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 static struct cftype cgroup_psi_files[];
221 
222 /* cgroup optional features */
223 enum cgroup_opt_features {
224 #ifdef CONFIG_PSI
225 	OPT_FEATURE_PRESSURE,
226 #endif
227 	OPT_FEATURE_COUNT
228 };
229 
230 static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
231 #ifdef CONFIG_PSI
232 	"pressure",
233 #endif
234 };
235 
236 static u16 cgroup_feature_disable_mask __read_mostly;
237 
238 static int cgroup_apply_control(struct cgroup *cgrp);
239 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
240 static void css_task_iter_skip(struct css_task_iter *it,
241 			       struct task_struct *task);
242 static int cgroup_destroy_locked(struct cgroup *cgrp);
243 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
244 					      struct cgroup_subsys *ss);
245 static void css_release(struct percpu_ref *ref);
246 static void kill_css(struct cgroup_subsys_state *css);
247 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
248 			      struct cgroup *cgrp, struct cftype cfts[],
249 			      bool is_add);
250 
251 /**
252  * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
253  * @ssid: subsys ID of interest
254  *
255  * cgroup_subsys_enabled() can only be used with literal subsys names which
256  * is fine for individual subsystems but unsuitable for cgroup core.  This
257  * is slower static_key_enabled() based test indexed by @ssid.
258  */
cgroup_ssid_enabled(int ssid)259 bool cgroup_ssid_enabled(int ssid)
260 {
261 	if (!CGROUP_HAS_SUBSYS_CONFIG)
262 		return false;
263 
264 	return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
265 }
266 
267 /**
268  * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
269  * @cgrp: the cgroup of interest
270  *
271  * The default hierarchy is the v2 interface of cgroup and this function
272  * can be used to test whether a cgroup is on the default hierarchy for
273  * cases where a subsystem should behave differently depending on the
274  * interface version.
275  *
276  * List of changed behaviors:
277  *
278  * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
279  *   and "name" are disallowed.
280  *
281  * - When mounting an existing superblock, mount options should match.
282  *
283  * - rename(2) is disallowed.
284  *
285  * - "tasks" is removed.  Everything should be at process granularity.  Use
286  *   "cgroup.procs" instead.
287  *
288  * - "cgroup.procs" is not sorted.  pids will be unique unless they got
289  *   recycled in-between reads.
290  *
291  * - "release_agent" and "notify_on_release" are removed.  Replacement
292  *   notification mechanism will be implemented.
293  *
294  * - "cgroup.clone_children" is removed.
295  *
296  * - "cgroup.subtree_populated" is available.  Its value is 0 if the cgroup
297  *   and its descendants contain no task; otherwise, 1.  The file also
298  *   generates kernfs notification which can be monitored through poll and
299  *   [di]notify when the value of the file changes.
300  *
301  * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
302  *   take masks of ancestors with non-empty cpus/mems, instead of being
303  *   moved to an ancestor.
304  *
305  * - cpuset: a task can be moved into an empty cpuset, and again it takes
306  *   masks of ancestors.
307  *
308  * - blkcg: blk-throttle becomes properly hierarchical.
309  *
310  * - debug: disallowed on the default hierarchy.
311  */
cgroup_on_dfl(const struct cgroup * cgrp)312 bool cgroup_on_dfl(const struct cgroup *cgrp)
313 {
314 	return cgrp->root == &cgrp_dfl_root;
315 }
316 
317 /* IDR wrappers which synchronize using cgroup_idr_lock */
cgroup_idr_alloc(struct idr * idr,void * ptr,int start,int end,gfp_t gfp_mask)318 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
319 			    gfp_t gfp_mask)
320 {
321 	int ret;
322 
323 	idr_preload(gfp_mask);
324 	spin_lock_bh(&cgroup_idr_lock);
325 	ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
326 	spin_unlock_bh(&cgroup_idr_lock);
327 	idr_preload_end();
328 	return ret;
329 }
330 
cgroup_idr_replace(struct idr * idr,void * ptr,int id)331 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
332 {
333 	void *ret;
334 
335 	spin_lock_bh(&cgroup_idr_lock);
336 	ret = idr_replace(idr, ptr, id);
337 	spin_unlock_bh(&cgroup_idr_lock);
338 	return ret;
339 }
340 
cgroup_idr_remove(struct idr * idr,int id)341 static void cgroup_idr_remove(struct idr *idr, int id)
342 {
343 	spin_lock_bh(&cgroup_idr_lock);
344 	idr_remove(idr, id);
345 	spin_unlock_bh(&cgroup_idr_lock);
346 }
347 
cgroup_has_tasks(struct cgroup * cgrp)348 static bool cgroup_has_tasks(struct cgroup *cgrp)
349 {
350 	return cgrp->nr_populated_csets;
351 }
352 
cgroup_is_threaded(struct cgroup * cgrp)353 bool cgroup_is_threaded(struct cgroup *cgrp)
354 {
355 	return cgrp->dom_cgrp != cgrp;
356 }
357 
358 /* can @cgrp host both domain and threaded children? */
cgroup_is_mixable(struct cgroup * cgrp)359 static bool cgroup_is_mixable(struct cgroup *cgrp)
360 {
361 	/*
362 	 * Root isn't under domain level resource control exempting it from
363 	 * the no-internal-process constraint, so it can serve as a thread
364 	 * root and a parent of resource domains at the same time.
365 	 */
366 	return !cgroup_parent(cgrp);
367 }
368 
369 /* can @cgrp become a thread root? Should always be true for a thread root */
cgroup_can_be_thread_root(struct cgroup * cgrp)370 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
371 {
372 	/* mixables don't care */
373 	if (cgroup_is_mixable(cgrp))
374 		return true;
375 
376 	/* domain roots can't be nested under threaded */
377 	if (cgroup_is_threaded(cgrp))
378 		return false;
379 
380 	/* can only have either domain or threaded children */
381 	if (cgrp->nr_populated_domain_children)
382 		return false;
383 
384 	/* and no domain controllers can be enabled */
385 	if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
386 		return false;
387 
388 	return true;
389 }
390 
391 /* is @cgrp root of a threaded subtree? */
cgroup_is_thread_root(struct cgroup * cgrp)392 bool cgroup_is_thread_root(struct cgroup *cgrp)
393 {
394 	/* thread root should be a domain */
395 	if (cgroup_is_threaded(cgrp))
396 		return false;
397 
398 	/* a domain w/ threaded children is a thread root */
399 	if (cgrp->nr_threaded_children)
400 		return true;
401 
402 	/*
403 	 * A domain which has tasks and explicit threaded controllers
404 	 * enabled is a thread root.
405 	 */
406 	if (cgroup_has_tasks(cgrp) &&
407 	    (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
408 		return true;
409 
410 	return false;
411 }
412 
413 /* a domain which isn't connected to the root w/o brekage can't be used */
cgroup_is_valid_domain(struct cgroup * cgrp)414 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
415 {
416 	/* the cgroup itself can be a thread root */
417 	if (cgroup_is_threaded(cgrp))
418 		return false;
419 
420 	/* but the ancestors can't be unless mixable */
421 	while ((cgrp = cgroup_parent(cgrp))) {
422 		if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
423 			return false;
424 		if (cgroup_is_threaded(cgrp))
425 			return false;
426 	}
427 
428 	return true;
429 }
430 
431 /* subsystems visibly enabled on a cgroup */
cgroup_control(struct cgroup * cgrp)432 static u16 cgroup_control(struct cgroup *cgrp)
433 {
434 	struct cgroup *parent = cgroup_parent(cgrp);
435 	u16 root_ss_mask = cgrp->root->subsys_mask;
436 
437 	if (parent) {
438 		u16 ss_mask = parent->subtree_control;
439 
440 		/* threaded cgroups can only have threaded controllers */
441 		if (cgroup_is_threaded(cgrp))
442 			ss_mask &= cgrp_dfl_threaded_ss_mask;
443 		return ss_mask;
444 	}
445 
446 	if (cgroup_on_dfl(cgrp))
447 		root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
448 				  cgrp_dfl_implicit_ss_mask);
449 	return root_ss_mask;
450 }
451 
452 /* subsystems enabled on a cgroup */
cgroup_ss_mask(struct cgroup * cgrp)453 static u16 cgroup_ss_mask(struct cgroup *cgrp)
454 {
455 	struct cgroup *parent = cgroup_parent(cgrp);
456 
457 	if (parent) {
458 		u16 ss_mask = parent->subtree_ss_mask;
459 
460 		/* threaded cgroups can only have threaded controllers */
461 		if (cgroup_is_threaded(cgrp))
462 			ss_mask &= cgrp_dfl_threaded_ss_mask;
463 		return ss_mask;
464 	}
465 
466 	return cgrp->root->subsys_mask;
467 }
468 
469 /**
470  * cgroup_css - obtain a cgroup's css for the specified subsystem
471  * @cgrp: the cgroup of interest
472  * @ss: the subsystem of interest (%NULL returns @cgrp->self)
473  *
474  * Return @cgrp's css (cgroup_subsys_state) associated with @ss.  This
475  * function must be called either under cgroup_mutex or rcu_read_lock() and
476  * the caller is responsible for pinning the returned css if it wants to
477  * keep accessing it outside the said locks.  This function may return
478  * %NULL if @cgrp doesn't have @subsys_id enabled.
479  */
cgroup_css(struct cgroup * cgrp,struct cgroup_subsys * ss)480 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
481 					      struct cgroup_subsys *ss)
482 {
483 	if (CGROUP_HAS_SUBSYS_CONFIG && ss)
484 		return rcu_dereference_check(cgrp->subsys[ss->id],
485 					lockdep_is_held(&cgroup_mutex));
486 	else
487 		return &cgrp->self;
488 }
489 
490 /**
491  * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
492  * @cgrp: the cgroup of interest
493  * @ss: the subsystem of interest
494  *
495  * Find and get @cgrp's css associated with @ss.  If the css doesn't exist
496  * or is offline, %NULL is returned.
497  */
cgroup_tryget_css(struct cgroup * cgrp,struct cgroup_subsys * ss)498 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
499 						     struct cgroup_subsys *ss)
500 {
501 	struct cgroup_subsys_state *css;
502 
503 	rcu_read_lock();
504 	css = cgroup_css(cgrp, ss);
505 	if (css && !css_tryget_online(css))
506 		css = NULL;
507 	rcu_read_unlock();
508 
509 	return css;
510 }
511 
512 /**
513  * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
514  * @cgrp: the cgroup of interest
515  * @ss: the subsystem of interest (%NULL returns @cgrp->self)
516  *
517  * Similar to cgroup_css() but returns the effective css, which is defined
518  * as the matching css of the nearest ancestor including self which has @ss
519  * enabled.  If @ss is associated with the hierarchy @cgrp is on, this
520  * function is guaranteed to return non-NULL css.
521  */
cgroup_e_css_by_mask(struct cgroup * cgrp,struct cgroup_subsys * ss)522 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
523 							struct cgroup_subsys *ss)
524 {
525 	lockdep_assert_held(&cgroup_mutex);
526 
527 	if (!ss)
528 		return &cgrp->self;
529 
530 	/*
531 	 * This function is used while updating css associations and thus
532 	 * can't test the csses directly.  Test ss_mask.
533 	 */
534 	while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
535 		cgrp = cgroup_parent(cgrp);
536 		if (!cgrp)
537 			return NULL;
538 	}
539 
540 	return cgroup_css(cgrp, ss);
541 }
542 
543 /**
544  * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
545  * @cgrp: the cgroup of interest
546  * @ss: the subsystem of interest
547  *
548  * Find and get the effective css of @cgrp for @ss.  The effective css is
549  * defined as the matching css of the nearest ancestor including self which
550  * has @ss enabled.  If @ss is not mounted on the hierarchy @cgrp is on,
551  * the root css is returned, so this function always returns a valid css.
552  *
553  * The returned css is not guaranteed to be online, and therefore it is the
554  * callers responsibility to try get a reference for it.
555  */
cgroup_e_css(struct cgroup * cgrp,struct cgroup_subsys * ss)556 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
557 					 struct cgroup_subsys *ss)
558 {
559 	struct cgroup_subsys_state *css;
560 
561 	if (!CGROUP_HAS_SUBSYS_CONFIG)
562 		return NULL;
563 
564 	do {
565 		css = cgroup_css(cgrp, ss);
566 
567 		if (css)
568 			return css;
569 		cgrp = cgroup_parent(cgrp);
570 	} while (cgrp);
571 
572 	return init_css_set.subsys[ss->id];
573 }
574 
575 /**
576  * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
577  * @cgrp: the cgroup of interest
578  * @ss: the subsystem of interest
579  *
580  * Find and get the effective css of @cgrp for @ss.  The effective css is
581  * defined as the matching css of the nearest ancestor including self which
582  * has @ss enabled.  If @ss is not mounted on the hierarchy @cgrp is on,
583  * the root css is returned, so this function always returns a valid css.
584  * The returned css must be put using css_put().
585  */
cgroup_get_e_css(struct cgroup * cgrp,struct cgroup_subsys * ss)586 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
587 					     struct cgroup_subsys *ss)
588 {
589 	struct cgroup_subsys_state *css;
590 
591 	if (!CGROUP_HAS_SUBSYS_CONFIG)
592 		return NULL;
593 
594 	rcu_read_lock();
595 
596 	do {
597 		css = cgroup_css(cgrp, ss);
598 
599 		if (css && css_tryget_online(css))
600 			goto out_unlock;
601 		cgrp = cgroup_parent(cgrp);
602 	} while (cgrp);
603 
604 	css = init_css_set.subsys[ss->id];
605 	css_get(css);
606 out_unlock:
607 	rcu_read_unlock();
608 	return css;
609 }
610 EXPORT_SYMBOL_GPL(cgroup_get_e_css);
611 
cgroup_get_live(struct cgroup * cgrp)612 static void cgroup_get_live(struct cgroup *cgrp)
613 {
614 	WARN_ON_ONCE(cgroup_is_dead(cgrp));
615 	css_get(&cgrp->self);
616 }
617 
618 /**
619  * __cgroup_task_count - count the number of tasks in a cgroup. The caller
620  * is responsible for taking the css_set_lock.
621  * @cgrp: the cgroup in question
622  */
__cgroup_task_count(const struct cgroup * cgrp)623 int __cgroup_task_count(const struct cgroup *cgrp)
624 {
625 	int count = 0;
626 	struct cgrp_cset_link *link;
627 
628 	lockdep_assert_held(&css_set_lock);
629 
630 	list_for_each_entry(link, &cgrp->cset_links, cset_link)
631 		count += link->cset->nr_tasks;
632 
633 	return count;
634 }
635 
636 /**
637  * cgroup_task_count - count the number of tasks in a cgroup.
638  * @cgrp: the cgroup in question
639  */
cgroup_task_count(const struct cgroup * cgrp)640 int cgroup_task_count(const struct cgroup *cgrp)
641 {
642 	int count;
643 
644 	spin_lock_irq(&css_set_lock);
645 	count = __cgroup_task_count(cgrp);
646 	spin_unlock_irq(&css_set_lock);
647 
648 	return count;
649 }
650 
of_css(struct kernfs_open_file * of)651 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
652 {
653 	struct cgroup *cgrp = of->kn->parent->priv;
654 	struct cftype *cft = of_cft(of);
655 
656 	/*
657 	 * This is open and unprotected implementation of cgroup_css().
658 	 * seq_css() is only called from a kernfs file operation which has
659 	 * an active reference on the file.  Because all the subsystem
660 	 * files are drained before a css is disassociated with a cgroup,
661 	 * the matching css from the cgroup's subsys table is guaranteed to
662 	 * be and stay valid until the enclosing operation is complete.
663 	 */
664 	if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
665 		return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
666 	else
667 		return &cgrp->self;
668 }
669 EXPORT_SYMBOL_GPL(of_css);
670 
671 /**
672  * for_each_css - iterate all css's of a cgroup
673  * @css: the iteration cursor
674  * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
675  * @cgrp: the target cgroup to iterate css's of
676  *
677  * Should be called under cgroup_[tree_]mutex.
678  */
679 #define for_each_css(css, ssid, cgrp)					\
680 	for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)	\
681 		if (!((css) = rcu_dereference_check(			\
682 				(cgrp)->subsys[(ssid)],			\
683 				lockdep_is_held(&cgroup_mutex)))) { }	\
684 		else
685 
686 /**
687  * for_each_e_css - iterate all effective css's of a cgroup
688  * @css: the iteration cursor
689  * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
690  * @cgrp: the target cgroup to iterate css's of
691  *
692  * Should be called under cgroup_[tree_]mutex.
693  */
694 #define for_each_e_css(css, ssid, cgrp)					    \
695 	for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)	    \
696 		if (!((css) = cgroup_e_css_by_mask(cgrp,		    \
697 						   cgroup_subsys[(ssid)]))) \
698 			;						    \
699 		else
700 
701 /**
702  * do_each_subsys_mask - filter for_each_subsys with a bitmask
703  * @ss: the iteration cursor
704  * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
705  * @ss_mask: the bitmask
706  *
707  * The block will only run for cases where the ssid-th bit (1 << ssid) of
708  * @ss_mask is set.
709  */
710 #define do_each_subsys_mask(ss, ssid, ss_mask) do {			\
711 	unsigned long __ss_mask = (ss_mask);				\
712 	if (!CGROUP_HAS_SUBSYS_CONFIG) {				\
713 		(ssid) = 0;						\
714 		break;							\
715 	}								\
716 	for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) {	\
717 		(ss) = cgroup_subsys[ssid];				\
718 		{
719 
720 #define while_each_subsys_mask()					\
721 		}							\
722 	}								\
723 } while (false)
724 
725 /* iterate over child cgrps, lock should be held throughout iteration */
726 #define cgroup_for_each_live_child(child, cgrp)				\
727 	list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
728 		if (({ lockdep_assert_held(&cgroup_mutex);		\
729 		       cgroup_is_dead(child); }))			\
730 			;						\
731 		else
732 
733 /* walk live descendants in pre order */
734 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)		\
735 	css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL))	\
736 		if (({ lockdep_assert_held(&cgroup_mutex);		\
737 		       (dsct) = (d_css)->cgroup;			\
738 		       cgroup_is_dead(dsct); }))			\
739 			;						\
740 		else
741 
742 /* walk live descendants in postorder */
743 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp)		\
744 	css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL))	\
745 		if (({ lockdep_assert_held(&cgroup_mutex);		\
746 		       (dsct) = (d_css)->cgroup;			\
747 		       cgroup_is_dead(dsct); }))			\
748 			;						\
749 		else
750 
751 /*
752  * The default css_set - used by init and its children prior to any
753  * hierarchies being mounted. It contains a pointer to the root state
754  * for each subsystem. Also used to anchor the list of css_sets. Not
755  * reference-counted, to improve performance when child cgroups
756  * haven't been created.
757  */
758 struct css_set init_css_set = {
759 	.refcount		= REFCOUNT_INIT(1),
760 	.dom_cset		= &init_css_set,
761 	.tasks			= LIST_HEAD_INIT(init_css_set.tasks),
762 	.mg_tasks		= LIST_HEAD_INIT(init_css_set.mg_tasks),
763 	.dying_tasks		= LIST_HEAD_INIT(init_css_set.dying_tasks),
764 	.task_iters		= LIST_HEAD_INIT(init_css_set.task_iters),
765 	.threaded_csets		= LIST_HEAD_INIT(init_css_set.threaded_csets),
766 	.cgrp_links		= LIST_HEAD_INIT(init_css_set.cgrp_links),
767 	.mg_src_preload_node	= LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
768 	.mg_dst_preload_node	= LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
769 	.mg_node		= LIST_HEAD_INIT(init_css_set.mg_node),
770 
771 	/*
772 	 * The following field is re-initialized when this cset gets linked
773 	 * in cgroup_init().  However, let's initialize the field
774 	 * statically too so that the default cgroup can be accessed safely
775 	 * early during boot.
776 	 */
777 	.dfl_cgrp		= &cgrp_dfl_root.cgrp,
778 };
779 
780 static int css_set_count	= 1;	/* 1 for init_css_set */
781 
css_set_threaded(struct css_set * cset)782 static bool css_set_threaded(struct css_set *cset)
783 {
784 	return cset->dom_cset != cset;
785 }
786 
787 /**
788  * css_set_populated - does a css_set contain any tasks?
789  * @cset: target css_set
790  *
791  * css_set_populated() should be the same as !!cset->nr_tasks at steady
792  * state. However, css_set_populated() can be called while a task is being
793  * added to or removed from the linked list before the nr_tasks is
794  * properly updated. Hence, we can't just look at ->nr_tasks here.
795  */
css_set_populated(struct css_set * cset)796 static bool css_set_populated(struct css_set *cset)
797 {
798 	lockdep_assert_held(&css_set_lock);
799 
800 	return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
801 }
802 
803 /**
804  * cgroup_update_populated - update the populated count of a cgroup
805  * @cgrp: the target cgroup
806  * @populated: inc or dec populated count
807  *
808  * One of the css_sets associated with @cgrp is either getting its first
809  * task or losing the last.  Update @cgrp->nr_populated_* accordingly.  The
810  * count is propagated towards root so that a given cgroup's
811  * nr_populated_children is zero iff none of its descendants contain any
812  * tasks.
813  *
814  * @cgrp's interface file "cgroup.populated" is zero if both
815  * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
816  * 1 otherwise.  When the sum changes from or to zero, userland is notified
817  * that the content of the interface file has changed.  This can be used to
818  * detect when @cgrp and its descendants become populated or empty.
819  */
cgroup_update_populated(struct cgroup * cgrp,bool populated)820 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
821 {
822 	struct cgroup *child = NULL;
823 	int adj = populated ? 1 : -1;
824 
825 	lockdep_assert_held(&css_set_lock);
826 
827 	do {
828 		bool was_populated = cgroup_is_populated(cgrp);
829 
830 		if (!child) {
831 			cgrp->nr_populated_csets += adj;
832 		} else {
833 			if (cgroup_is_threaded(child))
834 				cgrp->nr_populated_threaded_children += adj;
835 			else
836 				cgrp->nr_populated_domain_children += adj;
837 		}
838 
839 		if (was_populated == cgroup_is_populated(cgrp))
840 			break;
841 
842 		cgroup1_check_for_release(cgrp);
843 		TRACE_CGROUP_PATH(notify_populated, cgrp,
844 				  cgroup_is_populated(cgrp));
845 		cgroup_file_notify(&cgrp->events_file);
846 
847 		child = cgrp;
848 		cgrp = cgroup_parent(cgrp);
849 	} while (cgrp);
850 }
851 
852 /**
853  * css_set_update_populated - update populated state of a css_set
854  * @cset: target css_set
855  * @populated: whether @cset is populated or depopulated
856  *
857  * @cset is either getting the first task or losing the last.  Update the
858  * populated counters of all associated cgroups accordingly.
859  */
css_set_update_populated(struct css_set * cset,bool populated)860 static void css_set_update_populated(struct css_set *cset, bool populated)
861 {
862 	struct cgrp_cset_link *link;
863 
864 	lockdep_assert_held(&css_set_lock);
865 
866 	list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
867 		cgroup_update_populated(link->cgrp, populated);
868 }
869 
870 /*
871  * @task is leaving, advance task iterators which are pointing to it so
872  * that they can resume at the next position.  Advancing an iterator might
873  * remove it from the list, use safe walk.  See css_task_iter_skip() for
874  * details.
875  */
css_set_skip_task_iters(struct css_set * cset,struct task_struct * task)876 static void css_set_skip_task_iters(struct css_set *cset,
877 				    struct task_struct *task)
878 {
879 	struct css_task_iter *it, *pos;
880 
881 	list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
882 		css_task_iter_skip(it, task);
883 }
884 
885 /**
886  * css_set_move_task - move a task from one css_set to another
887  * @task: task being moved
888  * @from_cset: css_set @task currently belongs to (may be NULL)
889  * @to_cset: new css_set @task is being moved to (may be NULL)
890  * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
891  *
892  * Move @task from @from_cset to @to_cset.  If @task didn't belong to any
893  * css_set, @from_cset can be NULL.  If @task is being disassociated
894  * instead of moved, @to_cset can be NULL.
895  *
896  * This function automatically handles populated counter updates and
897  * css_task_iter adjustments but the caller is responsible for managing
898  * @from_cset and @to_cset's reference counts.
899  */
css_set_move_task(struct task_struct * task,struct css_set * from_cset,struct css_set * to_cset,bool use_mg_tasks)900 static void css_set_move_task(struct task_struct *task,
901 			      struct css_set *from_cset, struct css_set *to_cset,
902 			      bool use_mg_tasks)
903 {
904 	lockdep_assert_held(&css_set_lock);
905 
906 	if (to_cset && !css_set_populated(to_cset))
907 		css_set_update_populated(to_cset, true);
908 
909 	if (from_cset) {
910 		WARN_ON_ONCE(list_empty(&task->cg_list));
911 
912 		css_set_skip_task_iters(from_cset, task);
913 		list_del_init(&task->cg_list);
914 		if (!css_set_populated(from_cset))
915 			css_set_update_populated(from_cset, false);
916 	} else {
917 		WARN_ON_ONCE(!list_empty(&task->cg_list));
918 	}
919 
920 	if (to_cset) {
921 		/*
922 		 * We are synchronized through cgroup_threadgroup_rwsem
923 		 * against PF_EXITING setting such that we can't race
924 		 * against cgroup_exit()/cgroup_free() dropping the css_set.
925 		 */
926 		WARN_ON_ONCE(task->flags & PF_EXITING);
927 
928 		cgroup_move_task(task, to_cset);
929 		list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
930 							     &to_cset->tasks);
931 	}
932 }
933 
934 /*
935  * hash table for cgroup groups. This improves the performance to find
936  * an existing css_set. This hash doesn't (currently) take into
937  * account cgroups in empty hierarchies.
938  */
939 #define CSS_SET_HASH_BITS	7
940 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
941 
css_set_hash(struct cgroup_subsys_state * css[])942 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
943 {
944 	unsigned long key = 0UL;
945 	struct cgroup_subsys *ss;
946 	int i;
947 
948 	for_each_subsys(ss, i)
949 		key += (unsigned long)css[i];
950 	key = (key >> 16) ^ key;
951 
952 	return key;
953 }
954 
put_css_set_locked(struct css_set * cset)955 void put_css_set_locked(struct css_set *cset)
956 {
957 	struct cgrp_cset_link *link, *tmp_link;
958 	struct cgroup_subsys *ss;
959 	int ssid;
960 
961 	lockdep_assert_held(&css_set_lock);
962 
963 	if (!refcount_dec_and_test(&cset->refcount))
964 		return;
965 
966 	WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
967 
968 	/* This css_set is dead. Unlink it and release cgroup and css refs */
969 	for_each_subsys(ss, ssid) {
970 		list_del(&cset->e_cset_node[ssid]);
971 		css_put(cset->subsys[ssid]);
972 	}
973 	hash_del(&cset->hlist);
974 	css_set_count--;
975 
976 	list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
977 		list_del(&link->cset_link);
978 		list_del(&link->cgrp_link);
979 		if (cgroup_parent(link->cgrp))
980 			cgroup_put(link->cgrp);
981 		kfree(link);
982 	}
983 
984 	if (css_set_threaded(cset)) {
985 		list_del(&cset->threaded_csets_node);
986 		put_css_set_locked(cset->dom_cset);
987 	}
988 
989 	kfree_rcu(cset, rcu_head);
990 }
991 
992 /**
993  * compare_css_sets - helper function for find_existing_css_set().
994  * @cset: candidate css_set being tested
995  * @old_cset: existing css_set for a task
996  * @new_cgrp: cgroup that's being entered by the task
997  * @template: desired set of css pointers in css_set (pre-calculated)
998  *
999  * Returns true if "cset" matches "old_cset" except for the hierarchy
1000  * which "new_cgrp" belongs to, for which it should match "new_cgrp".
1001  */
compare_css_sets(struct css_set * cset,struct css_set * old_cset,struct cgroup * new_cgrp,struct cgroup_subsys_state * template[])1002 static bool compare_css_sets(struct css_set *cset,
1003 			     struct css_set *old_cset,
1004 			     struct cgroup *new_cgrp,
1005 			     struct cgroup_subsys_state *template[])
1006 {
1007 	struct cgroup *new_dfl_cgrp;
1008 	struct list_head *l1, *l2;
1009 
1010 	/*
1011 	 * On the default hierarchy, there can be csets which are
1012 	 * associated with the same set of cgroups but different csses.
1013 	 * Let's first ensure that csses match.
1014 	 */
1015 	if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1016 		return false;
1017 
1018 
1019 	/* @cset's domain should match the default cgroup's */
1020 	if (cgroup_on_dfl(new_cgrp))
1021 		new_dfl_cgrp = new_cgrp;
1022 	else
1023 		new_dfl_cgrp = old_cset->dfl_cgrp;
1024 
1025 	if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1026 		return false;
1027 
1028 	/*
1029 	 * Compare cgroup pointers in order to distinguish between
1030 	 * different cgroups in hierarchies.  As different cgroups may
1031 	 * share the same effective css, this comparison is always
1032 	 * necessary.
1033 	 */
1034 	l1 = &cset->cgrp_links;
1035 	l2 = &old_cset->cgrp_links;
1036 	while (1) {
1037 		struct cgrp_cset_link *link1, *link2;
1038 		struct cgroup *cgrp1, *cgrp2;
1039 
1040 		l1 = l1->next;
1041 		l2 = l2->next;
1042 		/* See if we reached the end - both lists are equal length. */
1043 		if (l1 == &cset->cgrp_links) {
1044 			BUG_ON(l2 != &old_cset->cgrp_links);
1045 			break;
1046 		} else {
1047 			BUG_ON(l2 == &old_cset->cgrp_links);
1048 		}
1049 		/* Locate the cgroups associated with these links. */
1050 		link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1051 		link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1052 		cgrp1 = link1->cgrp;
1053 		cgrp2 = link2->cgrp;
1054 		/* Hierarchies should be linked in the same order. */
1055 		BUG_ON(cgrp1->root != cgrp2->root);
1056 
1057 		/*
1058 		 * If this hierarchy is the hierarchy of the cgroup
1059 		 * that's changing, then we need to check that this
1060 		 * css_set points to the new cgroup; if it's any other
1061 		 * hierarchy, then this css_set should point to the
1062 		 * same cgroup as the old css_set.
1063 		 */
1064 		if (cgrp1->root == new_cgrp->root) {
1065 			if (cgrp1 != new_cgrp)
1066 				return false;
1067 		} else {
1068 			if (cgrp1 != cgrp2)
1069 				return false;
1070 		}
1071 	}
1072 	return true;
1073 }
1074 
1075 /**
1076  * find_existing_css_set - init css array and find the matching css_set
1077  * @old_cset: the css_set that we're using before the cgroup transition
1078  * @cgrp: the cgroup that we're moving into
1079  * @template: out param for the new set of csses, should be clear on entry
1080  */
find_existing_css_set(struct css_set * old_cset,struct cgroup * cgrp,struct cgroup_subsys_state * template[])1081 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1082 					struct cgroup *cgrp,
1083 					struct cgroup_subsys_state *template[])
1084 {
1085 	struct cgroup_root *root = cgrp->root;
1086 	struct cgroup_subsys *ss;
1087 	struct css_set *cset;
1088 	unsigned long key;
1089 	int i;
1090 
1091 	/*
1092 	 * Build the set of subsystem state objects that we want to see in the
1093 	 * new css_set. While subsystems can change globally, the entries here
1094 	 * won't change, so no need for locking.
1095 	 */
1096 	for_each_subsys(ss, i) {
1097 		if (root->subsys_mask & (1UL << i)) {
1098 			/*
1099 			 * @ss is in this hierarchy, so we want the
1100 			 * effective css from @cgrp.
1101 			 */
1102 			template[i] = cgroup_e_css_by_mask(cgrp, ss);
1103 		} else {
1104 			/*
1105 			 * @ss is not in this hierarchy, so we don't want
1106 			 * to change the css.
1107 			 */
1108 			template[i] = old_cset->subsys[i];
1109 		}
1110 	}
1111 
1112 	key = css_set_hash(template);
1113 	hash_for_each_possible(css_set_table, cset, hlist, key) {
1114 		if (!compare_css_sets(cset, old_cset, cgrp, template))
1115 			continue;
1116 
1117 		/* This css_set matches what we need */
1118 		return cset;
1119 	}
1120 
1121 	/* No existing cgroup group matched */
1122 	return NULL;
1123 }
1124 
free_cgrp_cset_links(struct list_head * links_to_free)1125 static void free_cgrp_cset_links(struct list_head *links_to_free)
1126 {
1127 	struct cgrp_cset_link *link, *tmp_link;
1128 
1129 	list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1130 		list_del(&link->cset_link);
1131 		kfree(link);
1132 	}
1133 }
1134 
1135 /**
1136  * allocate_cgrp_cset_links - allocate cgrp_cset_links
1137  * @count: the number of links to allocate
1138  * @tmp_links: list_head the allocated links are put on
1139  *
1140  * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1141  * through ->cset_link.  Returns 0 on success or -errno.
1142  */
allocate_cgrp_cset_links(int count,struct list_head * tmp_links)1143 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1144 {
1145 	struct cgrp_cset_link *link;
1146 	int i;
1147 
1148 	INIT_LIST_HEAD(tmp_links);
1149 
1150 	for (i = 0; i < count; i++) {
1151 		link = kzalloc(sizeof(*link), GFP_KERNEL);
1152 		if (!link) {
1153 			free_cgrp_cset_links(tmp_links);
1154 			return -ENOMEM;
1155 		}
1156 		list_add(&link->cset_link, tmp_links);
1157 	}
1158 	return 0;
1159 }
1160 
1161 /**
1162  * link_css_set - a helper function to link a css_set to a cgroup
1163  * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1164  * @cset: the css_set to be linked
1165  * @cgrp: the destination cgroup
1166  */
link_css_set(struct list_head * tmp_links,struct css_set * cset,struct cgroup * cgrp)1167 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1168 			 struct cgroup *cgrp)
1169 {
1170 	struct cgrp_cset_link *link;
1171 
1172 	BUG_ON(list_empty(tmp_links));
1173 
1174 	if (cgroup_on_dfl(cgrp))
1175 		cset->dfl_cgrp = cgrp;
1176 
1177 	link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1178 	link->cset = cset;
1179 	link->cgrp = cgrp;
1180 
1181 	/*
1182 	 * Always add links to the tail of the lists so that the lists are
1183 	 * in chronological order.
1184 	 */
1185 	list_move_tail(&link->cset_link, &cgrp->cset_links);
1186 	list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1187 
1188 	if (cgroup_parent(cgrp))
1189 		cgroup_get_live(cgrp);
1190 }
1191 
1192 /**
1193  * find_css_set - return a new css_set with one cgroup updated
1194  * @old_cset: the baseline css_set
1195  * @cgrp: the cgroup to be updated
1196  *
1197  * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1198  * substituted into the appropriate hierarchy.
1199  */
find_css_set(struct css_set * old_cset,struct cgroup * cgrp)1200 static struct css_set *find_css_set(struct css_set *old_cset,
1201 				    struct cgroup *cgrp)
1202 {
1203 	struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1204 	struct css_set *cset;
1205 	struct list_head tmp_links;
1206 	struct cgrp_cset_link *link;
1207 	struct cgroup_subsys *ss;
1208 	unsigned long key;
1209 	int ssid;
1210 
1211 	lockdep_assert_held(&cgroup_mutex);
1212 
1213 	/* First see if we already have a cgroup group that matches
1214 	 * the desired set */
1215 	spin_lock_irq(&css_set_lock);
1216 	cset = find_existing_css_set(old_cset, cgrp, template);
1217 	if (cset)
1218 		get_css_set(cset);
1219 	spin_unlock_irq(&css_set_lock);
1220 
1221 	if (cset)
1222 		return cset;
1223 
1224 	cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1225 	if (!cset)
1226 		return NULL;
1227 
1228 	/* Allocate all the cgrp_cset_link objects that we'll need */
1229 	if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1230 		kfree(cset);
1231 		return NULL;
1232 	}
1233 
1234 	refcount_set(&cset->refcount, 1);
1235 	cset->dom_cset = cset;
1236 	INIT_LIST_HEAD(&cset->tasks);
1237 	INIT_LIST_HEAD(&cset->mg_tasks);
1238 	INIT_LIST_HEAD(&cset->dying_tasks);
1239 	INIT_LIST_HEAD(&cset->task_iters);
1240 	INIT_LIST_HEAD(&cset->threaded_csets);
1241 	INIT_HLIST_NODE(&cset->hlist);
1242 	INIT_LIST_HEAD(&cset->cgrp_links);
1243 	INIT_LIST_HEAD(&cset->mg_src_preload_node);
1244 	INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1245 	INIT_LIST_HEAD(&cset->mg_node);
1246 
1247 	/* Copy the set of subsystem state objects generated in
1248 	 * find_existing_css_set() */
1249 	memcpy(cset->subsys, template, sizeof(cset->subsys));
1250 
1251 	spin_lock_irq(&css_set_lock);
1252 	/* Add reference counts and links from the new css_set. */
1253 	list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1254 		struct cgroup *c = link->cgrp;
1255 
1256 		if (c->root == cgrp->root)
1257 			c = cgrp;
1258 		link_css_set(&tmp_links, cset, c);
1259 	}
1260 
1261 	BUG_ON(!list_empty(&tmp_links));
1262 
1263 	css_set_count++;
1264 
1265 	/* Add @cset to the hash table */
1266 	key = css_set_hash(cset->subsys);
1267 	hash_add(css_set_table, &cset->hlist, key);
1268 
1269 	for_each_subsys(ss, ssid) {
1270 		struct cgroup_subsys_state *css = cset->subsys[ssid];
1271 
1272 		list_add_tail(&cset->e_cset_node[ssid],
1273 			      &css->cgroup->e_csets[ssid]);
1274 		css_get(css);
1275 	}
1276 
1277 	spin_unlock_irq(&css_set_lock);
1278 
1279 	/*
1280 	 * If @cset should be threaded, look up the matching dom_cset and
1281 	 * link them up.  We first fully initialize @cset then look for the
1282 	 * dom_cset.  It's simpler this way and safe as @cset is guaranteed
1283 	 * to stay empty until we return.
1284 	 */
1285 	if (cgroup_is_threaded(cset->dfl_cgrp)) {
1286 		struct css_set *dcset;
1287 
1288 		dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1289 		if (!dcset) {
1290 			put_css_set(cset);
1291 			return NULL;
1292 		}
1293 
1294 		spin_lock_irq(&css_set_lock);
1295 		cset->dom_cset = dcset;
1296 		list_add_tail(&cset->threaded_csets_node,
1297 			      &dcset->threaded_csets);
1298 		spin_unlock_irq(&css_set_lock);
1299 	}
1300 
1301 	return cset;
1302 }
1303 
cgroup_root_from_kf(struct kernfs_root * kf_root)1304 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1305 {
1306 	struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1307 
1308 	return root_cgrp->root;
1309 }
1310 
cgroup_favor_dynmods(struct cgroup_root * root,bool favor)1311 void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1312 {
1313 	bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1314 
1315 	/* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
1316 	if (favor && !favoring) {
1317 		rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1318 		root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1319 	} else if (!favor && favoring) {
1320 		rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1321 		root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1322 	}
1323 }
1324 
cgroup_init_root_id(struct cgroup_root * root)1325 static int cgroup_init_root_id(struct cgroup_root *root)
1326 {
1327 	int id;
1328 
1329 	lockdep_assert_held(&cgroup_mutex);
1330 
1331 	id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1332 	if (id < 0)
1333 		return id;
1334 
1335 	root->hierarchy_id = id;
1336 	return 0;
1337 }
1338 
cgroup_exit_root_id(struct cgroup_root * root)1339 static void cgroup_exit_root_id(struct cgroup_root *root)
1340 {
1341 	lockdep_assert_held(&cgroup_mutex);
1342 
1343 	idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1344 }
1345 
cgroup_free_root(struct cgroup_root * root)1346 void cgroup_free_root(struct cgroup_root *root)
1347 {
1348 	kfree(root);
1349 }
1350 
cgroup_destroy_root(struct cgroup_root * root)1351 static void cgroup_destroy_root(struct cgroup_root *root)
1352 {
1353 	struct cgroup *cgrp = &root->cgrp;
1354 	struct cgrp_cset_link *link, *tmp_link;
1355 
1356 	trace_cgroup_destroy_root(root);
1357 
1358 	cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1359 
1360 	BUG_ON(atomic_read(&root->nr_cgrps));
1361 	BUG_ON(!list_empty(&cgrp->self.children));
1362 
1363 	/* Rebind all subsystems back to the default hierarchy */
1364 	WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1365 
1366 	/*
1367 	 * Release all the links from cset_links to this hierarchy's
1368 	 * root cgroup
1369 	 */
1370 	spin_lock_irq(&css_set_lock);
1371 
1372 	list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1373 		list_del(&link->cset_link);
1374 		list_del(&link->cgrp_link);
1375 		kfree(link);
1376 	}
1377 
1378 	spin_unlock_irq(&css_set_lock);
1379 
1380 	if (!list_empty(&root->root_list)) {
1381 		list_del(&root->root_list);
1382 		cgroup_root_count--;
1383 	}
1384 
1385 	cgroup_favor_dynmods(root, false);
1386 	cgroup_exit_root_id(root);
1387 
1388 	mutex_unlock(&cgroup_mutex);
1389 
1390 	cgroup_rstat_exit(cgrp);
1391 	kernfs_destroy_root(root->kf_root);
1392 	cgroup_free_root(root);
1393 }
1394 
1395 /*
1396  * Returned cgroup is without refcount but it's valid as long as cset pins it.
1397  */
__cset_cgroup_from_root(struct css_set * cset,struct cgroup_root * root)1398 static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1399 					    struct cgroup_root *root)
1400 {
1401 	struct cgroup *res_cgroup = NULL;
1402 
1403 	if (cset == &init_css_set) {
1404 		res_cgroup = &root->cgrp;
1405 	} else if (root == &cgrp_dfl_root) {
1406 		res_cgroup = cset->dfl_cgrp;
1407 	} else {
1408 		struct cgrp_cset_link *link;
1409 		lockdep_assert_held(&css_set_lock);
1410 
1411 		list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1412 			struct cgroup *c = link->cgrp;
1413 
1414 			if (c->root == root) {
1415 				res_cgroup = c;
1416 				break;
1417 			}
1418 		}
1419 	}
1420 
1421 	BUG_ON(!res_cgroup);
1422 	return res_cgroup;
1423 }
1424 
1425 /*
1426  * look up cgroup associated with current task's cgroup namespace on the
1427  * specified hierarchy
1428  */
1429 static struct cgroup *
current_cgns_cgroup_from_root(struct cgroup_root * root)1430 current_cgns_cgroup_from_root(struct cgroup_root *root)
1431 {
1432 	struct cgroup *res = NULL;
1433 	struct css_set *cset;
1434 
1435 	lockdep_assert_held(&css_set_lock);
1436 
1437 	rcu_read_lock();
1438 
1439 	cset = current->nsproxy->cgroup_ns->root_cset;
1440 	res = __cset_cgroup_from_root(cset, root);
1441 
1442 	rcu_read_unlock();
1443 
1444 	return res;
1445 }
1446 
1447 /*
1448  * Look up cgroup associated with current task's cgroup namespace on the default
1449  * hierarchy.
1450  *
1451  * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1452  * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1453  *   pointers.
1454  * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1455  * - As a bonus returned cgrp is pinned with the current because it cannot
1456  *   switch cgroup_ns asynchronously.
1457  */
current_cgns_cgroup_dfl(void)1458 static struct cgroup *current_cgns_cgroup_dfl(void)
1459 {
1460 	struct css_set *cset;
1461 
1462 	cset = current->nsproxy->cgroup_ns->root_cset;
1463 	return __cset_cgroup_from_root(cset, &cgrp_dfl_root);
1464 }
1465 
1466 /* look up cgroup associated with given css_set on the specified hierarchy */
cset_cgroup_from_root(struct css_set * cset,struct cgroup_root * root)1467 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1468 					    struct cgroup_root *root)
1469 {
1470 	lockdep_assert_held(&cgroup_mutex);
1471 	lockdep_assert_held(&css_set_lock);
1472 
1473 	return __cset_cgroup_from_root(cset, root);
1474 }
1475 
1476 /*
1477  * Return the cgroup for "task" from the given hierarchy. Must be
1478  * called with cgroup_mutex and css_set_lock held.
1479  */
task_cgroup_from_root(struct task_struct * task,struct cgroup_root * root)1480 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1481 				     struct cgroup_root *root)
1482 {
1483 	/*
1484 	 * No need to lock the task - since we hold css_set_lock the
1485 	 * task can't change groups.
1486 	 */
1487 	return cset_cgroup_from_root(task_css_set(task), root);
1488 }
1489 
1490 /*
1491  * A task must hold cgroup_mutex to modify cgroups.
1492  *
1493  * Any task can increment and decrement the count field without lock.
1494  * So in general, code holding cgroup_mutex can't rely on the count
1495  * field not changing.  However, if the count goes to zero, then only
1496  * cgroup_attach_task() can increment it again.  Because a count of zero
1497  * means that no tasks are currently attached, therefore there is no
1498  * way a task attached to that cgroup can fork (the other way to
1499  * increment the count).  So code holding cgroup_mutex can safely
1500  * assume that if the count is zero, it will stay zero. Similarly, if
1501  * a task holds cgroup_mutex on a cgroup with zero count, it
1502  * knows that the cgroup won't be removed, as cgroup_rmdir()
1503  * needs that mutex.
1504  *
1505  * A cgroup can only be deleted if both its 'count' of using tasks
1506  * is zero, and its list of 'children' cgroups is empty.  Since all
1507  * tasks in the system use _some_ cgroup, and since there is always at
1508  * least one task in the system (init, pid == 1), therefore, root cgroup
1509  * always has either children cgroups and/or using tasks.  So we don't
1510  * need a special hack to ensure that root cgroup cannot be deleted.
1511  *
1512  * P.S.  One more locking exception.  RCU is used to guard the
1513  * update of a tasks cgroup pointer by cgroup_attach_task()
1514  */
1515 
1516 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1517 
cgroup_file_name(struct cgroup * cgrp,const struct cftype * cft,char * buf)1518 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1519 			      char *buf)
1520 {
1521 	struct cgroup_subsys *ss = cft->ss;
1522 
1523 	if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1524 	    !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1525 		const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1526 
1527 		snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1528 			 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1529 			 cft->name);
1530 	} else {
1531 		strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1532 	}
1533 	return buf;
1534 }
1535 
1536 /**
1537  * cgroup_file_mode - deduce file mode of a control file
1538  * @cft: the control file in question
1539  *
1540  * S_IRUGO for read, S_IWUSR for write.
1541  */
cgroup_file_mode(const struct cftype * cft)1542 static umode_t cgroup_file_mode(const struct cftype *cft)
1543 {
1544 	umode_t mode = 0;
1545 
1546 	if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1547 		mode |= S_IRUGO;
1548 
1549 	if (cft->write_u64 || cft->write_s64 || cft->write) {
1550 		if (cft->flags & CFTYPE_WORLD_WRITABLE)
1551 			mode |= S_IWUGO;
1552 		else
1553 			mode |= S_IWUSR;
1554 	}
1555 
1556 	return mode;
1557 }
1558 
1559 /**
1560  * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1561  * @subtree_control: the new subtree_control mask to consider
1562  * @this_ss_mask: available subsystems
1563  *
1564  * On the default hierarchy, a subsystem may request other subsystems to be
1565  * enabled together through its ->depends_on mask.  In such cases, more
1566  * subsystems than specified in "cgroup.subtree_control" may be enabled.
1567  *
1568  * This function calculates which subsystems need to be enabled if
1569  * @subtree_control is to be applied while restricted to @this_ss_mask.
1570  */
cgroup_calc_subtree_ss_mask(u16 subtree_control,u16 this_ss_mask)1571 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1572 {
1573 	u16 cur_ss_mask = subtree_control;
1574 	struct cgroup_subsys *ss;
1575 	int ssid;
1576 
1577 	lockdep_assert_held(&cgroup_mutex);
1578 
1579 	cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1580 
1581 	while (true) {
1582 		u16 new_ss_mask = cur_ss_mask;
1583 
1584 		do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1585 			new_ss_mask |= ss->depends_on;
1586 		} while_each_subsys_mask();
1587 
1588 		/*
1589 		 * Mask out subsystems which aren't available.  This can
1590 		 * happen only if some depended-upon subsystems were bound
1591 		 * to non-default hierarchies.
1592 		 */
1593 		new_ss_mask &= this_ss_mask;
1594 
1595 		if (new_ss_mask == cur_ss_mask)
1596 			break;
1597 		cur_ss_mask = new_ss_mask;
1598 	}
1599 
1600 	return cur_ss_mask;
1601 }
1602 
1603 /**
1604  * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1605  * @kn: the kernfs_node being serviced
1606  *
1607  * This helper undoes cgroup_kn_lock_live() and should be invoked before
1608  * the method finishes if locking succeeded.  Note that once this function
1609  * returns the cgroup returned by cgroup_kn_lock_live() may become
1610  * inaccessible any time.  If the caller intends to continue to access the
1611  * cgroup, it should pin it before invoking this function.
1612  */
cgroup_kn_unlock(struct kernfs_node * kn)1613 void cgroup_kn_unlock(struct kernfs_node *kn)
1614 {
1615 	struct cgroup *cgrp;
1616 
1617 	if (kernfs_type(kn) == KERNFS_DIR)
1618 		cgrp = kn->priv;
1619 	else
1620 		cgrp = kn->parent->priv;
1621 
1622 	mutex_unlock(&cgroup_mutex);
1623 
1624 	kernfs_unbreak_active_protection(kn);
1625 	cgroup_put(cgrp);
1626 }
1627 
1628 /**
1629  * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1630  * @kn: the kernfs_node being serviced
1631  * @drain_offline: perform offline draining on the cgroup
1632  *
1633  * This helper is to be used by a cgroup kernfs method currently servicing
1634  * @kn.  It breaks the active protection, performs cgroup locking and
1635  * verifies that the associated cgroup is alive.  Returns the cgroup if
1636  * alive; otherwise, %NULL.  A successful return should be undone by a
1637  * matching cgroup_kn_unlock() invocation.  If @drain_offline is %true, the
1638  * cgroup is drained of offlining csses before return.
1639  *
1640  * Any cgroup kernfs method implementation which requires locking the
1641  * associated cgroup should use this helper.  It avoids nesting cgroup
1642  * locking under kernfs active protection and allows all kernfs operations
1643  * including self-removal.
1644  */
cgroup_kn_lock_live(struct kernfs_node * kn,bool drain_offline)1645 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1646 {
1647 	struct cgroup *cgrp;
1648 
1649 	if (kernfs_type(kn) == KERNFS_DIR)
1650 		cgrp = kn->priv;
1651 	else
1652 		cgrp = kn->parent->priv;
1653 
1654 	/*
1655 	 * We're gonna grab cgroup_mutex which nests outside kernfs
1656 	 * active_ref.  cgroup liveliness check alone provides enough
1657 	 * protection against removal.  Ensure @cgrp stays accessible and
1658 	 * break the active_ref protection.
1659 	 */
1660 	if (!cgroup_tryget(cgrp))
1661 		return NULL;
1662 	kernfs_break_active_protection(kn);
1663 
1664 	if (drain_offline)
1665 		cgroup_lock_and_drain_offline(cgrp);
1666 	else
1667 		mutex_lock(&cgroup_mutex);
1668 
1669 	if (!cgroup_is_dead(cgrp))
1670 		return cgrp;
1671 
1672 	cgroup_kn_unlock(kn);
1673 	return NULL;
1674 }
1675 
cgroup_rm_file(struct cgroup * cgrp,const struct cftype * cft)1676 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1677 {
1678 	char name[CGROUP_FILE_NAME_MAX];
1679 
1680 	lockdep_assert_held(&cgroup_mutex);
1681 
1682 	if (cft->file_offset) {
1683 		struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1684 		struct cgroup_file *cfile = (void *)css + cft->file_offset;
1685 
1686 		spin_lock_irq(&cgroup_file_kn_lock);
1687 		cfile->kn = NULL;
1688 		spin_unlock_irq(&cgroup_file_kn_lock);
1689 
1690 		del_timer_sync(&cfile->notify_timer);
1691 	}
1692 
1693 	kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1694 }
1695 
1696 /**
1697  * css_clear_dir - remove subsys files in a cgroup directory
1698  * @css: target css
1699  */
css_clear_dir(struct cgroup_subsys_state * css)1700 static void css_clear_dir(struct cgroup_subsys_state *css)
1701 {
1702 	struct cgroup *cgrp = css->cgroup;
1703 	struct cftype *cfts;
1704 
1705 	if (!(css->flags & CSS_VISIBLE))
1706 		return;
1707 
1708 	css->flags &= ~CSS_VISIBLE;
1709 
1710 	if (!css->ss) {
1711 		if (cgroup_on_dfl(cgrp)) {
1712 			cgroup_addrm_files(css, cgrp,
1713 					   cgroup_base_files, false);
1714 			if (cgroup_psi_enabled())
1715 				cgroup_addrm_files(css, cgrp,
1716 						   cgroup_psi_files, false);
1717 		} else {
1718 			cgroup_addrm_files(css, cgrp,
1719 					   cgroup1_base_files, false);
1720 		}
1721 	} else {
1722 		list_for_each_entry(cfts, &css->ss->cfts, node)
1723 			cgroup_addrm_files(css, cgrp, cfts, false);
1724 	}
1725 }
1726 
1727 /**
1728  * css_populate_dir - create subsys files in a cgroup directory
1729  * @css: target css
1730  *
1731  * On failure, no file is added.
1732  */
css_populate_dir(struct cgroup_subsys_state * css)1733 static int css_populate_dir(struct cgroup_subsys_state *css)
1734 {
1735 	struct cgroup *cgrp = css->cgroup;
1736 	struct cftype *cfts, *failed_cfts;
1737 	int ret;
1738 
1739 	if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1740 		return 0;
1741 
1742 	if (!css->ss) {
1743 		if (cgroup_on_dfl(cgrp)) {
1744 			ret = cgroup_addrm_files(&cgrp->self, cgrp,
1745 						 cgroup_base_files, true);
1746 			if (ret < 0)
1747 				return ret;
1748 
1749 			if (cgroup_psi_enabled()) {
1750 				ret = cgroup_addrm_files(&cgrp->self, cgrp,
1751 							 cgroup_psi_files, true);
1752 				if (ret < 0)
1753 					return ret;
1754 			}
1755 		} else {
1756 			cgroup_addrm_files(css, cgrp,
1757 					   cgroup1_base_files, true);
1758 		}
1759 	} else {
1760 		list_for_each_entry(cfts, &css->ss->cfts, node) {
1761 			ret = cgroup_addrm_files(css, cgrp, cfts, true);
1762 			if (ret < 0) {
1763 				failed_cfts = cfts;
1764 				goto err;
1765 			}
1766 		}
1767 	}
1768 
1769 	css->flags |= CSS_VISIBLE;
1770 
1771 	return 0;
1772 err:
1773 	list_for_each_entry(cfts, &css->ss->cfts, node) {
1774 		if (cfts == failed_cfts)
1775 			break;
1776 		cgroup_addrm_files(css, cgrp, cfts, false);
1777 	}
1778 	return ret;
1779 }
1780 
rebind_subsystems(struct cgroup_root * dst_root,u16 ss_mask)1781 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1782 {
1783 	struct cgroup *dcgrp = &dst_root->cgrp;
1784 	struct cgroup_subsys *ss;
1785 	int ssid, i, ret;
1786 	u16 dfl_disable_ss_mask = 0;
1787 
1788 	lockdep_assert_held(&cgroup_mutex);
1789 
1790 	do_each_subsys_mask(ss, ssid, ss_mask) {
1791 		/*
1792 		 * If @ss has non-root csses attached to it, can't move.
1793 		 * If @ss is an implicit controller, it is exempt from this
1794 		 * rule and can be stolen.
1795 		 */
1796 		if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1797 		    !ss->implicit_on_dfl)
1798 			return -EBUSY;
1799 
1800 		/* can't move between two non-dummy roots either */
1801 		if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1802 			return -EBUSY;
1803 
1804 		/*
1805 		 * Collect ssid's that need to be disabled from default
1806 		 * hierarchy.
1807 		 */
1808 		if (ss->root == &cgrp_dfl_root)
1809 			dfl_disable_ss_mask |= 1 << ssid;
1810 
1811 	} while_each_subsys_mask();
1812 
1813 	if (dfl_disable_ss_mask) {
1814 		struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1815 
1816 		/*
1817 		 * Controllers from default hierarchy that need to be rebound
1818 		 * are all disabled together in one go.
1819 		 */
1820 		cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1821 		WARN_ON(cgroup_apply_control(scgrp));
1822 		cgroup_finalize_control(scgrp, 0);
1823 	}
1824 
1825 	do_each_subsys_mask(ss, ssid, ss_mask) {
1826 		struct cgroup_root *src_root = ss->root;
1827 		struct cgroup *scgrp = &src_root->cgrp;
1828 		struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1829 		struct css_set *cset;
1830 
1831 		WARN_ON(!css || cgroup_css(dcgrp, ss));
1832 
1833 		if (src_root != &cgrp_dfl_root) {
1834 			/* disable from the source */
1835 			src_root->subsys_mask &= ~(1 << ssid);
1836 			WARN_ON(cgroup_apply_control(scgrp));
1837 			cgroup_finalize_control(scgrp, 0);
1838 		}
1839 
1840 		/* rebind */
1841 		RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1842 		rcu_assign_pointer(dcgrp->subsys[ssid], css);
1843 		ss->root = dst_root;
1844 		css->cgroup = dcgrp;
1845 
1846 		spin_lock_irq(&css_set_lock);
1847 		hash_for_each(css_set_table, i, cset, hlist)
1848 			list_move_tail(&cset->e_cset_node[ss->id],
1849 				       &dcgrp->e_csets[ss->id]);
1850 		spin_unlock_irq(&css_set_lock);
1851 
1852 		if (ss->css_rstat_flush) {
1853 			list_del_rcu(&css->rstat_css_node);
1854 			synchronize_rcu();
1855 			list_add_rcu(&css->rstat_css_node,
1856 				     &dcgrp->rstat_css_list);
1857 		}
1858 
1859 		/* default hierarchy doesn't enable controllers by default */
1860 		dst_root->subsys_mask |= 1 << ssid;
1861 		if (dst_root == &cgrp_dfl_root) {
1862 			static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1863 		} else {
1864 			dcgrp->subtree_control |= 1 << ssid;
1865 			static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1866 		}
1867 
1868 		ret = cgroup_apply_control(dcgrp);
1869 		if (ret)
1870 			pr_warn("partial failure to rebind %s controller (err=%d)\n",
1871 				ss->name, ret);
1872 
1873 		if (ss->bind)
1874 			ss->bind(css);
1875 	} while_each_subsys_mask();
1876 
1877 	kernfs_activate(dcgrp->kn);
1878 	return 0;
1879 }
1880 
cgroup_show_path(struct seq_file * sf,struct kernfs_node * kf_node,struct kernfs_root * kf_root)1881 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1882 		     struct kernfs_root *kf_root)
1883 {
1884 	int len = 0;
1885 	char *buf = NULL;
1886 	struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1887 	struct cgroup *ns_cgroup;
1888 
1889 	buf = kmalloc(PATH_MAX, GFP_KERNEL);
1890 	if (!buf)
1891 		return -ENOMEM;
1892 
1893 	spin_lock_irq(&css_set_lock);
1894 	ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1895 	len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1896 	spin_unlock_irq(&css_set_lock);
1897 
1898 	if (len >= PATH_MAX)
1899 		len = -ERANGE;
1900 	else if (len > 0) {
1901 		seq_escape(sf, buf, " \t\n\\");
1902 		len = 0;
1903 	}
1904 	kfree(buf);
1905 	return len;
1906 }
1907 
1908 enum cgroup2_param {
1909 	Opt_nsdelegate,
1910 	Opt_favordynmods,
1911 	Opt_memory_localevents,
1912 	Opt_memory_recursiveprot,
1913 	nr__cgroup2_params
1914 };
1915 
1916 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1917 	fsparam_flag("nsdelegate",		Opt_nsdelegate),
1918 	fsparam_flag("favordynmods",		Opt_favordynmods),
1919 	fsparam_flag("memory_localevents",	Opt_memory_localevents),
1920 	fsparam_flag("memory_recursiveprot",	Opt_memory_recursiveprot),
1921 	{}
1922 };
1923 
cgroup2_parse_param(struct fs_context * fc,struct fs_parameter * param)1924 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1925 {
1926 	struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1927 	struct fs_parse_result result;
1928 	int opt;
1929 
1930 	opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1931 	if (opt < 0)
1932 		return opt;
1933 
1934 	switch (opt) {
1935 	case Opt_nsdelegate:
1936 		ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1937 		return 0;
1938 	case Opt_favordynmods:
1939 		ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1940 		return 0;
1941 	case Opt_memory_localevents:
1942 		ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1943 		return 0;
1944 	case Opt_memory_recursiveprot:
1945 		ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1946 		return 0;
1947 	}
1948 	return -EINVAL;
1949 }
1950 
apply_cgroup_root_flags(unsigned int root_flags)1951 static void apply_cgroup_root_flags(unsigned int root_flags)
1952 {
1953 	if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1954 		if (root_flags & CGRP_ROOT_NS_DELEGATE)
1955 			cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1956 		else
1957 			cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1958 
1959 		cgroup_favor_dynmods(&cgrp_dfl_root,
1960 				     root_flags & CGRP_ROOT_FAVOR_DYNMODS);
1961 
1962 		if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1963 			cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1964 		else
1965 			cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1966 
1967 		if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1968 			cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1969 		else
1970 			cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1971 	}
1972 }
1973 
cgroup_show_options(struct seq_file * seq,struct kernfs_root * kf_root)1974 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1975 {
1976 	if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1977 		seq_puts(seq, ",nsdelegate");
1978 	if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
1979 		seq_puts(seq, ",favordynmods");
1980 	if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1981 		seq_puts(seq, ",memory_localevents");
1982 	if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1983 		seq_puts(seq, ",memory_recursiveprot");
1984 	return 0;
1985 }
1986 
cgroup_reconfigure(struct fs_context * fc)1987 static int cgroup_reconfigure(struct fs_context *fc)
1988 {
1989 	struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1990 
1991 	apply_cgroup_root_flags(ctx->flags);
1992 	return 0;
1993 }
1994 
init_cgroup_housekeeping(struct cgroup * cgrp)1995 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1996 {
1997 	struct cgroup_subsys *ss;
1998 	int ssid;
1999 
2000 	INIT_LIST_HEAD(&cgrp->self.sibling);
2001 	INIT_LIST_HEAD(&cgrp->self.children);
2002 	INIT_LIST_HEAD(&cgrp->cset_links);
2003 	INIT_LIST_HEAD(&cgrp->pidlists);
2004 	mutex_init(&cgrp->pidlist_mutex);
2005 	cgrp->self.cgroup = cgrp;
2006 	cgrp->self.flags |= CSS_ONLINE;
2007 	cgrp->dom_cgrp = cgrp;
2008 	cgrp->max_descendants = INT_MAX;
2009 	cgrp->max_depth = INT_MAX;
2010 	INIT_LIST_HEAD(&cgrp->rstat_css_list);
2011 	prev_cputime_init(&cgrp->prev_cputime);
2012 
2013 	for_each_subsys(ss, ssid)
2014 		INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2015 
2016 	init_waitqueue_head(&cgrp->offline_waitq);
2017 	INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2018 }
2019 
init_cgroup_root(struct cgroup_fs_context * ctx)2020 void init_cgroup_root(struct cgroup_fs_context *ctx)
2021 {
2022 	struct cgroup_root *root = ctx->root;
2023 	struct cgroup *cgrp = &root->cgrp;
2024 
2025 	INIT_LIST_HEAD(&root->root_list);
2026 	atomic_set(&root->nr_cgrps, 1);
2027 	cgrp->root = root;
2028 	init_cgroup_housekeeping(cgrp);
2029 
2030 	/* DYNMODS must be modified through cgroup_favor_dynmods() */
2031 	root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2032 	if (ctx->release_agent)
2033 		strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2034 	if (ctx->name)
2035 		strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2036 	if (ctx->cpuset_clone_children)
2037 		set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2038 }
2039 
cgroup_setup_root(struct cgroup_root * root,u16 ss_mask)2040 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2041 {
2042 	LIST_HEAD(tmp_links);
2043 	struct cgroup *root_cgrp = &root->cgrp;
2044 	struct kernfs_syscall_ops *kf_sops;
2045 	struct css_set *cset;
2046 	int i, ret;
2047 
2048 	lockdep_assert_held(&cgroup_mutex);
2049 
2050 	ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2051 			      0, GFP_KERNEL);
2052 	if (ret)
2053 		goto out;
2054 
2055 	/*
2056 	 * We're accessing css_set_count without locking css_set_lock here,
2057 	 * but that's OK - it can only be increased by someone holding
2058 	 * cgroup_lock, and that's us.  Later rebinding may disable
2059 	 * controllers on the default hierarchy and thus create new csets,
2060 	 * which can't be more than the existing ones.  Allocate 2x.
2061 	 */
2062 	ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2063 	if (ret)
2064 		goto cancel_ref;
2065 
2066 	ret = cgroup_init_root_id(root);
2067 	if (ret)
2068 		goto cancel_ref;
2069 
2070 	kf_sops = root == &cgrp_dfl_root ?
2071 		&cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2072 
2073 	root->kf_root = kernfs_create_root(kf_sops,
2074 					   KERNFS_ROOT_CREATE_DEACTIVATED |
2075 					   KERNFS_ROOT_SUPPORT_EXPORTOP |
2076 					   KERNFS_ROOT_SUPPORT_USER_XATTR,
2077 					   root_cgrp);
2078 	if (IS_ERR(root->kf_root)) {
2079 		ret = PTR_ERR(root->kf_root);
2080 		goto exit_root_id;
2081 	}
2082 	root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2083 	WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2084 	root_cgrp->ancestors[0] = root_cgrp;
2085 
2086 	ret = css_populate_dir(&root_cgrp->self);
2087 	if (ret)
2088 		goto destroy_root;
2089 
2090 	ret = cgroup_rstat_init(root_cgrp);
2091 	if (ret)
2092 		goto destroy_root;
2093 
2094 	ret = rebind_subsystems(root, ss_mask);
2095 	if (ret)
2096 		goto exit_stats;
2097 
2098 	ret = cgroup_bpf_inherit(root_cgrp);
2099 	WARN_ON_ONCE(ret);
2100 
2101 	trace_cgroup_setup_root(root);
2102 
2103 	/*
2104 	 * There must be no failure case after here, since rebinding takes
2105 	 * care of subsystems' refcounts, which are explicitly dropped in
2106 	 * the failure exit path.
2107 	 */
2108 	list_add(&root->root_list, &cgroup_roots);
2109 	cgroup_root_count++;
2110 
2111 	/*
2112 	 * Link the root cgroup in this hierarchy into all the css_set
2113 	 * objects.
2114 	 */
2115 	spin_lock_irq(&css_set_lock);
2116 	hash_for_each(css_set_table, i, cset, hlist) {
2117 		link_css_set(&tmp_links, cset, root_cgrp);
2118 		if (css_set_populated(cset))
2119 			cgroup_update_populated(root_cgrp, true);
2120 	}
2121 	spin_unlock_irq(&css_set_lock);
2122 
2123 	BUG_ON(!list_empty(&root_cgrp->self.children));
2124 	BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2125 
2126 	ret = 0;
2127 	goto out;
2128 
2129 exit_stats:
2130 	cgroup_rstat_exit(root_cgrp);
2131 destroy_root:
2132 	kernfs_destroy_root(root->kf_root);
2133 	root->kf_root = NULL;
2134 exit_root_id:
2135 	cgroup_exit_root_id(root);
2136 cancel_ref:
2137 	percpu_ref_exit(&root_cgrp->self.refcnt);
2138 out:
2139 	free_cgrp_cset_links(&tmp_links);
2140 	return ret;
2141 }
2142 
cgroup_do_get_tree(struct fs_context * fc)2143 int cgroup_do_get_tree(struct fs_context *fc)
2144 {
2145 	struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2146 	int ret;
2147 
2148 	ctx->kfc.root = ctx->root->kf_root;
2149 	if (fc->fs_type == &cgroup2_fs_type)
2150 		ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2151 	else
2152 		ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2153 	ret = kernfs_get_tree(fc);
2154 
2155 	/*
2156 	 * In non-init cgroup namespace, instead of root cgroup's dentry,
2157 	 * we return the dentry corresponding to the cgroupns->root_cgrp.
2158 	 */
2159 	if (!ret && ctx->ns != &init_cgroup_ns) {
2160 		struct dentry *nsdentry;
2161 		struct super_block *sb = fc->root->d_sb;
2162 		struct cgroup *cgrp;
2163 
2164 		mutex_lock(&cgroup_mutex);
2165 		spin_lock_irq(&css_set_lock);
2166 
2167 		cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2168 
2169 		spin_unlock_irq(&css_set_lock);
2170 		mutex_unlock(&cgroup_mutex);
2171 
2172 		nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2173 		dput(fc->root);
2174 		if (IS_ERR(nsdentry)) {
2175 			deactivate_locked_super(sb);
2176 			ret = PTR_ERR(nsdentry);
2177 			nsdentry = NULL;
2178 		}
2179 		fc->root = nsdentry;
2180 	}
2181 
2182 	if (!ctx->kfc.new_sb_created)
2183 		cgroup_put(&ctx->root->cgrp);
2184 
2185 	return ret;
2186 }
2187 
2188 /*
2189  * Destroy a cgroup filesystem context.
2190  */
cgroup_fs_context_free(struct fs_context * fc)2191 static void cgroup_fs_context_free(struct fs_context *fc)
2192 {
2193 	struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2194 
2195 	kfree(ctx->name);
2196 	kfree(ctx->release_agent);
2197 	put_cgroup_ns(ctx->ns);
2198 	kernfs_free_fs_context(fc);
2199 	kfree(ctx);
2200 }
2201 
cgroup_get_tree(struct fs_context * fc)2202 static int cgroup_get_tree(struct fs_context *fc)
2203 {
2204 	struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2205 	int ret;
2206 
2207 	WRITE_ONCE(cgrp_dfl_visible, true);
2208 	cgroup_get_live(&cgrp_dfl_root.cgrp);
2209 	ctx->root = &cgrp_dfl_root;
2210 
2211 	ret = cgroup_do_get_tree(fc);
2212 	if (!ret)
2213 		apply_cgroup_root_flags(ctx->flags);
2214 	return ret;
2215 }
2216 
2217 static const struct fs_context_operations cgroup_fs_context_ops = {
2218 	.free		= cgroup_fs_context_free,
2219 	.parse_param	= cgroup2_parse_param,
2220 	.get_tree	= cgroup_get_tree,
2221 	.reconfigure	= cgroup_reconfigure,
2222 };
2223 
2224 static const struct fs_context_operations cgroup1_fs_context_ops = {
2225 	.free		= cgroup_fs_context_free,
2226 	.parse_param	= cgroup1_parse_param,
2227 	.get_tree	= cgroup1_get_tree,
2228 	.reconfigure	= cgroup1_reconfigure,
2229 };
2230 
2231 /*
2232  * Initialise the cgroup filesystem creation/reconfiguration context.  Notably,
2233  * we select the namespace we're going to use.
2234  */
cgroup_init_fs_context(struct fs_context * fc)2235 static int cgroup_init_fs_context(struct fs_context *fc)
2236 {
2237 	struct cgroup_fs_context *ctx;
2238 
2239 	ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2240 	if (!ctx)
2241 		return -ENOMEM;
2242 
2243 	ctx->ns = current->nsproxy->cgroup_ns;
2244 	get_cgroup_ns(ctx->ns);
2245 	fc->fs_private = &ctx->kfc;
2246 	if (fc->fs_type == &cgroup2_fs_type)
2247 		fc->ops = &cgroup_fs_context_ops;
2248 	else
2249 		fc->ops = &cgroup1_fs_context_ops;
2250 	put_user_ns(fc->user_ns);
2251 	fc->user_ns = get_user_ns(ctx->ns->user_ns);
2252 	fc->global = true;
2253 
2254 #ifdef CONFIG_CGROUP_FAVOR_DYNMODS
2255 	ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2256 #endif
2257 	return 0;
2258 }
2259 
cgroup_kill_sb(struct super_block * sb)2260 static void cgroup_kill_sb(struct super_block *sb)
2261 {
2262 	struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2263 	struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2264 
2265 	/*
2266 	 * If @root doesn't have any children, start killing it.
2267 	 * This prevents new mounts by disabling percpu_ref_tryget_live().
2268 	 *
2269 	 * And don't kill the default root.
2270 	 */
2271 	if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2272 	    !percpu_ref_is_dying(&root->cgrp.self.refcnt)) {
2273 		cgroup_bpf_offline(&root->cgrp);
2274 		percpu_ref_kill(&root->cgrp.self.refcnt);
2275 	}
2276 	cgroup_put(&root->cgrp);
2277 	kernfs_kill_sb(sb);
2278 }
2279 
2280 struct file_system_type cgroup_fs_type = {
2281 	.name			= "cgroup",
2282 	.init_fs_context	= cgroup_init_fs_context,
2283 	.parameters		= cgroup1_fs_parameters,
2284 	.kill_sb		= cgroup_kill_sb,
2285 	.fs_flags		= FS_USERNS_MOUNT,
2286 };
2287 
2288 static struct file_system_type cgroup2_fs_type = {
2289 	.name			= "cgroup2",
2290 	.init_fs_context	= cgroup_init_fs_context,
2291 	.parameters		= cgroup2_fs_parameters,
2292 	.kill_sb		= cgroup_kill_sb,
2293 	.fs_flags		= FS_USERNS_MOUNT,
2294 };
2295 
2296 #ifdef CONFIG_CPUSETS
2297 static const struct fs_context_operations cpuset_fs_context_ops = {
2298 	.get_tree	= cgroup1_get_tree,
2299 	.free		= cgroup_fs_context_free,
2300 };
2301 
2302 /*
2303  * This is ugly, but preserves the userspace API for existing cpuset
2304  * users. If someone tries to mount the "cpuset" filesystem, we
2305  * silently switch it to mount "cgroup" instead
2306  */
cpuset_init_fs_context(struct fs_context * fc)2307 static int cpuset_init_fs_context(struct fs_context *fc)
2308 {
2309 	char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2310 	struct cgroup_fs_context *ctx;
2311 	int err;
2312 
2313 	err = cgroup_init_fs_context(fc);
2314 	if (err) {
2315 		kfree(agent);
2316 		return err;
2317 	}
2318 
2319 	fc->ops = &cpuset_fs_context_ops;
2320 
2321 	ctx = cgroup_fc2context(fc);
2322 	ctx->subsys_mask = 1 << cpuset_cgrp_id;
2323 	ctx->flags |= CGRP_ROOT_NOPREFIX;
2324 	ctx->release_agent = agent;
2325 
2326 	get_filesystem(&cgroup_fs_type);
2327 	put_filesystem(fc->fs_type);
2328 	fc->fs_type = &cgroup_fs_type;
2329 
2330 	return 0;
2331 }
2332 
2333 static struct file_system_type cpuset_fs_type = {
2334 	.name			= "cpuset",
2335 	.init_fs_context	= cpuset_init_fs_context,
2336 	.fs_flags		= FS_USERNS_MOUNT,
2337 };
2338 #endif
2339 
cgroup_path_ns_locked(struct cgroup * cgrp,char * buf,size_t buflen,struct cgroup_namespace * ns)2340 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2341 			  struct cgroup_namespace *ns)
2342 {
2343 	struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2344 
2345 	return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2346 }
2347 
cgroup_path_ns(struct cgroup * cgrp,char * buf,size_t buflen,struct cgroup_namespace * ns)2348 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2349 		   struct cgroup_namespace *ns)
2350 {
2351 	int ret;
2352 
2353 	mutex_lock(&cgroup_mutex);
2354 	spin_lock_irq(&css_set_lock);
2355 
2356 	ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2357 
2358 	spin_unlock_irq(&css_set_lock);
2359 	mutex_unlock(&cgroup_mutex);
2360 
2361 	return ret;
2362 }
2363 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2364 
2365 /**
2366  * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2367  * @task: target task
2368  * @buf: the buffer to write the path into
2369  * @buflen: the length of the buffer
2370  *
2371  * Determine @task's cgroup on the first (the one with the lowest non-zero
2372  * hierarchy_id) cgroup hierarchy and copy its path into @buf.  This
2373  * function grabs cgroup_mutex and shouldn't be used inside locks used by
2374  * cgroup controller callbacks.
2375  *
2376  * Return value is the same as kernfs_path().
2377  */
task_cgroup_path(struct task_struct * task,char * buf,size_t buflen)2378 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2379 {
2380 	struct cgroup_root *root;
2381 	struct cgroup *cgrp;
2382 	int hierarchy_id = 1;
2383 	int ret;
2384 
2385 	mutex_lock(&cgroup_mutex);
2386 	spin_lock_irq(&css_set_lock);
2387 
2388 	root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2389 
2390 	if (root) {
2391 		cgrp = task_cgroup_from_root(task, root);
2392 		ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2393 	} else {
2394 		/* if no hierarchy exists, everyone is in "/" */
2395 		ret = strscpy(buf, "/", buflen);
2396 	}
2397 
2398 	spin_unlock_irq(&css_set_lock);
2399 	mutex_unlock(&cgroup_mutex);
2400 	return ret;
2401 }
2402 EXPORT_SYMBOL_GPL(task_cgroup_path);
2403 
2404 /**
2405  * cgroup_attach_lock - Lock for ->attach()
2406  * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2407  *
2408  * cgroup migration sometimes needs to stabilize threadgroups against forks and
2409  * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2410  * implementations (e.g. cpuset), also need to disable CPU hotplug.
2411  * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2412  * lead to deadlocks.
2413  *
2414  * Bringing up a CPU may involve creating and destroying tasks which requires
2415  * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2416  * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2417  * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2418  * waiting for an on-going CPU hotplug operation which in turn is waiting for
2419  * the threadgroup_rwsem to be released to create new tasks. For more details:
2420  *
2421  *   http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2422  *
2423  * Resolve the situation by always acquiring cpus_read_lock() before optionally
2424  * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2425  * CPU hotplug is disabled on entry.
2426  */
cgroup_attach_lock(bool lock_threadgroup)2427 void cgroup_attach_lock(bool lock_threadgroup)
2428 {
2429 	cpus_read_lock();
2430 	if (lock_threadgroup)
2431 		percpu_down_write(&cgroup_threadgroup_rwsem);
2432 }
2433 
2434 /**
2435  * cgroup_attach_unlock - Undo cgroup_attach_lock()
2436  * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2437  */
cgroup_attach_unlock(bool lock_threadgroup)2438 void cgroup_attach_unlock(bool lock_threadgroup)
2439 {
2440 	if (lock_threadgroup)
2441 		percpu_up_write(&cgroup_threadgroup_rwsem);
2442 	cpus_read_unlock();
2443 }
2444 
2445 /**
2446  * cgroup_migrate_add_task - add a migration target task to a migration context
2447  * @task: target task
2448  * @mgctx: target migration context
2449  *
2450  * Add @task, which is a migration target, to @mgctx->tset.  This function
2451  * becomes noop if @task doesn't need to be migrated.  @task's css_set
2452  * should have been added as a migration source and @task->cg_list will be
2453  * moved from the css_set's tasks list to mg_tasks one.
2454  */
cgroup_migrate_add_task(struct task_struct * task,struct cgroup_mgctx * mgctx)2455 static void cgroup_migrate_add_task(struct task_struct *task,
2456 				    struct cgroup_mgctx *mgctx)
2457 {
2458 	struct css_set *cset;
2459 
2460 	lockdep_assert_held(&css_set_lock);
2461 
2462 	/* @task either already exited or can't exit until the end */
2463 	if (task->flags & PF_EXITING)
2464 		return;
2465 
2466 	/* cgroup_threadgroup_rwsem protects racing against forks */
2467 	WARN_ON_ONCE(list_empty(&task->cg_list));
2468 
2469 	cset = task_css_set(task);
2470 	if (!cset->mg_src_cgrp)
2471 		return;
2472 
2473 	mgctx->tset.nr_tasks++;
2474 
2475 	list_move_tail(&task->cg_list, &cset->mg_tasks);
2476 	if (list_empty(&cset->mg_node))
2477 		list_add_tail(&cset->mg_node,
2478 			      &mgctx->tset.src_csets);
2479 	if (list_empty(&cset->mg_dst_cset->mg_node))
2480 		list_add_tail(&cset->mg_dst_cset->mg_node,
2481 			      &mgctx->tset.dst_csets);
2482 }
2483 
2484 /**
2485  * cgroup_taskset_first - reset taskset and return the first task
2486  * @tset: taskset of interest
2487  * @dst_cssp: output variable for the destination css
2488  *
2489  * @tset iteration is initialized and the first task is returned.
2490  */
cgroup_taskset_first(struct cgroup_taskset * tset,struct cgroup_subsys_state ** dst_cssp)2491 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2492 					 struct cgroup_subsys_state **dst_cssp)
2493 {
2494 	tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2495 	tset->cur_task = NULL;
2496 
2497 	return cgroup_taskset_next(tset, dst_cssp);
2498 }
2499 
2500 /**
2501  * cgroup_taskset_next - iterate to the next task in taskset
2502  * @tset: taskset of interest
2503  * @dst_cssp: output variable for the destination css
2504  *
2505  * Return the next task in @tset.  Iteration must have been initialized
2506  * with cgroup_taskset_first().
2507  */
cgroup_taskset_next(struct cgroup_taskset * tset,struct cgroup_subsys_state ** dst_cssp)2508 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2509 					struct cgroup_subsys_state **dst_cssp)
2510 {
2511 	struct css_set *cset = tset->cur_cset;
2512 	struct task_struct *task = tset->cur_task;
2513 
2514 	while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2515 		if (!task)
2516 			task = list_first_entry(&cset->mg_tasks,
2517 						struct task_struct, cg_list);
2518 		else
2519 			task = list_next_entry(task, cg_list);
2520 
2521 		if (&task->cg_list != &cset->mg_tasks) {
2522 			tset->cur_cset = cset;
2523 			tset->cur_task = task;
2524 
2525 			/*
2526 			 * This function may be called both before and
2527 			 * after cgroup_taskset_migrate().  The two cases
2528 			 * can be distinguished by looking at whether @cset
2529 			 * has its ->mg_dst_cset set.
2530 			 */
2531 			if (cset->mg_dst_cset)
2532 				*dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2533 			else
2534 				*dst_cssp = cset->subsys[tset->ssid];
2535 
2536 			return task;
2537 		}
2538 
2539 		cset = list_next_entry(cset, mg_node);
2540 		task = NULL;
2541 	}
2542 
2543 	return NULL;
2544 }
2545 
2546 /**
2547  * cgroup_migrate_execute - migrate a taskset
2548  * @mgctx: migration context
2549  *
2550  * Migrate tasks in @mgctx as setup by migration preparation functions.
2551  * This function fails iff one of the ->can_attach callbacks fails and
2552  * guarantees that either all or none of the tasks in @mgctx are migrated.
2553  * @mgctx is consumed regardless of success.
2554  */
cgroup_migrate_execute(struct cgroup_mgctx * mgctx)2555 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2556 {
2557 	struct cgroup_taskset *tset = &mgctx->tset;
2558 	struct cgroup_subsys *ss;
2559 	struct task_struct *task, *tmp_task;
2560 	struct css_set *cset, *tmp_cset;
2561 	int ssid, failed_ssid, ret;
2562 
2563 	/* check that we can legitimately attach to the cgroup */
2564 	if (tset->nr_tasks) {
2565 		do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2566 			if (ss->can_attach) {
2567 				tset->ssid = ssid;
2568 				ret = ss->can_attach(tset);
2569 				if (ret) {
2570 					failed_ssid = ssid;
2571 					goto out_cancel_attach;
2572 				}
2573 			}
2574 		} while_each_subsys_mask();
2575 	}
2576 
2577 	/*
2578 	 * Now that we're guaranteed success, proceed to move all tasks to
2579 	 * the new cgroup.  There are no failure cases after here, so this
2580 	 * is the commit point.
2581 	 */
2582 	spin_lock_irq(&css_set_lock);
2583 	list_for_each_entry(cset, &tset->src_csets, mg_node) {
2584 		list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2585 			struct css_set *from_cset = task_css_set(task);
2586 			struct css_set *to_cset = cset->mg_dst_cset;
2587 
2588 			get_css_set(to_cset);
2589 			to_cset->nr_tasks++;
2590 			css_set_move_task(task, from_cset, to_cset, true);
2591 			from_cset->nr_tasks--;
2592 			/*
2593 			 * If the source or destination cgroup is frozen,
2594 			 * the task might require to change its state.
2595 			 */
2596 			cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2597 						    to_cset->dfl_cgrp);
2598 			put_css_set_locked(from_cset);
2599 
2600 		}
2601 	}
2602 	spin_unlock_irq(&css_set_lock);
2603 
2604 	/*
2605 	 * Migration is committed, all target tasks are now on dst_csets.
2606 	 * Nothing is sensitive to fork() after this point.  Notify
2607 	 * controllers that migration is complete.
2608 	 */
2609 	tset->csets = &tset->dst_csets;
2610 
2611 	if (tset->nr_tasks) {
2612 		do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2613 			if (ss->attach) {
2614 				tset->ssid = ssid;
2615 				ss->attach(tset);
2616 			}
2617 		} while_each_subsys_mask();
2618 	}
2619 
2620 	ret = 0;
2621 	goto out_release_tset;
2622 
2623 out_cancel_attach:
2624 	if (tset->nr_tasks) {
2625 		do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2626 			if (ssid == failed_ssid)
2627 				break;
2628 			if (ss->cancel_attach) {
2629 				tset->ssid = ssid;
2630 				ss->cancel_attach(tset);
2631 			}
2632 		} while_each_subsys_mask();
2633 	}
2634 out_release_tset:
2635 	spin_lock_irq(&css_set_lock);
2636 	list_splice_init(&tset->dst_csets, &tset->src_csets);
2637 	list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2638 		list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2639 		list_del_init(&cset->mg_node);
2640 	}
2641 	spin_unlock_irq(&css_set_lock);
2642 
2643 	/*
2644 	 * Re-initialize the cgroup_taskset structure in case it is reused
2645 	 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2646 	 * iteration.
2647 	 */
2648 	tset->nr_tasks = 0;
2649 	tset->csets    = &tset->src_csets;
2650 	return ret;
2651 }
2652 
2653 /**
2654  * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2655  * @dst_cgrp: destination cgroup to test
2656  *
2657  * On the default hierarchy, except for the mixable, (possible) thread root
2658  * and threaded cgroups, subtree_control must be zero for migration
2659  * destination cgroups with tasks so that child cgroups don't compete
2660  * against tasks.
2661  */
cgroup_migrate_vet_dst(struct cgroup * dst_cgrp)2662 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2663 {
2664 	/* v1 doesn't have any restriction */
2665 	if (!cgroup_on_dfl(dst_cgrp))
2666 		return 0;
2667 
2668 	/* verify @dst_cgrp can host resources */
2669 	if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2670 		return -EOPNOTSUPP;
2671 
2672 	/*
2673 	 * If @dst_cgrp is already or can become a thread root or is
2674 	 * threaded, it doesn't matter.
2675 	 */
2676 	if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2677 		return 0;
2678 
2679 	/* apply no-internal-process constraint */
2680 	if (dst_cgrp->subtree_control)
2681 		return -EBUSY;
2682 
2683 	return 0;
2684 }
2685 
2686 /**
2687  * cgroup_migrate_finish - cleanup after attach
2688  * @mgctx: migration context
2689  *
2690  * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst().  See
2691  * those functions for details.
2692  */
cgroup_migrate_finish(struct cgroup_mgctx * mgctx)2693 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2694 {
2695 	struct css_set *cset, *tmp_cset;
2696 
2697 	lockdep_assert_held(&cgroup_mutex);
2698 
2699 	spin_lock_irq(&css_set_lock);
2700 
2701 	list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2702 				 mg_src_preload_node) {
2703 		cset->mg_src_cgrp = NULL;
2704 		cset->mg_dst_cgrp = NULL;
2705 		cset->mg_dst_cset = NULL;
2706 		list_del_init(&cset->mg_src_preload_node);
2707 		put_css_set_locked(cset);
2708 	}
2709 
2710 	list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2711 				 mg_dst_preload_node) {
2712 		cset->mg_src_cgrp = NULL;
2713 		cset->mg_dst_cgrp = NULL;
2714 		cset->mg_dst_cset = NULL;
2715 		list_del_init(&cset->mg_dst_preload_node);
2716 		put_css_set_locked(cset);
2717 	}
2718 
2719 	spin_unlock_irq(&css_set_lock);
2720 }
2721 
2722 /**
2723  * cgroup_migrate_add_src - add a migration source css_set
2724  * @src_cset: the source css_set to add
2725  * @dst_cgrp: the destination cgroup
2726  * @mgctx: migration context
2727  *
2728  * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp.  Pin
2729  * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2730  * up by cgroup_migrate_finish().
2731  *
2732  * This function may be called without holding cgroup_threadgroup_rwsem
2733  * even if the target is a process.  Threads may be created and destroyed
2734  * but as long as cgroup_mutex is not dropped, no new css_set can be put
2735  * into play and the preloaded css_sets are guaranteed to cover all
2736  * migrations.
2737  */
cgroup_migrate_add_src(struct css_set * src_cset,struct cgroup * dst_cgrp,struct cgroup_mgctx * mgctx)2738 void cgroup_migrate_add_src(struct css_set *src_cset,
2739 			    struct cgroup *dst_cgrp,
2740 			    struct cgroup_mgctx *mgctx)
2741 {
2742 	struct cgroup *src_cgrp;
2743 
2744 	lockdep_assert_held(&cgroup_mutex);
2745 	lockdep_assert_held(&css_set_lock);
2746 
2747 	/*
2748 	 * If ->dead, @src_set is associated with one or more dead cgroups
2749 	 * and doesn't contain any migratable tasks.  Ignore it early so
2750 	 * that the rest of migration path doesn't get confused by it.
2751 	 */
2752 	if (src_cset->dead)
2753 		return;
2754 
2755 	if (!list_empty(&src_cset->mg_src_preload_node))
2756 		return;
2757 
2758 	src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2759 
2760 	WARN_ON(src_cset->mg_src_cgrp);
2761 	WARN_ON(src_cset->mg_dst_cgrp);
2762 	WARN_ON(!list_empty(&src_cset->mg_tasks));
2763 	WARN_ON(!list_empty(&src_cset->mg_node));
2764 
2765 	src_cset->mg_src_cgrp = src_cgrp;
2766 	src_cset->mg_dst_cgrp = dst_cgrp;
2767 	get_css_set(src_cset);
2768 	list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2769 }
2770 
2771 /**
2772  * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2773  * @mgctx: migration context
2774  *
2775  * Tasks are about to be moved and all the source css_sets have been
2776  * preloaded to @mgctx->preloaded_src_csets.  This function looks up and
2777  * pins all destination css_sets, links each to its source, and append them
2778  * to @mgctx->preloaded_dst_csets.
2779  *
2780  * This function must be called after cgroup_migrate_add_src() has been
2781  * called on each migration source css_set.  After migration is performed
2782  * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2783  * @mgctx.
2784  */
cgroup_migrate_prepare_dst(struct cgroup_mgctx * mgctx)2785 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2786 {
2787 	struct css_set *src_cset, *tmp_cset;
2788 
2789 	lockdep_assert_held(&cgroup_mutex);
2790 
2791 	/* look up the dst cset for each src cset and link it to src */
2792 	list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2793 				 mg_src_preload_node) {
2794 		struct css_set *dst_cset;
2795 		struct cgroup_subsys *ss;
2796 		int ssid;
2797 
2798 		dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2799 		if (!dst_cset)
2800 			return -ENOMEM;
2801 
2802 		WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2803 
2804 		/*
2805 		 * If src cset equals dst, it's noop.  Drop the src.
2806 		 * cgroup_migrate() will skip the cset too.  Note that we
2807 		 * can't handle src == dst as some nodes are used by both.
2808 		 */
2809 		if (src_cset == dst_cset) {
2810 			src_cset->mg_src_cgrp = NULL;
2811 			src_cset->mg_dst_cgrp = NULL;
2812 			list_del_init(&src_cset->mg_src_preload_node);
2813 			put_css_set(src_cset);
2814 			put_css_set(dst_cset);
2815 			continue;
2816 		}
2817 
2818 		src_cset->mg_dst_cset = dst_cset;
2819 
2820 		if (list_empty(&dst_cset->mg_dst_preload_node))
2821 			list_add_tail(&dst_cset->mg_dst_preload_node,
2822 				      &mgctx->preloaded_dst_csets);
2823 		else
2824 			put_css_set(dst_cset);
2825 
2826 		for_each_subsys(ss, ssid)
2827 			if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2828 				mgctx->ss_mask |= 1 << ssid;
2829 	}
2830 
2831 	return 0;
2832 }
2833 
2834 /**
2835  * cgroup_migrate - migrate a process or task to a cgroup
2836  * @leader: the leader of the process or the task to migrate
2837  * @threadgroup: whether @leader points to the whole process or a single task
2838  * @mgctx: migration context
2839  *
2840  * Migrate a process or task denoted by @leader.  If migrating a process,
2841  * the caller must be holding cgroup_threadgroup_rwsem.  The caller is also
2842  * responsible for invoking cgroup_migrate_add_src() and
2843  * cgroup_migrate_prepare_dst() on the targets before invoking this
2844  * function and following up with cgroup_migrate_finish().
2845  *
2846  * As long as a controller's ->can_attach() doesn't fail, this function is
2847  * guaranteed to succeed.  This means that, excluding ->can_attach()
2848  * failure, when migrating multiple targets, the success or failure can be
2849  * decided for all targets by invoking group_migrate_prepare_dst() before
2850  * actually starting migrating.
2851  */
cgroup_migrate(struct task_struct * leader,bool threadgroup,struct cgroup_mgctx * mgctx)2852 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2853 		   struct cgroup_mgctx *mgctx)
2854 {
2855 	struct task_struct *task;
2856 
2857 	/*
2858 	 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2859 	 * already PF_EXITING could be freed from underneath us unless we
2860 	 * take an rcu_read_lock.
2861 	 */
2862 	spin_lock_irq(&css_set_lock);
2863 	rcu_read_lock();
2864 	task = leader;
2865 	do {
2866 		cgroup_migrate_add_task(task, mgctx);
2867 		if (!threadgroup)
2868 			break;
2869 	} while_each_thread(leader, task);
2870 	rcu_read_unlock();
2871 	spin_unlock_irq(&css_set_lock);
2872 
2873 	return cgroup_migrate_execute(mgctx);
2874 }
2875 
2876 /**
2877  * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2878  * @dst_cgrp: the cgroup to attach to
2879  * @leader: the task or the leader of the threadgroup to be attached
2880  * @threadgroup: attach the whole threadgroup?
2881  *
2882  * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2883  */
cgroup_attach_task(struct cgroup * dst_cgrp,struct task_struct * leader,bool threadgroup)2884 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2885 		       bool threadgroup)
2886 {
2887 	DEFINE_CGROUP_MGCTX(mgctx);
2888 	struct task_struct *task;
2889 	int ret = 0;
2890 
2891 	/* look up all src csets */
2892 	spin_lock_irq(&css_set_lock);
2893 	rcu_read_lock();
2894 	task = leader;
2895 	do {
2896 		cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2897 		if (!threadgroup)
2898 			break;
2899 	} while_each_thread(leader, task);
2900 	rcu_read_unlock();
2901 	spin_unlock_irq(&css_set_lock);
2902 
2903 	/* prepare dst csets and commit */
2904 	ret = cgroup_migrate_prepare_dst(&mgctx);
2905 	if (!ret)
2906 		ret = cgroup_migrate(leader, threadgroup, &mgctx);
2907 
2908 	cgroup_migrate_finish(&mgctx);
2909 
2910 	if (!ret)
2911 		TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2912 
2913 	return ret;
2914 }
2915 
cgroup_procs_write_start(char * buf,bool threadgroup,bool * threadgroup_locked)2916 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2917 					     bool *threadgroup_locked)
2918 {
2919 	struct task_struct *tsk;
2920 	pid_t pid;
2921 
2922 	if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2923 		return ERR_PTR(-EINVAL);
2924 
2925 	/*
2926 	 * If we migrate a single thread, we don't care about threadgroup
2927 	 * stability. If the thread is `current`, it won't exit(2) under our
2928 	 * hands or change PID through exec(2). We exclude
2929 	 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2930 	 * callers by cgroup_mutex.
2931 	 * Therefore, we can skip the global lock.
2932 	 */
2933 	lockdep_assert_held(&cgroup_mutex);
2934 	*threadgroup_locked = pid || threadgroup;
2935 	cgroup_attach_lock(*threadgroup_locked);
2936 
2937 	rcu_read_lock();
2938 	if (pid) {
2939 		tsk = find_task_by_vpid(pid);
2940 		if (!tsk) {
2941 			tsk = ERR_PTR(-ESRCH);
2942 			goto out_unlock_threadgroup;
2943 		}
2944 	} else {
2945 		tsk = current;
2946 	}
2947 
2948 	if (threadgroup)
2949 		tsk = tsk->group_leader;
2950 
2951 	/*
2952 	 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2953 	 * If userland migrates such a kthread to a non-root cgroup, it can
2954 	 * become trapped in a cpuset, or RT kthread may be born in a
2955 	 * cgroup with no rt_runtime allocated.  Just say no.
2956 	 */
2957 	if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2958 		tsk = ERR_PTR(-EINVAL);
2959 		goto out_unlock_threadgroup;
2960 	}
2961 
2962 	get_task_struct(tsk);
2963 	goto out_unlock_rcu;
2964 
2965 out_unlock_threadgroup:
2966 	cgroup_attach_unlock(*threadgroup_locked);
2967 	*threadgroup_locked = false;
2968 out_unlock_rcu:
2969 	rcu_read_unlock();
2970 	return tsk;
2971 }
2972 
cgroup_procs_write_finish(struct task_struct * task,bool threadgroup_locked)2973 void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2974 {
2975 	struct cgroup_subsys *ss;
2976 	int ssid;
2977 
2978 	/* release reference from cgroup_procs_write_start() */
2979 	put_task_struct(task);
2980 
2981 	cgroup_attach_unlock(threadgroup_locked);
2982 
2983 	for_each_subsys(ss, ssid)
2984 		if (ss->post_attach)
2985 			ss->post_attach();
2986 }
2987 
cgroup_print_ss_mask(struct seq_file * seq,u16 ss_mask)2988 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2989 {
2990 	struct cgroup_subsys *ss;
2991 	bool printed = false;
2992 	int ssid;
2993 
2994 	do_each_subsys_mask(ss, ssid, ss_mask) {
2995 		if (printed)
2996 			seq_putc(seq, ' ');
2997 		seq_puts(seq, ss->name);
2998 		printed = true;
2999 	} while_each_subsys_mask();
3000 	if (printed)
3001 		seq_putc(seq, '\n');
3002 }
3003 
3004 /* show controllers which are enabled from the parent */
cgroup_controllers_show(struct seq_file * seq,void * v)3005 static int cgroup_controllers_show(struct seq_file *seq, void *v)
3006 {
3007 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3008 
3009 	cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3010 	return 0;
3011 }
3012 
3013 /* show controllers which are enabled for a given cgroup's children */
cgroup_subtree_control_show(struct seq_file * seq,void * v)3014 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3015 {
3016 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3017 
3018 	cgroup_print_ss_mask(seq, cgrp->subtree_control);
3019 	return 0;
3020 }
3021 
3022 /**
3023  * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3024  * @cgrp: root of the subtree to update csses for
3025  *
3026  * @cgrp's control masks have changed and its subtree's css associations
3027  * need to be updated accordingly.  This function looks up all css_sets
3028  * which are attached to the subtree, creates the matching updated css_sets
3029  * and migrates the tasks to the new ones.
3030  */
cgroup_update_dfl_csses(struct cgroup * cgrp)3031 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3032 {
3033 	DEFINE_CGROUP_MGCTX(mgctx);
3034 	struct cgroup_subsys_state *d_css;
3035 	struct cgroup *dsct;
3036 	struct css_set *src_cset;
3037 	bool has_tasks;
3038 	int ret;
3039 
3040 	lockdep_assert_held(&cgroup_mutex);
3041 
3042 	/* look up all csses currently attached to @cgrp's subtree */
3043 	spin_lock_irq(&css_set_lock);
3044 	cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3045 		struct cgrp_cset_link *link;
3046 
3047 		/*
3048 		 * As cgroup_update_dfl_csses() is only called by
3049 		 * cgroup_apply_control(). The csses associated with the
3050 		 * given cgrp will not be affected by changes made to
3051 		 * its subtree_control file. We can skip them.
3052 		 */
3053 		if (dsct == cgrp)
3054 			continue;
3055 
3056 		list_for_each_entry(link, &dsct->cset_links, cset_link)
3057 			cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3058 	}
3059 	spin_unlock_irq(&css_set_lock);
3060 
3061 	/*
3062 	 * We need to write-lock threadgroup_rwsem while migrating tasks.
3063 	 * However, if there are no source csets for @cgrp, changing its
3064 	 * controllers isn't gonna produce any task migrations and the
3065 	 * write-locking can be skipped safely.
3066 	 */
3067 	has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3068 	cgroup_attach_lock(has_tasks);
3069 
3070 	/* NULL dst indicates self on default hierarchy */
3071 	ret = cgroup_migrate_prepare_dst(&mgctx);
3072 	if (ret)
3073 		goto out_finish;
3074 
3075 	spin_lock_irq(&css_set_lock);
3076 	list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3077 			    mg_src_preload_node) {
3078 		struct task_struct *task, *ntask;
3079 
3080 		/* all tasks in src_csets need to be migrated */
3081 		list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3082 			cgroup_migrate_add_task(task, &mgctx);
3083 	}
3084 	spin_unlock_irq(&css_set_lock);
3085 
3086 	ret = cgroup_migrate_execute(&mgctx);
3087 out_finish:
3088 	cgroup_migrate_finish(&mgctx);
3089 	cgroup_attach_unlock(has_tasks);
3090 	return ret;
3091 }
3092 
3093 /**
3094  * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3095  * @cgrp: root of the target subtree
3096  *
3097  * Because css offlining is asynchronous, userland may try to re-enable a
3098  * controller while the previous css is still around.  This function grabs
3099  * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3100  */
cgroup_lock_and_drain_offline(struct cgroup * cgrp)3101 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3102 	__acquires(&cgroup_mutex)
3103 {
3104 	struct cgroup *dsct;
3105 	struct cgroup_subsys_state *d_css;
3106 	struct cgroup_subsys *ss;
3107 	int ssid;
3108 
3109 restart:
3110 	mutex_lock(&cgroup_mutex);
3111 
3112 	cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3113 		for_each_subsys(ss, ssid) {
3114 			struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3115 			DEFINE_WAIT(wait);
3116 
3117 			if (!css || !percpu_ref_is_dying(&css->refcnt))
3118 				continue;
3119 
3120 			cgroup_get_live(dsct);
3121 			prepare_to_wait(&dsct->offline_waitq, &wait,
3122 					TASK_UNINTERRUPTIBLE);
3123 
3124 			mutex_unlock(&cgroup_mutex);
3125 			schedule();
3126 			finish_wait(&dsct->offline_waitq, &wait);
3127 
3128 			cgroup_put(dsct);
3129 			goto restart;
3130 		}
3131 	}
3132 }
3133 
3134 /**
3135  * cgroup_save_control - save control masks and dom_cgrp of a subtree
3136  * @cgrp: root of the target subtree
3137  *
3138  * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3139  * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3140  * itself.
3141  */
cgroup_save_control(struct cgroup * cgrp)3142 static void cgroup_save_control(struct cgroup *cgrp)
3143 {
3144 	struct cgroup *dsct;
3145 	struct cgroup_subsys_state *d_css;
3146 
3147 	cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3148 		dsct->old_subtree_control = dsct->subtree_control;
3149 		dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3150 		dsct->old_dom_cgrp = dsct->dom_cgrp;
3151 	}
3152 }
3153 
3154 /**
3155  * cgroup_propagate_control - refresh control masks of a subtree
3156  * @cgrp: root of the target subtree
3157  *
3158  * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3159  * ->subtree_control and propagate controller availability through the
3160  * subtree so that descendants don't have unavailable controllers enabled.
3161  */
cgroup_propagate_control(struct cgroup * cgrp)3162 static void cgroup_propagate_control(struct cgroup *cgrp)
3163 {
3164 	struct cgroup *dsct;
3165 	struct cgroup_subsys_state *d_css;
3166 
3167 	cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3168 		dsct->subtree_control &= cgroup_control(dsct);
3169 		dsct->subtree_ss_mask =
3170 			cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3171 						    cgroup_ss_mask(dsct));
3172 	}
3173 }
3174 
3175 /**
3176  * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3177  * @cgrp: root of the target subtree
3178  *
3179  * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3180  * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3181  * itself.
3182  */
cgroup_restore_control(struct cgroup * cgrp)3183 static void cgroup_restore_control(struct cgroup *cgrp)
3184 {
3185 	struct cgroup *dsct;
3186 	struct cgroup_subsys_state *d_css;
3187 
3188 	cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3189 		dsct->subtree_control = dsct->old_subtree_control;
3190 		dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3191 		dsct->dom_cgrp = dsct->old_dom_cgrp;
3192 	}
3193 }
3194 
css_visible(struct cgroup_subsys_state * css)3195 static bool css_visible(struct cgroup_subsys_state *css)
3196 {
3197 	struct cgroup_subsys *ss = css->ss;
3198 	struct cgroup *cgrp = css->cgroup;
3199 
3200 	if (cgroup_control(cgrp) & (1 << ss->id))
3201 		return true;
3202 	if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3203 		return false;
3204 	return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3205 }
3206 
3207 /**
3208  * cgroup_apply_control_enable - enable or show csses according to control
3209  * @cgrp: root of the target subtree
3210  *
3211  * Walk @cgrp's subtree and create new csses or make the existing ones
3212  * visible.  A css is created invisible if it's being implicitly enabled
3213  * through dependency.  An invisible css is made visible when the userland
3214  * explicitly enables it.
3215  *
3216  * Returns 0 on success, -errno on failure.  On failure, csses which have
3217  * been processed already aren't cleaned up.  The caller is responsible for
3218  * cleaning up with cgroup_apply_control_disable().
3219  */
cgroup_apply_control_enable(struct cgroup * cgrp)3220 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3221 {
3222 	struct cgroup *dsct;
3223 	struct cgroup_subsys_state *d_css;
3224 	struct cgroup_subsys *ss;
3225 	int ssid, ret;
3226 
3227 	cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3228 		for_each_subsys(ss, ssid) {
3229 			struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3230 
3231 			if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3232 				continue;
3233 
3234 			if (!css) {
3235 				css = css_create(dsct, ss);
3236 				if (IS_ERR(css))
3237 					return PTR_ERR(css);
3238 			}
3239 
3240 			WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3241 
3242 			if (css_visible(css)) {
3243 				ret = css_populate_dir(css);
3244 				if (ret)
3245 					return ret;
3246 			}
3247 		}
3248 	}
3249 
3250 	return 0;
3251 }
3252 
3253 /**
3254  * cgroup_apply_control_disable - kill or hide csses according to control
3255  * @cgrp: root of the target subtree
3256  *
3257  * Walk @cgrp's subtree and kill and hide csses so that they match
3258  * cgroup_ss_mask() and cgroup_visible_mask().
3259  *
3260  * A css is hidden when the userland requests it to be disabled while other
3261  * subsystems are still depending on it.  The css must not actively control
3262  * resources and be in the vanilla state if it's made visible again later.
3263  * Controllers which may be depended upon should provide ->css_reset() for
3264  * this purpose.
3265  */
cgroup_apply_control_disable(struct cgroup * cgrp)3266 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3267 {
3268 	struct cgroup *dsct;
3269 	struct cgroup_subsys_state *d_css;
3270 	struct cgroup_subsys *ss;
3271 	int ssid;
3272 
3273 	cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3274 		for_each_subsys(ss, ssid) {
3275 			struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3276 
3277 			if (!css)
3278 				continue;
3279 
3280 			WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3281 
3282 			if (css->parent &&
3283 			    !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3284 				kill_css(css);
3285 			} else if (!css_visible(css)) {
3286 				css_clear_dir(css);
3287 				if (ss->css_reset)
3288 					ss->css_reset(css);
3289 			}
3290 		}
3291 	}
3292 }
3293 
3294 /**
3295  * cgroup_apply_control - apply control mask updates to the subtree
3296  * @cgrp: root of the target subtree
3297  *
3298  * subsystems can be enabled and disabled in a subtree using the following
3299  * steps.
3300  *
3301  * 1. Call cgroup_save_control() to stash the current state.
3302  * 2. Update ->subtree_control masks in the subtree as desired.
3303  * 3. Call cgroup_apply_control() to apply the changes.
3304  * 4. Optionally perform other related operations.
3305  * 5. Call cgroup_finalize_control() to finish up.
3306  *
3307  * This function implements step 3 and propagates the mask changes
3308  * throughout @cgrp's subtree, updates csses accordingly and perform
3309  * process migrations.
3310  */
cgroup_apply_control(struct cgroup * cgrp)3311 static int cgroup_apply_control(struct cgroup *cgrp)
3312 {
3313 	int ret;
3314 
3315 	cgroup_propagate_control(cgrp);
3316 
3317 	ret = cgroup_apply_control_enable(cgrp);
3318 	if (ret)
3319 		return ret;
3320 
3321 	/*
3322 	 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3323 	 * making the following cgroup_update_dfl_csses() properly update
3324 	 * css associations of all tasks in the subtree.
3325 	 */
3326 	return cgroup_update_dfl_csses(cgrp);
3327 }
3328 
3329 /**
3330  * cgroup_finalize_control - finalize control mask update
3331  * @cgrp: root of the target subtree
3332  * @ret: the result of the update
3333  *
3334  * Finalize control mask update.  See cgroup_apply_control() for more info.
3335  */
cgroup_finalize_control(struct cgroup * cgrp,int ret)3336 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3337 {
3338 	if (ret) {
3339 		cgroup_restore_control(cgrp);
3340 		cgroup_propagate_control(cgrp);
3341 	}
3342 
3343 	cgroup_apply_control_disable(cgrp);
3344 }
3345 
cgroup_vet_subtree_control_enable(struct cgroup * cgrp,u16 enable)3346 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3347 {
3348 	u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3349 
3350 	/* if nothing is getting enabled, nothing to worry about */
3351 	if (!enable)
3352 		return 0;
3353 
3354 	/* can @cgrp host any resources? */
3355 	if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3356 		return -EOPNOTSUPP;
3357 
3358 	/* mixables don't care */
3359 	if (cgroup_is_mixable(cgrp))
3360 		return 0;
3361 
3362 	if (domain_enable) {
3363 		/* can't enable domain controllers inside a thread subtree */
3364 		if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3365 			return -EOPNOTSUPP;
3366 	} else {
3367 		/*
3368 		 * Threaded controllers can handle internal competitions
3369 		 * and are always allowed inside a (prospective) thread
3370 		 * subtree.
3371 		 */
3372 		if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3373 			return 0;
3374 	}
3375 
3376 	/*
3377 	 * Controllers can't be enabled for a cgroup with tasks to avoid
3378 	 * child cgroups competing against tasks.
3379 	 */
3380 	if (cgroup_has_tasks(cgrp))
3381 		return -EBUSY;
3382 
3383 	return 0;
3384 }
3385 
3386 /* 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)3387 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3388 					    char *buf, size_t nbytes,
3389 					    loff_t off)
3390 {
3391 	u16 enable = 0, disable = 0;
3392 	struct cgroup *cgrp, *child;
3393 	struct cgroup_subsys *ss;
3394 	char *tok;
3395 	int ssid, ret;
3396 
3397 	/*
3398 	 * Parse input - space separated list of subsystem names prefixed
3399 	 * with either + or -.
3400 	 */
3401 	buf = strstrip(buf);
3402 	while ((tok = strsep(&buf, " "))) {
3403 		if (tok[0] == '\0')
3404 			continue;
3405 		do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3406 			if (!cgroup_ssid_enabled(ssid) ||
3407 			    strcmp(tok + 1, ss->name))
3408 				continue;
3409 
3410 			if (*tok == '+') {
3411 				enable |= 1 << ssid;
3412 				disable &= ~(1 << ssid);
3413 			} else if (*tok == '-') {
3414 				disable |= 1 << ssid;
3415 				enable &= ~(1 << ssid);
3416 			} else {
3417 				return -EINVAL;
3418 			}
3419 			break;
3420 		} while_each_subsys_mask();
3421 		if (ssid == CGROUP_SUBSYS_COUNT)
3422 			return -EINVAL;
3423 	}
3424 
3425 	cgrp = cgroup_kn_lock_live(of->kn, true);
3426 	if (!cgrp)
3427 		return -ENODEV;
3428 
3429 	for_each_subsys(ss, ssid) {
3430 		if (enable & (1 << ssid)) {
3431 			if (cgrp->subtree_control & (1 << ssid)) {
3432 				enable &= ~(1 << ssid);
3433 				continue;
3434 			}
3435 
3436 			if (!(cgroup_control(cgrp) & (1 << ssid))) {
3437 				ret = -ENOENT;
3438 				goto out_unlock;
3439 			}
3440 		} else if (disable & (1 << ssid)) {
3441 			if (!(cgrp->subtree_control & (1 << ssid))) {
3442 				disable &= ~(1 << ssid);
3443 				continue;
3444 			}
3445 
3446 			/* a child has it enabled? */
3447 			cgroup_for_each_live_child(child, cgrp) {
3448 				if (child->subtree_control & (1 << ssid)) {
3449 					ret = -EBUSY;
3450 					goto out_unlock;
3451 				}
3452 			}
3453 		}
3454 	}
3455 
3456 	if (!enable && !disable) {
3457 		ret = 0;
3458 		goto out_unlock;
3459 	}
3460 
3461 	ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3462 	if (ret)
3463 		goto out_unlock;
3464 
3465 	/* save and update control masks and prepare csses */
3466 	cgroup_save_control(cgrp);
3467 
3468 	cgrp->subtree_control |= enable;
3469 	cgrp->subtree_control &= ~disable;
3470 
3471 	ret = cgroup_apply_control(cgrp);
3472 	cgroup_finalize_control(cgrp, ret);
3473 	if (ret)
3474 		goto out_unlock;
3475 
3476 	kernfs_activate(cgrp->kn);
3477 out_unlock:
3478 	cgroup_kn_unlock(of->kn);
3479 	return ret ?: nbytes;
3480 }
3481 
3482 /**
3483  * cgroup_enable_threaded - make @cgrp threaded
3484  * @cgrp: the target cgroup
3485  *
3486  * Called when "threaded" is written to the cgroup.type interface file and
3487  * tries to make @cgrp threaded and join the parent's resource domain.
3488  * This function is never called on the root cgroup as cgroup.type doesn't
3489  * exist on it.
3490  */
cgroup_enable_threaded(struct cgroup * cgrp)3491 static int cgroup_enable_threaded(struct cgroup *cgrp)
3492 {
3493 	struct cgroup *parent = cgroup_parent(cgrp);
3494 	struct cgroup *dom_cgrp = parent->dom_cgrp;
3495 	struct cgroup *dsct;
3496 	struct cgroup_subsys_state *d_css;
3497 	int ret;
3498 
3499 	lockdep_assert_held(&cgroup_mutex);
3500 
3501 	/* noop if already threaded */
3502 	if (cgroup_is_threaded(cgrp))
3503 		return 0;
3504 
3505 	/*
3506 	 * If @cgroup is populated or has domain controllers enabled, it
3507 	 * can't be switched.  While the below cgroup_can_be_thread_root()
3508 	 * test can catch the same conditions, that's only when @parent is
3509 	 * not mixable, so let's check it explicitly.
3510 	 */
3511 	if (cgroup_is_populated(cgrp) ||
3512 	    cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3513 		return -EOPNOTSUPP;
3514 
3515 	/* we're joining the parent's domain, ensure its validity */
3516 	if (!cgroup_is_valid_domain(dom_cgrp) ||
3517 	    !cgroup_can_be_thread_root(dom_cgrp))
3518 		return -EOPNOTSUPP;
3519 
3520 	/*
3521 	 * The following shouldn't cause actual migrations and should
3522 	 * always succeed.
3523 	 */
3524 	cgroup_save_control(cgrp);
3525 
3526 	cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3527 		if (dsct == cgrp || cgroup_is_threaded(dsct))
3528 			dsct->dom_cgrp = dom_cgrp;
3529 
3530 	ret = cgroup_apply_control(cgrp);
3531 	if (!ret)
3532 		parent->nr_threaded_children++;
3533 
3534 	cgroup_finalize_control(cgrp, ret);
3535 	return ret;
3536 }
3537 
cgroup_type_show(struct seq_file * seq,void * v)3538 static int cgroup_type_show(struct seq_file *seq, void *v)
3539 {
3540 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3541 
3542 	if (cgroup_is_threaded(cgrp))
3543 		seq_puts(seq, "threaded\n");
3544 	else if (!cgroup_is_valid_domain(cgrp))
3545 		seq_puts(seq, "domain invalid\n");
3546 	else if (cgroup_is_thread_root(cgrp))
3547 		seq_puts(seq, "domain threaded\n");
3548 	else
3549 		seq_puts(seq, "domain\n");
3550 
3551 	return 0;
3552 }
3553 
cgroup_type_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3554 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3555 				 size_t nbytes, loff_t off)
3556 {
3557 	struct cgroup *cgrp;
3558 	int ret;
3559 
3560 	/* only switching to threaded mode is supported */
3561 	if (strcmp(strstrip(buf), "threaded"))
3562 		return -EINVAL;
3563 
3564 	/* drain dying csses before we re-apply (threaded) subtree control */
3565 	cgrp = cgroup_kn_lock_live(of->kn, true);
3566 	if (!cgrp)
3567 		return -ENOENT;
3568 
3569 	/* threaded can only be enabled */
3570 	ret = cgroup_enable_threaded(cgrp);
3571 
3572 	cgroup_kn_unlock(of->kn);
3573 	return ret ?: nbytes;
3574 }
3575 
cgroup_max_descendants_show(struct seq_file * seq,void * v)3576 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3577 {
3578 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3579 	int descendants = READ_ONCE(cgrp->max_descendants);
3580 
3581 	if (descendants == INT_MAX)
3582 		seq_puts(seq, "max\n");
3583 	else
3584 		seq_printf(seq, "%d\n", descendants);
3585 
3586 	return 0;
3587 }
3588 
cgroup_max_descendants_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3589 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3590 					   char *buf, size_t nbytes, loff_t off)
3591 {
3592 	struct cgroup *cgrp;
3593 	int descendants;
3594 	ssize_t ret;
3595 
3596 	buf = strstrip(buf);
3597 	if (!strcmp(buf, "max")) {
3598 		descendants = INT_MAX;
3599 	} else {
3600 		ret = kstrtoint(buf, 0, &descendants);
3601 		if (ret)
3602 			return ret;
3603 	}
3604 
3605 	if (descendants < 0)
3606 		return -ERANGE;
3607 
3608 	cgrp = cgroup_kn_lock_live(of->kn, false);
3609 	if (!cgrp)
3610 		return -ENOENT;
3611 
3612 	cgrp->max_descendants = descendants;
3613 
3614 	cgroup_kn_unlock(of->kn);
3615 
3616 	return nbytes;
3617 }
3618 
cgroup_max_depth_show(struct seq_file * seq,void * v)3619 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3620 {
3621 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3622 	int depth = READ_ONCE(cgrp->max_depth);
3623 
3624 	if (depth == INT_MAX)
3625 		seq_puts(seq, "max\n");
3626 	else
3627 		seq_printf(seq, "%d\n", depth);
3628 
3629 	return 0;
3630 }
3631 
cgroup_max_depth_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3632 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3633 				      char *buf, size_t nbytes, loff_t off)
3634 {
3635 	struct cgroup *cgrp;
3636 	ssize_t ret;
3637 	int depth;
3638 
3639 	buf = strstrip(buf);
3640 	if (!strcmp(buf, "max")) {
3641 		depth = INT_MAX;
3642 	} else {
3643 		ret = kstrtoint(buf, 0, &depth);
3644 		if (ret)
3645 			return ret;
3646 	}
3647 
3648 	if (depth < 0)
3649 		return -ERANGE;
3650 
3651 	cgrp = cgroup_kn_lock_live(of->kn, false);
3652 	if (!cgrp)
3653 		return -ENOENT;
3654 
3655 	cgrp->max_depth = depth;
3656 
3657 	cgroup_kn_unlock(of->kn);
3658 
3659 	return nbytes;
3660 }
3661 
cgroup_events_show(struct seq_file * seq,void * v)3662 static int cgroup_events_show(struct seq_file *seq, void *v)
3663 {
3664 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3665 
3666 	seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3667 	seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3668 
3669 	return 0;
3670 }
3671 
cgroup_stat_show(struct seq_file * seq,void * v)3672 static int cgroup_stat_show(struct seq_file *seq, void *v)
3673 {
3674 	struct cgroup *cgroup = seq_css(seq)->cgroup;
3675 
3676 	seq_printf(seq, "nr_descendants %d\n",
3677 		   cgroup->nr_descendants);
3678 	seq_printf(seq, "nr_dying_descendants %d\n",
3679 		   cgroup->nr_dying_descendants);
3680 
3681 	return 0;
3682 }
3683 
cgroup_extra_stat_show(struct seq_file * seq,struct cgroup * cgrp,int ssid)3684 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3685 						 struct cgroup *cgrp, int ssid)
3686 {
3687 	struct cgroup_subsys *ss = cgroup_subsys[ssid];
3688 	struct cgroup_subsys_state *css;
3689 	int ret;
3690 
3691 	if (!ss->css_extra_stat_show)
3692 		return 0;
3693 
3694 	css = cgroup_tryget_css(cgrp, ss);
3695 	if (!css)
3696 		return 0;
3697 
3698 	ret = ss->css_extra_stat_show(seq, css);
3699 	css_put(css);
3700 	return ret;
3701 }
3702 
cpu_stat_show(struct seq_file * seq,void * v)3703 static int cpu_stat_show(struct seq_file *seq, void *v)
3704 {
3705 	struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3706 	int ret = 0;
3707 
3708 	cgroup_base_stat_cputime_show(seq);
3709 #ifdef CONFIG_CGROUP_SCHED
3710 	ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3711 #endif
3712 	return ret;
3713 }
3714 
3715 #ifdef CONFIG_PSI
cgroup_io_pressure_show(struct seq_file * seq,void * v)3716 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3717 {
3718 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3719 	struct psi_group *psi = cgroup_psi(cgrp);
3720 
3721 	return psi_show(seq, psi, PSI_IO);
3722 }
cgroup_memory_pressure_show(struct seq_file * seq,void * v)3723 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3724 {
3725 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3726 	struct psi_group *psi = cgroup_psi(cgrp);
3727 
3728 	return psi_show(seq, psi, PSI_MEM);
3729 }
cgroup_cpu_pressure_show(struct seq_file * seq,void * v)3730 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3731 {
3732 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3733 	struct psi_group *psi = cgroup_psi(cgrp);
3734 
3735 	return psi_show(seq, psi, PSI_CPU);
3736 }
3737 
pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,enum psi_res res)3738 static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3739 			      size_t nbytes, enum psi_res res)
3740 {
3741 	struct cgroup_file_ctx *ctx = of->priv;
3742 	struct psi_trigger *new;
3743 	struct cgroup *cgrp;
3744 	struct psi_group *psi;
3745 
3746 	cgrp = cgroup_kn_lock_live(of->kn, false);
3747 	if (!cgrp)
3748 		return -ENODEV;
3749 
3750 	cgroup_get(cgrp);
3751 	cgroup_kn_unlock(of->kn);
3752 
3753 	/* Allow only one trigger per file descriptor */
3754 	if (ctx->psi.trigger) {
3755 		cgroup_put(cgrp);
3756 		return -EBUSY;
3757 	}
3758 
3759 	psi = cgroup_psi(cgrp);
3760 	new = psi_trigger_create(psi, buf, res);
3761 	if (IS_ERR(new)) {
3762 		cgroup_put(cgrp);
3763 		return PTR_ERR(new);
3764 	}
3765 
3766 	smp_store_release(&ctx->psi.trigger, new);
3767 	cgroup_put(cgrp);
3768 
3769 	return nbytes;
3770 }
3771 
cgroup_io_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3772 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3773 					  char *buf, size_t nbytes,
3774 					  loff_t off)
3775 {
3776 	return pressure_write(of, buf, nbytes, PSI_IO);
3777 }
3778 
cgroup_memory_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3779 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3780 					  char *buf, size_t nbytes,
3781 					  loff_t off)
3782 {
3783 	return pressure_write(of, buf, nbytes, PSI_MEM);
3784 }
3785 
cgroup_cpu_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3786 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3787 					  char *buf, size_t nbytes,
3788 					  loff_t off)
3789 {
3790 	return pressure_write(of, buf, nbytes, PSI_CPU);
3791 }
3792 
3793 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
cgroup_irq_pressure_show(struct seq_file * seq,void * v)3794 static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
3795 {
3796 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3797 	struct psi_group *psi = cgroup_psi(cgrp);
3798 
3799 	return psi_show(seq, psi, PSI_IRQ);
3800 }
3801 
cgroup_irq_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3802 static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
3803 					 char *buf, size_t nbytes,
3804 					 loff_t off)
3805 {
3806 	return pressure_write(of, buf, nbytes, PSI_IRQ);
3807 }
3808 #endif
3809 
cgroup_pressure_show(struct seq_file * seq,void * v)3810 static int cgroup_pressure_show(struct seq_file *seq, void *v)
3811 {
3812 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3813 	struct psi_group *psi = cgroup_psi(cgrp);
3814 
3815 	seq_printf(seq, "%d\n", psi->enabled);
3816 
3817 	return 0;
3818 }
3819 
cgroup_pressure_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3820 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
3821 				     char *buf, size_t nbytes,
3822 				     loff_t off)
3823 {
3824 	ssize_t ret;
3825 	int enable;
3826 	struct cgroup *cgrp;
3827 	struct psi_group *psi;
3828 
3829 	ret = kstrtoint(strstrip(buf), 0, &enable);
3830 	if (ret)
3831 		return ret;
3832 
3833 	if (enable < 0 || enable > 1)
3834 		return -ERANGE;
3835 
3836 	cgrp = cgroup_kn_lock_live(of->kn, false);
3837 	if (!cgrp)
3838 		return -ENOENT;
3839 
3840 	psi = cgroup_psi(cgrp);
3841 	if (psi->enabled != enable) {
3842 		int i;
3843 
3844 		/* show or hide {cpu,memory,io,irq}.pressure files */
3845 		for (i = 0; i < NR_PSI_RESOURCES; i++)
3846 			cgroup_file_show(&cgrp->psi_files[i], enable);
3847 
3848 		psi->enabled = enable;
3849 		if (enable)
3850 			psi_cgroup_restart(psi);
3851 	}
3852 
3853 	cgroup_kn_unlock(of->kn);
3854 
3855 	return nbytes;
3856 }
3857 
cgroup_pressure_poll(struct kernfs_open_file * of,poll_table * pt)3858 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3859 					  poll_table *pt)
3860 {
3861 	struct cgroup_file_ctx *ctx = of->priv;
3862 
3863 	return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3864 }
3865 
cgroup_pressure_release(struct kernfs_open_file * of)3866 static void cgroup_pressure_release(struct kernfs_open_file *of)
3867 {
3868 	struct cgroup_file_ctx *ctx = of->priv;
3869 
3870 	psi_trigger_destroy(ctx->psi.trigger);
3871 }
3872 
cgroup_psi_enabled(void)3873 bool cgroup_psi_enabled(void)
3874 {
3875 	if (static_branch_likely(&psi_disabled))
3876 		return false;
3877 
3878 	return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3879 }
3880 
3881 #else /* CONFIG_PSI */
cgroup_psi_enabled(void)3882 bool cgroup_psi_enabled(void)
3883 {
3884 	return false;
3885 }
3886 
3887 #endif /* CONFIG_PSI */
3888 
cgroup_freeze_show(struct seq_file * seq,void * v)3889 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3890 {
3891 	struct cgroup *cgrp = seq_css(seq)->cgroup;
3892 
3893 	seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3894 
3895 	return 0;
3896 }
3897 
cgroup_freeze_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3898 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3899 				   char *buf, size_t nbytes, loff_t off)
3900 {
3901 	struct cgroup *cgrp;
3902 	ssize_t ret;
3903 	int freeze;
3904 
3905 	ret = kstrtoint(strstrip(buf), 0, &freeze);
3906 	if (ret)
3907 		return ret;
3908 
3909 	if (freeze < 0 || freeze > 1)
3910 		return -ERANGE;
3911 
3912 	cgrp = cgroup_kn_lock_live(of->kn, false);
3913 	if (!cgrp)
3914 		return -ENOENT;
3915 
3916 	cgroup_freeze(cgrp, freeze);
3917 
3918 	cgroup_kn_unlock(of->kn);
3919 
3920 	return nbytes;
3921 }
3922 
__cgroup_kill(struct cgroup * cgrp)3923 static void __cgroup_kill(struct cgroup *cgrp)
3924 {
3925 	struct css_task_iter it;
3926 	struct task_struct *task;
3927 
3928 	lockdep_assert_held(&cgroup_mutex);
3929 
3930 	spin_lock_irq(&css_set_lock);
3931 	set_bit(CGRP_KILL, &cgrp->flags);
3932 	spin_unlock_irq(&css_set_lock);
3933 
3934 	css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3935 	while ((task = css_task_iter_next(&it))) {
3936 		/* Ignore kernel threads here. */
3937 		if (task->flags & PF_KTHREAD)
3938 			continue;
3939 
3940 		/* Skip tasks that are already dying. */
3941 		if (__fatal_signal_pending(task))
3942 			continue;
3943 
3944 		send_sig(SIGKILL, task, 0);
3945 	}
3946 	css_task_iter_end(&it);
3947 
3948 	spin_lock_irq(&css_set_lock);
3949 	clear_bit(CGRP_KILL, &cgrp->flags);
3950 	spin_unlock_irq(&css_set_lock);
3951 }
3952 
cgroup_kill(struct cgroup * cgrp)3953 static void cgroup_kill(struct cgroup *cgrp)
3954 {
3955 	struct cgroup_subsys_state *css;
3956 	struct cgroup *dsct;
3957 
3958 	lockdep_assert_held(&cgroup_mutex);
3959 
3960 	cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3961 		__cgroup_kill(dsct);
3962 }
3963 
cgroup_kill_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3964 static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3965 				 size_t nbytes, loff_t off)
3966 {
3967 	ssize_t ret = 0;
3968 	int kill;
3969 	struct cgroup *cgrp;
3970 
3971 	ret = kstrtoint(strstrip(buf), 0, &kill);
3972 	if (ret)
3973 		return ret;
3974 
3975 	if (kill != 1)
3976 		return -ERANGE;
3977 
3978 	cgrp = cgroup_kn_lock_live(of->kn, false);
3979 	if (!cgrp)
3980 		return -ENOENT;
3981 
3982 	/*
3983 	 * Killing is a process directed operation, i.e. the whole thread-group
3984 	 * is taken down so act like we do for cgroup.procs and only make this
3985 	 * writable in non-threaded cgroups.
3986 	 */
3987 	if (cgroup_is_threaded(cgrp))
3988 		ret = -EOPNOTSUPP;
3989 	else
3990 		cgroup_kill(cgrp);
3991 
3992 	cgroup_kn_unlock(of->kn);
3993 
3994 	return ret ?: nbytes;
3995 }
3996 
cgroup_file_open(struct kernfs_open_file * of)3997 static int cgroup_file_open(struct kernfs_open_file *of)
3998 {
3999 	struct cftype *cft = of_cft(of);
4000 	struct cgroup_file_ctx *ctx;
4001 	int ret;
4002 
4003 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
4004 	if (!ctx)
4005 		return -ENOMEM;
4006 
4007 	ctx->ns = current->nsproxy->cgroup_ns;
4008 	get_cgroup_ns(ctx->ns);
4009 	of->priv = ctx;
4010 
4011 	if (!cft->open)
4012 		return 0;
4013 
4014 	ret = cft->open(of);
4015 	if (ret) {
4016 		put_cgroup_ns(ctx->ns);
4017 		kfree(ctx);
4018 	}
4019 	return ret;
4020 }
4021 
cgroup_file_release(struct kernfs_open_file * of)4022 static void cgroup_file_release(struct kernfs_open_file *of)
4023 {
4024 	struct cftype *cft = of_cft(of);
4025 	struct cgroup_file_ctx *ctx = of->priv;
4026 
4027 	if (cft->release)
4028 		cft->release(of);
4029 	put_cgroup_ns(ctx->ns);
4030 	kfree(ctx);
4031 }
4032 
cgroup_file_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)4033 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4034 				 size_t nbytes, loff_t off)
4035 {
4036 	struct cgroup_file_ctx *ctx = of->priv;
4037 	struct cgroup *cgrp = of->kn->parent->priv;
4038 	struct cftype *cft = of_cft(of);
4039 	struct cgroup_subsys_state *css;
4040 	int ret;
4041 
4042 	if (!nbytes)
4043 		return 0;
4044 
4045 	/*
4046 	 * If namespaces are delegation boundaries, disallow writes to
4047 	 * files in an non-init namespace root from inside the namespace
4048 	 * except for the files explicitly marked delegatable -
4049 	 * cgroup.procs and cgroup.subtree_control.
4050 	 */
4051 	if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4052 	    !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4053 	    ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4054 		return -EPERM;
4055 
4056 	if (cft->write)
4057 		return cft->write(of, buf, nbytes, off);
4058 
4059 	/*
4060 	 * kernfs guarantees that a file isn't deleted with operations in
4061 	 * flight, which means that the matching css is and stays alive and
4062 	 * doesn't need to be pinned.  The RCU locking is not necessary
4063 	 * either.  It's just for the convenience of using cgroup_css().
4064 	 */
4065 	rcu_read_lock();
4066 	css = cgroup_css(cgrp, cft->ss);
4067 	rcu_read_unlock();
4068 
4069 	if (cft->write_u64) {
4070 		unsigned long long v;
4071 		ret = kstrtoull(buf, 0, &v);
4072 		if (!ret)
4073 			ret = cft->write_u64(css, cft, v);
4074 	} else if (cft->write_s64) {
4075 		long long v;
4076 		ret = kstrtoll(buf, 0, &v);
4077 		if (!ret)
4078 			ret = cft->write_s64(css, cft, v);
4079 	} else {
4080 		ret = -EINVAL;
4081 	}
4082 
4083 	return ret ?: nbytes;
4084 }
4085 
cgroup_file_poll(struct kernfs_open_file * of,poll_table * pt)4086 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4087 {
4088 	struct cftype *cft = of_cft(of);
4089 
4090 	if (cft->poll)
4091 		return cft->poll(of, pt);
4092 
4093 	return kernfs_generic_poll(of, pt);
4094 }
4095 
cgroup_seqfile_start(struct seq_file * seq,loff_t * ppos)4096 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4097 {
4098 	return seq_cft(seq)->seq_start(seq, ppos);
4099 }
4100 
cgroup_seqfile_next(struct seq_file * seq,void * v,loff_t * ppos)4101 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4102 {
4103 	return seq_cft(seq)->seq_next(seq, v, ppos);
4104 }
4105 
cgroup_seqfile_stop(struct seq_file * seq,void * v)4106 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4107 {
4108 	if (seq_cft(seq)->seq_stop)
4109 		seq_cft(seq)->seq_stop(seq, v);
4110 }
4111 
cgroup_seqfile_show(struct seq_file * m,void * arg)4112 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4113 {
4114 	struct cftype *cft = seq_cft(m);
4115 	struct cgroup_subsys_state *css = seq_css(m);
4116 
4117 	if (cft->seq_show)
4118 		return cft->seq_show(m, arg);
4119 
4120 	if (cft->read_u64)
4121 		seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4122 	else if (cft->read_s64)
4123 		seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4124 	else
4125 		return -EINVAL;
4126 	return 0;
4127 }
4128 
4129 static struct kernfs_ops cgroup_kf_single_ops = {
4130 	.atomic_write_len	= PAGE_SIZE,
4131 	.open			= cgroup_file_open,
4132 	.release		= cgroup_file_release,
4133 	.write			= cgroup_file_write,
4134 	.poll			= cgroup_file_poll,
4135 	.seq_show		= cgroup_seqfile_show,
4136 };
4137 
4138 static struct kernfs_ops cgroup_kf_ops = {
4139 	.atomic_write_len	= PAGE_SIZE,
4140 	.open			= cgroup_file_open,
4141 	.release		= cgroup_file_release,
4142 	.write			= cgroup_file_write,
4143 	.poll			= cgroup_file_poll,
4144 	.seq_start		= cgroup_seqfile_start,
4145 	.seq_next		= cgroup_seqfile_next,
4146 	.seq_stop		= cgroup_seqfile_stop,
4147 	.seq_show		= cgroup_seqfile_show,
4148 };
4149 
4150 /* set uid and gid of cgroup dirs and files to that of the creator */
cgroup_kn_set_ugid(struct kernfs_node * kn)4151 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
4152 {
4153 	struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
4154 			       .ia_uid = current_fsuid(),
4155 			       .ia_gid = current_fsgid(), };
4156 
4157 	if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
4158 	    gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
4159 		return 0;
4160 
4161 	return kernfs_setattr(kn, &iattr);
4162 }
4163 
cgroup_file_notify_timer(struct timer_list * timer)4164 static void cgroup_file_notify_timer(struct timer_list *timer)
4165 {
4166 	cgroup_file_notify(container_of(timer, struct cgroup_file,
4167 					notify_timer));
4168 }
4169 
cgroup_add_file(struct cgroup_subsys_state * css,struct cgroup * cgrp,struct cftype * cft)4170 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4171 			   struct cftype *cft)
4172 {
4173 	char name[CGROUP_FILE_NAME_MAX];
4174 	struct kernfs_node *kn;
4175 	struct lock_class_key *key = NULL;
4176 	int ret;
4177 
4178 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4179 	key = &cft->lockdep_key;
4180 #endif
4181 	kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4182 				  cgroup_file_mode(cft),
4183 				  GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
4184 				  0, cft->kf_ops, cft,
4185 				  NULL, key);
4186 	if (IS_ERR(kn))
4187 		return PTR_ERR(kn);
4188 
4189 	ret = cgroup_kn_set_ugid(kn);
4190 	if (ret) {
4191 		kernfs_remove(kn);
4192 		return ret;
4193 	}
4194 
4195 	if (cft->file_offset) {
4196 		struct cgroup_file *cfile = (void *)css + cft->file_offset;
4197 
4198 		timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4199 
4200 		spin_lock_irq(&cgroup_file_kn_lock);
4201 		cfile->kn = kn;
4202 		spin_unlock_irq(&cgroup_file_kn_lock);
4203 	}
4204 
4205 	return 0;
4206 }
4207 
4208 /**
4209  * cgroup_addrm_files - add or remove files to a cgroup directory
4210  * @css: the target css
4211  * @cgrp: the target cgroup (usually css->cgroup)
4212  * @cfts: array of cftypes to be added
4213  * @is_add: whether to add or remove
4214  *
4215  * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4216  * For removals, this function never fails.
4217  */
cgroup_addrm_files(struct cgroup_subsys_state * css,struct cgroup * cgrp,struct cftype cfts[],bool is_add)4218 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4219 			      struct cgroup *cgrp, struct cftype cfts[],
4220 			      bool is_add)
4221 {
4222 	struct cftype *cft, *cft_end = NULL;
4223 	int ret = 0;
4224 
4225 	lockdep_assert_held(&cgroup_mutex);
4226 
4227 restart:
4228 	for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4229 		/* does cft->flags tell us to skip this file on @cgrp? */
4230 		if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4231 			continue;
4232 		if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4233 			continue;
4234 		if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4235 			continue;
4236 		if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4237 			continue;
4238 		if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4239 			continue;
4240 		if (is_add) {
4241 			ret = cgroup_add_file(css, cgrp, cft);
4242 			if (ret) {
4243 				pr_warn("%s: failed to add %s, err=%d\n",
4244 					__func__, cft->name, ret);
4245 				cft_end = cft;
4246 				is_add = false;
4247 				goto restart;
4248 			}
4249 		} else {
4250 			cgroup_rm_file(cgrp, cft);
4251 		}
4252 	}
4253 	return ret;
4254 }
4255 
cgroup_apply_cftypes(struct cftype * cfts,bool is_add)4256 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4257 {
4258 	struct cgroup_subsys *ss = cfts[0].ss;
4259 	struct cgroup *root = &ss->root->cgrp;
4260 	struct cgroup_subsys_state *css;
4261 	int ret = 0;
4262 
4263 	lockdep_assert_held(&cgroup_mutex);
4264 
4265 	/* add/rm files for all cgroups created before */
4266 	css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4267 		struct cgroup *cgrp = css->cgroup;
4268 
4269 		if (!(css->flags & CSS_VISIBLE))
4270 			continue;
4271 
4272 		ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4273 		if (ret)
4274 			break;
4275 	}
4276 
4277 	if (is_add && !ret)
4278 		kernfs_activate(root->kn);
4279 	return ret;
4280 }
4281 
cgroup_exit_cftypes(struct cftype * cfts)4282 static void cgroup_exit_cftypes(struct cftype *cfts)
4283 {
4284 	struct cftype *cft;
4285 
4286 	for (cft = cfts; cft->name[0] != '\0'; cft++) {
4287 		/* free copy for custom atomic_write_len, see init_cftypes() */
4288 		if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4289 			kfree(cft->kf_ops);
4290 		cft->kf_ops = NULL;
4291 		cft->ss = NULL;
4292 
4293 		/* revert flags set by cgroup core while adding @cfts */
4294 		cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4295 				__CFTYPE_ADDED);
4296 	}
4297 }
4298 
cgroup_init_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4299 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4300 {
4301 	struct cftype *cft;
4302 	int ret = 0;
4303 
4304 	for (cft = cfts; cft->name[0] != '\0'; cft++) {
4305 		struct kernfs_ops *kf_ops;
4306 
4307 		WARN_ON(cft->ss || cft->kf_ops);
4308 
4309 		if (cft->flags & __CFTYPE_ADDED) {
4310 			ret = -EBUSY;
4311 			break;
4312 		}
4313 
4314 		if (cft->seq_start)
4315 			kf_ops = &cgroup_kf_ops;
4316 		else
4317 			kf_ops = &cgroup_kf_single_ops;
4318 
4319 		/*
4320 		 * Ugh... if @cft wants a custom max_write_len, we need to
4321 		 * make a copy of kf_ops to set its atomic_write_len.
4322 		 */
4323 		if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4324 			kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4325 			if (!kf_ops) {
4326 				ret = -ENOMEM;
4327 				break;
4328 			}
4329 			kf_ops->atomic_write_len = cft->max_write_len;
4330 		}
4331 
4332 		cft->kf_ops = kf_ops;
4333 		cft->ss = ss;
4334 		cft->flags |= __CFTYPE_ADDED;
4335 	}
4336 
4337 	if (ret)
4338 		cgroup_exit_cftypes(cfts);
4339 	return ret;
4340 }
4341 
cgroup_rm_cftypes_locked(struct cftype * cfts)4342 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4343 {
4344 	lockdep_assert_held(&cgroup_mutex);
4345 
4346 	list_del(&cfts->node);
4347 	cgroup_apply_cftypes(cfts, false);
4348 	cgroup_exit_cftypes(cfts);
4349 	return 0;
4350 }
4351 
4352 /**
4353  * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4354  * @cfts: zero-length name terminated array of cftypes
4355  *
4356  * Unregister @cfts.  Files described by @cfts are removed from all
4357  * existing cgroups and all future cgroups won't have them either.  This
4358  * function can be called anytime whether @cfts' subsys is attached or not.
4359  *
4360  * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4361  * registered.
4362  */
cgroup_rm_cftypes(struct cftype * cfts)4363 int cgroup_rm_cftypes(struct cftype *cfts)
4364 {
4365 	int ret;
4366 
4367 	if (!cfts || cfts[0].name[0] == '\0')
4368 		return 0;
4369 
4370 	if (!(cfts[0].flags & __CFTYPE_ADDED))
4371 		return -ENOENT;
4372 
4373 	mutex_lock(&cgroup_mutex);
4374 	ret = cgroup_rm_cftypes_locked(cfts);
4375 	mutex_unlock(&cgroup_mutex);
4376 	return ret;
4377 }
4378 
4379 /**
4380  * cgroup_add_cftypes - add an array of cftypes to a subsystem
4381  * @ss: target cgroup subsystem
4382  * @cfts: zero-length name terminated array of cftypes
4383  *
4384  * Register @cfts to @ss.  Files described by @cfts are created for all
4385  * existing cgroups to which @ss is attached and all future cgroups will
4386  * have them too.  This function can be called anytime whether @ss is
4387  * attached or not.
4388  *
4389  * Returns 0 on successful registration, -errno on failure.  Note that this
4390  * function currently returns 0 as long as @cfts registration is successful
4391  * even if some file creation attempts on existing cgroups fail.
4392  */
cgroup_add_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4393 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4394 {
4395 	int ret;
4396 
4397 	if (!cgroup_ssid_enabled(ss->id))
4398 		return 0;
4399 
4400 	if (!cfts || cfts[0].name[0] == '\0')
4401 		return 0;
4402 
4403 	ret = cgroup_init_cftypes(ss, cfts);
4404 	if (ret)
4405 		return ret;
4406 
4407 	mutex_lock(&cgroup_mutex);
4408 
4409 	list_add_tail(&cfts->node, &ss->cfts);
4410 	ret = cgroup_apply_cftypes(cfts, true);
4411 	if (ret)
4412 		cgroup_rm_cftypes_locked(cfts);
4413 
4414 	mutex_unlock(&cgroup_mutex);
4415 	return ret;
4416 }
4417 
4418 /**
4419  * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4420  * @ss: target cgroup subsystem
4421  * @cfts: zero-length name terminated array of cftypes
4422  *
4423  * Similar to cgroup_add_cftypes() but the added files are only used for
4424  * the default hierarchy.
4425  */
cgroup_add_dfl_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4426 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4427 {
4428 	struct cftype *cft;
4429 
4430 	for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4431 		cft->flags |= __CFTYPE_ONLY_ON_DFL;
4432 	return cgroup_add_cftypes(ss, cfts);
4433 }
4434 
4435 /**
4436  * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4437  * @ss: target cgroup subsystem
4438  * @cfts: zero-length name terminated array of cftypes
4439  *
4440  * Similar to cgroup_add_cftypes() but the added files are only used for
4441  * the legacy hierarchies.
4442  */
cgroup_add_legacy_cftypes(struct cgroup_subsys * ss,struct cftype * cfts)4443 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4444 {
4445 	struct cftype *cft;
4446 
4447 	for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4448 		cft->flags |= __CFTYPE_NOT_ON_DFL;
4449 	return cgroup_add_cftypes(ss, cfts);
4450 }
4451 
4452 /**
4453  * cgroup_file_notify - generate a file modified event for a cgroup_file
4454  * @cfile: target cgroup_file
4455  *
4456  * @cfile must have been obtained by setting cftype->file_offset.
4457  */
cgroup_file_notify(struct cgroup_file * cfile)4458 void cgroup_file_notify(struct cgroup_file *cfile)
4459 {
4460 	unsigned long flags;
4461 
4462 	spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4463 	if (cfile->kn) {
4464 		unsigned long last = cfile->notified_at;
4465 		unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4466 
4467 		if (time_in_range(jiffies, last, next)) {
4468 			timer_reduce(&cfile->notify_timer, next);
4469 		} else {
4470 			kernfs_notify(cfile->kn);
4471 			cfile->notified_at = jiffies;
4472 		}
4473 	}
4474 	spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4475 }
4476 
4477 /**
4478  * cgroup_file_show - show or hide a hidden cgroup file
4479  * @cfile: target cgroup_file obtained by setting cftype->file_offset
4480  * @show: whether to show or hide
4481  */
cgroup_file_show(struct cgroup_file * cfile,bool show)4482 void cgroup_file_show(struct cgroup_file *cfile, bool show)
4483 {
4484 	struct kernfs_node *kn;
4485 
4486 	spin_lock_irq(&cgroup_file_kn_lock);
4487 	kn = cfile->kn;
4488 	kernfs_get(kn);
4489 	spin_unlock_irq(&cgroup_file_kn_lock);
4490 
4491 	if (kn)
4492 		kernfs_show(kn, show);
4493 
4494 	kernfs_put(kn);
4495 }
4496 
4497 /**
4498  * css_next_child - find the next child of a given css
4499  * @pos: the current position (%NULL to initiate traversal)
4500  * @parent: css whose children to walk
4501  *
4502  * This function returns the next child of @parent and should be called
4503  * under either cgroup_mutex or RCU read lock.  The only requirement is
4504  * that @parent and @pos are accessible.  The next sibling is guaranteed to
4505  * be returned regardless of their states.
4506  *
4507  * If a subsystem synchronizes ->css_online() and the start of iteration, a
4508  * css which finished ->css_online() is guaranteed to be visible in the
4509  * future iterations and will stay visible until the last reference is put.
4510  * A css which hasn't finished ->css_online() or already finished
4511  * ->css_offline() may show up during traversal.  It's each subsystem's
4512  * responsibility to synchronize against on/offlining.
4513  */
css_next_child(struct cgroup_subsys_state * pos,struct cgroup_subsys_state * parent)4514 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4515 					   struct cgroup_subsys_state *parent)
4516 {
4517 	struct cgroup_subsys_state *next;
4518 
4519 	cgroup_assert_mutex_or_rcu_locked();
4520 
4521 	/*
4522 	 * @pos could already have been unlinked from the sibling list.
4523 	 * Once a cgroup is removed, its ->sibling.next is no longer
4524 	 * updated when its next sibling changes.  CSS_RELEASED is set when
4525 	 * @pos is taken off list, at which time its next pointer is valid,
4526 	 * and, as releases are serialized, the one pointed to by the next
4527 	 * pointer is guaranteed to not have started release yet.  This
4528 	 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4529 	 * critical section, the one pointed to by its next pointer is
4530 	 * guaranteed to not have finished its RCU grace period even if we
4531 	 * have dropped rcu_read_lock() in-between iterations.
4532 	 *
4533 	 * If @pos has CSS_RELEASED set, its next pointer can't be
4534 	 * dereferenced; however, as each css is given a monotonically
4535 	 * increasing unique serial number and always appended to the
4536 	 * sibling list, the next one can be found by walking the parent's
4537 	 * children until the first css with higher serial number than
4538 	 * @pos's.  While this path can be slower, it happens iff iteration
4539 	 * races against release and the race window is very small.
4540 	 */
4541 	if (!pos) {
4542 		next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4543 	} else if (likely(!(pos->flags & CSS_RELEASED))) {
4544 		next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4545 	} else {
4546 		list_for_each_entry_rcu(next, &parent->children, sibling,
4547 					lockdep_is_held(&cgroup_mutex))
4548 			if (next->serial_nr > pos->serial_nr)
4549 				break;
4550 	}
4551 
4552 	/*
4553 	 * @next, if not pointing to the head, can be dereferenced and is
4554 	 * the next sibling.
4555 	 */
4556 	if (&next->sibling != &parent->children)
4557 		return next;
4558 	return NULL;
4559 }
4560 
4561 /**
4562  * css_next_descendant_pre - find the next descendant for pre-order walk
4563  * @pos: the current position (%NULL to initiate traversal)
4564  * @root: css whose descendants to walk
4565  *
4566  * To be used by css_for_each_descendant_pre().  Find the next descendant
4567  * to visit for pre-order traversal of @root's descendants.  @root is
4568  * included in the iteration and the first node to be visited.
4569  *
4570  * While this function requires cgroup_mutex or RCU read locking, it
4571  * doesn't require the whole traversal to be contained in a single critical
4572  * section.  This function will return the correct next descendant as long
4573  * as both @pos and @root are accessible and @pos is a descendant of @root.
4574  *
4575  * If a subsystem synchronizes ->css_online() and the start of iteration, a
4576  * css which finished ->css_online() is guaranteed to be visible in the
4577  * future iterations and will stay visible until the last reference is put.
4578  * A css which hasn't finished ->css_online() or already finished
4579  * ->css_offline() may show up during traversal.  It's each subsystem's
4580  * responsibility to synchronize against on/offlining.
4581  */
4582 struct cgroup_subsys_state *
css_next_descendant_pre(struct cgroup_subsys_state * pos,struct cgroup_subsys_state * root)4583 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4584 			struct cgroup_subsys_state *root)
4585 {
4586 	struct cgroup_subsys_state *next;
4587 
4588 	cgroup_assert_mutex_or_rcu_locked();
4589 
4590 	/* if first iteration, visit @root */
4591 	if (!pos)
4592 		return root;
4593 
4594 	/* visit the first child if exists */
4595 	next = css_next_child(NULL, pos);
4596 	if (next)
4597 		return next;
4598 
4599 	/* no child, visit my or the closest ancestor's next sibling */
4600 	while (pos != root) {
4601 		next = css_next_child(pos, pos->parent);
4602 		if (next)
4603 			return next;
4604 		pos = pos->parent;
4605 	}
4606 
4607 	return NULL;
4608 }
4609 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4610 
4611 /**
4612  * css_rightmost_descendant - return the rightmost descendant of a css
4613  * @pos: css of interest
4614  *
4615  * Return the rightmost descendant of @pos.  If there's no descendant, @pos
4616  * is returned.  This can be used during pre-order traversal to skip
4617  * subtree of @pos.
4618  *
4619  * While this function requires cgroup_mutex or RCU read locking, it
4620  * doesn't require the whole traversal to be contained in a single critical
4621  * section.  This function will return the correct rightmost descendant as
4622  * long as @pos is accessible.
4623  */
4624 struct cgroup_subsys_state *
css_rightmost_descendant(struct cgroup_subsys_state * pos)4625 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4626 {
4627 	struct cgroup_subsys_state *last, *tmp;
4628 
4629 	cgroup_assert_mutex_or_rcu_locked();
4630 
4631 	do {
4632 		last = pos;
4633 		/* ->prev isn't RCU safe, walk ->next till the end */
4634 		pos = NULL;
4635 		css_for_each_child(tmp, last)
4636 			pos = tmp;
4637 	} while (pos);
4638 
4639 	return last;
4640 }
4641 
4642 static struct cgroup_subsys_state *
css_leftmost_descendant(struct cgroup_subsys_state * pos)4643 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4644 {
4645 	struct cgroup_subsys_state *last;
4646 
4647 	do {
4648 		last = pos;
4649 		pos = css_next_child(NULL, pos);
4650 	} while (pos);
4651 
4652 	return last;
4653 }
4654 
4655 /**
4656  * css_next_descendant_post - find the next descendant for post-order walk
4657  * @pos: the current position (%NULL to initiate traversal)
4658  * @root: css whose descendants to walk
4659  *
4660  * To be used by css_for_each_descendant_post().  Find the next descendant
4661  * to visit for post-order traversal of @root's descendants.  @root is
4662  * included in the iteration and the last node to be visited.
4663  *
4664  * While this function requires cgroup_mutex or RCU read locking, it
4665  * doesn't require the whole traversal to be contained in a single critical
4666  * section.  This function will return the correct next descendant as long
4667  * as both @pos and @cgroup are accessible and @pos is a descendant of
4668  * @cgroup.
4669  *
4670  * If a subsystem synchronizes ->css_online() and the start of iteration, a
4671  * css which finished ->css_online() is guaranteed to be visible in the
4672  * future iterations and will stay visible until the last reference is put.
4673  * A css which hasn't finished ->css_online() or already finished
4674  * ->css_offline() may show up during traversal.  It's each subsystem's
4675  * responsibility to synchronize against on/offlining.
4676  */
4677 struct cgroup_subsys_state *
css_next_descendant_post(struct cgroup_subsys_state * pos,struct cgroup_subsys_state * root)4678 css_next_descendant_post(struct cgroup_subsys_state *pos,
4679 			 struct cgroup_subsys_state *root)
4680 {
4681 	struct cgroup_subsys_state *next;
4682 
4683 	cgroup_assert_mutex_or_rcu_locked();
4684 
4685 	/* if first iteration, visit leftmost descendant which may be @root */
4686 	if (!pos)
4687 		return css_leftmost_descendant(root);
4688 
4689 	/* if we visited @root, we're done */
4690 	if (pos == root)
4691 		return NULL;
4692 
4693 	/* if there's an unvisited sibling, visit its leftmost descendant */
4694 	next = css_next_child(pos, pos->parent);
4695 	if (next)
4696 		return css_leftmost_descendant(next);
4697 
4698 	/* no sibling left, visit parent */
4699 	return pos->parent;
4700 }
4701 
4702 /**
4703  * css_has_online_children - does a css have online children
4704  * @css: the target css
4705  *
4706  * Returns %true if @css has any online children; otherwise, %false.  This
4707  * function can be called from any context but the caller is responsible
4708  * for synchronizing against on/offlining as necessary.
4709  */
css_has_online_children(struct cgroup_subsys_state * css)4710 bool css_has_online_children(struct cgroup_subsys_state *css)
4711 {
4712 	struct cgroup_subsys_state *child;
4713 	bool ret = false;
4714 
4715 	rcu_read_lock();
4716 	css_for_each_child(child, css) {
4717 		if (child->flags & CSS_ONLINE) {
4718 			ret = true;
4719 			break;
4720 		}
4721 	}
4722 	rcu_read_unlock();
4723 	return ret;
4724 }
4725 
css_task_iter_next_css_set(struct css_task_iter * it)4726 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4727 {
4728 	struct list_head *l;
4729 	struct cgrp_cset_link *link;
4730 	struct css_set *cset;
4731 
4732 	lockdep_assert_held(&css_set_lock);
4733 
4734 	/* find the next threaded cset */
4735 	if (it->tcset_pos) {
4736 		l = it->tcset_pos->next;
4737 
4738 		if (l != it->tcset_head) {
4739 			it->tcset_pos = l;
4740 			return container_of(l, struct css_set,
4741 					    threaded_csets_node);
4742 		}
4743 
4744 		it->tcset_pos = NULL;
4745 	}
4746 
4747 	/* find the next cset */
4748 	l = it->cset_pos;
4749 	l = l->next;
4750 	if (l == it->cset_head) {
4751 		it->cset_pos = NULL;
4752 		return NULL;
4753 	}
4754 
4755 	if (it->ss) {
4756 		cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4757 	} else {
4758 		link = list_entry(l, struct cgrp_cset_link, cset_link);
4759 		cset = link->cset;
4760 	}
4761 
4762 	it->cset_pos = l;
4763 
4764 	/* initialize threaded css_set walking */
4765 	if (it->flags & CSS_TASK_ITER_THREADED) {
4766 		if (it->cur_dcset)
4767 			put_css_set_locked(it->cur_dcset);
4768 		it->cur_dcset = cset;
4769 		get_css_set(cset);
4770 
4771 		it->tcset_head = &cset->threaded_csets;
4772 		it->tcset_pos = &cset->threaded_csets;
4773 	}
4774 
4775 	return cset;
4776 }
4777 
4778 /**
4779  * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4780  * @it: the iterator to advance
4781  *
4782  * Advance @it to the next css_set to walk.
4783  */
css_task_iter_advance_css_set(struct css_task_iter * it)4784 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4785 {
4786 	struct css_set *cset;
4787 
4788 	lockdep_assert_held(&css_set_lock);
4789 
4790 	/* Advance to the next non-empty css_set and find first non-empty tasks list*/
4791 	while ((cset = css_task_iter_next_css_set(it))) {
4792 		if (!list_empty(&cset->tasks)) {
4793 			it->cur_tasks_head = &cset->tasks;
4794 			break;
4795 		} else if (!list_empty(&cset->mg_tasks)) {
4796 			it->cur_tasks_head = &cset->mg_tasks;
4797 			break;
4798 		} else if (!list_empty(&cset->dying_tasks)) {
4799 			it->cur_tasks_head = &cset->dying_tasks;
4800 			break;
4801 		}
4802 	}
4803 	if (!cset) {
4804 		it->task_pos = NULL;
4805 		return;
4806 	}
4807 	it->task_pos = it->cur_tasks_head->next;
4808 
4809 	/*
4810 	 * We don't keep css_sets locked across iteration steps and thus
4811 	 * need to take steps to ensure that iteration can be resumed after
4812 	 * the lock is re-acquired.  Iteration is performed at two levels -
4813 	 * css_sets and tasks in them.
4814 	 *
4815 	 * Once created, a css_set never leaves its cgroup lists, so a
4816 	 * pinned css_set is guaranteed to stay put and we can resume
4817 	 * iteration afterwards.
4818 	 *
4819 	 * Tasks may leave @cset across iteration steps.  This is resolved
4820 	 * by registering each iterator with the css_set currently being
4821 	 * walked and making css_set_move_task() advance iterators whose
4822 	 * next task is leaving.
4823 	 */
4824 	if (it->cur_cset) {
4825 		list_del(&it->iters_node);
4826 		put_css_set_locked(it->cur_cset);
4827 	}
4828 	get_css_set(cset);
4829 	it->cur_cset = cset;
4830 	list_add(&it->iters_node, &cset->task_iters);
4831 }
4832 
css_task_iter_skip(struct css_task_iter * it,struct task_struct * task)4833 static void css_task_iter_skip(struct css_task_iter *it,
4834 			       struct task_struct *task)
4835 {
4836 	lockdep_assert_held(&css_set_lock);
4837 
4838 	if (it->task_pos == &task->cg_list) {
4839 		it->task_pos = it->task_pos->next;
4840 		it->flags |= CSS_TASK_ITER_SKIPPED;
4841 	}
4842 }
4843 
css_task_iter_advance(struct css_task_iter * it)4844 static void css_task_iter_advance(struct css_task_iter *it)
4845 {
4846 	struct task_struct *task;
4847 
4848 	lockdep_assert_held(&css_set_lock);
4849 repeat:
4850 	if (it->task_pos) {
4851 		/*
4852 		 * Advance iterator to find next entry. We go through cset
4853 		 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4854 		 * the next cset.
4855 		 */
4856 		if (it->flags & CSS_TASK_ITER_SKIPPED)
4857 			it->flags &= ~CSS_TASK_ITER_SKIPPED;
4858 		else
4859 			it->task_pos = it->task_pos->next;
4860 
4861 		if (it->task_pos == &it->cur_cset->tasks) {
4862 			it->cur_tasks_head = &it->cur_cset->mg_tasks;
4863 			it->task_pos = it->cur_tasks_head->next;
4864 		}
4865 		if (it->task_pos == &it->cur_cset->mg_tasks) {
4866 			it->cur_tasks_head = &it->cur_cset->dying_tasks;
4867 			it->task_pos = it->cur_tasks_head->next;
4868 		}
4869 		if (it->task_pos == &it->cur_cset->dying_tasks)
4870 			css_task_iter_advance_css_set(it);
4871 	} else {
4872 		/* called from start, proceed to the first cset */
4873 		css_task_iter_advance_css_set(it);
4874 	}
4875 
4876 	if (!it->task_pos)
4877 		return;
4878 
4879 	task = list_entry(it->task_pos, struct task_struct, cg_list);
4880 
4881 	if (it->flags & CSS_TASK_ITER_PROCS) {
4882 		/* if PROCS, skip over tasks which aren't group leaders */
4883 		if (!thread_group_leader(task))
4884 			goto repeat;
4885 
4886 		/* and dying leaders w/o live member threads */
4887 		if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4888 		    !atomic_read(&task->signal->live))
4889 			goto repeat;
4890 	} else {
4891 		/* skip all dying ones */
4892 		if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4893 			goto repeat;
4894 	}
4895 }
4896 
4897 /**
4898  * css_task_iter_start - initiate task iteration
4899  * @css: the css to walk tasks of
4900  * @flags: CSS_TASK_ITER_* flags
4901  * @it: the task iterator to use
4902  *
4903  * Initiate iteration through the tasks of @css.  The caller can call
4904  * css_task_iter_next() to walk through the tasks until the function
4905  * returns NULL.  On completion of iteration, css_task_iter_end() must be
4906  * called.
4907  */
css_task_iter_start(struct cgroup_subsys_state * css,unsigned int flags,struct css_task_iter * it)4908 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4909 			 struct css_task_iter *it)
4910 {
4911 	memset(it, 0, sizeof(*it));
4912 
4913 	spin_lock_irq(&css_set_lock);
4914 
4915 	it->ss = css->ss;
4916 	it->flags = flags;
4917 
4918 	if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4919 		it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4920 	else
4921 		it->cset_pos = &css->cgroup->cset_links;
4922 
4923 	it->cset_head = it->cset_pos;
4924 
4925 	css_task_iter_advance(it);
4926 
4927 	spin_unlock_irq(&css_set_lock);
4928 }
4929 
4930 /**
4931  * css_task_iter_next - return the next task for the iterator
4932  * @it: the task iterator being iterated
4933  *
4934  * The "next" function for task iteration.  @it should have been
4935  * initialized via css_task_iter_start().  Returns NULL when the iteration
4936  * reaches the end.
4937  */
css_task_iter_next(struct css_task_iter * it)4938 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4939 {
4940 	if (it->cur_task) {
4941 		put_task_struct(it->cur_task);
4942 		it->cur_task = NULL;
4943 	}
4944 
4945 	spin_lock_irq(&css_set_lock);
4946 
4947 	/* @it may be half-advanced by skips, finish advancing */
4948 	if (it->flags & CSS_TASK_ITER_SKIPPED)
4949 		css_task_iter_advance(it);
4950 
4951 	if (it->task_pos) {
4952 		it->cur_task = list_entry(it->task_pos, struct task_struct,
4953 					  cg_list);
4954 		get_task_struct(it->cur_task);
4955 		css_task_iter_advance(it);
4956 	}
4957 
4958 	spin_unlock_irq(&css_set_lock);
4959 
4960 	return it->cur_task;
4961 }
4962 
4963 /**
4964  * css_task_iter_end - finish task iteration
4965  * @it: the task iterator to finish
4966  *
4967  * Finish task iteration started by css_task_iter_start().
4968  */
css_task_iter_end(struct css_task_iter * it)4969 void css_task_iter_end(struct css_task_iter *it)
4970 {
4971 	if (it->cur_cset) {
4972 		spin_lock_irq(&css_set_lock);
4973 		list_del(&it->iters_node);
4974 		put_css_set_locked(it->cur_cset);
4975 		spin_unlock_irq(&css_set_lock);
4976 	}
4977 
4978 	if (it->cur_dcset)
4979 		put_css_set(it->cur_dcset);
4980 
4981 	if (it->cur_task)
4982 		put_task_struct(it->cur_task);
4983 }
4984 
cgroup_procs_release(struct kernfs_open_file * of)4985 static void cgroup_procs_release(struct kernfs_open_file *of)
4986 {
4987 	struct cgroup_file_ctx *ctx = of->priv;
4988 
4989 	if (ctx->procs.started)
4990 		css_task_iter_end(&ctx->procs.iter);
4991 }
4992 
cgroup_procs_next(struct seq_file * s,void * v,loff_t * pos)4993 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4994 {
4995 	struct kernfs_open_file *of = s->private;
4996 	struct cgroup_file_ctx *ctx = of->priv;
4997 
4998 	if (pos)
4999 		(*pos)++;
5000 
5001 	return css_task_iter_next(&ctx->procs.iter);
5002 }
5003 
__cgroup_procs_start(struct seq_file * s,loff_t * pos,unsigned int iter_flags)5004 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5005 				  unsigned int iter_flags)
5006 {
5007 	struct kernfs_open_file *of = s->private;
5008 	struct cgroup *cgrp = seq_css(s)->cgroup;
5009 	struct cgroup_file_ctx *ctx = of->priv;
5010 	struct css_task_iter *it = &ctx->procs.iter;
5011 
5012 	/*
5013 	 * When a seq_file is seeked, it's always traversed sequentially
5014 	 * from position 0, so we can simply keep iterating on !0 *pos.
5015 	 */
5016 	if (!ctx->procs.started) {
5017 		if (WARN_ON_ONCE((*pos)))
5018 			return ERR_PTR(-EINVAL);
5019 		css_task_iter_start(&cgrp->self, iter_flags, it);
5020 		ctx->procs.started = true;
5021 	} else if (!(*pos)) {
5022 		css_task_iter_end(it);
5023 		css_task_iter_start(&cgrp->self, iter_flags, it);
5024 	} else
5025 		return it->cur_task;
5026 
5027 	return cgroup_procs_next(s, NULL, NULL);
5028 }
5029 
cgroup_procs_start(struct seq_file * s,loff_t * pos)5030 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5031 {
5032 	struct cgroup *cgrp = seq_css(s)->cgroup;
5033 
5034 	/*
5035 	 * All processes of a threaded subtree belong to the domain cgroup
5036 	 * of the subtree.  Only threads can be distributed across the
5037 	 * subtree.  Reject reads on cgroup.procs in the subtree proper.
5038 	 * They're always empty anyway.
5039 	 */
5040 	if (cgroup_is_threaded(cgrp))
5041 		return ERR_PTR(-EOPNOTSUPP);
5042 
5043 	return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5044 					    CSS_TASK_ITER_THREADED);
5045 }
5046 
cgroup_procs_show(struct seq_file * s,void * v)5047 static int cgroup_procs_show(struct seq_file *s, void *v)
5048 {
5049 	seq_printf(s, "%d\n", task_pid_vnr(v));
5050 	return 0;
5051 }
5052 
cgroup_may_write(const struct cgroup * cgrp,struct super_block * sb)5053 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5054 {
5055 	int ret;
5056 	struct inode *inode;
5057 
5058 	lockdep_assert_held(&cgroup_mutex);
5059 
5060 	inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5061 	if (!inode)
5062 		return -ENOMEM;
5063 
5064 	ret = inode_permission(&init_user_ns, inode, MAY_WRITE);
5065 	iput(inode);
5066 	return ret;
5067 }
5068 
cgroup_procs_write_permission(struct cgroup * src_cgrp,struct cgroup * dst_cgrp,struct super_block * sb,struct cgroup_namespace * ns)5069 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5070 					 struct cgroup *dst_cgrp,
5071 					 struct super_block *sb,
5072 					 struct cgroup_namespace *ns)
5073 {
5074 	struct cgroup *com_cgrp = src_cgrp;
5075 	int ret;
5076 
5077 	lockdep_assert_held(&cgroup_mutex);
5078 
5079 	/* find the common ancestor */
5080 	while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5081 		com_cgrp = cgroup_parent(com_cgrp);
5082 
5083 	/* %current should be authorized to migrate to the common ancestor */
5084 	ret = cgroup_may_write(com_cgrp, sb);
5085 	if (ret)
5086 		return ret;
5087 
5088 	/*
5089 	 * If namespaces are delegation boundaries, %current must be able
5090 	 * to see both source and destination cgroups from its namespace.
5091 	 */
5092 	if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5093 	    (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5094 	     !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5095 		return -ENOENT;
5096 
5097 	return 0;
5098 }
5099 
cgroup_attach_permissions(struct cgroup * src_cgrp,struct cgroup * dst_cgrp,struct super_block * sb,bool threadgroup,struct cgroup_namespace * ns)5100 static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5101 				     struct cgroup *dst_cgrp,
5102 				     struct super_block *sb, bool threadgroup,
5103 				     struct cgroup_namespace *ns)
5104 {
5105 	int ret = 0;
5106 
5107 	ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5108 	if (ret)
5109 		return ret;
5110 
5111 	ret = cgroup_migrate_vet_dst(dst_cgrp);
5112 	if (ret)
5113 		return ret;
5114 
5115 	if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5116 		ret = -EOPNOTSUPP;
5117 
5118 	return ret;
5119 }
5120 
__cgroup_procs_write(struct kernfs_open_file * of,char * buf,bool threadgroup)5121 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5122 				    bool threadgroup)
5123 {
5124 	struct cgroup_file_ctx *ctx = of->priv;
5125 	struct cgroup *src_cgrp, *dst_cgrp;
5126 	struct task_struct *task;
5127 	const struct cred *saved_cred;
5128 	ssize_t ret;
5129 	bool threadgroup_locked;
5130 
5131 	dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5132 	if (!dst_cgrp)
5133 		return -ENODEV;
5134 
5135 	task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
5136 	ret = PTR_ERR_OR_ZERO(task);
5137 	if (ret)
5138 		goto out_unlock;
5139 
5140 	/* find the source cgroup */
5141 	spin_lock_irq(&css_set_lock);
5142 	src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5143 	spin_unlock_irq(&css_set_lock);
5144 
5145 	/*
5146 	 * Process and thread migrations follow same delegation rule. Check
5147 	 * permissions using the credentials from file open to protect against
5148 	 * inherited fd attacks.
5149 	 */
5150 	saved_cred = override_creds(of->file->f_cred);
5151 	ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5152 					of->file->f_path.dentry->d_sb,
5153 					threadgroup, ctx->ns);
5154 	revert_creds(saved_cred);
5155 	if (ret)
5156 		goto out_finish;
5157 
5158 	ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5159 
5160 out_finish:
5161 	cgroup_procs_write_finish(task, threadgroup_locked);
5162 out_unlock:
5163 	cgroup_kn_unlock(of->kn);
5164 
5165 	return ret;
5166 }
5167 
cgroup_procs_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)5168 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5169 				  char *buf, size_t nbytes, loff_t off)
5170 {
5171 	return __cgroup_procs_write(of, buf, true) ?: nbytes;
5172 }
5173 
cgroup_threads_start(struct seq_file * s,loff_t * pos)5174 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5175 {
5176 	return __cgroup_procs_start(s, pos, 0);
5177 }
5178 
cgroup_threads_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)5179 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5180 				    char *buf, size_t nbytes, loff_t off)
5181 {
5182 	return __cgroup_procs_write(of, buf, false) ?: nbytes;
5183 }
5184 
5185 /* cgroup core interface files for the default hierarchy */
5186 static struct cftype cgroup_base_files[] = {
5187 	{
5188 		.name = "cgroup.type",
5189 		.flags = CFTYPE_NOT_ON_ROOT,
5190 		.seq_show = cgroup_type_show,
5191 		.write = cgroup_type_write,
5192 	},
5193 	{
5194 		.name = "cgroup.procs",
5195 		.flags = CFTYPE_NS_DELEGATABLE,
5196 		.file_offset = offsetof(struct cgroup, procs_file),
5197 		.release = cgroup_procs_release,
5198 		.seq_start = cgroup_procs_start,
5199 		.seq_next = cgroup_procs_next,
5200 		.seq_show = cgroup_procs_show,
5201 		.write = cgroup_procs_write,
5202 	},
5203 	{
5204 		.name = "cgroup.threads",
5205 		.flags = CFTYPE_NS_DELEGATABLE,
5206 		.release = cgroup_procs_release,
5207 		.seq_start = cgroup_threads_start,
5208 		.seq_next = cgroup_procs_next,
5209 		.seq_show = cgroup_procs_show,
5210 		.write = cgroup_threads_write,
5211 	},
5212 	{
5213 		.name = "cgroup.controllers",
5214 		.seq_show = cgroup_controllers_show,
5215 	},
5216 	{
5217 		.name = "cgroup.subtree_control",
5218 		.flags = CFTYPE_NS_DELEGATABLE,
5219 		.seq_show = cgroup_subtree_control_show,
5220 		.write = cgroup_subtree_control_write,
5221 	},
5222 	{
5223 		.name = "cgroup.events",
5224 		.flags = CFTYPE_NOT_ON_ROOT,
5225 		.file_offset = offsetof(struct cgroup, events_file),
5226 		.seq_show = cgroup_events_show,
5227 	},
5228 	{
5229 		.name = "cgroup.max.descendants",
5230 		.seq_show = cgroup_max_descendants_show,
5231 		.write = cgroup_max_descendants_write,
5232 	},
5233 	{
5234 		.name = "cgroup.max.depth",
5235 		.seq_show = cgroup_max_depth_show,
5236 		.write = cgroup_max_depth_write,
5237 	},
5238 	{
5239 		.name = "cgroup.stat",
5240 		.seq_show = cgroup_stat_show,
5241 	},
5242 	{
5243 		.name = "cgroup.freeze",
5244 		.flags = CFTYPE_NOT_ON_ROOT,
5245 		.seq_show = cgroup_freeze_show,
5246 		.write = cgroup_freeze_write,
5247 	},
5248 	{
5249 		.name = "cgroup.kill",
5250 		.flags = CFTYPE_NOT_ON_ROOT,
5251 		.write = cgroup_kill_write,
5252 	},
5253 	{
5254 		.name = "cpu.stat",
5255 		.seq_show = cpu_stat_show,
5256 	},
5257 	{ }	/* terminate */
5258 };
5259 
5260 static struct cftype cgroup_psi_files[] = {
5261 #ifdef CONFIG_PSI
5262 	{
5263 		.name = "io.pressure",
5264 		.file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5265 		.seq_show = cgroup_io_pressure_show,
5266 		.write = cgroup_io_pressure_write,
5267 		.poll = cgroup_pressure_poll,
5268 		.release = cgroup_pressure_release,
5269 	},
5270 	{
5271 		.name = "memory.pressure",
5272 		.file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5273 		.seq_show = cgroup_memory_pressure_show,
5274 		.write = cgroup_memory_pressure_write,
5275 		.poll = cgroup_pressure_poll,
5276 		.release = cgroup_pressure_release,
5277 	},
5278 	{
5279 		.name = "cpu.pressure",
5280 		.file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5281 		.seq_show = cgroup_cpu_pressure_show,
5282 		.write = cgroup_cpu_pressure_write,
5283 		.poll = cgroup_pressure_poll,
5284 		.release = cgroup_pressure_release,
5285 	},
5286 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
5287 	{
5288 		.name = "irq.pressure",
5289 		.file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5290 		.seq_show = cgroup_irq_pressure_show,
5291 		.write = cgroup_irq_pressure_write,
5292 		.poll = cgroup_pressure_poll,
5293 		.release = cgroup_pressure_release,
5294 	},
5295 #endif
5296 	{
5297 		.name = "cgroup.pressure",
5298 		.seq_show = cgroup_pressure_show,
5299 		.write = cgroup_pressure_write,
5300 	},
5301 #endif /* CONFIG_PSI */
5302 	{ }	/* terminate */
5303 };
5304 
5305 /*
5306  * css destruction is four-stage process.
5307  *
5308  * 1. Destruction starts.  Killing of the percpu_ref is initiated.
5309  *    Implemented in kill_css().
5310  *
5311  * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5312  *    and thus css_tryget_online() is guaranteed to fail, the css can be
5313  *    offlined by invoking offline_css().  After offlining, the base ref is
5314  *    put.  Implemented in css_killed_work_fn().
5315  *
5316  * 3. When the percpu_ref reaches zero, the only possible remaining
5317  *    accessors are inside RCU read sections.  css_release() schedules the
5318  *    RCU callback.
5319  *
5320  * 4. After the grace period, the css can be freed.  Implemented in
5321  *    css_free_work_fn().
5322  *
5323  * It is actually hairier because both step 2 and 4 require process context
5324  * and thus involve punting to css->destroy_work adding two additional
5325  * steps to the already complex sequence.
5326  */
css_free_rwork_fn(struct work_struct * work)5327 static void css_free_rwork_fn(struct work_struct *work)
5328 {
5329 	struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5330 				struct cgroup_subsys_state, destroy_rwork);
5331 	struct cgroup_subsys *ss = css->ss;
5332 	struct cgroup *cgrp = css->cgroup;
5333 
5334 	percpu_ref_exit(&css->refcnt);
5335 
5336 	if (ss) {
5337 		/* css free path */
5338 		struct cgroup_subsys_state *parent = css->parent;
5339 		int id = css->id;
5340 
5341 		ss->css_free(css);
5342 		cgroup_idr_remove(&ss->css_idr, id);
5343 		cgroup_put(cgrp);
5344 
5345 		if (parent)
5346 			css_put(parent);
5347 	} else {
5348 		/* cgroup free path */
5349 		atomic_dec(&cgrp->root->nr_cgrps);
5350 		cgroup1_pidlist_destroy_all(cgrp);
5351 		cancel_work_sync(&cgrp->release_agent_work);
5352 
5353 		if (cgroup_parent(cgrp)) {
5354 			/*
5355 			 * We get a ref to the parent, and put the ref when
5356 			 * this cgroup is being freed, so it's guaranteed
5357 			 * that the parent won't be destroyed before its
5358 			 * children.
5359 			 */
5360 			cgroup_put(cgroup_parent(cgrp));
5361 			kernfs_put(cgrp->kn);
5362 			psi_cgroup_free(cgrp);
5363 			cgroup_rstat_exit(cgrp);
5364 			kfree(cgrp);
5365 		} else {
5366 			/*
5367 			 * This is root cgroup's refcnt reaching zero,
5368 			 * which indicates that the root should be
5369 			 * released.
5370 			 */
5371 			cgroup_destroy_root(cgrp->root);
5372 		}
5373 	}
5374 }
5375 
css_release_work_fn(struct work_struct * work)5376 static void css_release_work_fn(struct work_struct *work)
5377 {
5378 	struct cgroup_subsys_state *css =
5379 		container_of(work, struct cgroup_subsys_state, destroy_work);
5380 	struct cgroup_subsys *ss = css->ss;
5381 	struct cgroup *cgrp = css->cgroup;
5382 
5383 	mutex_lock(&cgroup_mutex);
5384 
5385 	css->flags |= CSS_RELEASED;
5386 	list_del_rcu(&css->sibling);
5387 
5388 	if (ss) {
5389 		/* css release path */
5390 		if (!list_empty(&css->rstat_css_node)) {
5391 			cgroup_rstat_flush(cgrp);
5392 			list_del_rcu(&css->rstat_css_node);
5393 		}
5394 
5395 		cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5396 		if (ss->css_released)
5397 			ss->css_released(css);
5398 	} else {
5399 		struct cgroup *tcgrp;
5400 
5401 		/* cgroup release path */
5402 		TRACE_CGROUP_PATH(release, cgrp);
5403 
5404 		cgroup_rstat_flush(cgrp);
5405 
5406 		spin_lock_irq(&css_set_lock);
5407 		for (tcgrp = cgroup_parent(cgrp); tcgrp;
5408 		     tcgrp = cgroup_parent(tcgrp))
5409 			tcgrp->nr_dying_descendants--;
5410 		spin_unlock_irq(&css_set_lock);
5411 
5412 		/*
5413 		 * There are two control paths which try to determine
5414 		 * cgroup from dentry without going through kernfs -
5415 		 * cgroupstats_build() and css_tryget_online_from_dir().
5416 		 * Those are supported by RCU protecting clearing of
5417 		 * cgrp->kn->priv backpointer.
5418 		 */
5419 		if (cgrp->kn)
5420 			RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5421 					 NULL);
5422 	}
5423 
5424 	mutex_unlock(&cgroup_mutex);
5425 
5426 	INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5427 	queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5428 }
5429 
css_release(struct percpu_ref * ref)5430 static void css_release(struct percpu_ref *ref)
5431 {
5432 	struct cgroup_subsys_state *css =
5433 		container_of(ref, struct cgroup_subsys_state, refcnt);
5434 
5435 	INIT_WORK(&css->destroy_work, css_release_work_fn);
5436 	queue_work(cgroup_destroy_wq, &css->destroy_work);
5437 }
5438 
init_and_link_css(struct cgroup_subsys_state * css,struct cgroup_subsys * ss,struct cgroup * cgrp)5439 static void init_and_link_css(struct cgroup_subsys_state *css,
5440 			      struct cgroup_subsys *ss, struct cgroup *cgrp)
5441 {
5442 	lockdep_assert_held(&cgroup_mutex);
5443 
5444 	cgroup_get_live(cgrp);
5445 
5446 	memset(css, 0, sizeof(*css));
5447 	css->cgroup = cgrp;
5448 	css->ss = ss;
5449 	css->id = -1;
5450 	INIT_LIST_HEAD(&css->sibling);
5451 	INIT_LIST_HEAD(&css->children);
5452 	INIT_LIST_HEAD(&css->rstat_css_node);
5453 	css->serial_nr = css_serial_nr_next++;
5454 	atomic_set(&css->online_cnt, 0);
5455 
5456 	if (cgroup_parent(cgrp)) {
5457 		css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5458 		css_get(css->parent);
5459 	}
5460 
5461 	if (ss->css_rstat_flush)
5462 		list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5463 
5464 	BUG_ON(cgroup_css(cgrp, ss));
5465 }
5466 
5467 /* invoke ->css_online() on a new CSS and mark it online if successful */
online_css(struct cgroup_subsys_state * css)5468 static int online_css(struct cgroup_subsys_state *css)
5469 {
5470 	struct cgroup_subsys *ss = css->ss;
5471 	int ret = 0;
5472 
5473 	lockdep_assert_held(&cgroup_mutex);
5474 
5475 	if (ss->css_online)
5476 		ret = ss->css_online(css);
5477 	if (!ret) {
5478 		css->flags |= CSS_ONLINE;
5479 		rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5480 
5481 		atomic_inc(&css->online_cnt);
5482 		if (css->parent)
5483 			atomic_inc(&css->parent->online_cnt);
5484 	}
5485 	return ret;
5486 }
5487 
5488 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
offline_css(struct cgroup_subsys_state * css)5489 static void offline_css(struct cgroup_subsys_state *css)
5490 {
5491 	struct cgroup_subsys *ss = css->ss;
5492 
5493 	lockdep_assert_held(&cgroup_mutex);
5494 
5495 	if (!(css->flags & CSS_ONLINE))
5496 		return;
5497 
5498 	if (ss->css_offline)
5499 		ss->css_offline(css);
5500 
5501 	css->flags &= ~CSS_ONLINE;
5502 	RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5503 
5504 	wake_up_all(&css->cgroup->offline_waitq);
5505 }
5506 
5507 /**
5508  * css_create - create a cgroup_subsys_state
5509  * @cgrp: the cgroup new css will be associated with
5510  * @ss: the subsys of new css
5511  *
5512  * Create a new css associated with @cgrp - @ss pair.  On success, the new
5513  * css is online and installed in @cgrp.  This function doesn't create the
5514  * interface files.  Returns 0 on success, -errno on failure.
5515  */
css_create(struct cgroup * cgrp,struct cgroup_subsys * ss)5516 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5517 					      struct cgroup_subsys *ss)
5518 {
5519 	struct cgroup *parent = cgroup_parent(cgrp);
5520 	struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5521 	struct cgroup_subsys_state *css;
5522 	int err;
5523 
5524 	lockdep_assert_held(&cgroup_mutex);
5525 
5526 	css = ss->css_alloc(parent_css);
5527 	if (!css)
5528 		css = ERR_PTR(-ENOMEM);
5529 	if (IS_ERR(css))
5530 		return css;
5531 
5532 	init_and_link_css(css, ss, cgrp);
5533 
5534 	err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5535 	if (err)
5536 		goto err_free_css;
5537 
5538 	err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5539 	if (err < 0)
5540 		goto err_free_css;
5541 	css->id = err;
5542 
5543 	/* @css is ready to be brought online now, make it visible */
5544 	list_add_tail_rcu(&css->sibling, &parent_css->children);
5545 	cgroup_idr_replace(&ss->css_idr, css, css->id);
5546 
5547 	err = online_css(css);
5548 	if (err)
5549 		goto err_list_del;
5550 
5551 	return css;
5552 
5553 err_list_del:
5554 	list_del_rcu(&css->sibling);
5555 err_free_css:
5556 	list_del_rcu(&css->rstat_css_node);
5557 	INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5558 	queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5559 	return ERR_PTR(err);
5560 }
5561 
5562 /*
5563  * The returned cgroup is fully initialized including its control mask, but
5564  * it isn't associated with its kernfs_node and doesn't have the control
5565  * mask applied.
5566  */
cgroup_create(struct cgroup * parent,const char * name,umode_t mode)5567 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5568 				    umode_t mode)
5569 {
5570 	struct cgroup_root *root = parent->root;
5571 	struct cgroup *cgrp, *tcgrp;
5572 	struct kernfs_node *kn;
5573 	int level = parent->level + 1;
5574 	int ret;
5575 
5576 	/* allocate the cgroup and its ID, 0 is reserved for the root */
5577 	cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5578 	if (!cgrp)
5579 		return ERR_PTR(-ENOMEM);
5580 
5581 	ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5582 	if (ret)
5583 		goto out_free_cgrp;
5584 
5585 	ret = cgroup_rstat_init(cgrp);
5586 	if (ret)
5587 		goto out_cancel_ref;
5588 
5589 	/* create the directory */
5590 	kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5591 	if (IS_ERR(kn)) {
5592 		ret = PTR_ERR(kn);
5593 		goto out_stat_exit;
5594 	}
5595 	cgrp->kn = kn;
5596 
5597 	init_cgroup_housekeeping(cgrp);
5598 
5599 	cgrp->self.parent = &parent->self;
5600 	cgrp->root = root;
5601 	cgrp->level = level;
5602 
5603 	ret = psi_cgroup_alloc(cgrp);
5604 	if (ret)
5605 		goto out_kernfs_remove;
5606 
5607 	ret = cgroup_bpf_inherit(cgrp);
5608 	if (ret)
5609 		goto out_psi_free;
5610 
5611 	/*
5612 	 * New cgroup inherits effective freeze counter, and
5613 	 * if the parent has to be frozen, the child has too.
5614 	 */
5615 	cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5616 	if (cgrp->freezer.e_freeze) {
5617 		/*
5618 		 * Set the CGRP_FREEZE flag, so when a process will be
5619 		 * attached to the child cgroup, it will become frozen.
5620 		 * At this point the new cgroup is unpopulated, so we can
5621 		 * consider it frozen immediately.
5622 		 */
5623 		set_bit(CGRP_FREEZE, &cgrp->flags);
5624 		set_bit(CGRP_FROZEN, &cgrp->flags);
5625 	}
5626 
5627 	spin_lock_irq(&css_set_lock);
5628 	for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5629 		cgrp->ancestors[tcgrp->level] = tcgrp;
5630 
5631 		if (tcgrp != cgrp) {
5632 			tcgrp->nr_descendants++;
5633 
5634 			/*
5635 			 * If the new cgroup is frozen, all ancestor cgroups
5636 			 * get a new frozen descendant, but their state can't
5637 			 * change because of this.
5638 			 */
5639 			if (cgrp->freezer.e_freeze)
5640 				tcgrp->freezer.nr_frozen_descendants++;
5641 		}
5642 	}
5643 	spin_unlock_irq(&css_set_lock);
5644 
5645 	if (notify_on_release(parent))
5646 		set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5647 
5648 	if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5649 		set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5650 
5651 	cgrp->self.serial_nr = css_serial_nr_next++;
5652 
5653 	/* allocation complete, commit to creation */
5654 	list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5655 	atomic_inc(&root->nr_cgrps);
5656 	cgroup_get_live(parent);
5657 
5658 	/*
5659 	 * On the default hierarchy, a child doesn't automatically inherit
5660 	 * subtree_control from the parent.  Each is configured manually.
5661 	 */
5662 	if (!cgroup_on_dfl(cgrp))
5663 		cgrp->subtree_control = cgroup_control(cgrp);
5664 
5665 	cgroup_propagate_control(cgrp);
5666 
5667 	return cgrp;
5668 
5669 out_psi_free:
5670 	psi_cgroup_free(cgrp);
5671 out_kernfs_remove:
5672 	kernfs_remove(cgrp->kn);
5673 out_stat_exit:
5674 	cgroup_rstat_exit(cgrp);
5675 out_cancel_ref:
5676 	percpu_ref_exit(&cgrp->self.refcnt);
5677 out_free_cgrp:
5678 	kfree(cgrp);
5679 	return ERR_PTR(ret);
5680 }
5681 
cgroup_check_hierarchy_limits(struct cgroup * parent)5682 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5683 {
5684 	struct cgroup *cgroup;
5685 	int ret = false;
5686 	int level = 1;
5687 
5688 	lockdep_assert_held(&cgroup_mutex);
5689 
5690 	for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5691 		if (cgroup->nr_descendants >= cgroup->max_descendants)
5692 			goto fail;
5693 
5694 		if (level > cgroup->max_depth)
5695 			goto fail;
5696 
5697 		level++;
5698 	}
5699 
5700 	ret = true;
5701 fail:
5702 	return ret;
5703 }
5704 
cgroup_mkdir(struct kernfs_node * parent_kn,const char * name,umode_t mode)5705 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5706 {
5707 	struct cgroup *parent, *cgrp;
5708 	int ret;
5709 
5710 	/* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5711 	if (strchr(name, '\n'))
5712 		return -EINVAL;
5713 
5714 	parent = cgroup_kn_lock_live(parent_kn, false);
5715 	if (!parent)
5716 		return -ENODEV;
5717 
5718 	if (!cgroup_check_hierarchy_limits(parent)) {
5719 		ret = -EAGAIN;
5720 		goto out_unlock;
5721 	}
5722 
5723 	cgrp = cgroup_create(parent, name, mode);
5724 	if (IS_ERR(cgrp)) {
5725 		ret = PTR_ERR(cgrp);
5726 		goto out_unlock;
5727 	}
5728 
5729 	/*
5730 	 * This extra ref will be put in cgroup_free_fn() and guarantees
5731 	 * that @cgrp->kn is always accessible.
5732 	 */
5733 	kernfs_get(cgrp->kn);
5734 
5735 	ret = cgroup_kn_set_ugid(cgrp->kn);
5736 	if (ret)
5737 		goto out_destroy;
5738 
5739 	ret = css_populate_dir(&cgrp->self);
5740 	if (ret)
5741 		goto out_destroy;
5742 
5743 	ret = cgroup_apply_control_enable(cgrp);
5744 	if (ret)
5745 		goto out_destroy;
5746 
5747 	TRACE_CGROUP_PATH(mkdir, cgrp);
5748 
5749 	/* let's create and online css's */
5750 	kernfs_activate(cgrp->kn);
5751 
5752 	ret = 0;
5753 	goto out_unlock;
5754 
5755 out_destroy:
5756 	cgroup_destroy_locked(cgrp);
5757 out_unlock:
5758 	cgroup_kn_unlock(parent_kn);
5759 	return ret;
5760 }
5761 
5762 /*
5763  * This is called when the refcnt of a css is confirmed to be killed.
5764  * css_tryget_online() is now guaranteed to fail.  Tell the subsystem to
5765  * initiate destruction and put the css ref from kill_css().
5766  */
css_killed_work_fn(struct work_struct * work)5767 static void css_killed_work_fn(struct work_struct *work)
5768 {
5769 	struct cgroup_subsys_state *css =
5770 		container_of(work, struct cgroup_subsys_state, destroy_work);
5771 
5772 	mutex_lock(&cgroup_mutex);
5773 
5774 	do {
5775 		offline_css(css);
5776 		css_put(css);
5777 		/* @css can't go away while we're holding cgroup_mutex */
5778 		css = css->parent;
5779 	} while (css && atomic_dec_and_test(&css->online_cnt));
5780 
5781 	mutex_unlock(&cgroup_mutex);
5782 }
5783 
5784 /* css kill confirmation processing requires process context, bounce */
css_killed_ref_fn(struct percpu_ref * ref)5785 static void css_killed_ref_fn(struct percpu_ref *ref)
5786 {
5787 	struct cgroup_subsys_state *css =
5788 		container_of(ref, struct cgroup_subsys_state, refcnt);
5789 
5790 	if (atomic_dec_and_test(&css->online_cnt)) {
5791 		INIT_WORK(&css->destroy_work, css_killed_work_fn);
5792 		queue_work(cgroup_destroy_wq, &css->destroy_work);
5793 	}
5794 }
5795 
5796 /**
5797  * kill_css - destroy a css
5798  * @css: css to destroy
5799  *
5800  * This function initiates destruction of @css by removing cgroup interface
5801  * files and putting its base reference.  ->css_offline() will be invoked
5802  * asynchronously once css_tryget_online() is guaranteed to fail and when
5803  * the reference count reaches zero, @css will be released.
5804  */
kill_css(struct cgroup_subsys_state * css)5805 static void kill_css(struct cgroup_subsys_state *css)
5806 {
5807 	lockdep_assert_held(&cgroup_mutex);
5808 
5809 	if (css->flags & CSS_DYING)
5810 		return;
5811 
5812 	css->flags |= CSS_DYING;
5813 
5814 	/*
5815 	 * This must happen before css is disassociated with its cgroup.
5816 	 * See seq_css() for details.
5817 	 */
5818 	css_clear_dir(css);
5819 
5820 	/*
5821 	 * Killing would put the base ref, but we need to keep it alive
5822 	 * until after ->css_offline().
5823 	 */
5824 	css_get(css);
5825 
5826 	/*
5827 	 * cgroup core guarantees that, by the time ->css_offline() is
5828 	 * invoked, no new css reference will be given out via
5829 	 * css_tryget_online().  We can't simply call percpu_ref_kill() and
5830 	 * proceed to offlining css's because percpu_ref_kill() doesn't
5831 	 * guarantee that the ref is seen as killed on all CPUs on return.
5832 	 *
5833 	 * Use percpu_ref_kill_and_confirm() to get notifications as each
5834 	 * css is confirmed to be seen as killed on all CPUs.
5835 	 */
5836 	percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5837 }
5838 
5839 /**
5840  * cgroup_destroy_locked - the first stage of cgroup destruction
5841  * @cgrp: cgroup to be destroyed
5842  *
5843  * css's make use of percpu refcnts whose killing latency shouldn't be
5844  * exposed to userland and are RCU protected.  Also, cgroup core needs to
5845  * guarantee that css_tryget_online() won't succeed by the time
5846  * ->css_offline() is invoked.  To satisfy all the requirements,
5847  * destruction is implemented in the following two steps.
5848  *
5849  * s1. Verify @cgrp can be destroyed and mark it dying.  Remove all
5850  *     userland visible parts and start killing the percpu refcnts of
5851  *     css's.  Set up so that the next stage will be kicked off once all
5852  *     the percpu refcnts are confirmed to be killed.
5853  *
5854  * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5855  *     rest of destruction.  Once all cgroup references are gone, the
5856  *     cgroup is RCU-freed.
5857  *
5858  * This function implements s1.  After this step, @cgrp is gone as far as
5859  * the userland is concerned and a new cgroup with the same name may be
5860  * created.  As cgroup doesn't care about the names internally, this
5861  * doesn't cause any problem.
5862  */
cgroup_destroy_locked(struct cgroup * cgrp)5863 static int cgroup_destroy_locked(struct cgroup *cgrp)
5864 	__releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5865 {
5866 	struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5867 	struct cgroup_subsys_state *css;
5868 	struct cgrp_cset_link *link;
5869 	int ssid;
5870 
5871 	lockdep_assert_held(&cgroup_mutex);
5872 
5873 	/*
5874 	 * Only migration can raise populated from zero and we're already
5875 	 * holding cgroup_mutex.
5876 	 */
5877 	if (cgroup_is_populated(cgrp))
5878 		return -EBUSY;
5879 
5880 	/*
5881 	 * Make sure there's no live children.  We can't test emptiness of
5882 	 * ->self.children as dead children linger on it while being
5883 	 * drained; otherwise, "rmdir parent/child parent" may fail.
5884 	 */
5885 	if (css_has_online_children(&cgrp->self))
5886 		return -EBUSY;
5887 
5888 	/*
5889 	 * Mark @cgrp and the associated csets dead.  The former prevents
5890 	 * further task migration and child creation by disabling
5891 	 * cgroup_lock_live_group().  The latter makes the csets ignored by
5892 	 * the migration path.
5893 	 */
5894 	cgrp->self.flags &= ~CSS_ONLINE;
5895 
5896 	spin_lock_irq(&css_set_lock);
5897 	list_for_each_entry(link, &cgrp->cset_links, cset_link)
5898 		link->cset->dead = true;
5899 	spin_unlock_irq(&css_set_lock);
5900 
5901 	/* initiate massacre of all css's */
5902 	for_each_css(css, ssid, cgrp)
5903 		kill_css(css);
5904 
5905 	/* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5906 	css_clear_dir(&cgrp->self);
5907 	kernfs_remove(cgrp->kn);
5908 
5909 	if (cgroup_is_threaded(cgrp))
5910 		parent->nr_threaded_children--;
5911 
5912 	spin_lock_irq(&css_set_lock);
5913 	for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5914 		tcgrp->nr_descendants--;
5915 		tcgrp->nr_dying_descendants++;
5916 		/*
5917 		 * If the dying cgroup is frozen, decrease frozen descendants
5918 		 * counters of ancestor cgroups.
5919 		 */
5920 		if (test_bit(CGRP_FROZEN, &cgrp->flags))
5921 			tcgrp->freezer.nr_frozen_descendants--;
5922 	}
5923 	spin_unlock_irq(&css_set_lock);
5924 
5925 	cgroup1_check_for_release(parent);
5926 
5927 	cgroup_bpf_offline(cgrp);
5928 
5929 	/* put the base reference */
5930 	percpu_ref_kill(&cgrp->self.refcnt);
5931 
5932 	return 0;
5933 };
5934 
cgroup_rmdir(struct kernfs_node * kn)5935 int cgroup_rmdir(struct kernfs_node *kn)
5936 {
5937 	struct cgroup *cgrp;
5938 	int ret = 0;
5939 
5940 	cgrp = cgroup_kn_lock_live(kn, false);
5941 	if (!cgrp)
5942 		return 0;
5943 
5944 	ret = cgroup_destroy_locked(cgrp);
5945 	if (!ret)
5946 		TRACE_CGROUP_PATH(rmdir, cgrp);
5947 
5948 	cgroup_kn_unlock(kn);
5949 	return ret;
5950 }
5951 
5952 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5953 	.show_options		= cgroup_show_options,
5954 	.mkdir			= cgroup_mkdir,
5955 	.rmdir			= cgroup_rmdir,
5956 	.show_path		= cgroup_show_path,
5957 };
5958 
cgroup_init_subsys(struct cgroup_subsys * ss,bool early)5959 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5960 {
5961 	struct cgroup_subsys_state *css;
5962 
5963 	pr_debug("Initializing cgroup subsys %s\n", ss->name);
5964 
5965 	mutex_lock(&cgroup_mutex);
5966 
5967 	idr_init(&ss->css_idr);
5968 	INIT_LIST_HEAD(&ss->cfts);
5969 
5970 	/* Create the root cgroup state for this subsystem */
5971 	ss->root = &cgrp_dfl_root;
5972 	css = ss->css_alloc(NULL);
5973 	/* We don't handle early failures gracefully */
5974 	BUG_ON(IS_ERR(css));
5975 	init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5976 
5977 	/*
5978 	 * Root csses are never destroyed and we can't initialize
5979 	 * percpu_ref during early init.  Disable refcnting.
5980 	 */
5981 	css->flags |= CSS_NO_REF;
5982 
5983 	if (early) {
5984 		/* allocation can't be done safely during early init */
5985 		css->id = 1;
5986 	} else {
5987 		css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5988 		BUG_ON(css->id < 0);
5989 	}
5990 
5991 	/* Update the init_css_set to contain a subsys
5992 	 * pointer to this state - since the subsystem is
5993 	 * newly registered, all tasks and hence the
5994 	 * init_css_set is in the subsystem's root cgroup. */
5995 	init_css_set.subsys[ss->id] = css;
5996 
5997 	have_fork_callback |= (bool)ss->fork << ss->id;
5998 	have_exit_callback |= (bool)ss->exit << ss->id;
5999 	have_release_callback |= (bool)ss->release << ss->id;
6000 	have_canfork_callback |= (bool)ss->can_fork << ss->id;
6001 
6002 	/* At system boot, before all subsystems have been
6003 	 * registered, no tasks have been forked, so we don't
6004 	 * need to invoke fork callbacks here. */
6005 	BUG_ON(!list_empty(&init_task.tasks));
6006 
6007 	BUG_ON(online_css(css));
6008 
6009 	mutex_unlock(&cgroup_mutex);
6010 }
6011 
6012 /**
6013  * cgroup_init_early - cgroup initialization at system boot
6014  *
6015  * Initialize cgroups at system boot, and initialize any
6016  * subsystems that request early init.
6017  */
cgroup_init_early(void)6018 int __init cgroup_init_early(void)
6019 {
6020 	static struct cgroup_fs_context __initdata ctx;
6021 	struct cgroup_subsys *ss;
6022 	int i;
6023 
6024 	ctx.root = &cgrp_dfl_root;
6025 	init_cgroup_root(&ctx);
6026 	cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6027 
6028 	RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6029 
6030 	for_each_subsys(ss, i) {
6031 		WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6032 		     "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6033 		     i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6034 		     ss->id, ss->name);
6035 		WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6036 		     "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6037 
6038 		ss->id = i;
6039 		ss->name = cgroup_subsys_name[i];
6040 		if (!ss->legacy_name)
6041 			ss->legacy_name = cgroup_subsys_name[i];
6042 
6043 		if (ss->early_init)
6044 			cgroup_init_subsys(ss, true);
6045 	}
6046 	return 0;
6047 }
6048 
6049 /**
6050  * cgroup_init - cgroup initialization
6051  *
6052  * Register cgroup filesystem and /proc file, and initialize
6053  * any subsystems that didn't request early init.
6054  */
cgroup_init(void)6055 int __init cgroup_init(void)
6056 {
6057 	struct cgroup_subsys *ss;
6058 	int ssid;
6059 
6060 	BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6061 	BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6062 	BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6063 	BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6064 
6065 	cgroup_rstat_boot();
6066 
6067 	get_user_ns(init_cgroup_ns.user_ns);
6068 
6069 	mutex_lock(&cgroup_mutex);
6070 
6071 	/*
6072 	 * Add init_css_set to the hash table so that dfl_root can link to
6073 	 * it during init.
6074 	 */
6075 	hash_add(css_set_table, &init_css_set.hlist,
6076 		 css_set_hash(init_css_set.subsys));
6077 
6078 	BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6079 
6080 	mutex_unlock(&cgroup_mutex);
6081 
6082 	for_each_subsys(ss, ssid) {
6083 		if (ss->early_init) {
6084 			struct cgroup_subsys_state *css =
6085 				init_css_set.subsys[ss->id];
6086 
6087 			css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6088 						   GFP_KERNEL);
6089 			BUG_ON(css->id < 0);
6090 		} else {
6091 			cgroup_init_subsys(ss, false);
6092 		}
6093 
6094 		list_add_tail(&init_css_set.e_cset_node[ssid],
6095 			      &cgrp_dfl_root.cgrp.e_csets[ssid]);
6096 
6097 		/*
6098 		 * Setting dfl_root subsys_mask needs to consider the
6099 		 * disabled flag and cftype registration needs kmalloc,
6100 		 * both of which aren't available during early_init.
6101 		 */
6102 		if (!cgroup_ssid_enabled(ssid))
6103 			continue;
6104 
6105 		if (cgroup1_ssid_disabled(ssid))
6106 			printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
6107 			       ss->name);
6108 
6109 		cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6110 
6111 		/* implicit controllers must be threaded too */
6112 		WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6113 
6114 		if (ss->implicit_on_dfl)
6115 			cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6116 		else if (!ss->dfl_cftypes)
6117 			cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6118 
6119 		if (ss->threaded)
6120 			cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6121 
6122 		if (ss->dfl_cftypes == ss->legacy_cftypes) {
6123 			WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6124 		} else {
6125 			WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6126 			WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6127 		}
6128 
6129 		if (ss->bind)
6130 			ss->bind(init_css_set.subsys[ssid]);
6131 
6132 		mutex_lock(&cgroup_mutex);
6133 		css_populate_dir(init_css_set.subsys[ssid]);
6134 		mutex_unlock(&cgroup_mutex);
6135 	}
6136 
6137 	/* init_css_set.subsys[] has been updated, re-hash */
6138 	hash_del(&init_css_set.hlist);
6139 	hash_add(css_set_table, &init_css_set.hlist,
6140 		 css_set_hash(init_css_set.subsys));
6141 
6142 	WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6143 	WARN_ON(register_filesystem(&cgroup_fs_type));
6144 	WARN_ON(register_filesystem(&cgroup2_fs_type));
6145 	WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6146 #ifdef CONFIG_CPUSETS
6147 	WARN_ON(register_filesystem(&cpuset_fs_type));
6148 #endif
6149 
6150 	return 0;
6151 }
6152 
cgroup_wq_init(void)6153 static int __init cgroup_wq_init(void)
6154 {
6155 	/*
6156 	 * There isn't much point in executing destruction path in
6157 	 * parallel.  Good chunk is serialized with cgroup_mutex anyway.
6158 	 * Use 1 for @max_active.
6159 	 *
6160 	 * We would prefer to do this in cgroup_init() above, but that
6161 	 * is called before init_workqueues(): so leave this until after.
6162 	 */
6163 	cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
6164 	BUG_ON(!cgroup_destroy_wq);
6165 	return 0;
6166 }
6167 core_initcall(cgroup_wq_init);
6168 
cgroup_path_from_kernfs_id(u64 id,char * buf,size_t buflen)6169 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6170 {
6171 	struct kernfs_node *kn;
6172 
6173 	kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6174 	if (!kn)
6175 		return;
6176 	kernfs_path(kn, buf, buflen);
6177 	kernfs_put(kn);
6178 }
6179 
6180 /*
6181  * cgroup_get_from_id : get the cgroup associated with cgroup id
6182  * @id: cgroup id
6183  * On success return the cgrp or ERR_PTR on failure
6184  * Only cgroups within current task's cgroup NS are valid.
6185  */
cgroup_get_from_id(u64 id)6186 struct cgroup *cgroup_get_from_id(u64 id)
6187 {
6188 	struct kernfs_node *kn;
6189 	struct cgroup *cgrp, *root_cgrp;
6190 
6191 	kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6192 	if (!kn)
6193 		return ERR_PTR(-ENOENT);
6194 
6195 	if (kernfs_type(kn) != KERNFS_DIR) {
6196 		kernfs_put(kn);
6197 		return ERR_PTR(-ENOENT);
6198 	}
6199 
6200 	rcu_read_lock();
6201 
6202 	cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6203 	if (cgrp && !cgroup_tryget(cgrp))
6204 		cgrp = NULL;
6205 
6206 	rcu_read_unlock();
6207 	kernfs_put(kn);
6208 
6209 	if (!cgrp)
6210 		return ERR_PTR(-ENOENT);
6211 
6212 	root_cgrp = current_cgns_cgroup_dfl();
6213 	if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6214 		cgroup_put(cgrp);
6215 		return ERR_PTR(-ENOENT);
6216 	}
6217 
6218 	return cgrp;
6219 }
6220 EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6221 
6222 /*
6223  * proc_cgroup_show()
6224  *  - Print task's cgroup paths into seq_file, one line for each hierarchy
6225  *  - Used for /proc/<pid>/cgroup.
6226  */
proc_cgroup_show(struct seq_file * m,struct pid_namespace * ns,struct pid * pid,struct task_struct * tsk)6227 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6228 		     struct pid *pid, struct task_struct *tsk)
6229 {
6230 	char *buf;
6231 	int retval;
6232 	struct cgroup_root *root;
6233 
6234 	retval = -ENOMEM;
6235 	buf = kmalloc(PATH_MAX, GFP_KERNEL);
6236 	if (!buf)
6237 		goto out;
6238 
6239 	mutex_lock(&cgroup_mutex);
6240 	spin_lock_irq(&css_set_lock);
6241 
6242 	for_each_root(root) {
6243 		struct cgroup_subsys *ss;
6244 		struct cgroup *cgrp;
6245 		int ssid, count = 0;
6246 
6247 		if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6248 			continue;
6249 
6250 		seq_printf(m, "%d:", root->hierarchy_id);
6251 		if (root != &cgrp_dfl_root)
6252 			for_each_subsys(ss, ssid)
6253 				if (root->subsys_mask & (1 << ssid))
6254 					seq_printf(m, "%s%s", count++ ? "," : "",
6255 						   ss->legacy_name);
6256 		if (strlen(root->name))
6257 			seq_printf(m, "%sname=%s", count ? "," : "",
6258 				   root->name);
6259 		seq_putc(m, ':');
6260 
6261 		cgrp = task_cgroup_from_root(tsk, root);
6262 
6263 		/*
6264 		 * On traditional hierarchies, all zombie tasks show up as
6265 		 * belonging to the root cgroup.  On the default hierarchy,
6266 		 * while a zombie doesn't show up in "cgroup.procs" and
6267 		 * thus can't be migrated, its /proc/PID/cgroup keeps
6268 		 * reporting the cgroup it belonged to before exiting.  If
6269 		 * the cgroup is removed before the zombie is reaped,
6270 		 * " (deleted)" is appended to the cgroup path.
6271 		 */
6272 		if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6273 			retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6274 						current->nsproxy->cgroup_ns);
6275 			if (retval >= PATH_MAX)
6276 				retval = -ENAMETOOLONG;
6277 			if (retval < 0)
6278 				goto out_unlock;
6279 
6280 			seq_puts(m, buf);
6281 		} else {
6282 			seq_puts(m, "/");
6283 		}
6284 
6285 		if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6286 			seq_puts(m, " (deleted)\n");
6287 		else
6288 			seq_putc(m, '\n');
6289 	}
6290 
6291 	retval = 0;
6292 out_unlock:
6293 	spin_unlock_irq(&css_set_lock);
6294 	mutex_unlock(&cgroup_mutex);
6295 	kfree(buf);
6296 out:
6297 	return retval;
6298 }
6299 
6300 /**
6301  * cgroup_fork - initialize cgroup related fields during copy_process()
6302  * @child: pointer to task_struct of forking parent process.
6303  *
6304  * A task is associated with the init_css_set until cgroup_post_fork()
6305  * attaches it to the target css_set.
6306  */
cgroup_fork(struct task_struct * child)6307 void cgroup_fork(struct task_struct *child)
6308 {
6309 	RCU_INIT_POINTER(child->cgroups, &init_css_set);
6310 	INIT_LIST_HEAD(&child->cg_list);
6311 }
6312 
6313 /**
6314  * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6315  * @f: file corresponding to cgroup_dir
6316  *
6317  * Find the cgroup from a file pointer associated with a cgroup directory.
6318  * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6319  * cgroup cannot be found.
6320  */
cgroup_v1v2_get_from_file(struct file * f)6321 static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6322 {
6323 	struct cgroup_subsys_state *css;
6324 
6325 	css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6326 	if (IS_ERR(css))
6327 		return ERR_CAST(css);
6328 
6329 	return css->cgroup;
6330 }
6331 
6332 /**
6333  * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6334  * cgroup2.
6335  * @f: file corresponding to cgroup2_dir
6336  */
cgroup_get_from_file(struct file * f)6337 static struct cgroup *cgroup_get_from_file(struct file *f)
6338 {
6339 	struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6340 
6341 	if (IS_ERR(cgrp))
6342 		return ERR_CAST(cgrp);
6343 
6344 	if (!cgroup_on_dfl(cgrp)) {
6345 		cgroup_put(cgrp);
6346 		return ERR_PTR(-EBADF);
6347 	}
6348 
6349 	return cgrp;
6350 }
6351 
6352 /**
6353  * cgroup_css_set_fork - find or create a css_set for a child process
6354  * @kargs: the arguments passed to create the child process
6355  *
6356  * This functions finds or creates a new css_set which the child
6357  * process will be attached to in cgroup_post_fork(). By default,
6358  * the child process will be given the same css_set as its parent.
6359  *
6360  * If CLONE_INTO_CGROUP is specified this function will try to find an
6361  * existing css_set which includes the requested cgroup and if not create
6362  * a new css_set that the child will be attached to later. If this function
6363  * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6364  * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6365  * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6366  * to the target cgroup.
6367  */
cgroup_css_set_fork(struct kernel_clone_args * kargs)6368 static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6369 	__acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6370 {
6371 	int ret;
6372 	struct cgroup *dst_cgrp = NULL;
6373 	struct css_set *cset;
6374 	struct super_block *sb;
6375 	struct file *f;
6376 
6377 	if (kargs->flags & CLONE_INTO_CGROUP)
6378 		mutex_lock(&cgroup_mutex);
6379 
6380 	cgroup_threadgroup_change_begin(current);
6381 
6382 	spin_lock_irq(&css_set_lock);
6383 	cset = task_css_set(current);
6384 	get_css_set(cset);
6385 	spin_unlock_irq(&css_set_lock);
6386 
6387 	if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6388 		kargs->cset = cset;
6389 		return 0;
6390 	}
6391 
6392 	f = fget_raw(kargs->cgroup);
6393 	if (!f) {
6394 		ret = -EBADF;
6395 		goto err;
6396 	}
6397 	sb = f->f_path.dentry->d_sb;
6398 
6399 	dst_cgrp = cgroup_get_from_file(f);
6400 	if (IS_ERR(dst_cgrp)) {
6401 		ret = PTR_ERR(dst_cgrp);
6402 		dst_cgrp = NULL;
6403 		goto err;
6404 	}
6405 
6406 	if (cgroup_is_dead(dst_cgrp)) {
6407 		ret = -ENODEV;
6408 		goto err;
6409 	}
6410 
6411 	/*
6412 	 * Verify that we the target cgroup is writable for us. This is
6413 	 * usually done by the vfs layer but since we're not going through
6414 	 * the vfs layer here we need to do it "manually".
6415 	 */
6416 	ret = cgroup_may_write(dst_cgrp, sb);
6417 	if (ret)
6418 		goto err;
6419 
6420 	/*
6421 	 * Spawning a task directly into a cgroup works by passing a file
6422 	 * descriptor to the target cgroup directory. This can even be an O_PATH
6423 	 * file descriptor. But it can never be a cgroup.procs file descriptor.
6424 	 * This was done on purpose so spawning into a cgroup could be
6425 	 * conceptualized as an atomic
6426 	 *
6427 	 *   fd = openat(dfd_cgroup, "cgroup.procs", ...);
6428 	 *   write(fd, <child-pid>, ...);
6429 	 *
6430 	 * sequence, i.e. it's a shorthand for the caller opening and writing
6431 	 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6432 	 * to always use the caller's credentials.
6433 	 */
6434 	ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6435 					!(kargs->flags & CLONE_THREAD),
6436 					current->nsproxy->cgroup_ns);
6437 	if (ret)
6438 		goto err;
6439 
6440 	kargs->cset = find_css_set(cset, dst_cgrp);
6441 	if (!kargs->cset) {
6442 		ret = -ENOMEM;
6443 		goto err;
6444 	}
6445 
6446 	put_css_set(cset);
6447 	fput(f);
6448 	kargs->cgrp = dst_cgrp;
6449 	return ret;
6450 
6451 err:
6452 	cgroup_threadgroup_change_end(current);
6453 	mutex_unlock(&cgroup_mutex);
6454 	if (f)
6455 		fput(f);
6456 	if (dst_cgrp)
6457 		cgroup_put(dst_cgrp);
6458 	put_css_set(cset);
6459 	if (kargs->cset)
6460 		put_css_set(kargs->cset);
6461 	return ret;
6462 }
6463 
6464 /**
6465  * cgroup_css_set_put_fork - drop references we took during fork
6466  * @kargs: the arguments passed to create the child process
6467  *
6468  * Drop references to the prepared css_set and target cgroup if
6469  * CLONE_INTO_CGROUP was requested.
6470  */
cgroup_css_set_put_fork(struct kernel_clone_args * kargs)6471 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6472 	__releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6473 {
6474 	cgroup_threadgroup_change_end(current);
6475 
6476 	if (kargs->flags & CLONE_INTO_CGROUP) {
6477 		struct cgroup *cgrp = kargs->cgrp;
6478 		struct css_set *cset = kargs->cset;
6479 
6480 		mutex_unlock(&cgroup_mutex);
6481 
6482 		if (cset) {
6483 			put_css_set(cset);
6484 			kargs->cset = NULL;
6485 		}
6486 
6487 		if (cgrp) {
6488 			cgroup_put(cgrp);
6489 			kargs->cgrp = NULL;
6490 		}
6491 	}
6492 }
6493 
6494 /**
6495  * cgroup_can_fork - called on a new task before the process is exposed
6496  * @child: the child process
6497  * @kargs: the arguments passed to create the child process
6498  *
6499  * This prepares a new css_set for the child process which the child will
6500  * be attached to in cgroup_post_fork().
6501  * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6502  * callback returns an error, the fork aborts with that error code. This
6503  * allows for a cgroup subsystem to conditionally allow or deny new forks.
6504  */
cgroup_can_fork(struct task_struct * child,struct kernel_clone_args * kargs)6505 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6506 {
6507 	struct cgroup_subsys *ss;
6508 	int i, j, ret;
6509 
6510 	ret = cgroup_css_set_fork(kargs);
6511 	if (ret)
6512 		return ret;
6513 
6514 	do_each_subsys_mask(ss, i, have_canfork_callback) {
6515 		ret = ss->can_fork(child, kargs->cset);
6516 		if (ret)
6517 			goto out_revert;
6518 	} while_each_subsys_mask();
6519 
6520 	return 0;
6521 
6522 out_revert:
6523 	for_each_subsys(ss, j) {
6524 		if (j >= i)
6525 			break;
6526 		if (ss->cancel_fork)
6527 			ss->cancel_fork(child, kargs->cset);
6528 	}
6529 
6530 	cgroup_css_set_put_fork(kargs);
6531 
6532 	return ret;
6533 }
6534 
6535 /**
6536  * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6537  * @child: the child process
6538  * @kargs: the arguments passed to create the child process
6539  *
6540  * This calls the cancel_fork() callbacks if a fork failed *after*
6541  * cgroup_can_fork() succeeded and cleans up references we took to
6542  * prepare a new css_set for the child process in cgroup_can_fork().
6543  */
cgroup_cancel_fork(struct task_struct * child,struct kernel_clone_args * kargs)6544 void cgroup_cancel_fork(struct task_struct *child,
6545 			struct kernel_clone_args *kargs)
6546 {
6547 	struct cgroup_subsys *ss;
6548 	int i;
6549 
6550 	for_each_subsys(ss, i)
6551 		if (ss->cancel_fork)
6552 			ss->cancel_fork(child, kargs->cset);
6553 
6554 	cgroup_css_set_put_fork(kargs);
6555 }
6556 
6557 /**
6558  * cgroup_post_fork - finalize cgroup setup for the child process
6559  * @child: the child process
6560  * @kargs: the arguments passed to create the child process
6561  *
6562  * Attach the child process to its css_set calling the subsystem fork()
6563  * callbacks.
6564  */
cgroup_post_fork(struct task_struct * child,struct kernel_clone_args * kargs)6565 void cgroup_post_fork(struct task_struct *child,
6566 		      struct kernel_clone_args *kargs)
6567 	__releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6568 {
6569 	unsigned long cgrp_flags = 0;
6570 	bool kill = false;
6571 	struct cgroup_subsys *ss;
6572 	struct css_set *cset;
6573 	int i;
6574 
6575 	cset = kargs->cset;
6576 	kargs->cset = NULL;
6577 
6578 	spin_lock_irq(&css_set_lock);
6579 
6580 	/* init tasks are special, only link regular threads */
6581 	if (likely(child->pid)) {
6582 		if (kargs->cgrp)
6583 			cgrp_flags = kargs->cgrp->flags;
6584 		else
6585 			cgrp_flags = cset->dfl_cgrp->flags;
6586 
6587 		WARN_ON_ONCE(!list_empty(&child->cg_list));
6588 		cset->nr_tasks++;
6589 		css_set_move_task(child, NULL, cset, false);
6590 	} else {
6591 		put_css_set(cset);
6592 		cset = NULL;
6593 	}
6594 
6595 	if (!(child->flags & PF_KTHREAD)) {
6596 		if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6597 			/*
6598 			 * If the cgroup has to be frozen, the new task has
6599 			 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6600 			 * get the task into the frozen state.
6601 			 */
6602 			spin_lock(&child->sighand->siglock);
6603 			WARN_ON_ONCE(child->frozen);
6604 			child->jobctl |= JOBCTL_TRAP_FREEZE;
6605 			spin_unlock(&child->sighand->siglock);
6606 
6607 			/*
6608 			 * Calling cgroup_update_frozen() isn't required here,
6609 			 * because it will be called anyway a bit later from
6610 			 * do_freezer_trap(). So we avoid cgroup's transient
6611 			 * switch from the frozen state and back.
6612 			 */
6613 		}
6614 
6615 		/*
6616 		 * If the cgroup is to be killed notice it now and take the
6617 		 * child down right after we finished preparing it for
6618 		 * userspace.
6619 		 */
6620 		kill = test_bit(CGRP_KILL, &cgrp_flags);
6621 	}
6622 
6623 	spin_unlock_irq(&css_set_lock);
6624 
6625 	/*
6626 	 * Call ss->fork().  This must happen after @child is linked on
6627 	 * css_set; otherwise, @child might change state between ->fork()
6628 	 * and addition to css_set.
6629 	 */
6630 	do_each_subsys_mask(ss, i, have_fork_callback) {
6631 		ss->fork(child);
6632 	} while_each_subsys_mask();
6633 
6634 	/* Make the new cset the root_cset of the new cgroup namespace. */
6635 	if (kargs->flags & CLONE_NEWCGROUP) {
6636 		struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6637 
6638 		get_css_set(cset);
6639 		child->nsproxy->cgroup_ns->root_cset = cset;
6640 		put_css_set(rcset);
6641 	}
6642 
6643 	/* Cgroup has to be killed so take down child immediately. */
6644 	if (unlikely(kill))
6645 		do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6646 
6647 	cgroup_css_set_put_fork(kargs);
6648 }
6649 
6650 /**
6651  * cgroup_exit - detach cgroup from exiting task
6652  * @tsk: pointer to task_struct of exiting process
6653  *
6654  * Description: Detach cgroup from @tsk.
6655  *
6656  */
cgroup_exit(struct task_struct * tsk)6657 void cgroup_exit(struct task_struct *tsk)
6658 {
6659 	struct cgroup_subsys *ss;
6660 	struct css_set *cset;
6661 	int i;
6662 
6663 	spin_lock_irq(&css_set_lock);
6664 
6665 	WARN_ON_ONCE(list_empty(&tsk->cg_list));
6666 	cset = task_css_set(tsk);
6667 	css_set_move_task(tsk, cset, NULL, false);
6668 	list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6669 	cset->nr_tasks--;
6670 
6671 	WARN_ON_ONCE(cgroup_task_frozen(tsk));
6672 	if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6673 		     test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6674 		cgroup_update_frozen(task_dfl_cgroup(tsk));
6675 
6676 	spin_unlock_irq(&css_set_lock);
6677 
6678 	/* see cgroup_post_fork() for details */
6679 	do_each_subsys_mask(ss, i, have_exit_callback) {
6680 		ss->exit(tsk);
6681 	} while_each_subsys_mask();
6682 }
6683 
cgroup_release(struct task_struct * task)6684 void cgroup_release(struct task_struct *task)
6685 {
6686 	struct cgroup_subsys *ss;
6687 	int ssid;
6688 
6689 	do_each_subsys_mask(ss, ssid, have_release_callback) {
6690 		ss->release(task);
6691 	} while_each_subsys_mask();
6692 
6693 	spin_lock_irq(&css_set_lock);
6694 	css_set_skip_task_iters(task_css_set(task), task);
6695 	list_del_init(&task->cg_list);
6696 	spin_unlock_irq(&css_set_lock);
6697 }
6698 
cgroup_free(struct task_struct * task)6699 void cgroup_free(struct task_struct *task)
6700 {
6701 	struct css_set *cset = task_css_set(task);
6702 	put_css_set(cset);
6703 }
6704 
cgroup_disable(char * str)6705 static int __init cgroup_disable(char *str)
6706 {
6707 	struct cgroup_subsys *ss;
6708 	char *token;
6709 	int i;
6710 
6711 	while ((token = strsep(&str, ",")) != NULL) {
6712 		if (!*token)
6713 			continue;
6714 
6715 		for_each_subsys(ss, i) {
6716 			if (strcmp(token, ss->name) &&
6717 			    strcmp(token, ss->legacy_name))
6718 				continue;
6719 
6720 			static_branch_disable(cgroup_subsys_enabled_key[i]);
6721 			pr_info("Disabling %s control group subsystem\n",
6722 				ss->name);
6723 		}
6724 
6725 		for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6726 			if (strcmp(token, cgroup_opt_feature_names[i]))
6727 				continue;
6728 			cgroup_feature_disable_mask |= 1 << i;
6729 			pr_info("Disabling %s control group feature\n",
6730 				cgroup_opt_feature_names[i]);
6731 			break;
6732 		}
6733 	}
6734 	return 1;
6735 }
6736 __setup("cgroup_disable=", cgroup_disable);
6737 
enable_debug_cgroup(void)6738 void __init __weak enable_debug_cgroup(void) { }
6739 
enable_cgroup_debug(char * str)6740 static int __init enable_cgroup_debug(char *str)
6741 {
6742 	cgroup_debug = true;
6743 	enable_debug_cgroup();
6744 	return 1;
6745 }
6746 __setup("cgroup_debug", enable_cgroup_debug);
6747 
6748 /**
6749  * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6750  * @dentry: directory dentry of interest
6751  * @ss: subsystem of interest
6752  *
6753  * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6754  * to get the corresponding css and return it.  If such css doesn't exist
6755  * or can't be pinned, an ERR_PTR value is returned.
6756  */
css_tryget_online_from_dir(struct dentry * dentry,struct cgroup_subsys * ss)6757 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6758 						       struct cgroup_subsys *ss)
6759 {
6760 	struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6761 	struct file_system_type *s_type = dentry->d_sb->s_type;
6762 	struct cgroup_subsys_state *css = NULL;
6763 	struct cgroup *cgrp;
6764 
6765 	/* is @dentry a cgroup dir? */
6766 	if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6767 	    !kn || kernfs_type(kn) != KERNFS_DIR)
6768 		return ERR_PTR(-EBADF);
6769 
6770 	rcu_read_lock();
6771 
6772 	/*
6773 	 * This path doesn't originate from kernfs and @kn could already
6774 	 * have been or be removed at any point.  @kn->priv is RCU
6775 	 * protected for this access.  See css_release_work_fn() for details.
6776 	 */
6777 	cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6778 	if (cgrp)
6779 		css = cgroup_css(cgrp, ss);
6780 
6781 	if (!css || !css_tryget_online(css))
6782 		css = ERR_PTR(-ENOENT);
6783 
6784 	rcu_read_unlock();
6785 	return css;
6786 }
6787 
6788 /**
6789  * css_from_id - lookup css by id
6790  * @id: the cgroup id
6791  * @ss: cgroup subsys to be looked into
6792  *
6793  * Returns the css if there's valid one with @id, otherwise returns NULL.
6794  * Should be called under rcu_read_lock().
6795  */
css_from_id(int id,struct cgroup_subsys * ss)6796 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6797 {
6798 	WARN_ON_ONCE(!rcu_read_lock_held());
6799 	return idr_find(&ss->css_idr, id);
6800 }
6801 
6802 /**
6803  * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6804  * @path: path on the default hierarchy
6805  *
6806  * Find the cgroup at @path on the default hierarchy, increment its
6807  * reference count and return it.  Returns pointer to the found cgroup on
6808  * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6809  * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6810  */
cgroup_get_from_path(const char * path)6811 struct cgroup *cgroup_get_from_path(const char *path)
6812 {
6813 	struct kernfs_node *kn;
6814 	struct cgroup *cgrp = ERR_PTR(-ENOENT);
6815 	struct cgroup *root_cgrp;
6816 
6817 	root_cgrp = current_cgns_cgroup_dfl();
6818 	kn = kernfs_walk_and_get(root_cgrp->kn, path);
6819 	if (!kn)
6820 		goto out;
6821 
6822 	if (kernfs_type(kn) != KERNFS_DIR) {
6823 		cgrp = ERR_PTR(-ENOTDIR);
6824 		goto out_kernfs;
6825 	}
6826 
6827 	rcu_read_lock();
6828 
6829 	cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6830 	if (!cgrp || !cgroup_tryget(cgrp))
6831 		cgrp = ERR_PTR(-ENOENT);
6832 
6833 	rcu_read_unlock();
6834 
6835 out_kernfs:
6836 	kernfs_put(kn);
6837 out:
6838 	return cgrp;
6839 }
6840 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6841 
6842 /**
6843  * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
6844  * @fd: fd obtained by open(cgroup_dir)
6845  *
6846  * Find the cgroup from a fd which should be obtained
6847  * by opening a cgroup directory.  Returns a pointer to the
6848  * cgroup on success. ERR_PTR is returned if the cgroup
6849  * cannot be found.
6850  */
cgroup_v1v2_get_from_fd(int fd)6851 struct cgroup *cgroup_v1v2_get_from_fd(int fd)
6852 {
6853 	struct cgroup *cgrp;
6854 	struct file *f;
6855 
6856 	f = fget_raw(fd);
6857 	if (!f)
6858 		return ERR_PTR(-EBADF);
6859 
6860 	cgrp = cgroup_v1v2_get_from_file(f);
6861 	fput(f);
6862 	return cgrp;
6863 }
6864 
6865 /**
6866  * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
6867  * cgroup2.
6868  * @fd: fd obtained by open(cgroup2_dir)
6869  */
cgroup_get_from_fd(int fd)6870 struct cgroup *cgroup_get_from_fd(int fd)
6871 {
6872 	struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
6873 
6874 	if (IS_ERR(cgrp))
6875 		return ERR_CAST(cgrp);
6876 
6877 	if (!cgroup_on_dfl(cgrp)) {
6878 		cgroup_put(cgrp);
6879 		return ERR_PTR(-EBADF);
6880 	}
6881 	return cgrp;
6882 }
6883 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6884 
power_of_ten(int power)6885 static u64 power_of_ten(int power)
6886 {
6887 	u64 v = 1;
6888 	while (power--)
6889 		v *= 10;
6890 	return v;
6891 }
6892 
6893 /**
6894  * cgroup_parse_float - parse a floating number
6895  * @input: input string
6896  * @dec_shift: number of decimal digits to shift
6897  * @v: output
6898  *
6899  * Parse a decimal floating point number in @input and store the result in
6900  * @v with decimal point right shifted @dec_shift times.  For example, if
6901  * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6902  * Returns 0 on success, -errno otherwise.
6903  *
6904  * There's nothing cgroup specific about this function except that it's
6905  * currently the only user.
6906  */
cgroup_parse_float(const char * input,unsigned dec_shift,s64 * v)6907 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6908 {
6909 	s64 whole, frac = 0;
6910 	int fstart = 0, fend = 0, flen;
6911 
6912 	if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6913 		return -EINVAL;
6914 	if (frac < 0)
6915 		return -EINVAL;
6916 
6917 	flen = fend > fstart ? fend - fstart : 0;
6918 	if (flen < dec_shift)
6919 		frac *= power_of_ten(dec_shift - flen);
6920 	else
6921 		frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6922 
6923 	*v = whole * power_of_ten(dec_shift) + frac;
6924 	return 0;
6925 }
6926 
6927 /*
6928  * sock->sk_cgrp_data handling.  For more info, see sock_cgroup_data
6929  * definition in cgroup-defs.h.
6930  */
6931 #ifdef CONFIG_SOCK_CGROUP_DATA
6932 
cgroup_sk_alloc(struct sock_cgroup_data * skcd)6933 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6934 {
6935 	struct cgroup *cgroup;
6936 
6937 	rcu_read_lock();
6938 	/* Don't associate the sock with unrelated interrupted task's cgroup. */
6939 	if (in_interrupt()) {
6940 		cgroup = &cgrp_dfl_root.cgrp;
6941 		cgroup_get(cgroup);
6942 		goto out;
6943 	}
6944 
6945 	while (true) {
6946 		struct css_set *cset;
6947 
6948 		cset = task_css_set(current);
6949 		if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6950 			cgroup = cset->dfl_cgrp;
6951 			break;
6952 		}
6953 		cpu_relax();
6954 	}
6955 out:
6956 	skcd->cgroup = cgroup;
6957 	cgroup_bpf_get(cgroup);
6958 	rcu_read_unlock();
6959 }
6960 
cgroup_sk_clone(struct sock_cgroup_data * skcd)6961 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6962 {
6963 	struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6964 
6965 	/*
6966 	 * We might be cloning a socket which is left in an empty
6967 	 * cgroup and the cgroup might have already been rmdir'd.
6968 	 * Don't use cgroup_get_live().
6969 	 */
6970 	cgroup_get(cgrp);
6971 	cgroup_bpf_get(cgrp);
6972 }
6973 
cgroup_sk_free(struct sock_cgroup_data * skcd)6974 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6975 {
6976 	struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6977 
6978 	cgroup_bpf_put(cgrp);
6979 	cgroup_put(cgrp);
6980 }
6981 
6982 #endif	/* CONFIG_SOCK_CGROUP_DATA */
6983 
6984 #ifdef CONFIG_SYSFS
show_delegatable_files(struct cftype * files,char * buf,ssize_t size,const char * prefix)6985 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6986 				      ssize_t size, const char *prefix)
6987 {
6988 	struct cftype *cft;
6989 	ssize_t ret = 0;
6990 
6991 	for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6992 		if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6993 			continue;
6994 
6995 		if (prefix)
6996 			ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6997 
6998 		ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6999 
7000 		if (WARN_ON(ret >= size))
7001 			break;
7002 	}
7003 
7004 	return ret;
7005 }
7006 
delegate_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)7007 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7008 			      char *buf)
7009 {
7010 	struct cgroup_subsys *ss;
7011 	int ssid;
7012 	ssize_t ret = 0;
7013 
7014 	ret = show_delegatable_files(cgroup_base_files, buf + ret,
7015 				     PAGE_SIZE - ret, NULL);
7016 	if (cgroup_psi_enabled())
7017 		ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7018 					      PAGE_SIZE - ret, NULL);
7019 
7020 	for_each_subsys(ss, ssid)
7021 		ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7022 					      PAGE_SIZE - ret,
7023 					      cgroup_subsys_name[ssid]);
7024 
7025 	return ret;
7026 }
7027 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7028 
features_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)7029 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7030 			     char *buf)
7031 {
7032 	return snprintf(buf, PAGE_SIZE,
7033 			"nsdelegate\n"
7034 			"favordynmods\n"
7035 			"memory_localevents\n"
7036 			"memory_recursiveprot\n");
7037 }
7038 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7039 
7040 static struct attribute *cgroup_sysfs_attrs[] = {
7041 	&cgroup_delegate_attr.attr,
7042 	&cgroup_features_attr.attr,
7043 	NULL,
7044 };
7045 
7046 static const struct attribute_group cgroup_sysfs_attr_group = {
7047 	.attrs = cgroup_sysfs_attrs,
7048 	.name = "cgroup",
7049 };
7050 
cgroup_sysfs_init(void)7051 static int __init cgroup_sysfs_init(void)
7052 {
7053 	return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7054 }
7055 subsys_initcall(cgroup_sysfs_init);
7056 
7057 #endif /* CONFIG_SYSFS */
7058