xref: /openbmc/linux/mm/memcontrol.c (revision d7a3d85e)
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
13  * Kernel Memory Controller
14  * Copyright (C) 2012 Parallels Inc. and Google Inc.
15  * Authors: Glauber Costa and Suleiman Souhlal
16  *
17  * Native page reclaim
18  * Charge lifetime sanitation
19  * Lockless page tracking & accounting
20  * Unified hierarchy configuration model
21  * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
22  *
23  * This program is free software; you can redistribute it and/or modify
24  * it under the terms of the GNU General Public License as published by
25  * the Free Software Foundation; either version 2 of the License, or
26  * (at your option) any later version.
27  *
28  * This program is distributed in the hope that it will be useful,
29  * but WITHOUT ANY WARRANTY; without even the implied warranty of
30  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
31  * GNU General Public License for more details.
32  */
33 
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
37 #include <linux/mm.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
56 #include <linux/fs.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include "internal.h"
66 #include <net/sock.h>
67 #include <net/ip.h>
68 #include <net/tcp_memcontrol.h>
69 #include "slab.h"
70 
71 #include <asm/uaccess.h>
72 
73 #include <trace/events/vmscan.h>
74 
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
77 
78 #define MEM_CGROUP_RECLAIM_RETRIES	5
79 static struct mem_cgroup *root_mem_cgroup __read_mostly;
80 
81 /* Whether the swap controller is active */
82 #ifdef CONFIG_MEMCG_SWAP
83 int do_swap_account __read_mostly;
84 #else
85 #define do_swap_account		0
86 #endif
87 
88 static const char * const mem_cgroup_stat_names[] = {
89 	"cache",
90 	"rss",
91 	"rss_huge",
92 	"mapped_file",
93 	"writeback",
94 	"swap",
95 };
96 
97 static const char * const mem_cgroup_events_names[] = {
98 	"pgpgin",
99 	"pgpgout",
100 	"pgfault",
101 	"pgmajfault",
102 };
103 
104 static const char * const mem_cgroup_lru_names[] = {
105 	"inactive_anon",
106 	"active_anon",
107 	"inactive_file",
108 	"active_file",
109 	"unevictable",
110 };
111 
112 /*
113  * Per memcg event counter is incremented at every pagein/pageout. With THP,
114  * it will be incremated by the number of pages. This counter is used for
115  * for trigger some periodic events. This is straightforward and better
116  * than using jiffies etc. to handle periodic memcg event.
117  */
118 enum mem_cgroup_events_target {
119 	MEM_CGROUP_TARGET_THRESH,
120 	MEM_CGROUP_TARGET_SOFTLIMIT,
121 	MEM_CGROUP_TARGET_NUMAINFO,
122 	MEM_CGROUP_NTARGETS,
123 };
124 #define THRESHOLDS_EVENTS_TARGET 128
125 #define SOFTLIMIT_EVENTS_TARGET 1024
126 #define NUMAINFO_EVENTS_TARGET	1024
127 
128 struct mem_cgroup_stat_cpu {
129 	long count[MEM_CGROUP_STAT_NSTATS];
130 	unsigned long events[MEMCG_NR_EVENTS];
131 	unsigned long nr_page_events;
132 	unsigned long targets[MEM_CGROUP_NTARGETS];
133 };
134 
135 struct reclaim_iter {
136 	struct mem_cgroup *position;
137 	/* scan generation, increased every round-trip */
138 	unsigned int generation;
139 };
140 
141 /*
142  * per-zone information in memory controller.
143  */
144 struct mem_cgroup_per_zone {
145 	struct lruvec		lruvec;
146 	unsigned long		lru_size[NR_LRU_LISTS];
147 
148 	struct reclaim_iter	iter[DEF_PRIORITY + 1];
149 
150 	struct rb_node		tree_node;	/* RB tree node */
151 	unsigned long		usage_in_excess;/* Set to the value by which */
152 						/* the soft limit is exceeded*/
153 	bool			on_tree;
154 	struct mem_cgroup	*memcg;		/* Back pointer, we cannot */
155 						/* use container_of	   */
156 };
157 
158 struct mem_cgroup_per_node {
159 	struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
160 };
161 
162 /*
163  * Cgroups above their limits are maintained in a RB-Tree, independent of
164  * their hierarchy representation
165  */
166 
167 struct mem_cgroup_tree_per_zone {
168 	struct rb_root rb_root;
169 	spinlock_t lock;
170 };
171 
172 struct mem_cgroup_tree_per_node {
173 	struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
174 };
175 
176 struct mem_cgroup_tree {
177 	struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
178 };
179 
180 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
181 
182 struct mem_cgroup_threshold {
183 	struct eventfd_ctx *eventfd;
184 	unsigned long threshold;
185 };
186 
187 /* For threshold */
188 struct mem_cgroup_threshold_ary {
189 	/* An array index points to threshold just below or equal to usage. */
190 	int current_threshold;
191 	/* Size of entries[] */
192 	unsigned int size;
193 	/* Array of thresholds */
194 	struct mem_cgroup_threshold entries[0];
195 };
196 
197 struct mem_cgroup_thresholds {
198 	/* Primary thresholds array */
199 	struct mem_cgroup_threshold_ary *primary;
200 	/*
201 	 * Spare threshold array.
202 	 * This is needed to make mem_cgroup_unregister_event() "never fail".
203 	 * It must be able to store at least primary->size - 1 entries.
204 	 */
205 	struct mem_cgroup_threshold_ary *spare;
206 };
207 
208 /* for OOM */
209 struct mem_cgroup_eventfd_list {
210 	struct list_head list;
211 	struct eventfd_ctx *eventfd;
212 };
213 
214 /*
215  * cgroup_event represents events which userspace want to receive.
216  */
217 struct mem_cgroup_event {
218 	/*
219 	 * memcg which the event belongs to.
220 	 */
221 	struct mem_cgroup *memcg;
222 	/*
223 	 * eventfd to signal userspace about the event.
224 	 */
225 	struct eventfd_ctx *eventfd;
226 	/*
227 	 * Each of these stored in a list by the cgroup.
228 	 */
229 	struct list_head list;
230 	/*
231 	 * register_event() callback will be used to add new userspace
232 	 * waiter for changes related to this event.  Use eventfd_signal()
233 	 * on eventfd to send notification to userspace.
234 	 */
235 	int (*register_event)(struct mem_cgroup *memcg,
236 			      struct eventfd_ctx *eventfd, const char *args);
237 	/*
238 	 * unregister_event() callback will be called when userspace closes
239 	 * the eventfd or on cgroup removing.  This callback must be set,
240 	 * if you want provide notification functionality.
241 	 */
242 	void (*unregister_event)(struct mem_cgroup *memcg,
243 				 struct eventfd_ctx *eventfd);
244 	/*
245 	 * All fields below needed to unregister event when
246 	 * userspace closes eventfd.
247 	 */
248 	poll_table pt;
249 	wait_queue_head_t *wqh;
250 	wait_queue_t wait;
251 	struct work_struct remove;
252 };
253 
254 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
255 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
256 
257 /*
258  * The memory controller data structure. The memory controller controls both
259  * page cache and RSS per cgroup. We would eventually like to provide
260  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
261  * to help the administrator determine what knobs to tune.
262  */
263 struct mem_cgroup {
264 	struct cgroup_subsys_state css;
265 
266 	/* Accounted resources */
267 	struct page_counter memory;
268 	struct page_counter memsw;
269 	struct page_counter kmem;
270 
271 	/* Normal memory consumption range */
272 	unsigned long low;
273 	unsigned long high;
274 
275 	unsigned long soft_limit;
276 
277 	/* vmpressure notifications */
278 	struct vmpressure vmpressure;
279 
280 	/* css_online() has been completed */
281 	int initialized;
282 
283 	/*
284 	 * Should the accounting and control be hierarchical, per subtree?
285 	 */
286 	bool use_hierarchy;
287 
288 	bool		oom_lock;
289 	atomic_t	under_oom;
290 	atomic_t	oom_wakeups;
291 
292 	int	swappiness;
293 	/* OOM-Killer disable */
294 	int		oom_kill_disable;
295 
296 	/* protect arrays of thresholds */
297 	struct mutex thresholds_lock;
298 
299 	/* thresholds for memory usage. RCU-protected */
300 	struct mem_cgroup_thresholds thresholds;
301 
302 	/* thresholds for mem+swap usage. RCU-protected */
303 	struct mem_cgroup_thresholds memsw_thresholds;
304 
305 	/* For oom notifier event fd */
306 	struct list_head oom_notify;
307 
308 	/*
309 	 * Should we move charges of a task when a task is moved into this
310 	 * mem_cgroup ? And what type of charges should we move ?
311 	 */
312 	unsigned long move_charge_at_immigrate;
313 	/*
314 	 * set > 0 if pages under this cgroup are moving to other cgroup.
315 	 */
316 	atomic_t		moving_account;
317 	/* taken only while moving_account > 0 */
318 	spinlock_t		move_lock;
319 	struct task_struct	*move_lock_task;
320 	unsigned long		move_lock_flags;
321 	/*
322 	 * percpu counter.
323 	 */
324 	struct mem_cgroup_stat_cpu __percpu *stat;
325 	/*
326 	 * used when a cpu is offlined or other synchronizations
327 	 * See mem_cgroup_read_stat().
328 	 */
329 	struct mem_cgroup_stat_cpu nocpu_base;
330 	spinlock_t pcp_counter_lock;
331 
332 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
333 	struct cg_proto tcp_mem;
334 #endif
335 #if defined(CONFIG_MEMCG_KMEM)
336         /* Index in the kmem_cache->memcg_params.memcg_caches array */
337 	int kmemcg_id;
338 	bool kmem_acct_activated;
339 	bool kmem_acct_active;
340 #endif
341 
342 	int last_scanned_node;
343 #if MAX_NUMNODES > 1
344 	nodemask_t	scan_nodes;
345 	atomic_t	numainfo_events;
346 	atomic_t	numainfo_updating;
347 #endif
348 
349 	/* List of events which userspace want to receive */
350 	struct list_head event_list;
351 	spinlock_t event_list_lock;
352 
353 	struct mem_cgroup_per_node *nodeinfo[0];
354 	/* WARNING: nodeinfo must be the last member here */
355 };
356 
357 #ifdef CONFIG_MEMCG_KMEM
358 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
359 {
360 	return memcg->kmem_acct_active;
361 }
362 #endif
363 
364 /* Stuffs for move charges at task migration. */
365 /*
366  * Types of charges to be moved.
367  */
368 #define MOVE_ANON	0x1U
369 #define MOVE_FILE	0x2U
370 #define MOVE_MASK	(MOVE_ANON | MOVE_FILE)
371 
372 /* "mc" and its members are protected by cgroup_mutex */
373 static struct move_charge_struct {
374 	spinlock_t	  lock; /* for from, to */
375 	struct mem_cgroup *from;
376 	struct mem_cgroup *to;
377 	unsigned long flags;
378 	unsigned long precharge;
379 	unsigned long moved_charge;
380 	unsigned long moved_swap;
381 	struct task_struct *moving_task;	/* a task moving charges */
382 	wait_queue_head_t waitq;		/* a waitq for other context */
383 } mc = {
384 	.lock = __SPIN_LOCK_UNLOCKED(mc.lock),
385 	.waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
386 };
387 
388 /*
389  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
390  * limit reclaim to prevent infinite loops, if they ever occur.
391  */
392 #define	MEM_CGROUP_MAX_RECLAIM_LOOPS		100
393 #define	MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS	2
394 
395 enum charge_type {
396 	MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
397 	MEM_CGROUP_CHARGE_TYPE_ANON,
398 	MEM_CGROUP_CHARGE_TYPE_SWAPOUT,	/* for accounting swapcache */
399 	MEM_CGROUP_CHARGE_TYPE_DROP,	/* a page was unused swap cache */
400 	NR_CHARGE_TYPE,
401 };
402 
403 /* for encoding cft->private value on file */
404 enum res_type {
405 	_MEM,
406 	_MEMSWAP,
407 	_OOM_TYPE,
408 	_KMEM,
409 };
410 
411 #define MEMFILE_PRIVATE(x, val)	((x) << 16 | (val))
412 #define MEMFILE_TYPE(val)	((val) >> 16 & 0xffff)
413 #define MEMFILE_ATTR(val)	((val) & 0xffff)
414 /* Used for OOM nofiier */
415 #define OOM_CONTROL		(0)
416 
417 /*
418  * The memcg_create_mutex will be held whenever a new cgroup is created.
419  * As a consequence, any change that needs to protect against new child cgroups
420  * appearing has to hold it as well.
421  */
422 static DEFINE_MUTEX(memcg_create_mutex);
423 
424 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
425 {
426 	return s ? container_of(s, struct mem_cgroup, css) : NULL;
427 }
428 
429 /* Some nice accessors for the vmpressure. */
430 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
431 {
432 	if (!memcg)
433 		memcg = root_mem_cgroup;
434 	return &memcg->vmpressure;
435 }
436 
437 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
438 {
439 	return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
440 }
441 
442 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
443 {
444 	return (memcg == root_mem_cgroup);
445 }
446 
447 /*
448  * We restrict the id in the range of [1, 65535], so it can fit into
449  * an unsigned short.
450  */
451 #define MEM_CGROUP_ID_MAX	USHRT_MAX
452 
453 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
454 {
455 	return memcg->css.id;
456 }
457 
458 /*
459  * A helper function to get mem_cgroup from ID. must be called under
460  * rcu_read_lock().  The caller is responsible for calling
461  * css_tryget_online() if the mem_cgroup is used for charging. (dropping
462  * refcnt from swap can be called against removed memcg.)
463  */
464 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
465 {
466 	struct cgroup_subsys_state *css;
467 
468 	css = css_from_id(id, &memory_cgrp_subsys);
469 	return mem_cgroup_from_css(css);
470 }
471 
472 /* Writing them here to avoid exposing memcg's inner layout */
473 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
474 
475 void sock_update_memcg(struct sock *sk)
476 {
477 	if (mem_cgroup_sockets_enabled) {
478 		struct mem_cgroup *memcg;
479 		struct cg_proto *cg_proto;
480 
481 		BUG_ON(!sk->sk_prot->proto_cgroup);
482 
483 		/* Socket cloning can throw us here with sk_cgrp already
484 		 * filled. It won't however, necessarily happen from
485 		 * process context. So the test for root memcg given
486 		 * the current task's memcg won't help us in this case.
487 		 *
488 		 * Respecting the original socket's memcg is a better
489 		 * decision in this case.
490 		 */
491 		if (sk->sk_cgrp) {
492 			BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
493 			css_get(&sk->sk_cgrp->memcg->css);
494 			return;
495 		}
496 
497 		rcu_read_lock();
498 		memcg = mem_cgroup_from_task(current);
499 		cg_proto = sk->sk_prot->proto_cgroup(memcg);
500 		if (!mem_cgroup_is_root(memcg) &&
501 		    memcg_proto_active(cg_proto) &&
502 		    css_tryget_online(&memcg->css)) {
503 			sk->sk_cgrp = cg_proto;
504 		}
505 		rcu_read_unlock();
506 	}
507 }
508 EXPORT_SYMBOL(sock_update_memcg);
509 
510 void sock_release_memcg(struct sock *sk)
511 {
512 	if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
513 		struct mem_cgroup *memcg;
514 		WARN_ON(!sk->sk_cgrp->memcg);
515 		memcg = sk->sk_cgrp->memcg;
516 		css_put(&sk->sk_cgrp->memcg->css);
517 	}
518 }
519 
520 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
521 {
522 	if (!memcg || mem_cgroup_is_root(memcg))
523 		return NULL;
524 
525 	return &memcg->tcp_mem;
526 }
527 EXPORT_SYMBOL(tcp_proto_cgroup);
528 
529 #endif
530 
531 #ifdef CONFIG_MEMCG_KMEM
532 /*
533  * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
534  * The main reason for not using cgroup id for this:
535  *  this works better in sparse environments, where we have a lot of memcgs,
536  *  but only a few kmem-limited. Or also, if we have, for instance, 200
537  *  memcgs, and none but the 200th is kmem-limited, we'd have to have a
538  *  200 entry array for that.
539  *
540  * The current size of the caches array is stored in memcg_nr_cache_ids. It
541  * will double each time we have to increase it.
542  */
543 static DEFINE_IDA(memcg_cache_ida);
544 int memcg_nr_cache_ids;
545 
546 /* Protects memcg_nr_cache_ids */
547 static DECLARE_RWSEM(memcg_cache_ids_sem);
548 
549 void memcg_get_cache_ids(void)
550 {
551 	down_read(&memcg_cache_ids_sem);
552 }
553 
554 void memcg_put_cache_ids(void)
555 {
556 	up_read(&memcg_cache_ids_sem);
557 }
558 
559 /*
560  * MIN_SIZE is different than 1, because we would like to avoid going through
561  * the alloc/free process all the time. In a small machine, 4 kmem-limited
562  * cgroups is a reasonable guess. In the future, it could be a parameter or
563  * tunable, but that is strictly not necessary.
564  *
565  * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
566  * this constant directly from cgroup, but it is understandable that this is
567  * better kept as an internal representation in cgroup.c. In any case, the
568  * cgrp_id space is not getting any smaller, and we don't have to necessarily
569  * increase ours as well if it increases.
570  */
571 #define MEMCG_CACHES_MIN_SIZE 4
572 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
573 
574 /*
575  * A lot of the calls to the cache allocation functions are expected to be
576  * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
577  * conditional to this static branch, we'll have to allow modules that does
578  * kmem_cache_alloc and the such to see this symbol as well
579  */
580 struct static_key memcg_kmem_enabled_key;
581 EXPORT_SYMBOL(memcg_kmem_enabled_key);
582 
583 #endif /* CONFIG_MEMCG_KMEM */
584 
585 static struct mem_cgroup_per_zone *
586 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
587 {
588 	int nid = zone_to_nid(zone);
589 	int zid = zone_idx(zone);
590 
591 	return &memcg->nodeinfo[nid]->zoneinfo[zid];
592 }
593 
594 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
595 {
596 	return &memcg->css;
597 }
598 
599 static struct mem_cgroup_per_zone *
600 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
601 {
602 	int nid = page_to_nid(page);
603 	int zid = page_zonenum(page);
604 
605 	return &memcg->nodeinfo[nid]->zoneinfo[zid];
606 }
607 
608 static struct mem_cgroup_tree_per_zone *
609 soft_limit_tree_node_zone(int nid, int zid)
610 {
611 	return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
612 }
613 
614 static struct mem_cgroup_tree_per_zone *
615 soft_limit_tree_from_page(struct page *page)
616 {
617 	int nid = page_to_nid(page);
618 	int zid = page_zonenum(page);
619 
620 	return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
621 }
622 
623 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
624 					 struct mem_cgroup_tree_per_zone *mctz,
625 					 unsigned long new_usage_in_excess)
626 {
627 	struct rb_node **p = &mctz->rb_root.rb_node;
628 	struct rb_node *parent = NULL;
629 	struct mem_cgroup_per_zone *mz_node;
630 
631 	if (mz->on_tree)
632 		return;
633 
634 	mz->usage_in_excess = new_usage_in_excess;
635 	if (!mz->usage_in_excess)
636 		return;
637 	while (*p) {
638 		parent = *p;
639 		mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
640 					tree_node);
641 		if (mz->usage_in_excess < mz_node->usage_in_excess)
642 			p = &(*p)->rb_left;
643 		/*
644 		 * We can't avoid mem cgroups that are over their soft
645 		 * limit by the same amount
646 		 */
647 		else if (mz->usage_in_excess >= mz_node->usage_in_excess)
648 			p = &(*p)->rb_right;
649 	}
650 	rb_link_node(&mz->tree_node, parent, p);
651 	rb_insert_color(&mz->tree_node, &mctz->rb_root);
652 	mz->on_tree = true;
653 }
654 
655 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
656 					 struct mem_cgroup_tree_per_zone *mctz)
657 {
658 	if (!mz->on_tree)
659 		return;
660 	rb_erase(&mz->tree_node, &mctz->rb_root);
661 	mz->on_tree = false;
662 }
663 
664 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
665 				       struct mem_cgroup_tree_per_zone *mctz)
666 {
667 	unsigned long flags;
668 
669 	spin_lock_irqsave(&mctz->lock, flags);
670 	__mem_cgroup_remove_exceeded(mz, mctz);
671 	spin_unlock_irqrestore(&mctz->lock, flags);
672 }
673 
674 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
675 {
676 	unsigned long nr_pages = page_counter_read(&memcg->memory);
677 	unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
678 	unsigned long excess = 0;
679 
680 	if (nr_pages > soft_limit)
681 		excess = nr_pages - soft_limit;
682 
683 	return excess;
684 }
685 
686 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
687 {
688 	unsigned long excess;
689 	struct mem_cgroup_per_zone *mz;
690 	struct mem_cgroup_tree_per_zone *mctz;
691 
692 	mctz = soft_limit_tree_from_page(page);
693 	/*
694 	 * Necessary to update all ancestors when hierarchy is used.
695 	 * because their event counter is not touched.
696 	 */
697 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
698 		mz = mem_cgroup_page_zoneinfo(memcg, page);
699 		excess = soft_limit_excess(memcg);
700 		/*
701 		 * We have to update the tree if mz is on RB-tree or
702 		 * mem is over its softlimit.
703 		 */
704 		if (excess || mz->on_tree) {
705 			unsigned long flags;
706 
707 			spin_lock_irqsave(&mctz->lock, flags);
708 			/* if on-tree, remove it */
709 			if (mz->on_tree)
710 				__mem_cgroup_remove_exceeded(mz, mctz);
711 			/*
712 			 * Insert again. mz->usage_in_excess will be updated.
713 			 * If excess is 0, no tree ops.
714 			 */
715 			__mem_cgroup_insert_exceeded(mz, mctz, excess);
716 			spin_unlock_irqrestore(&mctz->lock, flags);
717 		}
718 	}
719 }
720 
721 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
722 {
723 	struct mem_cgroup_tree_per_zone *mctz;
724 	struct mem_cgroup_per_zone *mz;
725 	int nid, zid;
726 
727 	for_each_node(nid) {
728 		for (zid = 0; zid < MAX_NR_ZONES; zid++) {
729 			mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
730 			mctz = soft_limit_tree_node_zone(nid, zid);
731 			mem_cgroup_remove_exceeded(mz, mctz);
732 		}
733 	}
734 }
735 
736 static struct mem_cgroup_per_zone *
737 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
738 {
739 	struct rb_node *rightmost = NULL;
740 	struct mem_cgroup_per_zone *mz;
741 
742 retry:
743 	mz = NULL;
744 	rightmost = rb_last(&mctz->rb_root);
745 	if (!rightmost)
746 		goto done;		/* Nothing to reclaim from */
747 
748 	mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
749 	/*
750 	 * Remove the node now but someone else can add it back,
751 	 * we will to add it back at the end of reclaim to its correct
752 	 * position in the tree.
753 	 */
754 	__mem_cgroup_remove_exceeded(mz, mctz);
755 	if (!soft_limit_excess(mz->memcg) ||
756 	    !css_tryget_online(&mz->memcg->css))
757 		goto retry;
758 done:
759 	return mz;
760 }
761 
762 static struct mem_cgroup_per_zone *
763 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
764 {
765 	struct mem_cgroup_per_zone *mz;
766 
767 	spin_lock_irq(&mctz->lock);
768 	mz = __mem_cgroup_largest_soft_limit_node(mctz);
769 	spin_unlock_irq(&mctz->lock);
770 	return mz;
771 }
772 
773 /*
774  * Implementation Note: reading percpu statistics for memcg.
775  *
776  * Both of vmstat[] and percpu_counter has threshold and do periodic
777  * synchronization to implement "quick" read. There are trade-off between
778  * reading cost and precision of value. Then, we may have a chance to implement
779  * a periodic synchronizion of counter in memcg's counter.
780  *
781  * But this _read() function is used for user interface now. The user accounts
782  * memory usage by memory cgroup and he _always_ requires exact value because
783  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
784  * have to visit all online cpus and make sum. So, for now, unnecessary
785  * synchronization is not implemented. (just implemented for cpu hotplug)
786  *
787  * If there are kernel internal actions which can make use of some not-exact
788  * value, and reading all cpu value can be performance bottleneck in some
789  * common workload, threashold and synchonization as vmstat[] should be
790  * implemented.
791  */
792 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
793 				 enum mem_cgroup_stat_index idx)
794 {
795 	long val = 0;
796 	int cpu;
797 
798 	get_online_cpus();
799 	for_each_online_cpu(cpu)
800 		val += per_cpu(memcg->stat->count[idx], cpu);
801 #ifdef CONFIG_HOTPLUG_CPU
802 	spin_lock(&memcg->pcp_counter_lock);
803 	val += memcg->nocpu_base.count[idx];
804 	spin_unlock(&memcg->pcp_counter_lock);
805 #endif
806 	put_online_cpus();
807 	return val;
808 }
809 
810 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
811 					    enum mem_cgroup_events_index idx)
812 {
813 	unsigned long val = 0;
814 	int cpu;
815 
816 	get_online_cpus();
817 	for_each_online_cpu(cpu)
818 		val += per_cpu(memcg->stat->events[idx], cpu);
819 #ifdef CONFIG_HOTPLUG_CPU
820 	spin_lock(&memcg->pcp_counter_lock);
821 	val += memcg->nocpu_base.events[idx];
822 	spin_unlock(&memcg->pcp_counter_lock);
823 #endif
824 	put_online_cpus();
825 	return val;
826 }
827 
828 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
829 					 struct page *page,
830 					 int nr_pages)
831 {
832 	/*
833 	 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
834 	 * counted as CACHE even if it's on ANON LRU.
835 	 */
836 	if (PageAnon(page))
837 		__this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
838 				nr_pages);
839 	else
840 		__this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
841 				nr_pages);
842 
843 	if (PageTransHuge(page))
844 		__this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
845 				nr_pages);
846 
847 	/* pagein of a big page is an event. So, ignore page size */
848 	if (nr_pages > 0)
849 		__this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
850 	else {
851 		__this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
852 		nr_pages = -nr_pages; /* for event */
853 	}
854 
855 	__this_cpu_add(memcg->stat->nr_page_events, nr_pages);
856 }
857 
858 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
859 {
860 	struct mem_cgroup_per_zone *mz;
861 
862 	mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
863 	return mz->lru_size[lru];
864 }
865 
866 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
867 						  int nid,
868 						  unsigned int lru_mask)
869 {
870 	unsigned long nr = 0;
871 	int zid;
872 
873 	VM_BUG_ON((unsigned)nid >= nr_node_ids);
874 
875 	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
876 		struct mem_cgroup_per_zone *mz;
877 		enum lru_list lru;
878 
879 		for_each_lru(lru) {
880 			if (!(BIT(lru) & lru_mask))
881 				continue;
882 			mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
883 			nr += mz->lru_size[lru];
884 		}
885 	}
886 	return nr;
887 }
888 
889 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
890 			unsigned int lru_mask)
891 {
892 	unsigned long nr = 0;
893 	int nid;
894 
895 	for_each_node_state(nid, N_MEMORY)
896 		nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
897 	return nr;
898 }
899 
900 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
901 				       enum mem_cgroup_events_target target)
902 {
903 	unsigned long val, next;
904 
905 	val = __this_cpu_read(memcg->stat->nr_page_events);
906 	next = __this_cpu_read(memcg->stat->targets[target]);
907 	/* from time_after() in jiffies.h */
908 	if ((long)next - (long)val < 0) {
909 		switch (target) {
910 		case MEM_CGROUP_TARGET_THRESH:
911 			next = val + THRESHOLDS_EVENTS_TARGET;
912 			break;
913 		case MEM_CGROUP_TARGET_SOFTLIMIT:
914 			next = val + SOFTLIMIT_EVENTS_TARGET;
915 			break;
916 		case MEM_CGROUP_TARGET_NUMAINFO:
917 			next = val + NUMAINFO_EVENTS_TARGET;
918 			break;
919 		default:
920 			break;
921 		}
922 		__this_cpu_write(memcg->stat->targets[target], next);
923 		return true;
924 	}
925 	return false;
926 }
927 
928 /*
929  * Check events in order.
930  *
931  */
932 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
933 {
934 	/* threshold event is triggered in finer grain than soft limit */
935 	if (unlikely(mem_cgroup_event_ratelimit(memcg,
936 						MEM_CGROUP_TARGET_THRESH))) {
937 		bool do_softlimit;
938 		bool do_numainfo __maybe_unused;
939 
940 		do_softlimit = mem_cgroup_event_ratelimit(memcg,
941 						MEM_CGROUP_TARGET_SOFTLIMIT);
942 #if MAX_NUMNODES > 1
943 		do_numainfo = mem_cgroup_event_ratelimit(memcg,
944 						MEM_CGROUP_TARGET_NUMAINFO);
945 #endif
946 		mem_cgroup_threshold(memcg);
947 		if (unlikely(do_softlimit))
948 			mem_cgroup_update_tree(memcg, page);
949 #if MAX_NUMNODES > 1
950 		if (unlikely(do_numainfo))
951 			atomic_inc(&memcg->numainfo_events);
952 #endif
953 	}
954 }
955 
956 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
957 {
958 	/*
959 	 * mm_update_next_owner() may clear mm->owner to NULL
960 	 * if it races with swapoff, page migration, etc.
961 	 * So this can be called with p == NULL.
962 	 */
963 	if (unlikely(!p))
964 		return NULL;
965 
966 	return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
967 }
968 
969 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
970 {
971 	struct mem_cgroup *memcg = NULL;
972 
973 	rcu_read_lock();
974 	do {
975 		/*
976 		 * Page cache insertions can happen withou an
977 		 * actual mm context, e.g. during disk probing
978 		 * on boot, loopback IO, acct() writes etc.
979 		 */
980 		if (unlikely(!mm))
981 			memcg = root_mem_cgroup;
982 		else {
983 			memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
984 			if (unlikely(!memcg))
985 				memcg = root_mem_cgroup;
986 		}
987 	} while (!css_tryget_online(&memcg->css));
988 	rcu_read_unlock();
989 	return memcg;
990 }
991 
992 /**
993  * mem_cgroup_iter - iterate over memory cgroup hierarchy
994  * @root: hierarchy root
995  * @prev: previously returned memcg, NULL on first invocation
996  * @reclaim: cookie for shared reclaim walks, NULL for full walks
997  *
998  * Returns references to children of the hierarchy below @root, or
999  * @root itself, or %NULL after a full round-trip.
1000  *
1001  * Caller must pass the return value in @prev on subsequent
1002  * invocations for reference counting, or use mem_cgroup_iter_break()
1003  * to cancel a hierarchy walk before the round-trip is complete.
1004  *
1005  * Reclaimers can specify a zone and a priority level in @reclaim to
1006  * divide up the memcgs in the hierarchy among all concurrent
1007  * reclaimers operating on the same zone and priority.
1008  */
1009 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1010 				   struct mem_cgroup *prev,
1011 				   struct mem_cgroup_reclaim_cookie *reclaim)
1012 {
1013 	struct reclaim_iter *uninitialized_var(iter);
1014 	struct cgroup_subsys_state *css = NULL;
1015 	struct mem_cgroup *memcg = NULL;
1016 	struct mem_cgroup *pos = NULL;
1017 
1018 	if (mem_cgroup_disabled())
1019 		return NULL;
1020 
1021 	if (!root)
1022 		root = root_mem_cgroup;
1023 
1024 	if (prev && !reclaim)
1025 		pos = prev;
1026 
1027 	if (!root->use_hierarchy && root != root_mem_cgroup) {
1028 		if (prev)
1029 			goto out;
1030 		return root;
1031 	}
1032 
1033 	rcu_read_lock();
1034 
1035 	if (reclaim) {
1036 		struct mem_cgroup_per_zone *mz;
1037 
1038 		mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1039 		iter = &mz->iter[reclaim->priority];
1040 
1041 		if (prev && reclaim->generation != iter->generation)
1042 			goto out_unlock;
1043 
1044 		do {
1045 			pos = READ_ONCE(iter->position);
1046 			/*
1047 			 * A racing update may change the position and
1048 			 * put the last reference, hence css_tryget(),
1049 			 * or retry to see the updated position.
1050 			 */
1051 		} while (pos && !css_tryget(&pos->css));
1052 	}
1053 
1054 	if (pos)
1055 		css = &pos->css;
1056 
1057 	for (;;) {
1058 		css = css_next_descendant_pre(css, &root->css);
1059 		if (!css) {
1060 			/*
1061 			 * Reclaimers share the hierarchy walk, and a
1062 			 * new one might jump in right at the end of
1063 			 * the hierarchy - make sure they see at least
1064 			 * one group and restart from the beginning.
1065 			 */
1066 			if (!prev)
1067 				continue;
1068 			break;
1069 		}
1070 
1071 		/*
1072 		 * Verify the css and acquire a reference.  The root
1073 		 * is provided by the caller, so we know it's alive
1074 		 * and kicking, and don't take an extra reference.
1075 		 */
1076 		memcg = mem_cgroup_from_css(css);
1077 
1078 		if (css == &root->css)
1079 			break;
1080 
1081 		if (css_tryget(css)) {
1082 			/*
1083 			 * Make sure the memcg is initialized:
1084 			 * mem_cgroup_css_online() orders the the
1085 			 * initialization against setting the flag.
1086 			 */
1087 			if (smp_load_acquire(&memcg->initialized))
1088 				break;
1089 
1090 			css_put(css);
1091 		}
1092 
1093 		memcg = NULL;
1094 	}
1095 
1096 	if (reclaim) {
1097 		if (cmpxchg(&iter->position, pos, memcg) == pos) {
1098 			if (memcg)
1099 				css_get(&memcg->css);
1100 			if (pos)
1101 				css_put(&pos->css);
1102 		}
1103 
1104 		/*
1105 		 * pairs with css_tryget when dereferencing iter->position
1106 		 * above.
1107 		 */
1108 		if (pos)
1109 			css_put(&pos->css);
1110 
1111 		if (!memcg)
1112 			iter->generation++;
1113 		else if (!prev)
1114 			reclaim->generation = iter->generation;
1115 	}
1116 
1117 out_unlock:
1118 	rcu_read_unlock();
1119 out:
1120 	if (prev && prev != root)
1121 		css_put(&prev->css);
1122 
1123 	return memcg;
1124 }
1125 
1126 /**
1127  * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1128  * @root: hierarchy root
1129  * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1130  */
1131 void mem_cgroup_iter_break(struct mem_cgroup *root,
1132 			   struct mem_cgroup *prev)
1133 {
1134 	if (!root)
1135 		root = root_mem_cgroup;
1136 	if (prev && prev != root)
1137 		css_put(&prev->css);
1138 }
1139 
1140 /*
1141  * Iteration constructs for visiting all cgroups (under a tree).  If
1142  * loops are exited prematurely (break), mem_cgroup_iter_break() must
1143  * be used for reference counting.
1144  */
1145 #define for_each_mem_cgroup_tree(iter, root)		\
1146 	for (iter = mem_cgroup_iter(root, NULL, NULL);	\
1147 	     iter != NULL;				\
1148 	     iter = mem_cgroup_iter(root, iter, NULL))
1149 
1150 #define for_each_mem_cgroup(iter)			\
1151 	for (iter = mem_cgroup_iter(NULL, NULL, NULL);	\
1152 	     iter != NULL;				\
1153 	     iter = mem_cgroup_iter(NULL, iter, NULL))
1154 
1155 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1156 {
1157 	struct mem_cgroup *memcg;
1158 
1159 	rcu_read_lock();
1160 	memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1161 	if (unlikely(!memcg))
1162 		goto out;
1163 
1164 	switch (idx) {
1165 	case PGFAULT:
1166 		this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1167 		break;
1168 	case PGMAJFAULT:
1169 		this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1170 		break;
1171 	default:
1172 		BUG();
1173 	}
1174 out:
1175 	rcu_read_unlock();
1176 }
1177 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1178 
1179 /**
1180  * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1181  * @zone: zone of the wanted lruvec
1182  * @memcg: memcg of the wanted lruvec
1183  *
1184  * Returns the lru list vector holding pages for the given @zone and
1185  * @mem.  This can be the global zone lruvec, if the memory controller
1186  * is disabled.
1187  */
1188 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1189 				      struct mem_cgroup *memcg)
1190 {
1191 	struct mem_cgroup_per_zone *mz;
1192 	struct lruvec *lruvec;
1193 
1194 	if (mem_cgroup_disabled()) {
1195 		lruvec = &zone->lruvec;
1196 		goto out;
1197 	}
1198 
1199 	mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1200 	lruvec = &mz->lruvec;
1201 out:
1202 	/*
1203 	 * Since a node can be onlined after the mem_cgroup was created,
1204 	 * we have to be prepared to initialize lruvec->zone here;
1205 	 * and if offlined then reonlined, we need to reinitialize it.
1206 	 */
1207 	if (unlikely(lruvec->zone != zone))
1208 		lruvec->zone = zone;
1209 	return lruvec;
1210 }
1211 
1212 /**
1213  * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1214  * @page: the page
1215  * @zone: zone of the page
1216  *
1217  * This function is only safe when following the LRU page isolation
1218  * and putback protocol: the LRU lock must be held, and the page must
1219  * either be PageLRU() or the caller must have isolated/allocated it.
1220  */
1221 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1222 {
1223 	struct mem_cgroup_per_zone *mz;
1224 	struct mem_cgroup *memcg;
1225 	struct lruvec *lruvec;
1226 
1227 	if (mem_cgroup_disabled()) {
1228 		lruvec = &zone->lruvec;
1229 		goto out;
1230 	}
1231 
1232 	memcg = page->mem_cgroup;
1233 	/*
1234 	 * Swapcache readahead pages are added to the LRU - and
1235 	 * possibly migrated - before they are charged.
1236 	 */
1237 	if (!memcg)
1238 		memcg = root_mem_cgroup;
1239 
1240 	mz = mem_cgroup_page_zoneinfo(memcg, page);
1241 	lruvec = &mz->lruvec;
1242 out:
1243 	/*
1244 	 * Since a node can be onlined after the mem_cgroup was created,
1245 	 * we have to be prepared to initialize lruvec->zone here;
1246 	 * and if offlined then reonlined, we need to reinitialize it.
1247 	 */
1248 	if (unlikely(lruvec->zone != zone))
1249 		lruvec->zone = zone;
1250 	return lruvec;
1251 }
1252 
1253 /**
1254  * mem_cgroup_update_lru_size - account for adding or removing an lru page
1255  * @lruvec: mem_cgroup per zone lru vector
1256  * @lru: index of lru list the page is sitting on
1257  * @nr_pages: positive when adding or negative when removing
1258  *
1259  * This function must be called when a page is added to or removed from an
1260  * lru list.
1261  */
1262 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1263 				int nr_pages)
1264 {
1265 	struct mem_cgroup_per_zone *mz;
1266 	unsigned long *lru_size;
1267 
1268 	if (mem_cgroup_disabled())
1269 		return;
1270 
1271 	mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1272 	lru_size = mz->lru_size + lru;
1273 	*lru_size += nr_pages;
1274 	VM_BUG_ON((long)(*lru_size) < 0);
1275 }
1276 
1277 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1278 {
1279 	if (root == memcg)
1280 		return true;
1281 	if (!root->use_hierarchy)
1282 		return false;
1283 	return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1284 }
1285 
1286 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1287 {
1288 	struct mem_cgroup *task_memcg;
1289 	struct task_struct *p;
1290 	bool ret;
1291 
1292 	p = find_lock_task_mm(task);
1293 	if (p) {
1294 		task_memcg = get_mem_cgroup_from_mm(p->mm);
1295 		task_unlock(p);
1296 	} else {
1297 		/*
1298 		 * All threads may have already detached their mm's, but the oom
1299 		 * killer still needs to detect if they have already been oom
1300 		 * killed to prevent needlessly killing additional tasks.
1301 		 */
1302 		rcu_read_lock();
1303 		task_memcg = mem_cgroup_from_task(task);
1304 		css_get(&task_memcg->css);
1305 		rcu_read_unlock();
1306 	}
1307 	ret = mem_cgroup_is_descendant(task_memcg, memcg);
1308 	css_put(&task_memcg->css);
1309 	return ret;
1310 }
1311 
1312 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1313 {
1314 	unsigned long inactive_ratio;
1315 	unsigned long inactive;
1316 	unsigned long active;
1317 	unsigned long gb;
1318 
1319 	inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1320 	active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1321 
1322 	gb = (inactive + active) >> (30 - PAGE_SHIFT);
1323 	if (gb)
1324 		inactive_ratio = int_sqrt(10 * gb);
1325 	else
1326 		inactive_ratio = 1;
1327 
1328 	return inactive * inactive_ratio < active;
1329 }
1330 
1331 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1332 {
1333 	struct mem_cgroup_per_zone *mz;
1334 	struct mem_cgroup *memcg;
1335 
1336 	if (mem_cgroup_disabled())
1337 		return true;
1338 
1339 	mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1340 	memcg = mz->memcg;
1341 
1342 	return !!(memcg->css.flags & CSS_ONLINE);
1343 }
1344 
1345 #define mem_cgroup_from_counter(counter, member)	\
1346 	container_of(counter, struct mem_cgroup, member)
1347 
1348 /**
1349  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1350  * @memcg: the memory cgroup
1351  *
1352  * Returns the maximum amount of memory @mem can be charged with, in
1353  * pages.
1354  */
1355 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1356 {
1357 	unsigned long margin = 0;
1358 	unsigned long count;
1359 	unsigned long limit;
1360 
1361 	count = page_counter_read(&memcg->memory);
1362 	limit = READ_ONCE(memcg->memory.limit);
1363 	if (count < limit)
1364 		margin = limit - count;
1365 
1366 	if (do_swap_account) {
1367 		count = page_counter_read(&memcg->memsw);
1368 		limit = READ_ONCE(memcg->memsw.limit);
1369 		if (count <= limit)
1370 			margin = min(margin, limit - count);
1371 	}
1372 
1373 	return margin;
1374 }
1375 
1376 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1377 {
1378 	/* root ? */
1379 	if (mem_cgroup_disabled() || !memcg->css.parent)
1380 		return vm_swappiness;
1381 
1382 	return memcg->swappiness;
1383 }
1384 
1385 /*
1386  * A routine for checking "mem" is under move_account() or not.
1387  *
1388  * Checking a cgroup is mc.from or mc.to or under hierarchy of
1389  * moving cgroups. This is for waiting at high-memory pressure
1390  * caused by "move".
1391  */
1392 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1393 {
1394 	struct mem_cgroup *from;
1395 	struct mem_cgroup *to;
1396 	bool ret = false;
1397 	/*
1398 	 * Unlike task_move routines, we access mc.to, mc.from not under
1399 	 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1400 	 */
1401 	spin_lock(&mc.lock);
1402 	from = mc.from;
1403 	to = mc.to;
1404 	if (!from)
1405 		goto unlock;
1406 
1407 	ret = mem_cgroup_is_descendant(from, memcg) ||
1408 		mem_cgroup_is_descendant(to, memcg);
1409 unlock:
1410 	spin_unlock(&mc.lock);
1411 	return ret;
1412 }
1413 
1414 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1415 {
1416 	if (mc.moving_task && current != mc.moving_task) {
1417 		if (mem_cgroup_under_move(memcg)) {
1418 			DEFINE_WAIT(wait);
1419 			prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1420 			/* moving charge context might have finished. */
1421 			if (mc.moving_task)
1422 				schedule();
1423 			finish_wait(&mc.waitq, &wait);
1424 			return true;
1425 		}
1426 	}
1427 	return false;
1428 }
1429 
1430 #define K(x) ((x) << (PAGE_SHIFT-10))
1431 /**
1432  * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1433  * @memcg: The memory cgroup that went over limit
1434  * @p: Task that is going to be killed
1435  *
1436  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1437  * enabled
1438  */
1439 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1440 {
1441 	/* oom_info_lock ensures that parallel ooms do not interleave */
1442 	static DEFINE_MUTEX(oom_info_lock);
1443 	struct mem_cgroup *iter;
1444 	unsigned int i;
1445 
1446 	mutex_lock(&oom_info_lock);
1447 	rcu_read_lock();
1448 
1449 	if (p) {
1450 		pr_info("Task in ");
1451 		pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1452 		pr_cont(" killed as a result of limit of ");
1453 	} else {
1454 		pr_info("Memory limit reached of cgroup ");
1455 	}
1456 
1457 	pr_cont_cgroup_path(memcg->css.cgroup);
1458 	pr_cont("\n");
1459 
1460 	rcu_read_unlock();
1461 
1462 	pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1463 		K((u64)page_counter_read(&memcg->memory)),
1464 		K((u64)memcg->memory.limit), memcg->memory.failcnt);
1465 	pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1466 		K((u64)page_counter_read(&memcg->memsw)),
1467 		K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1468 	pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1469 		K((u64)page_counter_read(&memcg->kmem)),
1470 		K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1471 
1472 	for_each_mem_cgroup_tree(iter, memcg) {
1473 		pr_info("Memory cgroup stats for ");
1474 		pr_cont_cgroup_path(iter->css.cgroup);
1475 		pr_cont(":");
1476 
1477 		for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1478 			if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1479 				continue;
1480 			pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1481 				K(mem_cgroup_read_stat(iter, i)));
1482 		}
1483 
1484 		for (i = 0; i < NR_LRU_LISTS; i++)
1485 			pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1486 				K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1487 
1488 		pr_cont("\n");
1489 	}
1490 	mutex_unlock(&oom_info_lock);
1491 }
1492 
1493 /*
1494  * This function returns the number of memcg under hierarchy tree. Returns
1495  * 1(self count) if no children.
1496  */
1497 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1498 {
1499 	int num = 0;
1500 	struct mem_cgroup *iter;
1501 
1502 	for_each_mem_cgroup_tree(iter, memcg)
1503 		num++;
1504 	return num;
1505 }
1506 
1507 /*
1508  * Return the memory (and swap, if configured) limit for a memcg.
1509  */
1510 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1511 {
1512 	unsigned long limit;
1513 
1514 	limit = memcg->memory.limit;
1515 	if (mem_cgroup_swappiness(memcg)) {
1516 		unsigned long memsw_limit;
1517 
1518 		memsw_limit = memcg->memsw.limit;
1519 		limit = min(limit + total_swap_pages, memsw_limit);
1520 	}
1521 	return limit;
1522 }
1523 
1524 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1525 				     int order)
1526 {
1527 	struct mem_cgroup *iter;
1528 	unsigned long chosen_points = 0;
1529 	unsigned long totalpages;
1530 	unsigned int points = 0;
1531 	struct task_struct *chosen = NULL;
1532 
1533 	/*
1534 	 * If current has a pending SIGKILL or is exiting, then automatically
1535 	 * select it.  The goal is to allow it to allocate so that it may
1536 	 * quickly exit and free its memory.
1537 	 */
1538 	if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1539 		mark_tsk_oom_victim(current);
1540 		return;
1541 	}
1542 
1543 	check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL, memcg);
1544 	totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1545 	for_each_mem_cgroup_tree(iter, memcg) {
1546 		struct css_task_iter it;
1547 		struct task_struct *task;
1548 
1549 		css_task_iter_start(&iter->css, &it);
1550 		while ((task = css_task_iter_next(&it))) {
1551 			switch (oom_scan_process_thread(task, totalpages, NULL,
1552 							false)) {
1553 			case OOM_SCAN_SELECT:
1554 				if (chosen)
1555 					put_task_struct(chosen);
1556 				chosen = task;
1557 				chosen_points = ULONG_MAX;
1558 				get_task_struct(chosen);
1559 				/* fall through */
1560 			case OOM_SCAN_CONTINUE:
1561 				continue;
1562 			case OOM_SCAN_ABORT:
1563 				css_task_iter_end(&it);
1564 				mem_cgroup_iter_break(memcg, iter);
1565 				if (chosen)
1566 					put_task_struct(chosen);
1567 				return;
1568 			case OOM_SCAN_OK:
1569 				break;
1570 			};
1571 			points = oom_badness(task, memcg, NULL, totalpages);
1572 			if (!points || points < chosen_points)
1573 				continue;
1574 			/* Prefer thread group leaders for display purposes */
1575 			if (points == chosen_points &&
1576 			    thread_group_leader(chosen))
1577 				continue;
1578 
1579 			if (chosen)
1580 				put_task_struct(chosen);
1581 			chosen = task;
1582 			chosen_points = points;
1583 			get_task_struct(chosen);
1584 		}
1585 		css_task_iter_end(&it);
1586 	}
1587 
1588 	if (!chosen)
1589 		return;
1590 	points = chosen_points * 1000 / totalpages;
1591 	oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1592 			 NULL, "Memory cgroup out of memory");
1593 }
1594 
1595 #if MAX_NUMNODES > 1
1596 
1597 /**
1598  * test_mem_cgroup_node_reclaimable
1599  * @memcg: the target memcg
1600  * @nid: the node ID to be checked.
1601  * @noswap : specify true here if the user wants flle only information.
1602  *
1603  * This function returns whether the specified memcg contains any
1604  * reclaimable pages on a node. Returns true if there are any reclaimable
1605  * pages in the node.
1606  */
1607 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1608 		int nid, bool noswap)
1609 {
1610 	if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1611 		return true;
1612 	if (noswap || !total_swap_pages)
1613 		return false;
1614 	if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1615 		return true;
1616 	return false;
1617 
1618 }
1619 
1620 /*
1621  * Always updating the nodemask is not very good - even if we have an empty
1622  * list or the wrong list here, we can start from some node and traverse all
1623  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1624  *
1625  */
1626 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1627 {
1628 	int nid;
1629 	/*
1630 	 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1631 	 * pagein/pageout changes since the last update.
1632 	 */
1633 	if (!atomic_read(&memcg->numainfo_events))
1634 		return;
1635 	if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1636 		return;
1637 
1638 	/* make a nodemask where this memcg uses memory from */
1639 	memcg->scan_nodes = node_states[N_MEMORY];
1640 
1641 	for_each_node_mask(nid, node_states[N_MEMORY]) {
1642 
1643 		if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1644 			node_clear(nid, memcg->scan_nodes);
1645 	}
1646 
1647 	atomic_set(&memcg->numainfo_events, 0);
1648 	atomic_set(&memcg->numainfo_updating, 0);
1649 }
1650 
1651 /*
1652  * Selecting a node where we start reclaim from. Because what we need is just
1653  * reducing usage counter, start from anywhere is O,K. Considering
1654  * memory reclaim from current node, there are pros. and cons.
1655  *
1656  * Freeing memory from current node means freeing memory from a node which
1657  * we'll use or we've used. So, it may make LRU bad. And if several threads
1658  * hit limits, it will see a contention on a node. But freeing from remote
1659  * node means more costs for memory reclaim because of memory latency.
1660  *
1661  * Now, we use round-robin. Better algorithm is welcomed.
1662  */
1663 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1664 {
1665 	int node;
1666 
1667 	mem_cgroup_may_update_nodemask(memcg);
1668 	node = memcg->last_scanned_node;
1669 
1670 	node = next_node(node, memcg->scan_nodes);
1671 	if (node == MAX_NUMNODES)
1672 		node = first_node(memcg->scan_nodes);
1673 	/*
1674 	 * We call this when we hit limit, not when pages are added to LRU.
1675 	 * No LRU may hold pages because all pages are UNEVICTABLE or
1676 	 * memcg is too small and all pages are not on LRU. In that case,
1677 	 * we use curret node.
1678 	 */
1679 	if (unlikely(node == MAX_NUMNODES))
1680 		node = numa_node_id();
1681 
1682 	memcg->last_scanned_node = node;
1683 	return node;
1684 }
1685 #else
1686 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1687 {
1688 	return 0;
1689 }
1690 #endif
1691 
1692 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1693 				   struct zone *zone,
1694 				   gfp_t gfp_mask,
1695 				   unsigned long *total_scanned)
1696 {
1697 	struct mem_cgroup *victim = NULL;
1698 	int total = 0;
1699 	int loop = 0;
1700 	unsigned long excess;
1701 	unsigned long nr_scanned;
1702 	struct mem_cgroup_reclaim_cookie reclaim = {
1703 		.zone = zone,
1704 		.priority = 0,
1705 	};
1706 
1707 	excess = soft_limit_excess(root_memcg);
1708 
1709 	while (1) {
1710 		victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1711 		if (!victim) {
1712 			loop++;
1713 			if (loop >= 2) {
1714 				/*
1715 				 * If we have not been able to reclaim
1716 				 * anything, it might because there are
1717 				 * no reclaimable pages under this hierarchy
1718 				 */
1719 				if (!total)
1720 					break;
1721 				/*
1722 				 * We want to do more targeted reclaim.
1723 				 * excess >> 2 is not to excessive so as to
1724 				 * reclaim too much, nor too less that we keep
1725 				 * coming back to reclaim from this cgroup
1726 				 */
1727 				if (total >= (excess >> 2) ||
1728 					(loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1729 					break;
1730 			}
1731 			continue;
1732 		}
1733 		total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1734 						     zone, &nr_scanned);
1735 		*total_scanned += nr_scanned;
1736 		if (!soft_limit_excess(root_memcg))
1737 			break;
1738 	}
1739 	mem_cgroup_iter_break(root_memcg, victim);
1740 	return total;
1741 }
1742 
1743 #ifdef CONFIG_LOCKDEP
1744 static struct lockdep_map memcg_oom_lock_dep_map = {
1745 	.name = "memcg_oom_lock",
1746 };
1747 #endif
1748 
1749 static DEFINE_SPINLOCK(memcg_oom_lock);
1750 
1751 /*
1752  * Check OOM-Killer is already running under our hierarchy.
1753  * If someone is running, return false.
1754  */
1755 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1756 {
1757 	struct mem_cgroup *iter, *failed = NULL;
1758 
1759 	spin_lock(&memcg_oom_lock);
1760 
1761 	for_each_mem_cgroup_tree(iter, memcg) {
1762 		if (iter->oom_lock) {
1763 			/*
1764 			 * this subtree of our hierarchy is already locked
1765 			 * so we cannot give a lock.
1766 			 */
1767 			failed = iter;
1768 			mem_cgroup_iter_break(memcg, iter);
1769 			break;
1770 		} else
1771 			iter->oom_lock = true;
1772 	}
1773 
1774 	if (failed) {
1775 		/*
1776 		 * OK, we failed to lock the whole subtree so we have
1777 		 * to clean up what we set up to the failing subtree
1778 		 */
1779 		for_each_mem_cgroup_tree(iter, memcg) {
1780 			if (iter == failed) {
1781 				mem_cgroup_iter_break(memcg, iter);
1782 				break;
1783 			}
1784 			iter->oom_lock = false;
1785 		}
1786 	} else
1787 		mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1788 
1789 	spin_unlock(&memcg_oom_lock);
1790 
1791 	return !failed;
1792 }
1793 
1794 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1795 {
1796 	struct mem_cgroup *iter;
1797 
1798 	spin_lock(&memcg_oom_lock);
1799 	mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1800 	for_each_mem_cgroup_tree(iter, memcg)
1801 		iter->oom_lock = false;
1802 	spin_unlock(&memcg_oom_lock);
1803 }
1804 
1805 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1806 {
1807 	struct mem_cgroup *iter;
1808 
1809 	for_each_mem_cgroup_tree(iter, memcg)
1810 		atomic_inc(&iter->under_oom);
1811 }
1812 
1813 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1814 {
1815 	struct mem_cgroup *iter;
1816 
1817 	/*
1818 	 * When a new child is created while the hierarchy is under oom,
1819 	 * mem_cgroup_oom_lock() may not be called. We have to use
1820 	 * atomic_add_unless() here.
1821 	 */
1822 	for_each_mem_cgroup_tree(iter, memcg)
1823 		atomic_add_unless(&iter->under_oom, -1, 0);
1824 }
1825 
1826 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1827 
1828 struct oom_wait_info {
1829 	struct mem_cgroup *memcg;
1830 	wait_queue_t	wait;
1831 };
1832 
1833 static int memcg_oom_wake_function(wait_queue_t *wait,
1834 	unsigned mode, int sync, void *arg)
1835 {
1836 	struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1837 	struct mem_cgroup *oom_wait_memcg;
1838 	struct oom_wait_info *oom_wait_info;
1839 
1840 	oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1841 	oom_wait_memcg = oom_wait_info->memcg;
1842 
1843 	if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1844 	    !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1845 		return 0;
1846 	return autoremove_wake_function(wait, mode, sync, arg);
1847 }
1848 
1849 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1850 {
1851 	atomic_inc(&memcg->oom_wakeups);
1852 	/* for filtering, pass "memcg" as argument. */
1853 	__wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1854 }
1855 
1856 static void memcg_oom_recover(struct mem_cgroup *memcg)
1857 {
1858 	if (memcg && atomic_read(&memcg->under_oom))
1859 		memcg_wakeup_oom(memcg);
1860 }
1861 
1862 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1863 {
1864 	if (!current->memcg_oom.may_oom)
1865 		return;
1866 	/*
1867 	 * We are in the middle of the charge context here, so we
1868 	 * don't want to block when potentially sitting on a callstack
1869 	 * that holds all kinds of filesystem and mm locks.
1870 	 *
1871 	 * Also, the caller may handle a failed allocation gracefully
1872 	 * (like optional page cache readahead) and so an OOM killer
1873 	 * invocation might not even be necessary.
1874 	 *
1875 	 * That's why we don't do anything here except remember the
1876 	 * OOM context and then deal with it at the end of the page
1877 	 * fault when the stack is unwound, the locks are released,
1878 	 * and when we know whether the fault was overall successful.
1879 	 */
1880 	css_get(&memcg->css);
1881 	current->memcg_oom.memcg = memcg;
1882 	current->memcg_oom.gfp_mask = mask;
1883 	current->memcg_oom.order = order;
1884 }
1885 
1886 /**
1887  * mem_cgroup_oom_synchronize - complete memcg OOM handling
1888  * @handle: actually kill/wait or just clean up the OOM state
1889  *
1890  * This has to be called at the end of a page fault if the memcg OOM
1891  * handler was enabled.
1892  *
1893  * Memcg supports userspace OOM handling where failed allocations must
1894  * sleep on a waitqueue until the userspace task resolves the
1895  * situation.  Sleeping directly in the charge context with all kinds
1896  * of locks held is not a good idea, instead we remember an OOM state
1897  * in the task and mem_cgroup_oom_synchronize() has to be called at
1898  * the end of the page fault to complete the OOM handling.
1899  *
1900  * Returns %true if an ongoing memcg OOM situation was detected and
1901  * completed, %false otherwise.
1902  */
1903 bool mem_cgroup_oom_synchronize(bool handle)
1904 {
1905 	struct mem_cgroup *memcg = current->memcg_oom.memcg;
1906 	struct oom_wait_info owait;
1907 	bool locked;
1908 
1909 	/* OOM is global, do not handle */
1910 	if (!memcg)
1911 		return false;
1912 
1913 	if (!handle || oom_killer_disabled)
1914 		goto cleanup;
1915 
1916 	owait.memcg = memcg;
1917 	owait.wait.flags = 0;
1918 	owait.wait.func = memcg_oom_wake_function;
1919 	owait.wait.private = current;
1920 	INIT_LIST_HEAD(&owait.wait.task_list);
1921 
1922 	prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1923 	mem_cgroup_mark_under_oom(memcg);
1924 
1925 	locked = mem_cgroup_oom_trylock(memcg);
1926 
1927 	if (locked)
1928 		mem_cgroup_oom_notify(memcg);
1929 
1930 	if (locked && !memcg->oom_kill_disable) {
1931 		mem_cgroup_unmark_under_oom(memcg);
1932 		finish_wait(&memcg_oom_waitq, &owait.wait);
1933 		mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1934 					 current->memcg_oom.order);
1935 	} else {
1936 		schedule();
1937 		mem_cgroup_unmark_under_oom(memcg);
1938 		finish_wait(&memcg_oom_waitq, &owait.wait);
1939 	}
1940 
1941 	if (locked) {
1942 		mem_cgroup_oom_unlock(memcg);
1943 		/*
1944 		 * There is no guarantee that an OOM-lock contender
1945 		 * sees the wakeups triggered by the OOM kill
1946 		 * uncharges.  Wake any sleepers explicitely.
1947 		 */
1948 		memcg_oom_recover(memcg);
1949 	}
1950 cleanup:
1951 	current->memcg_oom.memcg = NULL;
1952 	css_put(&memcg->css);
1953 	return true;
1954 }
1955 
1956 /**
1957  * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1958  * @page: page that is going to change accounted state
1959  *
1960  * This function must mark the beginning of an accounted page state
1961  * change to prevent double accounting when the page is concurrently
1962  * being moved to another memcg:
1963  *
1964  *   memcg = mem_cgroup_begin_page_stat(page);
1965  *   if (TestClearPageState(page))
1966  *     mem_cgroup_update_page_stat(memcg, state, -1);
1967  *   mem_cgroup_end_page_stat(memcg);
1968  */
1969 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1970 {
1971 	struct mem_cgroup *memcg;
1972 	unsigned long flags;
1973 
1974 	/*
1975 	 * The RCU lock is held throughout the transaction.  The fast
1976 	 * path can get away without acquiring the memcg->move_lock
1977 	 * because page moving starts with an RCU grace period.
1978 	 *
1979 	 * The RCU lock also protects the memcg from being freed when
1980 	 * the page state that is going to change is the only thing
1981 	 * preventing the page from being uncharged.
1982 	 * E.g. end-writeback clearing PageWriteback(), which allows
1983 	 * migration to go ahead and uncharge the page before the
1984 	 * account transaction might be complete.
1985 	 */
1986 	rcu_read_lock();
1987 
1988 	if (mem_cgroup_disabled())
1989 		return NULL;
1990 again:
1991 	memcg = page->mem_cgroup;
1992 	if (unlikely(!memcg))
1993 		return NULL;
1994 
1995 	if (atomic_read(&memcg->moving_account) <= 0)
1996 		return memcg;
1997 
1998 	spin_lock_irqsave(&memcg->move_lock, flags);
1999 	if (memcg != page->mem_cgroup) {
2000 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2001 		goto again;
2002 	}
2003 
2004 	/*
2005 	 * When charge migration first begins, we can have locked and
2006 	 * unlocked page stat updates happening concurrently.  Track
2007 	 * the task who has the lock for mem_cgroup_end_page_stat().
2008 	 */
2009 	memcg->move_lock_task = current;
2010 	memcg->move_lock_flags = flags;
2011 
2012 	return memcg;
2013 }
2014 
2015 /**
2016  * mem_cgroup_end_page_stat - finish a page state statistics transaction
2017  * @memcg: the memcg that was accounted against
2018  */
2019 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2020 {
2021 	if (memcg && memcg->move_lock_task == current) {
2022 		unsigned long flags = memcg->move_lock_flags;
2023 
2024 		memcg->move_lock_task = NULL;
2025 		memcg->move_lock_flags = 0;
2026 
2027 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2028 	}
2029 
2030 	rcu_read_unlock();
2031 }
2032 
2033 /**
2034  * mem_cgroup_update_page_stat - update page state statistics
2035  * @memcg: memcg to account against
2036  * @idx: page state item to account
2037  * @val: number of pages (positive or negative)
2038  *
2039  * See mem_cgroup_begin_page_stat() for locking requirements.
2040  */
2041 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2042 				 enum mem_cgroup_stat_index idx, int val)
2043 {
2044 	VM_BUG_ON(!rcu_read_lock_held());
2045 
2046 	if (memcg)
2047 		this_cpu_add(memcg->stat->count[idx], val);
2048 }
2049 
2050 /*
2051  * size of first charge trial. "32" comes from vmscan.c's magic value.
2052  * TODO: maybe necessary to use big numbers in big irons.
2053  */
2054 #define CHARGE_BATCH	32U
2055 struct memcg_stock_pcp {
2056 	struct mem_cgroup *cached; /* this never be root cgroup */
2057 	unsigned int nr_pages;
2058 	struct work_struct work;
2059 	unsigned long flags;
2060 #define FLUSHING_CACHED_CHARGE	0
2061 };
2062 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2063 static DEFINE_MUTEX(percpu_charge_mutex);
2064 
2065 /**
2066  * consume_stock: Try to consume stocked charge on this cpu.
2067  * @memcg: memcg to consume from.
2068  * @nr_pages: how many pages to charge.
2069  *
2070  * The charges will only happen if @memcg matches the current cpu's memcg
2071  * stock, and at least @nr_pages are available in that stock.  Failure to
2072  * service an allocation will refill the stock.
2073  *
2074  * returns true if successful, false otherwise.
2075  */
2076 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2077 {
2078 	struct memcg_stock_pcp *stock;
2079 	bool ret = false;
2080 
2081 	if (nr_pages > CHARGE_BATCH)
2082 		return ret;
2083 
2084 	stock = &get_cpu_var(memcg_stock);
2085 	if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2086 		stock->nr_pages -= nr_pages;
2087 		ret = true;
2088 	}
2089 	put_cpu_var(memcg_stock);
2090 	return ret;
2091 }
2092 
2093 /*
2094  * Returns stocks cached in percpu and reset cached information.
2095  */
2096 static void drain_stock(struct memcg_stock_pcp *stock)
2097 {
2098 	struct mem_cgroup *old = stock->cached;
2099 
2100 	if (stock->nr_pages) {
2101 		page_counter_uncharge(&old->memory, stock->nr_pages);
2102 		if (do_swap_account)
2103 			page_counter_uncharge(&old->memsw, stock->nr_pages);
2104 		css_put_many(&old->css, stock->nr_pages);
2105 		stock->nr_pages = 0;
2106 	}
2107 	stock->cached = NULL;
2108 }
2109 
2110 /*
2111  * This must be called under preempt disabled or must be called by
2112  * a thread which is pinned to local cpu.
2113  */
2114 static void drain_local_stock(struct work_struct *dummy)
2115 {
2116 	struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2117 	drain_stock(stock);
2118 	clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2119 }
2120 
2121 /*
2122  * Cache charges(val) to local per_cpu area.
2123  * This will be consumed by consume_stock() function, later.
2124  */
2125 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2126 {
2127 	struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2128 
2129 	if (stock->cached != memcg) { /* reset if necessary */
2130 		drain_stock(stock);
2131 		stock->cached = memcg;
2132 	}
2133 	stock->nr_pages += nr_pages;
2134 	put_cpu_var(memcg_stock);
2135 }
2136 
2137 /*
2138  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2139  * of the hierarchy under it.
2140  */
2141 static void drain_all_stock(struct mem_cgroup *root_memcg)
2142 {
2143 	int cpu, curcpu;
2144 
2145 	/* If someone's already draining, avoid adding running more workers. */
2146 	if (!mutex_trylock(&percpu_charge_mutex))
2147 		return;
2148 	/* Notify other cpus that system-wide "drain" is running */
2149 	get_online_cpus();
2150 	curcpu = get_cpu();
2151 	for_each_online_cpu(cpu) {
2152 		struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2153 		struct mem_cgroup *memcg;
2154 
2155 		memcg = stock->cached;
2156 		if (!memcg || !stock->nr_pages)
2157 			continue;
2158 		if (!mem_cgroup_is_descendant(memcg, root_memcg))
2159 			continue;
2160 		if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2161 			if (cpu == curcpu)
2162 				drain_local_stock(&stock->work);
2163 			else
2164 				schedule_work_on(cpu, &stock->work);
2165 		}
2166 	}
2167 	put_cpu();
2168 	put_online_cpus();
2169 	mutex_unlock(&percpu_charge_mutex);
2170 }
2171 
2172 /*
2173  * This function drains percpu counter value from DEAD cpu and
2174  * move it to local cpu. Note that this function can be preempted.
2175  */
2176 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2177 {
2178 	int i;
2179 
2180 	spin_lock(&memcg->pcp_counter_lock);
2181 	for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2182 		long x = per_cpu(memcg->stat->count[i], cpu);
2183 
2184 		per_cpu(memcg->stat->count[i], cpu) = 0;
2185 		memcg->nocpu_base.count[i] += x;
2186 	}
2187 	for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2188 		unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2189 
2190 		per_cpu(memcg->stat->events[i], cpu) = 0;
2191 		memcg->nocpu_base.events[i] += x;
2192 	}
2193 	spin_unlock(&memcg->pcp_counter_lock);
2194 }
2195 
2196 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2197 					unsigned long action,
2198 					void *hcpu)
2199 {
2200 	int cpu = (unsigned long)hcpu;
2201 	struct memcg_stock_pcp *stock;
2202 	struct mem_cgroup *iter;
2203 
2204 	if (action == CPU_ONLINE)
2205 		return NOTIFY_OK;
2206 
2207 	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2208 		return NOTIFY_OK;
2209 
2210 	for_each_mem_cgroup(iter)
2211 		mem_cgroup_drain_pcp_counter(iter, cpu);
2212 
2213 	stock = &per_cpu(memcg_stock, cpu);
2214 	drain_stock(stock);
2215 	return NOTIFY_OK;
2216 }
2217 
2218 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2219 		      unsigned int nr_pages)
2220 {
2221 	unsigned int batch = max(CHARGE_BATCH, nr_pages);
2222 	int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2223 	struct mem_cgroup *mem_over_limit;
2224 	struct page_counter *counter;
2225 	unsigned long nr_reclaimed;
2226 	bool may_swap = true;
2227 	bool drained = false;
2228 	int ret = 0;
2229 
2230 	if (mem_cgroup_is_root(memcg))
2231 		goto done;
2232 retry:
2233 	if (consume_stock(memcg, nr_pages))
2234 		goto done;
2235 
2236 	if (!do_swap_account ||
2237 	    !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2238 		if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2239 			goto done_restock;
2240 		if (do_swap_account)
2241 			page_counter_uncharge(&memcg->memsw, batch);
2242 		mem_over_limit = mem_cgroup_from_counter(counter, memory);
2243 	} else {
2244 		mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2245 		may_swap = false;
2246 	}
2247 
2248 	if (batch > nr_pages) {
2249 		batch = nr_pages;
2250 		goto retry;
2251 	}
2252 
2253 	/*
2254 	 * Unlike in global OOM situations, memcg is not in a physical
2255 	 * memory shortage.  Allow dying and OOM-killed tasks to
2256 	 * bypass the last charges so that they can exit quickly and
2257 	 * free their memory.
2258 	 */
2259 	if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2260 		     fatal_signal_pending(current) ||
2261 		     current->flags & PF_EXITING))
2262 		goto bypass;
2263 
2264 	if (unlikely(task_in_memcg_oom(current)))
2265 		goto nomem;
2266 
2267 	if (!(gfp_mask & __GFP_WAIT))
2268 		goto nomem;
2269 
2270 	mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2271 
2272 	nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2273 						    gfp_mask, may_swap);
2274 
2275 	if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2276 		goto retry;
2277 
2278 	if (!drained) {
2279 		drain_all_stock(mem_over_limit);
2280 		drained = true;
2281 		goto retry;
2282 	}
2283 
2284 	if (gfp_mask & __GFP_NORETRY)
2285 		goto nomem;
2286 	/*
2287 	 * Even though the limit is exceeded at this point, reclaim
2288 	 * may have been able to free some pages.  Retry the charge
2289 	 * before killing the task.
2290 	 *
2291 	 * Only for regular pages, though: huge pages are rather
2292 	 * unlikely to succeed so close to the limit, and we fall back
2293 	 * to regular pages anyway in case of failure.
2294 	 */
2295 	if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2296 		goto retry;
2297 	/*
2298 	 * At task move, charge accounts can be doubly counted. So, it's
2299 	 * better to wait until the end of task_move if something is going on.
2300 	 */
2301 	if (mem_cgroup_wait_acct_move(mem_over_limit))
2302 		goto retry;
2303 
2304 	if (nr_retries--)
2305 		goto retry;
2306 
2307 	if (gfp_mask & __GFP_NOFAIL)
2308 		goto bypass;
2309 
2310 	if (fatal_signal_pending(current))
2311 		goto bypass;
2312 
2313 	mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2314 
2315 	mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2316 nomem:
2317 	if (!(gfp_mask & __GFP_NOFAIL))
2318 		return -ENOMEM;
2319 bypass:
2320 	return -EINTR;
2321 
2322 done_restock:
2323 	css_get_many(&memcg->css, batch);
2324 	if (batch > nr_pages)
2325 		refill_stock(memcg, batch - nr_pages);
2326 	/*
2327 	 * If the hierarchy is above the normal consumption range,
2328 	 * make the charging task trim their excess contribution.
2329 	 */
2330 	do {
2331 		if (page_counter_read(&memcg->memory) <= memcg->high)
2332 			continue;
2333 		mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2334 		try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2335 	} while ((memcg = parent_mem_cgroup(memcg)));
2336 done:
2337 	return ret;
2338 }
2339 
2340 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2341 {
2342 	if (mem_cgroup_is_root(memcg))
2343 		return;
2344 
2345 	page_counter_uncharge(&memcg->memory, nr_pages);
2346 	if (do_swap_account)
2347 		page_counter_uncharge(&memcg->memsw, nr_pages);
2348 
2349 	css_put_many(&memcg->css, nr_pages);
2350 }
2351 
2352 /*
2353  * try_get_mem_cgroup_from_page - look up page's memcg association
2354  * @page: the page
2355  *
2356  * Look up, get a css reference, and return the memcg that owns @page.
2357  *
2358  * The page must be locked to prevent racing with swap-in and page
2359  * cache charges.  If coming from an unlocked page table, the caller
2360  * must ensure the page is on the LRU or this can race with charging.
2361  */
2362 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2363 {
2364 	struct mem_cgroup *memcg;
2365 	unsigned short id;
2366 	swp_entry_t ent;
2367 
2368 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2369 
2370 	memcg = page->mem_cgroup;
2371 	if (memcg) {
2372 		if (!css_tryget_online(&memcg->css))
2373 			memcg = NULL;
2374 	} else if (PageSwapCache(page)) {
2375 		ent.val = page_private(page);
2376 		id = lookup_swap_cgroup_id(ent);
2377 		rcu_read_lock();
2378 		memcg = mem_cgroup_from_id(id);
2379 		if (memcg && !css_tryget_online(&memcg->css))
2380 			memcg = NULL;
2381 		rcu_read_unlock();
2382 	}
2383 	return memcg;
2384 }
2385 
2386 static void lock_page_lru(struct page *page, int *isolated)
2387 {
2388 	struct zone *zone = page_zone(page);
2389 
2390 	spin_lock_irq(&zone->lru_lock);
2391 	if (PageLRU(page)) {
2392 		struct lruvec *lruvec;
2393 
2394 		lruvec = mem_cgroup_page_lruvec(page, zone);
2395 		ClearPageLRU(page);
2396 		del_page_from_lru_list(page, lruvec, page_lru(page));
2397 		*isolated = 1;
2398 	} else
2399 		*isolated = 0;
2400 }
2401 
2402 static void unlock_page_lru(struct page *page, int isolated)
2403 {
2404 	struct zone *zone = page_zone(page);
2405 
2406 	if (isolated) {
2407 		struct lruvec *lruvec;
2408 
2409 		lruvec = mem_cgroup_page_lruvec(page, zone);
2410 		VM_BUG_ON_PAGE(PageLRU(page), page);
2411 		SetPageLRU(page);
2412 		add_page_to_lru_list(page, lruvec, page_lru(page));
2413 	}
2414 	spin_unlock_irq(&zone->lru_lock);
2415 }
2416 
2417 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2418 			  bool lrucare)
2419 {
2420 	int isolated;
2421 
2422 	VM_BUG_ON_PAGE(page->mem_cgroup, page);
2423 
2424 	/*
2425 	 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2426 	 * may already be on some other mem_cgroup's LRU.  Take care of it.
2427 	 */
2428 	if (lrucare)
2429 		lock_page_lru(page, &isolated);
2430 
2431 	/*
2432 	 * Nobody should be changing or seriously looking at
2433 	 * page->mem_cgroup at this point:
2434 	 *
2435 	 * - the page is uncharged
2436 	 *
2437 	 * - the page is off-LRU
2438 	 *
2439 	 * - an anonymous fault has exclusive page access, except for
2440 	 *   a locked page table
2441 	 *
2442 	 * - a page cache insertion, a swapin fault, or a migration
2443 	 *   have the page locked
2444 	 */
2445 	page->mem_cgroup = memcg;
2446 
2447 	if (lrucare)
2448 		unlock_page_lru(page, isolated);
2449 }
2450 
2451 #ifdef CONFIG_MEMCG_KMEM
2452 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2453 		      unsigned long nr_pages)
2454 {
2455 	struct page_counter *counter;
2456 	int ret = 0;
2457 
2458 	ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2459 	if (ret < 0)
2460 		return ret;
2461 
2462 	ret = try_charge(memcg, gfp, nr_pages);
2463 	if (ret == -EINTR)  {
2464 		/*
2465 		 * try_charge() chose to bypass to root due to OOM kill or
2466 		 * fatal signal.  Since our only options are to either fail
2467 		 * the allocation or charge it to this cgroup, do it as a
2468 		 * temporary condition. But we can't fail. From a kmem/slab
2469 		 * perspective, the cache has already been selected, by
2470 		 * mem_cgroup_kmem_get_cache(), so it is too late to change
2471 		 * our minds.
2472 		 *
2473 		 * This condition will only trigger if the task entered
2474 		 * memcg_charge_kmem in a sane state, but was OOM-killed
2475 		 * during try_charge() above. Tasks that were already dying
2476 		 * when the allocation triggers should have been already
2477 		 * directed to the root cgroup in memcontrol.h
2478 		 */
2479 		page_counter_charge(&memcg->memory, nr_pages);
2480 		if (do_swap_account)
2481 			page_counter_charge(&memcg->memsw, nr_pages);
2482 		css_get_many(&memcg->css, nr_pages);
2483 		ret = 0;
2484 	} else if (ret)
2485 		page_counter_uncharge(&memcg->kmem, nr_pages);
2486 
2487 	return ret;
2488 }
2489 
2490 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2491 {
2492 	page_counter_uncharge(&memcg->memory, nr_pages);
2493 	if (do_swap_account)
2494 		page_counter_uncharge(&memcg->memsw, nr_pages);
2495 
2496 	page_counter_uncharge(&memcg->kmem, nr_pages);
2497 
2498 	css_put_many(&memcg->css, nr_pages);
2499 }
2500 
2501 /*
2502  * helper for acessing a memcg's index. It will be used as an index in the
2503  * child cache array in kmem_cache, and also to derive its name. This function
2504  * will return -1 when this is not a kmem-limited memcg.
2505  */
2506 int memcg_cache_id(struct mem_cgroup *memcg)
2507 {
2508 	return memcg ? memcg->kmemcg_id : -1;
2509 }
2510 
2511 static int memcg_alloc_cache_id(void)
2512 {
2513 	int id, size;
2514 	int err;
2515 
2516 	id = ida_simple_get(&memcg_cache_ida,
2517 			    0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2518 	if (id < 0)
2519 		return id;
2520 
2521 	if (id < memcg_nr_cache_ids)
2522 		return id;
2523 
2524 	/*
2525 	 * There's no space for the new id in memcg_caches arrays,
2526 	 * so we have to grow them.
2527 	 */
2528 	down_write(&memcg_cache_ids_sem);
2529 
2530 	size = 2 * (id + 1);
2531 	if (size < MEMCG_CACHES_MIN_SIZE)
2532 		size = MEMCG_CACHES_MIN_SIZE;
2533 	else if (size > MEMCG_CACHES_MAX_SIZE)
2534 		size = MEMCG_CACHES_MAX_SIZE;
2535 
2536 	err = memcg_update_all_caches(size);
2537 	if (!err)
2538 		err = memcg_update_all_list_lrus(size);
2539 	if (!err)
2540 		memcg_nr_cache_ids = size;
2541 
2542 	up_write(&memcg_cache_ids_sem);
2543 
2544 	if (err) {
2545 		ida_simple_remove(&memcg_cache_ida, id);
2546 		return err;
2547 	}
2548 	return id;
2549 }
2550 
2551 static void memcg_free_cache_id(int id)
2552 {
2553 	ida_simple_remove(&memcg_cache_ida, id);
2554 }
2555 
2556 struct memcg_kmem_cache_create_work {
2557 	struct mem_cgroup *memcg;
2558 	struct kmem_cache *cachep;
2559 	struct work_struct work;
2560 };
2561 
2562 static void memcg_kmem_cache_create_func(struct work_struct *w)
2563 {
2564 	struct memcg_kmem_cache_create_work *cw =
2565 		container_of(w, struct memcg_kmem_cache_create_work, work);
2566 	struct mem_cgroup *memcg = cw->memcg;
2567 	struct kmem_cache *cachep = cw->cachep;
2568 
2569 	memcg_create_kmem_cache(memcg, cachep);
2570 
2571 	css_put(&memcg->css);
2572 	kfree(cw);
2573 }
2574 
2575 /*
2576  * Enqueue the creation of a per-memcg kmem_cache.
2577  */
2578 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2579 					       struct kmem_cache *cachep)
2580 {
2581 	struct memcg_kmem_cache_create_work *cw;
2582 
2583 	cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2584 	if (!cw)
2585 		return;
2586 
2587 	css_get(&memcg->css);
2588 
2589 	cw->memcg = memcg;
2590 	cw->cachep = cachep;
2591 	INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2592 
2593 	schedule_work(&cw->work);
2594 }
2595 
2596 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2597 					     struct kmem_cache *cachep)
2598 {
2599 	/*
2600 	 * We need to stop accounting when we kmalloc, because if the
2601 	 * corresponding kmalloc cache is not yet created, the first allocation
2602 	 * in __memcg_schedule_kmem_cache_create will recurse.
2603 	 *
2604 	 * However, it is better to enclose the whole function. Depending on
2605 	 * the debugging options enabled, INIT_WORK(), for instance, can
2606 	 * trigger an allocation. This too, will make us recurse. Because at
2607 	 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2608 	 * the safest choice is to do it like this, wrapping the whole function.
2609 	 */
2610 	current->memcg_kmem_skip_account = 1;
2611 	__memcg_schedule_kmem_cache_create(memcg, cachep);
2612 	current->memcg_kmem_skip_account = 0;
2613 }
2614 
2615 /*
2616  * Return the kmem_cache we're supposed to use for a slab allocation.
2617  * We try to use the current memcg's version of the cache.
2618  *
2619  * If the cache does not exist yet, if we are the first user of it,
2620  * we either create it immediately, if possible, or create it asynchronously
2621  * in a workqueue.
2622  * In the latter case, we will let the current allocation go through with
2623  * the original cache.
2624  *
2625  * Can't be called in interrupt context or from kernel threads.
2626  * This function needs to be called with rcu_read_lock() held.
2627  */
2628 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2629 {
2630 	struct mem_cgroup *memcg;
2631 	struct kmem_cache *memcg_cachep;
2632 	int kmemcg_id;
2633 
2634 	VM_BUG_ON(!is_root_cache(cachep));
2635 
2636 	if (current->memcg_kmem_skip_account)
2637 		return cachep;
2638 
2639 	memcg = get_mem_cgroup_from_mm(current->mm);
2640 	kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2641 	if (kmemcg_id < 0)
2642 		goto out;
2643 
2644 	memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2645 	if (likely(memcg_cachep))
2646 		return memcg_cachep;
2647 
2648 	/*
2649 	 * If we are in a safe context (can wait, and not in interrupt
2650 	 * context), we could be be predictable and return right away.
2651 	 * This would guarantee that the allocation being performed
2652 	 * already belongs in the new cache.
2653 	 *
2654 	 * However, there are some clashes that can arrive from locking.
2655 	 * For instance, because we acquire the slab_mutex while doing
2656 	 * memcg_create_kmem_cache, this means no further allocation
2657 	 * could happen with the slab_mutex held. So it's better to
2658 	 * defer everything.
2659 	 */
2660 	memcg_schedule_kmem_cache_create(memcg, cachep);
2661 out:
2662 	css_put(&memcg->css);
2663 	return cachep;
2664 }
2665 
2666 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2667 {
2668 	if (!is_root_cache(cachep))
2669 		css_put(&cachep->memcg_params.memcg->css);
2670 }
2671 
2672 /*
2673  * We need to verify if the allocation against current->mm->owner's memcg is
2674  * possible for the given order. But the page is not allocated yet, so we'll
2675  * need a further commit step to do the final arrangements.
2676  *
2677  * It is possible for the task to switch cgroups in this mean time, so at
2678  * commit time, we can't rely on task conversion any longer.  We'll then use
2679  * the handle argument to return to the caller which cgroup we should commit
2680  * against. We could also return the memcg directly and avoid the pointer
2681  * passing, but a boolean return value gives better semantics considering
2682  * the compiled-out case as well.
2683  *
2684  * Returning true means the allocation is possible.
2685  */
2686 bool
2687 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2688 {
2689 	struct mem_cgroup *memcg;
2690 	int ret;
2691 
2692 	*_memcg = NULL;
2693 
2694 	memcg = get_mem_cgroup_from_mm(current->mm);
2695 
2696 	if (!memcg_kmem_is_active(memcg)) {
2697 		css_put(&memcg->css);
2698 		return true;
2699 	}
2700 
2701 	ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2702 	if (!ret)
2703 		*_memcg = memcg;
2704 
2705 	css_put(&memcg->css);
2706 	return (ret == 0);
2707 }
2708 
2709 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2710 			      int order)
2711 {
2712 	VM_BUG_ON(mem_cgroup_is_root(memcg));
2713 
2714 	/* The page allocation failed. Revert */
2715 	if (!page) {
2716 		memcg_uncharge_kmem(memcg, 1 << order);
2717 		return;
2718 	}
2719 	page->mem_cgroup = memcg;
2720 }
2721 
2722 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2723 {
2724 	struct mem_cgroup *memcg = page->mem_cgroup;
2725 
2726 	if (!memcg)
2727 		return;
2728 
2729 	VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2730 
2731 	memcg_uncharge_kmem(memcg, 1 << order);
2732 	page->mem_cgroup = NULL;
2733 }
2734 
2735 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2736 {
2737 	struct mem_cgroup *memcg = NULL;
2738 	struct kmem_cache *cachep;
2739 	struct page *page;
2740 
2741 	page = virt_to_head_page(ptr);
2742 	if (PageSlab(page)) {
2743 		cachep = page->slab_cache;
2744 		if (!is_root_cache(cachep))
2745 			memcg = cachep->memcg_params.memcg;
2746 	} else
2747 		/* page allocated by alloc_kmem_pages */
2748 		memcg = page->mem_cgroup;
2749 
2750 	return memcg;
2751 }
2752 #endif /* CONFIG_MEMCG_KMEM */
2753 
2754 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2755 
2756 /*
2757  * Because tail pages are not marked as "used", set it. We're under
2758  * zone->lru_lock, 'splitting on pmd' and compound_lock.
2759  * charge/uncharge will be never happen and move_account() is done under
2760  * compound_lock(), so we don't have to take care of races.
2761  */
2762 void mem_cgroup_split_huge_fixup(struct page *head)
2763 {
2764 	int i;
2765 
2766 	if (mem_cgroup_disabled())
2767 		return;
2768 
2769 	for (i = 1; i < HPAGE_PMD_NR; i++)
2770 		head[i].mem_cgroup = head->mem_cgroup;
2771 
2772 	__this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2773 		       HPAGE_PMD_NR);
2774 }
2775 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2776 
2777 #ifdef CONFIG_MEMCG_SWAP
2778 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2779 					 bool charge)
2780 {
2781 	int val = (charge) ? 1 : -1;
2782 	this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2783 }
2784 
2785 /**
2786  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2787  * @entry: swap entry to be moved
2788  * @from:  mem_cgroup which the entry is moved from
2789  * @to:  mem_cgroup which the entry is moved to
2790  *
2791  * It succeeds only when the swap_cgroup's record for this entry is the same
2792  * as the mem_cgroup's id of @from.
2793  *
2794  * Returns 0 on success, -EINVAL on failure.
2795  *
2796  * The caller must have charged to @to, IOW, called page_counter_charge() about
2797  * both res and memsw, and called css_get().
2798  */
2799 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2800 				struct mem_cgroup *from, struct mem_cgroup *to)
2801 {
2802 	unsigned short old_id, new_id;
2803 
2804 	old_id = mem_cgroup_id(from);
2805 	new_id = mem_cgroup_id(to);
2806 
2807 	if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2808 		mem_cgroup_swap_statistics(from, false);
2809 		mem_cgroup_swap_statistics(to, true);
2810 		return 0;
2811 	}
2812 	return -EINVAL;
2813 }
2814 #else
2815 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2816 				struct mem_cgroup *from, struct mem_cgroup *to)
2817 {
2818 	return -EINVAL;
2819 }
2820 #endif
2821 
2822 static DEFINE_MUTEX(memcg_limit_mutex);
2823 
2824 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2825 				   unsigned long limit)
2826 {
2827 	unsigned long curusage;
2828 	unsigned long oldusage;
2829 	bool enlarge = false;
2830 	int retry_count;
2831 	int ret;
2832 
2833 	/*
2834 	 * For keeping hierarchical_reclaim simple, how long we should retry
2835 	 * is depends on callers. We set our retry-count to be function
2836 	 * of # of children which we should visit in this loop.
2837 	 */
2838 	retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2839 		      mem_cgroup_count_children(memcg);
2840 
2841 	oldusage = page_counter_read(&memcg->memory);
2842 
2843 	do {
2844 		if (signal_pending(current)) {
2845 			ret = -EINTR;
2846 			break;
2847 		}
2848 
2849 		mutex_lock(&memcg_limit_mutex);
2850 		if (limit > memcg->memsw.limit) {
2851 			mutex_unlock(&memcg_limit_mutex);
2852 			ret = -EINVAL;
2853 			break;
2854 		}
2855 		if (limit > memcg->memory.limit)
2856 			enlarge = true;
2857 		ret = page_counter_limit(&memcg->memory, limit);
2858 		mutex_unlock(&memcg_limit_mutex);
2859 
2860 		if (!ret)
2861 			break;
2862 
2863 		try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2864 
2865 		curusage = page_counter_read(&memcg->memory);
2866 		/* Usage is reduced ? */
2867 		if (curusage >= oldusage)
2868 			retry_count--;
2869 		else
2870 			oldusage = curusage;
2871 	} while (retry_count);
2872 
2873 	if (!ret && enlarge)
2874 		memcg_oom_recover(memcg);
2875 
2876 	return ret;
2877 }
2878 
2879 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2880 					 unsigned long limit)
2881 {
2882 	unsigned long curusage;
2883 	unsigned long oldusage;
2884 	bool enlarge = false;
2885 	int retry_count;
2886 	int ret;
2887 
2888 	/* see mem_cgroup_resize_res_limit */
2889 	retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2890 		      mem_cgroup_count_children(memcg);
2891 
2892 	oldusage = page_counter_read(&memcg->memsw);
2893 
2894 	do {
2895 		if (signal_pending(current)) {
2896 			ret = -EINTR;
2897 			break;
2898 		}
2899 
2900 		mutex_lock(&memcg_limit_mutex);
2901 		if (limit < memcg->memory.limit) {
2902 			mutex_unlock(&memcg_limit_mutex);
2903 			ret = -EINVAL;
2904 			break;
2905 		}
2906 		if (limit > memcg->memsw.limit)
2907 			enlarge = true;
2908 		ret = page_counter_limit(&memcg->memsw, limit);
2909 		mutex_unlock(&memcg_limit_mutex);
2910 
2911 		if (!ret)
2912 			break;
2913 
2914 		try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2915 
2916 		curusage = page_counter_read(&memcg->memsw);
2917 		/* Usage is reduced ? */
2918 		if (curusage >= oldusage)
2919 			retry_count--;
2920 		else
2921 			oldusage = curusage;
2922 	} while (retry_count);
2923 
2924 	if (!ret && enlarge)
2925 		memcg_oom_recover(memcg);
2926 
2927 	return ret;
2928 }
2929 
2930 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2931 					    gfp_t gfp_mask,
2932 					    unsigned long *total_scanned)
2933 {
2934 	unsigned long nr_reclaimed = 0;
2935 	struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2936 	unsigned long reclaimed;
2937 	int loop = 0;
2938 	struct mem_cgroup_tree_per_zone *mctz;
2939 	unsigned long excess;
2940 	unsigned long nr_scanned;
2941 
2942 	if (order > 0)
2943 		return 0;
2944 
2945 	mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2946 	/*
2947 	 * This loop can run a while, specially if mem_cgroup's continuously
2948 	 * keep exceeding their soft limit and putting the system under
2949 	 * pressure
2950 	 */
2951 	do {
2952 		if (next_mz)
2953 			mz = next_mz;
2954 		else
2955 			mz = mem_cgroup_largest_soft_limit_node(mctz);
2956 		if (!mz)
2957 			break;
2958 
2959 		nr_scanned = 0;
2960 		reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2961 						    gfp_mask, &nr_scanned);
2962 		nr_reclaimed += reclaimed;
2963 		*total_scanned += nr_scanned;
2964 		spin_lock_irq(&mctz->lock);
2965 		__mem_cgroup_remove_exceeded(mz, mctz);
2966 
2967 		/*
2968 		 * If we failed to reclaim anything from this memory cgroup
2969 		 * it is time to move on to the next cgroup
2970 		 */
2971 		next_mz = NULL;
2972 		if (!reclaimed)
2973 			next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2974 
2975 		excess = soft_limit_excess(mz->memcg);
2976 		/*
2977 		 * One school of thought says that we should not add
2978 		 * back the node to the tree if reclaim returns 0.
2979 		 * But our reclaim could return 0, simply because due
2980 		 * to priority we are exposing a smaller subset of
2981 		 * memory to reclaim from. Consider this as a longer
2982 		 * term TODO.
2983 		 */
2984 		/* If excess == 0, no tree ops */
2985 		__mem_cgroup_insert_exceeded(mz, mctz, excess);
2986 		spin_unlock_irq(&mctz->lock);
2987 		css_put(&mz->memcg->css);
2988 		loop++;
2989 		/*
2990 		 * Could not reclaim anything and there are no more
2991 		 * mem cgroups to try or we seem to be looping without
2992 		 * reclaiming anything.
2993 		 */
2994 		if (!nr_reclaimed &&
2995 			(next_mz == NULL ||
2996 			loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2997 			break;
2998 	} while (!nr_reclaimed);
2999 	if (next_mz)
3000 		css_put(&next_mz->memcg->css);
3001 	return nr_reclaimed;
3002 }
3003 
3004 /*
3005  * Test whether @memcg has children, dead or alive.  Note that this
3006  * function doesn't care whether @memcg has use_hierarchy enabled and
3007  * returns %true if there are child csses according to the cgroup
3008  * hierarchy.  Testing use_hierarchy is the caller's responsiblity.
3009  */
3010 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3011 {
3012 	bool ret;
3013 
3014 	/*
3015 	 * The lock does not prevent addition or deletion of children, but
3016 	 * it prevents a new child from being initialized based on this
3017 	 * parent in css_online(), so it's enough to decide whether
3018 	 * hierarchically inherited attributes can still be changed or not.
3019 	 */
3020 	lockdep_assert_held(&memcg_create_mutex);
3021 
3022 	rcu_read_lock();
3023 	ret = css_next_child(NULL, &memcg->css);
3024 	rcu_read_unlock();
3025 	return ret;
3026 }
3027 
3028 /*
3029  * Reclaims as many pages from the given memcg as possible and moves
3030  * the rest to the parent.
3031  *
3032  * Caller is responsible for holding css reference for memcg.
3033  */
3034 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3035 {
3036 	int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3037 
3038 	/* we call try-to-free pages for make this cgroup empty */
3039 	lru_add_drain_all();
3040 	/* try to free all pages in this cgroup */
3041 	while (nr_retries && page_counter_read(&memcg->memory)) {
3042 		int progress;
3043 
3044 		if (signal_pending(current))
3045 			return -EINTR;
3046 
3047 		progress = try_to_free_mem_cgroup_pages(memcg, 1,
3048 							GFP_KERNEL, true);
3049 		if (!progress) {
3050 			nr_retries--;
3051 			/* maybe some writeback is necessary */
3052 			congestion_wait(BLK_RW_ASYNC, HZ/10);
3053 		}
3054 
3055 	}
3056 
3057 	return 0;
3058 }
3059 
3060 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3061 					    char *buf, size_t nbytes,
3062 					    loff_t off)
3063 {
3064 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3065 
3066 	if (mem_cgroup_is_root(memcg))
3067 		return -EINVAL;
3068 	return mem_cgroup_force_empty(memcg) ?: nbytes;
3069 }
3070 
3071 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3072 				     struct cftype *cft)
3073 {
3074 	return mem_cgroup_from_css(css)->use_hierarchy;
3075 }
3076 
3077 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3078 				      struct cftype *cft, u64 val)
3079 {
3080 	int retval = 0;
3081 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3082 	struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3083 
3084 	mutex_lock(&memcg_create_mutex);
3085 
3086 	if (memcg->use_hierarchy == val)
3087 		goto out;
3088 
3089 	/*
3090 	 * If parent's use_hierarchy is set, we can't make any modifications
3091 	 * in the child subtrees. If it is unset, then the change can
3092 	 * occur, provided the current cgroup has no children.
3093 	 *
3094 	 * For the root cgroup, parent_mem is NULL, we allow value to be
3095 	 * set if there are no children.
3096 	 */
3097 	if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3098 				(val == 1 || val == 0)) {
3099 		if (!memcg_has_children(memcg))
3100 			memcg->use_hierarchy = val;
3101 		else
3102 			retval = -EBUSY;
3103 	} else
3104 		retval = -EINVAL;
3105 
3106 out:
3107 	mutex_unlock(&memcg_create_mutex);
3108 
3109 	return retval;
3110 }
3111 
3112 static unsigned long tree_stat(struct mem_cgroup *memcg,
3113 			       enum mem_cgroup_stat_index idx)
3114 {
3115 	struct mem_cgroup *iter;
3116 	long val = 0;
3117 
3118 	/* Per-cpu values can be negative, use a signed accumulator */
3119 	for_each_mem_cgroup_tree(iter, memcg)
3120 		val += mem_cgroup_read_stat(iter, idx);
3121 
3122 	if (val < 0) /* race ? */
3123 		val = 0;
3124 	return val;
3125 }
3126 
3127 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3128 {
3129 	u64 val;
3130 
3131 	if (mem_cgroup_is_root(memcg)) {
3132 		val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3133 		val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3134 		if (swap)
3135 			val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3136 	} else {
3137 		if (!swap)
3138 			val = page_counter_read(&memcg->memory);
3139 		else
3140 			val = page_counter_read(&memcg->memsw);
3141 	}
3142 	return val << PAGE_SHIFT;
3143 }
3144 
3145 enum {
3146 	RES_USAGE,
3147 	RES_LIMIT,
3148 	RES_MAX_USAGE,
3149 	RES_FAILCNT,
3150 	RES_SOFT_LIMIT,
3151 };
3152 
3153 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3154 			       struct cftype *cft)
3155 {
3156 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3157 	struct page_counter *counter;
3158 
3159 	switch (MEMFILE_TYPE(cft->private)) {
3160 	case _MEM:
3161 		counter = &memcg->memory;
3162 		break;
3163 	case _MEMSWAP:
3164 		counter = &memcg->memsw;
3165 		break;
3166 	case _KMEM:
3167 		counter = &memcg->kmem;
3168 		break;
3169 	default:
3170 		BUG();
3171 	}
3172 
3173 	switch (MEMFILE_ATTR(cft->private)) {
3174 	case RES_USAGE:
3175 		if (counter == &memcg->memory)
3176 			return mem_cgroup_usage(memcg, false);
3177 		if (counter == &memcg->memsw)
3178 			return mem_cgroup_usage(memcg, true);
3179 		return (u64)page_counter_read(counter) * PAGE_SIZE;
3180 	case RES_LIMIT:
3181 		return (u64)counter->limit * PAGE_SIZE;
3182 	case RES_MAX_USAGE:
3183 		return (u64)counter->watermark * PAGE_SIZE;
3184 	case RES_FAILCNT:
3185 		return counter->failcnt;
3186 	case RES_SOFT_LIMIT:
3187 		return (u64)memcg->soft_limit * PAGE_SIZE;
3188 	default:
3189 		BUG();
3190 	}
3191 }
3192 
3193 #ifdef CONFIG_MEMCG_KMEM
3194 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3195 			       unsigned long nr_pages)
3196 {
3197 	int err = 0;
3198 	int memcg_id;
3199 
3200 	BUG_ON(memcg->kmemcg_id >= 0);
3201 	BUG_ON(memcg->kmem_acct_activated);
3202 	BUG_ON(memcg->kmem_acct_active);
3203 
3204 	/*
3205 	 * For simplicity, we won't allow this to be disabled.  It also can't
3206 	 * be changed if the cgroup has children already, or if tasks had
3207 	 * already joined.
3208 	 *
3209 	 * If tasks join before we set the limit, a person looking at
3210 	 * kmem.usage_in_bytes will have no way to determine when it took
3211 	 * place, which makes the value quite meaningless.
3212 	 *
3213 	 * After it first became limited, changes in the value of the limit are
3214 	 * of course permitted.
3215 	 */
3216 	mutex_lock(&memcg_create_mutex);
3217 	if (cgroup_has_tasks(memcg->css.cgroup) ||
3218 	    (memcg->use_hierarchy && memcg_has_children(memcg)))
3219 		err = -EBUSY;
3220 	mutex_unlock(&memcg_create_mutex);
3221 	if (err)
3222 		goto out;
3223 
3224 	memcg_id = memcg_alloc_cache_id();
3225 	if (memcg_id < 0) {
3226 		err = memcg_id;
3227 		goto out;
3228 	}
3229 
3230 	/*
3231 	 * We couldn't have accounted to this cgroup, because it hasn't got
3232 	 * activated yet, so this should succeed.
3233 	 */
3234 	err = page_counter_limit(&memcg->kmem, nr_pages);
3235 	VM_BUG_ON(err);
3236 
3237 	static_key_slow_inc(&memcg_kmem_enabled_key);
3238 	/*
3239 	 * A memory cgroup is considered kmem-active as soon as it gets
3240 	 * kmemcg_id. Setting the id after enabling static branching will
3241 	 * guarantee no one starts accounting before all call sites are
3242 	 * patched.
3243 	 */
3244 	memcg->kmemcg_id = memcg_id;
3245 	memcg->kmem_acct_activated = true;
3246 	memcg->kmem_acct_active = true;
3247 out:
3248 	return err;
3249 }
3250 
3251 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3252 				   unsigned long limit)
3253 {
3254 	int ret;
3255 
3256 	mutex_lock(&memcg_limit_mutex);
3257 	if (!memcg_kmem_is_active(memcg))
3258 		ret = memcg_activate_kmem(memcg, limit);
3259 	else
3260 		ret = page_counter_limit(&memcg->kmem, limit);
3261 	mutex_unlock(&memcg_limit_mutex);
3262 	return ret;
3263 }
3264 
3265 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3266 {
3267 	int ret = 0;
3268 	struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3269 
3270 	if (!parent)
3271 		return 0;
3272 
3273 	mutex_lock(&memcg_limit_mutex);
3274 	/*
3275 	 * If the parent cgroup is not kmem-active now, it cannot be activated
3276 	 * after this point, because it has at least one child already.
3277 	 */
3278 	if (memcg_kmem_is_active(parent))
3279 		ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3280 	mutex_unlock(&memcg_limit_mutex);
3281 	return ret;
3282 }
3283 #else
3284 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3285 				   unsigned long limit)
3286 {
3287 	return -EINVAL;
3288 }
3289 #endif /* CONFIG_MEMCG_KMEM */
3290 
3291 /*
3292  * The user of this function is...
3293  * RES_LIMIT.
3294  */
3295 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3296 				char *buf, size_t nbytes, loff_t off)
3297 {
3298 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3299 	unsigned long nr_pages;
3300 	int ret;
3301 
3302 	buf = strstrip(buf);
3303 	ret = page_counter_memparse(buf, "-1", &nr_pages);
3304 	if (ret)
3305 		return ret;
3306 
3307 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
3308 	case RES_LIMIT:
3309 		if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3310 			ret = -EINVAL;
3311 			break;
3312 		}
3313 		switch (MEMFILE_TYPE(of_cft(of)->private)) {
3314 		case _MEM:
3315 			ret = mem_cgroup_resize_limit(memcg, nr_pages);
3316 			break;
3317 		case _MEMSWAP:
3318 			ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3319 			break;
3320 		case _KMEM:
3321 			ret = memcg_update_kmem_limit(memcg, nr_pages);
3322 			break;
3323 		}
3324 		break;
3325 	case RES_SOFT_LIMIT:
3326 		memcg->soft_limit = nr_pages;
3327 		ret = 0;
3328 		break;
3329 	}
3330 	return ret ?: nbytes;
3331 }
3332 
3333 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3334 				size_t nbytes, loff_t off)
3335 {
3336 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3337 	struct page_counter *counter;
3338 
3339 	switch (MEMFILE_TYPE(of_cft(of)->private)) {
3340 	case _MEM:
3341 		counter = &memcg->memory;
3342 		break;
3343 	case _MEMSWAP:
3344 		counter = &memcg->memsw;
3345 		break;
3346 	case _KMEM:
3347 		counter = &memcg->kmem;
3348 		break;
3349 	default:
3350 		BUG();
3351 	}
3352 
3353 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
3354 	case RES_MAX_USAGE:
3355 		page_counter_reset_watermark(counter);
3356 		break;
3357 	case RES_FAILCNT:
3358 		counter->failcnt = 0;
3359 		break;
3360 	default:
3361 		BUG();
3362 	}
3363 
3364 	return nbytes;
3365 }
3366 
3367 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3368 					struct cftype *cft)
3369 {
3370 	return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3371 }
3372 
3373 #ifdef CONFIG_MMU
3374 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3375 					struct cftype *cft, u64 val)
3376 {
3377 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3378 
3379 	if (val & ~MOVE_MASK)
3380 		return -EINVAL;
3381 
3382 	/*
3383 	 * No kind of locking is needed in here, because ->can_attach() will
3384 	 * check this value once in the beginning of the process, and then carry
3385 	 * on with stale data. This means that changes to this value will only
3386 	 * affect task migrations starting after the change.
3387 	 */
3388 	memcg->move_charge_at_immigrate = val;
3389 	return 0;
3390 }
3391 #else
3392 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3393 					struct cftype *cft, u64 val)
3394 {
3395 	return -ENOSYS;
3396 }
3397 #endif
3398 
3399 #ifdef CONFIG_NUMA
3400 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3401 {
3402 	struct numa_stat {
3403 		const char *name;
3404 		unsigned int lru_mask;
3405 	};
3406 
3407 	static const struct numa_stat stats[] = {
3408 		{ "total", LRU_ALL },
3409 		{ "file", LRU_ALL_FILE },
3410 		{ "anon", LRU_ALL_ANON },
3411 		{ "unevictable", BIT(LRU_UNEVICTABLE) },
3412 	};
3413 	const struct numa_stat *stat;
3414 	int nid;
3415 	unsigned long nr;
3416 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3417 
3418 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3419 		nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3420 		seq_printf(m, "%s=%lu", stat->name, nr);
3421 		for_each_node_state(nid, N_MEMORY) {
3422 			nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3423 							  stat->lru_mask);
3424 			seq_printf(m, " N%d=%lu", nid, nr);
3425 		}
3426 		seq_putc(m, '\n');
3427 	}
3428 
3429 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3430 		struct mem_cgroup *iter;
3431 
3432 		nr = 0;
3433 		for_each_mem_cgroup_tree(iter, memcg)
3434 			nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3435 		seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3436 		for_each_node_state(nid, N_MEMORY) {
3437 			nr = 0;
3438 			for_each_mem_cgroup_tree(iter, memcg)
3439 				nr += mem_cgroup_node_nr_lru_pages(
3440 					iter, nid, stat->lru_mask);
3441 			seq_printf(m, " N%d=%lu", nid, nr);
3442 		}
3443 		seq_putc(m, '\n');
3444 	}
3445 
3446 	return 0;
3447 }
3448 #endif /* CONFIG_NUMA */
3449 
3450 static int memcg_stat_show(struct seq_file *m, void *v)
3451 {
3452 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3453 	unsigned long memory, memsw;
3454 	struct mem_cgroup *mi;
3455 	unsigned int i;
3456 
3457 	BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3458 		     MEM_CGROUP_STAT_NSTATS);
3459 	BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3460 		     MEM_CGROUP_EVENTS_NSTATS);
3461 	BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3462 
3463 	for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3464 		if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3465 			continue;
3466 		seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3467 			   mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3468 	}
3469 
3470 	for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3471 		seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3472 			   mem_cgroup_read_events(memcg, i));
3473 
3474 	for (i = 0; i < NR_LRU_LISTS; i++)
3475 		seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3476 			   mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3477 
3478 	/* Hierarchical information */
3479 	memory = memsw = PAGE_COUNTER_MAX;
3480 	for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3481 		memory = min(memory, mi->memory.limit);
3482 		memsw = min(memsw, mi->memsw.limit);
3483 	}
3484 	seq_printf(m, "hierarchical_memory_limit %llu\n",
3485 		   (u64)memory * PAGE_SIZE);
3486 	if (do_swap_account)
3487 		seq_printf(m, "hierarchical_memsw_limit %llu\n",
3488 			   (u64)memsw * PAGE_SIZE);
3489 
3490 	for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3491 		long long val = 0;
3492 
3493 		if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3494 			continue;
3495 		for_each_mem_cgroup_tree(mi, memcg)
3496 			val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3497 		seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3498 	}
3499 
3500 	for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3501 		unsigned long long val = 0;
3502 
3503 		for_each_mem_cgroup_tree(mi, memcg)
3504 			val += mem_cgroup_read_events(mi, i);
3505 		seq_printf(m, "total_%s %llu\n",
3506 			   mem_cgroup_events_names[i], val);
3507 	}
3508 
3509 	for (i = 0; i < NR_LRU_LISTS; i++) {
3510 		unsigned long long val = 0;
3511 
3512 		for_each_mem_cgroup_tree(mi, memcg)
3513 			val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3514 		seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3515 	}
3516 
3517 #ifdef CONFIG_DEBUG_VM
3518 	{
3519 		int nid, zid;
3520 		struct mem_cgroup_per_zone *mz;
3521 		struct zone_reclaim_stat *rstat;
3522 		unsigned long recent_rotated[2] = {0, 0};
3523 		unsigned long recent_scanned[2] = {0, 0};
3524 
3525 		for_each_online_node(nid)
3526 			for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3527 				mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3528 				rstat = &mz->lruvec.reclaim_stat;
3529 
3530 				recent_rotated[0] += rstat->recent_rotated[0];
3531 				recent_rotated[1] += rstat->recent_rotated[1];
3532 				recent_scanned[0] += rstat->recent_scanned[0];
3533 				recent_scanned[1] += rstat->recent_scanned[1];
3534 			}
3535 		seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3536 		seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3537 		seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3538 		seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3539 	}
3540 #endif
3541 
3542 	return 0;
3543 }
3544 
3545 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3546 				      struct cftype *cft)
3547 {
3548 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3549 
3550 	return mem_cgroup_swappiness(memcg);
3551 }
3552 
3553 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3554 				       struct cftype *cft, u64 val)
3555 {
3556 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3557 
3558 	if (val > 100)
3559 		return -EINVAL;
3560 
3561 	if (css->parent)
3562 		memcg->swappiness = val;
3563 	else
3564 		vm_swappiness = val;
3565 
3566 	return 0;
3567 }
3568 
3569 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3570 {
3571 	struct mem_cgroup_threshold_ary *t;
3572 	unsigned long usage;
3573 	int i;
3574 
3575 	rcu_read_lock();
3576 	if (!swap)
3577 		t = rcu_dereference(memcg->thresholds.primary);
3578 	else
3579 		t = rcu_dereference(memcg->memsw_thresholds.primary);
3580 
3581 	if (!t)
3582 		goto unlock;
3583 
3584 	usage = mem_cgroup_usage(memcg, swap);
3585 
3586 	/*
3587 	 * current_threshold points to threshold just below or equal to usage.
3588 	 * If it's not true, a threshold was crossed after last
3589 	 * call of __mem_cgroup_threshold().
3590 	 */
3591 	i = t->current_threshold;
3592 
3593 	/*
3594 	 * Iterate backward over array of thresholds starting from
3595 	 * current_threshold and check if a threshold is crossed.
3596 	 * If none of thresholds below usage is crossed, we read
3597 	 * only one element of the array here.
3598 	 */
3599 	for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3600 		eventfd_signal(t->entries[i].eventfd, 1);
3601 
3602 	/* i = current_threshold + 1 */
3603 	i++;
3604 
3605 	/*
3606 	 * Iterate forward over array of thresholds starting from
3607 	 * current_threshold+1 and check if a threshold is crossed.
3608 	 * If none of thresholds above usage is crossed, we read
3609 	 * only one element of the array here.
3610 	 */
3611 	for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3612 		eventfd_signal(t->entries[i].eventfd, 1);
3613 
3614 	/* Update current_threshold */
3615 	t->current_threshold = i - 1;
3616 unlock:
3617 	rcu_read_unlock();
3618 }
3619 
3620 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3621 {
3622 	while (memcg) {
3623 		__mem_cgroup_threshold(memcg, false);
3624 		if (do_swap_account)
3625 			__mem_cgroup_threshold(memcg, true);
3626 
3627 		memcg = parent_mem_cgroup(memcg);
3628 	}
3629 }
3630 
3631 static int compare_thresholds(const void *a, const void *b)
3632 {
3633 	const struct mem_cgroup_threshold *_a = a;
3634 	const struct mem_cgroup_threshold *_b = b;
3635 
3636 	if (_a->threshold > _b->threshold)
3637 		return 1;
3638 
3639 	if (_a->threshold < _b->threshold)
3640 		return -1;
3641 
3642 	return 0;
3643 }
3644 
3645 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3646 {
3647 	struct mem_cgroup_eventfd_list *ev;
3648 
3649 	spin_lock(&memcg_oom_lock);
3650 
3651 	list_for_each_entry(ev, &memcg->oom_notify, list)
3652 		eventfd_signal(ev->eventfd, 1);
3653 
3654 	spin_unlock(&memcg_oom_lock);
3655 	return 0;
3656 }
3657 
3658 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3659 {
3660 	struct mem_cgroup *iter;
3661 
3662 	for_each_mem_cgroup_tree(iter, memcg)
3663 		mem_cgroup_oom_notify_cb(iter);
3664 }
3665 
3666 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3667 	struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3668 {
3669 	struct mem_cgroup_thresholds *thresholds;
3670 	struct mem_cgroup_threshold_ary *new;
3671 	unsigned long threshold;
3672 	unsigned long usage;
3673 	int i, size, ret;
3674 
3675 	ret = page_counter_memparse(args, "-1", &threshold);
3676 	if (ret)
3677 		return ret;
3678 
3679 	mutex_lock(&memcg->thresholds_lock);
3680 
3681 	if (type == _MEM) {
3682 		thresholds = &memcg->thresholds;
3683 		usage = mem_cgroup_usage(memcg, false);
3684 	} else if (type == _MEMSWAP) {
3685 		thresholds = &memcg->memsw_thresholds;
3686 		usage = mem_cgroup_usage(memcg, true);
3687 	} else
3688 		BUG();
3689 
3690 	/* Check if a threshold crossed before adding a new one */
3691 	if (thresholds->primary)
3692 		__mem_cgroup_threshold(memcg, type == _MEMSWAP);
3693 
3694 	size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3695 
3696 	/* Allocate memory for new array of thresholds */
3697 	new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3698 			GFP_KERNEL);
3699 	if (!new) {
3700 		ret = -ENOMEM;
3701 		goto unlock;
3702 	}
3703 	new->size = size;
3704 
3705 	/* Copy thresholds (if any) to new array */
3706 	if (thresholds->primary) {
3707 		memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3708 				sizeof(struct mem_cgroup_threshold));
3709 	}
3710 
3711 	/* Add new threshold */
3712 	new->entries[size - 1].eventfd = eventfd;
3713 	new->entries[size - 1].threshold = threshold;
3714 
3715 	/* Sort thresholds. Registering of new threshold isn't time-critical */
3716 	sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3717 			compare_thresholds, NULL);
3718 
3719 	/* Find current threshold */
3720 	new->current_threshold = -1;
3721 	for (i = 0; i < size; i++) {
3722 		if (new->entries[i].threshold <= usage) {
3723 			/*
3724 			 * new->current_threshold will not be used until
3725 			 * rcu_assign_pointer(), so it's safe to increment
3726 			 * it here.
3727 			 */
3728 			++new->current_threshold;
3729 		} else
3730 			break;
3731 	}
3732 
3733 	/* Free old spare buffer and save old primary buffer as spare */
3734 	kfree(thresholds->spare);
3735 	thresholds->spare = thresholds->primary;
3736 
3737 	rcu_assign_pointer(thresholds->primary, new);
3738 
3739 	/* To be sure that nobody uses thresholds */
3740 	synchronize_rcu();
3741 
3742 unlock:
3743 	mutex_unlock(&memcg->thresholds_lock);
3744 
3745 	return ret;
3746 }
3747 
3748 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3749 	struct eventfd_ctx *eventfd, const char *args)
3750 {
3751 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3752 }
3753 
3754 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3755 	struct eventfd_ctx *eventfd, const char *args)
3756 {
3757 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3758 }
3759 
3760 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3761 	struct eventfd_ctx *eventfd, enum res_type type)
3762 {
3763 	struct mem_cgroup_thresholds *thresholds;
3764 	struct mem_cgroup_threshold_ary *new;
3765 	unsigned long usage;
3766 	int i, j, size;
3767 
3768 	mutex_lock(&memcg->thresholds_lock);
3769 
3770 	if (type == _MEM) {
3771 		thresholds = &memcg->thresholds;
3772 		usage = mem_cgroup_usage(memcg, false);
3773 	} else if (type == _MEMSWAP) {
3774 		thresholds = &memcg->memsw_thresholds;
3775 		usage = mem_cgroup_usage(memcg, true);
3776 	} else
3777 		BUG();
3778 
3779 	if (!thresholds->primary)
3780 		goto unlock;
3781 
3782 	/* Check if a threshold crossed before removing */
3783 	__mem_cgroup_threshold(memcg, type == _MEMSWAP);
3784 
3785 	/* Calculate new number of threshold */
3786 	size = 0;
3787 	for (i = 0; i < thresholds->primary->size; i++) {
3788 		if (thresholds->primary->entries[i].eventfd != eventfd)
3789 			size++;
3790 	}
3791 
3792 	new = thresholds->spare;
3793 
3794 	/* Set thresholds array to NULL if we don't have thresholds */
3795 	if (!size) {
3796 		kfree(new);
3797 		new = NULL;
3798 		goto swap_buffers;
3799 	}
3800 
3801 	new->size = size;
3802 
3803 	/* Copy thresholds and find current threshold */
3804 	new->current_threshold = -1;
3805 	for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3806 		if (thresholds->primary->entries[i].eventfd == eventfd)
3807 			continue;
3808 
3809 		new->entries[j] = thresholds->primary->entries[i];
3810 		if (new->entries[j].threshold <= usage) {
3811 			/*
3812 			 * new->current_threshold will not be used
3813 			 * until rcu_assign_pointer(), so it's safe to increment
3814 			 * it here.
3815 			 */
3816 			++new->current_threshold;
3817 		}
3818 		j++;
3819 	}
3820 
3821 swap_buffers:
3822 	/* Swap primary and spare array */
3823 	thresholds->spare = thresholds->primary;
3824 	/* If all events are unregistered, free the spare array */
3825 	if (!new) {
3826 		kfree(thresholds->spare);
3827 		thresholds->spare = NULL;
3828 	}
3829 
3830 	rcu_assign_pointer(thresholds->primary, new);
3831 
3832 	/* To be sure that nobody uses thresholds */
3833 	synchronize_rcu();
3834 unlock:
3835 	mutex_unlock(&memcg->thresholds_lock);
3836 }
3837 
3838 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3839 	struct eventfd_ctx *eventfd)
3840 {
3841 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3842 }
3843 
3844 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3845 	struct eventfd_ctx *eventfd)
3846 {
3847 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3848 }
3849 
3850 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3851 	struct eventfd_ctx *eventfd, const char *args)
3852 {
3853 	struct mem_cgroup_eventfd_list *event;
3854 
3855 	event = kmalloc(sizeof(*event),	GFP_KERNEL);
3856 	if (!event)
3857 		return -ENOMEM;
3858 
3859 	spin_lock(&memcg_oom_lock);
3860 
3861 	event->eventfd = eventfd;
3862 	list_add(&event->list, &memcg->oom_notify);
3863 
3864 	/* already in OOM ? */
3865 	if (atomic_read(&memcg->under_oom))
3866 		eventfd_signal(eventfd, 1);
3867 	spin_unlock(&memcg_oom_lock);
3868 
3869 	return 0;
3870 }
3871 
3872 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3873 	struct eventfd_ctx *eventfd)
3874 {
3875 	struct mem_cgroup_eventfd_list *ev, *tmp;
3876 
3877 	spin_lock(&memcg_oom_lock);
3878 
3879 	list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3880 		if (ev->eventfd == eventfd) {
3881 			list_del(&ev->list);
3882 			kfree(ev);
3883 		}
3884 	}
3885 
3886 	spin_unlock(&memcg_oom_lock);
3887 }
3888 
3889 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3890 {
3891 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3892 
3893 	seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3894 	seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3895 	return 0;
3896 }
3897 
3898 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3899 	struct cftype *cft, u64 val)
3900 {
3901 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3902 
3903 	/* cannot set to root cgroup and only 0 and 1 are allowed */
3904 	if (!css->parent || !((val == 0) || (val == 1)))
3905 		return -EINVAL;
3906 
3907 	memcg->oom_kill_disable = val;
3908 	if (!val)
3909 		memcg_oom_recover(memcg);
3910 
3911 	return 0;
3912 }
3913 
3914 #ifdef CONFIG_MEMCG_KMEM
3915 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3916 {
3917 	int ret;
3918 
3919 	ret = memcg_propagate_kmem(memcg);
3920 	if (ret)
3921 		return ret;
3922 
3923 	return mem_cgroup_sockets_init(memcg, ss);
3924 }
3925 
3926 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3927 {
3928 	struct cgroup_subsys_state *css;
3929 	struct mem_cgroup *parent, *child;
3930 	int kmemcg_id;
3931 
3932 	if (!memcg->kmem_acct_active)
3933 		return;
3934 
3935 	/*
3936 	 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3937 	 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3938 	 * guarantees no cache will be created for this cgroup after we are
3939 	 * done (see memcg_create_kmem_cache()).
3940 	 */
3941 	memcg->kmem_acct_active = false;
3942 
3943 	memcg_deactivate_kmem_caches(memcg);
3944 
3945 	kmemcg_id = memcg->kmemcg_id;
3946 	BUG_ON(kmemcg_id < 0);
3947 
3948 	parent = parent_mem_cgroup(memcg);
3949 	if (!parent)
3950 		parent = root_mem_cgroup;
3951 
3952 	/*
3953 	 * Change kmemcg_id of this cgroup and all its descendants to the
3954 	 * parent's id, and then move all entries from this cgroup's list_lrus
3955 	 * to ones of the parent. After we have finished, all list_lrus
3956 	 * corresponding to this cgroup are guaranteed to remain empty. The
3957 	 * ordering is imposed by list_lru_node->lock taken by
3958 	 * memcg_drain_all_list_lrus().
3959 	 */
3960 	css_for_each_descendant_pre(css, &memcg->css) {
3961 		child = mem_cgroup_from_css(css);
3962 		BUG_ON(child->kmemcg_id != kmemcg_id);
3963 		child->kmemcg_id = parent->kmemcg_id;
3964 		if (!memcg->use_hierarchy)
3965 			break;
3966 	}
3967 	memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3968 
3969 	memcg_free_cache_id(kmemcg_id);
3970 }
3971 
3972 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3973 {
3974 	if (memcg->kmem_acct_activated) {
3975 		memcg_destroy_kmem_caches(memcg);
3976 		static_key_slow_dec(&memcg_kmem_enabled_key);
3977 		WARN_ON(page_counter_read(&memcg->kmem));
3978 	}
3979 	mem_cgroup_sockets_destroy(memcg);
3980 }
3981 #else
3982 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3983 {
3984 	return 0;
3985 }
3986 
3987 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3988 {
3989 }
3990 
3991 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3992 {
3993 }
3994 #endif
3995 
3996 /*
3997  * DO NOT USE IN NEW FILES.
3998  *
3999  * "cgroup.event_control" implementation.
4000  *
4001  * This is way over-engineered.  It tries to support fully configurable
4002  * events for each user.  Such level of flexibility is completely
4003  * unnecessary especially in the light of the planned unified hierarchy.
4004  *
4005  * Please deprecate this and replace with something simpler if at all
4006  * possible.
4007  */
4008 
4009 /*
4010  * Unregister event and free resources.
4011  *
4012  * Gets called from workqueue.
4013  */
4014 static void memcg_event_remove(struct work_struct *work)
4015 {
4016 	struct mem_cgroup_event *event =
4017 		container_of(work, struct mem_cgroup_event, remove);
4018 	struct mem_cgroup *memcg = event->memcg;
4019 
4020 	remove_wait_queue(event->wqh, &event->wait);
4021 
4022 	event->unregister_event(memcg, event->eventfd);
4023 
4024 	/* Notify userspace the event is going away. */
4025 	eventfd_signal(event->eventfd, 1);
4026 
4027 	eventfd_ctx_put(event->eventfd);
4028 	kfree(event);
4029 	css_put(&memcg->css);
4030 }
4031 
4032 /*
4033  * Gets called on POLLHUP on eventfd when user closes it.
4034  *
4035  * Called with wqh->lock held and interrupts disabled.
4036  */
4037 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4038 			    int sync, void *key)
4039 {
4040 	struct mem_cgroup_event *event =
4041 		container_of(wait, struct mem_cgroup_event, wait);
4042 	struct mem_cgroup *memcg = event->memcg;
4043 	unsigned long flags = (unsigned long)key;
4044 
4045 	if (flags & POLLHUP) {
4046 		/*
4047 		 * If the event has been detached at cgroup removal, we
4048 		 * can simply return knowing the other side will cleanup
4049 		 * for us.
4050 		 *
4051 		 * We can't race against event freeing since the other
4052 		 * side will require wqh->lock via remove_wait_queue(),
4053 		 * which we hold.
4054 		 */
4055 		spin_lock(&memcg->event_list_lock);
4056 		if (!list_empty(&event->list)) {
4057 			list_del_init(&event->list);
4058 			/*
4059 			 * We are in atomic context, but cgroup_event_remove()
4060 			 * may sleep, so we have to call it in workqueue.
4061 			 */
4062 			schedule_work(&event->remove);
4063 		}
4064 		spin_unlock(&memcg->event_list_lock);
4065 	}
4066 
4067 	return 0;
4068 }
4069 
4070 static void memcg_event_ptable_queue_proc(struct file *file,
4071 		wait_queue_head_t *wqh, poll_table *pt)
4072 {
4073 	struct mem_cgroup_event *event =
4074 		container_of(pt, struct mem_cgroup_event, pt);
4075 
4076 	event->wqh = wqh;
4077 	add_wait_queue(wqh, &event->wait);
4078 }
4079 
4080 /*
4081  * DO NOT USE IN NEW FILES.
4082  *
4083  * Parse input and register new cgroup event handler.
4084  *
4085  * Input must be in format '<event_fd> <control_fd> <args>'.
4086  * Interpretation of args is defined by control file implementation.
4087  */
4088 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4089 					 char *buf, size_t nbytes, loff_t off)
4090 {
4091 	struct cgroup_subsys_state *css = of_css(of);
4092 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4093 	struct mem_cgroup_event *event;
4094 	struct cgroup_subsys_state *cfile_css;
4095 	unsigned int efd, cfd;
4096 	struct fd efile;
4097 	struct fd cfile;
4098 	const char *name;
4099 	char *endp;
4100 	int ret;
4101 
4102 	buf = strstrip(buf);
4103 
4104 	efd = simple_strtoul(buf, &endp, 10);
4105 	if (*endp != ' ')
4106 		return -EINVAL;
4107 	buf = endp + 1;
4108 
4109 	cfd = simple_strtoul(buf, &endp, 10);
4110 	if ((*endp != ' ') && (*endp != '\0'))
4111 		return -EINVAL;
4112 	buf = endp + 1;
4113 
4114 	event = kzalloc(sizeof(*event), GFP_KERNEL);
4115 	if (!event)
4116 		return -ENOMEM;
4117 
4118 	event->memcg = memcg;
4119 	INIT_LIST_HEAD(&event->list);
4120 	init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4121 	init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4122 	INIT_WORK(&event->remove, memcg_event_remove);
4123 
4124 	efile = fdget(efd);
4125 	if (!efile.file) {
4126 		ret = -EBADF;
4127 		goto out_kfree;
4128 	}
4129 
4130 	event->eventfd = eventfd_ctx_fileget(efile.file);
4131 	if (IS_ERR(event->eventfd)) {
4132 		ret = PTR_ERR(event->eventfd);
4133 		goto out_put_efile;
4134 	}
4135 
4136 	cfile = fdget(cfd);
4137 	if (!cfile.file) {
4138 		ret = -EBADF;
4139 		goto out_put_eventfd;
4140 	}
4141 
4142 	/* the process need read permission on control file */
4143 	/* AV: shouldn't we check that it's been opened for read instead? */
4144 	ret = inode_permission(file_inode(cfile.file), MAY_READ);
4145 	if (ret < 0)
4146 		goto out_put_cfile;
4147 
4148 	/*
4149 	 * Determine the event callbacks and set them in @event.  This used
4150 	 * to be done via struct cftype but cgroup core no longer knows
4151 	 * about these events.  The following is crude but the whole thing
4152 	 * is for compatibility anyway.
4153 	 *
4154 	 * DO NOT ADD NEW FILES.
4155 	 */
4156 	name = cfile.file->f_path.dentry->d_name.name;
4157 
4158 	if (!strcmp(name, "memory.usage_in_bytes")) {
4159 		event->register_event = mem_cgroup_usage_register_event;
4160 		event->unregister_event = mem_cgroup_usage_unregister_event;
4161 	} else if (!strcmp(name, "memory.oom_control")) {
4162 		event->register_event = mem_cgroup_oom_register_event;
4163 		event->unregister_event = mem_cgroup_oom_unregister_event;
4164 	} else if (!strcmp(name, "memory.pressure_level")) {
4165 		event->register_event = vmpressure_register_event;
4166 		event->unregister_event = vmpressure_unregister_event;
4167 	} else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4168 		event->register_event = memsw_cgroup_usage_register_event;
4169 		event->unregister_event = memsw_cgroup_usage_unregister_event;
4170 	} else {
4171 		ret = -EINVAL;
4172 		goto out_put_cfile;
4173 	}
4174 
4175 	/*
4176 	 * Verify @cfile should belong to @css.  Also, remaining events are
4177 	 * automatically removed on cgroup destruction but the removal is
4178 	 * asynchronous, so take an extra ref on @css.
4179 	 */
4180 	cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4181 					       &memory_cgrp_subsys);
4182 	ret = -EINVAL;
4183 	if (IS_ERR(cfile_css))
4184 		goto out_put_cfile;
4185 	if (cfile_css != css) {
4186 		css_put(cfile_css);
4187 		goto out_put_cfile;
4188 	}
4189 
4190 	ret = event->register_event(memcg, event->eventfd, buf);
4191 	if (ret)
4192 		goto out_put_css;
4193 
4194 	efile.file->f_op->poll(efile.file, &event->pt);
4195 
4196 	spin_lock(&memcg->event_list_lock);
4197 	list_add(&event->list, &memcg->event_list);
4198 	spin_unlock(&memcg->event_list_lock);
4199 
4200 	fdput(cfile);
4201 	fdput(efile);
4202 
4203 	return nbytes;
4204 
4205 out_put_css:
4206 	css_put(css);
4207 out_put_cfile:
4208 	fdput(cfile);
4209 out_put_eventfd:
4210 	eventfd_ctx_put(event->eventfd);
4211 out_put_efile:
4212 	fdput(efile);
4213 out_kfree:
4214 	kfree(event);
4215 
4216 	return ret;
4217 }
4218 
4219 static struct cftype mem_cgroup_legacy_files[] = {
4220 	{
4221 		.name = "usage_in_bytes",
4222 		.private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4223 		.read_u64 = mem_cgroup_read_u64,
4224 	},
4225 	{
4226 		.name = "max_usage_in_bytes",
4227 		.private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4228 		.write = mem_cgroup_reset,
4229 		.read_u64 = mem_cgroup_read_u64,
4230 	},
4231 	{
4232 		.name = "limit_in_bytes",
4233 		.private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4234 		.write = mem_cgroup_write,
4235 		.read_u64 = mem_cgroup_read_u64,
4236 	},
4237 	{
4238 		.name = "soft_limit_in_bytes",
4239 		.private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4240 		.write = mem_cgroup_write,
4241 		.read_u64 = mem_cgroup_read_u64,
4242 	},
4243 	{
4244 		.name = "failcnt",
4245 		.private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4246 		.write = mem_cgroup_reset,
4247 		.read_u64 = mem_cgroup_read_u64,
4248 	},
4249 	{
4250 		.name = "stat",
4251 		.seq_show = memcg_stat_show,
4252 	},
4253 	{
4254 		.name = "force_empty",
4255 		.write = mem_cgroup_force_empty_write,
4256 	},
4257 	{
4258 		.name = "use_hierarchy",
4259 		.write_u64 = mem_cgroup_hierarchy_write,
4260 		.read_u64 = mem_cgroup_hierarchy_read,
4261 	},
4262 	{
4263 		.name = "cgroup.event_control",		/* XXX: for compat */
4264 		.write = memcg_write_event_control,
4265 		.flags = CFTYPE_NO_PREFIX,
4266 		.mode = S_IWUGO,
4267 	},
4268 	{
4269 		.name = "swappiness",
4270 		.read_u64 = mem_cgroup_swappiness_read,
4271 		.write_u64 = mem_cgroup_swappiness_write,
4272 	},
4273 	{
4274 		.name = "move_charge_at_immigrate",
4275 		.read_u64 = mem_cgroup_move_charge_read,
4276 		.write_u64 = mem_cgroup_move_charge_write,
4277 	},
4278 	{
4279 		.name = "oom_control",
4280 		.seq_show = mem_cgroup_oom_control_read,
4281 		.write_u64 = mem_cgroup_oom_control_write,
4282 		.private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4283 	},
4284 	{
4285 		.name = "pressure_level",
4286 	},
4287 #ifdef CONFIG_NUMA
4288 	{
4289 		.name = "numa_stat",
4290 		.seq_show = memcg_numa_stat_show,
4291 	},
4292 #endif
4293 #ifdef CONFIG_MEMCG_KMEM
4294 	{
4295 		.name = "kmem.limit_in_bytes",
4296 		.private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4297 		.write = mem_cgroup_write,
4298 		.read_u64 = mem_cgroup_read_u64,
4299 	},
4300 	{
4301 		.name = "kmem.usage_in_bytes",
4302 		.private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4303 		.read_u64 = mem_cgroup_read_u64,
4304 	},
4305 	{
4306 		.name = "kmem.failcnt",
4307 		.private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4308 		.write = mem_cgroup_reset,
4309 		.read_u64 = mem_cgroup_read_u64,
4310 	},
4311 	{
4312 		.name = "kmem.max_usage_in_bytes",
4313 		.private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4314 		.write = mem_cgroup_reset,
4315 		.read_u64 = mem_cgroup_read_u64,
4316 	},
4317 #ifdef CONFIG_SLABINFO
4318 	{
4319 		.name = "kmem.slabinfo",
4320 		.seq_start = slab_start,
4321 		.seq_next = slab_next,
4322 		.seq_stop = slab_stop,
4323 		.seq_show = memcg_slab_show,
4324 	},
4325 #endif
4326 #endif
4327 	{ },	/* terminate */
4328 };
4329 
4330 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4331 {
4332 	struct mem_cgroup_per_node *pn;
4333 	struct mem_cgroup_per_zone *mz;
4334 	int zone, tmp = node;
4335 	/*
4336 	 * This routine is called against possible nodes.
4337 	 * But it's BUG to call kmalloc() against offline node.
4338 	 *
4339 	 * TODO: this routine can waste much memory for nodes which will
4340 	 *       never be onlined. It's better to use memory hotplug callback
4341 	 *       function.
4342 	 */
4343 	if (!node_state(node, N_NORMAL_MEMORY))
4344 		tmp = -1;
4345 	pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4346 	if (!pn)
4347 		return 1;
4348 
4349 	for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4350 		mz = &pn->zoneinfo[zone];
4351 		lruvec_init(&mz->lruvec);
4352 		mz->usage_in_excess = 0;
4353 		mz->on_tree = false;
4354 		mz->memcg = memcg;
4355 	}
4356 	memcg->nodeinfo[node] = pn;
4357 	return 0;
4358 }
4359 
4360 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4361 {
4362 	kfree(memcg->nodeinfo[node]);
4363 }
4364 
4365 static struct mem_cgroup *mem_cgroup_alloc(void)
4366 {
4367 	struct mem_cgroup *memcg;
4368 	size_t size;
4369 
4370 	size = sizeof(struct mem_cgroup);
4371 	size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4372 
4373 	memcg = kzalloc(size, GFP_KERNEL);
4374 	if (!memcg)
4375 		return NULL;
4376 
4377 	memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4378 	if (!memcg->stat)
4379 		goto out_free;
4380 	spin_lock_init(&memcg->pcp_counter_lock);
4381 	return memcg;
4382 
4383 out_free:
4384 	kfree(memcg);
4385 	return NULL;
4386 }
4387 
4388 /*
4389  * At destroying mem_cgroup, references from swap_cgroup can remain.
4390  * (scanning all at force_empty is too costly...)
4391  *
4392  * Instead of clearing all references at force_empty, we remember
4393  * the number of reference from swap_cgroup and free mem_cgroup when
4394  * it goes down to 0.
4395  *
4396  * Removal of cgroup itself succeeds regardless of refs from swap.
4397  */
4398 
4399 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4400 {
4401 	int node;
4402 
4403 	mem_cgroup_remove_from_trees(memcg);
4404 
4405 	for_each_node(node)
4406 		free_mem_cgroup_per_zone_info(memcg, node);
4407 
4408 	free_percpu(memcg->stat);
4409 	kfree(memcg);
4410 }
4411 
4412 /*
4413  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4414  */
4415 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4416 {
4417 	if (!memcg->memory.parent)
4418 		return NULL;
4419 	return mem_cgroup_from_counter(memcg->memory.parent, memory);
4420 }
4421 EXPORT_SYMBOL(parent_mem_cgroup);
4422 
4423 static struct cgroup_subsys_state * __ref
4424 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4425 {
4426 	struct mem_cgroup *memcg;
4427 	long error = -ENOMEM;
4428 	int node;
4429 
4430 	memcg = mem_cgroup_alloc();
4431 	if (!memcg)
4432 		return ERR_PTR(error);
4433 
4434 	for_each_node(node)
4435 		if (alloc_mem_cgroup_per_zone_info(memcg, node))
4436 			goto free_out;
4437 
4438 	/* root ? */
4439 	if (parent_css == NULL) {
4440 		root_mem_cgroup = memcg;
4441 		page_counter_init(&memcg->memory, NULL);
4442 		memcg->high = PAGE_COUNTER_MAX;
4443 		memcg->soft_limit = PAGE_COUNTER_MAX;
4444 		page_counter_init(&memcg->memsw, NULL);
4445 		page_counter_init(&memcg->kmem, NULL);
4446 	}
4447 
4448 	memcg->last_scanned_node = MAX_NUMNODES;
4449 	INIT_LIST_HEAD(&memcg->oom_notify);
4450 	memcg->move_charge_at_immigrate = 0;
4451 	mutex_init(&memcg->thresholds_lock);
4452 	spin_lock_init(&memcg->move_lock);
4453 	vmpressure_init(&memcg->vmpressure);
4454 	INIT_LIST_HEAD(&memcg->event_list);
4455 	spin_lock_init(&memcg->event_list_lock);
4456 #ifdef CONFIG_MEMCG_KMEM
4457 	memcg->kmemcg_id = -1;
4458 #endif
4459 
4460 	return &memcg->css;
4461 
4462 free_out:
4463 	__mem_cgroup_free(memcg);
4464 	return ERR_PTR(error);
4465 }
4466 
4467 static int
4468 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4469 {
4470 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4471 	struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4472 	int ret;
4473 
4474 	if (css->id > MEM_CGROUP_ID_MAX)
4475 		return -ENOSPC;
4476 
4477 	if (!parent)
4478 		return 0;
4479 
4480 	mutex_lock(&memcg_create_mutex);
4481 
4482 	memcg->use_hierarchy = parent->use_hierarchy;
4483 	memcg->oom_kill_disable = parent->oom_kill_disable;
4484 	memcg->swappiness = mem_cgroup_swappiness(parent);
4485 
4486 	if (parent->use_hierarchy) {
4487 		page_counter_init(&memcg->memory, &parent->memory);
4488 		memcg->high = PAGE_COUNTER_MAX;
4489 		memcg->soft_limit = PAGE_COUNTER_MAX;
4490 		page_counter_init(&memcg->memsw, &parent->memsw);
4491 		page_counter_init(&memcg->kmem, &parent->kmem);
4492 
4493 		/*
4494 		 * No need to take a reference to the parent because cgroup
4495 		 * core guarantees its existence.
4496 		 */
4497 	} else {
4498 		page_counter_init(&memcg->memory, NULL);
4499 		memcg->high = PAGE_COUNTER_MAX;
4500 		memcg->soft_limit = PAGE_COUNTER_MAX;
4501 		page_counter_init(&memcg->memsw, NULL);
4502 		page_counter_init(&memcg->kmem, NULL);
4503 		/*
4504 		 * Deeper hierachy with use_hierarchy == false doesn't make
4505 		 * much sense so let cgroup subsystem know about this
4506 		 * unfortunate state in our controller.
4507 		 */
4508 		if (parent != root_mem_cgroup)
4509 			memory_cgrp_subsys.broken_hierarchy = true;
4510 	}
4511 	mutex_unlock(&memcg_create_mutex);
4512 
4513 	ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4514 	if (ret)
4515 		return ret;
4516 
4517 	/*
4518 	 * Make sure the memcg is initialized: mem_cgroup_iter()
4519 	 * orders reading memcg->initialized against its callers
4520 	 * reading the memcg members.
4521 	 */
4522 	smp_store_release(&memcg->initialized, 1);
4523 
4524 	return 0;
4525 }
4526 
4527 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4528 {
4529 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4530 	struct mem_cgroup_event *event, *tmp;
4531 
4532 	/*
4533 	 * Unregister events and notify userspace.
4534 	 * Notify userspace about cgroup removing only after rmdir of cgroup
4535 	 * directory to avoid race between userspace and kernelspace.
4536 	 */
4537 	spin_lock(&memcg->event_list_lock);
4538 	list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4539 		list_del_init(&event->list);
4540 		schedule_work(&event->remove);
4541 	}
4542 	spin_unlock(&memcg->event_list_lock);
4543 
4544 	vmpressure_cleanup(&memcg->vmpressure);
4545 
4546 	memcg_deactivate_kmem(memcg);
4547 }
4548 
4549 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4550 {
4551 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4552 
4553 	memcg_destroy_kmem(memcg);
4554 	__mem_cgroup_free(memcg);
4555 }
4556 
4557 /**
4558  * mem_cgroup_css_reset - reset the states of a mem_cgroup
4559  * @css: the target css
4560  *
4561  * Reset the states of the mem_cgroup associated with @css.  This is
4562  * invoked when the userland requests disabling on the default hierarchy
4563  * but the memcg is pinned through dependency.  The memcg should stop
4564  * applying policies and should revert to the vanilla state as it may be
4565  * made visible again.
4566  *
4567  * The current implementation only resets the essential configurations.
4568  * This needs to be expanded to cover all the visible parts.
4569  */
4570 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4571 {
4572 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4573 
4574 	mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4575 	mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4576 	memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4577 	memcg->low = 0;
4578 	memcg->high = PAGE_COUNTER_MAX;
4579 	memcg->soft_limit = PAGE_COUNTER_MAX;
4580 }
4581 
4582 #ifdef CONFIG_MMU
4583 /* Handlers for move charge at task migration. */
4584 static int mem_cgroup_do_precharge(unsigned long count)
4585 {
4586 	int ret;
4587 
4588 	/* Try a single bulk charge without reclaim first */
4589 	ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4590 	if (!ret) {
4591 		mc.precharge += count;
4592 		return ret;
4593 	}
4594 	if (ret == -EINTR) {
4595 		cancel_charge(root_mem_cgroup, count);
4596 		return ret;
4597 	}
4598 
4599 	/* Try charges one by one with reclaim */
4600 	while (count--) {
4601 		ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4602 		/*
4603 		 * In case of failure, any residual charges against
4604 		 * mc.to will be dropped by mem_cgroup_clear_mc()
4605 		 * later on.  However, cancel any charges that are
4606 		 * bypassed to root right away or they'll be lost.
4607 		 */
4608 		if (ret == -EINTR)
4609 			cancel_charge(root_mem_cgroup, 1);
4610 		if (ret)
4611 			return ret;
4612 		mc.precharge++;
4613 		cond_resched();
4614 	}
4615 	return 0;
4616 }
4617 
4618 /**
4619  * get_mctgt_type - get target type of moving charge
4620  * @vma: the vma the pte to be checked belongs
4621  * @addr: the address corresponding to the pte to be checked
4622  * @ptent: the pte to be checked
4623  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4624  *
4625  * Returns
4626  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4627  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4628  *     move charge. if @target is not NULL, the page is stored in target->page
4629  *     with extra refcnt got(Callers should handle it).
4630  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4631  *     target for charge migration. if @target is not NULL, the entry is stored
4632  *     in target->ent.
4633  *
4634  * Called with pte lock held.
4635  */
4636 union mc_target {
4637 	struct page	*page;
4638 	swp_entry_t	ent;
4639 };
4640 
4641 enum mc_target_type {
4642 	MC_TARGET_NONE = 0,
4643 	MC_TARGET_PAGE,
4644 	MC_TARGET_SWAP,
4645 };
4646 
4647 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4648 						unsigned long addr, pte_t ptent)
4649 {
4650 	struct page *page = vm_normal_page(vma, addr, ptent);
4651 
4652 	if (!page || !page_mapped(page))
4653 		return NULL;
4654 	if (PageAnon(page)) {
4655 		if (!(mc.flags & MOVE_ANON))
4656 			return NULL;
4657 	} else {
4658 		if (!(mc.flags & MOVE_FILE))
4659 			return NULL;
4660 	}
4661 	if (!get_page_unless_zero(page))
4662 		return NULL;
4663 
4664 	return page;
4665 }
4666 
4667 #ifdef CONFIG_SWAP
4668 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4669 			unsigned long addr, pte_t ptent, swp_entry_t *entry)
4670 {
4671 	struct page *page = NULL;
4672 	swp_entry_t ent = pte_to_swp_entry(ptent);
4673 
4674 	if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4675 		return NULL;
4676 	/*
4677 	 * Because lookup_swap_cache() updates some statistics counter,
4678 	 * we call find_get_page() with swapper_space directly.
4679 	 */
4680 	page = find_get_page(swap_address_space(ent), ent.val);
4681 	if (do_swap_account)
4682 		entry->val = ent.val;
4683 
4684 	return page;
4685 }
4686 #else
4687 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4688 			unsigned long addr, pte_t ptent, swp_entry_t *entry)
4689 {
4690 	return NULL;
4691 }
4692 #endif
4693 
4694 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4695 			unsigned long addr, pte_t ptent, swp_entry_t *entry)
4696 {
4697 	struct page *page = NULL;
4698 	struct address_space *mapping;
4699 	pgoff_t pgoff;
4700 
4701 	if (!vma->vm_file) /* anonymous vma */
4702 		return NULL;
4703 	if (!(mc.flags & MOVE_FILE))
4704 		return NULL;
4705 
4706 	mapping = vma->vm_file->f_mapping;
4707 	pgoff = linear_page_index(vma, addr);
4708 
4709 	/* page is moved even if it's not RSS of this task(page-faulted). */
4710 #ifdef CONFIG_SWAP
4711 	/* shmem/tmpfs may report page out on swap: account for that too. */
4712 	if (shmem_mapping(mapping)) {
4713 		page = find_get_entry(mapping, pgoff);
4714 		if (radix_tree_exceptional_entry(page)) {
4715 			swp_entry_t swp = radix_to_swp_entry(page);
4716 			if (do_swap_account)
4717 				*entry = swp;
4718 			page = find_get_page(swap_address_space(swp), swp.val);
4719 		}
4720 	} else
4721 		page = find_get_page(mapping, pgoff);
4722 #else
4723 	page = find_get_page(mapping, pgoff);
4724 #endif
4725 	return page;
4726 }
4727 
4728 /**
4729  * mem_cgroup_move_account - move account of the page
4730  * @page: the page
4731  * @nr_pages: number of regular pages (>1 for huge pages)
4732  * @from: mem_cgroup which the page is moved from.
4733  * @to:	mem_cgroup which the page is moved to. @from != @to.
4734  *
4735  * The caller must confirm following.
4736  * - page is not on LRU (isolate_page() is useful.)
4737  * - compound_lock is held when nr_pages > 1
4738  *
4739  * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4740  * from old cgroup.
4741  */
4742 static int mem_cgroup_move_account(struct page *page,
4743 				   unsigned int nr_pages,
4744 				   struct mem_cgroup *from,
4745 				   struct mem_cgroup *to)
4746 {
4747 	unsigned long flags;
4748 	int ret;
4749 
4750 	VM_BUG_ON(from == to);
4751 	VM_BUG_ON_PAGE(PageLRU(page), page);
4752 	/*
4753 	 * The page is isolated from LRU. So, collapse function
4754 	 * will not handle this page. But page splitting can happen.
4755 	 * Do this check under compound_page_lock(). The caller should
4756 	 * hold it.
4757 	 */
4758 	ret = -EBUSY;
4759 	if (nr_pages > 1 && !PageTransHuge(page))
4760 		goto out;
4761 
4762 	/*
4763 	 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4764 	 * of its source page while we change it: page migration takes
4765 	 * both pages off the LRU, but page cache replacement doesn't.
4766 	 */
4767 	if (!trylock_page(page))
4768 		goto out;
4769 
4770 	ret = -EINVAL;
4771 	if (page->mem_cgroup != from)
4772 		goto out_unlock;
4773 
4774 	spin_lock_irqsave(&from->move_lock, flags);
4775 
4776 	if (!PageAnon(page) && page_mapped(page)) {
4777 		__this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4778 			       nr_pages);
4779 		__this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4780 			       nr_pages);
4781 	}
4782 
4783 	if (PageWriteback(page)) {
4784 		__this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4785 			       nr_pages);
4786 		__this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4787 			       nr_pages);
4788 	}
4789 
4790 	/*
4791 	 * It is safe to change page->mem_cgroup here because the page
4792 	 * is referenced, charged, and isolated - we can't race with
4793 	 * uncharging, charging, migration, or LRU putback.
4794 	 */
4795 
4796 	/* caller should have done css_get */
4797 	page->mem_cgroup = to;
4798 	spin_unlock_irqrestore(&from->move_lock, flags);
4799 
4800 	ret = 0;
4801 
4802 	local_irq_disable();
4803 	mem_cgroup_charge_statistics(to, page, nr_pages);
4804 	memcg_check_events(to, page);
4805 	mem_cgroup_charge_statistics(from, page, -nr_pages);
4806 	memcg_check_events(from, page);
4807 	local_irq_enable();
4808 out_unlock:
4809 	unlock_page(page);
4810 out:
4811 	return ret;
4812 }
4813 
4814 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4815 		unsigned long addr, pte_t ptent, union mc_target *target)
4816 {
4817 	struct page *page = NULL;
4818 	enum mc_target_type ret = MC_TARGET_NONE;
4819 	swp_entry_t ent = { .val = 0 };
4820 
4821 	if (pte_present(ptent))
4822 		page = mc_handle_present_pte(vma, addr, ptent);
4823 	else if (is_swap_pte(ptent))
4824 		page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4825 	else if (pte_none(ptent))
4826 		page = mc_handle_file_pte(vma, addr, ptent, &ent);
4827 
4828 	if (!page && !ent.val)
4829 		return ret;
4830 	if (page) {
4831 		/*
4832 		 * Do only loose check w/o serialization.
4833 		 * mem_cgroup_move_account() checks the page is valid or
4834 		 * not under LRU exclusion.
4835 		 */
4836 		if (page->mem_cgroup == mc.from) {
4837 			ret = MC_TARGET_PAGE;
4838 			if (target)
4839 				target->page = page;
4840 		}
4841 		if (!ret || !target)
4842 			put_page(page);
4843 	}
4844 	/* There is a swap entry and a page doesn't exist or isn't charged */
4845 	if (ent.val && !ret &&
4846 	    mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4847 		ret = MC_TARGET_SWAP;
4848 		if (target)
4849 			target->ent = ent;
4850 	}
4851 	return ret;
4852 }
4853 
4854 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4855 /*
4856  * We don't consider swapping or file mapped pages because THP does not
4857  * support them for now.
4858  * Caller should make sure that pmd_trans_huge(pmd) is true.
4859  */
4860 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4861 		unsigned long addr, pmd_t pmd, union mc_target *target)
4862 {
4863 	struct page *page = NULL;
4864 	enum mc_target_type ret = MC_TARGET_NONE;
4865 
4866 	page = pmd_page(pmd);
4867 	VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4868 	if (!(mc.flags & MOVE_ANON))
4869 		return ret;
4870 	if (page->mem_cgroup == mc.from) {
4871 		ret = MC_TARGET_PAGE;
4872 		if (target) {
4873 			get_page(page);
4874 			target->page = page;
4875 		}
4876 	}
4877 	return ret;
4878 }
4879 #else
4880 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4881 		unsigned long addr, pmd_t pmd, union mc_target *target)
4882 {
4883 	return MC_TARGET_NONE;
4884 }
4885 #endif
4886 
4887 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4888 					unsigned long addr, unsigned long end,
4889 					struct mm_walk *walk)
4890 {
4891 	struct vm_area_struct *vma = walk->vma;
4892 	pte_t *pte;
4893 	spinlock_t *ptl;
4894 
4895 	if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4896 		if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4897 			mc.precharge += HPAGE_PMD_NR;
4898 		spin_unlock(ptl);
4899 		return 0;
4900 	}
4901 
4902 	if (pmd_trans_unstable(pmd))
4903 		return 0;
4904 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4905 	for (; addr != end; pte++, addr += PAGE_SIZE)
4906 		if (get_mctgt_type(vma, addr, *pte, NULL))
4907 			mc.precharge++;	/* increment precharge temporarily */
4908 	pte_unmap_unlock(pte - 1, ptl);
4909 	cond_resched();
4910 
4911 	return 0;
4912 }
4913 
4914 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4915 {
4916 	unsigned long precharge;
4917 
4918 	struct mm_walk mem_cgroup_count_precharge_walk = {
4919 		.pmd_entry = mem_cgroup_count_precharge_pte_range,
4920 		.mm = mm,
4921 	};
4922 	down_read(&mm->mmap_sem);
4923 	walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4924 	up_read(&mm->mmap_sem);
4925 
4926 	precharge = mc.precharge;
4927 	mc.precharge = 0;
4928 
4929 	return precharge;
4930 }
4931 
4932 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4933 {
4934 	unsigned long precharge = mem_cgroup_count_precharge(mm);
4935 
4936 	VM_BUG_ON(mc.moving_task);
4937 	mc.moving_task = current;
4938 	return mem_cgroup_do_precharge(precharge);
4939 }
4940 
4941 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4942 static void __mem_cgroup_clear_mc(void)
4943 {
4944 	struct mem_cgroup *from = mc.from;
4945 	struct mem_cgroup *to = mc.to;
4946 
4947 	/* we must uncharge all the leftover precharges from mc.to */
4948 	if (mc.precharge) {
4949 		cancel_charge(mc.to, mc.precharge);
4950 		mc.precharge = 0;
4951 	}
4952 	/*
4953 	 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4954 	 * we must uncharge here.
4955 	 */
4956 	if (mc.moved_charge) {
4957 		cancel_charge(mc.from, mc.moved_charge);
4958 		mc.moved_charge = 0;
4959 	}
4960 	/* we must fixup refcnts and charges */
4961 	if (mc.moved_swap) {
4962 		/* uncharge swap account from the old cgroup */
4963 		if (!mem_cgroup_is_root(mc.from))
4964 			page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4965 
4966 		/*
4967 		 * we charged both to->memory and to->memsw, so we
4968 		 * should uncharge to->memory.
4969 		 */
4970 		if (!mem_cgroup_is_root(mc.to))
4971 			page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4972 
4973 		css_put_many(&mc.from->css, mc.moved_swap);
4974 
4975 		/* we've already done css_get(mc.to) */
4976 		mc.moved_swap = 0;
4977 	}
4978 	memcg_oom_recover(from);
4979 	memcg_oom_recover(to);
4980 	wake_up_all(&mc.waitq);
4981 }
4982 
4983 static void mem_cgroup_clear_mc(void)
4984 {
4985 	/*
4986 	 * we must clear moving_task before waking up waiters at the end of
4987 	 * task migration.
4988 	 */
4989 	mc.moving_task = NULL;
4990 	__mem_cgroup_clear_mc();
4991 	spin_lock(&mc.lock);
4992 	mc.from = NULL;
4993 	mc.to = NULL;
4994 	spin_unlock(&mc.lock);
4995 }
4996 
4997 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4998 				 struct cgroup_taskset *tset)
4999 {
5000 	struct task_struct *p = cgroup_taskset_first(tset);
5001 	int ret = 0;
5002 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5003 	unsigned long move_flags;
5004 
5005 	/*
5006 	 * We are now commited to this value whatever it is. Changes in this
5007 	 * tunable will only affect upcoming migrations, not the current one.
5008 	 * So we need to save it, and keep it going.
5009 	 */
5010 	move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5011 	if (move_flags) {
5012 		struct mm_struct *mm;
5013 		struct mem_cgroup *from = mem_cgroup_from_task(p);
5014 
5015 		VM_BUG_ON(from == memcg);
5016 
5017 		mm = get_task_mm(p);
5018 		if (!mm)
5019 			return 0;
5020 		/* We move charges only when we move a owner of the mm */
5021 		if (mm->owner == p) {
5022 			VM_BUG_ON(mc.from);
5023 			VM_BUG_ON(mc.to);
5024 			VM_BUG_ON(mc.precharge);
5025 			VM_BUG_ON(mc.moved_charge);
5026 			VM_BUG_ON(mc.moved_swap);
5027 
5028 			spin_lock(&mc.lock);
5029 			mc.from = from;
5030 			mc.to = memcg;
5031 			mc.flags = move_flags;
5032 			spin_unlock(&mc.lock);
5033 			/* We set mc.moving_task later */
5034 
5035 			ret = mem_cgroup_precharge_mc(mm);
5036 			if (ret)
5037 				mem_cgroup_clear_mc();
5038 		}
5039 		mmput(mm);
5040 	}
5041 	return ret;
5042 }
5043 
5044 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5045 				     struct cgroup_taskset *tset)
5046 {
5047 	if (mc.to)
5048 		mem_cgroup_clear_mc();
5049 }
5050 
5051 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5052 				unsigned long addr, unsigned long end,
5053 				struct mm_walk *walk)
5054 {
5055 	int ret = 0;
5056 	struct vm_area_struct *vma = walk->vma;
5057 	pte_t *pte;
5058 	spinlock_t *ptl;
5059 	enum mc_target_type target_type;
5060 	union mc_target target;
5061 	struct page *page;
5062 
5063 	/*
5064 	 * We don't take compound_lock() here but no race with splitting thp
5065 	 * happens because:
5066 	 *  - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5067 	 *    under splitting, which means there's no concurrent thp split,
5068 	 *  - if another thread runs into split_huge_page() just after we
5069 	 *    entered this if-block, the thread must wait for page table lock
5070 	 *    to be unlocked in __split_huge_page_splitting(), where the main
5071 	 *    part of thp split is not executed yet.
5072 	 */
5073 	if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5074 		if (mc.precharge < HPAGE_PMD_NR) {
5075 			spin_unlock(ptl);
5076 			return 0;
5077 		}
5078 		target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5079 		if (target_type == MC_TARGET_PAGE) {
5080 			page = target.page;
5081 			if (!isolate_lru_page(page)) {
5082 				if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5083 							     mc.from, mc.to)) {
5084 					mc.precharge -= HPAGE_PMD_NR;
5085 					mc.moved_charge += HPAGE_PMD_NR;
5086 				}
5087 				putback_lru_page(page);
5088 			}
5089 			put_page(page);
5090 		}
5091 		spin_unlock(ptl);
5092 		return 0;
5093 	}
5094 
5095 	if (pmd_trans_unstable(pmd))
5096 		return 0;
5097 retry:
5098 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5099 	for (; addr != end; addr += PAGE_SIZE) {
5100 		pte_t ptent = *(pte++);
5101 		swp_entry_t ent;
5102 
5103 		if (!mc.precharge)
5104 			break;
5105 
5106 		switch (get_mctgt_type(vma, addr, ptent, &target)) {
5107 		case MC_TARGET_PAGE:
5108 			page = target.page;
5109 			if (isolate_lru_page(page))
5110 				goto put;
5111 			if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5112 				mc.precharge--;
5113 				/* we uncharge from mc.from later. */
5114 				mc.moved_charge++;
5115 			}
5116 			putback_lru_page(page);
5117 put:			/* get_mctgt_type() gets the page */
5118 			put_page(page);
5119 			break;
5120 		case MC_TARGET_SWAP:
5121 			ent = target.ent;
5122 			if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5123 				mc.precharge--;
5124 				/* we fixup refcnts and charges later. */
5125 				mc.moved_swap++;
5126 			}
5127 			break;
5128 		default:
5129 			break;
5130 		}
5131 	}
5132 	pte_unmap_unlock(pte - 1, ptl);
5133 	cond_resched();
5134 
5135 	if (addr != end) {
5136 		/*
5137 		 * We have consumed all precharges we got in can_attach().
5138 		 * We try charge one by one, but don't do any additional
5139 		 * charges to mc.to if we have failed in charge once in attach()
5140 		 * phase.
5141 		 */
5142 		ret = mem_cgroup_do_precharge(1);
5143 		if (!ret)
5144 			goto retry;
5145 	}
5146 
5147 	return ret;
5148 }
5149 
5150 static void mem_cgroup_move_charge(struct mm_struct *mm)
5151 {
5152 	struct mm_walk mem_cgroup_move_charge_walk = {
5153 		.pmd_entry = mem_cgroup_move_charge_pte_range,
5154 		.mm = mm,
5155 	};
5156 
5157 	lru_add_drain_all();
5158 	/*
5159 	 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5160 	 * move_lock while we're moving its pages to another memcg.
5161 	 * Then wait for already started RCU-only updates to finish.
5162 	 */
5163 	atomic_inc(&mc.from->moving_account);
5164 	synchronize_rcu();
5165 retry:
5166 	if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5167 		/*
5168 		 * Someone who are holding the mmap_sem might be waiting in
5169 		 * waitq. So we cancel all extra charges, wake up all waiters,
5170 		 * and retry. Because we cancel precharges, we might not be able
5171 		 * to move enough charges, but moving charge is a best-effort
5172 		 * feature anyway, so it wouldn't be a big problem.
5173 		 */
5174 		__mem_cgroup_clear_mc();
5175 		cond_resched();
5176 		goto retry;
5177 	}
5178 	/*
5179 	 * When we have consumed all precharges and failed in doing
5180 	 * additional charge, the page walk just aborts.
5181 	 */
5182 	walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5183 	up_read(&mm->mmap_sem);
5184 	atomic_dec(&mc.from->moving_account);
5185 }
5186 
5187 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5188 				 struct cgroup_taskset *tset)
5189 {
5190 	struct task_struct *p = cgroup_taskset_first(tset);
5191 	struct mm_struct *mm = get_task_mm(p);
5192 
5193 	if (mm) {
5194 		if (mc.to)
5195 			mem_cgroup_move_charge(mm);
5196 		mmput(mm);
5197 	}
5198 	if (mc.to)
5199 		mem_cgroup_clear_mc();
5200 }
5201 #else	/* !CONFIG_MMU */
5202 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5203 				 struct cgroup_taskset *tset)
5204 {
5205 	return 0;
5206 }
5207 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5208 				     struct cgroup_taskset *tset)
5209 {
5210 }
5211 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5212 				 struct cgroup_taskset *tset)
5213 {
5214 }
5215 #endif
5216 
5217 /*
5218  * Cgroup retains root cgroups across [un]mount cycles making it necessary
5219  * to verify whether we're attached to the default hierarchy on each mount
5220  * attempt.
5221  */
5222 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5223 {
5224 	/*
5225 	 * use_hierarchy is forced on the default hierarchy.  cgroup core
5226 	 * guarantees that @root doesn't have any children, so turning it
5227 	 * on for the root memcg is enough.
5228 	 */
5229 	if (cgroup_on_dfl(root_css->cgroup))
5230 		root_mem_cgroup->use_hierarchy = true;
5231 	else
5232 		root_mem_cgroup->use_hierarchy = false;
5233 }
5234 
5235 static u64 memory_current_read(struct cgroup_subsys_state *css,
5236 			       struct cftype *cft)
5237 {
5238 	return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5239 }
5240 
5241 static int memory_low_show(struct seq_file *m, void *v)
5242 {
5243 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5244 	unsigned long low = READ_ONCE(memcg->low);
5245 
5246 	if (low == PAGE_COUNTER_MAX)
5247 		seq_puts(m, "max\n");
5248 	else
5249 		seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5250 
5251 	return 0;
5252 }
5253 
5254 static ssize_t memory_low_write(struct kernfs_open_file *of,
5255 				char *buf, size_t nbytes, loff_t off)
5256 {
5257 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5258 	unsigned long low;
5259 	int err;
5260 
5261 	buf = strstrip(buf);
5262 	err = page_counter_memparse(buf, "max", &low);
5263 	if (err)
5264 		return err;
5265 
5266 	memcg->low = low;
5267 
5268 	return nbytes;
5269 }
5270 
5271 static int memory_high_show(struct seq_file *m, void *v)
5272 {
5273 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5274 	unsigned long high = READ_ONCE(memcg->high);
5275 
5276 	if (high == PAGE_COUNTER_MAX)
5277 		seq_puts(m, "max\n");
5278 	else
5279 		seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5280 
5281 	return 0;
5282 }
5283 
5284 static ssize_t memory_high_write(struct kernfs_open_file *of,
5285 				 char *buf, size_t nbytes, loff_t off)
5286 {
5287 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5288 	unsigned long high;
5289 	int err;
5290 
5291 	buf = strstrip(buf);
5292 	err = page_counter_memparse(buf, "max", &high);
5293 	if (err)
5294 		return err;
5295 
5296 	memcg->high = high;
5297 
5298 	return nbytes;
5299 }
5300 
5301 static int memory_max_show(struct seq_file *m, void *v)
5302 {
5303 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5304 	unsigned long max = READ_ONCE(memcg->memory.limit);
5305 
5306 	if (max == PAGE_COUNTER_MAX)
5307 		seq_puts(m, "max\n");
5308 	else
5309 		seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5310 
5311 	return 0;
5312 }
5313 
5314 static ssize_t memory_max_write(struct kernfs_open_file *of,
5315 				char *buf, size_t nbytes, loff_t off)
5316 {
5317 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5318 	unsigned long max;
5319 	int err;
5320 
5321 	buf = strstrip(buf);
5322 	err = page_counter_memparse(buf, "max", &max);
5323 	if (err)
5324 		return err;
5325 
5326 	err = mem_cgroup_resize_limit(memcg, max);
5327 	if (err)
5328 		return err;
5329 
5330 	return nbytes;
5331 }
5332 
5333 static int memory_events_show(struct seq_file *m, void *v)
5334 {
5335 	struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5336 
5337 	seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5338 	seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5339 	seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5340 	seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5341 
5342 	return 0;
5343 }
5344 
5345 static struct cftype memory_files[] = {
5346 	{
5347 		.name = "current",
5348 		.read_u64 = memory_current_read,
5349 	},
5350 	{
5351 		.name = "low",
5352 		.flags = CFTYPE_NOT_ON_ROOT,
5353 		.seq_show = memory_low_show,
5354 		.write = memory_low_write,
5355 	},
5356 	{
5357 		.name = "high",
5358 		.flags = CFTYPE_NOT_ON_ROOT,
5359 		.seq_show = memory_high_show,
5360 		.write = memory_high_write,
5361 	},
5362 	{
5363 		.name = "max",
5364 		.flags = CFTYPE_NOT_ON_ROOT,
5365 		.seq_show = memory_max_show,
5366 		.write = memory_max_write,
5367 	},
5368 	{
5369 		.name = "events",
5370 		.flags = CFTYPE_NOT_ON_ROOT,
5371 		.seq_show = memory_events_show,
5372 	},
5373 	{ }	/* terminate */
5374 };
5375 
5376 struct cgroup_subsys memory_cgrp_subsys = {
5377 	.css_alloc = mem_cgroup_css_alloc,
5378 	.css_online = mem_cgroup_css_online,
5379 	.css_offline = mem_cgroup_css_offline,
5380 	.css_free = mem_cgroup_css_free,
5381 	.css_reset = mem_cgroup_css_reset,
5382 	.can_attach = mem_cgroup_can_attach,
5383 	.cancel_attach = mem_cgroup_cancel_attach,
5384 	.attach = mem_cgroup_move_task,
5385 	.bind = mem_cgroup_bind,
5386 	.dfl_cftypes = memory_files,
5387 	.legacy_cftypes = mem_cgroup_legacy_files,
5388 	.early_init = 0,
5389 };
5390 
5391 /**
5392  * mem_cgroup_events - count memory events against a cgroup
5393  * @memcg: the memory cgroup
5394  * @idx: the event index
5395  * @nr: the number of events to account for
5396  */
5397 void mem_cgroup_events(struct mem_cgroup *memcg,
5398 		       enum mem_cgroup_events_index idx,
5399 		       unsigned int nr)
5400 {
5401 	this_cpu_add(memcg->stat->events[idx], nr);
5402 }
5403 
5404 /**
5405  * mem_cgroup_low - check if memory consumption is below the normal range
5406  * @root: the highest ancestor to consider
5407  * @memcg: the memory cgroup to check
5408  *
5409  * Returns %true if memory consumption of @memcg, and that of all
5410  * configurable ancestors up to @root, is below the normal range.
5411  */
5412 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5413 {
5414 	if (mem_cgroup_disabled())
5415 		return false;
5416 
5417 	/*
5418 	 * The toplevel group doesn't have a configurable range, so
5419 	 * it's never low when looked at directly, and it is not
5420 	 * considered an ancestor when assessing the hierarchy.
5421 	 */
5422 
5423 	if (memcg == root_mem_cgroup)
5424 		return false;
5425 
5426 	if (page_counter_read(&memcg->memory) >= memcg->low)
5427 		return false;
5428 
5429 	while (memcg != root) {
5430 		memcg = parent_mem_cgroup(memcg);
5431 
5432 		if (memcg == root_mem_cgroup)
5433 			break;
5434 
5435 		if (page_counter_read(&memcg->memory) >= memcg->low)
5436 			return false;
5437 	}
5438 	return true;
5439 }
5440 
5441 /**
5442  * mem_cgroup_try_charge - try charging a page
5443  * @page: page to charge
5444  * @mm: mm context of the victim
5445  * @gfp_mask: reclaim mode
5446  * @memcgp: charged memcg return
5447  *
5448  * Try to charge @page to the memcg that @mm belongs to, reclaiming
5449  * pages according to @gfp_mask if necessary.
5450  *
5451  * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5452  * Otherwise, an error code is returned.
5453  *
5454  * After page->mapping has been set up, the caller must finalize the
5455  * charge with mem_cgroup_commit_charge().  Or abort the transaction
5456  * with mem_cgroup_cancel_charge() in case page instantiation fails.
5457  */
5458 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5459 			  gfp_t gfp_mask, struct mem_cgroup **memcgp)
5460 {
5461 	struct mem_cgroup *memcg = NULL;
5462 	unsigned int nr_pages = 1;
5463 	int ret = 0;
5464 
5465 	if (mem_cgroup_disabled())
5466 		goto out;
5467 
5468 	if (PageSwapCache(page)) {
5469 		/*
5470 		 * Every swap fault against a single page tries to charge the
5471 		 * page, bail as early as possible.  shmem_unuse() encounters
5472 		 * already charged pages, too.  The USED bit is protected by
5473 		 * the page lock, which serializes swap cache removal, which
5474 		 * in turn serializes uncharging.
5475 		 */
5476 		if (page->mem_cgroup)
5477 			goto out;
5478 	}
5479 
5480 	if (PageTransHuge(page)) {
5481 		nr_pages <<= compound_order(page);
5482 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5483 	}
5484 
5485 	if (do_swap_account && PageSwapCache(page))
5486 		memcg = try_get_mem_cgroup_from_page(page);
5487 	if (!memcg)
5488 		memcg = get_mem_cgroup_from_mm(mm);
5489 
5490 	ret = try_charge(memcg, gfp_mask, nr_pages);
5491 
5492 	css_put(&memcg->css);
5493 
5494 	if (ret == -EINTR) {
5495 		memcg = root_mem_cgroup;
5496 		ret = 0;
5497 	}
5498 out:
5499 	*memcgp = memcg;
5500 	return ret;
5501 }
5502 
5503 /**
5504  * mem_cgroup_commit_charge - commit a page charge
5505  * @page: page to charge
5506  * @memcg: memcg to charge the page to
5507  * @lrucare: page might be on LRU already
5508  *
5509  * Finalize a charge transaction started by mem_cgroup_try_charge(),
5510  * after page->mapping has been set up.  This must happen atomically
5511  * as part of the page instantiation, i.e. under the page table lock
5512  * for anonymous pages, under the page lock for page and swap cache.
5513  *
5514  * In addition, the page must not be on the LRU during the commit, to
5515  * prevent racing with task migration.  If it might be, use @lrucare.
5516  *
5517  * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5518  */
5519 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5520 			      bool lrucare)
5521 {
5522 	unsigned int nr_pages = 1;
5523 
5524 	VM_BUG_ON_PAGE(!page->mapping, page);
5525 	VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5526 
5527 	if (mem_cgroup_disabled())
5528 		return;
5529 	/*
5530 	 * Swap faults will attempt to charge the same page multiple
5531 	 * times.  But reuse_swap_page() might have removed the page
5532 	 * from swapcache already, so we can't check PageSwapCache().
5533 	 */
5534 	if (!memcg)
5535 		return;
5536 
5537 	commit_charge(page, memcg, lrucare);
5538 
5539 	if (PageTransHuge(page)) {
5540 		nr_pages <<= compound_order(page);
5541 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5542 	}
5543 
5544 	local_irq_disable();
5545 	mem_cgroup_charge_statistics(memcg, page, nr_pages);
5546 	memcg_check_events(memcg, page);
5547 	local_irq_enable();
5548 
5549 	if (do_swap_account && PageSwapCache(page)) {
5550 		swp_entry_t entry = { .val = page_private(page) };
5551 		/*
5552 		 * The swap entry might not get freed for a long time,
5553 		 * let's not wait for it.  The page already received a
5554 		 * memory+swap charge, drop the swap entry duplicate.
5555 		 */
5556 		mem_cgroup_uncharge_swap(entry);
5557 	}
5558 }
5559 
5560 /**
5561  * mem_cgroup_cancel_charge - cancel a page charge
5562  * @page: page to charge
5563  * @memcg: memcg to charge the page to
5564  *
5565  * Cancel a charge transaction started by mem_cgroup_try_charge().
5566  */
5567 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5568 {
5569 	unsigned int nr_pages = 1;
5570 
5571 	if (mem_cgroup_disabled())
5572 		return;
5573 	/*
5574 	 * Swap faults will attempt to charge the same page multiple
5575 	 * times.  But reuse_swap_page() might have removed the page
5576 	 * from swapcache already, so we can't check PageSwapCache().
5577 	 */
5578 	if (!memcg)
5579 		return;
5580 
5581 	if (PageTransHuge(page)) {
5582 		nr_pages <<= compound_order(page);
5583 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5584 	}
5585 
5586 	cancel_charge(memcg, nr_pages);
5587 }
5588 
5589 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5590 			   unsigned long nr_anon, unsigned long nr_file,
5591 			   unsigned long nr_huge, struct page *dummy_page)
5592 {
5593 	unsigned long nr_pages = nr_anon + nr_file;
5594 	unsigned long flags;
5595 
5596 	if (!mem_cgroup_is_root(memcg)) {
5597 		page_counter_uncharge(&memcg->memory, nr_pages);
5598 		if (do_swap_account)
5599 			page_counter_uncharge(&memcg->memsw, nr_pages);
5600 		memcg_oom_recover(memcg);
5601 	}
5602 
5603 	local_irq_save(flags);
5604 	__this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5605 	__this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5606 	__this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5607 	__this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5608 	__this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5609 	memcg_check_events(memcg, dummy_page);
5610 	local_irq_restore(flags);
5611 
5612 	if (!mem_cgroup_is_root(memcg))
5613 		css_put_many(&memcg->css, nr_pages);
5614 }
5615 
5616 static void uncharge_list(struct list_head *page_list)
5617 {
5618 	struct mem_cgroup *memcg = NULL;
5619 	unsigned long nr_anon = 0;
5620 	unsigned long nr_file = 0;
5621 	unsigned long nr_huge = 0;
5622 	unsigned long pgpgout = 0;
5623 	struct list_head *next;
5624 	struct page *page;
5625 
5626 	next = page_list->next;
5627 	do {
5628 		unsigned int nr_pages = 1;
5629 
5630 		page = list_entry(next, struct page, lru);
5631 		next = page->lru.next;
5632 
5633 		VM_BUG_ON_PAGE(PageLRU(page), page);
5634 		VM_BUG_ON_PAGE(page_count(page), page);
5635 
5636 		if (!page->mem_cgroup)
5637 			continue;
5638 
5639 		/*
5640 		 * Nobody should be changing or seriously looking at
5641 		 * page->mem_cgroup at this point, we have fully
5642 		 * exclusive access to the page.
5643 		 */
5644 
5645 		if (memcg != page->mem_cgroup) {
5646 			if (memcg) {
5647 				uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5648 					       nr_huge, page);
5649 				pgpgout = nr_anon = nr_file = nr_huge = 0;
5650 			}
5651 			memcg = page->mem_cgroup;
5652 		}
5653 
5654 		if (PageTransHuge(page)) {
5655 			nr_pages <<= compound_order(page);
5656 			VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5657 			nr_huge += nr_pages;
5658 		}
5659 
5660 		if (PageAnon(page))
5661 			nr_anon += nr_pages;
5662 		else
5663 			nr_file += nr_pages;
5664 
5665 		page->mem_cgroup = NULL;
5666 
5667 		pgpgout++;
5668 	} while (next != page_list);
5669 
5670 	if (memcg)
5671 		uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5672 			       nr_huge, page);
5673 }
5674 
5675 /**
5676  * mem_cgroup_uncharge - uncharge a page
5677  * @page: page to uncharge
5678  *
5679  * Uncharge a page previously charged with mem_cgroup_try_charge() and
5680  * mem_cgroup_commit_charge().
5681  */
5682 void mem_cgroup_uncharge(struct page *page)
5683 {
5684 	if (mem_cgroup_disabled())
5685 		return;
5686 
5687 	/* Don't touch page->lru of any random page, pre-check: */
5688 	if (!page->mem_cgroup)
5689 		return;
5690 
5691 	INIT_LIST_HEAD(&page->lru);
5692 	uncharge_list(&page->lru);
5693 }
5694 
5695 /**
5696  * mem_cgroup_uncharge_list - uncharge a list of page
5697  * @page_list: list of pages to uncharge
5698  *
5699  * Uncharge a list of pages previously charged with
5700  * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5701  */
5702 void mem_cgroup_uncharge_list(struct list_head *page_list)
5703 {
5704 	if (mem_cgroup_disabled())
5705 		return;
5706 
5707 	if (!list_empty(page_list))
5708 		uncharge_list(page_list);
5709 }
5710 
5711 /**
5712  * mem_cgroup_migrate - migrate a charge to another page
5713  * @oldpage: currently charged page
5714  * @newpage: page to transfer the charge to
5715  * @lrucare: either or both pages might be on the LRU already
5716  *
5717  * Migrate the charge from @oldpage to @newpage.
5718  *
5719  * Both pages must be locked, @newpage->mapping must be set up.
5720  */
5721 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5722 			bool lrucare)
5723 {
5724 	struct mem_cgroup *memcg;
5725 	int isolated;
5726 
5727 	VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5728 	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5729 	VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5730 	VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5731 	VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5732 	VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5733 		       newpage);
5734 
5735 	if (mem_cgroup_disabled())
5736 		return;
5737 
5738 	/* Page cache replacement: new page already charged? */
5739 	if (newpage->mem_cgroup)
5740 		return;
5741 
5742 	/*
5743 	 * Swapcache readahead pages can get migrated before being
5744 	 * charged, and migration from compaction can happen to an
5745 	 * uncharged page when the PFN walker finds a page that
5746 	 * reclaim just put back on the LRU but has not released yet.
5747 	 */
5748 	memcg = oldpage->mem_cgroup;
5749 	if (!memcg)
5750 		return;
5751 
5752 	if (lrucare)
5753 		lock_page_lru(oldpage, &isolated);
5754 
5755 	oldpage->mem_cgroup = NULL;
5756 
5757 	if (lrucare)
5758 		unlock_page_lru(oldpage, isolated);
5759 
5760 	commit_charge(newpage, memcg, lrucare);
5761 }
5762 
5763 /*
5764  * subsys_initcall() for memory controller.
5765  *
5766  * Some parts like hotcpu_notifier() have to be initialized from this context
5767  * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5768  * everything that doesn't depend on a specific mem_cgroup structure should
5769  * be initialized from here.
5770  */
5771 static int __init mem_cgroup_init(void)
5772 {
5773 	int cpu, node;
5774 
5775 	hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5776 
5777 	for_each_possible_cpu(cpu)
5778 		INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5779 			  drain_local_stock);
5780 
5781 	for_each_node(node) {
5782 		struct mem_cgroup_tree_per_node *rtpn;
5783 		int zone;
5784 
5785 		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5786 				    node_online(node) ? node : NUMA_NO_NODE);
5787 
5788 		for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5789 			struct mem_cgroup_tree_per_zone *rtpz;
5790 
5791 			rtpz = &rtpn->rb_tree_per_zone[zone];
5792 			rtpz->rb_root = RB_ROOT;
5793 			spin_lock_init(&rtpz->lock);
5794 		}
5795 		soft_limit_tree.rb_tree_per_node[node] = rtpn;
5796 	}
5797 
5798 	return 0;
5799 }
5800 subsys_initcall(mem_cgroup_init);
5801 
5802 #ifdef CONFIG_MEMCG_SWAP
5803 /**
5804  * mem_cgroup_swapout - transfer a memsw charge to swap
5805  * @page: page whose memsw charge to transfer
5806  * @entry: swap entry to move the charge to
5807  *
5808  * Transfer the memsw charge of @page to @entry.
5809  */
5810 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5811 {
5812 	struct mem_cgroup *memcg;
5813 	unsigned short oldid;
5814 
5815 	VM_BUG_ON_PAGE(PageLRU(page), page);
5816 	VM_BUG_ON_PAGE(page_count(page), page);
5817 
5818 	if (!do_swap_account)
5819 		return;
5820 
5821 	memcg = page->mem_cgroup;
5822 
5823 	/* Readahead page, never charged */
5824 	if (!memcg)
5825 		return;
5826 
5827 	oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5828 	VM_BUG_ON_PAGE(oldid, page);
5829 	mem_cgroup_swap_statistics(memcg, true);
5830 
5831 	page->mem_cgroup = NULL;
5832 
5833 	if (!mem_cgroup_is_root(memcg))
5834 		page_counter_uncharge(&memcg->memory, 1);
5835 
5836 	/* XXX: caller holds IRQ-safe mapping->tree_lock */
5837 	VM_BUG_ON(!irqs_disabled());
5838 
5839 	mem_cgroup_charge_statistics(memcg, page, -1);
5840 	memcg_check_events(memcg, page);
5841 }
5842 
5843 /**
5844  * mem_cgroup_uncharge_swap - uncharge a swap entry
5845  * @entry: swap entry to uncharge
5846  *
5847  * Drop the memsw charge associated with @entry.
5848  */
5849 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5850 {
5851 	struct mem_cgroup *memcg;
5852 	unsigned short id;
5853 
5854 	if (!do_swap_account)
5855 		return;
5856 
5857 	id = swap_cgroup_record(entry, 0);
5858 	rcu_read_lock();
5859 	memcg = mem_cgroup_from_id(id);
5860 	if (memcg) {
5861 		if (!mem_cgroup_is_root(memcg))
5862 			page_counter_uncharge(&memcg->memsw, 1);
5863 		mem_cgroup_swap_statistics(memcg, false);
5864 		css_put(&memcg->css);
5865 	}
5866 	rcu_read_unlock();
5867 }
5868 
5869 /* for remember boot option*/
5870 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5871 static int really_do_swap_account __initdata = 1;
5872 #else
5873 static int really_do_swap_account __initdata;
5874 #endif
5875 
5876 static int __init enable_swap_account(char *s)
5877 {
5878 	if (!strcmp(s, "1"))
5879 		really_do_swap_account = 1;
5880 	else if (!strcmp(s, "0"))
5881 		really_do_swap_account = 0;
5882 	return 1;
5883 }
5884 __setup("swapaccount=", enable_swap_account);
5885 
5886 static struct cftype memsw_cgroup_files[] = {
5887 	{
5888 		.name = "memsw.usage_in_bytes",
5889 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5890 		.read_u64 = mem_cgroup_read_u64,
5891 	},
5892 	{
5893 		.name = "memsw.max_usage_in_bytes",
5894 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5895 		.write = mem_cgroup_reset,
5896 		.read_u64 = mem_cgroup_read_u64,
5897 	},
5898 	{
5899 		.name = "memsw.limit_in_bytes",
5900 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5901 		.write = mem_cgroup_write,
5902 		.read_u64 = mem_cgroup_read_u64,
5903 	},
5904 	{
5905 		.name = "memsw.failcnt",
5906 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5907 		.write = mem_cgroup_reset,
5908 		.read_u64 = mem_cgroup_read_u64,
5909 	},
5910 	{ },	/* terminate */
5911 };
5912 
5913 static int __init mem_cgroup_swap_init(void)
5914 {
5915 	if (!mem_cgroup_disabled() && really_do_swap_account) {
5916 		do_swap_account = 1;
5917 		WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5918 						  memsw_cgroup_files));
5919 	}
5920 	return 0;
5921 }
5922 subsys_initcall(mem_cgroup_swap_init);
5923 
5924 #endif /* CONFIG_MEMCG_SWAP */
5925