163 * link_css_set - a helper function to link a css_set to a cgroup
1164 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1165 * @cset: the css_set to be linked
1166 * @cgrp: the destination cgroup
1167 */
link_css_set(struct list_head * tmp_links,struct css_set * cset,struct cgroup * cgrp)1168 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1169 struct cgroup *cgrp)
1170 {
1171 struct cgrp_cset_link *link;
1172
1173 BUG_ON(list_empty(tmp_links));
1174
1175 if (cgroup_on_dfl(cgrp))
1176 cset->dfl_cgrp = cgrp;
1177
1178 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1179 link->cset = cset;
1180 link->cgrp = cgrp;
1181
1182 /*
1183 * Always add links to the tail of the lists so that the lists are
1184 * in chronological order.
1185 */
1186 list_move_tail(&link->cset_link, &cgrp->cset_links);
1187 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1188
1189 if (cgroup_parent(cgrp))
1190 cgroup_get_live(cgrp);
1191 }
1192
1193 /**
1194 * find_css_set - return a new css_set with one cgroup updated
1195 * @old_cset: the baseline css_set
1196 * @cgrp: the cgroup to be updated
1197 *
1198 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1199 * substituted into the appropriate hierarchy.
1200 */
find_css_set(struct css_set * old_cset,struct cgroup * cgrp)1201 static struct css_set *find_css_set(struct css_set *old_cset,
1202 struct cgroup *cgrp)
1203 {
1204 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1205 struct css_set *cset;
1206 struct list_head tmp_links;
1207 struct cgrp_cset_link *link;
1208 struct cgroup_subsys *ss;
1209 unsigned long key;
1210 int ssid;
1211
1212 lockdep_assert_held(&cgroup_mutex);
1213
1214 /* First see if we already have a cgroup group that matches
1215 * the desired set */
1216 spin_lock_irq(&css_set_lock);
1217 cset = find_existing_css_set(old_cset, cgrp, template);
1218 if (cset)
1219 get_css_set(cset);
1220 spin_unlock_irq(&css_set_lock);
1221
1222 if (cset)
1223 return cset;
1224
1225 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1226 if (!cset)
1227 return NULL;
1228
1229 /* Allocate all the cgrp_cset_link objects that we'll need */
1230 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1231 kfree(cset);
1232 return NULL;
1233 }
1234
1235 refcount_set(&cset->refcount, 1);
1236 cset->dom_cset = cset;
1237 INIT_LIST_HEAD(&cset->tasks);
1238 INIT_LIST_HEAD(&cset->mg_tasks);
1239 INIT_LIST_HEAD(&cset->dying_tasks);
1240 INIT_LIST_HEAD(&cset->task_iters);
1241 INIT_LIST_HEAD(&cset->threaded_csets);
1242 INIT_HLIST_NODE(&cset->hlist);
1243 INIT_LIST_HEAD(&cset->cgrp_links);
1244 INIT_LIST_HEAD(&cset->mg_src_preload_node);
1245 INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1246 INIT_LIST_HEAD(&cset->mg_node);
1247
1248 /* Copy the set of subsystem state objects generated in
1249 * find_existing_css_set() */
1250 memcpy(cset->subsys, template, sizeof(cset->subsys));
1251
1252 spin_lock_irq(&css_set_lock);
1253 /* Add reference counts and links from the new css_set. */
1254 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1255 struct cgroup *c = link->cgrp;
1256
1257 if (c->root == cgrp->root)
1258 c = cgrp;
1259 link_css_set(&tmp_links, cset, c);
1260 }
1261
1262 BUG_ON(!list_empty(&tmp_links));
1263
1264 css_set_count++;
1265
1266 /* Add @cset to the hash table */
1267 key = css_set_hash(cset->subsys);
1268 hash_add(css_set_table, &cset->hlist, key);
1269
1270 for_each_subsys(ss, ssid) {
1271 struct cgroup_subsys_state *css = cset->subsys[ssid];
1272
1273 list_add_tail(&cset->e_cset_node[ssid],
1274 &css->cgroup->e_csets[ssid]);
1275 css_get(css);
1276 }
1277
1278 spin_unlock_irq(&css_set_lock);
1279
1280 /*
1281 * If @cset should be threaded, look up the matching dom_cset and
1282 * link them up. We first fully initialize @cset then look for the
1283 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1284 * to stay empty until we return.
1285 */
1286 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1287 struct css_set *dcset;
1288
1289 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1290 if (!dcset) {
1291 put_css_set(cset);
1292 return NULL;
1293 }
1294
1295 spin_lock_irq(&css_set_lock);
1296 cset->dom_cset = dcset;
1297 list_add_tail(&cset->threaded_csets_node,
1298 &dcset->threaded_csets);
1299 spin_unlock_irq(&css_set_lock);
1300 }
1301
1302 return cset;
1303 }
1304
cgroup_root_from_kf(struct kernfs_root * kf_root)1305 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1306 {
1307 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1308
1309 return root_cgrp->root;
1310 }
1311
cgroup_init_root_id(struct cgroup_root * root)1312 static int cgroup_init_root_id(struct cgroup_root *root)
1313 {
1314 int id;
1315
1316 lockdep_assert_held(&cgroup_mutex);
1317
1318 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1319 if (id < 0)
1320 return id;
1321
1322 root->hierarchy_id = id;
1323 return 0;
1324 }
1325
cgroup_exit_root_id(struct cgroup_root * root)1326 static void cgroup_exit_root_id(struct cgroup_root *root)
1327 {
1328 lockdep_assert_held(&cgroup_mutex);
1329
1330 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1331 }
1332
cgroup_free_root(struct cgroup_root * root)1333 void cgroup_free_root(struct cgroup_root *root)
1334 {
1335 kfree(root);
1336 }
1337
cgroup_destroy_root(struct cgroup_root * root)1338 static void cgroup_destroy_root(struct cgroup_root *root)
1339 {
1340 struct cgroup *cgrp = &root->cgrp;
1341 struct cgrp_cset_link *link, *tmp_link;
1342
1343 trace_cgroup_destroy_root(root);
1344
1345 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1346
1347 BUG_ON(atomic_read(&root->nr_cgrps));
1348 BUG_ON(!list_empty(&cgrp->self.children));
1349
1350 /* Rebind all subsystems back to the default hierarchy */
1351 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1352
1353 /*
1354 * Release all the links from cset_links to this hierarchy's
1355 * root cgroup
1356 */
1357 spin_lock_irq(&css_set_lock);
1358
1359 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1360 list_del(&link->cset_link);
1361 list_del(&link->cgrp_link);
1362 kfree(link);
1363 }
1364
1365 spin_unlock_irq(&css_set_lock);
1366
1367 if (!list_empty(&root->root_list)) {
1368 list_del(&root->root_list);
1369 cgroup_root_count--;
1370 }
1371
1372 cgroup_exit_root_id(root);
1373
1374 mutex_unlock(&cgroup_mutex);
1375
1376 cgroup_rstat_exit(cgrp);
1377 kernfs_destroy_root(root->kf_root);
1378 cgroup_free_root(root);
1379 }
1380
1381 /*
1382 * look up cgroup associated with current task's cgroup namespace on the
1383 * specified hierarchy
1384 */
1385 static struct cgroup *
current_cgns_cgroup_from_root(struct cgroup_root * root)1386 current_cgns_cgroup_from_root(struct cgroup_root *root)
1387 {
1388 struct cgroup *res = NULL;
1389 struct css_set *cset;
1390
1391 lockdep_assert_held(&css_set_lock);
1392
1393 rcu_read_lock();
1394
1395 cset = current->nsproxy->cgroup_ns->root_cset;
1396 if (cset == &init_css_set) {
1397 res = &root->cgrp;
1398 } else if (root == &cgrp_dfl_root) {
1399 res = cset->dfl_cgrp;
1400 } else {
1401 struct cgrp_cset_link *link;
1402
1403 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1404 struct cgroup *c = link->cgrp;
1405
1406 if (c->root == root) {
1407 res = c;
1408 break;
1409 }
1410 }
1411 }
1412 rcu_read_unlock();
1413
1414 BUG_ON(!res);
1415 return res;
1416 }
1417
1418 /* look up cgroup associated with given css_set on the specified hierarchy */
cset_cgroup_from_root(struct css_set * cset,struct cgroup_root * root)1419 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1420 struct cgroup_root *root)
1421 {
1422 struct cgroup *res = NULL;
1423
1424 lockdep_assert_held(&cgroup_mutex);
1425 lockdep_assert_held(&css_set_lock);
1426
1427 if (cset == &init_css_set) {
1428 res = &root->cgrp;
1429 } else if (root == &cgrp_dfl_root) {
1430 res = cset->dfl_cgrp;
1431 } else {
1432 struct cgrp_cset_link *link;
1433
1434 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1435 struct cgroup *c = link->cgrp;
1436
1437 if (c->root == root) {
1438 res = c;
1439 break;
1440 }
1441 }
1442 }
1443
1444 BUG_ON(!res);
1445 return res;
1446 }
1447
1448 /*
1449 * Return the cgroup for "task" from the given hierarchy. Must be
1450 * called with cgroup_mutex and css_set_lock held.
1451 */
task_cgroup_from_root(struct task_struct * task,struct cgroup_root * root)1452 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1453 struct cgroup_root *root)
1454 {
1455 /*
1456 * No need to lock the task - since we hold css_set_lock the
1457 * task can't change groups.
1458 */
1459 return cset_cgroup_from_root(task_css_set(task), root);
1460 }
1461
1462 /*
1463 * A task must hold cgroup_mutex to modify cgroups.
1464 *
1465 * Any task can increment and decrement the count field without lock.
1466 * So in general, code holding cgroup_mutex can't rely on the count
1467 * field not changing. However, if the count goes to zero, then only
1468 * cgroup_attach_task() can increment it again. Because a count of zero
1469 * means that no tasks are currently attached, therefore there is no
1470 * way a task attached to that cgroup can fork (the other way to
1471 * increment the count). So code holding cgroup_mutex can safely
1472 * assume that if the count is zero, it will stay zero. Similarly, if
1473 * a task holds cgroup_mutex on a cgroup with zero count, it
1474 * knows that the cgroup won't be removed, as cgroup_rmdir()
1475 * needs that mutex.
1476 *
1477 * A cgroup can only be deleted if both its 'count' of using tasks
1478 * is zero, and its list of 'children' cgroups is empty. Since all
1479 * tasks in the system use _some_ cgroup, and since there is always at
1480 * least one task in the system (init, pid == 1), therefore, root cgroup
1481 * always has either children cgroups and/or using tasks. So we don't
1482 * need a special hack to ensure that root cgroup cannot be deleted.
1483 *
1484 * P.S. One more locking exception. RCU is used to guard the
1485 * update of a tasks cgroup pointer by cgroup_attach_task()
1486 */
1487
1488 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1489
cgroup_file_name(struct cgroup * cgrp,const struct cftype * cft,char * buf)1490 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1491 char *buf)
1492 {
1493 struct cgroup_subsys *ss = cft->ss;
1494
1495 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1496 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1497 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1498
1499 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1500 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1501 cft->name);
1502 } else {
1503 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1504 }
1505 return buf;
1506 }
1507
1508 /**
1509 * cgroup_file_mode - deduce file mode of a control file
1510 * @cft: the control file in question
1511 *
1512 * S_IRUGO for read, S_IWUSR for write.
1513 */
cgroup_file_mode(const struct cftype * cft)1514 static umode_t cgroup_file_mode(const struct cftype *cft)
1515 {
1516 umode_t mode = 0;
1517
1518 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1519 mode |= S_IRUGO;
1520
1521 if (cft->write_u64 || cft->write_s64 || cft->write) {
1522 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1523 mode |= S_IWUGO;
1524 else
1525 mode |= S_IWUSR;
1526 }
1527
1528 return mode;
1529 }
1530
1531 /**
1532 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1533 * @subtree_control: the new subtree_control mask to consider
1534 * @this_ss_mask: available subsystems
1535 *
1536 * On the default hierarchy, a subsystem may request other subsystems to be
1537 * enabled together through its ->depends_on mask. In such cases, more
1538 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1539 *
1540 * This function calculates which subsystems need to be enabled if
1541 * @subtree_control is to be applied while restricted to @this_ss_mask.
1542 */
cgroup_calc_subtree_ss_mask(u16 subtree_control,u16 this_ss_mask)1543 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1544 {
1545 u16 cur_ss_mask = subtree_control;
1546 struct cgroup_subsys *ss;
1547 int ssid;
1548
1549 lockdep_assert_held(&cgroup_mutex);
1550
1551 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1552
1553 while (true) {
1554 u16 new_ss_mask = cur_ss_mask;
1555
1556 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1557 new_ss_mask |= ss->depends_on;
1558 } while_each_subsys_mask();
1559
1560 /*
1561 * Mask out subsystems which aren't available. This can
1562 * happen only if some depended-upon subsystems were bound
1563 * to non-default hierarchies.
1564 */
1565 new_ss_mask &= this_ss_mask;
1566
1567 if (new_ss_mask == cur_ss_mask)
1568 break;
1569 cur_ss_mask = new_ss_mask;
1570 }
1571
1572 return cur_ss_mask;
1573 }
1574
1575 /**
1576 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1577 * @kn: the kernfs_node being serviced
1578 *
1579 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1580 * the method finishes if locking succeeded. Note that once this function
1581 * returns the cgroup returned by cgroup_kn_lock_live() may become
1582 * inaccessible any time. If the caller intends to continue to access the
1583 * cgroup, it should pin it before invoking this function.
1584 */
cgroup_kn_unlock(struct kernfs_node * kn)1585 void cgroup_kn_unlock(struct kernfs_node *kn)
1586 {
1587 struct cgroup *cgrp;
1588
1589 if (kernfs_type(kn) == KERNFS_DIR)
1590 cgrp = kn->priv;
1591 else
1592 cgrp = kn->parent->priv;
1593
1594 mutex_unlock(&cgroup_mutex);
1595
1596 kernfs_unbreak_active_protection(kn);
1597 cgroup_put(cgrp);
1598 }
1599
1600 /**
1601 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1602 * @kn: the kernfs_node being serviced
1603 * @drain_offline: perform offline draining on the cgroup
1604 *
1605 * This helper is to be used by a cgroup kernfs method currently servicing
1606 * @kn. It breaks the active protection, performs cgroup locking and
1607 * verifies that the associated cgroup is alive. Returns the cgroup if
1608 * alive; otherwise, %NULL. A successful return should be undone by a
1609 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1610 * cgroup is drained of offlining csses before return.
1611 *
1612 * Any cgroup kernfs method implementation which requires locking the
1613 * associated cgroup should use this helper. It avoids nesting cgroup
1614 * locking under kernfs active protection and allows all kernfs operations
1615 * including self-removal.
1616 */
cgroup_kn_lock_live(struct kernfs_node * kn,bool drain_offline)1617 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1618 {
1619 struct cgroup *cgrp;
1620
1621 if (kernfs_type(kn) == KERNFS_DIR)
1622 cgrp = kn->priv;
1623 else
1624 cgrp = kn->parent->priv;
1625
1626 /*
1627 * We're gonna grab cgroup_mutex which nests outside kernfs
1628 * active_ref. cgroup liveliness check alone provides enough
1629 * protection against removal. Ensure @cgrp stays accessible and
1630 * break the active_ref protection.
1631 */
1632 if (!cgroup_tryget(cgrp))
1633 return NULL;
1634 kernfs_break_active_protection(kn);
1635
1636 if (drain_offline)
1637 cgroup_lock_and_drain_offline(cgrp);
1638 else
1639 mutex_lock(&cgroup_mutex);
1640
1641 if (!cgroup_is_dead(cgrp))
1642 return cgrp;
1643
1644 cgroup_kn_unlock(kn);
1645 return NULL;
1646 }
1647
cgroup_rm_file(struct cgroup * cgrp,const struct cftype * cft)1648 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1649 {
1650 char name[CGROUP_FILE_NAME_MAX];
1651
1652 lockdep_assert_held(&cgroup_mutex);
1653
1654 if (cft->file_offset) {
1655 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1656 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1657
1658 spin_lock_irq(&cgroup_file_kn_lock);
1659 cfile->kn = NULL;
1660 spin_unlock_irq(&cgroup_file_kn_lock);
1661
1662 del_timer_sync(&cfile->notify_timer);
1663 }
1664
1665 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1666 }
1667
1668 /**
1669 * css_clear_dir - remove subsys files in a cgroup directory
1670 * @css: target css
1671 */
css_clear_dir(struct cgroup_subsys_state * css)1672 static void css_clear_dir(struct cgroup_subsys_state *css)
1673 {
1674 struct cgroup *cgrp = css->cgroup;
1675 struct cftype *cfts;
1676
1677 if (!(css->flags & CSS_VISIBLE))
1678 return;
1679
1680 css->flags &= ~CSS_VISIBLE;
1681
1682 if (!css->ss) {
1683 if (cgroup_on_dfl(cgrp))
1684 cfts = cgroup_base_files;
1685 else
1686 cfts = cgroup1_base_files;
1687
1688 cgroup_addrm_files(css, cgrp, cfts, false);
1689 } else {
1690 list_for_each_entry(cfts, &css->ss->cfts, node)
1691 cgroup_addrm_files(css, cgrp, cfts, false);
1692 }
1693 }
1694
1695 /**
1696 * css_populate_dir - create subsys files in a cgroup directory
1697 * @css: target css
1698 *
1699 * On failure, no file is added.
1700 */
css_populate_dir(struct cgroup_subsys_state * css)1701 static int css_populate_dir(struct cgroup_subsys_state *css)
1702 {
1703 struct cgroup *cgrp = css->cgroup;
1704 struct cftype *cfts, *failed_cfts;
1705 int ret;
1706
1707 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1708 return 0;
1709
1710 if (!css->ss) {
1711 if (cgroup_on_dfl(cgrp))
1712 cfts = cgroup_base_files;
1713 else
1714 cfts = cgroup1_base_files;
1715
1716 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1717 if (ret < 0)
1718 return ret;
1719 } else {
1720 list_for_each_entry(cfts, &css->ss->cfts, node) {
1721 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1722 if (ret < 0) {
1723 failed_cfts = cfts;
1724 goto err;
1725 }
1726 }
1727 }
1728
1729 css->flags |= CSS_VISIBLE;
1730
1731 return 0;
1732 err:
1733 list_for_each_entry(cfts, &css->ss->cfts, node) {
1734 if (cfts == failed_cfts)
1735 break;
1736 cgroup_addrm_files(css, cgrp, cfts, false);
1737 }
1738 return ret;
1739 }
1740
rebind_subsystems(struct cgroup_root * dst_root,u16 ss_mask)1741 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1742 {
1743 struct cgroup *dcgrp = &dst_root->cgrp;
1744 struct cgroup_subsys *ss;
1745 int ssid, i, ret;
1746 u16 dfl_disable_ss_mask = 0;
1747
1748 lockdep_assert_held(&cgroup_mutex);
1749
1750 do_each_subsys_mask(ss, ssid, ss_mask) {
1751 /*
1752 * If @ss has non-root csses attached to it, can't move.
1753 * If @ss is an implicit controller, it is exempt from this
1754 * rule and can be stolen.
1755 */
1756 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1757 !ss->implicit_on_dfl)
1758 return -EBUSY;
1759
1760 /* can't move between two non-dummy roots either */
1761 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1762 return -EBUSY;
1763
1764 /*
1765 * Collect ssid's that need to be disabled from default
1766 * hierarchy.
1767 */
1768 if (ss->root == &cgrp_dfl_root)
1769 dfl_disable_ss_mask |= 1 << ssid;
1770
1771 } while_each_subsys_mask();
1772
1773 if (dfl_disable_ss_mask) {
1774 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1775
1776 /*
1777 * Controllers from default hierarchy that need to be rebound
1778 * are all disabled together in one go.
1779 */
1780 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1781 WARN_ON(cgroup_apply_control(scgrp));
1782 cgroup_finalize_control(scgrp, 0);
1783 }
1784
1785 do_each_subsys_mask(ss, ssid, ss_mask) {
1786 struct cgroup_root *src_root = ss->root;
1787 struct cgroup *scgrp = &src_root->cgrp;
1788 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1789 struct css_set *cset;
1790
1791 WARN_ON(!css || cgroup_css(dcgrp, ss));
1792
1793 if (src_root != &cgrp_dfl_root) {
1794 /* disable from the source */
1795 src_root->subsys_mask &= ~(1 << ssid);
1796 WARN_ON(cgroup_apply_control(scgrp));
1797 cgroup_finalize_control(scgrp, 0);
1798 }
1799
1800 /* rebind */
1801 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1802 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1803 ss->root = dst_root;
1804 css->cgroup = dcgrp;
1805
1806 spin_lock_irq(&css_set_lock);
1807 hash_for_each(css_set_table, i, cset, hlist)
1808 list_move_tail(&cset->e_cset_node[ss->id],
1809 &dcgrp->e_csets[ss->id]);
1810 spin_unlock_irq(&css_set_lock);
1811
1812 if (ss->css_rstat_flush) {
1813 list_del_rcu(&css->rstat_css_node);
1814 synchronize_rcu();
1815 list_add_rcu(&css->rstat_css_node,
1816 &dcgrp->rstat_css_list);
1817 }
1818
1819 /* default hierarchy doesn't enable controllers by default */
1820 dst_root->subsys_mask |= 1 << ssid;
1821 if (dst_root == &cgrp_dfl_root) {
1822 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1823 } else {
1824 dcgrp->subtree_control |= 1 << ssid;
1825 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1826 }
1827
1828 ret = cgroup_apply_control(dcgrp);
1829 if (ret)
1830 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1831 ss->name, ret);
1832
1833 if (ss->bind)
1834 ss->bind(css);
1835 } while_each_subsys_mask();
1836
1837 kernfs_activate(dcgrp->kn);
1838 return 0;
1839 }
1840
cgroup_show_path(struct seq_file * sf,struct kernfs_node * kf_node,struct kernfs_root * kf_root)1841 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1842 struct kernfs_root *kf_root)
1843 {
1844 int len = 0;
1845 char *buf = NULL;
1846 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1847 struct cgroup *ns_cgroup;
1848
1849 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1850 if (!buf)
1851 return -ENOMEM;
1852
1853 spin_lock_irq(&css_set_lock);
1854 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1855 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1856 spin_unlock_irq(&css_set_lock);
1857
1858 if (len >= PATH_MAX)
1859 len = -ERANGE;
1860 else if (len > 0) {
1861 seq_escape(sf, buf, " \t\n\\");
1862 len = 0;
1863 }
1864 kfree(buf);
1865 return len;
1866 }
1867
1868 enum cgroup2_param {
1869 Opt_nsdelegate,
1870 Opt_memory_localevents,
1871 Opt_memory_recursiveprot,
1872 nr__cgroup2_params
1873 };
1874
1875 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1876 fsparam_flag("nsdelegate", Opt_nsdelegate),
1877 fsparam_flag("memory_localevents", Opt_memory_localevents),
1878 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1879 {}
1880 };
1881
cgroup2_parse_param(struct fs_context * fc,struct fs_parameter * param)1882 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1883 {
1884 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1885 struct fs_parse_result result;
1886 int opt;
1887
1888 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1889 if (opt < 0)
1890 return opt;
1891
1892 switch (opt) {
1893 case Opt_nsdelegate:
1894 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1895 return 0;
1896 case Opt_memory_localevents:
1897 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1898 return 0;
1899 case Opt_memory_recursiveprot:
1900 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1901 return 0;
1902 }
1903 return -EINVAL;
1904 }
1905
apply_cgroup_root_flags(unsigned int root_flags)1906 static void apply_cgroup_root_flags(unsigned int root_flags)
1907 {
1908 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1909 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1910 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1911 else
1912 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1913
1914 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1915 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1916 else
1917 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1918
1919 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1920 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1921 else
1922 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1923 }
1924 }
1925
cgroup_show_options(struct seq_file * seq,struct kernfs_root * kf_root)1926 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1927 {
1928 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1929 seq_puts(seq, ",nsdelegate");
1930 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1931 seq_puts(seq, ",memory_localevents");
1932 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1933 seq_puts(seq, ",memory_recursiveprot");
1934 return 0;
1935 }
1936
cgroup_reconfigure(struct fs_context * fc)1937 static int cgroup_reconfigure(struct fs_context *fc)
1938 {
1939 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1940
1941 apply_cgroup_root_flags(ctx->flags);
1942 return 0;
1943 }
1944
init_cgroup_housekeeping(struct cgroup * cgrp)1945 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1946 {
1947 struct cgroup_subsys *ss;
1948 int ssid;
1949
1950 INIT_LIST_HEAD(&cgrp->self.sibling);
1951 INIT_LIST_HEAD(&cgrp->self.children);
1952 INIT_LIST_HEAD(&cgrp->cset_links);
1953 INIT_LIST_HEAD(&cgrp->pidlists);
1954 mutex_init(&cgrp->pidlist_mutex);
1955 cgrp->self.cgroup = cgrp;
1956 cgrp->self.flags |= CSS_ONLINE;
1957 cgrp->dom_cgrp = cgrp;
1958 cgrp->max_descendants = INT_MAX;
1959 cgrp->max_depth = INT_MAX;
1960 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1961 prev_cputime_init(&cgrp->prev_cputime);
1962
1963 for_each_subsys(ss, ssid)
1964 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1965
1966 init_waitqueue_head(&cgrp->offline_waitq);
1967 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1968 }
1969
init_cgroup_root(struct cgroup_fs_context * ctx)1970 void init_cgroup_root(struct cgroup_fs_context *ctx)
1971 {
1972 struct cgroup_root *root = ctx->root;
1973 struct cgroup *cgrp = &root->cgrp;
1974
1975 INIT_LIST_HEAD(&root->root_list);
1976 atomic_set(&root->nr_cgrps, 1);
1977 cgrp->root = root;
1978 init_cgroup_housekeeping(cgrp);
1979
1980 root->flags = ctx->flags;
1981 if (ctx->release_agent)
1982 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1983 if (ctx->name)
1984 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1985 if (ctx->cpuset_clone_children)
1986 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1987 }
1988
cgroup_setup_root(struct cgroup_root * root,u16 ss_mask)1989 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1990 {
1991 LIST_HEAD(tmp_links);
1992 struct cgroup *root_cgrp = &root->cgrp;
1993 struct kernfs_syscall_ops *kf_sops;
1994 struct css_set *cset;
1995 int i, ret;
1996
1997 lockdep_assert_held(&cgroup_mutex);
1998
1999 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2000 0, GFP_KERNEL);
2001 if (ret)
2002 goto out;
2003
2004 /*
2005 * We're accessing css_set_count without locking css_set_lock here,
2006 * but that's OK - it can only be increased by someone holding
2007 * cgroup_lock, and that's us. Later rebinding may disable
2008 * controllers on the default hierarchy and thus create new csets,
2009 * which can't be more than the existing ones. Allocate 2x.
2010 */
2011 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2012 if (ret)
2013 goto cancel_ref;
2014
2015 ret = cgroup_init_root_id(root);
2016 if (ret)
2017 goto cancel_ref;
2018
2019 kf_sops = root == &cgrp_dfl_root ?
2020 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2021
2022 root->kf_root = kernfs_create_root(kf_sops,
2023 KERNFS_ROOT_CREATE_DEACTIVATED |
2024 KERNFS_ROOT_SUPPORT_EXPORTOP |
2025 KERNFS_ROOT_SUPPORT_USER_XATTR,
2026 root_cgrp);
2027 if (IS_ERR(root->kf_root)) {
2028 ret = PTR_ERR(root->kf_root);
2029 goto exit_root_id;
2030 }
2031 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2032 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2033 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
2034
2035 ret = css_populate_dir(&root_cgrp->self);
2036 if (ret)
2037 goto destroy_root;
2038
2039 ret = cgroup_rstat_init(root_cgrp);
2040 if (ret)
2041 goto destroy_root;
2042
2043 ret = rebind_subsystems(root, ss_mask);
2044 if (ret)
2045 goto exit_stats;
2046
2047 ret = cgroup_bpf_inherit(root_cgrp);
2048 WARN_ON_ONCE(ret);
2049
2050 trace_cgroup_setup_root(root);
2051
2052 /*
2053 * There must be no failure case after here, since rebinding takes
2054 * care of subsystems' refcounts, which are explicitly dropped in
2055 * the failure exit path.
2056 */
2057 list_add(&root->root_list, &cgroup_roots);
2058 cgroup_root_count++;
2059
2060 /*
2061 * Link the root cgroup in this hierarchy into all the css_set
2062 * objects.
2063 */
2064 spin_lock_irq(&css_set_lock);
2065 hash_for_each(css_set_table, i, cset, hlist) {
2066 link_css_set(&tmp_links, cset, root_cgrp);
2067 if (css_set_populated(cset))
2068 cgroup_update_populated(root_cgrp, true);
2069 }
2070 spin_unlock_irq(&css_set_lock);
2071
2072 BUG_ON(!list_empty(&root_cgrp->self.children));
2073 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2074
2075 ret = 0;
2076 goto out;
2077
2078 exit_stats:
2079 cgroup_rstat_exit(root_cgrp);
2080 destroy_root:
2081 kernfs_destroy_root(root->kf_root);
2082 root->kf_root = NULL;
2083 exit_root_id:
2084 cgroup_exit_root_id(root);
2085 cancel_ref:
2086 percpu_ref_exit(&root_cgrp->self.refcnt);
2087 out:
2088 free_cgrp_cset_links(&tmp_links);
2089 return ret;
2090 }
2091
cgroup_do_get_tree(struct fs_context * fc)2092 int cgroup_do_get_tree(struct fs_context *fc)
2093 {
2094 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2095 int ret;
2096
2097 ctx->kfc.root = ctx->root->kf_root;
2098 if (fc->fs_type == &cgroup2_fs_type)
2099 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2100 else
2101 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2102 ret = kernfs_get_tree(fc);
2103
2104 /*
2105 * In non-init cgroup namespace, instead of root cgroup's dentry,
2106 * we return the dentry corresponding to the cgroupns->root_cgrp.
2107 */
2108 if (!ret && ctx->ns != &init_cgroup_ns) {
2109 struct dentry *nsdentry;
2110 struct super_block *sb = fc->root->d_sb;
2111 struct cgroup *cgrp;
2112
2113 mutex_lock(&cgroup_mutex);
2114 spin_lock_irq(&css_set_lock);
2115
2116 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2117
2118 spin_unlock_irq(&css_set_lock);
2119 mutex_unlock(&cgroup_mutex);
2120
2121 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2122 dput(fc->root);
2123 if (IS_ERR(nsdentry)) {
2124 deactivate_locked_super(sb);
2125 ret = PTR_ERR(nsdentry);
2126 nsdentry = NULL;
2127 }
2128 fc->root = nsdentry;
2129 }
2130
2131 if (!ctx->kfc.new_sb_created)
2132 cgroup_put(&ctx->root->cgrp);
2133
2134 return ret;
2135 }
2136
2137 /*
2138 * Destroy a cgroup filesystem context.
2139 */
cgroup_fs_context_free(struct fs_context * fc)2140 static void cgroup_fs_context_free(struct fs_context *fc)
2141 {
2142 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2143
2144 kfree(ctx->name);
2145 kfree(ctx->release_agent);
2146 put_cgroup_ns(ctx->ns);
2147 kernfs_free_fs_context(fc);
2148 kfree(ctx);
2149 }
2150
cgroup_get_tree(struct fs_context * fc)2151 static int cgroup_get_tree(struct fs_context *fc)
2152 {
2153 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2154 int ret;
2155
2156 cgrp_dfl_visible = true;
2157 cgroup_get_live(&cgrp_dfl_root.cgrp);
2158 ctx->root = &cgrp_dfl_root;
2159
2160 ret = cgroup_do_get_tree(fc);
2161 if (!ret)
2162 apply_cgroup_root_flags(ctx->flags);
2163 return ret;
2164 }
2165
2166 static const struct fs_context_operations cgroup_fs_context_ops = {
2167 .free = cgroup_fs_context_free,
2168 .parse_param = cgroup2_parse_param,
2169 .get_tree = cgroup_get_tree,
2170 .reconfigure = cgroup_reconfigure,
2171 };
2172
2173 static const struct fs_context_operations cgroup1_fs_context_ops = {
2174 .free = cgroup_fs_context_free,
2175 .parse_param = cgroup1_parse_param,
2176 .get_tree = cgroup1_get_tree,
2177 .reconfigure = cgroup1_reconfigure,
2178 };
2179
2180 /*
2181 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2182 * we select the namespace we're going to use.
2183 */
cgroup_init_fs_context(struct fs_context * fc)2184 static int cgroup_init_fs_context(struct fs_context *fc)
2185 {
2186 struct cgroup_fs_context *ctx;
2187
2188 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2189 if (!ctx)
2190 return -ENOMEM;
2191
2192 ctx->ns = current->nsproxy->cgroup_ns;
2193 get_cgroup_ns(ctx->ns);
2194 fc->fs_private = &ctx->kfc;
2195 if (fc->fs_type == &cgroup2_fs_type)
2196 fc->ops = &cgroup_fs_context_ops;
2197 else
2198 fc->ops = &cgroup1_fs_context_ops;
2199 put_user_ns(fc->user_ns);
2200 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2201 fc->global = true;
2202 return 0;
2203 }
2204
cgroup_kill_sb(struct super_block * sb)2205 static void cgroup_kill_sb(struct super_block *sb)
2206 {
2207 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2208 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2209
2210 /*
2211 * If @root doesn't have any children, start killing it.
2212 * This prevents new mounts by disabling percpu_ref_tryget_live().
2213 *
2214 * And don't kill the default root.
2215 */
2216 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2217 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) {
2218 cgroup_bpf_offline(&root->cgrp);
2219 percpu_ref_kill(&root->cgrp.self.refcnt);
2220 }
2221 cgroup_put(&root->cgrp);
2222 kernfs_kill_sb(sb);
2223 }
2224
2225 struct file_system_type cgroup_fs_type = {
2226 .name = "cgroup",
2227 .init_fs_context = cgroup_init_fs_context,
2228 .parameters = cgroup1_fs_parameters,
2229 .kill_sb = cgroup_kill_sb,
2230 .fs_flags = FS_USERNS_MOUNT,
2231 };
2232
2233 static struct file_system_type cgroup2_fs_type = {
2234 .name = "cgroup2",
2235 .init_fs_context = cgroup_init_fs_context,
2236 .parameters = cgroup2_fs_parameters,
2237 .kill_sb = cgroup_kill_sb,
2238 .fs_flags = FS_USERNS_MOUNT,
2239 };
2240
2241 #ifdef CONFIG_CPUSETS
2242 static const struct fs_context_operations cpuset_fs_context_ops = {
2243 .get_tree = cgroup1_get_tree,
2244 .free = cgroup_fs_context_free,
2245 };
2246
2247 /*
2248 * This is ugly, but preserves the userspace API for existing cpuset
2249 * users. If someone tries to mount the "cpuset" filesystem, we
2250 * silently switch it to mount "cgroup" instead
2251 */
cpuset_init_fs_context(struct fs_context * fc)2252 static int cpuset_init_fs_context(struct fs_context *fc)
2253 {
2254 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2255 struct cgroup_fs_context *ctx;
2256 int err;
2257
2258 err = cgroup_init_fs_context(fc);
2259 if (err) {
2260 kfree(agent);
2261 return err;
2262 }
2263
2264 fc->ops = &cpuset_fs_context_ops;
2265
2266 ctx = cgroup_fc2context(fc);
2267 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2268 ctx->flags |= CGRP_ROOT_NOPREFIX;
2269 ctx->release_agent = agent;
2270
2271 get_filesystem(&cgroup_fs_type);
2272 put_filesystem(fc->fs_type);
2273 fc->fs_type = &cgroup_fs_type;
2274
2275 return 0;
2276 }
2277
2278 static struct file_system_type cpuset_fs_type = {
2279 .name = "cpuset",
2280 .init_fs_context = cpuset_init_fs_context,
2281 .fs_flags = FS_USERNS_MOUNT,
2282 };
2283 #endif
2284
cgroup_path_ns_locked(struct cgroup * cgrp,char * buf,size_t buflen,struct cgroup_namespace * ns)2285 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2286 struct cgroup_namespace *ns)
2287 {
2288 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2289
2290 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2291 }
2292
cgroup_path_ns(struct cgroup * cgrp,char * buf,size_t buflen,struct cgroup_namespace * ns)2293 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2294 struct cgroup_namespace *ns)
2295 {
2296 int ret;
2297
2298 mutex_lock(&cgroup_mutex);
2299 spin_lock_irq(&css_set_lock);
2300
2301 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2302
2303 spin_unlock_irq(&css_set_lock);
2304 mutex_unlock(&cgroup_mutex);
2305
2306 return ret;
2307 }
2308 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2309
2310 /**
2311 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2312 * @task: target task
2313 * @buf: the buffer to write the path into
2314 * @buflen: the length of the buffer
2315 *
2316 * Determine @task's cgroup on the first (the one with the lowest non-zero
2317 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2318 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2319 * cgroup controller callbacks.
2320 *
2321 * Return value is the same as kernfs_path().
2322 */
task_cgroup_path(struct task_struct * task,char * buf,size_t buflen)2323 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2324 {
2325 struct cgroup_root *root;
2326 struct cgroup *cgrp;
2327 int hierarchy_id = 1;
2328 int ret;
2329
2330 mutex_lock(&cgroup_mutex);
2331 spin_lock_irq(&css_set_lock);
2332
2333 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2334
2335 if (root) {
2336 cgrp = task_cgroup_from_root(task, root);
2337 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2338 } else {
2339 /* if no hierarchy exists, everyone is in "/" */
2340 ret = strlcpy(buf, "/", buflen);
2341 }
2342
2343 spin_unlock_irq(&css_set_lock);
2344 mutex_unlock(&cgroup_mutex);
2345 return ret;
2346 }
2347 EXPORT_SYMBOL_GPL(task_cgroup_path);
2348
2349 /**
2350 * cgroup_attach_lock - Lock for ->attach()
2351 * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2352 *
2353 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2354 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2355 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2356 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2357 * lead to deadlocks.
2358 *
2359 * Bringing up a CPU may involve creating and destroying tasks which requires
2360 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2361 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2362 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2363 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2364 * the threadgroup_rwsem to be released to create new tasks. For more details:
2365 *
2366 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2367 *
2368 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2369 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2370 * CPU hotplug is disabled on entry.
2371 */
cgroup_attach_lock(bool lock_threadgroup)2372 static void cgroup_attach_lock(bool lock_threadgroup)
2373 {
2374 cpus_read_lock();
2375 if (lock_threadgroup)
2376 percpu_down_write(&cgroup_threadgroup_rwsem);
2377 }
2378
2379 /**
2380 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2381 * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2382 */
cgroup_attach_unlock(bool lock_threadgroup)2383 static void cgroup_attach_unlock(bool lock_threadgroup)
2384 {
2385 if (lock_threadgroup)
2386 percpu_up_write(&cgroup_threadgroup_rwsem);
2387 cpus_read_unlock();
2388 }
2389
2390 /**
2391 * cgroup_migrate_add_task - add a migration target task to a migration context
2392 * @task: target task
2393 * @mgctx: target migration context
2394 *
2395 * Add @task, which is a migration target, to @mgctx->tset. This function
2396 * becomes noop if @task doesn't need to be migrated. @task's css_set
2397 * should have been added as a migration source and @task->cg_list will be
2398 * moved from the css_set's tasks list to mg_tasks one.
2399 */
cgroup_migrate_add_task(struct task_struct * task,struct cgroup_mgctx * mgctx)2400 static void cgroup_migrate_add_task(struct task_struct *task,
2401 struct cgroup_mgctx *mgctx)
2402 {
2403 struct css_set *cset;
2404
2405 lockdep_assert_held(&css_set_lock);
2406
2407 /* @task either already exited or can't exit until the end */
2408 if (task->flags & PF_EXITING)
2409 return;
2410
2411 /* cgroup_threadgroup_rwsem protects racing against forks */
2412 WARN_ON_ONCE(list_empty(&task->cg_list));
2413
2414 cset = task_css_set(task);
2415 if (!cset->mg_src_cgrp)
2416 return;
2417
2418 mgctx->tset.nr_tasks++;
2419
2420 list_move_tail(&task->cg_list, &cset->mg_tasks);
2421 if (list_empty(&cset->mg_node))
2422 list_add_tail(&cset->mg_node,
2423 &mgctx->tset.src_csets);
2424 if (list_empty(&cset->mg_dst_cset->mg_node))
2425 list_add_tail(&cset->mg_dst_cset->mg_node,
2426 &mgctx->tset.dst_csets);
2427 }
2428
2429 /**
2430 * cgroup_taskset_first - reset taskset and return the first task
2431 * @tset: taskset of interest
2432 * @dst_cssp: output variable for the destination css
2433 *
2434 * @tset iteration is initialized and the first task is returned.
2435 */
cgroup_taskset_first(struct cgroup_taskset * tset,struct cgroup_subsys_state ** dst_cssp)2436 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2437 struct cgroup_subsys_state **dst_cssp)
2438 {
2439 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2440 tset->cur_task = NULL;
2441
2442 return cgroup_taskset_next(tset, dst_cssp);
2443 }
2444
2445 /**
2446 * cgroup_taskset_next - iterate to the next task in taskset
2447 * @tset: taskset of interest
2448 * @dst_cssp: output variable for the destination css
2449 *
2450 * Return the next task in @tset. Iteration must have been initialized
2451 * with cgroup_taskset_first().
2452 */
cgroup_taskset_next(struct cgroup_taskset * tset,struct cgroup_subsys_state ** dst_cssp)2453 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2454 struct cgroup_subsys_state **dst_cssp)
2455 {
2456 struct css_set *cset = tset->cur_cset;
2457 struct task_struct *task = tset->cur_task;
2458
2459 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2460 if (!task)
2461 task = list_first_entry(&cset->mg_tasks,
2462 struct task_struct, cg_list);
2463 else
2464 task = list_next_entry(task, cg_list);
2465
2466 if (&task->cg_list != &cset->mg_tasks) {
2467 tset->cur_cset = cset;
2468 tset->cur_task = task;
2469
2470 /*
2471 * This function may be called both before and
2472 * after cgroup_taskset_migrate(). The two cases
2473 * can be distinguished by looking at whether @cset
2474 * has its ->mg_dst_cset set.
2475 */
2476 if (cset->mg_dst_cset)
2477 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2478 else
2479 *dst_cssp = cset->subsys[tset->ssid];
2480
2481 return task;
2482 }
2483
2484 cset = list_next_entry(cset, mg_node);
2485 task = NULL;
2486 }
2487
2488 return NULL;
2489 }
2490
2491 /**
2492 * cgroup_migrate_execute - migrate a taskset
2493 * @mgctx: migration context
2494 *
2495 * Migrate tasks in @mgctx as setup by migration preparation functions.
2496 * This function fails iff one of the ->can_attach callbacks fails and
2497 * guarantees that either all or none of the tasks in @mgctx are migrated.
2498 * @mgctx is consumed regardless of success.
2499 */
cgroup_migrate_execute(struct cgroup_mgctx * mgctx)2500 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2501 {
2502 struct cgroup_taskset *tset = &mgctx->tset;
2503 struct cgroup_subsys *ss;
2504 struct task_struct *task, *tmp_task;
2505 struct css_set *cset, *tmp_cset;
2506 int ssid, failed_ssid, ret;
2507
2508 /* check that we can legitimately attach to the cgroup */
2509 if (tset->nr_tasks) {
2510 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2511 if (ss->can_attach) {
2512 tset->ssid = ssid;
2513 ret = ss->can_attach(tset);
2514 if (ret) {
2515 failed_ssid = ssid;
2516 goto out_cancel_attach;
2517 }
2518 }
2519 } while_each_subsys_mask();
2520 }
2521
2522 /*
2523 * Now that we're guaranteed success, proceed to move all tasks to
2524 * the new cgroup. There are no failure cases after here, so this
2525 * is the commit point.
2526 */
2527 spin_lock_irq(&css_set_lock);
2528 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2529 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2530 struct css_set *from_cset = task_css_set(task);
2531 struct css_set *to_cset = cset->mg_dst_cset;
2532
2533 get_css_set(to_cset);
2534 to_cset->nr_tasks++;
2535 css_set_move_task(task, from_cset, to_cset, true);
2536 from_cset->nr_tasks--;
2537 /*
2538 * If the source or destination cgroup is frozen,
2539 * the task might require to change its state.
2540 */
2541 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2542 to_cset->dfl_cgrp);
2543 put_css_set_locked(from_cset);
2544
2545 }
2546 }
2547 spin_unlock_irq(&css_set_lock);
2548
2549 /*
2550 * Migration is committed, all target tasks are now on dst_csets.
2551 * Nothing is sensitive to fork() after this point. Notify
2552 * controllers that migration is complete.
2553 */
2554 tset->csets = &tset->dst_csets;
2555
2556 if (tset->nr_tasks) {
2557 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2558 if (ss->attach) {
2559 tset->ssid = ssid;
2560 ss->attach(tset);
2561 }
2562 } while_each_subsys_mask();
2563 }
2564
2565 ret = 0;
2566 goto out_release_tset;
2567
2568 out_cancel_attach:
2569 if (tset->nr_tasks) {
2570 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2571 if (ssid == failed_ssid)
2572 break;
2573 if (ss->cancel_attach) {
2574 tset->ssid = ssid;
2575 ss->cancel_attach(tset);
2576 }
2577 } while_each_subsys_mask();
2578 }
2579 out_release_tset:
2580 spin_lock_irq(&css_set_lock);
2581 list_splice_init(&tset->dst_csets, &tset->src_csets);
2582 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2583 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2584 list_del_init(&cset->mg_node);
2585 }
2586 spin_unlock_irq(&css_set_lock);
2587
2588 /*
2589 * Re-initialize the cgroup_taskset structure in case it is reused
2590 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2591 * iteration.
2592 */
2593 tset->nr_tasks = 0;
2594 tset->csets = &tset->src_csets;
2595 return ret;
2596 }
2597
2598 /**
2599 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2600 * @dst_cgrp: destination cgroup to test
2601 *
2602 * On the default hierarchy, except for the mixable, (possible) thread root
2603 * and threaded cgroups, subtree_control must be zero for migration
2604 * destination cgroups with tasks so that child cgroups don't compete
2605 * against tasks.
2606 */
cgroup_migrate_vet_dst(struct cgroup * dst_cgrp)2607 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2608 {
2609 /* v1 doesn't have any restriction */
2610 if (!cgroup_on_dfl(dst_cgrp))
2611 return 0;
2612
2613 /* verify @dst_cgrp can host resources */
2614 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2615 return -EOPNOTSUPP;
2616
2617 /* mixables don't care */
2618 if (cgroup_is_mixable(dst_cgrp))
2619 return 0;
2620
2621 /*
2622 * If @dst_cgrp is already or can become a thread root or is
2623 * threaded, it doesn't matter.
2624 */
2625 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2626 return 0;
2627
2628 /* apply no-internal-process constraint */
2629 if (dst_cgrp->subtree_control)
2630 return -EBUSY;
2631
2632 return 0;
2633 }
2634
2635 /**
2636 * cgroup_migrate_finish - cleanup after attach
2637 * @mgctx: migration context
2638 *
2639 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2640 * those functions for details.
2641 */
cgroup_migrate_finish(struct cgroup_mgctx * mgctx)2642 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2643 {
2644 struct css_set *cset, *tmp_cset;
2645
2646 lockdep_assert_held(&cgroup_mutex);
2647
2648 spin_lock_irq(&css_set_lock);
2649
2650 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2651 mg_src_preload_node) {
2652 cset->mg_src_cgrp = NULL;
2653 cset->mg_dst_cgrp = NULL;
2654 cset->mg_dst_cset = NULL;
2655 list_del_init(&cset->mg_src_preload_node);
2656 put_css_set_locked(cset);
2657 }
2658
2659 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2660 mg_dst_preload_node) {
2661 cset->mg_src_cgrp = NULL;
2662 cset->mg_dst_cgrp = NULL;
2663 cset->mg_dst_cset = NULL;
2664 list_del_init(&cset->mg_dst_preload_node);
2665 put_css_set_locked(cset);
2666 }
2667
2668 spin_unlock_irq(&css_set_lock);
2669 }
2670
2671 /**
2672 * cgroup_migrate_add_src - add a migration source css_set
2673 * @src_cset: the source css_set to add
2674 * @dst_cgrp: the destination cgroup
2675 * @mgctx: migration context
2676 *
2677 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2678 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2679 * up by cgroup_migrate_finish().
2680 *
2681 * This function may be called without holding cgroup_threadgroup_rwsem
2682 * even if the target is a process. Threads may be created and destroyed
2683 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2684 * into play and the preloaded css_sets are guaranteed to cover all
2685 * migrations.
2686 */
cgroup_migrate_add_src(struct css_set * src_cset,struct cgroup * dst_cgrp,struct cgroup_mgctx * mgctx)2687 void cgroup_migrate_add_src(struct css_set *src_cset,
2688 struct cgroup *dst_cgrp,
2689 struct cgroup_mgctx *mgctx)
2690 {
2691 struct cgroup *src_cgrp;
2692
2693 lockdep_assert_held(&cgroup_mutex);
2694 lockdep_assert_held(&css_set_lock);
2695
2696 /*
2697 * If ->dead, @src_set is associated with one or more dead cgroups
2698 * and doesn't contain any migratable tasks. Ignore it early so
2699 * that the rest of migration path doesn't get confused by it.
2700 */
2701 if (src_cset->dead)
2702 return;
2703
2704 if (!list_empty(&src_cset->mg_src_preload_node))
2705 return;
2706
2707 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2708
2709 WARN_ON(src_cset->mg_src_cgrp);
2710 WARN_ON(src_cset->mg_dst_cgrp);
2711 WARN_ON(!list_empty(&src_cset->mg_tasks));
2712 WARN_ON(!list_empty(&src_cset->mg_node));
2713
2714 src_cset->mg_src_cgrp = src_cgrp;
2715 src_cset->mg_dst_cgrp = dst_cgrp;
2716 get_css_set(src_cset);
2717 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2718 }
2719
2720 /**
2721 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2722 * @mgctx: migration context
2723 *
2724 * Tasks are about to be moved and all the source css_sets have been
2725 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2726 * pins all destination css_sets, links each to its source, and append them
2727 * to @mgctx->preloaded_dst_csets.
2728 *
2729 * This function must be called after cgroup_migrate_add_src() has been
2730 * called on each migration source css_set. After migration is performed
2731 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2732 * @mgctx.
2733 */
cgroup_migrate_prepare_dst(struct cgroup_mgctx * mgctx)2734 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2735 {
2736 struct css_set *src_cset, *tmp_cset;
2737
2738 lockdep_assert_held(&cgroup_mutex);
2739
2740 /* look up the dst cset for each src cset and link it to src */
2741 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2742 mg_src_preload_node) {
2743 struct css_set *dst_cset;
2744 struct cgroup_subsys *ss;
2745 int ssid;
2746
2747 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2748 if (!dst_cset)
2749 return -ENOMEM;
2750
2751 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2752
2753 /*
2754 * If src cset equals dst, it's noop. Drop the src.
2755 * cgroup_migrate() will skip the cset too. Note that we
2756 * can't handle src == dst as some nodes are used by both.
2757 */
2758 if (src_cset == dst_cset) {
2759 src_cset->mg_src_cgrp = NULL;
2760 src_cset->mg_dst_cgrp = NULL;
2761 list_del_init(&src_cset->mg_src_preload_node);
2762 put_css_set(src_cset);
2763 put_css_set(dst_cset);
2764 continue;
2765 }
2766
2767 src_cset->mg_dst_cset = dst_cset;
2768
2769 if (list_empty(&dst_cset->mg_dst_preload_node))
2770 list_add_tail(&dst_cset->mg_dst_preload_node,
2771 &mgctx->preloaded_dst_csets);
2772 else
2773 put_css_set(dst_cset);
2774
2775 for_each_subsys(ss, ssid)
2776 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2777 mgctx->ss_mask |= 1 << ssid;
2778 }
2779
2780 return 0;
2781 }
2782
2783 /**
2784 * cgroup_migrate - migrate a process or task to a cgroup
2785 * @leader: the leader of the process or the task to migrate
2786 * @threadgroup: whether @leader points to the whole process or a single task
2787 * @mgctx: migration context
2788 *
2789 * Migrate a process or task denoted by @leader. If migrating a process,
2790 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2791 * responsible for invoking cgroup_migrate_add_src() and
2792 * cgroup_migrate_prepare_dst() on the targets before invoking this
2793 * function and following up with cgroup_migrate_finish().
2794 *
2795 * As long as a controller's ->can_attach() doesn't fail, this function is
2796 * guaranteed to succeed. This means that, excluding ->can_attach()
2797 * failure, when migrating multiple targets, the success or failure can be
2798 * decided for all targets by invoking group_migrate_prepare_dst() before
2799 * actually starting migrating.
2800 */
cgroup_migrate(struct task_struct * leader,bool threadgroup,struct cgroup_mgctx * mgctx)2801 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2802 struct cgroup_mgctx *mgctx)
2803 {
2804 struct task_struct *task;
2805
2806 /*
2807 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2808 * already PF_EXITING could be freed from underneath us unless we
2809 * take an rcu_read_lock.
2810 */
2811 spin_lock_irq(&css_set_lock);
2812 rcu_read_lock();
2813 task = leader;
2814 do {
2815 cgroup_migrate_add_task(task, mgctx);
2816 if (!threadgroup)
2817 break;
2818 } while_each_thread(leader, task);
2819 rcu_read_unlock();
2820 spin_unlock_irq(&css_set_lock);
2821
2822 return cgroup_migrate_execute(mgctx);
2823 }
2824
2825 /**
2826 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2827 * @dst_cgrp: the cgroup to attach to
2828 * @leader: the task or the leader of the threadgroup to be attached
2829 * @threadgroup: attach the whole threadgroup?
2830 *
2831 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2832 */
cgroup_attach_task(struct cgroup * dst_cgrp,struct task_struct * leader,bool threadgroup)2833 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2834 bool threadgroup)
2835 {
2836 DEFINE_CGROUP_MGCTX(mgctx);
2837 struct task_struct *task;
2838 int ret = 0;
2839
2840 /* look up all src csets */
2841 spin_lock_irq(&css_set_lock);
2842 rcu_read_lock();
2843 task = leader;
2844 do {
2845 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2846 if (!threadgroup)
2847 break;
2848 } while_each_thread(leader, task);
2849 rcu_read_unlock();
2850 spin_unlock_irq(&css_set_lock);
2851
2852 /* prepare dst csets and commit */
2853 ret = cgroup_migrate_prepare_dst(&mgctx);
2854 if (!ret)
2855 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2856
2857 cgroup_migrate_finish(&mgctx);
2858
2859 if (!ret)
2860 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2861
2862 return ret;
2863 }
2864
cgroup_procs_write_start(char * buf,bool threadgroup,bool * threadgroup_locked)2865 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2866 bool *threadgroup_locked)
2867 {
2868 struct task_struct *tsk;
2869 pid_t pid;
2870
2871 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2872 return ERR_PTR(-EINVAL);
2873
2874 /*
2875 * If we migrate a single thread, we don't care about threadgroup
2876 * stability. If the thread is `current`, it won't exit(2) under our
2877 * hands or change PID through exec(2). We exclude
2878 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2879 * callers by cgroup_mutex.
2880 * Therefore, we can skip the global lock.
2881 */
2882 lockdep_assert_held(&cgroup_mutex);
2883 *threadgroup_locked = pid || threadgroup;
2884 cgroup_attach_lock(*threadgroup_locked);
2885
2886 rcu_read_lock();
2887 if (pid) {
2888 tsk = find_task_by_vpid(pid);
2889 if (!tsk) {
2890 tsk = ERR_PTR(-ESRCH);
2891 goto out_unlock_threadgroup;
2892 }
2893 } else {
2894 tsk = current;
2895 }
2896
2897 if (threadgroup)
2898 tsk = tsk->group_leader;
2899
2900 /*
2901 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2902 * If userland migrates such a kthread to a non-root cgroup, it can
2903 * become trapped in a cpuset, or RT kthread may be born in a
2904 * cgroup with no rt_runtime allocated. Just say no.
2905 */
2906 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2907 tsk = ERR_PTR(-EINVAL);
2908 goto out_unlock_threadgroup;
2909 }
2910
2911 get_task_struct(tsk);
2912 goto out_unlock_rcu;
2913
2914 out_unlock_threadgroup:
2915 cgroup_attach_unlock(*threadgroup_locked);
2916 *threadgroup_locked = false;
2917 out_unlock_rcu:
2918 rcu_read_unlock();
2919 return tsk;
2920 }
2921
cgroup_procs_write_finish(struct task_struct * task,bool threadgroup_locked)2922 void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2923 {
2924 struct cgroup_subsys *ss;
2925 int ssid;
2926
2927 /* release reference from cgroup_procs_write_start() */
2928 put_task_struct(task);
2929
2930 cgroup_attach_unlock(threadgroup_locked);
2931
2932 for_each_subsys(ss, ssid)
2933 if (ss->post_attach)
2934 ss->post_attach();
2935 }
2936
cgroup_print_ss_mask(struct seq_file * seq,u16 ss_mask)2937 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2938 {
2939 struct cgroup_subsys *ss;
2940 bool printed = false;
2941 int ssid;
2942
2943 do_each_subsys_mask(ss, ssid, ss_mask) {
2944 if (printed)
2945 seq_putc(seq, ' ');
2946 seq_puts(seq, ss->name);
2947 printed = true;
2948 } while_each_subsys_mask();
2949 if (printed)
2950 seq_putc(seq, '\n');
2951 }
2952
2953 /* show controllers which are enabled from the parent */
cgroup_controllers_show(struct seq_file * seq,void * v)2954 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2955 {
2956 struct cgroup *cgrp = seq_css(seq)->cgroup;
2957
2958 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2959 return 0;
2960 }
2961
2962 /* show controllers which are enabled for a given cgroup's children */
cgroup_subtree_control_show(struct seq_file * seq,void * v)2963 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2964 {
2965 struct cgroup *cgrp = seq_css(seq)->cgroup;
2966
2967 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2968 return 0;
2969 }
2970
2971 /**
2972 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2973 * @cgrp: root of the subtree to update csses for
2974 *
2975 * @cgrp's control masks have changed and its subtree's css associations
2976 * need to be updated accordingly. This function looks up all css_sets
2977 * which are attached to the subtree, creates the matching updated css_sets
2978 * and migrates the tasks to the new ones.
2979 */
cgroup_update_dfl_csses(struct cgroup * cgrp)2980 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2981 {
2982 DEFINE_CGROUP_MGCTX(mgctx);
2983 struct cgroup_subsys_state *d_css;
2984 struct cgroup *dsct;
2985 struct css_set *src_cset;
2986 bool has_tasks;
2987 int ret;
2988
2989 lockdep_assert_held(&cgroup_mutex);
2990
2991 /* look up all csses currently attached to @cgrp's subtree */
2992 spin_lock_irq(&css_set_lock);
2993 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2994 struct cgrp_cset_link *link;
2995
2996 list_for_each_entry(link, &dsct->cset_links, cset_link)
2997 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2998 }
2999 spin_unlock_irq(&css_set_lock);
3000
3001 /*
3002 * We need to write-lock threadgroup_rwsem while migrating tasks.
3003 * However, if there are no source csets for @cgrp, changing its
3004 * controllers isn't gonna produce any task migrations and the
3005 * write-locking can be skipped safely.
3006 */
3007 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3008 cgroup_attach_lock(has_tasks);
3009
3010 /* NULL dst indicates self on default hierarchy */
3011 ret = cgroup_migrate_prepare_dst(&mgctx);
3012 if (ret)
3013 goto out_finish;
3014
3015 spin_lock_irq(&css_set_lock);
3016 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3017 mg_src_preload_node) {
3018 struct task_struct *task, *ntask;
3019
3020 /* all tasks in src_csets need to be migrated */
3021 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3022 cgroup_migrate_add_task(task, &mgctx);
3023 }
3024 spin_unlock_irq(&css_set_lock);
3025
3026 ret = cgroup_migrate_execute(&mgctx);
3027 out_finish:
3028 cgroup_migrate_finish(&mgctx);
3029 cgroup_attach_unlock(has_tasks);
3030 return ret;
3031 }
3032
3033 /**
3034 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3035 * @cgrp: root of the target subtree
3036 *
3037 * Because css offlining is asynchronous, userland may try to re-enable a
3038 * controller while the previous css is still around. This function grabs
3039 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3040 */
cgroup_lock_and_drain_offline(struct cgroup * cgrp)3041 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3042 __acquires(&cgroup_mutex)
3043 {
3044 struct cgroup *dsct;
3045 struct cgroup_subsys_state *d_css;
3046 struct cgroup_subsys *ss;
3047 int ssid;
3048
3049 restart:
3050 mutex_lock(&cgroup_mutex);
3051
3052 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3053 for_each_subsys(ss, ssid) {
3054 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3055 DEFINE_WAIT(wait);
3056
3057 if (!css || !percpu_ref_is_dying(&css->refcnt))
3058 continue;
3059
3060 cgroup_get_live(dsct);
3061 prepare_to_wait(&dsct->offline_waitq, &wait,
3062 TASK_UNINTERRUPTIBLE);
3063
3064 mutex_unlock(&cgroup_mutex);
3065 schedule();
3066 finish_wait(&dsct->offline_waitq, &wait);
3067
3068 cgroup_put(dsct);
3069 goto restart;
3070 }
3071 }
3072 }
3073
3074 /**
3075 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3076 * @cgrp: root of the target subtree
3077 *
3078 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3079 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3080 * itself.
3081 */
cgroup_save_control(struct cgroup * cgrp)3082 static void cgroup_save_control(struct cgroup *cgrp)
3083 {
3084 struct cgroup *dsct;
3085 struct cgroup_subsys_state *d_css;
3086
3087 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3088 dsct->old_subtree_control = dsct->subtree_control;
3089 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3090 dsct->old_dom_cgrp = dsct->dom_cgrp;
3091 }
3092 }
3093
3094 /**
3095 * cgroup_propagate_control - refresh control masks of a subtree
3096 * @cgrp: root of the target subtree
3097 *
3098 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3099 * ->subtree_control and propagate controller availability through the
3100 * subtree so that descendants don't have unavailable controllers enabled.
3101 */
cgroup_propagate_control(struct cgroup * cgrp)3102 static void cgroup_propagate_control(struct cgroup *cgrp)
3103 {
3104 struct cgroup *dsct;
3105 struct cgroup_subsys_state *d_css;
3106
3107 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3108 dsct->subtree_control &= cgroup_control(dsct);
3109 dsct->subtree_ss_mask =
3110 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3111 cgroup_ss_mask(dsct));
3112 }
3113 }
3114
3115 /**
3116 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3117 * @cgrp: root of the target subtree
3118 *
3119 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3120 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3121 * itself.
3122 */
cgroup_restore_control(struct cgroup * cgrp)3123 static void cgroup_restore_control(struct cgroup *cgrp)
3124 {
3125 struct cgroup *dsct;
3126 struct cgroup_subsys_state *d_css;
3127
3128 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3129 dsct->subtree_control = dsct->old_subtree_control;
3130 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3131 dsct->dom_cgrp = dsct->old_dom_cgrp;
3132 }
3133 }
3134
css_visible(struct cgroup_subsys_state * css)3135 static bool css_visible(struct cgroup_subsys_state *css)
3136 {
3137 struct cgroup_subsys *ss = css->ss;
3138 struct cgroup *cgrp = css->cgroup;
3139
3140 if (cgroup_control(cgrp) & (1 << ss->id))
3141 return true;
3142 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3143 return false;
3144 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3145 }
3146
3147 /**
3148 * cgroup_apply_control_enable - enable or show csses according to control
3149 * @cgrp: root of the target subtree
3150 *
3151 * Walk @cgrp's subtree and create new csses or make the existing ones
3152 * visible. A css is created invisible if it's being implicitly enabled
3153 * through dependency. An invisible css is made visible when the userland
3154 * explicitly enables it.
3155 *
3156 * Returns 0 on success, -errno on failure. On failure, csses which have
3157 * been processed already aren't cleaned up. The caller is responsible for
3158 * cleaning up with cgroup_apply_control_disable().
3159 */
cgroup_apply_control_enable(struct cgroup * cgrp)3160 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3161 {
3162 struct cgroup *dsct;
3163 struct cgroup_subsys_state *d_css;
3164 struct cgroup_subsys *ss;
3165 int ssid, ret;
3166
3167 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3168 for_each_subsys(ss, ssid) {
3169 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3170
3171 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3172 continue;
3173
3174 if (!css) {
3175 css = css_create(dsct, ss);
3176 if (IS_ERR(css))
3177 return PTR_ERR(css);
3178 }
3179
3180 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3181
3182 if (css_visible(css)) {
3183 ret = css_populate_dir(css);
3184 if (ret)
3185 return ret;
3186 }
3187 }
3188 }
3189
3190 return 0;
3191 }
3192
3193 /**
3194 * cgroup_apply_control_disable - kill or hide csses according to control
3195 * @cgrp: root of the target subtree
3196 *
3197 * Walk @cgrp's subtree and kill and hide csses so that they match
3198 * cgroup_ss_mask() and cgroup_visible_mask().
3199 *
3200 * A css is hidden when the userland requests it to be disabled while other
3201 * subsystems are still depending on it. The css must not actively control
3202 * resources and be in the vanilla state if it's made visible again later.
3203 * Controllers which may be depended upon should provide ->css_reset() for
3204 * this purpose.
3205 */
cgroup_apply_control_disable(struct cgroup * cgrp)3206 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3207 {
3208 struct cgroup *dsct;
3209 struct cgroup_subsys_state *d_css;
3210 struct cgroup_subsys *ss;
3211 int ssid;
3212
3213 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3214 for_each_subsys(ss, ssid) {
3215 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3216
3217 if (!css)
3218 continue;
3219
3220 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3221
3222 if (css->parent &&
3223 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3224 kill_css(css);
3225 } else if (!css_visible(css)) {
3226 css_clear_dir(css);
3227 if (ss->css_reset)
3228 ss->css_reset(css);
3229 }
3230 }
3231 }
3232 }
3233
3234 /**
3235 * cgroup_apply_control - apply control mask updates to the subtree
3236 * @cgrp: root of the target subtree
3237 *
3238 * subsystems can be enabled and disabled in a subtree using the following
3239 * steps.
3240 *
3241 * 1. Call cgroup_save_control() to stash the current state.
3242 * 2. Update ->subtree_control masks in the subtree as desired.
3243 * 3. Call cgroup_apply_control() to apply the changes.
3244 * 4. Optionally perform other related operations.
3245 * 5. Call cgroup_finalize_control() to finish up.
3246 *
3247 * This function implements step 3 and propagates the mask changes
3248 * throughout @cgrp's subtree, updates csses accordingly and perform
3249 * process migrations.
3250 */
cgroup_apply_control(struct cgroup * cgrp)3251 static int cgroup_apply_control(struct cgroup *cgrp)
3252 {
3253 int ret;
3254
3255 cgroup_propagate_control(cgrp);
3256
3257 ret = cgroup_apply_control_enable(cgrp);
3258 if (ret)
3259 return ret;
3260
3261 /*
3262 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3263 * making the following cgroup_update_dfl_csses() properly update
3264 * css associations of all tasks in the subtree.
3265 */
3266 ret = cgroup_update_dfl_csses(cgrp);
3267 if (ret)
3268 return ret;
3269
3270 return 0;
3271 }
3272
3273 /**
3274 * cgroup_finalize_control - finalize control mask update
3275 * @cgrp: root of the target subtree
3276 * @ret: the result of the update
3277 *
3278 * Finalize control mask update. See cgroup_apply_control() for more info.
3279 */
cgroup_finalize_control(struct cgroup * cgrp,int ret)3280 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3281 {
3282 if (ret) {
3283 cgroup_restore_control(cgrp);
3284 cgroup_propagate_control(cgrp);
3285 }
3286
3287 cgroup_apply_control_disable(cgrp);
3288 }
3289
cgroup_vet_subtree_control_enable(struct cgroup * cgrp,u16 enable)3290 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3291 {
3292 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3293
3294 /* if nothing is getting enabled, nothing to worry about */
3295 if (!enable)
3296 return 0;
3297
3298 /* can @cgrp host any resources? */
3299 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3300 return -EOPNOTSUPP;
3301
3302 /* mixables don't care */
3303 if (cgroup_is_mixable(cgrp))
3304 return 0;
3305
3306 if (domain_enable) {
3307 /* can't enable domain controllers inside a thread subtree */
3308 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3309 return -EOPNOTSUPP;
3310 } else {
3311 /*
3312 * Threaded controllers can handle internal competitions
3313 * and are always allowed inside a (prospective) thread
3314 * subtree.
3315 */
3316 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3317 return 0;
3318 }
3319
3320 /*
3321 * Controllers can't be enabled for a cgroup with tasks to avoid
3322 * child cgroups competing against tasks.
3323 */
3324 if (cgroup_has_tasks(cgrp))
3325 return -EBUSY;
3326
3327 return 0;
3328 }
3329
3330 /* change the enabled child controllers for a cgroup in the default hierarchy */
cgroup_subtree_control_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3331 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3332 char *buf, size_t nbytes,
3333 loff_t off)
3334 {
3335 u16 enable = 0, disable = 0;
3336 struct cgroup *cgrp, *child;
3337 struct cgroup_subsys *ss;
3338 char *tok;
3339 int ssid, ret;
3340
3341 /*
3342 * Parse input - space separated list of subsystem names prefixed
3343 * with either + or -.
3344 */
3345 buf = strstrip(buf);
3346 while ((tok = strsep(&buf, " "))) {
3347 if (tok[0] == '\0')
3348 continue;
3349 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3350 if (!cgroup_ssid_enabled(ssid) ||
3351 strcmp(tok + 1, ss->name))
3352 continue;
3353
3354 if (*tok == '+') {
3355 enable |= 1 << ssid;
3356 disable &= ~(1 << ssid);
3357 } else if (*tok == '-') {
3358 disable |= 1 << ssid;
3359 enable &= ~(1 << ssid);
3360 } else {
3361 return -EINVAL;
3362 }
3363 break;
3364 } while_each_subsys_mask();
3365 if (ssid == CGROUP_SUBSYS_COUNT)
3366 return -EINVAL;
3367 }
3368
3369 cgrp = cgroup_kn_lock_live(of->kn, true);
3370 if (!cgrp)
3371 return -ENODEV;
3372
3373 for_each_subsys(ss, ssid) {
3374 if (enable & (1 << ssid)) {
3375 if (cgrp->subtree_control & (1 << ssid)) {
3376 enable &= ~(1 << ssid);
3377 continue;
3378 }
3379
3380 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3381 ret = -ENOENT;
3382 goto out_unlock;
3383 }
3384 } else if (disable & (1 << ssid)) {
3385 if (!(cgrp->subtree_control & (1 << ssid))) {
3386 disable &= ~(1 << ssid);
3387 continue;
3388 }
3389
3390 /* a child has it enabled? */
3391 cgroup_for_each_live_child(child, cgrp) {
3392 if (child->subtree_control & (1 << ssid)) {
3393 ret = -EBUSY;
3394 goto out_unlock;
3395 }
3396 }
3397 }
3398 }
3399
3400 if (!enable && !disable) {
3401 ret = 0;
3402 goto out_unlock;
3403 }
3404
3405 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3406 if (ret)
3407 goto out_unlock;
3408
3409 /* save and update control masks and prepare csses */
3410 cgroup_save_control(cgrp);
3411
3412 cgrp->subtree_control |= enable;
3413 cgrp->subtree_control &= ~disable;
3414
3415 ret = cgroup_apply_control(cgrp);
3416 cgroup_finalize_control(cgrp, ret);
3417 if (ret)
3418 goto out_unlock;
3419
3420 kernfs_activate(cgrp->kn);
3421 out_unlock:
3422 cgroup_kn_unlock(of->kn);
3423 return ret ?: nbytes;
3424 }
3425
3426 /**
3427 * cgroup_enable_threaded - make @cgrp threaded
3428 * @cgrp: the target cgroup
3429 *
3430 * Called when "threaded" is written to the cgroup.type interface file and
3431 * tries to make @cgrp threaded and join the parent's resource domain.
3432 * This function is never called on the root cgroup as cgroup.type doesn't
3433 * exist on it.
3434 */
cgroup_enable_threaded(struct cgroup * cgrp)3435 static int cgroup_enable_threaded(struct cgroup *cgrp)
3436 {
3437 struct cgroup *parent = cgroup_parent(cgrp);
3438 struct cgroup *dom_cgrp = parent->dom_cgrp;
3439 struct cgroup *dsct;
3440 struct cgroup_subsys_state *d_css;
3441 int ret;
3442
3443 lockdep_assert_held(&cgroup_mutex);
3444
3445 /* noop if already threaded */
3446 if (cgroup_is_threaded(cgrp))
3447 return 0;
3448
3449 /*
3450 * If @cgroup is populated or has domain controllers enabled, it
3451 * can't be switched. While the below cgroup_can_be_thread_root()
3452 * test can catch the same conditions, that's only when @parent is
3453 * not mixable, so let's check it explicitly.
3454 */
3455 if (cgroup_is_populated(cgrp) ||
3456 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3457 return -EOPNOTSUPP;
3458
3459 /* we're joining the parent's domain, ensure its validity */
3460 if (!cgroup_is_valid_domain(dom_cgrp) ||
3461 !cgroup_can_be_thread_root(dom_cgrp))
3462 return -EOPNOTSUPP;
3463
3464 /*
3465 * The following shouldn't cause actual migrations and should
3466 * always succeed.
3467 */
3468 cgroup_save_control(cgrp);
3469
3470 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3471 if (dsct == cgrp || cgroup_is_threaded(dsct))
3472 dsct->dom_cgrp = dom_cgrp;
3473
3474 ret = cgroup_apply_control(cgrp);
3475 if (!ret)
3476 parent->nr_threaded_children++;
3477
3478 cgroup_finalize_control(cgrp, ret);
3479 return ret;
3480 }
3481
cgroup_type_show(struct seq_file * seq,void * v)3482 static int cgroup_type_show(struct seq_file *seq, void *v)
3483 {
3484 struct cgroup *cgrp = seq_css(seq)->cgroup;
3485
3486 if (cgroup_is_threaded(cgrp))
3487 seq_puts(seq, "threaded\n");
3488 else if (!cgroup_is_valid_domain(cgrp))
3489 seq_puts(seq, "domain invalid\n");
3490 else if (cgroup_is_thread_root(cgrp))
3491 seq_puts(seq, "domain threaded\n");
3492 else
3493 seq_puts(seq, "domain\n");
3494
3495 return 0;
3496 }
3497
cgroup_type_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3498 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3499 size_t nbytes, loff_t off)
3500 {
3501 struct cgroup *cgrp;
3502 int ret;
3503
3504 /* only switching to threaded mode is supported */
3505 if (strcmp(strstrip(buf), "threaded"))
3506 return -EINVAL;
3507
3508 /* drain dying csses before we re-apply (threaded) subtree control */
3509 cgrp = cgroup_kn_lock_live(of->kn, true);
3510 if (!cgrp)
3511 return -ENOENT;
3512
3513 /* threaded can only be enabled */
3514 ret = cgroup_enable_threaded(cgrp);
3515
3516 cgroup_kn_unlock(of->kn);
3517 return ret ?: nbytes;
3518 }
3519
cgroup_max_descendants_show(struct seq_file * seq,void * v)3520 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3521 {
3522 struct cgroup *cgrp = seq_css(seq)->cgroup;
3523 int descendants = READ_ONCE(cgrp->max_descendants);
3524
3525 if (descendants == INT_MAX)
3526 seq_puts(seq, "max\n");
3527 else
3528 seq_printf(seq, "%d\n", descendants);
3529
3530 return 0;
3531 }
3532
cgroup_max_descendants_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3533 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3534 char *buf, size_t nbytes, loff_t off)
3535 {
3536 struct cgroup *cgrp;
3537 int descendants;
3538 ssize_t ret;
3539
3540 buf = strstrip(buf);
3541 if (!strcmp(buf, "max")) {
3542 descendants = INT_MAX;
3543 } else {
3544 ret = kstrtoint(buf, 0, &descendants);
3545 if (ret)
3546 return ret;
3547 }
3548
3549 if (descendants < 0)
3550 return -ERANGE;
3551
3552 cgrp = cgroup_kn_lock_live(of->kn, false);
3553 if (!cgrp)
3554 return -ENOENT;
3555
3556 cgrp->max_descendants = descendants;
3557
3558 cgroup_kn_unlock(of->kn);
3559
3560 return nbytes;
3561 }
3562
cgroup_max_depth_show(struct seq_file * seq,void * v)3563 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3564 {
3565 struct cgroup *cgrp = seq_css(seq)->cgroup;
3566 int depth = READ_ONCE(cgrp->max_depth);
3567
3568 if (depth == INT_MAX)
3569 seq_puts(seq, "max\n");
3570 else
3571 seq_printf(seq, "%d\n", depth);
3572
3573 return 0;
3574 }
3575
cgroup_max_depth_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3576 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3577 char *buf, size_t nbytes, loff_t off)
3578 {
3579 struct cgroup *cgrp;
3580 ssize_t ret;
3581 int depth;
3582
3583 buf = strstrip(buf);
3584 if (!strcmp(buf, "max")) {
3585 depth = INT_MAX;
3586 } else {
3587 ret = kstrtoint(buf, 0, &depth);
3588 if (ret)
3589 return ret;
3590 }
3591
3592 if (depth < 0)
3593 return -ERANGE;
3594
3595 cgrp = cgroup_kn_lock_live(of->kn, false);
3596 if (!cgrp)
3597 return -ENOENT;
3598
3599 cgrp->max_depth = depth;
3600
3601 cgroup_kn_unlock(of->kn);
3602
3603 return nbytes;
3604 }
3605
cgroup_events_show(struct seq_file * seq,void * v)3606 static int cgroup_events_show(struct seq_file *seq, void *v)
3607 {
3608 struct cgroup *cgrp = seq_css(seq)->cgroup;
3609
3610 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3611 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3612
3613 return 0;
3614 }
3615
cgroup_stat_show(struct seq_file * seq,void * v)3616 static int cgroup_stat_show(struct seq_file *seq, void *v)
3617 {
3618 struct cgroup *cgroup = seq_css(seq)->cgroup;
3619
3620 seq_printf(seq, "nr_descendants %d\n",
3621 cgroup->nr_descendants);
3622 seq_printf(seq, "nr_dying_descendants %d\n",
3623 cgroup->nr_dying_descendants);
3624
3625 return 0;
3626 }
3627
cgroup_extra_stat_show(struct seq_file * seq,struct cgroup * cgrp,int ssid)3628 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3629 struct cgroup *cgrp, int ssid)
3630 {
3631 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3632 struct cgroup_subsys_state *css;
3633 int ret;
3634
3635 if (!ss->css_extra_stat_show)
3636 return 0;
3637
3638 css = cgroup_tryget_css(cgrp, ss);
3639 if (!css)
3640 return 0;
3641
3642 ret = ss->css_extra_stat_show(seq, css);
3643 css_put(css);
3644 return ret;
3645 }
3646
cpu_stat_show(struct seq_file * seq,void * v)3647 static int cpu_stat_show(struct seq_file *seq, void *v)
3648 {
3649 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3650 int ret = 0;
3651
3652 cgroup_base_stat_cputime_show(seq);
3653 #ifdef CONFIG_CGROUP_SCHED
3654 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3655 #endif
3656 return ret;
3657 }
3658
3659 #ifdef CONFIG_PSI
cgroup_io_pressure_show(struct seq_file * seq,void * v)3660 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3661 {
3662 struct cgroup *cgrp = seq_css(seq)->cgroup;
3663 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3664
3665 return psi_show(seq, psi, PSI_IO);
3666 }
cgroup_memory_pressure_show(struct seq_file * seq,void * v)3667 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3668 {
3669 struct cgroup *cgrp = seq_css(seq)->cgroup;
3670 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3671
3672 return psi_show(seq, psi, PSI_MEM);
3673 }
cgroup_cpu_pressure_show(struct seq_file * seq,void * v)3674 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3675 {
3676 struct cgroup *cgrp = seq_css(seq)->cgroup;
3677 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3678
3679 return psi_show(seq, psi, PSI_CPU);
3680 }
3681
cgroup_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,enum psi_res res)3682 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3683 size_t nbytes, enum psi_res res)
3684 {
3685 struct cgroup_file_ctx *ctx = of->priv;
3686 struct psi_trigger *new;
3687 struct cgroup *cgrp;
3688 struct psi_group *psi;
3689
3690 cgrp = cgroup_kn_lock_live(of->kn, false);
3691 if (!cgrp)
3692 return -ENODEV;
3693
3694 cgroup_get(cgrp);
3695 cgroup_kn_unlock(of->kn);
3696
3697 /* Allow only one trigger per file descriptor */
3698 if (ctx->psi.trigger) {
3699 cgroup_put(cgrp);
3700 return -EBUSY;
3701 }
3702
3703 psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3704 new = psi_trigger_create(psi, buf, nbytes, res);
3705 if (IS_ERR(new)) {
3706 cgroup_put(cgrp);
3707 return PTR_ERR(new);
3708 }
3709
3710 smp_store_release(&ctx->psi.trigger, new);
3711 cgroup_put(cgrp);
3712
3713 return nbytes;
3714 }
3715
cgroup_io_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3716 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3717 char *buf, size_t nbytes,
3718 loff_t off)
3719 {
3720 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3721 }
3722
cgroup_memory_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3723 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3724 char *buf, size_t nbytes,
3725 loff_t off)
3726 {
3727 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3728 }
3729
cgroup_cpu_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3730 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3731 char *buf, size_t nbytes,
3732 loff_t off)
3733 {
3734 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3735 }
3736
cgroup_pressure_poll(struct kernfs_open_file * of,poll_table * pt)3737 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3738 poll_table *pt)
3739 {
3740 struct cgroup_file_ctx *ctx = of->priv;
3741
3742 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3743 }
3744
cgroup_pressure_release(struct kernfs_open_file * of)3745 static void cgroup_pressure_release(struct kernfs_open_file *of)
3746 {
3747 struct cgroup_file_ctx *ctx = of->priv;
3748
3749 psi_trigger_destroy(ctx->psi.trigger);
3750 }
3751
cgroup_psi_enabled(void)3752 bool cgroup_psi_enabled(void)
3753 {
3754 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3755 }
3756
3757 #else /* CONFIG_PSI */
cgroup_psi_enabled(void)3758 bool cgroup_psi_enabled(void)
3759 {
3760 return false;
3761 }
3762
3763 #endif /* CONFIG_PSI */
3764
cgroup_freeze_show(struct seq_file * seq,void * v)3765 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3766 {
3767 struct cgroup *cgrp = seq_css(seq)->cgroup;
3768
3769 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3770
3771 return 0;
3772 }
3773
cgroup_freeze_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3774 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3775 char *buf, size_t nbytes, loff_t off)
3776 {
3777 struct cgroup *cgrp;
3778 ssize_t ret;
3779 int freeze;
3780
3781 ret = kstrtoint(strstrip(buf), 0, &freeze);
3782 if (ret)
3783 return ret;
3784
3785 if (freeze < 0 || freeze > 1)
3786 return -ERANGE;
3787
3788 cgrp = cgroup_kn_lock_live(of->kn, false);
3789 if (!cgrp)
3790 return -ENOENT;
3791
3792 cgroup_freeze(cgrp, freeze);
3793
3794 cgroup_kn_unlock(of->kn);
3795
3796 return nbytes;
3797 }
3798
__cgroup_kill(struct cgroup * cgrp)3799 static void __cgroup_kill(struct cgroup *cgrp)
3800 {
3801 struct css_task_iter it;
3802 struct task_struct *task;
3803
3804 lockdep_assert_held(&cgroup_mutex);
3805
3806 spin_lock_irq(&css_set_lock);
3807 set_bit(CGRP_KILL, &cgrp->flags);
3808 spin_unlock_irq(&css_set_lock);
3809
3810 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3811 while ((task = css_task_iter_next(&it))) {
3812 /* Ignore kernel threads here. */
3813 if (task->flags & PF_KTHREAD)
3814 continue;
3815
3816 /* Skip tasks that are already dying. */
3817 if (__fatal_signal_pending(task))
3818 continue;
3819
3820 send_sig(SIGKILL, task, 0);
3821 }
3822 css_task_iter_end(&it);
3823
3824 spin_lock_irq(&css_set_lock);
3825 clear_bit(CGRP_KILL, &cgrp->flags);
3826 spin_unlock_irq(&css_set_lock);
3827 }
3828
cgroup_kill(struct cgroup * cgrp)3829 static void cgroup_kill(struct cgroup *cgrp)
3830 {
3831 struct cgroup_subsys_state *css;
3832 struct cgroup *dsct;
3833
3834 lockdep_assert_held(&cgroup_mutex);
3835
3836 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3837 __cgroup_kill(dsct);
3838 }
3839
cgroup_kill_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3840 static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3841 size_t nbytes, loff_t off)
3842 {
3843 ssize_t ret = 0;
3844 int kill;
3845 struct cgroup *cgrp;
3846
3847 ret = kstrtoint(strstrip(buf), 0, &kill);
3848 if (ret)
3849 return ret;
3850
3851 if (kill != 1)
3852 return -ERANGE;
3853
3854 cgrp = cgroup_kn_lock_live(of->kn, false);
3855 if (!cgrp)
3856 return -ENOENT;
3857
3858 /*
3859 * Killing is a process directed operation, i.e. the whole thread-group
3860 * is taken down so act like we do for cgroup.procs and only make this
3861 * writable in non-threaded cgroups.
3862 */
3863 if (cgroup_is_threaded(cgrp))
3864 ret = -EOPNOTSUPP;
3865 else
3866 cgroup_kill(cgrp);
3867
3868 cgroup_kn_unlock(of->kn);
3869
3870 return ret ?: nbytes;
3871 }
3872
cgroup_file_open(struct kernfs_open_file * of)3873 static int cgroup_file_open(struct kernfs_open_file *of)
3874 {
3875 struct cftype *cft = of_cft(of);
3876 struct cgroup_file_ctx *ctx;
3877 int ret;
3878
3879 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
3880 if (!ctx)
3881 return -ENOMEM;
3882
3883 ctx->ns = current->nsproxy->cgroup_ns;
3884 get_cgroup_ns(ctx->ns);
3885 of->priv = ctx;
3886
3887 if (!cft->open)
3888 return 0;
3889
3890 ret = cft->open(of);
3891 if (ret) {
3892 put_cgroup_ns(ctx->ns);
3893 kfree(ctx);
3894 }
3895 return ret;
3896 }
3897
cgroup_file_release(struct kernfs_open_file * of)3898 static void cgroup_file_release(struct kernfs_open_file *of)
3899 {
3900 struct cftype *cft = of_cft(of);
3901 struct cgroup_file_ctx *ctx = of->priv;
3902
3903 if (cft->release)
3904 cft->release(of);
3905 put_cgroup_ns(ctx->ns);
3906 kfree(ctx);
3907 }
3908
cgroup_file_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3909 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3910 size_t nbytes, loff_t off)
3911 {
3912 struct cgroup_file_ctx *ctx = of->priv;
3913 struct cgroup *cgrp = of->kn->parent->priv;
3914 struct cftype *cft = of_cft(of);
3915 struct cgroup_subsys_state *css;
3916 int ret;
3917
3918 if (!nbytes)
3919 return 0;
3920
3921 /*
3922 * If namespaces are delegation boundaries, disallow writes to
3923 * files in an non-init namespace root from inside the namespace
3924 * except for the files explicitly marked delegatable -
3925 * cgroup.procs and cgroup.subtree_control.
3926 */
3927 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3928 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3929 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
3930 return -EPERM;
3931
3932 if (cft->write)
3933 return cft->write(of, buf, nbytes, off);
3934
3935 /*
3936 * kernfs guarantees that a file isn't deleted with operations in
3937 * flight, which means that the matching css is and stays alive and
3938 * doesn't need to be pinned. The RCU locking is not necessary
3939 * either. It's just for the convenience of using cgroup_css().
3940 */
3941 rcu_read_lock();
3942 css = cgroup_css(cgrp, cft->ss);
3943 rcu_read_unlock();
3944
3945 if (cft->write_u64) {
3946 unsigned long long v;
3947 ret = kstrtoull(buf, 0, &v);
3948 if (!ret)
3949 ret = cft->write_u64(css, cft, v);
3950 } else if (cft->write_s64) {
3951 long long v;
3952 ret = kstrtoll(buf, 0, &v);
3953 if (!ret)
3954 ret = cft->write_s64(css, cft, v);
3955 } else {
3956 ret = -EINVAL;
3957 }
3958
3959 return ret ?: nbytes;
3960 }
3961
cgroup_file_poll(struct kernfs_open_file * of,poll_table * pt)3962 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3963 {
3964 struct cftype *cft = of_cft(of);
3965
3966 if (cft->poll)
3967 return cft->poll(of, pt);
3968
3969 return kernfs_generic_poll(of, pt);
3970 }
3971
cgroup_seqfile_start(struct seq_file * seq,loff_t * ppos)3972 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3973 {
3974 return seq_cft(seq)->seq_start(seq, ppos);
3975 }
3976
cgroup_seqfile_next(struct seq_file * seq,void * v,loff_t * ppos)3977 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3978 {
3979 return seq_cft(seq)->seq_next(seq, v, ppos);
3980 }
3981
cgroup_seqfile_stop(struct seq_file * seq,void * v)3982 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3983 {
3984 if (seq_cft(seq)->seq_stop)
3985 seq_cft(seq)->seq_stop(seq, v);
3986 }
3987
cgroup_seqfile_show(struct seq_file * m,void * arg)3988 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3989 {
3990 struct cftype *cft = seq_cft(m);
3991 struct cgroup_subsys_state *css = seq_css(m);
3992
3993 if (cft->seq_show)
3994 return cft->seq_show(m, arg);
3995
3996 if (cft->read_u64)
3997 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3998 else if (cft->read_s64)
3999 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4000 else
4001 return -EINVAL;
4002 return 0;
4003 }
4004
4005 static struct kernfs_ops cgroup_kf_single_ops = {
4006 .atomic_write_len = PAGE_SIZE,
4007 .open = cgroup_file_open,
4008 .release = cgroup_file_release,
4009 .write = cgroup_file_write,
4010 .poll = cgroup_file_poll,
4011 .seq_show = cgroup_seqfile_show,
4012 };
4013
4014 static struct kernfs_ops cgroup_kf_ops = {
4015 .atomic_write_len = PAGE_SIZE,
4016 .open = cgroup_file_open,
4017 .release = cgroup_file_release,
4018 .write = cgroup_file_write,
4019 .poll = cgroup_file_poll,
4020 .seq_start = cgroup_seqfile_start,
4021 .seq_next = cgroup_seqfile_next,
4022 .seq_stop = cgroup_seqfile_stop,
4023 .seq_show = cgroup_seqfile_show,
4024 };
4025
4026 /* set uid and gid of cgroup dirs and files to that of the creator */
cgroup_kn_set_ugid(struct kernfs_node * kn)4027 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
4028 {
4029 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
4030 .ia_uid = current_fsuid(),
4031 .ia_gid = current_fsgid(), };
4032
4033 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
4034 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
4035 return 0;
4036
4037 return kernfs_setattr(kn, &iattr);
4038 }
4039
cgroup_file_notify_timer(struct timer_list * timer)4040 static void cgroup_file_notify_timer(struct timer_list *timer)
4041 {
4042 cgroup_file_notify(container_of(timer, struct cgroup_file,
4043 notify_timer));
4044 }
4045
cgroup_add_file(struct cgroup_subsys_state * css,struct cgroup * cgrp,struct cftype * cft)4046 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4047 struct cftype *cft)
4048 {
4049 char name[CGROUP_FILE_NAME_MAX];
4050 struct kernfs_node *kn;
4051 struct lock_class_key *key = NULL;
4052 int ret;
4053
4054 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4055 key = &cft->lockdep_key;
4056 #endif
4057 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4058 cgroup_file_mode(cft),
4059 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
4060 0, cft->kf_ops, cft,
4061 NULL, key);
4062 if (IS_ERR(kn))
4063 return PTR_ERR(kn);
4064
4065 ret = cgroup_kn_set_ugid(kn);
4066 if (ret) {
4067 kernfs_remove(kn);
4068 return ret;
4069 }
4070
4071 if (cft->file_offset) {
4072 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4073
4074 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4075
4076 spin_lock_irq(&cgroup_file_kn_lock);
4077 cfile->kn = kn;
4078 spin_unlock_irq(&cgroup_file_kn_lock);
4079 }
4080
4081 return 0;
4082 }
4083
4084 /**
4085 * cgroup_addrm_files - add or remove files to a cgroup directory
4086 * @css: the target css
4087 * @cgrp: the target cgroup (usually css->cgroup)
4088 * @cfts: array of cftypes to be added
4089 * @is_add: whether to add or remove
4090 *
4091 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4092 * For removals, this function never fails.
4093 */
cgroup_addrm_files(struct cgroup_subsys_state * css,struct cgroup * cgrp,struct cftype cfts[],bool is_add)4094 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4095 struct cgroup *cgrp, struct cftype cfts[],
4096 bool is_add)
4097 {
4098 struct cftype *cft, *cft_end = NULL;
4099 int ret = 0;
4100
4101 lockdep_assert_held(&cgroup_mutex);
4102
4103 restart:
4104 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4105 /* does cft->flags tell us to skip this file on @cgrp? */
4106 if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
4107 continue;
4108 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4109 continue;
4110 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4111 continue;
4112 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4113 continue;
4114 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4115 continue;
4116 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4117 continue;
4118 if (is_add) {
4119 ret = cgroup_add_file(css, cgrp, cft);
4120 if (ret) {
4121 pr_warn("%s: failed to add %s, err=%d\n",
4122 __func__, cft->name, ret);
4123 cft_end = cft;
4124 is_add = false;
4125 goto restart;
4126 }
4127 } else {
4128 cgroup_rm_file(cgrp, cft);
4129 }
4130 }
4131 return ret;
4132 }
4133
cgroup_apply_cftypes(struct cftype * cfts,bool is_add)4134 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4135 {
4136 struct cgroup_subsys *ss = cfts[0].ss;
4137 struct cgroup *root = &ss->root->cgrp;
4138 struct cgroup_subsys_state *css;
4139 int ret = 0;
4140
4141 lockdep_assert_held(&cgroup_mutex);
4142
4143 /* add/rm files for all cgroups created before */
4144 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4145 struct cgroup *cgrp = css->cgroup;
4146
4147 if (!(css->flags & CSS_VISIBLE))
4148 continue;
4149
4150 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4151 if (ret)
4152 break;
4153 }
4154
4155 if (is_add && !ret)
4156 kernfs_activate(root->kn);
4157 return ret;
4158 }
4159
cgroup_exit_cftypes(struct cftype * cfts)4160 static void cgroup_exit_cftypes(struct cftype *cfts)
4161 {
4162 struct cftype *cft;
4163
4164 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4165 /* free copy for custom atomic_write_len, see init_cftypes() */
4166 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4167 kfree(cft->kf_ops);
4168 cft->kf_ops = NULL;
4169 cft->ss = NULL;
4170
4171 /* revert flags set by cgroup core while adding @cfts */
4172 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
4173 }
4174 }
4175
cgroup_init_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4176 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4177 {
4178 struct cftype *cft;
4179
4180 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4181 struct kernfs_ops *kf_ops;
4182
4183 WARN_ON(cft->ss || cft->kf_ops);
4184
4185 if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
4186 continue;
4187
4188 if (cft->seq_start)
4189 kf_ops = &cgroup_kf_ops;
4190 else
4191 kf_ops = &cgroup_kf_single_ops;
4192
4193 /*
4194 * Ugh... if @cft wants a custom max_write_len, we need to
4195 * make a copy of kf_ops to set its atomic_write_len.
4196 */
4197 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4198 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4199 if (!kf_ops) {
4200 cgroup_exit_cftypes(cfts);
4201 return -ENOMEM;
4202 }
4203 kf_ops->atomic_write_len = cft->max_write_len;
4204 }
4205
4206 cft->kf_ops = kf_ops;
4207 cft->ss = ss;
4208 }
4209
4210 return 0;
4211 }
4212
cgroup_rm_cftypes_locked(struct cftype * cfts)4213 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4214 {
4215 lockdep_assert_held(&cgroup_mutex);
4216
4217 if (!cfts || !cfts[0].ss)
4218 return -ENOENT;
4219
4220 list_del(&cfts->node);
4221 cgroup_apply_cftypes(cfts, false);
4222 cgroup_exit_cftypes(cfts);
4223 return 0;
4224 }
4225
4226 /**
4227 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4228 * @cfts: zero-length name terminated array of cftypes
4229 *
4230 * Unregister @cfts. Files described by @cfts are removed from all
4231 * existing cgroups and all future cgroups won't have them either. This
4232 * function can be called anytime whether @cfts' subsys is attached or not.
4233 *
4234 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4235 * registered.
4236 */
cgroup_rm_cftypes(struct cftype * cfts)4237 int cgroup_rm_cftypes(struct cftype *cfts)
4238 {
4239 int ret;
4240
4241 mutex_lock(&cgroup_mutex);
4242 ret = cgroup_rm_cftypes_locked(cfts);
4243 mutex_unlock(&cgroup_mutex);
4244 return ret;
4245 }
4246
4247 /**
4248 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4249 * @ss: target cgroup subsystem
4250 * @cfts: zero-length name terminated array of cftypes
4251 *
4252 * Register @cfts to @ss. Files described by @cfts are created for all
4253 * existing cgroups to which @ss is attached and all future cgroups will
4254 * have them too. This function can be called anytime whether @ss is
4255 * attached or not.
4256 *
4257 * Returns 0 on successful registration, -errno on failure. Note that this
4258 * function currently returns 0 as long as @cfts registration is successful
4259 * even if some file creation attempts on existing cgroups fail.
4260 */
cgroup_add_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4261 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4262 {
4263 int ret;
4264
4265 if (!cgroup_ssid_enabled(ss->id))
4266 return 0;
4267
4268 if (!cfts || cfts[0].name[0] == '\0')
4269 return 0;
4270
4271 ret = cgroup_init_cftypes(ss, cfts);
4272 if (ret)
4273 return ret;
4274
4275 mutex_lock(&cgroup_mutex);
4276
4277 list_add_tail(&cfts->node, &ss->cfts);
4278 ret = cgroup_apply_cftypes(cfts, true);
4279 if (ret)
4280 cgroup_rm_cftypes_locked(cfts);
4281
4282 mutex_unlock(&cgroup_mutex);
4283 return ret;
4284 }
4285
4286 /**
4287 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4288 * @ss: target cgroup subsystem
4289 * @cfts: zero-length name terminated array of cftypes
4290 *
4291 * Similar to cgroup_add_cftypes() but the added files are only used for
4292 * the default hierarchy.
4293 */
cgroup_add_dfl_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4294 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4295 {
4296 struct cftype *cft;
4297
4298 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4299 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4300 return cgroup_add_cftypes(ss, cfts);
4301 }
4302
4303 /**
4304 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4305 * @ss: target cgroup subsystem
4306 * @cfts: zero-length name terminated array of cftypes
4307 *
4308 * Similar to cgroup_add_cftypes() but the added files are only used for
4309 * the legacy hierarchies.
4310 */
cgroup_add_legacy_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4311 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4312 {
4313 struct cftype *cft;
4314
4315 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4316 cft->flags |= __CFTYPE_NOT_ON_DFL;
4317 return cgroup_add_cftypes(ss, cfts);
4318 }
4319
4320 /**
4321 * cgroup_file_notify - generate a file modified event for a cgroup_file
4322 * @cfile: target cgroup_file
4323 *
4324 * @cfile must have been obtained by setting cftype->file_offset.
4325 */
cgroup_file_notify(struct cgroup_file * cfile)4326 void cgroup_file_notify(struct cgroup_file *cfile)
4327 {
4328 unsigned long flags;
4329
4330 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4331 if (cfile->kn) {
4332 unsigned long last = cfile->notified_at;
4333 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4334
4335 if (time_in_range(jiffies, last, next)) {
4336 timer_reduce(&cfile->notify_timer, next);
4337 } else {
4338 kernfs_notify(cfile->kn);
4339 cfile->notified_at = jiffies;
4340 }
4341 }
4342 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4343 }
4344
4345 /**
4346 * css_next_child - find the next child of a given css
4347 * @pos: the current position (%NULL to initiate traversal)
4348 * @parent: css whose children to walk
4349 *
4350 * This function returns the next child of @parent and should be called
4351 * under either cgroup_mutex or RCU read lock. The only requirement is
4352 * that @parent and @pos are accessible. The next sibling is guaranteed to
4353 * be returned regardless of their states.
4354 *
4355 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4356 * css which finished ->css_online() is guaranteed to be visible in the
4357 * future iterations and will stay visible until the last reference is put.
4358 * A css which hasn't finished ->css_online() or already finished
4359 * ->css_offline() may show up during traversal. It's each subsystem's
4360 * responsibility to synchronize against on/offlining.
4361 */
css_next_child(struct cgroup_subsys_state * pos,struct cgroup_subsys_state * parent)4362 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4363 struct cgroup_subsys_state *parent)
4364 {
4365 struct cgroup_subsys_state *next;
4366
4367 cgroup_assert_mutex_or_rcu_locked();
4368
4369 /*
4370 * @pos could already have been unlinked from the sibling list.
4371 * Once a cgroup is removed, its ->sibling.next is no longer
4372 * updated when its next sibling changes. CSS_RELEASED is set when
4373 * @pos is taken off list, at which time its next pointer is valid,
4374 * and, as releases are serialized, the one pointed to by the next
4375 * pointer is guaranteed to not have started release yet. This
4376 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4377 * critical section, the one pointed to by its next pointer is
4378 * guaranteed to not have finished its RCU grace period even if we
4379 * have dropped rcu_read_lock() in-between iterations.
4380 *
4381 * If @pos has CSS_RELEASED set, its next pointer can't be
4382 * dereferenced; however, as each css is given a monotonically
4383 * increasing unique serial number and always appended to the
4384 * sibling list, the next one can be found by walking the parent's
4385 * children until the first css with higher serial number than
4386 * @pos's. While this path can be slower, it happens iff iteration
4387 * races against release and the race window is very small.
4388 */
4389 if (!pos) {
4390 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4391 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4392 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4393 } else {
4394 list_for_each_entry_rcu(next, &parent->children, sibling,
4395 lockdep_is_held(&cgroup_mutex))
4396 if (next->serial_nr > pos->serial_nr)
4397 break;
4398 }
4399
4400 /*
4401 * @next, if not pointing to the head, can be dereferenced and is
4402 * the next sibling.
4403 */
4404 if (&next->sibling != &parent->children)
4405 return next;
4406 return NULL;
4407 }
4408
4409 /**
4410 * css_next_descendant_pre - find the next descendant for pre-order walk
4411 * @pos: the current position (%NULL to initiate traversal)
4412 * @root: css whose descendants to walk
4413 *
4414 * To be used by css_for_each_descendant_pre(). Find the next descendant
4415 * to visit for pre-order traversal of @root's descendants. @root is
4416 * included in the iteration and the first node to be visited.
4417 *
4418 * While this function requires cgroup_mutex or RCU read locking, it
4419 * doesn't require the whole traversal to be contained in a single critical
4420 * section. This function will return the correct next descendant as long
4421 * as both @pos and @root are accessible and @pos is a descendant of @root.
4422 *
4423 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4424 * css which finished ->css_online() is guaranteed to be visible in the
4425 * future iterations and will stay visible until the last reference is put.
4426 * A css which hasn't finished ->css_online() or already finished
4427 * ->css_offline() may show up during traversal. It's each subsystem's
4428 * responsibility to synchronize against on/offlining.
4429 */
4430 struct cgroup_subsys_state *
css_next_descendant_pre(struct cgroup_subsys_state * pos,struct cgroup_subsys_state * root)4431 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4432 struct cgroup_subsys_state *root)
4433 {
4434 struct cgroup_subsys_state *next;
4435
4436 cgroup_assert_mutex_or_rcu_locked();
4437
4438 /* if first iteration, visit @root */
4439 if (!pos)
4440 return root;
4441
4442 /* visit the first child if exists */
4443 next = css_next_child(NULL, pos);
4444 if (next)
4445 return next;
4446
4447 /* no child, visit my or the closest ancestor's next sibling */
4448 while (pos != root) {
4449 next = css_next_child(pos, pos->parent);
4450 if (next)
4451 return next;
4452 pos = pos->parent;
4453 }
4454
4455 return NULL;
4456 }
4457 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4458
4459 /**
4460 * css_rightmost_descendant - return the rightmost descendant of a css
4461 * @pos: css of interest
4462 *
4463 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4464 * is returned. This can be used during pre-order traversal to skip
4465 * subtree of @pos.
4466 *
4467 * While this function requires cgroup_mutex or RCU read locking, it
4468 * doesn't require the whole traversal to be contained in a single critical
4469 * section. This function will return the correct rightmost descendant as
4470 * long as @pos is accessible.
4471 */
4472 struct cgroup_subsys_state *
css_rightmost_descendant(struct cgroup_subsys_state * pos)4473 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4474 {
4475 struct cgroup_subsys_state *last, *tmp;
4476
4477 cgroup_assert_mutex_or_rcu_locked();
4478
4479 do {
4480 last = pos;
4481 /* ->prev isn't RCU safe, walk ->next till the end */
4482 pos = NULL;
4483 css_for_each_child(tmp, last)
4484 pos = tmp;
4485 } while (pos);
4486
4487 return last;
4488 }
4489
4490 static struct cgroup_subsys_state *
css_leftmost_descendant(struct cgroup_subsys_state * pos)4491 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4492 {
4493 struct cgroup_subsys_state *last;
4494
4495 do {
4496 last = pos;
4497 pos = css_next_child(NULL, pos);
4498 } while (pos);
4499
4500 return last;
4501 }
4502
4503 /**
4504 * css_next_descendant_post - find the next descendant for post-order walk
4505 * @pos: the current position (%NULL to initiate traversal)
4506 * @root: css whose descendants to walk
4507 *
4508 * To be used by css_for_each_descendant_post(). Find the next descendant
4509 * to visit for post-order traversal of @root's descendants. @root is
4510 * included in the iteration and the last node to be visited.
4511 *
4512 * While this function requires cgroup_mutex or RCU read locking, it
4513 * doesn't require the whole traversal to be contained in a single critical
4514 * section. This function will return the correct next descendant as long
4515 * as both @pos and @cgroup are accessible and @pos is a descendant of
4516 * @cgroup.
4517 *
4518 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4519 * css which finished ->css_online() is guaranteed to be visible in the
4520 * future iterations and will stay visible until the last reference is put.
4521 * A css which hasn't finished ->css_online() or already finished
4522 * ->css_offline() may show up during traversal. It's each subsystem's
4523 * responsibility to synchronize against on/offlining.
4524 */
4525 struct cgroup_subsys_state *
css_next_descendant_post(struct cgroup_subsys_state * pos,struct cgroup_subsys_state * root)4526 css_next_descendant_post(struct cgroup_subsys_state *pos,
4527 struct cgroup_subsys_state *root)
4528 {
4529 struct cgroup_subsys_state *next;
4530
4531 cgroup_assert_mutex_or_rcu_locked();
4532
4533 /* if first iteration, visit leftmost descendant which may be @root */
4534 if (!pos)
4535 return css_leftmost_descendant(root);
4536
4537 /* if we visited @root, we're done */
4538 if (pos == root)
4539 return NULL;
4540
4541 /* if there's an unvisited sibling, visit its leftmost descendant */
4542 next = css_next_child(pos, pos->parent);
4543 if (next)
4544 return css_leftmost_descendant(next);
4545
4546 /* no sibling left, visit parent */
4547 return pos->parent;
4548 }
4549
4550 /**
4551 * css_has_online_children - does a css have online children
4552 * @css: the target css
4553 *
4554 * Returns %true if @css has any online children; otherwise, %false. This
4555 * function can be called from any context but the caller is responsible
4556 * for synchronizing against on/offlining as necessary.
4557 */
css_has_online_children(struct cgroup_subsys_state * css)4558 bool css_has_online_children(struct cgroup_subsys_state *css)
4559 {
4560 struct cgroup_subsys_state *child;
4561 bool ret = false;
4562
4563 rcu_read_lock();
4564 css_for_each_child(child, css) {
4565 if (child->flags & CSS_ONLINE) {
4566 ret = true;
4567 break;
4568 }
4569 }
4570 rcu_read_unlock();
4571 return ret;
4572 }
4573
css_task_iter_next_css_set(struct css_task_iter * it)4574 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4575 {
4576 struct list_head *l;
4577 struct cgrp_cset_link *link;
4578 struct css_set *cset;
4579
4580 lockdep_assert_held(&css_set_lock);
4581
4582 /* find the next threaded cset */
4583 if (it->tcset_pos) {
4584 l = it->tcset_pos->next;
4585
4586 if (l != it->tcset_head) {
4587 it->tcset_pos = l;
4588 return container_of(l, struct css_set,
4589 threaded_csets_node);
4590 }
4591
4592 it->tcset_pos = NULL;
4593 }
4594
4595 /* find the next cset */
4596 l = it->cset_pos;
4597 l = l->next;
4598 if (l == it->cset_head) {
4599 it->cset_pos = NULL;
4600 return NULL;
4601 }
4602
4603 if (it->ss) {
4604 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4605 } else {
4606 link = list_entry(l, struct cgrp_cset_link, cset_link);
4607 cset = link->cset;
4608 }
4609
4610 it->cset_pos = l;
4611
4612 /* initialize threaded css_set walking */
4613 if (it->flags & CSS_TASK_ITER_THREADED) {
4614 if (it->cur_dcset)
4615 put_css_set_locked(it->cur_dcset);
4616 it->cur_dcset = cset;
4617 get_css_set(cset);
4618
4619 it->tcset_head = &cset->threaded_csets;
4620 it->tcset_pos = &cset->threaded_csets;
4621 }
4622
4623 return cset;
4624 }
4625
4626 /**
4627 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4628 * @it: the iterator to advance
4629 *
4630 * Advance @it to the next css_set to walk.
4631 */
css_task_iter_advance_css_set(struct css_task_iter * it)4632 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4633 {
4634 struct css_set *cset;
4635
4636 lockdep_assert_held(&css_set_lock);
4637
4638 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4639 while ((cset = css_task_iter_next_css_set(it))) {
4640 if (!list_empty(&cset->tasks)) {
4641 it->cur_tasks_head = &cset->tasks;
4642 break;
4643 } else if (!list_empty(&cset->mg_tasks)) {
4644 it->cur_tasks_head = &cset->mg_tasks;
4645 break;
4646 } else if (!list_empty(&cset->dying_tasks)) {
4647 it->cur_tasks_head = &cset->dying_tasks;
4648 break;
4649 }
4650 }
4651 if (!cset) {
4652 it->task_pos = NULL;
4653 return;
4654 }
4655 it->task_pos = it->cur_tasks_head->next;
4656
4657 /*
4658 * We don't keep css_sets locked across iteration steps and thus
4659 * need to take steps to ensure that iteration can be resumed after
4660 * the lock is re-acquired. Iteration is performed at two levels -
4661 * css_sets and tasks in them.
4662 *
4663 * Once created, a css_set never leaves its cgroup lists, so a
4664 * pinned css_set is guaranteed to stay put and we can resume
4665 * iteration afterwards.
4666 *
4667 * Tasks may leave @cset across iteration steps. This is resolved
4668 * by registering each iterator with the css_set currently being
4669 * walked and making css_set_move_task() advance iterators whose
4670 * next task is leaving.
4671 */
4672 if (it->cur_cset) {
4673 list_del(&it->iters_node);
4674 put_css_set_locked(it->cur_cset);
4675 }
4676 get_css_set(cset);
4677 it->cur_cset = cset;
4678 list_add(&it->iters_node, &cset->task_iters);
4679 }
4680
css_task_iter_skip(struct css_task_iter * it,struct task_struct * task)4681 static void css_task_iter_skip(struct css_task_iter *it,
4682 struct task_struct *task)
4683 {
4684 lockdep_assert_held(&css_set_lock);
4685
4686 if (it->task_pos == &task->cg_list) {
4687 it->task_pos = it->task_pos->next;
4688 it->flags |= CSS_TASK_ITER_SKIPPED;
4689 }
4690 }
4691
css_task_iter_advance(struct css_task_iter * it)4692 static void css_task_iter_advance(struct css_task_iter *it)
4693 {
4694 struct task_struct *task;
4695
4696 lockdep_assert_held(&css_set_lock);
4697 repeat:
4698 if (it->task_pos) {
4699 /*
4700 * Advance iterator to find next entry. We go through cset
4701 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4702 * the next cset.
4703 */
4704 if (it->flags & CSS_TASK_ITER_SKIPPED)
4705 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4706 else
4707 it->task_pos = it->task_pos->next;
4708
4709 if (it->task_pos == &it->cur_cset->tasks) {
4710 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4711 it->task_pos = it->cur_tasks_head->next;
4712 }
4713 if (it->task_pos == &it->cur_cset->mg_tasks) {
4714 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4715 it->task_pos = it->cur_tasks_head->next;
4716 }
4717 if (it->task_pos == &it->cur_cset->dying_tasks)
4718 css_task_iter_advance_css_set(it);
4719 } else {
4720 /* called from start, proceed to the first cset */
4721 css_task_iter_advance_css_set(it);
4722 }
4723
4724 if (!it->task_pos)
4725 return;
4726
4727 task = list_entry(it->task_pos, struct task_struct, cg_list);
4728
4729 if (it->flags & CSS_TASK_ITER_PROCS) {
4730 /* if PROCS, skip over tasks which aren't group leaders */
4731 if (!thread_group_leader(task))
4732 goto repeat;
4733
4734 /* and dying leaders w/o live member threads */
4735 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4736 !atomic_read(&task->signal->live))
4737 goto repeat;
4738 } else {
4739 /* skip all dying ones */
4740 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4741 goto repeat;
4742 }
4743 }
4744
4745 /**
4746 * css_task_iter_start - initiate task iteration
4747 * @css: the css to walk tasks of
4748 * @flags: CSS_TASK_ITER_* flags
4749 * @it: the task iterator to use
4750 *
4751 * Initiate iteration through the tasks of @css. The caller can call
4752 * css_task_iter_next() to walk through the tasks until the function
4753 * returns NULL. On completion of iteration, css_task_iter_end() must be
4754 * called.
4755 */
css_task_iter_start(struct cgroup_subsys_state * css,unsigned int flags,struct css_task_iter * it)4756 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4757 struct css_task_iter *it)
4758 {
4759 memset(it, 0, sizeof(*it));
4760
4761 spin_lock_irq(&css_set_lock);
4762
4763 it->ss = css->ss;
4764 it->flags = flags;
4765
4766 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4767 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4768 else
4769 it->cset_pos = &css->cgroup->cset_links;
4770
4771 it->cset_head = it->cset_pos;
4772
4773 css_task_iter_advance(it);
4774
4775 spin_unlock_irq(&css_set_lock);
4776 }
4777
4778 /**
4779 * css_task_iter_next - return the next task for the iterator
4780 * @it: the task iterator being iterated
4781 *
4782 * The "next" function for task iteration. @it should have been
4783 * initialized via css_task_iter_start(). Returns NULL when the iteration
4784 * reaches the end.
4785 */
css_task_iter_next(struct css_task_iter * it)4786 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4787 {
4788 if (it->cur_task) {
4789 put_task_struct(it->cur_task);
4790 it->cur_task = NULL;
4791 }
4792
4793 spin_lock_irq(&css_set_lock);
4794
4795 /* @it may be half-advanced by skips, finish advancing */
4796 if (it->flags & CSS_TASK_ITER_SKIPPED)
4797 css_task_iter_advance(it);
4798
4799 if (it->task_pos) {
4800 it->cur_task = list_entry(it->task_pos, struct task_struct,
4801 cg_list);
4802 get_task_struct(it->cur_task);
4803 css_task_iter_advance(it);
4804 }
4805
4806 spin_unlock_irq(&css_set_lock);
4807
4808 return it->cur_task;
4809 }
4810
4811 /**
4812 * css_task_iter_end - finish task iteration
4813 * @it: the task iterator to finish
4814 *
4815 * Finish task iteration started by css_task_iter_start().
4816 */
css_task_iter_end(struct css_task_iter * it)4817 void css_task_iter_end(struct css_task_iter *it)
4818 {
4819 if (it->cur_cset) {
4820 spin_lock_irq(&css_set_lock);
4821 list_del(&it->iters_node);
4822 put_css_set_locked(it->cur_cset);
4823 spin_unlock_irq(&css_set_lock);
4824 }
4825
4826 if (it->cur_dcset)
4827 put_css_set(it->cur_dcset);
4828
4829 if (it->cur_task)
4830 put_task_struct(it->cur_task);
4831 }
4832
cgroup_procs_release(struct kernfs_open_file * of)4833 static void cgroup_procs_release(struct kernfs_open_file *of)
4834 {
4835 struct cgroup_file_ctx *ctx = of->priv;
4836
4837 if (ctx->procs.started)
4838 css_task_iter_end(&ctx->procs.iter);
4839 }
4840
cgroup_procs_next(struct seq_file * s,void * v,loff_t * pos)4841 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4842 {
4843 struct kernfs_open_file *of = s->private;
4844 struct cgroup_file_ctx *ctx = of->priv;
4845
4846 if (pos)
4847 (*pos)++;
4848
4849 return css_task_iter_next(&ctx->procs.iter);
4850 }
4851
__cgroup_procs_start(struct seq_file * s,loff_t * pos,unsigned int iter_flags)4852 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4853 unsigned int iter_flags)
4854 {
4855 struct kernfs_open_file *of = s->private;
4856 struct cgroup *cgrp = seq_css(s)->cgroup;
4857 struct cgroup_file_ctx *ctx = of->priv;
4858 struct css_task_iter *it = &ctx->procs.iter;
4859
4860 /*
4861 * When a seq_file is seeked, it's always traversed sequentially
4862 * from position 0, so we can simply keep iterating on !0 *pos.
4863 */
4864 if (!ctx->procs.started) {
4865 if (WARN_ON_ONCE((*pos)))
4866 return ERR_PTR(-EINVAL);
4867 css_task_iter_start(&cgrp->self, iter_flags, it);
4868 ctx->procs.started = true;
4869 } else if (!(*pos)) {
4870 css_task_iter_end(it);
4871 css_task_iter_start(&cgrp->self, iter_flags, it);
4872 } else
4873 return it->cur_task;
4874
4875 return cgroup_procs_next(s, NULL, NULL);
4876 }
4877
cgroup_procs_start(struct seq_file * s,loff_t * pos)4878 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4879 {
4880 struct cgroup *cgrp = seq_css(s)->cgroup;
4881
4882 /*
4883 * All processes of a threaded subtree belong to the domain cgroup
4884 * of the subtree. Only threads can be distributed across the
4885 * subtree. Reject reads on cgroup.procs in the subtree proper.
4886 * They're always empty anyway.
4887 */
4888 if (cgroup_is_threaded(cgrp))
4889 return ERR_PTR(-EOPNOTSUPP);
4890
4891 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4892 CSS_TASK_ITER_THREADED);
4893 }
4894
cgroup_procs_show(struct seq_file * s,void * v)4895 static int cgroup_procs_show(struct seq_file *s, void *v)
4896 {
4897 seq_printf(s, "%d\n", task_pid_vnr(v));
4898 return 0;
4899 }
4900
cgroup_may_write(const struct cgroup * cgrp,struct super_block * sb)4901 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
4902 {
4903 int ret;
4904 struct inode *inode;
4905
4906 lockdep_assert_held(&cgroup_mutex);
4907
4908 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
4909 if (!inode)
4910 return -ENOMEM;
4911
4912 ret = inode_permission(&init_user_ns, inode, MAY_WRITE);
4913 iput(inode);
4914 return ret;
4915 }
4916
cgroup_procs_write_permission(struct cgroup * src_cgrp,struct cgroup * dst_cgrp,struct super_block * sb,struct cgroup_namespace * ns)4917 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4918 struct cgroup *dst_cgrp,
4919 struct super_block *sb,
4920 struct cgroup_namespace *ns)
4921 {
4922 struct cgroup *com_cgrp = src_cgrp;
4923 int ret;
4924
4925 lockdep_assert_held(&cgroup_mutex);
4926
4927 /* find the common ancestor */
4928 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4929 com_cgrp = cgroup_parent(com_cgrp);
4930
4931 /* %current should be authorized to migrate to the common ancestor */
4932 ret = cgroup_may_write(com_cgrp, sb);
4933 if (ret)
4934 return ret;
4935
4936 /*
4937 * If namespaces are delegation boundaries, %current must be able
4938 * to see both source and destination cgroups from its namespace.
4939 */
4940 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4941 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4942 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4943 return -ENOENT;
4944
4945 return 0;
4946 }
4947
cgroup_attach_permissions(struct cgroup * src_cgrp,struct cgroup * dst_cgrp,struct super_block * sb,bool threadgroup,struct cgroup_namespace * ns)4948 static int cgroup_attach_permissions(struct cgroup *src_cgrp,
4949 struct cgroup *dst_cgrp,
4950 struct super_block *sb, bool threadgroup,
4951 struct cgroup_namespace *ns)
4952 {
4953 int ret = 0;
4954
4955 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
4956 if (ret)
4957 return ret;
4958
4959 ret = cgroup_migrate_vet_dst(dst_cgrp);
4960 if (ret)
4961 return ret;
4962
4963 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
4964 ret = -EOPNOTSUPP;
4965
4966 return ret;
4967 }
4968
__cgroup_procs_write(struct kernfs_open_file * of,char * buf,bool threadgroup)4969 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
4970 bool threadgroup)
4971 {
4972 struct cgroup_file_ctx *ctx = of->priv;
4973 struct cgroup *src_cgrp, *dst_cgrp;
4974 struct task_struct *task;
4975 const struct cred *saved_cred;
4976 ssize_t ret;
4977 bool threadgroup_locked;
4978
4979 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4980 if (!dst_cgrp)
4981 return -ENODEV;
4982
4983 task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
4984 ret = PTR_ERR_OR_ZERO(task);
4985 if (ret)
4986 goto out_unlock;
4987
4988 /* find the source cgroup */
4989 spin_lock_irq(&css_set_lock);
4990 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4991 spin_unlock_irq(&css_set_lock);
4992
4993 /*
4994 * Process and thread migrations follow same delegation rule. Check
4995 * permissions using the credentials from file open to protect against
4996 * inherited fd attacks.
4997 */
4998 saved_cred = override_creds(of->file->f_cred);
4999 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5000 of->file->f_path.dentry->d_sb,
5001 threadgroup, ctx->ns);
5002 revert_creds(saved_cred);
5003 if (ret)
5004 goto out_finish;
5005
5006 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5007
5008 out_finish:
5009 cgroup_procs_write_finish(task, threadgroup_locked);
5010 out_unlock:
5011 cgroup_kn_unlock(of->kn);
5012
5013 return ret;
5014 }
5015
cgroup_procs_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)5016 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5017 char *buf, size_t nbytes, loff_t off)
5018 {
5019 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5020 }
5021
cgroup_threads_start(struct seq_file * s,loff_t * pos)5022 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5023 {
5024 return __cgroup_procs_start(s, pos, 0);
5025 }
5026
cgroup_threads_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)5027 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5028 char *buf, size_t nbytes, loff_t off)
5029 {
5030 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5031 }
5032
5033 /* cgroup core interface files for the default hierarchy */
5034 static struct cftype cgroup_base_files[] = {
5035 {
5036 .name = "cgroup.type",
5037 .flags = CFTYPE_NOT_ON_ROOT,
5038 .seq_show = cgroup_type_show,
5039 .write = cgroup_type_write,
5040 },
5041 {
5042 .name = "cgroup.procs",
5043 .flags = CFTYPE_NS_DELEGATABLE,
5044 .file_offset = offsetof(struct cgroup, procs_file),
5045 .release = cgroup_procs_release,
5046 .seq_start = cgroup_procs_start,
5047 .seq_next = cgroup_procs_next,
5048 .seq_show = cgroup_procs_show,
5049 .write = cgroup_procs_write,
5050 },
5051 {
5052 .name = "cgroup.threads",
5053 .flags = CFTYPE_NS_DELEGATABLE,
5054 .release = cgroup_procs_release,
5055 .seq_start = cgroup_threads_start,
5056 .seq_next = cgroup_procs_next,
5057 .seq_show = cgroup_procs_show,
5058 .write = cgroup_threads_write,
5059 },
5060 {
5061 .name = "cgroup.controllers",
5062 .seq_show = cgroup_controllers_show,
5063 },
5064 {
5065 .name = "cgroup.subtree_control",
5066 .flags = CFTYPE_NS_DELEGATABLE,
5067 .seq_show = cgroup_subtree_control_show,
5068 .write = cgroup_subtree_control_write,
5069 },
5070 {
5071 .name = "cgroup.events",
5072 .flags = CFTYPE_NOT_ON_ROOT,
5073 .file_offset = offsetof(struct cgroup, events_file),
5074 .seq_show = cgroup_events_show,
5075 },
5076 {
5077 .name = "cgroup.max.descendants",
5078 .seq_show = cgroup_max_descendants_show,
5079 .write = cgroup_max_descendants_write,
5080 },
5081 {
5082 .name = "cgroup.max.depth",
5083 .seq_show = cgroup_max_depth_show,
5084 .write = cgroup_max_depth_write,
5085 },
5086 {
5087 .name = "cgroup.stat",
5088 .seq_show = cgroup_stat_show,
5089 },
5090 {
5091 .name = "cgroup.freeze",
5092 .flags = CFTYPE_NOT_ON_ROOT,
5093 .seq_show = cgroup_freeze_show,
5094 .write = cgroup_freeze_write,
5095 },
5096 {
5097 .name = "cgroup.kill",
5098 .flags = CFTYPE_NOT_ON_ROOT,
5099 .write = cgroup_kill_write,
5100 },
5101 {
5102 .name = "cpu.stat",
5103 .seq_show = cpu_stat_show,
5104 },
5105 #ifdef CONFIG_PSI
5106 {
5107 .name = "io.pressure",
5108 .flags = CFTYPE_PRESSURE,
5109 .seq_show = cgroup_io_pressure_show,
5110 .write = cgroup_io_pressure_write,
5111 .poll = cgroup_pressure_poll,
5112 .release = cgroup_pressure_release,
5113 },
5114 {
5115 .name = "memory.pressure",
5116 .flags = CFTYPE_PRESSURE,
5117 .seq_show = cgroup_memory_pressure_show,
5118 .write = cgroup_memory_pressure_write,
5119 .poll = cgroup_pressure_poll,
5120 .release = cgroup_pressure_release,
5121 },
5122 {
5123 .name = "cpu.pressure",
5124 .flags = CFTYPE_PRESSURE,
5125 .seq_show = cgroup_cpu_pressure_show,
5126 .write = cgroup_cpu_pressure_write,
5127 .poll = cgroup_pressure_poll,
5128 .release = cgroup_pressure_release,
5129 },
5130 #endif /* CONFIG_PSI */
5131 { } /* terminate */
5132 };
5133
5134 /*
5135 * css destruction is four-stage process.
5136 *
5137 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5138 * Implemented in kill_css().
5139 *
5140 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5141 * and thus css_tryget_online() is guaranteed to fail, the css can be
5142 * offlined by invoking offline_css(). After offlining, the base ref is
5143 * put. Implemented in css_killed_work_fn().
5144 *
5145 * 3. When the percpu_ref reaches zero, the only possible remaining
5146 * accessors are inside RCU read sections. css_release() schedules the
5147 * RCU callback.
5148 *
5149 * 4. After the grace period, the css can be freed. Implemented in
5150 * css_free_work_fn().
5151 *
5152 * It is actually hairier because both step 2 and 4 require process context
5153 * and thus involve punting to css->destroy_work adding two additional
5154 * steps to the already complex sequence.
5155 */
css_free_rwork_fn(struct work_struct * work)5156 static void css_free_rwork_fn(struct work_struct *work)
5157 {
5158 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5159 struct cgroup_subsys_state, destroy_rwork);
5160 struct cgroup_subsys *ss = css->ss;
5161 struct cgroup *cgrp = css->cgroup;
5162
5163 percpu_ref_exit(&css->refcnt);
5164
5165 if (ss) {
5166 /* css free path */
5167 struct cgroup_subsys_state *parent = css->parent;
5168 int id = css->id;
5169
5170 ss->css_free(css);
5171 cgroup_idr_remove(&ss->css_idr, id);
5172 cgroup_put(cgrp);
5173
5174 if (parent)
5175 css_put(parent);
5176 } else {
5177 /* cgroup free path */
5178 atomic_dec(&cgrp->root->nr_cgrps);
5179 cgroup1_pidlist_destroy_all(cgrp);
5180 cancel_work_sync(&cgrp->release_agent_work);
5181
5182 if (cgroup_parent(cgrp)) {
5183 /*
5184 * We get a ref to the parent, and put the ref when
5185 * this cgroup is being freed, so it's guaranteed
5186 * that the parent won't be destroyed before its
5187 * children.
5188 */
5189 cgroup_put(cgroup_parent(cgrp));
5190 kernfs_put(cgrp->kn);
5191 psi_cgroup_free(cgrp);
5192 cgroup_rstat_exit(cgrp);
5193 kfree(cgrp);
5194 } else {
5195 /*
5196 * This is root cgroup's refcnt reaching zero,
5197 * which indicates that the root should be
5198 * released.
5199 */
5200 cgroup_destroy_root(cgrp->root);
5201 }
5202 }
5203 }
5204
css_release_work_fn(struct work_struct * work)5205 static void css_release_work_fn(struct work_struct *work)
5206 {
5207 struct cgroup_subsys_state *css =
5208 container_of(work, struct cgroup_subsys_state, destroy_work);
5209 struct cgroup_subsys *ss = css->ss;
5210 struct cgroup *cgrp = css->cgroup;
5211
5212 mutex_lock(&cgroup_mutex);
5213
5214 css->flags |= CSS_RELEASED;
5215 list_del_rcu(&css->sibling);
5216
5217 if (ss) {
5218 /* css release path */
5219 if (!list_empty(&css->rstat_css_node)) {
5220 cgroup_rstat_flush(cgrp);
5221 list_del_rcu(&css->rstat_css_node);
5222 }
5223
5224 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5225 if (ss->css_released)
5226 ss->css_released(css);
5227 } else {
5228 struct cgroup *tcgrp;
5229
5230 /* cgroup release path */
5231 TRACE_CGROUP_PATH(release, cgrp);
5232
5233 cgroup_rstat_flush(cgrp);
5234
5235 spin_lock_irq(&css_set_lock);
5236 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5237 tcgrp = cgroup_parent(tcgrp))
5238 tcgrp->nr_dying_descendants--;
5239 spin_unlock_irq(&css_set_lock);
5240
5241 /*
5242 * There are two control paths which try to determine
5243 * cgroup from dentry without going through kernfs -
5244 * cgroupstats_build() and css_tryget_online_from_dir().
5245 * Those are supported by RCU protecting clearing of
5246 * cgrp->kn->priv backpointer.
5247 */
5248 if (cgrp->kn)
5249 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5250 NULL);
5251 }
5252
5253 mutex_unlock(&cgroup_mutex);
5254
5255 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5256 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5257 }
5258
css_release(struct percpu_ref * ref)5259 static void css_release(struct percpu_ref *ref)
5260 {
5261 struct cgroup_subsys_state *css =
5262 container_of(ref, struct cgroup_subsys_state, refcnt);
5263
5264 INIT_WORK(&css->destroy_work, css_release_work_fn);
5265 queue_work(cgroup_destroy_wq, &css->destroy_work);
5266 }
5267
init_and_link_css(struct cgroup_subsys_state * css,struct cgroup_subsys * ss,struct cgroup * cgrp)5268 static void init_and_link_css(struct cgroup_subsys_state *css,
5269 struct cgroup_subsys *ss, struct cgroup *cgrp)
5270 {
5271 lockdep_assert_held(&cgroup_mutex);
5272
5273 cgroup_get_live(cgrp);
5274
5275 memset(css, 0, sizeof(*css));
5276 css->cgroup = cgrp;
5277 css->ss = ss;
5278 css->id = -1;
5279 INIT_LIST_HEAD(&css->sibling);
5280 INIT_LIST_HEAD(&css->children);
5281 INIT_LIST_HEAD(&css->rstat_css_node);
5282 css->serial_nr = css_serial_nr_next++;
5283 atomic_set(&css->online_cnt, 0);
5284
5285 if (cgroup_parent(cgrp)) {
5286 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5287 css_get(css->parent);
5288 }
5289
5290 if (ss->css_rstat_flush)
5291 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5292
5293 BUG_ON(cgroup_css(cgrp, ss));
5294 }
5295
5296 /* invoke ->css_online() on a new CSS and mark it online if successful */
online_css(struct cgroup_subsys_state * css)5297 static int online_css(struct cgroup_subsys_state *css)
5298 {
5299 struct cgroup_subsys *ss = css->ss;
5300 int ret = 0;
5301
5302 lockdep_assert_held(&cgroup_mutex);
5303
5304 if (ss->css_online)
5305 ret = ss->css_online(css);
5306 if (!ret) {
5307 css->flags |= CSS_ONLINE;
5308 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5309
5310 atomic_inc(&css->online_cnt);
5311 if (css->parent)
5312 atomic_inc(&css->parent->online_cnt);
5313 }
5314 return ret;
5315 }
5316
5317 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
offline_css(struct cgroup_subsys_state * css)5318 static void offline_css(struct cgroup_subsys_state *css)
5319 {
5320 struct cgroup_subsys *ss = css->ss;
5321
5322 lockdep_assert_held(&cgroup_mutex);
5323
5324 if (!(css->flags & CSS_ONLINE))
5325 return;
5326
5327 if (ss->css_offline)
5328 ss->css_offline(css);
5329
5330 css->flags &= ~CSS_ONLINE;
5331 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5332
5333 wake_up_all(&css->cgroup->offline_waitq);
5334 }
5335
5336 /**
5337 * css_create - create a cgroup_subsys_state
5338 * @cgrp: the cgroup new css will be associated with
5339 * @ss: the subsys of new css
5340 *
5341 * Create a new css associated with @cgrp - @ss pair. On success, the new
5342 * css is online and installed in @cgrp. This function doesn't create the
5343 * interface files. Returns 0 on success, -errno on failure.
5344 */
css_create(struct cgroup * cgrp,struct cgroup_subsys * ss)5345 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5346 struct cgroup_subsys *ss)
5347 {
5348 struct cgroup *parent = cgroup_parent(cgrp);
5349 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5350 struct cgroup_subsys_state *css;
5351 int err;
5352
5353 lockdep_assert_held(&cgroup_mutex);
5354
5355 css = ss->css_alloc(parent_css);
5356 if (!css)
5357 css = ERR_PTR(-ENOMEM);
5358 if (IS_ERR(css))
5359 return css;
5360
5361 init_and_link_css(css, ss, cgrp);
5362
5363 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5364 if (err)
5365 goto err_free_css;
5366
5367 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5368 if (err < 0)
5369 goto err_free_css;
5370 css->id = err;
5371
5372 /* @css is ready to be brought online now, make it visible */
5373 list_add_tail_rcu(&css->sibling, &parent_css->children);
5374 cgroup_idr_replace(&ss->css_idr, css, css->id);
5375
5376 err = online_css(css);
5377 if (err)
5378 goto err_list_del;
5379
5380 return css;
5381
5382 err_list_del:
5383 list_del_rcu(&css->sibling);
5384 err_free_css:
5385 list_del_rcu(&css->rstat_css_node);
5386 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5387 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5388 return ERR_PTR(err);
5389 }
5390
5391 /*
5392 * The returned cgroup is fully initialized including its control mask, but
5393 * it isn't associated with its kernfs_node and doesn't have the control
5394 * mask applied.
5395 */
cgroup_create(struct cgroup * parent,const char * name,umode_t mode)5396 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5397 umode_t mode)
5398 {
5399 struct cgroup_root *root = parent->root;
5400 struct cgroup *cgrp, *tcgrp;
5401 struct kernfs_node *kn;
5402 int level = parent->level + 1;
5403 int ret;
5404
5405 /* allocate the cgroup and its ID, 0 is reserved for the root */
5406 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5407 GFP_KERNEL);
5408 if (!cgrp)
5409 return ERR_PTR(-ENOMEM);
5410
5411 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5412 if (ret)
5413 goto out_free_cgrp;
5414
5415 ret = cgroup_rstat_init(cgrp);
5416 if (ret)
5417 goto out_cancel_ref;
5418
5419 /* create the directory */
5420 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5421 if (IS_ERR(kn)) {
5422 ret = PTR_ERR(kn);
5423 goto out_stat_exit;
5424 }
5425 cgrp->kn = kn;
5426
5427 init_cgroup_housekeeping(cgrp);
5428
5429 cgrp->self.parent = &parent->self;
5430 cgrp->root = root;
5431 cgrp->level = level;
5432
5433 ret = psi_cgroup_alloc(cgrp);
5434 if (ret)
5435 goto out_kernfs_remove;
5436
5437 ret = cgroup_bpf_inherit(cgrp);
5438 if (ret)
5439 goto out_psi_free;
5440
5441 /*
5442 * New cgroup inherits effective freeze counter, and
5443 * if the parent has to be frozen, the child has too.
5444 */
5445 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5446 if (cgrp->freezer.e_freeze) {
5447 /*
5448 * Set the CGRP_FREEZE flag, so when a process will be
5449 * attached to the child cgroup, it will become frozen.
5450 * At this point the new cgroup is unpopulated, so we can
5451 * consider it frozen immediately.
5452 */
5453 set_bit(CGRP_FREEZE, &cgrp->flags);
5454 set_bit(CGRP_FROZEN, &cgrp->flags);
5455 }
5456
5457 spin_lock_irq(&css_set_lock);
5458 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5459 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp);
5460
5461 if (tcgrp != cgrp) {
5462 tcgrp->nr_descendants++;
5463
5464 /*
5465 * If the new cgroup is frozen, all ancestor cgroups
5466 * get a new frozen descendant, but their state can't
5467 * change because of this.
5468 */
5469 if (cgrp->freezer.e_freeze)
5470 tcgrp->freezer.nr_frozen_descendants++;
5471 }
5472 }
5473 spin_unlock_irq(&css_set_lock);
5474
5475 if (notify_on_release(parent))
5476 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5477
5478 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5479 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5480
5481 cgrp->self.serial_nr = css_serial_nr_next++;
5482
5483 /* allocation complete, commit to creation */
5484 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5485 atomic_inc(&root->nr_cgrps);
5486 cgroup_get_live(parent);
5487
5488 /*
5489 * On the default hierarchy, a child doesn't automatically inherit
5490 * subtree_control from the parent. Each is configured manually.
5491 */
5492 if (!cgroup_on_dfl(cgrp))
5493 cgrp->subtree_control = cgroup_control(cgrp);
5494
5495 cgroup_propagate_control(cgrp);
5496
5497 return cgrp;
5498
5499 out_psi_free:
5500 psi_cgroup_free(cgrp);
5501 out_kernfs_remove:
5502 kernfs_remove(cgrp->kn);
5503 out_stat_exit:
5504 cgroup_rstat_exit(cgrp);
5505 out_cancel_ref:
5506 percpu_ref_exit(&cgrp->self.refcnt);
5507 out_free_cgrp:
5508 kfree(cgrp);
5509 return ERR_PTR(ret);
5510 }
5511
cgroup_check_hierarchy_limits(struct cgroup * parent)5512 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5513 {
5514 struct cgroup *cgroup;
5515 int ret = false;
5516 int level = 1;
5517
5518 lockdep_assert_held(&cgroup_mutex);
5519
5520 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5521 if (cgroup->nr_descendants >= cgroup->max_descendants)
5522 goto fail;
5523
5524 if (level > cgroup->max_depth)
5525 goto fail;
5526
5527 level++;
5528 }
5529
5530 ret = true;
5531 fail:
5532 return ret;
5533 }
5534
cgroup_mkdir(struct kernfs_node * parent_kn,const char * name,umode_t mode)5535 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5536 {
5537 struct cgroup *parent, *cgrp;
5538 int ret;
5539
5540 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5541 if (strchr(name, '\n'))
5542 return -EINVAL;
5543
5544 parent = cgroup_kn_lock_live(parent_kn, false);
5545 if (!parent)
5546 return -ENODEV;
5547
5548 if (!cgroup_check_hierarchy_limits(parent)) {
5549 ret = -EAGAIN;
5550 goto out_unlock;
5551 }
5552
5553 cgrp = cgroup_create(parent, name, mode);
5554 if (IS_ERR(cgrp)) {
5555 ret = PTR_ERR(cgrp);
5556 goto out_unlock;
5557 }
5558
5559 /*
5560 * This extra ref will be put in cgroup_free_fn() and guarantees
5561 * that @cgrp->kn is always accessible.
5562 */
5563 kernfs_get(cgrp->kn);
5564
5565 ret = cgroup_kn_set_ugid(cgrp->kn);
5566 if (ret)
5567 goto out_destroy;
5568
5569 ret = css_populate_dir(&cgrp->self);
5570 if (ret)
5571 goto out_destroy;
5572
5573 ret = cgroup_apply_control_enable(cgrp);
5574 if (ret)
5575 goto out_destroy;
5576
5577 TRACE_CGROUP_PATH(mkdir, cgrp);
5578
5579 /* let's create and online css's */
5580 kernfs_activate(cgrp->kn);
5581
5582 ret = 0;
5583 goto out_unlock;
5584
5585 out_destroy:
5586 cgroup_destroy_locked(cgrp);
5587 out_unlock:
5588 cgroup_kn_unlock(parent_kn);
5589 return ret;
5590 }
5591
5592 /*
5593 * This is called when the refcnt of a css is confirmed to be killed.
5594 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5595 * initiate destruction and put the css ref from kill_css().
5596 */
css_killed_work_fn(struct work_struct * work)5597 static void css_killed_work_fn(struct work_struct *work)
5598 {
5599 struct cgroup_subsys_state *css =
5600 container_of(work, struct cgroup_subsys_state, destroy_work);
5601
5602 mutex_lock(&cgroup_mutex);
5603
5604 do {
5605 offline_css(css);
5606 css_put(css);
5607 /* @css can't go away while we're holding cgroup_mutex */
5608 css = css->parent;
5609 } while (css && atomic_dec_and_test(&css->online_cnt));
5610
5611 mutex_unlock(&cgroup_mutex);
5612 }
5613
5614 /* css kill confirmation processing requires process context, bounce */
css_killed_ref_fn(struct percpu_ref * ref)5615 static void css_killed_ref_fn(struct percpu_ref *ref)
5616 {
5617 struct cgroup_subsys_state *css =
5618 container_of(ref, struct cgroup_subsys_state, refcnt);
5619
5620 if (atomic_dec_and_test(&css->online_cnt)) {
5621 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5622 queue_work(cgroup_destroy_wq, &css->destroy_work);
5623 }
5624 }
5625
5626 /**
5627 * kill_css - destroy a css
5628 * @css: css to destroy
5629 *
5630 * This function initiates destruction of @css by removing cgroup interface
5631 * files and putting its base reference. ->css_offline() will be invoked
5632 * asynchronously once css_tryget_online() is guaranteed to fail and when
5633 * the reference count reaches zero, @css will be released.
5634 */
kill_css(struct cgroup_subsys_state * css)5635 static void kill_css(struct cgroup_subsys_state *css)
5636 {
5637 lockdep_assert_held(&cgroup_mutex);
5638
5639 if (css->flags & CSS_DYING)
5640 return;
5641
5642 css->flags |= CSS_DYING;
5643
5644 /*
5645 * This must happen before css is disassociated with its cgroup.
5646 * See seq_css() for details.
5647 */
5648 css_clear_dir(css);
5649
5650 /*
5651 * Killing would put the base ref, but we need to keep it alive
5652 * until after ->css_offline().
5653 */
5654 css_get(css);
5655
5656 /*
5657 * cgroup core guarantees that, by the time ->css_offline() is
5658 * invoked, no new css reference will be given out via
5659 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5660 * proceed to offlining css's because percpu_ref_kill() doesn't
5661 * guarantee that the ref is seen as killed on all CPUs on return.
5662 *
5663 * Use percpu_ref_kill_and_confirm() to get notifications as each
5664 * css is confirmed to be seen as killed on all CPUs.
5665 */
5666 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5667 }
5668
5669 /**
5670 * cgroup_destroy_locked - the first stage of cgroup destruction
5671 * @cgrp: cgroup to be destroyed
5672 *
5673 * css's make use of percpu refcnts whose killing latency shouldn't be
5674 * exposed to userland and are RCU protected. Also, cgroup core needs to
5675 * guarantee that css_tryget_online() won't succeed by the time
5676 * ->css_offline() is invoked. To satisfy all the requirements,
5677 * destruction is implemented in the following two steps.
5678 *
5679 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5680 * userland visible parts and start killing the percpu refcnts of
5681 * css's. Set up so that the next stage will be kicked off once all
5682 * the percpu refcnts are confirmed to be killed.
5683 *
5684 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5685 * rest of destruction. Once all cgroup references are gone, the
5686 * cgroup is RCU-freed.
5687 *
5688 * This function implements s1. After this step, @cgrp is gone as far as
5689 * the userland is concerned and a new cgroup with the same name may be
5690 * created. As cgroup doesn't care about the names internally, this
5691 * doesn't cause any problem.
5692 */
cgroup_destroy_locked(struct cgroup * cgrp)5693 static int cgroup_destroy_locked(struct cgroup *cgrp)
5694 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5695 {
5696 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5697 struct cgroup_subsys_state *css;
5698 struct cgrp_cset_link *link;
5699 int ssid;
5700
5701 lockdep_assert_held(&cgroup_mutex);
5702
5703 /*
5704 * Only migration can raise populated from zero and we're already
5705 * holding cgroup_mutex.
5706 */
5707 if (cgroup_is_populated(cgrp))
5708 return -EBUSY;
5709
5710 /*
5711 * Make sure there's no live children. We can't test emptiness of
5712 * ->self.children as dead children linger on it while being
5713 * drained; otherwise, "rmdir parent/child parent" may fail.
5714 */
5715 if (css_has_online_children(&cgrp->self))
5716 return -EBUSY;
5717
5718 /*
5719 * Mark @cgrp and the associated csets dead. The former prevents
5720 * further task migration and child creation by disabling
5721 * cgroup_lock_live_group(). The latter makes the csets ignored by
5722 * the migration path.
5723 */
5724 cgrp->self.flags &= ~CSS_ONLINE;
5725
5726 spin_lock_irq(&css_set_lock);
5727 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5728 link->cset->dead = true;
5729 spin_unlock_irq(&css_set_lock);
5730
5731 /* initiate massacre of all css's */
5732 for_each_css(css, ssid, cgrp)
5733 kill_css(css);
5734
5735 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5736 css_clear_dir(&cgrp->self);
5737 kernfs_remove(cgrp->kn);
5738
5739 if (cgroup_is_threaded(cgrp))
5740 parent->nr_threaded_children--;
5741
5742 spin_lock_irq(&css_set_lock);
5743 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5744 tcgrp->nr_descendants--;
5745 tcgrp->nr_dying_descendants++;
5746 /*
5747 * If the dying cgroup is frozen, decrease frozen descendants
5748 * counters of ancestor cgroups.
5749 */
5750 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5751 tcgrp->freezer.nr_frozen_descendants--;
5752 }
5753 spin_unlock_irq(&css_set_lock);
5754
5755 cgroup1_check_for_release(parent);
5756
5757 cgroup_bpf_offline(cgrp);
5758
5759 /* put the base reference */
5760 percpu_ref_kill(&cgrp->self.refcnt);
5761
5762 return 0;
5763 };
5764
cgroup_rmdir(struct kernfs_node * kn)5765 int cgroup_rmdir(struct kernfs_node *kn)
5766 {
5767 struct cgroup *cgrp;
5768 int ret = 0;
5769
5770 cgrp = cgroup_kn_lock_live(kn, false);
5771 if (!cgrp)
5772 return 0;
5773
5774 ret = cgroup_destroy_locked(cgrp);
5775 if (!ret)
5776 TRACE_CGROUP_PATH(rmdir, cgrp);
5777
5778 cgroup_kn_unlock(kn);
5779 return ret;
5780 }
5781
5782 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5783 .show_options = cgroup_show_options,
5784 .mkdir = cgroup_mkdir,
5785 .rmdir = cgroup_rmdir,
5786 .show_path = cgroup_show_path,
5787 };
5788
cgroup_init_subsys(struct cgroup_subsys * ss,bool early)5789 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5790 {
5791 struct cgroup_subsys_state *css;
5792
5793 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5794
5795 mutex_lock(&cgroup_mutex);
5796
5797 idr_init(&ss->css_idr);
5798 INIT_LIST_HEAD(&ss->cfts);
5799
5800 /* Create the root cgroup state for this subsystem */
5801 ss->root = &cgrp_dfl_root;
5802 css = ss->css_alloc(NULL);
5803 /* We don't handle early failures gracefully */
5804 BUG_ON(IS_ERR(css));
5805 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5806
5807 /*
5808 * Root csses are never destroyed and we can't initialize
5809 * percpu_ref during early init. Disable refcnting.
5810 */
5811 css->flags |= CSS_NO_REF;
5812
5813 if (early) {
5814 /* allocation can't be done safely during early init */
5815 css->id = 1;
5816 } else {
5817 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5818 BUG_ON(css->id < 0);
5819 }
5820
5821 /* Update the init_css_set to contain a subsys
5822 * pointer to this state - since the subsystem is
5823 * newly registered, all tasks and hence the
5824 * init_css_set is in the subsystem's root cgroup. */
5825 init_css_set.subsys[ss->id] = css;
5826
5827 have_fork_callback |= (bool)ss->fork << ss->id;
5828 have_exit_callback |= (bool)ss->exit << ss->id;
5829 have_release_callback |= (bool)ss->release << ss->id;
5830 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5831
5832 /* At system boot, before all subsystems have been
5833 * registered, no tasks have been forked, so we don't
5834 * need to invoke fork callbacks here. */
5835 BUG_ON(!list_empty(&init_task.tasks));
5836
5837 BUG_ON(online_css(css));
5838
5839 mutex_unlock(&cgroup_mutex);
5840 }
5841
5842 /**
5843 * cgroup_init_early - cgroup initialization at system boot
5844 *
5845 * Initialize cgroups at system boot, and initialize any
5846 * subsystems that request early init.
5847 */
cgroup_init_early(void)5848 int __init cgroup_init_early(void)
5849 {
5850 static struct cgroup_fs_context __initdata ctx;
5851 struct cgroup_subsys *ss;
5852 int i;
5853
5854 ctx.root = &cgrp_dfl_root;
5855 init_cgroup_root(&ctx);
5856 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5857
5858 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5859
5860 for_each_subsys(ss, i) {
5861 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5862 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5863 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5864 ss->id, ss->name);
5865 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5866 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5867
5868 ss->id = i;
5869 ss->name = cgroup_subsys_name[i];
5870 if (!ss->legacy_name)
5871 ss->legacy_name = cgroup_subsys_name[i];
5872
5873 if (ss->early_init)
5874 cgroup_init_subsys(ss, true);
5875 }
5876 return 0;
5877 }
5878
5879 /**
5880 * cgroup_init - cgroup initialization
5881 *
5882 * Register cgroup filesystem and /proc file, and initialize
5883 * any subsystems that didn't request early init.
5884 */
cgroup_init(void)5885 int __init cgroup_init(void)
5886 {
5887 struct cgroup_subsys *ss;
5888 int ssid;
5889
5890 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5891 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5892 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5893
5894 cgroup_rstat_boot();
5895
5896 /*
5897 * The latency of the synchronize_rcu() is too high for cgroups,
5898 * avoid it at the cost of forcing all readers into the slow path.
5899 */
5900 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5901
5902 get_user_ns(init_cgroup_ns.user_ns);
5903
5904 mutex_lock(&cgroup_mutex);
5905
5906 /*
5907 * Add init_css_set to the hash table so that dfl_root can link to
5908 * it during init.
5909 */
5910 hash_add(css_set_table, &init_css_set.hlist,
5911 css_set_hash(init_css_set.subsys));
5912
5913 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5914
5915 mutex_unlock(&cgroup_mutex);
5916
5917 for_each_subsys(ss, ssid) {
5918 if (ss->early_init) {
5919 struct cgroup_subsys_state *css =
5920 init_css_set.subsys[ss->id];
5921
5922 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5923 GFP_KERNEL);
5924 BUG_ON(css->id < 0);
5925 } else {
5926 cgroup_init_subsys(ss, false);
5927 }
5928
5929 list_add_tail(&init_css_set.e_cset_node[ssid],
5930 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5931
5932 /*
5933 * Setting dfl_root subsys_mask needs to consider the
5934 * disabled flag and cftype registration needs kmalloc,
5935 * both of which aren't available during early_init.
5936 */
5937 if (!cgroup_ssid_enabled(ssid))
5938 continue;
5939
5940 if (cgroup1_ssid_disabled(ssid))
5941 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5942 ss->name);
5943
5944 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5945
5946 /* implicit controllers must be threaded too */
5947 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5948
5949 if (ss->implicit_on_dfl)
5950 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5951 else if (!ss->dfl_cftypes)
5952 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5953
5954 if (ss->threaded)
5955 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5956
5957 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5958 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5959 } else {
5960 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5961 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5962 }
5963
5964 if (ss->bind)
5965 ss->bind(init_css_set.subsys[ssid]);
5966
5967 mutex_lock(&cgroup_mutex);
5968 css_populate_dir(init_css_set.subsys[ssid]);
5969 mutex_unlock(&cgroup_mutex);
5970 }
5971
5972 /* init_css_set.subsys[] has been updated, re-hash */
5973 hash_del(&init_css_set.hlist);
5974 hash_add(css_set_table, &init_css_set.hlist,
5975 css_set_hash(init_css_set.subsys));
5976
5977 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5978 WARN_ON(register_filesystem(&cgroup_fs_type));
5979 WARN_ON(register_filesystem(&cgroup2_fs_type));
5980 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5981 #ifdef CONFIG_CPUSETS
5982 WARN_ON(register_filesystem(&cpuset_fs_type));
5983 #endif
5984
5985 return 0;
5986 }
5987
cgroup_wq_init(void)5988 static int __init cgroup_wq_init(void)
5989 {
5990 /*
5991 * There isn't much point in executing destruction path in
5992 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5993 * Use 1 for @max_active.
5994 *
5995 * We would prefer to do this in cgroup_init() above, but that
5996 * is called before init_workqueues(): so leave this until after.
5997 */
5998 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5999 BUG_ON(!cgroup_destroy_wq);
6000 return 0;
6001 }
6002 core_initcall(cgroup_wq_init);
6003
cgroup_path_from_kernfs_id(u64 id,char * buf,size_t buflen)6004 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6005 {
6006 struct kernfs_node *kn;
6007
6008 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6009 if (!kn)
6010 return;
6011 kernfs_path(kn, buf, buflen);
6012 kernfs_put(kn);
6013 }
6014
6015 /*
6016 * cgroup_get_from_id : get the cgroup associated with cgroup id
6017 * @id: cgroup id
6018 * On success return the cgrp, on failure return NULL
6019 */
cgroup_get_from_id(u64 id)6020 struct cgroup *cgroup_get_from_id(u64 id)
6021 {
6022 struct kernfs_node *kn;
6023 struct cgroup *cgrp = NULL;
6024
6025 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6026 if (!kn)
6027 goto out;
6028
6029 rcu_read_lock();
6030
6031 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6032 if (cgrp && !cgroup_tryget(cgrp))
6033 cgrp = NULL;
6034
6035 rcu_read_unlock();
6036
6037 kernfs_put(kn);
6038 out:
6039 return cgrp;
6040 }
6041 EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6042
6043 /*
6044 * proc_cgroup_show()
6045 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6046 * - Used for /proc/<pid>/cgroup.
6047 */
proc_cgroup_show(struct seq_file * m,struct pid_namespace * ns,struct pid * pid,struct task_struct * tsk)6048 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6049 struct pid *pid, struct task_struct *tsk)
6050 {
6051 char *buf;
6052 int retval;
6053 struct cgroup_root *root;
6054
6055 retval = -ENOMEM;
6056 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6057 if (!buf)
6058 goto out;
6059
6060 mutex_lock(&cgroup_mutex);
6061 spin_lock_irq(&css_set_lock);
6062
6063 for_each_root(root) {
6064 struct cgroup_subsys *ss;
6065 struct cgroup *cgrp;
6066 int ssid, count = 0;
6067
6068 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
6069 continue;
6070
6071 seq_printf(m, "%d:", root->hierarchy_id);
6072 if (root != &cgrp_dfl_root)
6073 for_each_subsys(ss, ssid)
6074 if (root->subsys_mask & (1 << ssid))
6075 seq_printf(m, "%s%s", count++ ? "," : "",
6076 ss->legacy_name);
6077 if (strlen(root->name))
6078 seq_printf(m, "%sname=%s", count ? "," : "",
6079 root->name);
6080 seq_putc(m, ':');
6081
6082 cgrp = task_cgroup_from_root(tsk, root);
6083
6084 /*
6085 * On traditional hierarchies, all zombie tasks show up as
6086 * belonging to the root cgroup. On the default hierarchy,
6087 * while a zombie doesn't show up in "cgroup.procs" and
6088 * thus can't be migrated, its /proc/PID/cgroup keeps
6089 * reporting the cgroup it belonged to before exiting. If
6090 * the cgroup is removed before the zombie is reaped,
6091 * " (deleted)" is appended to the cgroup path.
6092 */
6093 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6094 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6095 current->nsproxy->cgroup_ns);
6096 if (retval >= PATH_MAX)
6097 retval = -ENAMETOOLONG;
6098 if (retval < 0)
6099 goto out_unlock;
6100
6101 seq_puts(m, buf);
6102 } else {
6103 seq_puts(m, "/");
6104 }
6105
6106 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6107 seq_puts(m, " (deleted)\n");
6108 else
6109 seq_putc(m, '\n');
6110 }
6111
6112 retval = 0;
6113 out_unlock:
6114 spin_unlock_irq(&css_set_lock);
6115 mutex_unlock(&cgroup_mutex);
6116 kfree(buf);
6117 out:
6118 return retval;
6119 }
6120
6121 /**
6122 * cgroup_fork - initialize cgroup related fields during copy_process()
6123 * @child: pointer to task_struct of forking parent process.
6124 *
6125 * A task is associated with the init_css_set until cgroup_post_fork()
6126 * attaches it to the target css_set.
6127 */
cgroup_fork(struct task_struct * child)6128 void cgroup_fork(struct task_struct *child)
6129 {
6130 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6131 INIT_LIST_HEAD(&child->cg_list);
6132 }
6133
cgroup_get_from_file(struct file * f)6134 static struct cgroup *cgroup_get_from_file(struct file *f)
6135 {
6136 struct cgroup_subsys_state *css;
6137 struct cgroup *cgrp;
6138
6139 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6140 if (IS_ERR(css))
6141 return ERR_CAST(css);
6142
6143 cgrp = css->cgroup;
6144 if (!cgroup_on_dfl(cgrp)) {
6145 cgroup_put(cgrp);
6146 return ERR_PTR(-EBADF);
6147 }
6148
6149 return cgrp;
6150 }
6151
6152 /**
6153 * cgroup_css_set_fork - find or create a css_set for a child process
6154 * @kargs: the arguments passed to create the child process
6155 *
6156 * This functions finds or creates a new css_set which the child
6157 * process will be attached to in cgroup_post_fork(). By default,
6158 * the child process will be given the same css_set as its parent.
6159 *
6160 * If CLONE_INTO_CGROUP is specified this function will try to find an
6161 * existing css_set which includes the requested cgroup and if not create
6162 * a new css_set that the child will be attached to later. If this function
6163 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6164 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6165 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6166 * to the target cgroup.
6167 */
cgroup_css_set_fork(struct kernel_clone_args * kargs)6168 static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6169 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6170 {
6171 int ret;
6172 struct cgroup *dst_cgrp = NULL;
6173 struct css_set *cset;
6174 struct super_block *sb;
6175 struct file *f;
6176
6177 if (kargs->flags & CLONE_INTO_CGROUP)
6178 mutex_lock(&cgroup_mutex);
6179
6180 cgroup_threadgroup_change_begin(current);
6181
6182 spin_lock_irq(&css_set_lock);
6183 cset = task_css_set(current);
6184 get_css_set(cset);
6185 spin_unlock_irq(&css_set_lock);
6186
6187 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6188 kargs->cset = cset;
6189 return 0;
6190 }
6191
6192 f = fget_raw(kargs->cgroup);
6193 if (!f) {
6194 ret = -EBADF;
6195 goto err;
6196 }
6197 sb = f->f_path.dentry->d_sb;
6198
6199 dst_cgrp = cgroup_get_from_file(f);
6200 if (IS_ERR(dst_cgrp)) {
6201 ret = PTR_ERR(dst_cgrp);
6202 dst_cgrp = NULL;
6203 goto err;
6204 }
6205
6206 if (cgroup_is_dead(dst_cgrp)) {
6207 ret = -ENODEV;
6208 goto err;
6209 }
6210
6211 /*
6212 * Verify that we the target cgroup is writable for us. This is
6213 * usually done by the vfs layer but since we're not going through
6214 * the vfs layer here we need to do it "manually".
6215 */
6216 ret = cgroup_may_write(dst_cgrp, sb);
6217 if (ret)
6218 goto err;
6219
6220 /*
6221 * Spawning a task directly into a cgroup works by passing a file
6222 * descriptor to the target cgroup directory. This can even be an O_PATH
6223 * file descriptor. But it can never be a cgroup.procs file descriptor.
6224 * This was done on purpose so spawning into a cgroup could be
6225 * conceptualized as an atomic
6226 *
6227 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6228 * write(fd, <child-pid>, ...);
6229 *
6230 * sequence, i.e. it's a shorthand for the caller opening and writing
6231 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6232 * to always use the caller's credentials.
6233 */
6234 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6235 !(kargs->flags & CLONE_THREAD),
6236 current->nsproxy->cgroup_ns);
6237 if (ret)
6238 goto err;
6239
6240 kargs->cset = find_css_set(cset, dst_cgrp);
6241 if (!kargs->cset) {
6242 ret = -ENOMEM;
6243 goto err;
6244 }
6245
6246 put_css_set(cset);
6247 fput(f);
6248 kargs->cgrp = dst_cgrp;
6249 return ret;
6250
6251 err:
6252 cgroup_threadgroup_change_end(current);
6253 mutex_unlock(&cgroup_mutex);
6254 if (f)
6255 fput(f);
6256 if (dst_cgrp)
6257 cgroup_put(dst_cgrp);
6258 put_css_set(cset);
6259 if (kargs->cset)
6260 put_css_set(kargs->cset);
6261 return ret;
6262 }
6263
6264 /**
6265 * cgroup_css_set_put_fork - drop references we took during fork
6266 * @kargs: the arguments passed to create the child process
6267 *
6268 * Drop references to the prepared css_set and target cgroup if
6269 * CLONE_INTO_CGROUP was requested.
6270 */
cgroup_css_set_put_fork(struct kernel_clone_args * kargs)6271 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6272 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6273 {
6274 cgroup_threadgroup_change_end(current);
6275
6276 if (kargs->flags & CLONE_INTO_CGROUP) {
6277 struct cgroup *cgrp = kargs->cgrp;
6278 struct css_set *cset = kargs->cset;
6279
6280 mutex_unlock(&cgroup_mutex);
6281
6282 if (cset) {
6283 put_css_set(cset);
6284 kargs->cset = NULL;
6285 }
6286
6287 if (cgrp) {
6288 cgroup_put(cgrp);
6289 kargs->cgrp = NULL;
6290 }
6291 }
6292 }
6293
6294 /**
6295 * cgroup_can_fork - called on a new task before the process is exposed
6296 * @child: the child process
6297 * @kargs: the arguments passed to create the child process
6298 *
6299 * This prepares a new css_set for the child process which the child will
6300 * be attached to in cgroup_post_fork().
6301 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6302 * callback returns an error, the fork aborts with that error code. This
6303 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6304 */
cgroup_can_fork(struct task_struct * child,struct kernel_clone_args * kargs)6305 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6306 {
6307 struct cgroup_subsys *ss;
6308 int i, j, ret;
6309
6310 ret = cgroup_css_set_fork(kargs);
6311 if (ret)
6312 return ret;
6313
6314 do_each_subsys_mask(ss, i, have_canfork_callback) {
6315 ret = ss->can_fork(child, kargs->cset);
6316 if (ret)
6317 goto out_revert;
6318 } while_each_subsys_mask();
6319
6320 return 0;
6321
6322 out_revert:
6323 for_each_subsys(ss, j) {
6324 if (j >= i)
6325 break;
6326 if (ss->cancel_fork)
6327 ss->cancel_fork(child, kargs->cset);
6328 }
6329
6330 cgroup_css_set_put_fork(kargs);
6331
6332 return ret;
6333 }
6334
6335 /**
6336 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6337 * @child: the child process
6338 * @kargs: the arguments passed to create the child process
6339 *
6340 * This calls the cancel_fork() callbacks if a fork failed *after*
6341 * cgroup_can_fork() succeeded and cleans up references we took to
6342 * prepare a new css_set for the child process in cgroup_can_fork().
6343 */
cgroup_cancel_fork(struct task_struct * child,struct kernel_clone_args * kargs)6344 void cgroup_cancel_fork(struct task_struct *child,
6345 struct kernel_clone_args *kargs)
6346 {
6347 struct cgroup_subsys *ss;
6348 int i;
6349
6350 for_each_subsys(ss, i)
6351 if (ss->cancel_fork)
6352 ss->cancel_fork(child, kargs->cset);
6353
6354 cgroup_css_set_put_fork(kargs);
6355 }
6356
6357 /**
6358 * cgroup_post_fork - finalize cgroup setup for the child process
6359 * @child: the child process
6360 * @kargs: the arguments passed to create the child process
6361 *
6362 * Attach the child process to its css_set calling the subsystem fork()
6363 * callbacks.
6364 */
cgroup_post_fork(struct task_struct * child,struct kernel_clone_args * kargs)6365 void cgroup_post_fork(struct task_struct *child,
6366 struct kernel_clone_args *kargs)
6367 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6368 {
6369 unsigned long cgrp_flags = 0;
6370 bool kill = false;
6371 struct cgroup_subsys *ss;
6372 struct css_set *cset;
6373 int i;
6374
6375 cset = kargs->cset;
6376 kargs->cset = NULL;
6377
6378 spin_lock_irq(&css_set_lock);
6379
6380 /* init tasks are special, only link regular threads */
6381 if (likely(child->pid)) {
6382 if (kargs->cgrp)
6383 cgrp_flags = kargs->cgrp->flags;
6384 else
6385 cgrp_flags = cset->dfl_cgrp->flags;
6386
6387 WARN_ON_ONCE(!list_empty(&child->cg_list));
6388 cset->nr_tasks++;
6389 css_set_move_task(child, NULL, cset, false);
6390 } else {
6391 put_css_set(cset);
6392 cset = NULL;
6393 }
6394
6395 if (!(child->flags & PF_KTHREAD)) {
6396 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6397 /*
6398 * If the cgroup has to be frozen, the new task has
6399 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6400 * get the task into the frozen state.
6401 */
6402 spin_lock(&child->sighand->siglock);
6403 WARN_ON_ONCE(child->frozen);
6404 child->jobctl |= JOBCTL_TRAP_FREEZE;
6405 spin_unlock(&child->sighand->siglock);
6406
6407 /*
6408 * Calling cgroup_update_frozen() isn't required here,
6409 * because it will be called anyway a bit later from
6410 * do_freezer_trap(). So we avoid cgroup's transient
6411 * switch from the frozen state and back.
6412 */
6413 }
6414
6415 /*
6416 * If the cgroup is to be killed notice it now and take the
6417 * child down right after we finished preparing it for
6418 * userspace.
6419 */
6420 kill = test_bit(CGRP_KILL, &cgrp_flags);
6421 }
6422
6423 spin_unlock_irq(&css_set_lock);
6424
6425 /*
6426 * Call ss->fork(). This must happen after @child is linked on
6427 * css_set; otherwise, @child might change state between ->fork()
6428 * and addition to css_set.
6429 */
6430 do_each_subsys_mask(ss, i, have_fork_callback) {
6431 ss->fork(child);
6432 } while_each_subsys_mask();
6433
6434 /* Make the new cset the root_cset of the new cgroup namespace. */
6435 if (kargs->flags & CLONE_NEWCGROUP) {
6436 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6437
6438 get_css_set(cset);
6439 child->nsproxy->cgroup_ns->root_cset = cset;
6440 put_css_set(rcset);
6441 }
6442
6443 /* Cgroup has to be killed so take down child immediately. */
6444 if (unlikely(kill))
6445 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6446
6447 cgroup_css_set_put_fork(kargs);
6448 }
6449
6450 /**
6451 * cgroup_exit - detach cgroup from exiting task
6452 * @tsk: pointer to task_struct of exiting process
6453 *
6454 * Description: Detach cgroup from @tsk.
6455 *
6456 */
cgroup_exit(struct task_struct * tsk)6457 void cgroup_exit(struct task_struct *tsk)
6458 {
6459 struct cgroup_subsys *ss;
6460 struct css_set *cset;
6461 int i;
6462
6463 spin_lock_irq(&css_set_lock);
6464
6465 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6466 cset = task_css_set(tsk);
6467 css_set_move_task(tsk, cset, NULL, false);
6468 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6469 cset->nr_tasks--;
6470
6471 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6472 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6473 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6474 cgroup_update_frozen(task_dfl_cgroup(tsk));
6475
6476 spin_unlock_irq(&css_set_lock);
6477
6478 /* see cgroup_post_fork() for details */
6479 do_each_subsys_mask(ss, i, have_exit_callback) {
6480 ss->exit(tsk);
6481 } while_each_subsys_mask();
6482 }
6483
cgroup_release(struct task_struct * task)6484 void cgroup_release(struct task_struct *task)
6485 {
6486 struct cgroup_subsys *ss;
6487 int ssid;
6488
6489 do_each_subsys_mask(ss, ssid, have_release_callback) {
6490 ss->release(task);
6491 } while_each_subsys_mask();
6492
6493 spin_lock_irq(&css_set_lock);
6494 css_set_skip_task_iters(task_css_set(task), task);
6495 list_del_init(&task->cg_list);
6496 spin_unlock_irq(&css_set_lock);
6497 }
6498
cgroup_free(struct task_struct * task)6499 void cgroup_free(struct task_struct *task)
6500 {
6501 struct css_set *cset = task_css_set(task);
6502 put_css_set(cset);
6503 }
6504
cgroup_disable(char * str)6505 static int __init cgroup_disable(char *str)
6506 {
6507 struct cgroup_subsys *ss;
6508 char *token;
6509 int i;
6510
6511 while ((token = strsep(&str, ",")) != NULL) {
6512 if (!*token)
6513 continue;
6514
6515 for_each_subsys(ss, i) {
6516 if (strcmp(token, ss->name) &&
6517 strcmp(token, ss->legacy_name))
6518 continue;
6519
6520 static_branch_disable(cgroup_subsys_enabled_key[i]);
6521 pr_info("Disabling %s control group subsystem\n",
6522 ss->name);
6523 }
6524
6525 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6526 if (strcmp(token, cgroup_opt_feature_names[i]))
6527 continue;
6528 cgroup_feature_disable_mask |= 1 << i;
6529 pr_info("Disabling %s control group feature\n",
6530 cgroup_opt_feature_names[i]);
6531 break;
6532 }
6533 }
6534 return 1;
6535 }
6536 __setup("cgroup_disable=", cgroup_disable);
6537
enable_debug_cgroup(void)6538 void __init __weak enable_debug_cgroup(void) { }
6539
enable_cgroup_debug(char * str)6540 static int __init enable_cgroup_debug(char *str)
6541 {
6542 cgroup_debug = true;
6543 enable_debug_cgroup();
6544 return 1;
6545 }
6546 __setup("cgroup_debug", enable_cgroup_debug);
6547
6548 /**
6549 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6550 * @dentry: directory dentry of interest
6551 * @ss: subsystem of interest
6552 *
6553 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6554 * to get the corresponding css and return it. If such css doesn't exist
6555 * or can't be pinned, an ERR_PTR value is returned.
6556 */
css_tryget_online_from_dir(struct dentry * dentry,struct cgroup_subsys * ss)6557 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6558 struct cgroup_subsys *ss)
6559 {
6560 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6561 struct file_system_type *s_type = dentry->d_sb->s_type;
6562 struct cgroup_subsys_state *css = NULL;
6563 struct cgroup *cgrp;
6564
6565 /* is @dentry a cgroup dir? */
6566 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6567 !kn || kernfs_type(kn) != KERNFS_DIR)
6568 return ERR_PTR(-EBADF);
6569
6570 rcu_read_lock();
6571
6572 /*
6573 * This path doesn't originate from kernfs and @kn could already
6574 * have been or be removed at any point. @kn->priv is RCU
6575 * protected for this access. See css_release_work_fn() for details.
6576 */
6577 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6578 if (cgrp)
6579 css = cgroup_css(cgrp, ss);
6580
6581 if (!css || !css_tryget_online(css))
6582 css = ERR_PTR(-ENOENT);
6583
6584 rcu_read_unlock();
6585 return css;
6586 }
6587
6588 /**
6589 * css_from_id - lookup css by id
6590 * @id: the cgroup id
6591 * @ss: cgroup subsys to be looked into
6592 *
6593 * Returns the css if there's valid one with @id, otherwise returns NULL.
6594 * Should be called under rcu_read_lock().
6595 */
css_from_id(int id,struct cgroup_subsys * ss)6596 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6597 {
6598 WARN_ON_ONCE(!rcu_read_lock_held());
6599 return idr_find(&ss->css_idr, id);
6600 }
6601
6602 /**
6603 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6604 * @path: path on the default hierarchy
6605 *
6606 * Find the cgroup at @path on the default hierarchy, increment its
6607 * reference count and return it. Returns pointer to the found cgroup on
6608 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6609 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6610 */
cgroup_get_from_path(const char * path)6611 struct cgroup *cgroup_get_from_path(const char *path)
6612 {
6613 struct kernfs_node *kn;
6614 struct cgroup *cgrp = ERR_PTR(-ENOENT);
6615
6616 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6617 if (!kn)
6618 goto out;
6619
6620 if (kernfs_type(kn) != KERNFS_DIR) {
6621 cgrp = ERR_PTR(-ENOTDIR);
6622 goto out_kernfs;
6623 }
6624
6625 rcu_read_lock();
6626
6627 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6628 if (!cgrp || !cgroup_tryget(cgrp))
6629 cgrp = ERR_PTR(-ENOENT);
6630
6631 rcu_read_unlock();
6632
6633 out_kernfs:
6634 kernfs_put(kn);
6635 out:
6636 return cgrp;
6637 }
6638 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6639
6640 /**
6641 * cgroup_get_from_fd - get a cgroup pointer from a fd
6642 * @fd: fd obtained by open(cgroup2_dir)
6643 *
6644 * Find the cgroup from a fd which should be obtained
6645 * by opening a cgroup directory. Returns a pointer to the
6646 * cgroup on success. ERR_PTR is returned if the cgroup
6647 * cannot be found.
6648 */
cgroup_get_from_fd(int fd)6649 struct cgroup *cgroup_get_from_fd(int fd)
6650 {
6651 struct cgroup *cgrp;
6652 struct file *f;
6653
6654 f = fget_raw(fd);
6655 if (!f)
6656 return ERR_PTR(-EBADF);
6657
6658 cgrp = cgroup_get_from_file(f);
6659 fput(f);
6660 return cgrp;
6661 }
6662 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6663
power_of_ten(int power)6664 static u64 power_of_ten(int power)
6665 {
6666 u64 v = 1;
6667 while (power--)
6668 v *= 10;
6669 return v;
6670 }
6671
6672 /**
6673 * cgroup_parse_float - parse a floating number
6674 * @input: input string
6675 * @dec_shift: number of decimal digits to shift
6676 * @v: output
6677 *
6678 * Parse a decimal floating point number in @input and store the result in
6679 * @v with decimal point right shifted @dec_shift times. For example, if
6680 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6681 * Returns 0 on success, -errno otherwise.
6682 *
6683 * There's nothing cgroup specific about this function except that it's
6684 * currently the only user.
6685 */
cgroup_parse_float(const char * input,unsigned dec_shift,s64 * v)6686 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6687 {
6688 s64 whole, frac = 0;
6689 int fstart = 0, fend = 0, flen;
6690
6691 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6692 return -EINVAL;
6693 if (frac < 0)
6694 return -EINVAL;
6695
6696 flen = fend > fstart ? fend - fstart : 0;
6697 if (flen < dec_shift)
6698 frac *= power_of_ten(dec_shift - flen);
6699 else
6700 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6701
6702 *v = whole * power_of_ten(dec_shift) + frac;
6703 return 0;
6704 }
6705
6706 /*
6707 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6708 * definition in cgroup-defs.h.
6709 */
6710 #ifdef CONFIG_SOCK_CGROUP_DATA
6711
cgroup_sk_alloc(struct sock_cgroup_data * skcd)6712 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6713 {
6714 struct cgroup *cgroup;
6715
6716 rcu_read_lock();
6717 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6718 if (in_interrupt()) {
6719 cgroup = &cgrp_dfl_root.cgrp;
6720 cgroup_get(cgroup);
6721 goto out;
6722 }
6723
6724 while (true) {
6725 struct css_set *cset;
6726
6727 cset = task_css_set(current);
6728 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6729 cgroup = cset->dfl_cgrp;
6730 break;
6731 }
6732 cpu_relax();
6733 }
6734 out:
6735 skcd->cgroup = cgroup;
6736 cgroup_bpf_get(cgroup);
6737 rcu_read_unlock();
6738 }
6739
cgroup_sk_clone(struct sock_cgroup_data * skcd)6740 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6741 {
6742 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6743
6744 /*
6745 * We might be cloning a socket which is left in an empty
6746 * cgroup and the cgroup might have already been rmdir'd.
6747 * Don't use cgroup_get_live().
6748 */
6749 cgroup_get(cgrp);
6750 cgroup_bpf_get(cgrp);
6751 }
6752
cgroup_sk_free(struct sock_cgroup_data * skcd)6753 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6754 {
6755 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6756
6757 cgroup_bpf_put(cgrp);
6758 cgroup_put(cgrp);
6759 }
6760
6761 #endif /* CONFIG_SOCK_CGROUP_DATA */
6762
6763 #ifdef CONFIG_SYSFS
show_delegatable_files(struct cftype * files,char * buf,ssize_t size,const char * prefix)6764 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6765 ssize_t size, const char *prefix)
6766 {
6767 struct cftype *cft;
6768 ssize_t ret = 0;
6769
6770 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6771 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6772 continue;
6773
6774 if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
6775 continue;
6776
6777 if (prefix)
6778 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6779
6780 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6781
6782 if (WARN_ON(ret >= size))
6783 break;
6784 }
6785
6786 return ret;
6787 }
6788
delegate_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)6789 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6790 char *buf)
6791 {
6792 struct cgroup_subsys *ss;
6793 int ssid;
6794 ssize_t ret = 0;
6795
6796 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6797 NULL);
6798
6799 for_each_subsys(ss, ssid)
6800 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6801 PAGE_SIZE - ret,
6802 cgroup_subsys_name[ssid]);
6803
6804 return ret;
6805 }
6806 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6807
features_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)6808 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6809 char *buf)
6810 {
6811 return snprintf(buf, PAGE_SIZE,
6812 "nsdelegate\n"
6813 "memory_localevents\n"
6814 "memory_recursiveprot\n");
6815 }
6816 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6817
6818 static struct attribute *cgroup_sysfs_attrs[] = {
6819 &cgroup_delegate_attr.attr,
6820 &cgroup_features_attr.attr,
6821 NULL,
6822 };
6823
6824 static const struct attribute_group cgroup_sysfs_attr_group = {
6825 .attrs = cgroup_sysfs_attrs,
6826 .name = "cgroup",
6827 };
6828
cgroup_sysfs_init(void)6829 static int __init cgroup_sysfs_init(void)
6830 {
6831 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6832 }
6833 subsys_initcall(cgroup_sysfs_init);
6834
6835 #endif /* CONFIG_SYSFS */
6836