xref: /openbmc/linux/mm/memcontrol.c (revision a9d85efb)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
3  *
4  * Copyright IBM Corporation, 2007
5  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6  *
7  * Copyright 2007 OpenVZ SWsoft Inc
8  * Author: Pavel Emelianov <xemul@openvz.org>
9  *
10  * Memory thresholds
11  * Copyright (C) 2009 Nokia Corporation
12  * Author: Kirill A. Shutemov
13  *
14  * Kernel Memory Controller
15  * Copyright (C) 2012 Parallels Inc. and Google Inc.
16  * Authors: Glauber Costa and Suleiman Souhlal
17  *
18  * Native page reclaim
19  * Charge lifetime sanitation
20  * Lockless page tracking & accounting
21  * Unified hierarchy configuration model
22  * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23  *
24  * Per memcg lru locking
25  * Copyright (C) 2020 Alibaba, Inc, Alex Shi
26  */
27 
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
53 #include <linux/fs.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
65 #include "internal.h"
66 #include <net/sock.h>
67 #include <net/ip.h>
68 #include "slab.h"
69 
70 #include <linux/uaccess.h>
71 
72 #include <trace/events/vmscan.h>
73 
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
76 
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
78 
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
81 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
82 
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket __ro_after_init;
85 
86 /* Kernel memory accounting disabled? */
87 bool cgroup_memory_nokmem __ro_after_init;
88 
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 bool cgroup_memory_noswap __ro_after_init;
92 #else
93 #define cgroup_memory_noswap		1
94 #endif
95 
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
98 #endif
99 
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
102 {
103 	return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
104 }
105 
106 /* memcg and lruvec stats flushing */
107 static void flush_memcg_stats_dwork(struct work_struct *w);
108 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
109 static DEFINE_SPINLOCK(stats_flush_lock);
110 
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
113 
114 /*
115  * Cgroups above their limits are maintained in a RB-Tree, independent of
116  * their hierarchy representation
117  */
118 
119 struct mem_cgroup_tree_per_node {
120 	struct rb_root rb_root;
121 	struct rb_node *rb_rightmost;
122 	spinlock_t lock;
123 };
124 
125 struct mem_cgroup_tree {
126 	struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
127 };
128 
129 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
130 
131 /* for OOM */
132 struct mem_cgroup_eventfd_list {
133 	struct list_head list;
134 	struct eventfd_ctx *eventfd;
135 };
136 
137 /*
138  * cgroup_event represents events which userspace want to receive.
139  */
140 struct mem_cgroup_event {
141 	/*
142 	 * memcg which the event belongs to.
143 	 */
144 	struct mem_cgroup *memcg;
145 	/*
146 	 * eventfd to signal userspace about the event.
147 	 */
148 	struct eventfd_ctx *eventfd;
149 	/*
150 	 * Each of these stored in a list by the cgroup.
151 	 */
152 	struct list_head list;
153 	/*
154 	 * register_event() callback will be used to add new userspace
155 	 * waiter for changes related to this event.  Use eventfd_signal()
156 	 * on eventfd to send notification to userspace.
157 	 */
158 	int (*register_event)(struct mem_cgroup *memcg,
159 			      struct eventfd_ctx *eventfd, const char *args);
160 	/*
161 	 * unregister_event() callback will be called when userspace closes
162 	 * the eventfd or on cgroup removing.  This callback must be set,
163 	 * if you want provide notification functionality.
164 	 */
165 	void (*unregister_event)(struct mem_cgroup *memcg,
166 				 struct eventfd_ctx *eventfd);
167 	/*
168 	 * All fields below needed to unregister event when
169 	 * userspace closes eventfd.
170 	 */
171 	poll_table pt;
172 	wait_queue_head_t *wqh;
173 	wait_queue_entry_t wait;
174 	struct work_struct remove;
175 };
176 
177 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
178 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
179 
180 /* Stuffs for move charges at task migration. */
181 /*
182  * Types of charges to be moved.
183  */
184 #define MOVE_ANON	0x1U
185 #define MOVE_FILE	0x2U
186 #define MOVE_MASK	(MOVE_ANON | MOVE_FILE)
187 
188 /* "mc" and its members are protected by cgroup_mutex */
189 static struct move_charge_struct {
190 	spinlock_t	  lock; /* for from, to */
191 	struct mm_struct  *mm;
192 	struct mem_cgroup *from;
193 	struct mem_cgroup *to;
194 	unsigned long flags;
195 	unsigned long precharge;
196 	unsigned long moved_charge;
197 	unsigned long moved_swap;
198 	struct task_struct *moving_task;	/* a task moving charges */
199 	wait_queue_head_t waitq;		/* a waitq for other context */
200 } mc = {
201 	.lock = __SPIN_LOCK_UNLOCKED(mc.lock),
202 	.waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
203 };
204 
205 /*
206  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
207  * limit reclaim to prevent infinite loops, if they ever occur.
208  */
209 #define	MEM_CGROUP_MAX_RECLAIM_LOOPS		100
210 #define	MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS	2
211 
212 /* for encoding cft->private value on file */
213 enum res_type {
214 	_MEM,
215 	_MEMSWAP,
216 	_OOM_TYPE,
217 	_KMEM,
218 	_TCP,
219 };
220 
221 #define MEMFILE_PRIVATE(x, val)	((x) << 16 | (val))
222 #define MEMFILE_TYPE(val)	((val) >> 16 & 0xffff)
223 #define MEMFILE_ATTR(val)	((val) & 0xffff)
224 /* Used for OOM notifier */
225 #define OOM_CONTROL		(0)
226 
227 /*
228  * Iteration constructs for visiting all cgroups (under a tree).  If
229  * loops are exited prematurely (break), mem_cgroup_iter_break() must
230  * be used for reference counting.
231  */
232 #define for_each_mem_cgroup_tree(iter, root)		\
233 	for (iter = mem_cgroup_iter(root, NULL, NULL);	\
234 	     iter != NULL;				\
235 	     iter = mem_cgroup_iter(root, iter, NULL))
236 
237 #define for_each_mem_cgroup(iter)			\
238 	for (iter = mem_cgroup_iter(NULL, NULL, NULL);	\
239 	     iter != NULL;				\
240 	     iter = mem_cgroup_iter(NULL, iter, NULL))
241 
242 static inline bool should_force_charge(void)
243 {
244 	return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
245 		(current->flags & PF_EXITING);
246 }
247 
248 /* Some nice accessors for the vmpressure. */
249 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 {
251 	if (!memcg)
252 		memcg = root_mem_cgroup;
253 	return &memcg->vmpressure;
254 }
255 
256 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
257 {
258 	return container_of(vmpr, struct mem_cgroup, vmpressure);
259 }
260 
261 #ifdef CONFIG_MEMCG_KMEM
262 extern spinlock_t css_set_lock;
263 
264 bool mem_cgroup_kmem_disabled(void)
265 {
266 	return cgroup_memory_nokmem;
267 }
268 
269 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
270 				      unsigned int nr_pages);
271 
272 static void obj_cgroup_release(struct percpu_ref *ref)
273 {
274 	struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
275 	unsigned int nr_bytes;
276 	unsigned int nr_pages;
277 	unsigned long flags;
278 
279 	/*
280 	 * At this point all allocated objects are freed, and
281 	 * objcg->nr_charged_bytes can't have an arbitrary byte value.
282 	 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
283 	 *
284 	 * The following sequence can lead to it:
285 	 * 1) CPU0: objcg == stock->cached_objcg
286 	 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
287 	 *          PAGE_SIZE bytes are charged
288 	 * 3) CPU1: a process from another memcg is allocating something,
289 	 *          the stock if flushed,
290 	 *          objcg->nr_charged_bytes = PAGE_SIZE - 92
291 	 * 5) CPU0: we do release this object,
292 	 *          92 bytes are added to stock->nr_bytes
293 	 * 6) CPU0: stock is flushed,
294 	 *          92 bytes are added to objcg->nr_charged_bytes
295 	 *
296 	 * In the result, nr_charged_bytes == PAGE_SIZE.
297 	 * This page will be uncharged in obj_cgroup_release().
298 	 */
299 	nr_bytes = atomic_read(&objcg->nr_charged_bytes);
300 	WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
301 	nr_pages = nr_bytes >> PAGE_SHIFT;
302 
303 	if (nr_pages)
304 		obj_cgroup_uncharge_pages(objcg, nr_pages);
305 
306 	spin_lock_irqsave(&css_set_lock, flags);
307 	list_del(&objcg->list);
308 	spin_unlock_irqrestore(&css_set_lock, flags);
309 
310 	percpu_ref_exit(ref);
311 	kfree_rcu(objcg, rcu);
312 }
313 
314 static struct obj_cgroup *obj_cgroup_alloc(void)
315 {
316 	struct obj_cgroup *objcg;
317 	int ret;
318 
319 	objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
320 	if (!objcg)
321 		return NULL;
322 
323 	ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
324 			      GFP_KERNEL);
325 	if (ret) {
326 		kfree(objcg);
327 		return NULL;
328 	}
329 	INIT_LIST_HEAD(&objcg->list);
330 	return objcg;
331 }
332 
333 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
334 				  struct mem_cgroup *parent)
335 {
336 	struct obj_cgroup *objcg, *iter;
337 
338 	objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
339 
340 	spin_lock_irq(&css_set_lock);
341 
342 	/* 1) Ready to reparent active objcg. */
343 	list_add(&objcg->list, &memcg->objcg_list);
344 	/* 2) Reparent active objcg and already reparented objcgs to parent. */
345 	list_for_each_entry(iter, &memcg->objcg_list, list)
346 		WRITE_ONCE(iter->memcg, parent);
347 	/* 3) Move already reparented objcgs to the parent's list */
348 	list_splice(&memcg->objcg_list, &parent->objcg_list);
349 
350 	spin_unlock_irq(&css_set_lock);
351 
352 	percpu_ref_kill(&objcg->refcnt);
353 }
354 
355 /*
356  * This will be used as a shrinker list's index.
357  * The main reason for not using cgroup id for this:
358  *  this works better in sparse environments, where we have a lot of memcgs,
359  *  but only a few kmem-limited. Or also, if we have, for instance, 200
360  *  memcgs, and none but the 200th is kmem-limited, we'd have to have a
361  *  200 entry array for that.
362  *
363  * The current size of the caches array is stored in memcg_nr_cache_ids. It
364  * will double each time we have to increase it.
365  */
366 static DEFINE_IDA(memcg_cache_ida);
367 int memcg_nr_cache_ids;
368 
369 /* Protects memcg_nr_cache_ids */
370 static DECLARE_RWSEM(memcg_cache_ids_sem);
371 
372 void memcg_get_cache_ids(void)
373 {
374 	down_read(&memcg_cache_ids_sem);
375 }
376 
377 void memcg_put_cache_ids(void)
378 {
379 	up_read(&memcg_cache_ids_sem);
380 }
381 
382 /*
383  * MIN_SIZE is different than 1, because we would like to avoid going through
384  * the alloc/free process all the time. In a small machine, 4 kmem-limited
385  * cgroups is a reasonable guess. In the future, it could be a parameter or
386  * tunable, but that is strictly not necessary.
387  *
388  * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
389  * this constant directly from cgroup, but it is understandable that this is
390  * better kept as an internal representation in cgroup.c. In any case, the
391  * cgrp_id space is not getting any smaller, and we don't have to necessarily
392  * increase ours as well if it increases.
393  */
394 #define MEMCG_CACHES_MIN_SIZE 4
395 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
396 
397 /*
398  * A lot of the calls to the cache allocation functions are expected to be
399  * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
400  * conditional to this static branch, we'll have to allow modules that does
401  * kmem_cache_alloc and the such to see this symbol as well
402  */
403 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
404 EXPORT_SYMBOL(memcg_kmem_enabled_key);
405 #endif
406 
407 /**
408  * mem_cgroup_css_from_page - css of the memcg associated with a page
409  * @page: page of interest
410  *
411  * If memcg is bound to the default hierarchy, css of the memcg associated
412  * with @page is returned.  The returned css remains associated with @page
413  * until it is released.
414  *
415  * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
416  * is returned.
417  */
418 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
419 {
420 	struct mem_cgroup *memcg;
421 
422 	memcg = page_memcg(page);
423 
424 	if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
425 		memcg = root_mem_cgroup;
426 
427 	return &memcg->css;
428 }
429 
430 /**
431  * page_cgroup_ino - return inode number of the memcg a page is charged to
432  * @page: the page
433  *
434  * Look up the closest online ancestor of the memory cgroup @page is charged to
435  * and return its inode number or 0 if @page is not charged to any cgroup. It
436  * is safe to call this function without holding a reference to @page.
437  *
438  * Note, this function is inherently racy, because there is nothing to prevent
439  * the cgroup inode from getting torn down and potentially reallocated a moment
440  * after page_cgroup_ino() returns, so it only should be used by callers that
441  * do not care (such as procfs interfaces).
442  */
443 ino_t page_cgroup_ino(struct page *page)
444 {
445 	struct mem_cgroup *memcg;
446 	unsigned long ino = 0;
447 
448 	rcu_read_lock();
449 	memcg = page_memcg_check(page);
450 
451 	while (memcg && !(memcg->css.flags & CSS_ONLINE))
452 		memcg = parent_mem_cgroup(memcg);
453 	if (memcg)
454 		ino = cgroup_ino(memcg->css.cgroup);
455 	rcu_read_unlock();
456 	return ino;
457 }
458 
459 static struct mem_cgroup_per_node *
460 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
461 {
462 	int nid = page_to_nid(page);
463 
464 	return memcg->nodeinfo[nid];
465 }
466 
467 static struct mem_cgroup_tree_per_node *
468 soft_limit_tree_node(int nid)
469 {
470 	return soft_limit_tree.rb_tree_per_node[nid];
471 }
472 
473 static struct mem_cgroup_tree_per_node *
474 soft_limit_tree_from_page(struct page *page)
475 {
476 	int nid = page_to_nid(page);
477 
478 	return soft_limit_tree.rb_tree_per_node[nid];
479 }
480 
481 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
482 					 struct mem_cgroup_tree_per_node *mctz,
483 					 unsigned long new_usage_in_excess)
484 {
485 	struct rb_node **p = &mctz->rb_root.rb_node;
486 	struct rb_node *parent = NULL;
487 	struct mem_cgroup_per_node *mz_node;
488 	bool rightmost = true;
489 
490 	if (mz->on_tree)
491 		return;
492 
493 	mz->usage_in_excess = new_usage_in_excess;
494 	if (!mz->usage_in_excess)
495 		return;
496 	while (*p) {
497 		parent = *p;
498 		mz_node = rb_entry(parent, struct mem_cgroup_per_node,
499 					tree_node);
500 		if (mz->usage_in_excess < mz_node->usage_in_excess) {
501 			p = &(*p)->rb_left;
502 			rightmost = false;
503 		} else {
504 			p = &(*p)->rb_right;
505 		}
506 	}
507 
508 	if (rightmost)
509 		mctz->rb_rightmost = &mz->tree_node;
510 
511 	rb_link_node(&mz->tree_node, parent, p);
512 	rb_insert_color(&mz->tree_node, &mctz->rb_root);
513 	mz->on_tree = true;
514 }
515 
516 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
517 					 struct mem_cgroup_tree_per_node *mctz)
518 {
519 	if (!mz->on_tree)
520 		return;
521 
522 	if (&mz->tree_node == mctz->rb_rightmost)
523 		mctz->rb_rightmost = rb_prev(&mz->tree_node);
524 
525 	rb_erase(&mz->tree_node, &mctz->rb_root);
526 	mz->on_tree = false;
527 }
528 
529 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
530 				       struct mem_cgroup_tree_per_node *mctz)
531 {
532 	unsigned long flags;
533 
534 	spin_lock_irqsave(&mctz->lock, flags);
535 	__mem_cgroup_remove_exceeded(mz, mctz);
536 	spin_unlock_irqrestore(&mctz->lock, flags);
537 }
538 
539 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
540 {
541 	unsigned long nr_pages = page_counter_read(&memcg->memory);
542 	unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
543 	unsigned long excess = 0;
544 
545 	if (nr_pages > soft_limit)
546 		excess = nr_pages - soft_limit;
547 
548 	return excess;
549 }
550 
551 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
552 {
553 	unsigned long excess;
554 	struct mem_cgroup_per_node *mz;
555 	struct mem_cgroup_tree_per_node *mctz;
556 
557 	mctz = soft_limit_tree_from_page(page);
558 	if (!mctz)
559 		return;
560 	/*
561 	 * Necessary to update all ancestors when hierarchy is used.
562 	 * because their event counter is not touched.
563 	 */
564 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
565 		mz = mem_cgroup_page_nodeinfo(memcg, page);
566 		excess = soft_limit_excess(memcg);
567 		/*
568 		 * We have to update the tree if mz is on RB-tree or
569 		 * mem is over its softlimit.
570 		 */
571 		if (excess || mz->on_tree) {
572 			unsigned long flags;
573 
574 			spin_lock_irqsave(&mctz->lock, flags);
575 			/* if on-tree, remove it */
576 			if (mz->on_tree)
577 				__mem_cgroup_remove_exceeded(mz, mctz);
578 			/*
579 			 * Insert again. mz->usage_in_excess will be updated.
580 			 * If excess is 0, no tree ops.
581 			 */
582 			__mem_cgroup_insert_exceeded(mz, mctz, excess);
583 			spin_unlock_irqrestore(&mctz->lock, flags);
584 		}
585 	}
586 }
587 
588 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
589 {
590 	struct mem_cgroup_tree_per_node *mctz;
591 	struct mem_cgroup_per_node *mz;
592 	int nid;
593 
594 	for_each_node(nid) {
595 		mz = memcg->nodeinfo[nid];
596 		mctz = soft_limit_tree_node(nid);
597 		if (mctz)
598 			mem_cgroup_remove_exceeded(mz, mctz);
599 	}
600 }
601 
602 static struct mem_cgroup_per_node *
603 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
604 {
605 	struct mem_cgroup_per_node *mz;
606 
607 retry:
608 	mz = NULL;
609 	if (!mctz->rb_rightmost)
610 		goto done;		/* Nothing to reclaim from */
611 
612 	mz = rb_entry(mctz->rb_rightmost,
613 		      struct mem_cgroup_per_node, tree_node);
614 	/*
615 	 * Remove the node now but someone else can add it back,
616 	 * we will to add it back at the end of reclaim to its correct
617 	 * position in the tree.
618 	 */
619 	__mem_cgroup_remove_exceeded(mz, mctz);
620 	if (!soft_limit_excess(mz->memcg) ||
621 	    !css_tryget(&mz->memcg->css))
622 		goto retry;
623 done:
624 	return mz;
625 }
626 
627 static struct mem_cgroup_per_node *
628 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
629 {
630 	struct mem_cgroup_per_node *mz;
631 
632 	spin_lock_irq(&mctz->lock);
633 	mz = __mem_cgroup_largest_soft_limit_node(mctz);
634 	spin_unlock_irq(&mctz->lock);
635 	return mz;
636 }
637 
638 /**
639  * __mod_memcg_state - update cgroup memory statistics
640  * @memcg: the memory cgroup
641  * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
642  * @val: delta to add to the counter, can be negative
643  */
644 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
645 {
646 	if (mem_cgroup_disabled())
647 		return;
648 
649 	__this_cpu_add(memcg->vmstats_percpu->state[idx], val);
650 	cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
651 }
652 
653 /* idx can be of type enum memcg_stat_item or node_stat_item. */
654 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
655 {
656 	long x = 0;
657 	int cpu;
658 
659 	for_each_possible_cpu(cpu)
660 		x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
661 #ifdef CONFIG_SMP
662 	if (x < 0)
663 		x = 0;
664 #endif
665 	return x;
666 }
667 
668 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
669 			      int val)
670 {
671 	struct mem_cgroup_per_node *pn;
672 	struct mem_cgroup *memcg;
673 
674 	pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
675 	memcg = pn->memcg;
676 
677 	/* Update memcg */
678 	__mod_memcg_state(memcg, idx, val);
679 
680 	/* Update lruvec */
681 	__this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
682 }
683 
684 /**
685  * __mod_lruvec_state - update lruvec memory statistics
686  * @lruvec: the lruvec
687  * @idx: the stat item
688  * @val: delta to add to the counter, can be negative
689  *
690  * The lruvec is the intersection of the NUMA node and a cgroup. This
691  * function updates the all three counters that are affected by a
692  * change of state at this level: per-node, per-cgroup, per-lruvec.
693  */
694 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
695 			int val)
696 {
697 	/* Update node */
698 	__mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
699 
700 	/* Update memcg and lruvec */
701 	if (!mem_cgroup_disabled())
702 		__mod_memcg_lruvec_state(lruvec, idx, val);
703 }
704 
705 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
706 			     int val)
707 {
708 	struct page *head = compound_head(page); /* rmap on tail pages */
709 	struct mem_cgroup *memcg;
710 	pg_data_t *pgdat = page_pgdat(page);
711 	struct lruvec *lruvec;
712 
713 	rcu_read_lock();
714 	memcg = page_memcg(head);
715 	/* Untracked pages have no memcg, no lruvec. Update only the node */
716 	if (!memcg) {
717 		rcu_read_unlock();
718 		__mod_node_page_state(pgdat, idx, val);
719 		return;
720 	}
721 
722 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
723 	__mod_lruvec_state(lruvec, idx, val);
724 	rcu_read_unlock();
725 }
726 EXPORT_SYMBOL(__mod_lruvec_page_state);
727 
728 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
729 {
730 	pg_data_t *pgdat = page_pgdat(virt_to_page(p));
731 	struct mem_cgroup *memcg;
732 	struct lruvec *lruvec;
733 
734 	rcu_read_lock();
735 	memcg = mem_cgroup_from_obj(p);
736 
737 	/*
738 	 * Untracked pages have no memcg, no lruvec. Update only the
739 	 * node. If we reparent the slab objects to the root memcg,
740 	 * when we free the slab object, we need to update the per-memcg
741 	 * vmstats to keep it correct for the root memcg.
742 	 */
743 	if (!memcg) {
744 		__mod_node_page_state(pgdat, idx, val);
745 	} else {
746 		lruvec = mem_cgroup_lruvec(memcg, pgdat);
747 		__mod_lruvec_state(lruvec, idx, val);
748 	}
749 	rcu_read_unlock();
750 }
751 
752 /*
753  * mod_objcg_mlstate() may be called with irq enabled, so
754  * mod_memcg_lruvec_state() should be used.
755  */
756 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
757 				     struct pglist_data *pgdat,
758 				     enum node_stat_item idx, int nr)
759 {
760 	struct mem_cgroup *memcg;
761 	struct lruvec *lruvec;
762 
763 	rcu_read_lock();
764 	memcg = obj_cgroup_memcg(objcg);
765 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
766 	mod_memcg_lruvec_state(lruvec, idx, nr);
767 	rcu_read_unlock();
768 }
769 
770 /**
771  * __count_memcg_events - account VM events in a cgroup
772  * @memcg: the memory cgroup
773  * @idx: the event item
774  * @count: the number of events that occurred
775  */
776 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
777 			  unsigned long count)
778 {
779 	if (mem_cgroup_disabled())
780 		return;
781 
782 	__this_cpu_add(memcg->vmstats_percpu->events[idx], count);
783 	cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
784 }
785 
786 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
787 {
788 	return READ_ONCE(memcg->vmstats.events[event]);
789 }
790 
791 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
792 {
793 	long x = 0;
794 	int cpu;
795 
796 	for_each_possible_cpu(cpu)
797 		x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
798 	return x;
799 }
800 
801 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
802 					 struct page *page,
803 					 int nr_pages)
804 {
805 	/* pagein of a big page is an event. So, ignore page size */
806 	if (nr_pages > 0)
807 		__count_memcg_events(memcg, PGPGIN, 1);
808 	else {
809 		__count_memcg_events(memcg, PGPGOUT, 1);
810 		nr_pages = -nr_pages; /* for event */
811 	}
812 
813 	__this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
814 }
815 
816 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
817 				       enum mem_cgroup_events_target target)
818 {
819 	unsigned long val, next;
820 
821 	val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
822 	next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
823 	/* from time_after() in jiffies.h */
824 	if ((long)(next - val) < 0) {
825 		switch (target) {
826 		case MEM_CGROUP_TARGET_THRESH:
827 			next = val + THRESHOLDS_EVENTS_TARGET;
828 			break;
829 		case MEM_CGROUP_TARGET_SOFTLIMIT:
830 			next = val + SOFTLIMIT_EVENTS_TARGET;
831 			break;
832 		default:
833 			break;
834 		}
835 		__this_cpu_write(memcg->vmstats_percpu->targets[target], next);
836 		return true;
837 	}
838 	return false;
839 }
840 
841 /*
842  * Check events in order.
843  *
844  */
845 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
846 {
847 	/* threshold event is triggered in finer grain than soft limit */
848 	if (unlikely(mem_cgroup_event_ratelimit(memcg,
849 						MEM_CGROUP_TARGET_THRESH))) {
850 		bool do_softlimit;
851 
852 		do_softlimit = mem_cgroup_event_ratelimit(memcg,
853 						MEM_CGROUP_TARGET_SOFTLIMIT);
854 		mem_cgroup_threshold(memcg);
855 		if (unlikely(do_softlimit))
856 			mem_cgroup_update_tree(memcg, page);
857 	}
858 }
859 
860 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
861 {
862 	/*
863 	 * mm_update_next_owner() may clear mm->owner to NULL
864 	 * if it races with swapoff, page migration, etc.
865 	 * So this can be called with p == NULL.
866 	 */
867 	if (unlikely(!p))
868 		return NULL;
869 
870 	return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
871 }
872 EXPORT_SYMBOL(mem_cgroup_from_task);
873 
874 static __always_inline struct mem_cgroup *active_memcg(void)
875 {
876 	if (!in_task())
877 		return this_cpu_read(int_active_memcg);
878 	else
879 		return current->active_memcg;
880 }
881 
882 /**
883  * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
884  * @mm: mm from which memcg should be extracted. It can be NULL.
885  *
886  * Obtain a reference on mm->memcg and returns it if successful. If mm
887  * is NULL, then the memcg is chosen as follows:
888  * 1) The active memcg, if set.
889  * 2) current->mm->memcg, if available
890  * 3) root memcg
891  * If mem_cgroup is disabled, NULL is returned.
892  */
893 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
894 {
895 	struct mem_cgroup *memcg;
896 
897 	if (mem_cgroup_disabled())
898 		return NULL;
899 
900 	/*
901 	 * Page cache insertions can happen without an
902 	 * actual mm context, e.g. during disk probing
903 	 * on boot, loopback IO, acct() writes etc.
904 	 *
905 	 * No need to css_get on root memcg as the reference
906 	 * counting is disabled on the root level in the
907 	 * cgroup core. See CSS_NO_REF.
908 	 */
909 	if (unlikely(!mm)) {
910 		memcg = active_memcg();
911 		if (unlikely(memcg)) {
912 			/* remote memcg must hold a ref */
913 			css_get(&memcg->css);
914 			return memcg;
915 		}
916 		mm = current->mm;
917 		if (unlikely(!mm))
918 			return root_mem_cgroup;
919 	}
920 
921 	rcu_read_lock();
922 	do {
923 		memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
924 		if (unlikely(!memcg))
925 			memcg = root_mem_cgroup;
926 	} while (!css_tryget(&memcg->css));
927 	rcu_read_unlock();
928 	return memcg;
929 }
930 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
931 
932 static __always_inline bool memcg_kmem_bypass(void)
933 {
934 	/* Allow remote memcg charging from any context. */
935 	if (unlikely(active_memcg()))
936 		return false;
937 
938 	/* Memcg to charge can't be determined. */
939 	if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
940 		return true;
941 
942 	return false;
943 }
944 
945 /**
946  * mem_cgroup_iter - iterate over memory cgroup hierarchy
947  * @root: hierarchy root
948  * @prev: previously returned memcg, NULL on first invocation
949  * @reclaim: cookie for shared reclaim walks, NULL for full walks
950  *
951  * Returns references to children of the hierarchy below @root, or
952  * @root itself, or %NULL after a full round-trip.
953  *
954  * Caller must pass the return value in @prev on subsequent
955  * invocations for reference counting, or use mem_cgroup_iter_break()
956  * to cancel a hierarchy walk before the round-trip is complete.
957  *
958  * Reclaimers can specify a node in @reclaim to divide up the memcgs
959  * in the hierarchy among all concurrent reclaimers operating on the
960  * same node.
961  */
962 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
963 				   struct mem_cgroup *prev,
964 				   struct mem_cgroup_reclaim_cookie *reclaim)
965 {
966 	struct mem_cgroup_reclaim_iter *iter;
967 	struct cgroup_subsys_state *css = NULL;
968 	struct mem_cgroup *memcg = NULL;
969 	struct mem_cgroup *pos = NULL;
970 
971 	if (mem_cgroup_disabled())
972 		return NULL;
973 
974 	if (!root)
975 		root = root_mem_cgroup;
976 
977 	if (prev && !reclaim)
978 		pos = prev;
979 
980 	rcu_read_lock();
981 
982 	if (reclaim) {
983 		struct mem_cgroup_per_node *mz;
984 
985 		mz = root->nodeinfo[reclaim->pgdat->node_id];
986 		iter = &mz->iter;
987 
988 		if (prev && reclaim->generation != iter->generation)
989 			goto out_unlock;
990 
991 		while (1) {
992 			pos = READ_ONCE(iter->position);
993 			if (!pos || css_tryget(&pos->css))
994 				break;
995 			/*
996 			 * css reference reached zero, so iter->position will
997 			 * be cleared by ->css_released. However, we should not
998 			 * rely on this happening soon, because ->css_released
999 			 * is called from a work queue, and by busy-waiting we
1000 			 * might block it. So we clear iter->position right
1001 			 * away.
1002 			 */
1003 			(void)cmpxchg(&iter->position, pos, NULL);
1004 		}
1005 	}
1006 
1007 	if (pos)
1008 		css = &pos->css;
1009 
1010 	for (;;) {
1011 		css = css_next_descendant_pre(css, &root->css);
1012 		if (!css) {
1013 			/*
1014 			 * Reclaimers share the hierarchy walk, and a
1015 			 * new one might jump in right at the end of
1016 			 * the hierarchy - make sure they see at least
1017 			 * one group and restart from the beginning.
1018 			 */
1019 			if (!prev)
1020 				continue;
1021 			break;
1022 		}
1023 
1024 		/*
1025 		 * Verify the css and acquire a reference.  The root
1026 		 * is provided by the caller, so we know it's alive
1027 		 * and kicking, and don't take an extra reference.
1028 		 */
1029 		memcg = mem_cgroup_from_css(css);
1030 
1031 		if (css == &root->css)
1032 			break;
1033 
1034 		if (css_tryget(css))
1035 			break;
1036 
1037 		memcg = NULL;
1038 	}
1039 
1040 	if (reclaim) {
1041 		/*
1042 		 * The position could have already been updated by a competing
1043 		 * thread, so check that the value hasn't changed since we read
1044 		 * it to avoid reclaiming from the same cgroup twice.
1045 		 */
1046 		(void)cmpxchg(&iter->position, pos, memcg);
1047 
1048 		if (pos)
1049 			css_put(&pos->css);
1050 
1051 		if (!memcg)
1052 			iter->generation++;
1053 		else if (!prev)
1054 			reclaim->generation = iter->generation;
1055 	}
1056 
1057 out_unlock:
1058 	rcu_read_unlock();
1059 	if (prev && prev != root)
1060 		css_put(&prev->css);
1061 
1062 	return memcg;
1063 }
1064 
1065 /**
1066  * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1067  * @root: hierarchy root
1068  * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1069  */
1070 void mem_cgroup_iter_break(struct mem_cgroup *root,
1071 			   struct mem_cgroup *prev)
1072 {
1073 	if (!root)
1074 		root = root_mem_cgroup;
1075 	if (prev && prev != root)
1076 		css_put(&prev->css);
1077 }
1078 
1079 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1080 					struct mem_cgroup *dead_memcg)
1081 {
1082 	struct mem_cgroup_reclaim_iter *iter;
1083 	struct mem_cgroup_per_node *mz;
1084 	int nid;
1085 
1086 	for_each_node(nid) {
1087 		mz = from->nodeinfo[nid];
1088 		iter = &mz->iter;
1089 		cmpxchg(&iter->position, dead_memcg, NULL);
1090 	}
1091 }
1092 
1093 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1094 {
1095 	struct mem_cgroup *memcg = dead_memcg;
1096 	struct mem_cgroup *last;
1097 
1098 	do {
1099 		__invalidate_reclaim_iterators(memcg, dead_memcg);
1100 		last = memcg;
1101 	} while ((memcg = parent_mem_cgroup(memcg)));
1102 
1103 	/*
1104 	 * When cgruop1 non-hierarchy mode is used,
1105 	 * parent_mem_cgroup() does not walk all the way up to the
1106 	 * cgroup root (root_mem_cgroup). So we have to handle
1107 	 * dead_memcg from cgroup root separately.
1108 	 */
1109 	if (last != root_mem_cgroup)
1110 		__invalidate_reclaim_iterators(root_mem_cgroup,
1111 						dead_memcg);
1112 }
1113 
1114 /**
1115  * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1116  * @memcg: hierarchy root
1117  * @fn: function to call for each task
1118  * @arg: argument passed to @fn
1119  *
1120  * This function iterates over tasks attached to @memcg or to any of its
1121  * descendants and calls @fn for each task. If @fn returns a non-zero
1122  * value, the function breaks the iteration loop and returns the value.
1123  * Otherwise, it will iterate over all tasks and return 0.
1124  *
1125  * This function must not be called for the root memory cgroup.
1126  */
1127 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1128 			  int (*fn)(struct task_struct *, void *), void *arg)
1129 {
1130 	struct mem_cgroup *iter;
1131 	int ret = 0;
1132 
1133 	BUG_ON(memcg == root_mem_cgroup);
1134 
1135 	for_each_mem_cgroup_tree(iter, memcg) {
1136 		struct css_task_iter it;
1137 		struct task_struct *task;
1138 
1139 		css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1140 		while (!ret && (task = css_task_iter_next(&it)))
1141 			ret = fn(task, arg);
1142 		css_task_iter_end(&it);
1143 		if (ret) {
1144 			mem_cgroup_iter_break(memcg, iter);
1145 			break;
1146 		}
1147 	}
1148 	return ret;
1149 }
1150 
1151 #ifdef CONFIG_DEBUG_VM
1152 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1153 {
1154 	struct mem_cgroup *memcg;
1155 
1156 	if (mem_cgroup_disabled())
1157 		return;
1158 
1159 	memcg = page_memcg(page);
1160 
1161 	if (!memcg)
1162 		VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1163 	else
1164 		VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1165 }
1166 #endif
1167 
1168 /**
1169  * lock_page_lruvec - lock and return lruvec for a given page.
1170  * @page: the page
1171  *
1172  * These functions are safe to use under any of the following conditions:
1173  * - page locked
1174  * - PageLRU cleared
1175  * - lock_page_memcg()
1176  * - page->_refcount is zero
1177  */
1178 struct lruvec *lock_page_lruvec(struct page *page)
1179 {
1180 	struct lruvec *lruvec;
1181 
1182 	lruvec = mem_cgroup_page_lruvec(page);
1183 	spin_lock(&lruvec->lru_lock);
1184 
1185 	lruvec_memcg_debug(lruvec, page);
1186 
1187 	return lruvec;
1188 }
1189 
1190 struct lruvec *lock_page_lruvec_irq(struct page *page)
1191 {
1192 	struct lruvec *lruvec;
1193 
1194 	lruvec = mem_cgroup_page_lruvec(page);
1195 	spin_lock_irq(&lruvec->lru_lock);
1196 
1197 	lruvec_memcg_debug(lruvec, page);
1198 
1199 	return lruvec;
1200 }
1201 
1202 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1203 {
1204 	struct lruvec *lruvec;
1205 
1206 	lruvec = mem_cgroup_page_lruvec(page);
1207 	spin_lock_irqsave(&lruvec->lru_lock, *flags);
1208 
1209 	lruvec_memcg_debug(lruvec, page);
1210 
1211 	return lruvec;
1212 }
1213 
1214 /**
1215  * mem_cgroup_update_lru_size - account for adding or removing an lru page
1216  * @lruvec: mem_cgroup per zone lru vector
1217  * @lru: index of lru list the page is sitting on
1218  * @zid: zone id of the accounted pages
1219  * @nr_pages: positive when adding or negative when removing
1220  *
1221  * This function must be called under lru_lock, just before a page is added
1222  * to or just after a page is removed from an lru list (that ordering being
1223  * so as to allow it to check that lru_size 0 is consistent with list_empty).
1224  */
1225 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1226 				int zid, int nr_pages)
1227 {
1228 	struct mem_cgroup_per_node *mz;
1229 	unsigned long *lru_size;
1230 	long size;
1231 
1232 	if (mem_cgroup_disabled())
1233 		return;
1234 
1235 	mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1236 	lru_size = &mz->lru_zone_size[zid][lru];
1237 
1238 	if (nr_pages < 0)
1239 		*lru_size += nr_pages;
1240 
1241 	size = *lru_size;
1242 	if (WARN_ONCE(size < 0,
1243 		"%s(%p, %d, %d): lru_size %ld\n",
1244 		__func__, lruvec, lru, nr_pages, size)) {
1245 		VM_BUG_ON(1);
1246 		*lru_size = 0;
1247 	}
1248 
1249 	if (nr_pages > 0)
1250 		*lru_size += nr_pages;
1251 }
1252 
1253 /**
1254  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1255  * @memcg: the memory cgroup
1256  *
1257  * Returns the maximum amount of memory @mem can be charged with, in
1258  * pages.
1259  */
1260 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1261 {
1262 	unsigned long margin = 0;
1263 	unsigned long count;
1264 	unsigned long limit;
1265 
1266 	count = page_counter_read(&memcg->memory);
1267 	limit = READ_ONCE(memcg->memory.max);
1268 	if (count < limit)
1269 		margin = limit - count;
1270 
1271 	if (do_memsw_account()) {
1272 		count = page_counter_read(&memcg->memsw);
1273 		limit = READ_ONCE(memcg->memsw.max);
1274 		if (count < limit)
1275 			margin = min(margin, limit - count);
1276 		else
1277 			margin = 0;
1278 	}
1279 
1280 	return margin;
1281 }
1282 
1283 /*
1284  * A routine for checking "mem" is under move_account() or not.
1285  *
1286  * Checking a cgroup is mc.from or mc.to or under hierarchy of
1287  * moving cgroups. This is for waiting at high-memory pressure
1288  * caused by "move".
1289  */
1290 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1291 {
1292 	struct mem_cgroup *from;
1293 	struct mem_cgroup *to;
1294 	bool ret = false;
1295 	/*
1296 	 * Unlike task_move routines, we access mc.to, mc.from not under
1297 	 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1298 	 */
1299 	spin_lock(&mc.lock);
1300 	from = mc.from;
1301 	to = mc.to;
1302 	if (!from)
1303 		goto unlock;
1304 
1305 	ret = mem_cgroup_is_descendant(from, memcg) ||
1306 		mem_cgroup_is_descendant(to, memcg);
1307 unlock:
1308 	spin_unlock(&mc.lock);
1309 	return ret;
1310 }
1311 
1312 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1313 {
1314 	if (mc.moving_task && current != mc.moving_task) {
1315 		if (mem_cgroup_under_move(memcg)) {
1316 			DEFINE_WAIT(wait);
1317 			prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1318 			/* moving charge context might have finished. */
1319 			if (mc.moving_task)
1320 				schedule();
1321 			finish_wait(&mc.waitq, &wait);
1322 			return true;
1323 		}
1324 	}
1325 	return false;
1326 }
1327 
1328 struct memory_stat {
1329 	const char *name;
1330 	unsigned int idx;
1331 };
1332 
1333 static const struct memory_stat memory_stats[] = {
1334 	{ "anon",			NR_ANON_MAPPED			},
1335 	{ "file",			NR_FILE_PAGES			},
1336 	{ "kernel_stack",		NR_KERNEL_STACK_KB		},
1337 	{ "pagetables",			NR_PAGETABLE			},
1338 	{ "percpu",			MEMCG_PERCPU_B			},
1339 	{ "sock",			MEMCG_SOCK			},
1340 	{ "shmem",			NR_SHMEM			},
1341 	{ "file_mapped",		NR_FILE_MAPPED			},
1342 	{ "file_dirty",			NR_FILE_DIRTY			},
1343 	{ "file_writeback",		NR_WRITEBACK			},
1344 #ifdef CONFIG_SWAP
1345 	{ "swapcached",			NR_SWAPCACHE			},
1346 #endif
1347 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1348 	{ "anon_thp",			NR_ANON_THPS			},
1349 	{ "file_thp",			NR_FILE_THPS			},
1350 	{ "shmem_thp",			NR_SHMEM_THPS			},
1351 #endif
1352 	{ "inactive_anon",		NR_INACTIVE_ANON		},
1353 	{ "active_anon",		NR_ACTIVE_ANON			},
1354 	{ "inactive_file",		NR_INACTIVE_FILE		},
1355 	{ "active_file",		NR_ACTIVE_FILE			},
1356 	{ "unevictable",		NR_UNEVICTABLE			},
1357 	{ "slab_reclaimable",		NR_SLAB_RECLAIMABLE_B		},
1358 	{ "slab_unreclaimable",		NR_SLAB_UNRECLAIMABLE_B		},
1359 
1360 	/* The memory events */
1361 	{ "workingset_refault_anon",	WORKINGSET_REFAULT_ANON		},
1362 	{ "workingset_refault_file",	WORKINGSET_REFAULT_FILE		},
1363 	{ "workingset_activate_anon",	WORKINGSET_ACTIVATE_ANON	},
1364 	{ "workingset_activate_file",	WORKINGSET_ACTIVATE_FILE	},
1365 	{ "workingset_restore_anon",	WORKINGSET_RESTORE_ANON		},
1366 	{ "workingset_restore_file",	WORKINGSET_RESTORE_FILE		},
1367 	{ "workingset_nodereclaim",	WORKINGSET_NODERECLAIM		},
1368 };
1369 
1370 /* Translate stat items to the correct unit for memory.stat output */
1371 static int memcg_page_state_unit(int item)
1372 {
1373 	switch (item) {
1374 	case MEMCG_PERCPU_B:
1375 	case NR_SLAB_RECLAIMABLE_B:
1376 	case NR_SLAB_UNRECLAIMABLE_B:
1377 	case WORKINGSET_REFAULT_ANON:
1378 	case WORKINGSET_REFAULT_FILE:
1379 	case WORKINGSET_ACTIVATE_ANON:
1380 	case WORKINGSET_ACTIVATE_FILE:
1381 	case WORKINGSET_RESTORE_ANON:
1382 	case WORKINGSET_RESTORE_FILE:
1383 	case WORKINGSET_NODERECLAIM:
1384 		return 1;
1385 	case NR_KERNEL_STACK_KB:
1386 		return SZ_1K;
1387 	default:
1388 		return PAGE_SIZE;
1389 	}
1390 }
1391 
1392 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1393 						    int item)
1394 {
1395 	return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1396 }
1397 
1398 static char *memory_stat_format(struct mem_cgroup *memcg)
1399 {
1400 	struct seq_buf s;
1401 	int i;
1402 
1403 	seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1404 	if (!s.buffer)
1405 		return NULL;
1406 
1407 	/*
1408 	 * Provide statistics on the state of the memory subsystem as
1409 	 * well as cumulative event counters that show past behavior.
1410 	 *
1411 	 * This list is ordered following a combination of these gradients:
1412 	 * 1) generic big picture -> specifics and details
1413 	 * 2) reflecting userspace activity -> reflecting kernel heuristics
1414 	 *
1415 	 * Current memory state:
1416 	 */
1417 	cgroup_rstat_flush(memcg->css.cgroup);
1418 
1419 	for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1420 		u64 size;
1421 
1422 		size = memcg_page_state_output(memcg, memory_stats[i].idx);
1423 		seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1424 
1425 		if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1426 			size += memcg_page_state_output(memcg,
1427 							NR_SLAB_RECLAIMABLE_B);
1428 			seq_buf_printf(&s, "slab %llu\n", size);
1429 		}
1430 	}
1431 
1432 	/* Accumulated memory events */
1433 
1434 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1435 		       memcg_events(memcg, PGFAULT));
1436 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1437 		       memcg_events(memcg, PGMAJFAULT));
1438 	seq_buf_printf(&s, "%s %lu\n",  vm_event_name(PGREFILL),
1439 		       memcg_events(memcg, PGREFILL));
1440 	seq_buf_printf(&s, "pgscan %lu\n",
1441 		       memcg_events(memcg, PGSCAN_KSWAPD) +
1442 		       memcg_events(memcg, PGSCAN_DIRECT));
1443 	seq_buf_printf(&s, "pgsteal %lu\n",
1444 		       memcg_events(memcg, PGSTEAL_KSWAPD) +
1445 		       memcg_events(memcg, PGSTEAL_DIRECT));
1446 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1447 		       memcg_events(memcg, PGACTIVATE));
1448 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1449 		       memcg_events(memcg, PGDEACTIVATE));
1450 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1451 		       memcg_events(memcg, PGLAZYFREE));
1452 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1453 		       memcg_events(memcg, PGLAZYFREED));
1454 
1455 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1456 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1457 		       memcg_events(memcg, THP_FAULT_ALLOC));
1458 	seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1459 		       memcg_events(memcg, THP_COLLAPSE_ALLOC));
1460 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1461 
1462 	/* The above should easily fit into one page */
1463 	WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1464 
1465 	return s.buffer;
1466 }
1467 
1468 #define K(x) ((x) << (PAGE_SHIFT-10))
1469 /**
1470  * mem_cgroup_print_oom_context: Print OOM information relevant to
1471  * memory controller.
1472  * @memcg: The memory cgroup that went over limit
1473  * @p: Task that is going to be killed
1474  *
1475  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1476  * enabled
1477  */
1478 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1479 {
1480 	rcu_read_lock();
1481 
1482 	if (memcg) {
1483 		pr_cont(",oom_memcg=");
1484 		pr_cont_cgroup_path(memcg->css.cgroup);
1485 	} else
1486 		pr_cont(",global_oom");
1487 	if (p) {
1488 		pr_cont(",task_memcg=");
1489 		pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1490 	}
1491 	rcu_read_unlock();
1492 }
1493 
1494 /**
1495  * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1496  * memory controller.
1497  * @memcg: The memory cgroup that went over limit
1498  */
1499 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1500 {
1501 	char *buf;
1502 
1503 	pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1504 		K((u64)page_counter_read(&memcg->memory)),
1505 		K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1506 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1507 		pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1508 			K((u64)page_counter_read(&memcg->swap)),
1509 			K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1510 	else {
1511 		pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1512 			K((u64)page_counter_read(&memcg->memsw)),
1513 			K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1514 		pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1515 			K((u64)page_counter_read(&memcg->kmem)),
1516 			K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1517 	}
1518 
1519 	pr_info("Memory cgroup stats for ");
1520 	pr_cont_cgroup_path(memcg->css.cgroup);
1521 	pr_cont(":");
1522 	buf = memory_stat_format(memcg);
1523 	if (!buf)
1524 		return;
1525 	pr_info("%s", buf);
1526 	kfree(buf);
1527 }
1528 
1529 /*
1530  * Return the memory (and swap, if configured) limit for a memcg.
1531  */
1532 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1533 {
1534 	unsigned long max = READ_ONCE(memcg->memory.max);
1535 
1536 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1537 		if (mem_cgroup_swappiness(memcg))
1538 			max += min(READ_ONCE(memcg->swap.max),
1539 				   (unsigned long)total_swap_pages);
1540 	} else { /* v1 */
1541 		if (mem_cgroup_swappiness(memcg)) {
1542 			/* Calculate swap excess capacity from memsw limit */
1543 			unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1544 
1545 			max += min(swap, (unsigned long)total_swap_pages);
1546 		}
1547 	}
1548 	return max;
1549 }
1550 
1551 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1552 {
1553 	return page_counter_read(&memcg->memory);
1554 }
1555 
1556 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1557 				     int order)
1558 {
1559 	struct oom_control oc = {
1560 		.zonelist = NULL,
1561 		.nodemask = NULL,
1562 		.memcg = memcg,
1563 		.gfp_mask = gfp_mask,
1564 		.order = order,
1565 	};
1566 	bool ret = true;
1567 
1568 	if (mutex_lock_killable(&oom_lock))
1569 		return true;
1570 
1571 	if (mem_cgroup_margin(memcg) >= (1 << order))
1572 		goto unlock;
1573 
1574 	/*
1575 	 * A few threads which were not waiting at mutex_lock_killable() can
1576 	 * fail to bail out. Therefore, check again after holding oom_lock.
1577 	 */
1578 	ret = should_force_charge() || out_of_memory(&oc);
1579 
1580 unlock:
1581 	mutex_unlock(&oom_lock);
1582 	return ret;
1583 }
1584 
1585 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1586 				   pg_data_t *pgdat,
1587 				   gfp_t gfp_mask,
1588 				   unsigned long *total_scanned)
1589 {
1590 	struct mem_cgroup *victim = NULL;
1591 	int total = 0;
1592 	int loop = 0;
1593 	unsigned long excess;
1594 	unsigned long nr_scanned;
1595 	struct mem_cgroup_reclaim_cookie reclaim = {
1596 		.pgdat = pgdat,
1597 	};
1598 
1599 	excess = soft_limit_excess(root_memcg);
1600 
1601 	while (1) {
1602 		victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1603 		if (!victim) {
1604 			loop++;
1605 			if (loop >= 2) {
1606 				/*
1607 				 * If we have not been able to reclaim
1608 				 * anything, it might because there are
1609 				 * no reclaimable pages under this hierarchy
1610 				 */
1611 				if (!total)
1612 					break;
1613 				/*
1614 				 * We want to do more targeted reclaim.
1615 				 * excess >> 2 is not to excessive so as to
1616 				 * reclaim too much, nor too less that we keep
1617 				 * coming back to reclaim from this cgroup
1618 				 */
1619 				if (total >= (excess >> 2) ||
1620 					(loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1621 					break;
1622 			}
1623 			continue;
1624 		}
1625 		total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1626 					pgdat, &nr_scanned);
1627 		*total_scanned += nr_scanned;
1628 		if (!soft_limit_excess(root_memcg))
1629 			break;
1630 	}
1631 	mem_cgroup_iter_break(root_memcg, victim);
1632 	return total;
1633 }
1634 
1635 #ifdef CONFIG_LOCKDEP
1636 static struct lockdep_map memcg_oom_lock_dep_map = {
1637 	.name = "memcg_oom_lock",
1638 };
1639 #endif
1640 
1641 static DEFINE_SPINLOCK(memcg_oom_lock);
1642 
1643 /*
1644  * Check OOM-Killer is already running under our hierarchy.
1645  * If someone is running, return false.
1646  */
1647 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1648 {
1649 	struct mem_cgroup *iter, *failed = NULL;
1650 
1651 	spin_lock(&memcg_oom_lock);
1652 
1653 	for_each_mem_cgroup_tree(iter, memcg) {
1654 		if (iter->oom_lock) {
1655 			/*
1656 			 * this subtree of our hierarchy is already locked
1657 			 * so we cannot give a lock.
1658 			 */
1659 			failed = iter;
1660 			mem_cgroup_iter_break(memcg, iter);
1661 			break;
1662 		} else
1663 			iter->oom_lock = true;
1664 	}
1665 
1666 	if (failed) {
1667 		/*
1668 		 * OK, we failed to lock the whole subtree so we have
1669 		 * to clean up what we set up to the failing subtree
1670 		 */
1671 		for_each_mem_cgroup_tree(iter, memcg) {
1672 			if (iter == failed) {
1673 				mem_cgroup_iter_break(memcg, iter);
1674 				break;
1675 			}
1676 			iter->oom_lock = false;
1677 		}
1678 	} else
1679 		mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1680 
1681 	spin_unlock(&memcg_oom_lock);
1682 
1683 	return !failed;
1684 }
1685 
1686 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1687 {
1688 	struct mem_cgroup *iter;
1689 
1690 	spin_lock(&memcg_oom_lock);
1691 	mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1692 	for_each_mem_cgroup_tree(iter, memcg)
1693 		iter->oom_lock = false;
1694 	spin_unlock(&memcg_oom_lock);
1695 }
1696 
1697 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1698 {
1699 	struct mem_cgroup *iter;
1700 
1701 	spin_lock(&memcg_oom_lock);
1702 	for_each_mem_cgroup_tree(iter, memcg)
1703 		iter->under_oom++;
1704 	spin_unlock(&memcg_oom_lock);
1705 }
1706 
1707 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1708 {
1709 	struct mem_cgroup *iter;
1710 
1711 	/*
1712 	 * Be careful about under_oom underflows because a child memcg
1713 	 * could have been added after mem_cgroup_mark_under_oom.
1714 	 */
1715 	spin_lock(&memcg_oom_lock);
1716 	for_each_mem_cgroup_tree(iter, memcg)
1717 		if (iter->under_oom > 0)
1718 			iter->under_oom--;
1719 	spin_unlock(&memcg_oom_lock);
1720 }
1721 
1722 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1723 
1724 struct oom_wait_info {
1725 	struct mem_cgroup *memcg;
1726 	wait_queue_entry_t	wait;
1727 };
1728 
1729 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1730 	unsigned mode, int sync, void *arg)
1731 {
1732 	struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1733 	struct mem_cgroup *oom_wait_memcg;
1734 	struct oom_wait_info *oom_wait_info;
1735 
1736 	oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1737 	oom_wait_memcg = oom_wait_info->memcg;
1738 
1739 	if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1740 	    !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1741 		return 0;
1742 	return autoremove_wake_function(wait, mode, sync, arg);
1743 }
1744 
1745 static void memcg_oom_recover(struct mem_cgroup *memcg)
1746 {
1747 	/*
1748 	 * For the following lockless ->under_oom test, the only required
1749 	 * guarantee is that it must see the state asserted by an OOM when
1750 	 * this function is called as a result of userland actions
1751 	 * triggered by the notification of the OOM.  This is trivially
1752 	 * achieved by invoking mem_cgroup_mark_under_oom() before
1753 	 * triggering notification.
1754 	 */
1755 	if (memcg && memcg->under_oom)
1756 		__wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1757 }
1758 
1759 enum oom_status {
1760 	OOM_SUCCESS,
1761 	OOM_FAILED,
1762 	OOM_ASYNC,
1763 	OOM_SKIPPED
1764 };
1765 
1766 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1767 {
1768 	enum oom_status ret;
1769 	bool locked;
1770 
1771 	if (order > PAGE_ALLOC_COSTLY_ORDER)
1772 		return OOM_SKIPPED;
1773 
1774 	memcg_memory_event(memcg, MEMCG_OOM);
1775 
1776 	/*
1777 	 * We are in the middle of the charge context here, so we
1778 	 * don't want to block when potentially sitting on a callstack
1779 	 * that holds all kinds of filesystem and mm locks.
1780 	 *
1781 	 * cgroup1 allows disabling the OOM killer and waiting for outside
1782 	 * handling until the charge can succeed; remember the context and put
1783 	 * the task to sleep at the end of the page fault when all locks are
1784 	 * released.
1785 	 *
1786 	 * On the other hand, in-kernel OOM killer allows for an async victim
1787 	 * memory reclaim (oom_reaper) and that means that we are not solely
1788 	 * relying on the oom victim to make a forward progress and we can
1789 	 * invoke the oom killer here.
1790 	 *
1791 	 * Please note that mem_cgroup_out_of_memory might fail to find a
1792 	 * victim and then we have to bail out from the charge path.
1793 	 */
1794 	if (memcg->oom_kill_disable) {
1795 		if (!current->in_user_fault)
1796 			return OOM_SKIPPED;
1797 		css_get(&memcg->css);
1798 		current->memcg_in_oom = memcg;
1799 		current->memcg_oom_gfp_mask = mask;
1800 		current->memcg_oom_order = order;
1801 
1802 		return OOM_ASYNC;
1803 	}
1804 
1805 	mem_cgroup_mark_under_oom(memcg);
1806 
1807 	locked = mem_cgroup_oom_trylock(memcg);
1808 
1809 	if (locked)
1810 		mem_cgroup_oom_notify(memcg);
1811 
1812 	mem_cgroup_unmark_under_oom(memcg);
1813 	if (mem_cgroup_out_of_memory(memcg, mask, order))
1814 		ret = OOM_SUCCESS;
1815 	else
1816 		ret = OOM_FAILED;
1817 
1818 	if (locked)
1819 		mem_cgroup_oom_unlock(memcg);
1820 
1821 	return ret;
1822 }
1823 
1824 /**
1825  * mem_cgroup_oom_synchronize - complete memcg OOM handling
1826  * @handle: actually kill/wait or just clean up the OOM state
1827  *
1828  * This has to be called at the end of a page fault if the memcg OOM
1829  * handler was enabled.
1830  *
1831  * Memcg supports userspace OOM handling where failed allocations must
1832  * sleep on a waitqueue until the userspace task resolves the
1833  * situation.  Sleeping directly in the charge context with all kinds
1834  * of locks held is not a good idea, instead we remember an OOM state
1835  * in the task and mem_cgroup_oom_synchronize() has to be called at
1836  * the end of the page fault to complete the OOM handling.
1837  *
1838  * Returns %true if an ongoing memcg OOM situation was detected and
1839  * completed, %false otherwise.
1840  */
1841 bool mem_cgroup_oom_synchronize(bool handle)
1842 {
1843 	struct mem_cgroup *memcg = current->memcg_in_oom;
1844 	struct oom_wait_info owait;
1845 	bool locked;
1846 
1847 	/* OOM is global, do not handle */
1848 	if (!memcg)
1849 		return false;
1850 
1851 	if (!handle)
1852 		goto cleanup;
1853 
1854 	owait.memcg = memcg;
1855 	owait.wait.flags = 0;
1856 	owait.wait.func = memcg_oom_wake_function;
1857 	owait.wait.private = current;
1858 	INIT_LIST_HEAD(&owait.wait.entry);
1859 
1860 	prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1861 	mem_cgroup_mark_under_oom(memcg);
1862 
1863 	locked = mem_cgroup_oom_trylock(memcg);
1864 
1865 	if (locked)
1866 		mem_cgroup_oom_notify(memcg);
1867 
1868 	if (locked && !memcg->oom_kill_disable) {
1869 		mem_cgroup_unmark_under_oom(memcg);
1870 		finish_wait(&memcg_oom_waitq, &owait.wait);
1871 		mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1872 					 current->memcg_oom_order);
1873 	} else {
1874 		schedule();
1875 		mem_cgroup_unmark_under_oom(memcg);
1876 		finish_wait(&memcg_oom_waitq, &owait.wait);
1877 	}
1878 
1879 	if (locked) {
1880 		mem_cgroup_oom_unlock(memcg);
1881 		/*
1882 		 * There is no guarantee that an OOM-lock contender
1883 		 * sees the wakeups triggered by the OOM kill
1884 		 * uncharges.  Wake any sleepers explicitly.
1885 		 */
1886 		memcg_oom_recover(memcg);
1887 	}
1888 cleanup:
1889 	current->memcg_in_oom = NULL;
1890 	css_put(&memcg->css);
1891 	return true;
1892 }
1893 
1894 /**
1895  * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1896  * @victim: task to be killed by the OOM killer
1897  * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1898  *
1899  * Returns a pointer to a memory cgroup, which has to be cleaned up
1900  * by killing all belonging OOM-killable tasks.
1901  *
1902  * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1903  */
1904 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1905 					    struct mem_cgroup *oom_domain)
1906 {
1907 	struct mem_cgroup *oom_group = NULL;
1908 	struct mem_cgroup *memcg;
1909 
1910 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1911 		return NULL;
1912 
1913 	if (!oom_domain)
1914 		oom_domain = root_mem_cgroup;
1915 
1916 	rcu_read_lock();
1917 
1918 	memcg = mem_cgroup_from_task(victim);
1919 	if (memcg == root_mem_cgroup)
1920 		goto out;
1921 
1922 	/*
1923 	 * If the victim task has been asynchronously moved to a different
1924 	 * memory cgroup, we might end up killing tasks outside oom_domain.
1925 	 * In this case it's better to ignore memory.group.oom.
1926 	 */
1927 	if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1928 		goto out;
1929 
1930 	/*
1931 	 * Traverse the memory cgroup hierarchy from the victim task's
1932 	 * cgroup up to the OOMing cgroup (or root) to find the
1933 	 * highest-level memory cgroup with oom.group set.
1934 	 */
1935 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1936 		if (memcg->oom_group)
1937 			oom_group = memcg;
1938 
1939 		if (memcg == oom_domain)
1940 			break;
1941 	}
1942 
1943 	if (oom_group)
1944 		css_get(&oom_group->css);
1945 out:
1946 	rcu_read_unlock();
1947 
1948 	return oom_group;
1949 }
1950 
1951 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1952 {
1953 	pr_info("Tasks in ");
1954 	pr_cont_cgroup_path(memcg->css.cgroup);
1955 	pr_cont(" are going to be killed due to memory.oom.group set\n");
1956 }
1957 
1958 /**
1959  * lock_page_memcg - lock a page and memcg binding
1960  * @page: the page
1961  *
1962  * This function protects unlocked LRU pages from being moved to
1963  * another cgroup.
1964  *
1965  * It ensures lifetime of the locked memcg. Caller is responsible
1966  * for the lifetime of the page.
1967  */
1968 void lock_page_memcg(struct page *page)
1969 {
1970 	struct page *head = compound_head(page); /* rmap on tail pages */
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 	rcu_read_lock();
1980 
1981 	if (mem_cgroup_disabled())
1982 		return;
1983 again:
1984 	memcg = page_memcg(head);
1985 	if (unlikely(!memcg))
1986 		return;
1987 
1988 #ifdef CONFIG_PROVE_LOCKING
1989 	local_irq_save(flags);
1990 	might_lock(&memcg->move_lock);
1991 	local_irq_restore(flags);
1992 #endif
1993 
1994 	if (atomic_read(&memcg->moving_account) <= 0)
1995 		return;
1996 
1997 	spin_lock_irqsave(&memcg->move_lock, flags);
1998 	if (memcg != page_memcg(head)) {
1999 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2000 		goto again;
2001 	}
2002 
2003 	/*
2004 	 * When charge migration first begins, we can have multiple
2005 	 * critical sections holding the fast-path RCU lock and one
2006 	 * holding the slowpath move_lock. Track the task who has the
2007 	 * move_lock for unlock_page_memcg().
2008 	 */
2009 	memcg->move_lock_task = current;
2010 	memcg->move_lock_flags = flags;
2011 }
2012 EXPORT_SYMBOL(lock_page_memcg);
2013 
2014 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2015 {
2016 	if (memcg && memcg->move_lock_task == current) {
2017 		unsigned long flags = memcg->move_lock_flags;
2018 
2019 		memcg->move_lock_task = NULL;
2020 		memcg->move_lock_flags = 0;
2021 
2022 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2023 	}
2024 
2025 	rcu_read_unlock();
2026 }
2027 
2028 /**
2029  * unlock_page_memcg - unlock a page and memcg binding
2030  * @page: the page
2031  */
2032 void unlock_page_memcg(struct page *page)
2033 {
2034 	struct page *head = compound_head(page);
2035 
2036 	__unlock_page_memcg(page_memcg(head));
2037 }
2038 EXPORT_SYMBOL(unlock_page_memcg);
2039 
2040 struct obj_stock {
2041 #ifdef CONFIG_MEMCG_KMEM
2042 	struct obj_cgroup *cached_objcg;
2043 	struct pglist_data *cached_pgdat;
2044 	unsigned int nr_bytes;
2045 	int nr_slab_reclaimable_b;
2046 	int nr_slab_unreclaimable_b;
2047 #else
2048 	int dummy[0];
2049 #endif
2050 };
2051 
2052 struct memcg_stock_pcp {
2053 	struct mem_cgroup *cached; /* this never be root cgroup */
2054 	unsigned int nr_pages;
2055 	struct obj_stock task_obj;
2056 	struct obj_stock irq_obj;
2057 
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 #ifdef CONFIG_MEMCG_KMEM
2066 static void drain_obj_stock(struct obj_stock *stock);
2067 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2068 				     struct mem_cgroup *root_memcg);
2069 
2070 #else
2071 static inline void drain_obj_stock(struct obj_stock *stock)
2072 {
2073 }
2074 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2075 				     struct mem_cgroup *root_memcg)
2076 {
2077 	return false;
2078 }
2079 #endif
2080 
2081 /*
2082  * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2083  * sequence used in this case to access content from object stock is slow.
2084  * To optimize for user context access, there are now two object stocks for
2085  * task context and interrupt context access respectively.
2086  *
2087  * The task context object stock can be accessed by disabling preemption only
2088  * which is cheap in non-preempt kernel. The interrupt context object stock
2089  * can only be accessed after disabling interrupt. User context code can
2090  * access interrupt object stock, but not vice versa.
2091  */
2092 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2093 {
2094 	struct memcg_stock_pcp *stock;
2095 
2096 	if (likely(in_task())) {
2097 		*pflags = 0UL;
2098 		preempt_disable();
2099 		stock = this_cpu_ptr(&memcg_stock);
2100 		return &stock->task_obj;
2101 	}
2102 
2103 	local_irq_save(*pflags);
2104 	stock = this_cpu_ptr(&memcg_stock);
2105 	return &stock->irq_obj;
2106 }
2107 
2108 static inline void put_obj_stock(unsigned long flags)
2109 {
2110 	if (likely(in_task()))
2111 		preempt_enable();
2112 	else
2113 		local_irq_restore(flags);
2114 }
2115 
2116 /**
2117  * consume_stock: Try to consume stocked charge on this cpu.
2118  * @memcg: memcg to consume from.
2119  * @nr_pages: how many pages to charge.
2120  *
2121  * The charges will only happen if @memcg matches the current cpu's memcg
2122  * stock, and at least @nr_pages are available in that stock.  Failure to
2123  * service an allocation will refill the stock.
2124  *
2125  * returns true if successful, false otherwise.
2126  */
2127 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2128 {
2129 	struct memcg_stock_pcp *stock;
2130 	unsigned long flags;
2131 	bool ret = false;
2132 
2133 	if (nr_pages > MEMCG_CHARGE_BATCH)
2134 		return ret;
2135 
2136 	local_irq_save(flags);
2137 
2138 	stock = this_cpu_ptr(&memcg_stock);
2139 	if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2140 		stock->nr_pages -= nr_pages;
2141 		ret = true;
2142 	}
2143 
2144 	local_irq_restore(flags);
2145 
2146 	return ret;
2147 }
2148 
2149 /*
2150  * Returns stocks cached in percpu and reset cached information.
2151  */
2152 static void drain_stock(struct memcg_stock_pcp *stock)
2153 {
2154 	struct mem_cgroup *old = stock->cached;
2155 
2156 	if (!old)
2157 		return;
2158 
2159 	if (stock->nr_pages) {
2160 		page_counter_uncharge(&old->memory, stock->nr_pages);
2161 		if (do_memsw_account())
2162 			page_counter_uncharge(&old->memsw, stock->nr_pages);
2163 		stock->nr_pages = 0;
2164 	}
2165 
2166 	css_put(&old->css);
2167 	stock->cached = NULL;
2168 }
2169 
2170 static void drain_local_stock(struct work_struct *dummy)
2171 {
2172 	struct memcg_stock_pcp *stock;
2173 	unsigned long flags;
2174 
2175 	/*
2176 	 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2177 	 * drain_stock races is that we always operate on local CPU stock
2178 	 * here with IRQ disabled
2179 	 */
2180 	local_irq_save(flags);
2181 
2182 	stock = this_cpu_ptr(&memcg_stock);
2183 	drain_obj_stock(&stock->irq_obj);
2184 	if (in_task())
2185 		drain_obj_stock(&stock->task_obj);
2186 	drain_stock(stock);
2187 	clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2188 
2189 	local_irq_restore(flags);
2190 }
2191 
2192 /*
2193  * Cache charges(val) to local per_cpu area.
2194  * This will be consumed by consume_stock() function, later.
2195  */
2196 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2197 {
2198 	struct memcg_stock_pcp *stock;
2199 	unsigned long flags;
2200 
2201 	local_irq_save(flags);
2202 
2203 	stock = this_cpu_ptr(&memcg_stock);
2204 	if (stock->cached != memcg) { /* reset if necessary */
2205 		drain_stock(stock);
2206 		css_get(&memcg->css);
2207 		stock->cached = memcg;
2208 	}
2209 	stock->nr_pages += nr_pages;
2210 
2211 	if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2212 		drain_stock(stock);
2213 
2214 	local_irq_restore(flags);
2215 }
2216 
2217 /*
2218  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2219  * of the hierarchy under it.
2220  */
2221 static void drain_all_stock(struct mem_cgroup *root_memcg)
2222 {
2223 	int cpu, curcpu;
2224 
2225 	/* If someone's already draining, avoid adding running more workers. */
2226 	if (!mutex_trylock(&percpu_charge_mutex))
2227 		return;
2228 	/*
2229 	 * Notify other cpus that system-wide "drain" is running
2230 	 * We do not care about races with the cpu hotplug because cpu down
2231 	 * as well as workers from this path always operate on the local
2232 	 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2233 	 */
2234 	curcpu = get_cpu();
2235 	for_each_online_cpu(cpu) {
2236 		struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2237 		struct mem_cgroup *memcg;
2238 		bool flush = false;
2239 
2240 		rcu_read_lock();
2241 		memcg = stock->cached;
2242 		if (memcg && stock->nr_pages &&
2243 		    mem_cgroup_is_descendant(memcg, root_memcg))
2244 			flush = true;
2245 		else if (obj_stock_flush_required(stock, root_memcg))
2246 			flush = true;
2247 		rcu_read_unlock();
2248 
2249 		if (flush &&
2250 		    !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2251 			if (cpu == curcpu)
2252 				drain_local_stock(&stock->work);
2253 			else
2254 				schedule_work_on(cpu, &stock->work);
2255 		}
2256 	}
2257 	put_cpu();
2258 	mutex_unlock(&percpu_charge_mutex);
2259 }
2260 
2261 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2262 {
2263 	struct memcg_stock_pcp *stock;
2264 
2265 	stock = &per_cpu(memcg_stock, cpu);
2266 	drain_stock(stock);
2267 
2268 	return 0;
2269 }
2270 
2271 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2272 				  unsigned int nr_pages,
2273 				  gfp_t gfp_mask)
2274 {
2275 	unsigned long nr_reclaimed = 0;
2276 
2277 	do {
2278 		unsigned long pflags;
2279 
2280 		if (page_counter_read(&memcg->memory) <=
2281 		    READ_ONCE(memcg->memory.high))
2282 			continue;
2283 
2284 		memcg_memory_event(memcg, MEMCG_HIGH);
2285 
2286 		psi_memstall_enter(&pflags);
2287 		nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2288 							     gfp_mask, true);
2289 		psi_memstall_leave(&pflags);
2290 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2291 		 !mem_cgroup_is_root(memcg));
2292 
2293 	return nr_reclaimed;
2294 }
2295 
2296 static void high_work_func(struct work_struct *work)
2297 {
2298 	struct mem_cgroup *memcg;
2299 
2300 	memcg = container_of(work, struct mem_cgroup, high_work);
2301 	reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2302 }
2303 
2304 /*
2305  * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2306  * enough to still cause a significant slowdown in most cases, while still
2307  * allowing diagnostics and tracing to proceed without becoming stuck.
2308  */
2309 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2310 
2311 /*
2312  * When calculating the delay, we use these either side of the exponentiation to
2313  * maintain precision and scale to a reasonable number of jiffies (see the table
2314  * below.
2315  *
2316  * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2317  *   overage ratio to a delay.
2318  * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2319  *   proposed penalty in order to reduce to a reasonable number of jiffies, and
2320  *   to produce a reasonable delay curve.
2321  *
2322  * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2323  * reasonable delay curve compared to precision-adjusted overage, not
2324  * penalising heavily at first, but still making sure that growth beyond the
2325  * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2326  * example, with a high of 100 megabytes:
2327  *
2328  *  +-------+------------------------+
2329  *  | usage | time to allocate in ms |
2330  *  +-------+------------------------+
2331  *  | 100M  |                      0 |
2332  *  | 101M  |                      6 |
2333  *  | 102M  |                     25 |
2334  *  | 103M  |                     57 |
2335  *  | 104M  |                    102 |
2336  *  | 105M  |                    159 |
2337  *  | 106M  |                    230 |
2338  *  | 107M  |                    313 |
2339  *  | 108M  |                    409 |
2340  *  | 109M  |                    518 |
2341  *  | 110M  |                    639 |
2342  *  | 111M  |                    774 |
2343  *  | 112M  |                    921 |
2344  *  | 113M  |                   1081 |
2345  *  | 114M  |                   1254 |
2346  *  | 115M  |                   1439 |
2347  *  | 116M  |                   1638 |
2348  *  | 117M  |                   1849 |
2349  *  | 118M  |                   2000 |
2350  *  | 119M  |                   2000 |
2351  *  | 120M  |                   2000 |
2352  *  +-------+------------------------+
2353  */
2354  #define MEMCG_DELAY_PRECISION_SHIFT 20
2355  #define MEMCG_DELAY_SCALING_SHIFT 14
2356 
2357 static u64 calculate_overage(unsigned long usage, unsigned long high)
2358 {
2359 	u64 overage;
2360 
2361 	if (usage <= high)
2362 		return 0;
2363 
2364 	/*
2365 	 * Prevent division by 0 in overage calculation by acting as if
2366 	 * it was a threshold of 1 page
2367 	 */
2368 	high = max(high, 1UL);
2369 
2370 	overage = usage - high;
2371 	overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2372 	return div64_u64(overage, high);
2373 }
2374 
2375 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2376 {
2377 	u64 overage, max_overage = 0;
2378 
2379 	do {
2380 		overage = calculate_overage(page_counter_read(&memcg->memory),
2381 					    READ_ONCE(memcg->memory.high));
2382 		max_overage = max(overage, max_overage);
2383 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2384 		 !mem_cgroup_is_root(memcg));
2385 
2386 	return max_overage;
2387 }
2388 
2389 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2390 {
2391 	u64 overage, max_overage = 0;
2392 
2393 	do {
2394 		overage = calculate_overage(page_counter_read(&memcg->swap),
2395 					    READ_ONCE(memcg->swap.high));
2396 		if (overage)
2397 			memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2398 		max_overage = max(overage, max_overage);
2399 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2400 		 !mem_cgroup_is_root(memcg));
2401 
2402 	return max_overage;
2403 }
2404 
2405 /*
2406  * Get the number of jiffies that we should penalise a mischievous cgroup which
2407  * is exceeding its memory.high by checking both it and its ancestors.
2408  */
2409 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2410 					  unsigned int nr_pages,
2411 					  u64 max_overage)
2412 {
2413 	unsigned long penalty_jiffies;
2414 
2415 	if (!max_overage)
2416 		return 0;
2417 
2418 	/*
2419 	 * We use overage compared to memory.high to calculate the number of
2420 	 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2421 	 * fairly lenient on small overages, and increasingly harsh when the
2422 	 * memcg in question makes it clear that it has no intention of stopping
2423 	 * its crazy behaviour, so we exponentially increase the delay based on
2424 	 * overage amount.
2425 	 */
2426 	penalty_jiffies = max_overage * max_overage * HZ;
2427 	penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2428 	penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2429 
2430 	/*
2431 	 * Factor in the task's own contribution to the overage, such that four
2432 	 * N-sized allocations are throttled approximately the same as one
2433 	 * 4N-sized allocation.
2434 	 *
2435 	 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2436 	 * larger the current charge patch is than that.
2437 	 */
2438 	return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2439 }
2440 
2441 /*
2442  * Scheduled by try_charge() to be executed from the userland return path
2443  * and reclaims memory over the high limit.
2444  */
2445 void mem_cgroup_handle_over_high(void)
2446 {
2447 	unsigned long penalty_jiffies;
2448 	unsigned long pflags;
2449 	unsigned long nr_reclaimed;
2450 	unsigned int nr_pages = current->memcg_nr_pages_over_high;
2451 	int nr_retries = MAX_RECLAIM_RETRIES;
2452 	struct mem_cgroup *memcg;
2453 	bool in_retry = false;
2454 
2455 	if (likely(!nr_pages))
2456 		return;
2457 
2458 	memcg = get_mem_cgroup_from_mm(current->mm);
2459 	current->memcg_nr_pages_over_high = 0;
2460 
2461 retry_reclaim:
2462 	/*
2463 	 * The allocating task should reclaim at least the batch size, but for
2464 	 * subsequent retries we only want to do what's necessary to prevent oom
2465 	 * or breaching resource isolation.
2466 	 *
2467 	 * This is distinct from memory.max or page allocator behaviour because
2468 	 * memory.high is currently batched, whereas memory.max and the page
2469 	 * allocator run every time an allocation is made.
2470 	 */
2471 	nr_reclaimed = reclaim_high(memcg,
2472 				    in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2473 				    GFP_KERNEL);
2474 
2475 	/*
2476 	 * memory.high is breached and reclaim is unable to keep up. Throttle
2477 	 * allocators proactively to slow down excessive growth.
2478 	 */
2479 	penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2480 					       mem_find_max_overage(memcg));
2481 
2482 	penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2483 						swap_find_max_overage(memcg));
2484 
2485 	/*
2486 	 * Clamp the max delay per usermode return so as to still keep the
2487 	 * application moving forwards and also permit diagnostics, albeit
2488 	 * extremely slowly.
2489 	 */
2490 	penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2491 
2492 	/*
2493 	 * Don't sleep if the amount of jiffies this memcg owes us is so low
2494 	 * that it's not even worth doing, in an attempt to be nice to those who
2495 	 * go only a small amount over their memory.high value and maybe haven't
2496 	 * been aggressively reclaimed enough yet.
2497 	 */
2498 	if (penalty_jiffies <= HZ / 100)
2499 		goto out;
2500 
2501 	/*
2502 	 * If reclaim is making forward progress but we're still over
2503 	 * memory.high, we want to encourage that rather than doing allocator
2504 	 * throttling.
2505 	 */
2506 	if (nr_reclaimed || nr_retries--) {
2507 		in_retry = true;
2508 		goto retry_reclaim;
2509 	}
2510 
2511 	/*
2512 	 * If we exit early, we're guaranteed to die (since
2513 	 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2514 	 * need to account for any ill-begotten jiffies to pay them off later.
2515 	 */
2516 	psi_memstall_enter(&pflags);
2517 	schedule_timeout_killable(penalty_jiffies);
2518 	psi_memstall_leave(&pflags);
2519 
2520 out:
2521 	css_put(&memcg->css);
2522 }
2523 
2524 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2525 			unsigned int nr_pages)
2526 {
2527 	unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2528 	int nr_retries = MAX_RECLAIM_RETRIES;
2529 	struct mem_cgroup *mem_over_limit;
2530 	struct page_counter *counter;
2531 	enum oom_status oom_status;
2532 	unsigned long nr_reclaimed;
2533 	bool may_swap = true;
2534 	bool drained = false;
2535 	unsigned long pflags;
2536 
2537 retry:
2538 	if (consume_stock(memcg, nr_pages))
2539 		return 0;
2540 
2541 	if (!do_memsw_account() ||
2542 	    page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2543 		if (page_counter_try_charge(&memcg->memory, batch, &counter))
2544 			goto done_restock;
2545 		if (do_memsw_account())
2546 			page_counter_uncharge(&memcg->memsw, batch);
2547 		mem_over_limit = mem_cgroup_from_counter(counter, memory);
2548 	} else {
2549 		mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2550 		may_swap = false;
2551 	}
2552 
2553 	if (batch > nr_pages) {
2554 		batch = nr_pages;
2555 		goto retry;
2556 	}
2557 
2558 	/*
2559 	 * Memcg doesn't have a dedicated reserve for atomic
2560 	 * allocations. But like the global atomic pool, we need to
2561 	 * put the burden of reclaim on regular allocation requests
2562 	 * and let these go through as privileged allocations.
2563 	 */
2564 	if (gfp_mask & __GFP_ATOMIC)
2565 		goto force;
2566 
2567 	/*
2568 	 * Unlike in global OOM situations, memcg is not in a physical
2569 	 * memory shortage.  Allow dying and OOM-killed tasks to
2570 	 * bypass the last charges so that they can exit quickly and
2571 	 * free their memory.
2572 	 */
2573 	if (unlikely(should_force_charge()))
2574 		goto force;
2575 
2576 	/*
2577 	 * Prevent unbounded recursion when reclaim operations need to
2578 	 * allocate memory. This might exceed the limits temporarily,
2579 	 * but we prefer facilitating memory reclaim and getting back
2580 	 * under the limit over triggering OOM kills in these cases.
2581 	 */
2582 	if (unlikely(current->flags & PF_MEMALLOC))
2583 		goto force;
2584 
2585 	if (unlikely(task_in_memcg_oom(current)))
2586 		goto nomem;
2587 
2588 	if (!gfpflags_allow_blocking(gfp_mask))
2589 		goto nomem;
2590 
2591 	memcg_memory_event(mem_over_limit, MEMCG_MAX);
2592 
2593 	psi_memstall_enter(&pflags);
2594 	nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2595 						    gfp_mask, may_swap);
2596 	psi_memstall_leave(&pflags);
2597 
2598 	if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2599 		goto retry;
2600 
2601 	if (!drained) {
2602 		drain_all_stock(mem_over_limit);
2603 		drained = true;
2604 		goto retry;
2605 	}
2606 
2607 	if (gfp_mask & __GFP_NORETRY)
2608 		goto nomem;
2609 	/*
2610 	 * Even though the limit is exceeded at this point, reclaim
2611 	 * may have been able to free some pages.  Retry the charge
2612 	 * before killing the task.
2613 	 *
2614 	 * Only for regular pages, though: huge pages are rather
2615 	 * unlikely to succeed so close to the limit, and we fall back
2616 	 * to regular pages anyway in case of failure.
2617 	 */
2618 	if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2619 		goto retry;
2620 	/*
2621 	 * At task move, charge accounts can be doubly counted. So, it's
2622 	 * better to wait until the end of task_move if something is going on.
2623 	 */
2624 	if (mem_cgroup_wait_acct_move(mem_over_limit))
2625 		goto retry;
2626 
2627 	if (nr_retries--)
2628 		goto retry;
2629 
2630 	if (gfp_mask & __GFP_RETRY_MAYFAIL)
2631 		goto nomem;
2632 
2633 	if (fatal_signal_pending(current))
2634 		goto force;
2635 
2636 	/*
2637 	 * keep retrying as long as the memcg oom killer is able to make
2638 	 * a forward progress or bypass the charge if the oom killer
2639 	 * couldn't make any progress.
2640 	 */
2641 	oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2642 		       get_order(nr_pages * PAGE_SIZE));
2643 	switch (oom_status) {
2644 	case OOM_SUCCESS:
2645 		nr_retries = MAX_RECLAIM_RETRIES;
2646 		goto retry;
2647 	case OOM_FAILED:
2648 		goto force;
2649 	default:
2650 		goto nomem;
2651 	}
2652 nomem:
2653 	if (!(gfp_mask & __GFP_NOFAIL))
2654 		return -ENOMEM;
2655 force:
2656 	/*
2657 	 * The allocation either can't fail or will lead to more memory
2658 	 * being freed very soon.  Allow memory usage go over the limit
2659 	 * temporarily by force charging it.
2660 	 */
2661 	page_counter_charge(&memcg->memory, nr_pages);
2662 	if (do_memsw_account())
2663 		page_counter_charge(&memcg->memsw, nr_pages);
2664 
2665 	return 0;
2666 
2667 done_restock:
2668 	if (batch > nr_pages)
2669 		refill_stock(memcg, batch - nr_pages);
2670 
2671 	/*
2672 	 * If the hierarchy is above the normal consumption range, schedule
2673 	 * reclaim on returning to userland.  We can perform reclaim here
2674 	 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2675 	 * GFP_KERNEL can consistently be used during reclaim.  @memcg is
2676 	 * not recorded as it most likely matches current's and won't
2677 	 * change in the meantime.  As high limit is checked again before
2678 	 * reclaim, the cost of mismatch is negligible.
2679 	 */
2680 	do {
2681 		bool mem_high, swap_high;
2682 
2683 		mem_high = page_counter_read(&memcg->memory) >
2684 			READ_ONCE(memcg->memory.high);
2685 		swap_high = page_counter_read(&memcg->swap) >
2686 			READ_ONCE(memcg->swap.high);
2687 
2688 		/* Don't bother a random interrupted task */
2689 		if (in_interrupt()) {
2690 			if (mem_high) {
2691 				schedule_work(&memcg->high_work);
2692 				break;
2693 			}
2694 			continue;
2695 		}
2696 
2697 		if (mem_high || swap_high) {
2698 			/*
2699 			 * The allocating tasks in this cgroup will need to do
2700 			 * reclaim or be throttled to prevent further growth
2701 			 * of the memory or swap footprints.
2702 			 *
2703 			 * Target some best-effort fairness between the tasks,
2704 			 * and distribute reclaim work and delay penalties
2705 			 * based on how much each task is actually allocating.
2706 			 */
2707 			current->memcg_nr_pages_over_high += batch;
2708 			set_notify_resume(current);
2709 			break;
2710 		}
2711 	} while ((memcg = parent_mem_cgroup(memcg)));
2712 
2713 	return 0;
2714 }
2715 
2716 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2717 			     unsigned int nr_pages)
2718 {
2719 	if (mem_cgroup_is_root(memcg))
2720 		return 0;
2721 
2722 	return try_charge_memcg(memcg, gfp_mask, nr_pages);
2723 }
2724 
2725 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2726 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2727 {
2728 	if (mem_cgroup_is_root(memcg))
2729 		return;
2730 
2731 	page_counter_uncharge(&memcg->memory, nr_pages);
2732 	if (do_memsw_account())
2733 		page_counter_uncharge(&memcg->memsw, nr_pages);
2734 }
2735 #endif
2736 
2737 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2738 {
2739 	VM_BUG_ON_PAGE(page_memcg(page), page);
2740 	/*
2741 	 * Any of the following ensures page's memcg stability:
2742 	 *
2743 	 * - the page lock
2744 	 * - LRU isolation
2745 	 * - lock_page_memcg()
2746 	 * - exclusive reference
2747 	 */
2748 	page->memcg_data = (unsigned long)memcg;
2749 }
2750 
2751 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2752 {
2753 	struct mem_cgroup *memcg;
2754 
2755 	rcu_read_lock();
2756 retry:
2757 	memcg = obj_cgroup_memcg(objcg);
2758 	if (unlikely(!css_tryget(&memcg->css)))
2759 		goto retry;
2760 	rcu_read_unlock();
2761 
2762 	return memcg;
2763 }
2764 
2765 #ifdef CONFIG_MEMCG_KMEM
2766 /*
2767  * The allocated objcg pointers array is not accounted directly.
2768  * Moreover, it should not come from DMA buffer and is not readily
2769  * reclaimable. So those GFP bits should be masked off.
2770  */
2771 #define OBJCGS_CLEAR_MASK	(__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2772 
2773 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2774 				 gfp_t gfp, bool new_page)
2775 {
2776 	unsigned int objects = objs_per_slab_page(s, page);
2777 	unsigned long memcg_data;
2778 	void *vec;
2779 
2780 	gfp &= ~OBJCGS_CLEAR_MASK;
2781 	vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2782 			   page_to_nid(page));
2783 	if (!vec)
2784 		return -ENOMEM;
2785 
2786 	memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2787 	if (new_page) {
2788 		/*
2789 		 * If the slab page is brand new and nobody can yet access
2790 		 * it's memcg_data, no synchronization is required and
2791 		 * memcg_data can be simply assigned.
2792 		 */
2793 		page->memcg_data = memcg_data;
2794 	} else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2795 		/*
2796 		 * If the slab page is already in use, somebody can allocate
2797 		 * and assign obj_cgroups in parallel. In this case the existing
2798 		 * objcg vector should be reused.
2799 		 */
2800 		kfree(vec);
2801 		return 0;
2802 	}
2803 
2804 	kmemleak_not_leak(vec);
2805 	return 0;
2806 }
2807 
2808 /*
2809  * Returns a pointer to the memory cgroup to which the kernel object is charged.
2810  *
2811  * A passed kernel object can be a slab object or a generic kernel page, so
2812  * different mechanisms for getting the memory cgroup pointer should be used.
2813  * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2814  * can not know for sure how the kernel object is implemented.
2815  * mem_cgroup_from_obj() can be safely used in such cases.
2816  *
2817  * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2818  * cgroup_mutex, etc.
2819  */
2820 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2821 {
2822 	struct page *page;
2823 
2824 	if (mem_cgroup_disabled())
2825 		return NULL;
2826 
2827 	page = virt_to_head_page(p);
2828 
2829 	/*
2830 	 * Slab objects are accounted individually, not per-page.
2831 	 * Memcg membership data for each individual object is saved in
2832 	 * the page->obj_cgroups.
2833 	 */
2834 	if (page_objcgs_check(page)) {
2835 		struct obj_cgroup *objcg;
2836 		unsigned int off;
2837 
2838 		off = obj_to_index(page->slab_cache, page, p);
2839 		objcg = page_objcgs(page)[off];
2840 		if (objcg)
2841 			return obj_cgroup_memcg(objcg);
2842 
2843 		return NULL;
2844 	}
2845 
2846 	/*
2847 	 * page_memcg_check() is used here, because page_has_obj_cgroups()
2848 	 * check above could fail because the object cgroups vector wasn't set
2849 	 * at that moment, but it can be set concurrently.
2850 	 * page_memcg_check(page) will guarantee that a proper memory
2851 	 * cgroup pointer or NULL will be returned.
2852 	 */
2853 	return page_memcg_check(page);
2854 }
2855 
2856 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2857 {
2858 	struct obj_cgroup *objcg = NULL;
2859 	struct mem_cgroup *memcg;
2860 
2861 	if (memcg_kmem_bypass())
2862 		return NULL;
2863 
2864 	rcu_read_lock();
2865 	if (unlikely(active_memcg()))
2866 		memcg = active_memcg();
2867 	else
2868 		memcg = mem_cgroup_from_task(current);
2869 
2870 	for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2871 		objcg = rcu_dereference(memcg->objcg);
2872 		if (objcg && obj_cgroup_tryget(objcg))
2873 			break;
2874 		objcg = NULL;
2875 	}
2876 	rcu_read_unlock();
2877 
2878 	return objcg;
2879 }
2880 
2881 static int memcg_alloc_cache_id(void)
2882 {
2883 	int id, size;
2884 	int err;
2885 
2886 	id = ida_simple_get(&memcg_cache_ida,
2887 			    0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2888 	if (id < 0)
2889 		return id;
2890 
2891 	if (id < memcg_nr_cache_ids)
2892 		return id;
2893 
2894 	/*
2895 	 * There's no space for the new id in memcg_caches arrays,
2896 	 * so we have to grow them.
2897 	 */
2898 	down_write(&memcg_cache_ids_sem);
2899 
2900 	size = 2 * (id + 1);
2901 	if (size < MEMCG_CACHES_MIN_SIZE)
2902 		size = MEMCG_CACHES_MIN_SIZE;
2903 	else if (size > MEMCG_CACHES_MAX_SIZE)
2904 		size = MEMCG_CACHES_MAX_SIZE;
2905 
2906 	err = memcg_update_all_list_lrus(size);
2907 	if (!err)
2908 		memcg_nr_cache_ids = size;
2909 
2910 	up_write(&memcg_cache_ids_sem);
2911 
2912 	if (err) {
2913 		ida_simple_remove(&memcg_cache_ida, id);
2914 		return err;
2915 	}
2916 	return id;
2917 }
2918 
2919 static void memcg_free_cache_id(int id)
2920 {
2921 	ida_simple_remove(&memcg_cache_ida, id);
2922 }
2923 
2924 /*
2925  * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2926  * @objcg: object cgroup to uncharge
2927  * @nr_pages: number of pages to uncharge
2928  */
2929 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2930 				      unsigned int nr_pages)
2931 {
2932 	struct mem_cgroup *memcg;
2933 
2934 	memcg = get_mem_cgroup_from_objcg(objcg);
2935 
2936 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2937 		page_counter_uncharge(&memcg->kmem, nr_pages);
2938 	refill_stock(memcg, nr_pages);
2939 
2940 	css_put(&memcg->css);
2941 }
2942 
2943 /*
2944  * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2945  * @objcg: object cgroup to charge
2946  * @gfp: reclaim mode
2947  * @nr_pages: number of pages to charge
2948  *
2949  * Returns 0 on success, an error code on failure.
2950  */
2951 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2952 				   unsigned int nr_pages)
2953 {
2954 	struct page_counter *counter;
2955 	struct mem_cgroup *memcg;
2956 	int ret;
2957 
2958 	memcg = get_mem_cgroup_from_objcg(objcg);
2959 
2960 	ret = try_charge_memcg(memcg, gfp, nr_pages);
2961 	if (ret)
2962 		goto out;
2963 
2964 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2965 	    !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2966 
2967 		/*
2968 		 * Enforce __GFP_NOFAIL allocation because callers are not
2969 		 * prepared to see failures and likely do not have any failure
2970 		 * handling code.
2971 		 */
2972 		if (gfp & __GFP_NOFAIL) {
2973 			page_counter_charge(&memcg->kmem, nr_pages);
2974 			goto out;
2975 		}
2976 		cancel_charge(memcg, nr_pages);
2977 		ret = -ENOMEM;
2978 	}
2979 out:
2980 	css_put(&memcg->css);
2981 
2982 	return ret;
2983 }
2984 
2985 /**
2986  * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2987  * @page: page to charge
2988  * @gfp: reclaim mode
2989  * @order: allocation order
2990  *
2991  * Returns 0 on success, an error code on failure.
2992  */
2993 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
2994 {
2995 	struct obj_cgroup *objcg;
2996 	int ret = 0;
2997 
2998 	objcg = get_obj_cgroup_from_current();
2999 	if (objcg) {
3000 		ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3001 		if (!ret) {
3002 			page->memcg_data = (unsigned long)objcg |
3003 				MEMCG_DATA_KMEM;
3004 			return 0;
3005 		}
3006 		obj_cgroup_put(objcg);
3007 	}
3008 	return ret;
3009 }
3010 
3011 /**
3012  * __memcg_kmem_uncharge_page: uncharge a kmem page
3013  * @page: page to uncharge
3014  * @order: allocation order
3015  */
3016 void __memcg_kmem_uncharge_page(struct page *page, int order)
3017 {
3018 	struct obj_cgroup *objcg;
3019 	unsigned int nr_pages = 1 << order;
3020 
3021 	if (!PageMemcgKmem(page))
3022 		return;
3023 
3024 	objcg = __page_objcg(page);
3025 	obj_cgroup_uncharge_pages(objcg, nr_pages);
3026 	page->memcg_data = 0;
3027 	obj_cgroup_put(objcg);
3028 }
3029 
3030 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3031 		     enum node_stat_item idx, int nr)
3032 {
3033 	unsigned long flags;
3034 	struct obj_stock *stock = get_obj_stock(&flags);
3035 	int *bytes;
3036 
3037 	/*
3038 	 * Save vmstat data in stock and skip vmstat array update unless
3039 	 * accumulating over a page of vmstat data or when pgdat or idx
3040 	 * changes.
3041 	 */
3042 	if (stock->cached_objcg != objcg) {
3043 		drain_obj_stock(stock);
3044 		obj_cgroup_get(objcg);
3045 		stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3046 				? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3047 		stock->cached_objcg = objcg;
3048 		stock->cached_pgdat = pgdat;
3049 	} else if (stock->cached_pgdat != pgdat) {
3050 		/* Flush the existing cached vmstat data */
3051 		struct pglist_data *oldpg = stock->cached_pgdat;
3052 
3053 		if (stock->nr_slab_reclaimable_b) {
3054 			mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3055 					  stock->nr_slab_reclaimable_b);
3056 			stock->nr_slab_reclaimable_b = 0;
3057 		}
3058 		if (stock->nr_slab_unreclaimable_b) {
3059 			mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3060 					  stock->nr_slab_unreclaimable_b);
3061 			stock->nr_slab_unreclaimable_b = 0;
3062 		}
3063 		stock->cached_pgdat = pgdat;
3064 	}
3065 
3066 	bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3067 					       : &stock->nr_slab_unreclaimable_b;
3068 	/*
3069 	 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3070 	 * cached locally at least once before pushing it out.
3071 	 */
3072 	if (!*bytes) {
3073 		*bytes = nr;
3074 		nr = 0;
3075 	} else {
3076 		*bytes += nr;
3077 		if (abs(*bytes) > PAGE_SIZE) {
3078 			nr = *bytes;
3079 			*bytes = 0;
3080 		} else {
3081 			nr = 0;
3082 		}
3083 	}
3084 	if (nr)
3085 		mod_objcg_mlstate(objcg, pgdat, idx, nr);
3086 
3087 	put_obj_stock(flags);
3088 }
3089 
3090 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3091 {
3092 	unsigned long flags;
3093 	struct obj_stock *stock = get_obj_stock(&flags);
3094 	bool ret = false;
3095 
3096 	if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3097 		stock->nr_bytes -= nr_bytes;
3098 		ret = true;
3099 	}
3100 
3101 	put_obj_stock(flags);
3102 
3103 	return ret;
3104 }
3105 
3106 static void drain_obj_stock(struct obj_stock *stock)
3107 {
3108 	struct obj_cgroup *old = stock->cached_objcg;
3109 
3110 	if (!old)
3111 		return;
3112 
3113 	if (stock->nr_bytes) {
3114 		unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3115 		unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3116 
3117 		if (nr_pages)
3118 			obj_cgroup_uncharge_pages(old, nr_pages);
3119 
3120 		/*
3121 		 * The leftover is flushed to the centralized per-memcg value.
3122 		 * On the next attempt to refill obj stock it will be moved
3123 		 * to a per-cpu stock (probably, on an other CPU), see
3124 		 * refill_obj_stock().
3125 		 *
3126 		 * How often it's flushed is a trade-off between the memory
3127 		 * limit enforcement accuracy and potential CPU contention,
3128 		 * so it might be changed in the future.
3129 		 */
3130 		atomic_add(nr_bytes, &old->nr_charged_bytes);
3131 		stock->nr_bytes = 0;
3132 	}
3133 
3134 	/*
3135 	 * Flush the vmstat data in current stock
3136 	 */
3137 	if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3138 		if (stock->nr_slab_reclaimable_b) {
3139 			mod_objcg_mlstate(old, stock->cached_pgdat,
3140 					  NR_SLAB_RECLAIMABLE_B,
3141 					  stock->nr_slab_reclaimable_b);
3142 			stock->nr_slab_reclaimable_b = 0;
3143 		}
3144 		if (stock->nr_slab_unreclaimable_b) {
3145 			mod_objcg_mlstate(old, stock->cached_pgdat,
3146 					  NR_SLAB_UNRECLAIMABLE_B,
3147 					  stock->nr_slab_unreclaimable_b);
3148 			stock->nr_slab_unreclaimable_b = 0;
3149 		}
3150 		stock->cached_pgdat = NULL;
3151 	}
3152 
3153 	obj_cgroup_put(old);
3154 	stock->cached_objcg = NULL;
3155 }
3156 
3157 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3158 				     struct mem_cgroup *root_memcg)
3159 {
3160 	struct mem_cgroup *memcg;
3161 
3162 	if (in_task() && stock->task_obj.cached_objcg) {
3163 		memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3164 		if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3165 			return true;
3166 	}
3167 	if (stock->irq_obj.cached_objcg) {
3168 		memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3169 		if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3170 			return true;
3171 	}
3172 
3173 	return false;
3174 }
3175 
3176 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3177 			     bool allow_uncharge)
3178 {
3179 	unsigned long flags;
3180 	struct obj_stock *stock = get_obj_stock(&flags);
3181 	unsigned int nr_pages = 0;
3182 
3183 	if (stock->cached_objcg != objcg) { /* reset if necessary */
3184 		drain_obj_stock(stock);
3185 		obj_cgroup_get(objcg);
3186 		stock->cached_objcg = objcg;
3187 		stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3188 				? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3189 		allow_uncharge = true;	/* Allow uncharge when objcg changes */
3190 	}
3191 	stock->nr_bytes += nr_bytes;
3192 
3193 	if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3194 		nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3195 		stock->nr_bytes &= (PAGE_SIZE - 1);
3196 	}
3197 
3198 	put_obj_stock(flags);
3199 
3200 	if (nr_pages)
3201 		obj_cgroup_uncharge_pages(objcg, nr_pages);
3202 }
3203 
3204 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3205 {
3206 	unsigned int nr_pages, nr_bytes;
3207 	int ret;
3208 
3209 	if (consume_obj_stock(objcg, size))
3210 		return 0;
3211 
3212 	/*
3213 	 * In theory, objcg->nr_charged_bytes can have enough
3214 	 * pre-charged bytes to satisfy the allocation. However,
3215 	 * flushing objcg->nr_charged_bytes requires two atomic
3216 	 * operations, and objcg->nr_charged_bytes can't be big.
3217 	 * The shared objcg->nr_charged_bytes can also become a
3218 	 * performance bottleneck if all tasks of the same memcg are
3219 	 * trying to update it. So it's better to ignore it and try
3220 	 * grab some new pages. The stock's nr_bytes will be flushed to
3221 	 * objcg->nr_charged_bytes later on when objcg changes.
3222 	 *
3223 	 * The stock's nr_bytes may contain enough pre-charged bytes
3224 	 * to allow one less page from being charged, but we can't rely
3225 	 * on the pre-charged bytes not being changed outside of
3226 	 * consume_obj_stock() or refill_obj_stock(). So ignore those
3227 	 * pre-charged bytes as well when charging pages. To avoid a
3228 	 * page uncharge right after a page charge, we set the
3229 	 * allow_uncharge flag to false when calling refill_obj_stock()
3230 	 * to temporarily allow the pre-charged bytes to exceed the page
3231 	 * size limit. The maximum reachable value of the pre-charged
3232 	 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3233 	 * race.
3234 	 */
3235 	nr_pages = size >> PAGE_SHIFT;
3236 	nr_bytes = size & (PAGE_SIZE - 1);
3237 
3238 	if (nr_bytes)
3239 		nr_pages += 1;
3240 
3241 	ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3242 	if (!ret && nr_bytes)
3243 		refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3244 
3245 	return ret;
3246 }
3247 
3248 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3249 {
3250 	refill_obj_stock(objcg, size, true);
3251 }
3252 
3253 #endif /* CONFIG_MEMCG_KMEM */
3254 
3255 /*
3256  * Because page_memcg(head) is not set on tails, set it now.
3257  */
3258 void split_page_memcg(struct page *head, unsigned int nr)
3259 {
3260 	struct mem_cgroup *memcg = page_memcg(head);
3261 	int i;
3262 
3263 	if (mem_cgroup_disabled() || !memcg)
3264 		return;
3265 
3266 	for (i = 1; i < nr; i++)
3267 		head[i].memcg_data = head->memcg_data;
3268 
3269 	if (PageMemcgKmem(head))
3270 		obj_cgroup_get_many(__page_objcg(head), nr - 1);
3271 	else
3272 		css_get_many(&memcg->css, nr - 1);
3273 }
3274 
3275 #ifdef CONFIG_MEMCG_SWAP
3276 /**
3277  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3278  * @entry: swap entry to be moved
3279  * @from:  mem_cgroup which the entry is moved from
3280  * @to:  mem_cgroup which the entry is moved to
3281  *
3282  * It succeeds only when the swap_cgroup's record for this entry is the same
3283  * as the mem_cgroup's id of @from.
3284  *
3285  * Returns 0 on success, -EINVAL on failure.
3286  *
3287  * The caller must have charged to @to, IOW, called page_counter_charge() about
3288  * both res and memsw, and called css_get().
3289  */
3290 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3291 				struct mem_cgroup *from, struct mem_cgroup *to)
3292 {
3293 	unsigned short old_id, new_id;
3294 
3295 	old_id = mem_cgroup_id(from);
3296 	new_id = mem_cgroup_id(to);
3297 
3298 	if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3299 		mod_memcg_state(from, MEMCG_SWAP, -1);
3300 		mod_memcg_state(to, MEMCG_SWAP, 1);
3301 		return 0;
3302 	}
3303 	return -EINVAL;
3304 }
3305 #else
3306 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3307 				struct mem_cgroup *from, struct mem_cgroup *to)
3308 {
3309 	return -EINVAL;
3310 }
3311 #endif
3312 
3313 static DEFINE_MUTEX(memcg_max_mutex);
3314 
3315 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3316 				 unsigned long max, bool memsw)
3317 {
3318 	bool enlarge = false;
3319 	bool drained = false;
3320 	int ret;
3321 	bool limits_invariant;
3322 	struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3323 
3324 	do {
3325 		if (signal_pending(current)) {
3326 			ret = -EINTR;
3327 			break;
3328 		}
3329 
3330 		mutex_lock(&memcg_max_mutex);
3331 		/*
3332 		 * Make sure that the new limit (memsw or memory limit) doesn't
3333 		 * break our basic invariant rule memory.max <= memsw.max.
3334 		 */
3335 		limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3336 					   max <= memcg->memsw.max;
3337 		if (!limits_invariant) {
3338 			mutex_unlock(&memcg_max_mutex);
3339 			ret = -EINVAL;
3340 			break;
3341 		}
3342 		if (max > counter->max)
3343 			enlarge = true;
3344 		ret = page_counter_set_max(counter, max);
3345 		mutex_unlock(&memcg_max_mutex);
3346 
3347 		if (!ret)
3348 			break;
3349 
3350 		if (!drained) {
3351 			drain_all_stock(memcg);
3352 			drained = true;
3353 			continue;
3354 		}
3355 
3356 		if (!try_to_free_mem_cgroup_pages(memcg, 1,
3357 					GFP_KERNEL, !memsw)) {
3358 			ret = -EBUSY;
3359 			break;
3360 		}
3361 	} while (true);
3362 
3363 	if (!ret && enlarge)
3364 		memcg_oom_recover(memcg);
3365 
3366 	return ret;
3367 }
3368 
3369 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3370 					    gfp_t gfp_mask,
3371 					    unsigned long *total_scanned)
3372 {
3373 	unsigned long nr_reclaimed = 0;
3374 	struct mem_cgroup_per_node *mz, *next_mz = NULL;
3375 	unsigned long reclaimed;
3376 	int loop = 0;
3377 	struct mem_cgroup_tree_per_node *mctz;
3378 	unsigned long excess;
3379 	unsigned long nr_scanned;
3380 
3381 	if (order > 0)
3382 		return 0;
3383 
3384 	mctz = soft_limit_tree_node(pgdat->node_id);
3385 
3386 	/*
3387 	 * Do not even bother to check the largest node if the root
3388 	 * is empty. Do it lockless to prevent lock bouncing. Races
3389 	 * are acceptable as soft limit is best effort anyway.
3390 	 */
3391 	if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3392 		return 0;
3393 
3394 	/*
3395 	 * This loop can run a while, specially if mem_cgroup's continuously
3396 	 * keep exceeding their soft limit and putting the system under
3397 	 * pressure
3398 	 */
3399 	do {
3400 		if (next_mz)
3401 			mz = next_mz;
3402 		else
3403 			mz = mem_cgroup_largest_soft_limit_node(mctz);
3404 		if (!mz)
3405 			break;
3406 
3407 		nr_scanned = 0;
3408 		reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3409 						    gfp_mask, &nr_scanned);
3410 		nr_reclaimed += reclaimed;
3411 		*total_scanned += nr_scanned;
3412 		spin_lock_irq(&mctz->lock);
3413 		__mem_cgroup_remove_exceeded(mz, mctz);
3414 
3415 		/*
3416 		 * If we failed to reclaim anything from this memory cgroup
3417 		 * it is time to move on to the next cgroup
3418 		 */
3419 		next_mz = NULL;
3420 		if (!reclaimed)
3421 			next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3422 
3423 		excess = soft_limit_excess(mz->memcg);
3424 		/*
3425 		 * One school of thought says that we should not add
3426 		 * back the node to the tree if reclaim returns 0.
3427 		 * But our reclaim could return 0, simply because due
3428 		 * to priority we are exposing a smaller subset of
3429 		 * memory to reclaim from. Consider this as a longer
3430 		 * term TODO.
3431 		 */
3432 		/* If excess == 0, no tree ops */
3433 		__mem_cgroup_insert_exceeded(mz, mctz, excess);
3434 		spin_unlock_irq(&mctz->lock);
3435 		css_put(&mz->memcg->css);
3436 		loop++;
3437 		/*
3438 		 * Could not reclaim anything and there are no more
3439 		 * mem cgroups to try or we seem to be looping without
3440 		 * reclaiming anything.
3441 		 */
3442 		if (!nr_reclaimed &&
3443 			(next_mz == NULL ||
3444 			loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3445 			break;
3446 	} while (!nr_reclaimed);
3447 	if (next_mz)
3448 		css_put(&next_mz->memcg->css);
3449 	return nr_reclaimed;
3450 }
3451 
3452 /*
3453  * Reclaims as many pages from the given memcg as possible.
3454  *
3455  * Caller is responsible for holding css reference for memcg.
3456  */
3457 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3458 {
3459 	int nr_retries = MAX_RECLAIM_RETRIES;
3460 
3461 	/* we call try-to-free pages for make this cgroup empty */
3462 	lru_add_drain_all();
3463 
3464 	drain_all_stock(memcg);
3465 
3466 	/* try to free all pages in this cgroup */
3467 	while (nr_retries && page_counter_read(&memcg->memory)) {
3468 		int progress;
3469 
3470 		if (signal_pending(current))
3471 			return -EINTR;
3472 
3473 		progress = try_to_free_mem_cgroup_pages(memcg, 1,
3474 							GFP_KERNEL, true);
3475 		if (!progress) {
3476 			nr_retries--;
3477 			/* maybe some writeback is necessary */
3478 			congestion_wait(BLK_RW_ASYNC, HZ/10);
3479 		}
3480 
3481 	}
3482 
3483 	return 0;
3484 }
3485 
3486 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3487 					    char *buf, size_t nbytes,
3488 					    loff_t off)
3489 {
3490 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3491 
3492 	if (mem_cgroup_is_root(memcg))
3493 		return -EINVAL;
3494 	return mem_cgroup_force_empty(memcg) ?: nbytes;
3495 }
3496 
3497 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3498 				     struct cftype *cft)
3499 {
3500 	return 1;
3501 }
3502 
3503 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3504 				      struct cftype *cft, u64 val)
3505 {
3506 	if (val == 1)
3507 		return 0;
3508 
3509 	pr_warn_once("Non-hierarchical mode is deprecated. "
3510 		     "Please report your usecase to linux-mm@kvack.org if you "
3511 		     "depend on this functionality.\n");
3512 
3513 	return -EINVAL;
3514 }
3515 
3516 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3517 {
3518 	unsigned long val;
3519 
3520 	if (mem_cgroup_is_root(memcg)) {
3521 		/* mem_cgroup_threshold() calls here from irqsafe context */
3522 		cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
3523 		val = memcg_page_state(memcg, NR_FILE_PAGES) +
3524 			memcg_page_state(memcg, NR_ANON_MAPPED);
3525 		if (swap)
3526 			val += memcg_page_state(memcg, MEMCG_SWAP);
3527 	} else {
3528 		if (!swap)
3529 			val = page_counter_read(&memcg->memory);
3530 		else
3531 			val = page_counter_read(&memcg->memsw);
3532 	}
3533 	return val;
3534 }
3535 
3536 enum {
3537 	RES_USAGE,
3538 	RES_LIMIT,
3539 	RES_MAX_USAGE,
3540 	RES_FAILCNT,
3541 	RES_SOFT_LIMIT,
3542 };
3543 
3544 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3545 			       struct cftype *cft)
3546 {
3547 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3548 	struct page_counter *counter;
3549 
3550 	switch (MEMFILE_TYPE(cft->private)) {
3551 	case _MEM:
3552 		counter = &memcg->memory;
3553 		break;
3554 	case _MEMSWAP:
3555 		counter = &memcg->memsw;
3556 		break;
3557 	case _KMEM:
3558 		counter = &memcg->kmem;
3559 		break;
3560 	case _TCP:
3561 		counter = &memcg->tcpmem;
3562 		break;
3563 	default:
3564 		BUG();
3565 	}
3566 
3567 	switch (MEMFILE_ATTR(cft->private)) {
3568 	case RES_USAGE:
3569 		if (counter == &memcg->memory)
3570 			return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3571 		if (counter == &memcg->memsw)
3572 			return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3573 		return (u64)page_counter_read(counter) * PAGE_SIZE;
3574 	case RES_LIMIT:
3575 		return (u64)counter->max * PAGE_SIZE;
3576 	case RES_MAX_USAGE:
3577 		return (u64)counter->watermark * PAGE_SIZE;
3578 	case RES_FAILCNT:
3579 		return counter->failcnt;
3580 	case RES_SOFT_LIMIT:
3581 		return (u64)memcg->soft_limit * PAGE_SIZE;
3582 	default:
3583 		BUG();
3584 	}
3585 }
3586 
3587 #ifdef CONFIG_MEMCG_KMEM
3588 static int memcg_online_kmem(struct mem_cgroup *memcg)
3589 {
3590 	struct obj_cgroup *objcg;
3591 	int memcg_id;
3592 
3593 	if (cgroup_memory_nokmem)
3594 		return 0;
3595 
3596 	BUG_ON(memcg->kmemcg_id >= 0);
3597 	BUG_ON(memcg->kmem_state);
3598 
3599 	memcg_id = memcg_alloc_cache_id();
3600 	if (memcg_id < 0)
3601 		return memcg_id;
3602 
3603 	objcg = obj_cgroup_alloc();
3604 	if (!objcg) {
3605 		memcg_free_cache_id(memcg_id);
3606 		return -ENOMEM;
3607 	}
3608 	objcg->memcg = memcg;
3609 	rcu_assign_pointer(memcg->objcg, objcg);
3610 
3611 	static_branch_enable(&memcg_kmem_enabled_key);
3612 
3613 	memcg->kmemcg_id = memcg_id;
3614 	memcg->kmem_state = KMEM_ONLINE;
3615 
3616 	return 0;
3617 }
3618 
3619 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3620 {
3621 	struct cgroup_subsys_state *css;
3622 	struct mem_cgroup *parent, *child;
3623 	int kmemcg_id;
3624 
3625 	if (memcg->kmem_state != KMEM_ONLINE)
3626 		return;
3627 
3628 	memcg->kmem_state = KMEM_ALLOCATED;
3629 
3630 	parent = parent_mem_cgroup(memcg);
3631 	if (!parent)
3632 		parent = root_mem_cgroup;
3633 
3634 	memcg_reparent_objcgs(memcg, parent);
3635 
3636 	kmemcg_id = memcg->kmemcg_id;
3637 	BUG_ON(kmemcg_id < 0);
3638 
3639 	/*
3640 	 * Change kmemcg_id of this cgroup and all its descendants to the
3641 	 * parent's id, and then move all entries from this cgroup's list_lrus
3642 	 * to ones of the parent. After we have finished, all list_lrus
3643 	 * corresponding to this cgroup are guaranteed to remain empty. The
3644 	 * ordering is imposed by list_lru_node->lock taken by
3645 	 * memcg_drain_all_list_lrus().
3646 	 */
3647 	rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3648 	css_for_each_descendant_pre(css, &memcg->css) {
3649 		child = mem_cgroup_from_css(css);
3650 		BUG_ON(child->kmemcg_id != kmemcg_id);
3651 		child->kmemcg_id = parent->kmemcg_id;
3652 	}
3653 	rcu_read_unlock();
3654 
3655 	memcg_drain_all_list_lrus(kmemcg_id, parent);
3656 
3657 	memcg_free_cache_id(kmemcg_id);
3658 }
3659 
3660 static void memcg_free_kmem(struct mem_cgroup *memcg)
3661 {
3662 	/* css_alloc() failed, offlining didn't happen */
3663 	if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3664 		memcg_offline_kmem(memcg);
3665 }
3666 #else
3667 static int memcg_online_kmem(struct mem_cgroup *memcg)
3668 {
3669 	return 0;
3670 }
3671 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3672 {
3673 }
3674 static void memcg_free_kmem(struct mem_cgroup *memcg)
3675 {
3676 }
3677 #endif /* CONFIG_MEMCG_KMEM */
3678 
3679 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3680 				 unsigned long max)
3681 {
3682 	int ret;
3683 
3684 	mutex_lock(&memcg_max_mutex);
3685 	ret = page_counter_set_max(&memcg->kmem, max);
3686 	mutex_unlock(&memcg_max_mutex);
3687 	return ret;
3688 }
3689 
3690 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3691 {
3692 	int ret;
3693 
3694 	mutex_lock(&memcg_max_mutex);
3695 
3696 	ret = page_counter_set_max(&memcg->tcpmem, max);
3697 	if (ret)
3698 		goto out;
3699 
3700 	if (!memcg->tcpmem_active) {
3701 		/*
3702 		 * The active flag needs to be written after the static_key
3703 		 * update. This is what guarantees that the socket activation
3704 		 * function is the last one to run. See mem_cgroup_sk_alloc()
3705 		 * for details, and note that we don't mark any socket as
3706 		 * belonging to this memcg until that flag is up.
3707 		 *
3708 		 * We need to do this, because static_keys will span multiple
3709 		 * sites, but we can't control their order. If we mark a socket
3710 		 * as accounted, but the accounting functions are not patched in
3711 		 * yet, we'll lose accounting.
3712 		 *
3713 		 * We never race with the readers in mem_cgroup_sk_alloc(),
3714 		 * because when this value change, the code to process it is not
3715 		 * patched in yet.
3716 		 */
3717 		static_branch_inc(&memcg_sockets_enabled_key);
3718 		memcg->tcpmem_active = true;
3719 	}
3720 out:
3721 	mutex_unlock(&memcg_max_mutex);
3722 	return ret;
3723 }
3724 
3725 /*
3726  * The user of this function is...
3727  * RES_LIMIT.
3728  */
3729 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3730 				char *buf, size_t nbytes, loff_t off)
3731 {
3732 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3733 	unsigned long nr_pages;
3734 	int ret;
3735 
3736 	buf = strstrip(buf);
3737 	ret = page_counter_memparse(buf, "-1", &nr_pages);
3738 	if (ret)
3739 		return ret;
3740 
3741 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
3742 	case RES_LIMIT:
3743 		if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3744 			ret = -EINVAL;
3745 			break;
3746 		}
3747 		switch (MEMFILE_TYPE(of_cft(of)->private)) {
3748 		case _MEM:
3749 			ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3750 			break;
3751 		case _MEMSWAP:
3752 			ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3753 			break;
3754 		case _KMEM:
3755 			pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3756 				     "Please report your usecase to linux-mm@kvack.org if you "
3757 				     "depend on this functionality.\n");
3758 			ret = memcg_update_kmem_max(memcg, nr_pages);
3759 			break;
3760 		case _TCP:
3761 			ret = memcg_update_tcp_max(memcg, nr_pages);
3762 			break;
3763 		}
3764 		break;
3765 	case RES_SOFT_LIMIT:
3766 		memcg->soft_limit = nr_pages;
3767 		ret = 0;
3768 		break;
3769 	}
3770 	return ret ?: nbytes;
3771 }
3772 
3773 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3774 				size_t nbytes, loff_t off)
3775 {
3776 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3777 	struct page_counter *counter;
3778 
3779 	switch (MEMFILE_TYPE(of_cft(of)->private)) {
3780 	case _MEM:
3781 		counter = &memcg->memory;
3782 		break;
3783 	case _MEMSWAP:
3784 		counter = &memcg->memsw;
3785 		break;
3786 	case _KMEM:
3787 		counter = &memcg->kmem;
3788 		break;
3789 	case _TCP:
3790 		counter = &memcg->tcpmem;
3791 		break;
3792 	default:
3793 		BUG();
3794 	}
3795 
3796 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
3797 	case RES_MAX_USAGE:
3798 		page_counter_reset_watermark(counter);
3799 		break;
3800 	case RES_FAILCNT:
3801 		counter->failcnt = 0;
3802 		break;
3803 	default:
3804 		BUG();
3805 	}
3806 
3807 	return nbytes;
3808 }
3809 
3810 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3811 					struct cftype *cft)
3812 {
3813 	return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3814 }
3815 
3816 #ifdef CONFIG_MMU
3817 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3818 					struct cftype *cft, u64 val)
3819 {
3820 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3821 
3822 	if (val & ~MOVE_MASK)
3823 		return -EINVAL;
3824 
3825 	/*
3826 	 * No kind of locking is needed in here, because ->can_attach() will
3827 	 * check this value once in the beginning of the process, and then carry
3828 	 * on with stale data. This means that changes to this value will only
3829 	 * affect task migrations starting after the change.
3830 	 */
3831 	memcg->move_charge_at_immigrate = val;
3832 	return 0;
3833 }
3834 #else
3835 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3836 					struct cftype *cft, u64 val)
3837 {
3838 	return -ENOSYS;
3839 }
3840 #endif
3841 
3842 #ifdef CONFIG_NUMA
3843 
3844 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3845 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3846 #define LRU_ALL	     ((1 << NR_LRU_LISTS) - 1)
3847 
3848 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3849 				int nid, unsigned int lru_mask, bool tree)
3850 {
3851 	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3852 	unsigned long nr = 0;
3853 	enum lru_list lru;
3854 
3855 	VM_BUG_ON((unsigned)nid >= nr_node_ids);
3856 
3857 	for_each_lru(lru) {
3858 		if (!(BIT(lru) & lru_mask))
3859 			continue;
3860 		if (tree)
3861 			nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3862 		else
3863 			nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3864 	}
3865 	return nr;
3866 }
3867 
3868 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3869 					     unsigned int lru_mask,
3870 					     bool tree)
3871 {
3872 	unsigned long nr = 0;
3873 	enum lru_list lru;
3874 
3875 	for_each_lru(lru) {
3876 		if (!(BIT(lru) & lru_mask))
3877 			continue;
3878 		if (tree)
3879 			nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3880 		else
3881 			nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3882 	}
3883 	return nr;
3884 }
3885 
3886 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3887 {
3888 	struct numa_stat {
3889 		const char *name;
3890 		unsigned int lru_mask;
3891 	};
3892 
3893 	static const struct numa_stat stats[] = {
3894 		{ "total", LRU_ALL },
3895 		{ "file", LRU_ALL_FILE },
3896 		{ "anon", LRU_ALL_ANON },
3897 		{ "unevictable", BIT(LRU_UNEVICTABLE) },
3898 	};
3899 	const struct numa_stat *stat;
3900 	int nid;
3901 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3902 
3903 	cgroup_rstat_flush(memcg->css.cgroup);
3904 
3905 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3906 		seq_printf(m, "%s=%lu", stat->name,
3907 			   mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3908 						   false));
3909 		for_each_node_state(nid, N_MEMORY)
3910 			seq_printf(m, " N%d=%lu", nid,
3911 				   mem_cgroup_node_nr_lru_pages(memcg, nid,
3912 							stat->lru_mask, false));
3913 		seq_putc(m, '\n');
3914 	}
3915 
3916 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3917 
3918 		seq_printf(m, "hierarchical_%s=%lu", stat->name,
3919 			   mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3920 						   true));
3921 		for_each_node_state(nid, N_MEMORY)
3922 			seq_printf(m, " N%d=%lu", nid,
3923 				   mem_cgroup_node_nr_lru_pages(memcg, nid,
3924 							stat->lru_mask, true));
3925 		seq_putc(m, '\n');
3926 	}
3927 
3928 	return 0;
3929 }
3930 #endif /* CONFIG_NUMA */
3931 
3932 static const unsigned int memcg1_stats[] = {
3933 	NR_FILE_PAGES,
3934 	NR_ANON_MAPPED,
3935 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3936 	NR_ANON_THPS,
3937 #endif
3938 	NR_SHMEM,
3939 	NR_FILE_MAPPED,
3940 	NR_FILE_DIRTY,
3941 	NR_WRITEBACK,
3942 	MEMCG_SWAP,
3943 };
3944 
3945 static const char *const memcg1_stat_names[] = {
3946 	"cache",
3947 	"rss",
3948 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3949 	"rss_huge",
3950 #endif
3951 	"shmem",
3952 	"mapped_file",
3953 	"dirty",
3954 	"writeback",
3955 	"swap",
3956 };
3957 
3958 /* Universal VM events cgroup1 shows, original sort order */
3959 static const unsigned int memcg1_events[] = {
3960 	PGPGIN,
3961 	PGPGOUT,
3962 	PGFAULT,
3963 	PGMAJFAULT,
3964 };
3965 
3966 static int memcg_stat_show(struct seq_file *m, void *v)
3967 {
3968 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3969 	unsigned long memory, memsw;
3970 	struct mem_cgroup *mi;
3971 	unsigned int i;
3972 
3973 	BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3974 
3975 	cgroup_rstat_flush(memcg->css.cgroup);
3976 
3977 	for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3978 		unsigned long nr;
3979 
3980 		if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3981 			continue;
3982 		nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3983 		seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3984 	}
3985 
3986 	for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3987 		seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3988 			   memcg_events_local(memcg, memcg1_events[i]));
3989 
3990 	for (i = 0; i < NR_LRU_LISTS; i++)
3991 		seq_printf(m, "%s %lu\n", lru_list_name(i),
3992 			   memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3993 			   PAGE_SIZE);
3994 
3995 	/* Hierarchical information */
3996 	memory = memsw = PAGE_COUNTER_MAX;
3997 	for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3998 		memory = min(memory, READ_ONCE(mi->memory.max));
3999 		memsw = min(memsw, READ_ONCE(mi->memsw.max));
4000 	}
4001 	seq_printf(m, "hierarchical_memory_limit %llu\n",
4002 		   (u64)memory * PAGE_SIZE);
4003 	if (do_memsw_account())
4004 		seq_printf(m, "hierarchical_memsw_limit %llu\n",
4005 			   (u64)memsw * PAGE_SIZE);
4006 
4007 	for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4008 		unsigned long nr;
4009 
4010 		if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4011 			continue;
4012 		nr = memcg_page_state(memcg, memcg1_stats[i]);
4013 		seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4014 						(u64)nr * PAGE_SIZE);
4015 	}
4016 
4017 	for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4018 		seq_printf(m, "total_%s %llu\n",
4019 			   vm_event_name(memcg1_events[i]),
4020 			   (u64)memcg_events(memcg, memcg1_events[i]));
4021 
4022 	for (i = 0; i < NR_LRU_LISTS; i++)
4023 		seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4024 			   (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4025 			   PAGE_SIZE);
4026 
4027 #ifdef CONFIG_DEBUG_VM
4028 	{
4029 		pg_data_t *pgdat;
4030 		struct mem_cgroup_per_node *mz;
4031 		unsigned long anon_cost = 0;
4032 		unsigned long file_cost = 0;
4033 
4034 		for_each_online_pgdat(pgdat) {
4035 			mz = memcg->nodeinfo[pgdat->node_id];
4036 
4037 			anon_cost += mz->lruvec.anon_cost;
4038 			file_cost += mz->lruvec.file_cost;
4039 		}
4040 		seq_printf(m, "anon_cost %lu\n", anon_cost);
4041 		seq_printf(m, "file_cost %lu\n", file_cost);
4042 	}
4043 #endif
4044 
4045 	return 0;
4046 }
4047 
4048 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4049 				      struct cftype *cft)
4050 {
4051 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4052 
4053 	return mem_cgroup_swappiness(memcg);
4054 }
4055 
4056 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4057 				       struct cftype *cft, u64 val)
4058 {
4059 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4060 
4061 	if (val > 200)
4062 		return -EINVAL;
4063 
4064 	if (!mem_cgroup_is_root(memcg))
4065 		memcg->swappiness = val;
4066 	else
4067 		vm_swappiness = val;
4068 
4069 	return 0;
4070 }
4071 
4072 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4073 {
4074 	struct mem_cgroup_threshold_ary *t;
4075 	unsigned long usage;
4076 	int i;
4077 
4078 	rcu_read_lock();
4079 	if (!swap)
4080 		t = rcu_dereference(memcg->thresholds.primary);
4081 	else
4082 		t = rcu_dereference(memcg->memsw_thresholds.primary);
4083 
4084 	if (!t)
4085 		goto unlock;
4086 
4087 	usage = mem_cgroup_usage(memcg, swap);
4088 
4089 	/*
4090 	 * current_threshold points to threshold just below or equal to usage.
4091 	 * If it's not true, a threshold was crossed after last
4092 	 * call of __mem_cgroup_threshold().
4093 	 */
4094 	i = t->current_threshold;
4095 
4096 	/*
4097 	 * Iterate backward over array of thresholds starting from
4098 	 * current_threshold and check if a threshold is crossed.
4099 	 * If none of thresholds below usage is crossed, we read
4100 	 * only one element of the array here.
4101 	 */
4102 	for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4103 		eventfd_signal(t->entries[i].eventfd, 1);
4104 
4105 	/* i = current_threshold + 1 */
4106 	i++;
4107 
4108 	/*
4109 	 * Iterate forward over array of thresholds starting from
4110 	 * current_threshold+1 and check if a threshold is crossed.
4111 	 * If none of thresholds above usage is crossed, we read
4112 	 * only one element of the array here.
4113 	 */
4114 	for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4115 		eventfd_signal(t->entries[i].eventfd, 1);
4116 
4117 	/* Update current_threshold */
4118 	t->current_threshold = i - 1;
4119 unlock:
4120 	rcu_read_unlock();
4121 }
4122 
4123 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4124 {
4125 	while (memcg) {
4126 		__mem_cgroup_threshold(memcg, false);
4127 		if (do_memsw_account())
4128 			__mem_cgroup_threshold(memcg, true);
4129 
4130 		memcg = parent_mem_cgroup(memcg);
4131 	}
4132 }
4133 
4134 static int compare_thresholds(const void *a, const void *b)
4135 {
4136 	const struct mem_cgroup_threshold *_a = a;
4137 	const struct mem_cgroup_threshold *_b = b;
4138 
4139 	if (_a->threshold > _b->threshold)
4140 		return 1;
4141 
4142 	if (_a->threshold < _b->threshold)
4143 		return -1;
4144 
4145 	return 0;
4146 }
4147 
4148 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4149 {
4150 	struct mem_cgroup_eventfd_list *ev;
4151 
4152 	spin_lock(&memcg_oom_lock);
4153 
4154 	list_for_each_entry(ev, &memcg->oom_notify, list)
4155 		eventfd_signal(ev->eventfd, 1);
4156 
4157 	spin_unlock(&memcg_oom_lock);
4158 	return 0;
4159 }
4160 
4161 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4162 {
4163 	struct mem_cgroup *iter;
4164 
4165 	for_each_mem_cgroup_tree(iter, memcg)
4166 		mem_cgroup_oom_notify_cb(iter);
4167 }
4168 
4169 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4170 	struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4171 {
4172 	struct mem_cgroup_thresholds *thresholds;
4173 	struct mem_cgroup_threshold_ary *new;
4174 	unsigned long threshold;
4175 	unsigned long usage;
4176 	int i, size, ret;
4177 
4178 	ret = page_counter_memparse(args, "-1", &threshold);
4179 	if (ret)
4180 		return ret;
4181 
4182 	mutex_lock(&memcg->thresholds_lock);
4183 
4184 	if (type == _MEM) {
4185 		thresholds = &memcg->thresholds;
4186 		usage = mem_cgroup_usage(memcg, false);
4187 	} else if (type == _MEMSWAP) {
4188 		thresholds = &memcg->memsw_thresholds;
4189 		usage = mem_cgroup_usage(memcg, true);
4190 	} else
4191 		BUG();
4192 
4193 	/* Check if a threshold crossed before adding a new one */
4194 	if (thresholds->primary)
4195 		__mem_cgroup_threshold(memcg, type == _MEMSWAP);
4196 
4197 	size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4198 
4199 	/* Allocate memory for new array of thresholds */
4200 	new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4201 	if (!new) {
4202 		ret = -ENOMEM;
4203 		goto unlock;
4204 	}
4205 	new->size = size;
4206 
4207 	/* Copy thresholds (if any) to new array */
4208 	if (thresholds->primary)
4209 		memcpy(new->entries, thresholds->primary->entries,
4210 		       flex_array_size(new, entries, size - 1));
4211 
4212 	/* Add new threshold */
4213 	new->entries[size - 1].eventfd = eventfd;
4214 	new->entries[size - 1].threshold = threshold;
4215 
4216 	/* Sort thresholds. Registering of new threshold isn't time-critical */
4217 	sort(new->entries, size, sizeof(*new->entries),
4218 			compare_thresholds, NULL);
4219 
4220 	/* Find current threshold */
4221 	new->current_threshold = -1;
4222 	for (i = 0; i < size; i++) {
4223 		if (new->entries[i].threshold <= usage) {
4224 			/*
4225 			 * new->current_threshold will not be used until
4226 			 * rcu_assign_pointer(), so it's safe to increment
4227 			 * it here.
4228 			 */
4229 			++new->current_threshold;
4230 		} else
4231 			break;
4232 	}
4233 
4234 	/* Free old spare buffer and save old primary buffer as spare */
4235 	kfree(thresholds->spare);
4236 	thresholds->spare = thresholds->primary;
4237 
4238 	rcu_assign_pointer(thresholds->primary, new);
4239 
4240 	/* To be sure that nobody uses thresholds */
4241 	synchronize_rcu();
4242 
4243 unlock:
4244 	mutex_unlock(&memcg->thresholds_lock);
4245 
4246 	return ret;
4247 }
4248 
4249 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4250 	struct eventfd_ctx *eventfd, const char *args)
4251 {
4252 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4253 }
4254 
4255 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4256 	struct eventfd_ctx *eventfd, const char *args)
4257 {
4258 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4259 }
4260 
4261 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4262 	struct eventfd_ctx *eventfd, enum res_type type)
4263 {
4264 	struct mem_cgroup_thresholds *thresholds;
4265 	struct mem_cgroup_threshold_ary *new;
4266 	unsigned long usage;
4267 	int i, j, size, entries;
4268 
4269 	mutex_lock(&memcg->thresholds_lock);
4270 
4271 	if (type == _MEM) {
4272 		thresholds = &memcg->thresholds;
4273 		usage = mem_cgroup_usage(memcg, false);
4274 	} else if (type == _MEMSWAP) {
4275 		thresholds = &memcg->memsw_thresholds;
4276 		usage = mem_cgroup_usage(memcg, true);
4277 	} else
4278 		BUG();
4279 
4280 	if (!thresholds->primary)
4281 		goto unlock;
4282 
4283 	/* Check if a threshold crossed before removing */
4284 	__mem_cgroup_threshold(memcg, type == _MEMSWAP);
4285 
4286 	/* Calculate new number of threshold */
4287 	size = entries = 0;
4288 	for (i = 0; i < thresholds->primary->size; i++) {
4289 		if (thresholds->primary->entries[i].eventfd != eventfd)
4290 			size++;
4291 		else
4292 			entries++;
4293 	}
4294 
4295 	new = thresholds->spare;
4296 
4297 	/* If no items related to eventfd have been cleared, nothing to do */
4298 	if (!entries)
4299 		goto unlock;
4300 
4301 	/* Set thresholds array to NULL if we don't have thresholds */
4302 	if (!size) {
4303 		kfree(new);
4304 		new = NULL;
4305 		goto swap_buffers;
4306 	}
4307 
4308 	new->size = size;
4309 
4310 	/* Copy thresholds and find current threshold */
4311 	new->current_threshold = -1;
4312 	for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4313 		if (thresholds->primary->entries[i].eventfd == eventfd)
4314 			continue;
4315 
4316 		new->entries[j] = thresholds->primary->entries[i];
4317 		if (new->entries[j].threshold <= usage) {
4318 			/*
4319 			 * new->current_threshold will not be used
4320 			 * until rcu_assign_pointer(), so it's safe to increment
4321 			 * it here.
4322 			 */
4323 			++new->current_threshold;
4324 		}
4325 		j++;
4326 	}
4327 
4328 swap_buffers:
4329 	/* Swap primary and spare array */
4330 	thresholds->spare = thresholds->primary;
4331 
4332 	rcu_assign_pointer(thresholds->primary, new);
4333 
4334 	/* To be sure that nobody uses thresholds */
4335 	synchronize_rcu();
4336 
4337 	/* If all events are unregistered, free the spare array */
4338 	if (!new) {
4339 		kfree(thresholds->spare);
4340 		thresholds->spare = NULL;
4341 	}
4342 unlock:
4343 	mutex_unlock(&memcg->thresholds_lock);
4344 }
4345 
4346 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4347 	struct eventfd_ctx *eventfd)
4348 {
4349 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4350 }
4351 
4352 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4353 	struct eventfd_ctx *eventfd)
4354 {
4355 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4356 }
4357 
4358 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4359 	struct eventfd_ctx *eventfd, const char *args)
4360 {
4361 	struct mem_cgroup_eventfd_list *event;
4362 
4363 	event = kmalloc(sizeof(*event),	GFP_KERNEL);
4364 	if (!event)
4365 		return -ENOMEM;
4366 
4367 	spin_lock(&memcg_oom_lock);
4368 
4369 	event->eventfd = eventfd;
4370 	list_add(&event->list, &memcg->oom_notify);
4371 
4372 	/* already in OOM ? */
4373 	if (memcg->under_oom)
4374 		eventfd_signal(eventfd, 1);
4375 	spin_unlock(&memcg_oom_lock);
4376 
4377 	return 0;
4378 }
4379 
4380 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4381 	struct eventfd_ctx *eventfd)
4382 {
4383 	struct mem_cgroup_eventfd_list *ev, *tmp;
4384 
4385 	spin_lock(&memcg_oom_lock);
4386 
4387 	list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4388 		if (ev->eventfd == eventfd) {
4389 			list_del(&ev->list);
4390 			kfree(ev);
4391 		}
4392 	}
4393 
4394 	spin_unlock(&memcg_oom_lock);
4395 }
4396 
4397 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4398 {
4399 	struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4400 
4401 	seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4402 	seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4403 	seq_printf(sf, "oom_kill %lu\n",
4404 		   atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4405 	return 0;
4406 }
4407 
4408 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4409 	struct cftype *cft, u64 val)
4410 {
4411 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4412 
4413 	/* cannot set to root cgroup and only 0 and 1 are allowed */
4414 	if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4415 		return -EINVAL;
4416 
4417 	memcg->oom_kill_disable = val;
4418 	if (!val)
4419 		memcg_oom_recover(memcg);
4420 
4421 	return 0;
4422 }
4423 
4424 #ifdef CONFIG_CGROUP_WRITEBACK
4425 
4426 #include <trace/events/writeback.h>
4427 
4428 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4429 {
4430 	return wb_domain_init(&memcg->cgwb_domain, gfp);
4431 }
4432 
4433 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4434 {
4435 	wb_domain_exit(&memcg->cgwb_domain);
4436 }
4437 
4438 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4439 {
4440 	wb_domain_size_changed(&memcg->cgwb_domain);
4441 }
4442 
4443 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4444 {
4445 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4446 
4447 	if (!memcg->css.parent)
4448 		return NULL;
4449 
4450 	return &memcg->cgwb_domain;
4451 }
4452 
4453 /**
4454  * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4455  * @wb: bdi_writeback in question
4456  * @pfilepages: out parameter for number of file pages
4457  * @pheadroom: out parameter for number of allocatable pages according to memcg
4458  * @pdirty: out parameter for number of dirty pages
4459  * @pwriteback: out parameter for number of pages under writeback
4460  *
4461  * Determine the numbers of file, headroom, dirty, and writeback pages in
4462  * @wb's memcg.  File, dirty and writeback are self-explanatory.  Headroom
4463  * is a bit more involved.
4464  *
4465  * A memcg's headroom is "min(max, high) - used".  In the hierarchy, the
4466  * headroom is calculated as the lowest headroom of itself and the
4467  * ancestors.  Note that this doesn't consider the actual amount of
4468  * available memory in the system.  The caller should further cap
4469  * *@pheadroom accordingly.
4470  */
4471 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4472 			 unsigned long *pheadroom, unsigned long *pdirty,
4473 			 unsigned long *pwriteback)
4474 {
4475 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4476 	struct mem_cgroup *parent;
4477 
4478 	cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4479 
4480 	*pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4481 	*pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4482 	*pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4483 			memcg_page_state(memcg, NR_ACTIVE_FILE);
4484 
4485 	*pheadroom = PAGE_COUNTER_MAX;
4486 	while ((parent = parent_mem_cgroup(memcg))) {
4487 		unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4488 					    READ_ONCE(memcg->memory.high));
4489 		unsigned long used = page_counter_read(&memcg->memory);
4490 
4491 		*pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4492 		memcg = parent;
4493 	}
4494 }
4495 
4496 /*
4497  * Foreign dirty flushing
4498  *
4499  * There's an inherent mismatch between memcg and writeback.  The former
4500  * tracks ownership per-page while the latter per-inode.  This was a
4501  * deliberate design decision because honoring per-page ownership in the
4502  * writeback path is complicated, may lead to higher CPU and IO overheads
4503  * and deemed unnecessary given that write-sharing an inode across
4504  * different cgroups isn't a common use-case.
4505  *
4506  * Combined with inode majority-writer ownership switching, this works well
4507  * enough in most cases but there are some pathological cases.  For
4508  * example, let's say there are two cgroups A and B which keep writing to
4509  * different but confined parts of the same inode.  B owns the inode and
4510  * A's memory is limited far below B's.  A's dirty ratio can rise enough to
4511  * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4512  * triggering background writeback.  A will be slowed down without a way to
4513  * make writeback of the dirty pages happen.
4514  *
4515  * Conditions like the above can lead to a cgroup getting repeatedly and
4516  * severely throttled after making some progress after each
4517  * dirty_expire_interval while the underlying IO device is almost
4518  * completely idle.
4519  *
4520  * Solving this problem completely requires matching the ownership tracking
4521  * granularities between memcg and writeback in either direction.  However,
4522  * the more egregious behaviors can be avoided by simply remembering the
4523  * most recent foreign dirtying events and initiating remote flushes on
4524  * them when local writeback isn't enough to keep the memory clean enough.
4525  *
4526  * The following two functions implement such mechanism.  When a foreign
4527  * page - a page whose memcg and writeback ownerships don't match - is
4528  * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4529  * bdi_writeback on the page owning memcg.  When balance_dirty_pages()
4530  * decides that the memcg needs to sleep due to high dirty ratio, it calls
4531  * mem_cgroup_flush_foreign() which queues writeback on the recorded
4532  * foreign bdi_writebacks which haven't expired.  Both the numbers of
4533  * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4534  * limited to MEMCG_CGWB_FRN_CNT.
4535  *
4536  * The mechanism only remembers IDs and doesn't hold any object references.
4537  * As being wrong occasionally doesn't matter, updates and accesses to the
4538  * records are lockless and racy.
4539  */
4540 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4541 					     struct bdi_writeback *wb)
4542 {
4543 	struct mem_cgroup *memcg = page_memcg(page);
4544 	struct memcg_cgwb_frn *frn;
4545 	u64 now = get_jiffies_64();
4546 	u64 oldest_at = now;
4547 	int oldest = -1;
4548 	int i;
4549 
4550 	trace_track_foreign_dirty(page, wb);
4551 
4552 	/*
4553 	 * Pick the slot to use.  If there is already a slot for @wb, keep
4554 	 * using it.  If not replace the oldest one which isn't being
4555 	 * written out.
4556 	 */
4557 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4558 		frn = &memcg->cgwb_frn[i];
4559 		if (frn->bdi_id == wb->bdi->id &&
4560 		    frn->memcg_id == wb->memcg_css->id)
4561 			break;
4562 		if (time_before64(frn->at, oldest_at) &&
4563 		    atomic_read(&frn->done.cnt) == 1) {
4564 			oldest = i;
4565 			oldest_at = frn->at;
4566 		}
4567 	}
4568 
4569 	if (i < MEMCG_CGWB_FRN_CNT) {
4570 		/*
4571 		 * Re-using an existing one.  Update timestamp lazily to
4572 		 * avoid making the cacheline hot.  We want them to be
4573 		 * reasonably up-to-date and significantly shorter than
4574 		 * dirty_expire_interval as that's what expires the record.
4575 		 * Use the shorter of 1s and dirty_expire_interval / 8.
4576 		 */
4577 		unsigned long update_intv =
4578 			min_t(unsigned long, HZ,
4579 			      msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4580 
4581 		if (time_before64(frn->at, now - update_intv))
4582 			frn->at = now;
4583 	} else if (oldest >= 0) {
4584 		/* replace the oldest free one */
4585 		frn = &memcg->cgwb_frn[oldest];
4586 		frn->bdi_id = wb->bdi->id;
4587 		frn->memcg_id = wb->memcg_css->id;
4588 		frn->at = now;
4589 	}
4590 }
4591 
4592 /* issue foreign writeback flushes for recorded foreign dirtying events */
4593 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4594 {
4595 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4596 	unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4597 	u64 now = jiffies_64;
4598 	int i;
4599 
4600 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4601 		struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4602 
4603 		/*
4604 		 * If the record is older than dirty_expire_interval,
4605 		 * writeback on it has already started.  No need to kick it
4606 		 * off again.  Also, don't start a new one if there's
4607 		 * already one in flight.
4608 		 */
4609 		if (time_after64(frn->at, now - intv) &&
4610 		    atomic_read(&frn->done.cnt) == 1) {
4611 			frn->at = 0;
4612 			trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4613 			cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4614 					       WB_REASON_FOREIGN_FLUSH,
4615 					       &frn->done);
4616 		}
4617 	}
4618 }
4619 
4620 #else	/* CONFIG_CGROUP_WRITEBACK */
4621 
4622 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4623 {
4624 	return 0;
4625 }
4626 
4627 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4628 {
4629 }
4630 
4631 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4632 {
4633 }
4634 
4635 #endif	/* CONFIG_CGROUP_WRITEBACK */
4636 
4637 /*
4638  * DO NOT USE IN NEW FILES.
4639  *
4640  * "cgroup.event_control" implementation.
4641  *
4642  * This is way over-engineered.  It tries to support fully configurable
4643  * events for each user.  Such level of flexibility is completely
4644  * unnecessary especially in the light of the planned unified hierarchy.
4645  *
4646  * Please deprecate this and replace with something simpler if at all
4647  * possible.
4648  */
4649 
4650 /*
4651  * Unregister event and free resources.
4652  *
4653  * Gets called from workqueue.
4654  */
4655 static void memcg_event_remove(struct work_struct *work)
4656 {
4657 	struct mem_cgroup_event *event =
4658 		container_of(work, struct mem_cgroup_event, remove);
4659 	struct mem_cgroup *memcg = event->memcg;
4660 
4661 	remove_wait_queue(event->wqh, &event->wait);
4662 
4663 	event->unregister_event(memcg, event->eventfd);
4664 
4665 	/* Notify userspace the event is going away. */
4666 	eventfd_signal(event->eventfd, 1);
4667 
4668 	eventfd_ctx_put(event->eventfd);
4669 	kfree(event);
4670 	css_put(&memcg->css);
4671 }
4672 
4673 /*
4674  * Gets called on EPOLLHUP on eventfd when user closes it.
4675  *
4676  * Called with wqh->lock held and interrupts disabled.
4677  */
4678 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4679 			    int sync, void *key)
4680 {
4681 	struct mem_cgroup_event *event =
4682 		container_of(wait, struct mem_cgroup_event, wait);
4683 	struct mem_cgroup *memcg = event->memcg;
4684 	__poll_t flags = key_to_poll(key);
4685 
4686 	if (flags & EPOLLHUP) {
4687 		/*
4688 		 * If the event has been detached at cgroup removal, we
4689 		 * can simply return knowing the other side will cleanup
4690 		 * for us.
4691 		 *
4692 		 * We can't race against event freeing since the other
4693 		 * side will require wqh->lock via remove_wait_queue(),
4694 		 * which we hold.
4695 		 */
4696 		spin_lock(&memcg->event_list_lock);
4697 		if (!list_empty(&event->list)) {
4698 			list_del_init(&event->list);
4699 			/*
4700 			 * We are in atomic context, but cgroup_event_remove()
4701 			 * may sleep, so we have to call it in workqueue.
4702 			 */
4703 			schedule_work(&event->remove);
4704 		}
4705 		spin_unlock(&memcg->event_list_lock);
4706 	}
4707 
4708 	return 0;
4709 }
4710 
4711 static void memcg_event_ptable_queue_proc(struct file *file,
4712 		wait_queue_head_t *wqh, poll_table *pt)
4713 {
4714 	struct mem_cgroup_event *event =
4715 		container_of(pt, struct mem_cgroup_event, pt);
4716 
4717 	event->wqh = wqh;
4718 	add_wait_queue(wqh, &event->wait);
4719 }
4720 
4721 /*
4722  * DO NOT USE IN NEW FILES.
4723  *
4724  * Parse input and register new cgroup event handler.
4725  *
4726  * Input must be in format '<event_fd> <control_fd> <args>'.
4727  * Interpretation of args is defined by control file implementation.
4728  */
4729 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4730 					 char *buf, size_t nbytes, loff_t off)
4731 {
4732 	struct cgroup_subsys_state *css = of_css(of);
4733 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4734 	struct mem_cgroup_event *event;
4735 	struct cgroup_subsys_state *cfile_css;
4736 	unsigned int efd, cfd;
4737 	struct fd efile;
4738 	struct fd cfile;
4739 	const char *name;
4740 	char *endp;
4741 	int ret;
4742 
4743 	buf = strstrip(buf);
4744 
4745 	efd = simple_strtoul(buf, &endp, 10);
4746 	if (*endp != ' ')
4747 		return -EINVAL;
4748 	buf = endp + 1;
4749 
4750 	cfd = simple_strtoul(buf, &endp, 10);
4751 	if ((*endp != ' ') && (*endp != '\0'))
4752 		return -EINVAL;
4753 	buf = endp + 1;
4754 
4755 	event = kzalloc(sizeof(*event), GFP_KERNEL);
4756 	if (!event)
4757 		return -ENOMEM;
4758 
4759 	event->memcg = memcg;
4760 	INIT_LIST_HEAD(&event->list);
4761 	init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4762 	init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4763 	INIT_WORK(&event->remove, memcg_event_remove);
4764 
4765 	efile = fdget(efd);
4766 	if (!efile.file) {
4767 		ret = -EBADF;
4768 		goto out_kfree;
4769 	}
4770 
4771 	event->eventfd = eventfd_ctx_fileget(efile.file);
4772 	if (IS_ERR(event->eventfd)) {
4773 		ret = PTR_ERR(event->eventfd);
4774 		goto out_put_efile;
4775 	}
4776 
4777 	cfile = fdget(cfd);
4778 	if (!cfile.file) {
4779 		ret = -EBADF;
4780 		goto out_put_eventfd;
4781 	}
4782 
4783 	/* the process need read permission on control file */
4784 	/* AV: shouldn't we check that it's been opened for read instead? */
4785 	ret = file_permission(cfile.file, MAY_READ);
4786 	if (ret < 0)
4787 		goto out_put_cfile;
4788 
4789 	/*
4790 	 * Determine the event callbacks and set them in @event.  This used
4791 	 * to be done via struct cftype but cgroup core no longer knows
4792 	 * about these events.  The following is crude but the whole thing
4793 	 * is for compatibility anyway.
4794 	 *
4795 	 * DO NOT ADD NEW FILES.
4796 	 */
4797 	name = cfile.file->f_path.dentry->d_name.name;
4798 
4799 	if (!strcmp(name, "memory.usage_in_bytes")) {
4800 		event->register_event = mem_cgroup_usage_register_event;
4801 		event->unregister_event = mem_cgroup_usage_unregister_event;
4802 	} else if (!strcmp(name, "memory.oom_control")) {
4803 		event->register_event = mem_cgroup_oom_register_event;
4804 		event->unregister_event = mem_cgroup_oom_unregister_event;
4805 	} else if (!strcmp(name, "memory.pressure_level")) {
4806 		event->register_event = vmpressure_register_event;
4807 		event->unregister_event = vmpressure_unregister_event;
4808 	} else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4809 		event->register_event = memsw_cgroup_usage_register_event;
4810 		event->unregister_event = memsw_cgroup_usage_unregister_event;
4811 	} else {
4812 		ret = -EINVAL;
4813 		goto out_put_cfile;
4814 	}
4815 
4816 	/*
4817 	 * Verify @cfile should belong to @css.  Also, remaining events are
4818 	 * automatically removed on cgroup destruction but the removal is
4819 	 * asynchronous, so take an extra ref on @css.
4820 	 */
4821 	cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4822 					       &memory_cgrp_subsys);
4823 	ret = -EINVAL;
4824 	if (IS_ERR(cfile_css))
4825 		goto out_put_cfile;
4826 	if (cfile_css != css) {
4827 		css_put(cfile_css);
4828 		goto out_put_cfile;
4829 	}
4830 
4831 	ret = event->register_event(memcg, event->eventfd, buf);
4832 	if (ret)
4833 		goto out_put_css;
4834 
4835 	vfs_poll(efile.file, &event->pt);
4836 
4837 	spin_lock_irq(&memcg->event_list_lock);
4838 	list_add(&event->list, &memcg->event_list);
4839 	spin_unlock_irq(&memcg->event_list_lock);
4840 
4841 	fdput(cfile);
4842 	fdput(efile);
4843 
4844 	return nbytes;
4845 
4846 out_put_css:
4847 	css_put(css);
4848 out_put_cfile:
4849 	fdput(cfile);
4850 out_put_eventfd:
4851 	eventfd_ctx_put(event->eventfd);
4852 out_put_efile:
4853 	fdput(efile);
4854 out_kfree:
4855 	kfree(event);
4856 
4857 	return ret;
4858 }
4859 
4860 static struct cftype mem_cgroup_legacy_files[] = {
4861 	{
4862 		.name = "usage_in_bytes",
4863 		.private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4864 		.read_u64 = mem_cgroup_read_u64,
4865 	},
4866 	{
4867 		.name = "max_usage_in_bytes",
4868 		.private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4869 		.write = mem_cgroup_reset,
4870 		.read_u64 = mem_cgroup_read_u64,
4871 	},
4872 	{
4873 		.name = "limit_in_bytes",
4874 		.private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4875 		.write = mem_cgroup_write,
4876 		.read_u64 = mem_cgroup_read_u64,
4877 	},
4878 	{
4879 		.name = "soft_limit_in_bytes",
4880 		.private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4881 		.write = mem_cgroup_write,
4882 		.read_u64 = mem_cgroup_read_u64,
4883 	},
4884 	{
4885 		.name = "failcnt",
4886 		.private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4887 		.write = mem_cgroup_reset,
4888 		.read_u64 = mem_cgroup_read_u64,
4889 	},
4890 	{
4891 		.name = "stat",
4892 		.seq_show = memcg_stat_show,
4893 	},
4894 	{
4895 		.name = "force_empty",
4896 		.write = mem_cgroup_force_empty_write,
4897 	},
4898 	{
4899 		.name = "use_hierarchy",
4900 		.write_u64 = mem_cgroup_hierarchy_write,
4901 		.read_u64 = mem_cgroup_hierarchy_read,
4902 	},
4903 	{
4904 		.name = "cgroup.event_control",		/* XXX: for compat */
4905 		.write = memcg_write_event_control,
4906 		.flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4907 	},
4908 	{
4909 		.name = "swappiness",
4910 		.read_u64 = mem_cgroup_swappiness_read,
4911 		.write_u64 = mem_cgroup_swappiness_write,
4912 	},
4913 	{
4914 		.name = "move_charge_at_immigrate",
4915 		.read_u64 = mem_cgroup_move_charge_read,
4916 		.write_u64 = mem_cgroup_move_charge_write,
4917 	},
4918 	{
4919 		.name = "oom_control",
4920 		.seq_show = mem_cgroup_oom_control_read,
4921 		.write_u64 = mem_cgroup_oom_control_write,
4922 		.private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4923 	},
4924 	{
4925 		.name = "pressure_level",
4926 	},
4927 #ifdef CONFIG_NUMA
4928 	{
4929 		.name = "numa_stat",
4930 		.seq_show = memcg_numa_stat_show,
4931 	},
4932 #endif
4933 	{
4934 		.name = "kmem.limit_in_bytes",
4935 		.private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4936 		.write = mem_cgroup_write,
4937 		.read_u64 = mem_cgroup_read_u64,
4938 	},
4939 	{
4940 		.name = "kmem.usage_in_bytes",
4941 		.private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4942 		.read_u64 = mem_cgroup_read_u64,
4943 	},
4944 	{
4945 		.name = "kmem.failcnt",
4946 		.private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4947 		.write = mem_cgroup_reset,
4948 		.read_u64 = mem_cgroup_read_u64,
4949 	},
4950 	{
4951 		.name = "kmem.max_usage_in_bytes",
4952 		.private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4953 		.write = mem_cgroup_reset,
4954 		.read_u64 = mem_cgroup_read_u64,
4955 	},
4956 #if defined(CONFIG_MEMCG_KMEM) && \
4957 	(defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4958 	{
4959 		.name = "kmem.slabinfo",
4960 		.seq_show = memcg_slab_show,
4961 	},
4962 #endif
4963 	{
4964 		.name = "kmem.tcp.limit_in_bytes",
4965 		.private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4966 		.write = mem_cgroup_write,
4967 		.read_u64 = mem_cgroup_read_u64,
4968 	},
4969 	{
4970 		.name = "kmem.tcp.usage_in_bytes",
4971 		.private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4972 		.read_u64 = mem_cgroup_read_u64,
4973 	},
4974 	{
4975 		.name = "kmem.tcp.failcnt",
4976 		.private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4977 		.write = mem_cgroup_reset,
4978 		.read_u64 = mem_cgroup_read_u64,
4979 	},
4980 	{
4981 		.name = "kmem.tcp.max_usage_in_bytes",
4982 		.private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4983 		.write = mem_cgroup_reset,
4984 		.read_u64 = mem_cgroup_read_u64,
4985 	},
4986 	{ },	/* terminate */
4987 };
4988 
4989 /*
4990  * Private memory cgroup IDR
4991  *
4992  * Swap-out records and page cache shadow entries need to store memcg
4993  * references in constrained space, so we maintain an ID space that is
4994  * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4995  * memory-controlled cgroups to 64k.
4996  *
4997  * However, there usually are many references to the offline CSS after
4998  * the cgroup has been destroyed, such as page cache or reclaimable
4999  * slab objects, that don't need to hang on to the ID. We want to keep
5000  * those dead CSS from occupying IDs, or we might quickly exhaust the
5001  * relatively small ID space and prevent the creation of new cgroups
5002  * even when there are much fewer than 64k cgroups - possibly none.
5003  *
5004  * Maintain a private 16-bit ID space for memcg, and allow the ID to
5005  * be freed and recycled when it's no longer needed, which is usually
5006  * when the CSS is offlined.
5007  *
5008  * The only exception to that are records of swapped out tmpfs/shmem
5009  * pages that need to be attributed to live ancestors on swapin. But
5010  * those references are manageable from userspace.
5011  */
5012 
5013 static DEFINE_IDR(mem_cgroup_idr);
5014 
5015 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5016 {
5017 	if (memcg->id.id > 0) {
5018 		idr_remove(&mem_cgroup_idr, memcg->id.id);
5019 		memcg->id.id = 0;
5020 	}
5021 }
5022 
5023 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5024 						  unsigned int n)
5025 {
5026 	refcount_add(n, &memcg->id.ref);
5027 }
5028 
5029 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5030 {
5031 	if (refcount_sub_and_test(n, &memcg->id.ref)) {
5032 		mem_cgroup_id_remove(memcg);
5033 
5034 		/* Memcg ID pins CSS */
5035 		css_put(&memcg->css);
5036 	}
5037 }
5038 
5039 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5040 {
5041 	mem_cgroup_id_put_many(memcg, 1);
5042 }
5043 
5044 /**
5045  * mem_cgroup_from_id - look up a memcg from a memcg id
5046  * @id: the memcg id to look up
5047  *
5048  * Caller must hold rcu_read_lock().
5049  */
5050 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5051 {
5052 	WARN_ON_ONCE(!rcu_read_lock_held());
5053 	return idr_find(&mem_cgroup_idr, id);
5054 }
5055 
5056 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5057 {
5058 	struct mem_cgroup_per_node *pn;
5059 	int tmp = node;
5060 	/*
5061 	 * This routine is called against possible nodes.
5062 	 * But it's BUG to call kmalloc() against offline node.
5063 	 *
5064 	 * TODO: this routine can waste much memory for nodes which will
5065 	 *       never be onlined. It's better to use memory hotplug callback
5066 	 *       function.
5067 	 */
5068 	if (!node_state(node, N_NORMAL_MEMORY))
5069 		tmp = -1;
5070 	pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5071 	if (!pn)
5072 		return 1;
5073 
5074 	pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5075 						   GFP_KERNEL_ACCOUNT);
5076 	if (!pn->lruvec_stats_percpu) {
5077 		kfree(pn);
5078 		return 1;
5079 	}
5080 
5081 	lruvec_init(&pn->lruvec);
5082 	pn->usage_in_excess = 0;
5083 	pn->on_tree = false;
5084 	pn->memcg = memcg;
5085 
5086 	memcg->nodeinfo[node] = pn;
5087 	return 0;
5088 }
5089 
5090 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5091 {
5092 	struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5093 
5094 	if (!pn)
5095 		return;
5096 
5097 	free_percpu(pn->lruvec_stats_percpu);
5098 	kfree(pn);
5099 }
5100 
5101 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5102 {
5103 	int node;
5104 
5105 	for_each_node(node)
5106 		free_mem_cgroup_per_node_info(memcg, node);
5107 	free_percpu(memcg->vmstats_percpu);
5108 	kfree(memcg);
5109 }
5110 
5111 static void mem_cgroup_free(struct mem_cgroup *memcg)
5112 {
5113 	memcg_wb_domain_exit(memcg);
5114 	__mem_cgroup_free(memcg);
5115 }
5116 
5117 static struct mem_cgroup *mem_cgroup_alloc(void)
5118 {
5119 	struct mem_cgroup *memcg;
5120 	unsigned int size;
5121 	int node;
5122 	int __maybe_unused i;
5123 	long error = -ENOMEM;
5124 
5125 	size = sizeof(struct mem_cgroup);
5126 	size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5127 
5128 	memcg = kzalloc(size, GFP_KERNEL);
5129 	if (!memcg)
5130 		return ERR_PTR(error);
5131 
5132 	memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5133 				 1, MEM_CGROUP_ID_MAX,
5134 				 GFP_KERNEL);
5135 	if (memcg->id.id < 0) {
5136 		error = memcg->id.id;
5137 		goto fail;
5138 	}
5139 
5140 	memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5141 						 GFP_KERNEL_ACCOUNT);
5142 	if (!memcg->vmstats_percpu)
5143 		goto fail;
5144 
5145 	for_each_node(node)
5146 		if (alloc_mem_cgroup_per_node_info(memcg, node))
5147 			goto fail;
5148 
5149 	if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5150 		goto fail;
5151 
5152 	INIT_WORK(&memcg->high_work, high_work_func);
5153 	INIT_LIST_HEAD(&memcg->oom_notify);
5154 	mutex_init(&memcg->thresholds_lock);
5155 	spin_lock_init(&memcg->move_lock);
5156 	vmpressure_init(&memcg->vmpressure);
5157 	INIT_LIST_HEAD(&memcg->event_list);
5158 	spin_lock_init(&memcg->event_list_lock);
5159 	memcg->socket_pressure = jiffies;
5160 #ifdef CONFIG_MEMCG_KMEM
5161 	memcg->kmemcg_id = -1;
5162 	INIT_LIST_HEAD(&memcg->objcg_list);
5163 #endif
5164 #ifdef CONFIG_CGROUP_WRITEBACK
5165 	INIT_LIST_HEAD(&memcg->cgwb_list);
5166 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5167 		memcg->cgwb_frn[i].done =
5168 			__WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5169 #endif
5170 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5171 	spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5172 	INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5173 	memcg->deferred_split_queue.split_queue_len = 0;
5174 #endif
5175 	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5176 	return memcg;
5177 fail:
5178 	mem_cgroup_id_remove(memcg);
5179 	__mem_cgroup_free(memcg);
5180 	return ERR_PTR(error);
5181 }
5182 
5183 static struct cgroup_subsys_state * __ref
5184 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5185 {
5186 	struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5187 	struct mem_cgroup *memcg, *old_memcg;
5188 	long error = -ENOMEM;
5189 
5190 	old_memcg = set_active_memcg(parent);
5191 	memcg = mem_cgroup_alloc();
5192 	set_active_memcg(old_memcg);
5193 	if (IS_ERR(memcg))
5194 		return ERR_CAST(memcg);
5195 
5196 	page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5197 	memcg->soft_limit = PAGE_COUNTER_MAX;
5198 	page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5199 	if (parent) {
5200 		memcg->swappiness = mem_cgroup_swappiness(parent);
5201 		memcg->oom_kill_disable = parent->oom_kill_disable;
5202 
5203 		page_counter_init(&memcg->memory, &parent->memory);
5204 		page_counter_init(&memcg->swap, &parent->swap);
5205 		page_counter_init(&memcg->kmem, &parent->kmem);
5206 		page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5207 	} else {
5208 		page_counter_init(&memcg->memory, NULL);
5209 		page_counter_init(&memcg->swap, NULL);
5210 		page_counter_init(&memcg->kmem, NULL);
5211 		page_counter_init(&memcg->tcpmem, NULL);
5212 
5213 		root_mem_cgroup = memcg;
5214 		return &memcg->css;
5215 	}
5216 
5217 	/* The following stuff does not apply to the root */
5218 	error = memcg_online_kmem(memcg);
5219 	if (error)
5220 		goto fail;
5221 
5222 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5223 		static_branch_inc(&memcg_sockets_enabled_key);
5224 
5225 	return &memcg->css;
5226 fail:
5227 	mem_cgroup_id_remove(memcg);
5228 	mem_cgroup_free(memcg);
5229 	return ERR_PTR(error);
5230 }
5231 
5232 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5233 {
5234 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5235 
5236 	/*
5237 	 * A memcg must be visible for expand_shrinker_info()
5238 	 * by the time the maps are allocated. So, we allocate maps
5239 	 * here, when for_each_mem_cgroup() can't skip it.
5240 	 */
5241 	if (alloc_shrinker_info(memcg)) {
5242 		mem_cgroup_id_remove(memcg);
5243 		return -ENOMEM;
5244 	}
5245 
5246 	/* Online state pins memcg ID, memcg ID pins CSS */
5247 	refcount_set(&memcg->id.ref, 1);
5248 	css_get(css);
5249 
5250 	if (unlikely(mem_cgroup_is_root(memcg)))
5251 		queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5252 				   2UL*HZ);
5253 	return 0;
5254 }
5255 
5256 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5257 {
5258 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5259 	struct mem_cgroup_event *event, *tmp;
5260 
5261 	/*
5262 	 * Unregister events and notify userspace.
5263 	 * Notify userspace about cgroup removing only after rmdir of cgroup
5264 	 * directory to avoid race between userspace and kernelspace.
5265 	 */
5266 	spin_lock_irq(&memcg->event_list_lock);
5267 	list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5268 		list_del_init(&event->list);
5269 		schedule_work(&event->remove);
5270 	}
5271 	spin_unlock_irq(&memcg->event_list_lock);
5272 
5273 	page_counter_set_min(&memcg->memory, 0);
5274 	page_counter_set_low(&memcg->memory, 0);
5275 
5276 	memcg_offline_kmem(memcg);
5277 	reparent_shrinker_deferred(memcg);
5278 	wb_memcg_offline(memcg);
5279 
5280 	drain_all_stock(memcg);
5281 
5282 	mem_cgroup_id_put(memcg);
5283 }
5284 
5285 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5286 {
5287 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5288 
5289 	invalidate_reclaim_iterators(memcg);
5290 }
5291 
5292 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5293 {
5294 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5295 	int __maybe_unused i;
5296 
5297 #ifdef CONFIG_CGROUP_WRITEBACK
5298 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5299 		wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5300 #endif
5301 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5302 		static_branch_dec(&memcg_sockets_enabled_key);
5303 
5304 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5305 		static_branch_dec(&memcg_sockets_enabled_key);
5306 
5307 	vmpressure_cleanup(&memcg->vmpressure);
5308 	cancel_work_sync(&memcg->high_work);
5309 	mem_cgroup_remove_from_trees(memcg);
5310 	free_shrinker_info(memcg);
5311 	memcg_free_kmem(memcg);
5312 	mem_cgroup_free(memcg);
5313 }
5314 
5315 /**
5316  * mem_cgroup_css_reset - reset the states of a mem_cgroup
5317  * @css: the target css
5318  *
5319  * Reset the states of the mem_cgroup associated with @css.  This is
5320  * invoked when the userland requests disabling on the default hierarchy
5321  * but the memcg is pinned through dependency.  The memcg should stop
5322  * applying policies and should revert to the vanilla state as it may be
5323  * made visible again.
5324  *
5325  * The current implementation only resets the essential configurations.
5326  * This needs to be expanded to cover all the visible parts.
5327  */
5328 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5329 {
5330 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5331 
5332 	page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5333 	page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5334 	page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5335 	page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5336 	page_counter_set_min(&memcg->memory, 0);
5337 	page_counter_set_low(&memcg->memory, 0);
5338 	page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5339 	memcg->soft_limit = PAGE_COUNTER_MAX;
5340 	page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5341 	memcg_wb_domain_size_changed(memcg);
5342 }
5343 
5344 void mem_cgroup_flush_stats(void)
5345 {
5346 	if (!spin_trylock(&stats_flush_lock))
5347 		return;
5348 
5349 	cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup);
5350 	spin_unlock(&stats_flush_lock);
5351 }
5352 
5353 static void flush_memcg_stats_dwork(struct work_struct *w)
5354 {
5355 	mem_cgroup_flush_stats();
5356 	queue_delayed_work(system_unbound_wq, &stats_flush_dwork, 2UL*HZ);
5357 }
5358 
5359 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5360 {
5361 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5362 	struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5363 	struct memcg_vmstats_percpu *statc;
5364 	long delta, v;
5365 	int i, nid;
5366 
5367 	statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5368 
5369 	for (i = 0; i < MEMCG_NR_STAT; i++) {
5370 		/*
5371 		 * Collect the aggregated propagation counts of groups
5372 		 * below us. We're in a per-cpu loop here and this is
5373 		 * a global counter, so the first cycle will get them.
5374 		 */
5375 		delta = memcg->vmstats.state_pending[i];
5376 		if (delta)
5377 			memcg->vmstats.state_pending[i] = 0;
5378 
5379 		/* Add CPU changes on this level since the last flush */
5380 		v = READ_ONCE(statc->state[i]);
5381 		if (v != statc->state_prev[i]) {
5382 			delta += v - statc->state_prev[i];
5383 			statc->state_prev[i] = v;
5384 		}
5385 
5386 		if (!delta)
5387 			continue;
5388 
5389 		/* Aggregate counts on this level and propagate upwards */
5390 		memcg->vmstats.state[i] += delta;
5391 		if (parent)
5392 			parent->vmstats.state_pending[i] += delta;
5393 	}
5394 
5395 	for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5396 		delta = memcg->vmstats.events_pending[i];
5397 		if (delta)
5398 			memcg->vmstats.events_pending[i] = 0;
5399 
5400 		v = READ_ONCE(statc->events[i]);
5401 		if (v != statc->events_prev[i]) {
5402 			delta += v - statc->events_prev[i];
5403 			statc->events_prev[i] = v;
5404 		}
5405 
5406 		if (!delta)
5407 			continue;
5408 
5409 		memcg->vmstats.events[i] += delta;
5410 		if (parent)
5411 			parent->vmstats.events_pending[i] += delta;
5412 	}
5413 
5414 	for_each_node_state(nid, N_MEMORY) {
5415 		struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5416 		struct mem_cgroup_per_node *ppn = NULL;
5417 		struct lruvec_stats_percpu *lstatc;
5418 
5419 		if (parent)
5420 			ppn = parent->nodeinfo[nid];
5421 
5422 		lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5423 
5424 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5425 			delta = pn->lruvec_stats.state_pending[i];
5426 			if (delta)
5427 				pn->lruvec_stats.state_pending[i] = 0;
5428 
5429 			v = READ_ONCE(lstatc->state[i]);
5430 			if (v != lstatc->state_prev[i]) {
5431 				delta += v - lstatc->state_prev[i];
5432 				lstatc->state_prev[i] = v;
5433 			}
5434 
5435 			if (!delta)
5436 				continue;
5437 
5438 			pn->lruvec_stats.state[i] += delta;
5439 			if (ppn)
5440 				ppn->lruvec_stats.state_pending[i] += delta;
5441 		}
5442 	}
5443 }
5444 
5445 #ifdef CONFIG_MMU
5446 /* Handlers for move charge at task migration. */
5447 static int mem_cgroup_do_precharge(unsigned long count)
5448 {
5449 	int ret;
5450 
5451 	/* Try a single bulk charge without reclaim first, kswapd may wake */
5452 	ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5453 	if (!ret) {
5454 		mc.precharge += count;
5455 		return ret;
5456 	}
5457 
5458 	/* Try charges one by one with reclaim, but do not retry */
5459 	while (count--) {
5460 		ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5461 		if (ret)
5462 			return ret;
5463 		mc.precharge++;
5464 		cond_resched();
5465 	}
5466 	return 0;
5467 }
5468 
5469 union mc_target {
5470 	struct page	*page;
5471 	swp_entry_t	ent;
5472 };
5473 
5474 enum mc_target_type {
5475 	MC_TARGET_NONE = 0,
5476 	MC_TARGET_PAGE,
5477 	MC_TARGET_SWAP,
5478 	MC_TARGET_DEVICE,
5479 };
5480 
5481 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5482 						unsigned long addr, pte_t ptent)
5483 {
5484 	struct page *page = vm_normal_page(vma, addr, ptent);
5485 
5486 	if (!page || !page_mapped(page))
5487 		return NULL;
5488 	if (PageAnon(page)) {
5489 		if (!(mc.flags & MOVE_ANON))
5490 			return NULL;
5491 	} else {
5492 		if (!(mc.flags & MOVE_FILE))
5493 			return NULL;
5494 	}
5495 	if (!get_page_unless_zero(page))
5496 		return NULL;
5497 
5498 	return page;
5499 }
5500 
5501 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5502 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5503 			pte_t ptent, swp_entry_t *entry)
5504 {
5505 	struct page *page = NULL;
5506 	swp_entry_t ent = pte_to_swp_entry(ptent);
5507 
5508 	if (!(mc.flags & MOVE_ANON))
5509 		return NULL;
5510 
5511 	/*
5512 	 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5513 	 * a device and because they are not accessible by CPU they are store
5514 	 * as special swap entry in the CPU page table.
5515 	 */
5516 	if (is_device_private_entry(ent)) {
5517 		page = pfn_swap_entry_to_page(ent);
5518 		/*
5519 		 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5520 		 * a refcount of 1 when free (unlike normal page)
5521 		 */
5522 		if (!page_ref_add_unless(page, 1, 1))
5523 			return NULL;
5524 		return page;
5525 	}
5526 
5527 	if (non_swap_entry(ent))
5528 		return NULL;
5529 
5530 	/*
5531 	 * Because lookup_swap_cache() updates some statistics counter,
5532 	 * we call find_get_page() with swapper_space directly.
5533 	 */
5534 	page = find_get_page(swap_address_space(ent), swp_offset(ent));
5535 	entry->val = ent.val;
5536 
5537 	return page;
5538 }
5539 #else
5540 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5541 			pte_t ptent, swp_entry_t *entry)
5542 {
5543 	return NULL;
5544 }
5545 #endif
5546 
5547 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5548 			unsigned long addr, pte_t ptent, swp_entry_t *entry)
5549 {
5550 	if (!vma->vm_file) /* anonymous vma */
5551 		return NULL;
5552 	if (!(mc.flags & MOVE_FILE))
5553 		return NULL;
5554 
5555 	/* page is moved even if it's not RSS of this task(page-faulted). */
5556 	/* shmem/tmpfs may report page out on swap: account for that too. */
5557 	return find_get_incore_page(vma->vm_file->f_mapping,
5558 			linear_page_index(vma, addr));
5559 }
5560 
5561 /**
5562  * mem_cgroup_move_account - move account of the page
5563  * @page: the page
5564  * @compound: charge the page as compound or small page
5565  * @from: mem_cgroup which the page is moved from.
5566  * @to:	mem_cgroup which the page is moved to. @from != @to.
5567  *
5568  * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5569  *
5570  * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5571  * from old cgroup.
5572  */
5573 static int mem_cgroup_move_account(struct page *page,
5574 				   bool compound,
5575 				   struct mem_cgroup *from,
5576 				   struct mem_cgroup *to)
5577 {
5578 	struct lruvec *from_vec, *to_vec;
5579 	struct pglist_data *pgdat;
5580 	unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5581 	int ret;
5582 
5583 	VM_BUG_ON(from == to);
5584 	VM_BUG_ON_PAGE(PageLRU(page), page);
5585 	VM_BUG_ON(compound && !PageTransHuge(page));
5586 
5587 	/*
5588 	 * Prevent mem_cgroup_migrate() from looking at
5589 	 * page's memory cgroup of its source page while we change it.
5590 	 */
5591 	ret = -EBUSY;
5592 	if (!trylock_page(page))
5593 		goto out;
5594 
5595 	ret = -EINVAL;
5596 	if (page_memcg(page) != from)
5597 		goto out_unlock;
5598 
5599 	pgdat = page_pgdat(page);
5600 	from_vec = mem_cgroup_lruvec(from, pgdat);
5601 	to_vec = mem_cgroup_lruvec(to, pgdat);
5602 
5603 	lock_page_memcg(page);
5604 
5605 	if (PageAnon(page)) {
5606 		if (page_mapped(page)) {
5607 			__mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5608 			__mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5609 			if (PageTransHuge(page)) {
5610 				__mod_lruvec_state(from_vec, NR_ANON_THPS,
5611 						   -nr_pages);
5612 				__mod_lruvec_state(to_vec, NR_ANON_THPS,
5613 						   nr_pages);
5614 			}
5615 		}
5616 	} else {
5617 		__mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5618 		__mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5619 
5620 		if (PageSwapBacked(page)) {
5621 			__mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5622 			__mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5623 		}
5624 
5625 		if (page_mapped(page)) {
5626 			__mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5627 			__mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5628 		}
5629 
5630 		if (PageDirty(page)) {
5631 			struct address_space *mapping = page_mapping(page);
5632 
5633 			if (mapping_can_writeback(mapping)) {
5634 				__mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5635 						   -nr_pages);
5636 				__mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5637 						   nr_pages);
5638 			}
5639 		}
5640 	}
5641 
5642 	if (PageWriteback(page)) {
5643 		__mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5644 		__mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5645 	}
5646 
5647 	/*
5648 	 * All state has been migrated, let's switch to the new memcg.
5649 	 *
5650 	 * It is safe to change page's memcg here because the page
5651 	 * is referenced, charged, isolated, and locked: we can't race
5652 	 * with (un)charging, migration, LRU putback, or anything else
5653 	 * that would rely on a stable page's memory cgroup.
5654 	 *
5655 	 * Note that lock_page_memcg is a memcg lock, not a page lock,
5656 	 * to save space. As soon as we switch page's memory cgroup to a
5657 	 * new memcg that isn't locked, the above state can change
5658 	 * concurrently again. Make sure we're truly done with it.
5659 	 */
5660 	smp_mb();
5661 
5662 	css_get(&to->css);
5663 	css_put(&from->css);
5664 
5665 	page->memcg_data = (unsigned long)to;
5666 
5667 	__unlock_page_memcg(from);
5668 
5669 	ret = 0;
5670 
5671 	local_irq_disable();
5672 	mem_cgroup_charge_statistics(to, page, nr_pages);
5673 	memcg_check_events(to, page);
5674 	mem_cgroup_charge_statistics(from, page, -nr_pages);
5675 	memcg_check_events(from, page);
5676 	local_irq_enable();
5677 out_unlock:
5678 	unlock_page(page);
5679 out:
5680 	return ret;
5681 }
5682 
5683 /**
5684  * get_mctgt_type - get target type of moving charge
5685  * @vma: the vma the pte to be checked belongs
5686  * @addr: the address corresponding to the pte to be checked
5687  * @ptent: the pte to be checked
5688  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5689  *
5690  * Returns
5691  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5692  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5693  *     move charge. if @target is not NULL, the page is stored in target->page
5694  *     with extra refcnt got(Callers should handle it).
5695  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5696  *     target for charge migration. if @target is not NULL, the entry is stored
5697  *     in target->ent.
5698  *   3(MC_TARGET_DEVICE): like MC_TARGET_PAGE  but page is MEMORY_DEVICE_PRIVATE
5699  *     (so ZONE_DEVICE page and thus not on the lru).
5700  *     For now we such page is charge like a regular page would be as for all
5701  *     intent and purposes it is just special memory taking the place of a
5702  *     regular page.
5703  *
5704  *     See Documentations/vm/hmm.txt and include/linux/hmm.h
5705  *
5706  * Called with pte lock held.
5707  */
5708 
5709 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5710 		unsigned long addr, pte_t ptent, union mc_target *target)
5711 {
5712 	struct page *page = NULL;
5713 	enum mc_target_type ret = MC_TARGET_NONE;
5714 	swp_entry_t ent = { .val = 0 };
5715 
5716 	if (pte_present(ptent))
5717 		page = mc_handle_present_pte(vma, addr, ptent);
5718 	else if (is_swap_pte(ptent))
5719 		page = mc_handle_swap_pte(vma, ptent, &ent);
5720 	else if (pte_none(ptent))
5721 		page = mc_handle_file_pte(vma, addr, ptent, &ent);
5722 
5723 	if (!page && !ent.val)
5724 		return ret;
5725 	if (page) {
5726 		/*
5727 		 * Do only loose check w/o serialization.
5728 		 * mem_cgroup_move_account() checks the page is valid or
5729 		 * not under LRU exclusion.
5730 		 */
5731 		if (page_memcg(page) == mc.from) {
5732 			ret = MC_TARGET_PAGE;
5733 			if (is_device_private_page(page))
5734 				ret = MC_TARGET_DEVICE;
5735 			if (target)
5736 				target->page = page;
5737 		}
5738 		if (!ret || !target)
5739 			put_page(page);
5740 	}
5741 	/*
5742 	 * There is a swap entry and a page doesn't exist or isn't charged.
5743 	 * But we cannot move a tail-page in a THP.
5744 	 */
5745 	if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5746 	    mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5747 		ret = MC_TARGET_SWAP;
5748 		if (target)
5749 			target->ent = ent;
5750 	}
5751 	return ret;
5752 }
5753 
5754 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5755 /*
5756  * We don't consider PMD mapped swapping or file mapped pages because THP does
5757  * not support them for now.
5758  * Caller should make sure that pmd_trans_huge(pmd) is true.
5759  */
5760 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5761 		unsigned long addr, pmd_t pmd, union mc_target *target)
5762 {
5763 	struct page *page = NULL;
5764 	enum mc_target_type ret = MC_TARGET_NONE;
5765 
5766 	if (unlikely(is_swap_pmd(pmd))) {
5767 		VM_BUG_ON(thp_migration_supported() &&
5768 				  !is_pmd_migration_entry(pmd));
5769 		return ret;
5770 	}
5771 	page = pmd_page(pmd);
5772 	VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5773 	if (!(mc.flags & MOVE_ANON))
5774 		return ret;
5775 	if (page_memcg(page) == mc.from) {
5776 		ret = MC_TARGET_PAGE;
5777 		if (target) {
5778 			get_page(page);
5779 			target->page = page;
5780 		}
5781 	}
5782 	return ret;
5783 }
5784 #else
5785 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5786 		unsigned long addr, pmd_t pmd, union mc_target *target)
5787 {
5788 	return MC_TARGET_NONE;
5789 }
5790 #endif
5791 
5792 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5793 					unsigned long addr, unsigned long end,
5794 					struct mm_walk *walk)
5795 {
5796 	struct vm_area_struct *vma = walk->vma;
5797 	pte_t *pte;
5798 	spinlock_t *ptl;
5799 
5800 	ptl = pmd_trans_huge_lock(pmd, vma);
5801 	if (ptl) {
5802 		/*
5803 		 * Note their can not be MC_TARGET_DEVICE for now as we do not
5804 		 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5805 		 * this might change.
5806 		 */
5807 		if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5808 			mc.precharge += HPAGE_PMD_NR;
5809 		spin_unlock(ptl);
5810 		return 0;
5811 	}
5812 
5813 	if (pmd_trans_unstable(pmd))
5814 		return 0;
5815 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5816 	for (; addr != end; pte++, addr += PAGE_SIZE)
5817 		if (get_mctgt_type(vma, addr, *pte, NULL))
5818 			mc.precharge++;	/* increment precharge temporarily */
5819 	pte_unmap_unlock(pte - 1, ptl);
5820 	cond_resched();
5821 
5822 	return 0;
5823 }
5824 
5825 static const struct mm_walk_ops precharge_walk_ops = {
5826 	.pmd_entry	= mem_cgroup_count_precharge_pte_range,
5827 };
5828 
5829 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5830 {
5831 	unsigned long precharge;
5832 
5833 	mmap_read_lock(mm);
5834 	walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5835 	mmap_read_unlock(mm);
5836 
5837 	precharge = mc.precharge;
5838 	mc.precharge = 0;
5839 
5840 	return precharge;
5841 }
5842 
5843 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5844 {
5845 	unsigned long precharge = mem_cgroup_count_precharge(mm);
5846 
5847 	VM_BUG_ON(mc.moving_task);
5848 	mc.moving_task = current;
5849 	return mem_cgroup_do_precharge(precharge);
5850 }
5851 
5852 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5853 static void __mem_cgroup_clear_mc(void)
5854 {
5855 	struct mem_cgroup *from = mc.from;
5856 	struct mem_cgroup *to = mc.to;
5857 
5858 	/* we must uncharge all the leftover precharges from mc.to */
5859 	if (mc.precharge) {
5860 		cancel_charge(mc.to, mc.precharge);
5861 		mc.precharge = 0;
5862 	}
5863 	/*
5864 	 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5865 	 * we must uncharge here.
5866 	 */
5867 	if (mc.moved_charge) {
5868 		cancel_charge(mc.from, mc.moved_charge);
5869 		mc.moved_charge = 0;
5870 	}
5871 	/* we must fixup refcnts and charges */
5872 	if (mc.moved_swap) {
5873 		/* uncharge swap account from the old cgroup */
5874 		if (!mem_cgroup_is_root(mc.from))
5875 			page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5876 
5877 		mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5878 
5879 		/*
5880 		 * we charged both to->memory and to->memsw, so we
5881 		 * should uncharge to->memory.
5882 		 */
5883 		if (!mem_cgroup_is_root(mc.to))
5884 			page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5885 
5886 		mc.moved_swap = 0;
5887 	}
5888 	memcg_oom_recover(from);
5889 	memcg_oom_recover(to);
5890 	wake_up_all(&mc.waitq);
5891 }
5892 
5893 static void mem_cgroup_clear_mc(void)
5894 {
5895 	struct mm_struct *mm = mc.mm;
5896 
5897 	/*
5898 	 * we must clear moving_task before waking up waiters at the end of
5899 	 * task migration.
5900 	 */
5901 	mc.moving_task = NULL;
5902 	__mem_cgroup_clear_mc();
5903 	spin_lock(&mc.lock);
5904 	mc.from = NULL;
5905 	mc.to = NULL;
5906 	mc.mm = NULL;
5907 	spin_unlock(&mc.lock);
5908 
5909 	mmput(mm);
5910 }
5911 
5912 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5913 {
5914 	struct cgroup_subsys_state *css;
5915 	struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5916 	struct mem_cgroup *from;
5917 	struct task_struct *leader, *p;
5918 	struct mm_struct *mm;
5919 	unsigned long move_flags;
5920 	int ret = 0;
5921 
5922 	/* charge immigration isn't supported on the default hierarchy */
5923 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5924 		return 0;
5925 
5926 	/*
5927 	 * Multi-process migrations only happen on the default hierarchy
5928 	 * where charge immigration is not used.  Perform charge
5929 	 * immigration if @tset contains a leader and whine if there are
5930 	 * multiple.
5931 	 */
5932 	p = NULL;
5933 	cgroup_taskset_for_each_leader(leader, css, tset) {
5934 		WARN_ON_ONCE(p);
5935 		p = leader;
5936 		memcg = mem_cgroup_from_css(css);
5937 	}
5938 	if (!p)
5939 		return 0;
5940 
5941 	/*
5942 	 * We are now committed to this value whatever it is. Changes in this
5943 	 * tunable will only affect upcoming migrations, not the current one.
5944 	 * So we need to save it, and keep it going.
5945 	 */
5946 	move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5947 	if (!move_flags)
5948 		return 0;
5949 
5950 	from = mem_cgroup_from_task(p);
5951 
5952 	VM_BUG_ON(from == memcg);
5953 
5954 	mm = get_task_mm(p);
5955 	if (!mm)
5956 		return 0;
5957 	/* We move charges only when we move a owner of the mm */
5958 	if (mm->owner == p) {
5959 		VM_BUG_ON(mc.from);
5960 		VM_BUG_ON(mc.to);
5961 		VM_BUG_ON(mc.precharge);
5962 		VM_BUG_ON(mc.moved_charge);
5963 		VM_BUG_ON(mc.moved_swap);
5964 
5965 		spin_lock(&mc.lock);
5966 		mc.mm = mm;
5967 		mc.from = from;
5968 		mc.to = memcg;
5969 		mc.flags = move_flags;
5970 		spin_unlock(&mc.lock);
5971 		/* We set mc.moving_task later */
5972 
5973 		ret = mem_cgroup_precharge_mc(mm);
5974 		if (ret)
5975 			mem_cgroup_clear_mc();
5976 	} else {
5977 		mmput(mm);
5978 	}
5979 	return ret;
5980 }
5981 
5982 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5983 {
5984 	if (mc.to)
5985 		mem_cgroup_clear_mc();
5986 }
5987 
5988 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5989 				unsigned long addr, unsigned long end,
5990 				struct mm_walk *walk)
5991 {
5992 	int ret = 0;
5993 	struct vm_area_struct *vma = walk->vma;
5994 	pte_t *pte;
5995 	spinlock_t *ptl;
5996 	enum mc_target_type target_type;
5997 	union mc_target target;
5998 	struct page *page;
5999 
6000 	ptl = pmd_trans_huge_lock(pmd, vma);
6001 	if (ptl) {
6002 		if (mc.precharge < HPAGE_PMD_NR) {
6003 			spin_unlock(ptl);
6004 			return 0;
6005 		}
6006 		target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6007 		if (target_type == MC_TARGET_PAGE) {
6008 			page = target.page;
6009 			if (!isolate_lru_page(page)) {
6010 				if (!mem_cgroup_move_account(page, true,
6011 							     mc.from, mc.to)) {
6012 					mc.precharge -= HPAGE_PMD_NR;
6013 					mc.moved_charge += HPAGE_PMD_NR;
6014 				}
6015 				putback_lru_page(page);
6016 			}
6017 			put_page(page);
6018 		} else if (target_type == MC_TARGET_DEVICE) {
6019 			page = target.page;
6020 			if (!mem_cgroup_move_account(page, true,
6021 						     mc.from, mc.to)) {
6022 				mc.precharge -= HPAGE_PMD_NR;
6023 				mc.moved_charge += HPAGE_PMD_NR;
6024 			}
6025 			put_page(page);
6026 		}
6027 		spin_unlock(ptl);
6028 		return 0;
6029 	}
6030 
6031 	if (pmd_trans_unstable(pmd))
6032 		return 0;
6033 retry:
6034 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6035 	for (; addr != end; addr += PAGE_SIZE) {
6036 		pte_t ptent = *(pte++);
6037 		bool device = false;
6038 		swp_entry_t ent;
6039 
6040 		if (!mc.precharge)
6041 			break;
6042 
6043 		switch (get_mctgt_type(vma, addr, ptent, &target)) {
6044 		case MC_TARGET_DEVICE:
6045 			device = true;
6046 			fallthrough;
6047 		case MC_TARGET_PAGE:
6048 			page = target.page;
6049 			/*
6050 			 * We can have a part of the split pmd here. Moving it
6051 			 * can be done but it would be too convoluted so simply
6052 			 * ignore such a partial THP and keep it in original
6053 			 * memcg. There should be somebody mapping the head.
6054 			 */
6055 			if (PageTransCompound(page))
6056 				goto put;
6057 			if (!device && isolate_lru_page(page))
6058 				goto put;
6059 			if (!mem_cgroup_move_account(page, false,
6060 						mc.from, mc.to)) {
6061 				mc.precharge--;
6062 				/* we uncharge from mc.from later. */
6063 				mc.moved_charge++;
6064 			}
6065 			if (!device)
6066 				putback_lru_page(page);
6067 put:			/* get_mctgt_type() gets the page */
6068 			put_page(page);
6069 			break;
6070 		case MC_TARGET_SWAP:
6071 			ent = target.ent;
6072 			if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6073 				mc.precharge--;
6074 				mem_cgroup_id_get_many(mc.to, 1);
6075 				/* we fixup other refcnts and charges later. */
6076 				mc.moved_swap++;
6077 			}
6078 			break;
6079 		default:
6080 			break;
6081 		}
6082 	}
6083 	pte_unmap_unlock(pte - 1, ptl);
6084 	cond_resched();
6085 
6086 	if (addr != end) {
6087 		/*
6088 		 * We have consumed all precharges we got in can_attach().
6089 		 * We try charge one by one, but don't do any additional
6090 		 * charges to mc.to if we have failed in charge once in attach()
6091 		 * phase.
6092 		 */
6093 		ret = mem_cgroup_do_precharge(1);
6094 		if (!ret)
6095 			goto retry;
6096 	}
6097 
6098 	return ret;
6099 }
6100 
6101 static const struct mm_walk_ops charge_walk_ops = {
6102 	.pmd_entry	= mem_cgroup_move_charge_pte_range,
6103 };
6104 
6105 static void mem_cgroup_move_charge(void)
6106 {
6107 	lru_add_drain_all();
6108 	/*
6109 	 * Signal lock_page_memcg() to take the memcg's move_lock
6110 	 * while we're moving its pages to another memcg. Then wait
6111 	 * for already started RCU-only updates to finish.
6112 	 */
6113 	atomic_inc(&mc.from->moving_account);
6114 	synchronize_rcu();
6115 retry:
6116 	if (unlikely(!mmap_read_trylock(mc.mm))) {
6117 		/*
6118 		 * Someone who are holding the mmap_lock might be waiting in
6119 		 * waitq. So we cancel all extra charges, wake up all waiters,
6120 		 * and retry. Because we cancel precharges, we might not be able
6121 		 * to move enough charges, but moving charge is a best-effort
6122 		 * feature anyway, so it wouldn't be a big problem.
6123 		 */
6124 		__mem_cgroup_clear_mc();
6125 		cond_resched();
6126 		goto retry;
6127 	}
6128 	/*
6129 	 * When we have consumed all precharges and failed in doing
6130 	 * additional charge, the page walk just aborts.
6131 	 */
6132 	walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6133 			NULL);
6134 
6135 	mmap_read_unlock(mc.mm);
6136 	atomic_dec(&mc.from->moving_account);
6137 }
6138 
6139 static void mem_cgroup_move_task(void)
6140 {
6141 	if (mc.to) {
6142 		mem_cgroup_move_charge();
6143 		mem_cgroup_clear_mc();
6144 	}
6145 }
6146 #else	/* !CONFIG_MMU */
6147 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6148 {
6149 	return 0;
6150 }
6151 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6152 {
6153 }
6154 static void mem_cgroup_move_task(void)
6155 {
6156 }
6157 #endif
6158 
6159 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6160 {
6161 	if (value == PAGE_COUNTER_MAX)
6162 		seq_puts(m, "max\n");
6163 	else
6164 		seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6165 
6166 	return 0;
6167 }
6168 
6169 static u64 memory_current_read(struct cgroup_subsys_state *css,
6170 			       struct cftype *cft)
6171 {
6172 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6173 
6174 	return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6175 }
6176 
6177 static int memory_min_show(struct seq_file *m, void *v)
6178 {
6179 	return seq_puts_memcg_tunable(m,
6180 		READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6181 }
6182 
6183 static ssize_t memory_min_write(struct kernfs_open_file *of,
6184 				char *buf, size_t nbytes, loff_t off)
6185 {
6186 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6187 	unsigned long min;
6188 	int err;
6189 
6190 	buf = strstrip(buf);
6191 	err = page_counter_memparse(buf, "max", &min);
6192 	if (err)
6193 		return err;
6194 
6195 	page_counter_set_min(&memcg->memory, min);
6196 
6197 	return nbytes;
6198 }
6199 
6200 static int memory_low_show(struct seq_file *m, void *v)
6201 {
6202 	return seq_puts_memcg_tunable(m,
6203 		READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6204 }
6205 
6206 static ssize_t memory_low_write(struct kernfs_open_file *of,
6207 				char *buf, size_t nbytes, loff_t off)
6208 {
6209 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6210 	unsigned long low;
6211 	int err;
6212 
6213 	buf = strstrip(buf);
6214 	err = page_counter_memparse(buf, "max", &low);
6215 	if (err)
6216 		return err;
6217 
6218 	page_counter_set_low(&memcg->memory, low);
6219 
6220 	return nbytes;
6221 }
6222 
6223 static int memory_high_show(struct seq_file *m, void *v)
6224 {
6225 	return seq_puts_memcg_tunable(m,
6226 		READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6227 }
6228 
6229 static ssize_t memory_high_write(struct kernfs_open_file *of,
6230 				 char *buf, size_t nbytes, loff_t off)
6231 {
6232 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6233 	unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6234 	bool drained = false;
6235 	unsigned long high;
6236 	int err;
6237 
6238 	buf = strstrip(buf);
6239 	err = page_counter_memparse(buf, "max", &high);
6240 	if (err)
6241 		return err;
6242 
6243 	page_counter_set_high(&memcg->memory, high);
6244 
6245 	for (;;) {
6246 		unsigned long nr_pages = page_counter_read(&memcg->memory);
6247 		unsigned long reclaimed;
6248 
6249 		if (nr_pages <= high)
6250 			break;
6251 
6252 		if (signal_pending(current))
6253 			break;
6254 
6255 		if (!drained) {
6256 			drain_all_stock(memcg);
6257 			drained = true;
6258 			continue;
6259 		}
6260 
6261 		reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6262 							 GFP_KERNEL, true);
6263 
6264 		if (!reclaimed && !nr_retries--)
6265 			break;
6266 	}
6267 
6268 	memcg_wb_domain_size_changed(memcg);
6269 	return nbytes;
6270 }
6271 
6272 static int memory_max_show(struct seq_file *m, void *v)
6273 {
6274 	return seq_puts_memcg_tunable(m,
6275 		READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6276 }
6277 
6278 static ssize_t memory_max_write(struct kernfs_open_file *of,
6279 				char *buf, size_t nbytes, loff_t off)
6280 {
6281 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6282 	unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6283 	bool drained = false;
6284 	unsigned long max;
6285 	int err;
6286 
6287 	buf = strstrip(buf);
6288 	err = page_counter_memparse(buf, "max", &max);
6289 	if (err)
6290 		return err;
6291 
6292 	xchg(&memcg->memory.max, max);
6293 
6294 	for (;;) {
6295 		unsigned long nr_pages = page_counter_read(&memcg->memory);
6296 
6297 		if (nr_pages <= max)
6298 			break;
6299 
6300 		if (signal_pending(current))
6301 			break;
6302 
6303 		if (!drained) {
6304 			drain_all_stock(memcg);
6305 			drained = true;
6306 			continue;
6307 		}
6308 
6309 		if (nr_reclaims) {
6310 			if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6311 							  GFP_KERNEL, true))
6312 				nr_reclaims--;
6313 			continue;
6314 		}
6315 
6316 		memcg_memory_event(memcg, MEMCG_OOM);
6317 		if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6318 			break;
6319 	}
6320 
6321 	memcg_wb_domain_size_changed(memcg);
6322 	return nbytes;
6323 }
6324 
6325 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6326 {
6327 	seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6328 	seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6329 	seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6330 	seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6331 	seq_printf(m, "oom_kill %lu\n",
6332 		   atomic_long_read(&events[MEMCG_OOM_KILL]));
6333 }
6334 
6335 static int memory_events_show(struct seq_file *m, void *v)
6336 {
6337 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6338 
6339 	__memory_events_show(m, memcg->memory_events);
6340 	return 0;
6341 }
6342 
6343 static int memory_events_local_show(struct seq_file *m, void *v)
6344 {
6345 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6346 
6347 	__memory_events_show(m, memcg->memory_events_local);
6348 	return 0;
6349 }
6350 
6351 static int memory_stat_show(struct seq_file *m, void *v)
6352 {
6353 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6354 	char *buf;
6355 
6356 	buf = memory_stat_format(memcg);
6357 	if (!buf)
6358 		return -ENOMEM;
6359 	seq_puts(m, buf);
6360 	kfree(buf);
6361 	return 0;
6362 }
6363 
6364 #ifdef CONFIG_NUMA
6365 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6366 						     int item)
6367 {
6368 	return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6369 }
6370 
6371 static int memory_numa_stat_show(struct seq_file *m, void *v)
6372 {
6373 	int i;
6374 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6375 
6376 	cgroup_rstat_flush(memcg->css.cgroup);
6377 
6378 	for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6379 		int nid;
6380 
6381 		if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6382 			continue;
6383 
6384 		seq_printf(m, "%s", memory_stats[i].name);
6385 		for_each_node_state(nid, N_MEMORY) {
6386 			u64 size;
6387 			struct lruvec *lruvec;
6388 
6389 			lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6390 			size = lruvec_page_state_output(lruvec,
6391 							memory_stats[i].idx);
6392 			seq_printf(m, " N%d=%llu", nid, size);
6393 		}
6394 		seq_putc(m, '\n');
6395 	}
6396 
6397 	return 0;
6398 }
6399 #endif
6400 
6401 static int memory_oom_group_show(struct seq_file *m, void *v)
6402 {
6403 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6404 
6405 	seq_printf(m, "%d\n", memcg->oom_group);
6406 
6407 	return 0;
6408 }
6409 
6410 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6411 				      char *buf, size_t nbytes, loff_t off)
6412 {
6413 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6414 	int ret, oom_group;
6415 
6416 	buf = strstrip(buf);
6417 	if (!buf)
6418 		return -EINVAL;
6419 
6420 	ret = kstrtoint(buf, 0, &oom_group);
6421 	if (ret)
6422 		return ret;
6423 
6424 	if (oom_group != 0 && oom_group != 1)
6425 		return -EINVAL;
6426 
6427 	memcg->oom_group = oom_group;
6428 
6429 	return nbytes;
6430 }
6431 
6432 static struct cftype memory_files[] = {
6433 	{
6434 		.name = "current",
6435 		.flags = CFTYPE_NOT_ON_ROOT,
6436 		.read_u64 = memory_current_read,
6437 	},
6438 	{
6439 		.name = "min",
6440 		.flags = CFTYPE_NOT_ON_ROOT,
6441 		.seq_show = memory_min_show,
6442 		.write = memory_min_write,
6443 	},
6444 	{
6445 		.name = "low",
6446 		.flags = CFTYPE_NOT_ON_ROOT,
6447 		.seq_show = memory_low_show,
6448 		.write = memory_low_write,
6449 	},
6450 	{
6451 		.name = "high",
6452 		.flags = CFTYPE_NOT_ON_ROOT,
6453 		.seq_show = memory_high_show,
6454 		.write = memory_high_write,
6455 	},
6456 	{
6457 		.name = "max",
6458 		.flags = CFTYPE_NOT_ON_ROOT,
6459 		.seq_show = memory_max_show,
6460 		.write = memory_max_write,
6461 	},
6462 	{
6463 		.name = "events",
6464 		.flags = CFTYPE_NOT_ON_ROOT,
6465 		.file_offset = offsetof(struct mem_cgroup, events_file),
6466 		.seq_show = memory_events_show,
6467 	},
6468 	{
6469 		.name = "events.local",
6470 		.flags = CFTYPE_NOT_ON_ROOT,
6471 		.file_offset = offsetof(struct mem_cgroup, events_local_file),
6472 		.seq_show = memory_events_local_show,
6473 	},
6474 	{
6475 		.name = "stat",
6476 		.seq_show = memory_stat_show,
6477 	},
6478 #ifdef CONFIG_NUMA
6479 	{
6480 		.name = "numa_stat",
6481 		.seq_show = memory_numa_stat_show,
6482 	},
6483 #endif
6484 	{
6485 		.name = "oom.group",
6486 		.flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6487 		.seq_show = memory_oom_group_show,
6488 		.write = memory_oom_group_write,
6489 	},
6490 	{ }	/* terminate */
6491 };
6492 
6493 struct cgroup_subsys memory_cgrp_subsys = {
6494 	.css_alloc = mem_cgroup_css_alloc,
6495 	.css_online = mem_cgroup_css_online,
6496 	.css_offline = mem_cgroup_css_offline,
6497 	.css_released = mem_cgroup_css_released,
6498 	.css_free = mem_cgroup_css_free,
6499 	.css_reset = mem_cgroup_css_reset,
6500 	.css_rstat_flush = mem_cgroup_css_rstat_flush,
6501 	.can_attach = mem_cgroup_can_attach,
6502 	.cancel_attach = mem_cgroup_cancel_attach,
6503 	.post_attach = mem_cgroup_move_task,
6504 	.dfl_cftypes = memory_files,
6505 	.legacy_cftypes = mem_cgroup_legacy_files,
6506 	.early_init = 0,
6507 };
6508 
6509 /*
6510  * This function calculates an individual cgroup's effective
6511  * protection which is derived from its own memory.min/low, its
6512  * parent's and siblings' settings, as well as the actual memory
6513  * distribution in the tree.
6514  *
6515  * The following rules apply to the effective protection values:
6516  *
6517  * 1. At the first level of reclaim, effective protection is equal to
6518  *    the declared protection in memory.min and memory.low.
6519  *
6520  * 2. To enable safe delegation of the protection configuration, at
6521  *    subsequent levels the effective protection is capped to the
6522  *    parent's effective protection.
6523  *
6524  * 3. To make complex and dynamic subtrees easier to configure, the
6525  *    user is allowed to overcommit the declared protection at a given
6526  *    level. If that is the case, the parent's effective protection is
6527  *    distributed to the children in proportion to how much protection
6528  *    they have declared and how much of it they are utilizing.
6529  *
6530  *    This makes distribution proportional, but also work-conserving:
6531  *    if one cgroup claims much more protection than it uses memory,
6532  *    the unused remainder is available to its siblings.
6533  *
6534  * 4. Conversely, when the declared protection is undercommitted at a
6535  *    given level, the distribution of the larger parental protection
6536  *    budget is NOT proportional. A cgroup's protection from a sibling
6537  *    is capped to its own memory.min/low setting.
6538  *
6539  * 5. However, to allow protecting recursive subtrees from each other
6540  *    without having to declare each individual cgroup's fixed share
6541  *    of the ancestor's claim to protection, any unutilized -
6542  *    "floating" - protection from up the tree is distributed in
6543  *    proportion to each cgroup's *usage*. This makes the protection
6544  *    neutral wrt sibling cgroups and lets them compete freely over
6545  *    the shared parental protection budget, but it protects the
6546  *    subtree as a whole from neighboring subtrees.
6547  *
6548  * Note that 4. and 5. are not in conflict: 4. is about protecting
6549  * against immediate siblings whereas 5. is about protecting against
6550  * neighboring subtrees.
6551  */
6552 static unsigned long effective_protection(unsigned long usage,
6553 					  unsigned long parent_usage,
6554 					  unsigned long setting,
6555 					  unsigned long parent_effective,
6556 					  unsigned long siblings_protected)
6557 {
6558 	unsigned long protected;
6559 	unsigned long ep;
6560 
6561 	protected = min(usage, setting);
6562 	/*
6563 	 * If all cgroups at this level combined claim and use more
6564 	 * protection then what the parent affords them, distribute
6565 	 * shares in proportion to utilization.
6566 	 *
6567 	 * We are using actual utilization rather than the statically
6568 	 * claimed protection in order to be work-conserving: claimed
6569 	 * but unused protection is available to siblings that would
6570 	 * otherwise get a smaller chunk than what they claimed.
6571 	 */
6572 	if (siblings_protected > parent_effective)
6573 		return protected * parent_effective / siblings_protected;
6574 
6575 	/*
6576 	 * Ok, utilized protection of all children is within what the
6577 	 * parent affords them, so we know whatever this child claims
6578 	 * and utilizes is effectively protected.
6579 	 *
6580 	 * If there is unprotected usage beyond this value, reclaim
6581 	 * will apply pressure in proportion to that amount.
6582 	 *
6583 	 * If there is unutilized protection, the cgroup will be fully
6584 	 * shielded from reclaim, but we do return a smaller value for
6585 	 * protection than what the group could enjoy in theory. This
6586 	 * is okay. With the overcommit distribution above, effective
6587 	 * protection is always dependent on how memory is actually
6588 	 * consumed among the siblings anyway.
6589 	 */
6590 	ep = protected;
6591 
6592 	/*
6593 	 * If the children aren't claiming (all of) the protection
6594 	 * afforded to them by the parent, distribute the remainder in
6595 	 * proportion to the (unprotected) memory of each cgroup. That
6596 	 * way, cgroups that aren't explicitly prioritized wrt each
6597 	 * other compete freely over the allowance, but they are
6598 	 * collectively protected from neighboring trees.
6599 	 *
6600 	 * We're using unprotected memory for the weight so that if
6601 	 * some cgroups DO claim explicit protection, we don't protect
6602 	 * the same bytes twice.
6603 	 *
6604 	 * Check both usage and parent_usage against the respective
6605 	 * protected values. One should imply the other, but they
6606 	 * aren't read atomically - make sure the division is sane.
6607 	 */
6608 	if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6609 		return ep;
6610 	if (parent_effective > siblings_protected &&
6611 	    parent_usage > siblings_protected &&
6612 	    usage > protected) {
6613 		unsigned long unclaimed;
6614 
6615 		unclaimed = parent_effective - siblings_protected;
6616 		unclaimed *= usage - protected;
6617 		unclaimed /= parent_usage - siblings_protected;
6618 
6619 		ep += unclaimed;
6620 	}
6621 
6622 	return ep;
6623 }
6624 
6625 /**
6626  * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6627  * @root: the top ancestor of the sub-tree being checked
6628  * @memcg: the memory cgroup to check
6629  *
6630  * WARNING: This function is not stateless! It can only be used as part
6631  *          of a top-down tree iteration, not for isolated queries.
6632  */
6633 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6634 				     struct mem_cgroup *memcg)
6635 {
6636 	unsigned long usage, parent_usage;
6637 	struct mem_cgroup *parent;
6638 
6639 	if (mem_cgroup_disabled())
6640 		return;
6641 
6642 	if (!root)
6643 		root = root_mem_cgroup;
6644 
6645 	/*
6646 	 * Effective values of the reclaim targets are ignored so they
6647 	 * can be stale. Have a look at mem_cgroup_protection for more
6648 	 * details.
6649 	 * TODO: calculation should be more robust so that we do not need
6650 	 * that special casing.
6651 	 */
6652 	if (memcg == root)
6653 		return;
6654 
6655 	usage = page_counter_read(&memcg->memory);
6656 	if (!usage)
6657 		return;
6658 
6659 	parent = parent_mem_cgroup(memcg);
6660 	/* No parent means a non-hierarchical mode on v1 memcg */
6661 	if (!parent)
6662 		return;
6663 
6664 	if (parent == root) {
6665 		memcg->memory.emin = READ_ONCE(memcg->memory.min);
6666 		memcg->memory.elow = READ_ONCE(memcg->memory.low);
6667 		return;
6668 	}
6669 
6670 	parent_usage = page_counter_read(&parent->memory);
6671 
6672 	WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6673 			READ_ONCE(memcg->memory.min),
6674 			READ_ONCE(parent->memory.emin),
6675 			atomic_long_read(&parent->memory.children_min_usage)));
6676 
6677 	WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6678 			READ_ONCE(memcg->memory.low),
6679 			READ_ONCE(parent->memory.elow),
6680 			atomic_long_read(&parent->memory.children_low_usage)));
6681 }
6682 
6683 static int charge_memcg(struct page *page, struct mem_cgroup *memcg, gfp_t gfp)
6684 {
6685 	unsigned int nr_pages = thp_nr_pages(page);
6686 	int ret;
6687 
6688 	ret = try_charge(memcg, gfp, nr_pages);
6689 	if (ret)
6690 		goto out;
6691 
6692 	css_get(&memcg->css);
6693 	commit_charge(page, memcg);
6694 
6695 	local_irq_disable();
6696 	mem_cgroup_charge_statistics(memcg, page, nr_pages);
6697 	memcg_check_events(memcg, page);
6698 	local_irq_enable();
6699 out:
6700 	return ret;
6701 }
6702 
6703 /**
6704  * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6705  * @page: page to charge
6706  * @mm: mm context of the victim
6707  * @gfp_mask: reclaim mode
6708  *
6709  * Try to charge @page to the memcg that @mm belongs to, reclaiming
6710  * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6711  * charge to the active memcg.
6712  *
6713  * Do not use this for pages allocated for swapin.
6714  *
6715  * Returns 0 on success. Otherwise, an error code is returned.
6716  */
6717 int __mem_cgroup_charge(struct page *page, struct mm_struct *mm,
6718 			gfp_t gfp_mask)
6719 {
6720 	struct mem_cgroup *memcg;
6721 	int ret;
6722 
6723 	memcg = get_mem_cgroup_from_mm(mm);
6724 	ret = charge_memcg(page, memcg, gfp_mask);
6725 	css_put(&memcg->css);
6726 
6727 	return ret;
6728 }
6729 
6730 /**
6731  * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6732  * @page: page to charge
6733  * @mm: mm context of the victim
6734  * @gfp: reclaim mode
6735  * @entry: swap entry for which the page is allocated
6736  *
6737  * This function charges a page allocated for swapin. Please call this before
6738  * adding the page to the swapcache.
6739  *
6740  * Returns 0 on success. Otherwise, an error code is returned.
6741  */
6742 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6743 				  gfp_t gfp, swp_entry_t entry)
6744 {
6745 	struct mem_cgroup *memcg;
6746 	unsigned short id;
6747 	int ret;
6748 
6749 	if (mem_cgroup_disabled())
6750 		return 0;
6751 
6752 	id = lookup_swap_cgroup_id(entry);
6753 	rcu_read_lock();
6754 	memcg = mem_cgroup_from_id(id);
6755 	if (!memcg || !css_tryget_online(&memcg->css))
6756 		memcg = get_mem_cgroup_from_mm(mm);
6757 	rcu_read_unlock();
6758 
6759 	ret = charge_memcg(page, memcg, gfp);
6760 
6761 	css_put(&memcg->css);
6762 	return ret;
6763 }
6764 
6765 /*
6766  * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6767  * @entry: swap entry for which the page is charged
6768  *
6769  * Call this function after successfully adding the charged page to swapcache.
6770  *
6771  * Note: This function assumes the page for which swap slot is being uncharged
6772  * is order 0 page.
6773  */
6774 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6775 {
6776 	/*
6777 	 * Cgroup1's unified memory+swap counter has been charged with the
6778 	 * new swapcache page, finish the transfer by uncharging the swap
6779 	 * slot. The swap slot would also get uncharged when it dies, but
6780 	 * it can stick around indefinitely and we'd count the page twice
6781 	 * the entire time.
6782 	 *
6783 	 * Cgroup2 has separate resource counters for memory and swap,
6784 	 * so this is a non-issue here. Memory and swap charge lifetimes
6785 	 * correspond 1:1 to page and swap slot lifetimes: we charge the
6786 	 * page to memory here, and uncharge swap when the slot is freed.
6787 	 */
6788 	if (!mem_cgroup_disabled() && do_memsw_account()) {
6789 		/*
6790 		 * The swap entry might not get freed for a long time,
6791 		 * let's not wait for it.  The page already received a
6792 		 * memory+swap charge, drop the swap entry duplicate.
6793 		 */
6794 		mem_cgroup_uncharge_swap(entry, 1);
6795 	}
6796 }
6797 
6798 struct uncharge_gather {
6799 	struct mem_cgroup *memcg;
6800 	unsigned long nr_memory;
6801 	unsigned long pgpgout;
6802 	unsigned long nr_kmem;
6803 	struct page *dummy_page;
6804 };
6805 
6806 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6807 {
6808 	memset(ug, 0, sizeof(*ug));
6809 }
6810 
6811 static void uncharge_batch(const struct uncharge_gather *ug)
6812 {
6813 	unsigned long flags;
6814 
6815 	if (ug->nr_memory) {
6816 		page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6817 		if (do_memsw_account())
6818 			page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6819 		if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6820 			page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6821 		memcg_oom_recover(ug->memcg);
6822 	}
6823 
6824 	local_irq_save(flags);
6825 	__count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6826 	__this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6827 	memcg_check_events(ug->memcg, ug->dummy_page);
6828 	local_irq_restore(flags);
6829 
6830 	/* drop reference from uncharge_page */
6831 	css_put(&ug->memcg->css);
6832 }
6833 
6834 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6835 {
6836 	unsigned long nr_pages;
6837 	struct mem_cgroup *memcg;
6838 	struct obj_cgroup *objcg;
6839 	bool use_objcg = PageMemcgKmem(page);
6840 
6841 	VM_BUG_ON_PAGE(PageLRU(page), page);
6842 
6843 	/*
6844 	 * Nobody should be changing or seriously looking at
6845 	 * page memcg or objcg at this point, we have fully
6846 	 * exclusive access to the page.
6847 	 */
6848 	if (use_objcg) {
6849 		objcg = __page_objcg(page);
6850 		/*
6851 		 * This get matches the put at the end of the function and
6852 		 * kmem pages do not hold memcg references anymore.
6853 		 */
6854 		memcg = get_mem_cgroup_from_objcg(objcg);
6855 	} else {
6856 		memcg = __page_memcg(page);
6857 	}
6858 
6859 	if (!memcg)
6860 		return;
6861 
6862 	if (ug->memcg != memcg) {
6863 		if (ug->memcg) {
6864 			uncharge_batch(ug);
6865 			uncharge_gather_clear(ug);
6866 		}
6867 		ug->memcg = memcg;
6868 		ug->dummy_page = page;
6869 
6870 		/* pairs with css_put in uncharge_batch */
6871 		css_get(&memcg->css);
6872 	}
6873 
6874 	nr_pages = compound_nr(page);
6875 
6876 	if (use_objcg) {
6877 		ug->nr_memory += nr_pages;
6878 		ug->nr_kmem += nr_pages;
6879 
6880 		page->memcg_data = 0;
6881 		obj_cgroup_put(objcg);
6882 	} else {
6883 		/* LRU pages aren't accounted at the root level */
6884 		if (!mem_cgroup_is_root(memcg))
6885 			ug->nr_memory += nr_pages;
6886 		ug->pgpgout++;
6887 
6888 		page->memcg_data = 0;
6889 	}
6890 
6891 	css_put(&memcg->css);
6892 }
6893 
6894 /**
6895  * __mem_cgroup_uncharge - uncharge a page
6896  * @page: page to uncharge
6897  *
6898  * Uncharge a page previously charged with __mem_cgroup_charge().
6899  */
6900 void __mem_cgroup_uncharge(struct page *page)
6901 {
6902 	struct uncharge_gather ug;
6903 
6904 	/* Don't touch page->lru of any random page, pre-check: */
6905 	if (!page_memcg(page))
6906 		return;
6907 
6908 	uncharge_gather_clear(&ug);
6909 	uncharge_page(page, &ug);
6910 	uncharge_batch(&ug);
6911 }
6912 
6913 /**
6914  * __mem_cgroup_uncharge_list - uncharge a list of page
6915  * @page_list: list of pages to uncharge
6916  *
6917  * Uncharge a list of pages previously charged with
6918  * __mem_cgroup_charge().
6919  */
6920 void __mem_cgroup_uncharge_list(struct list_head *page_list)
6921 {
6922 	struct uncharge_gather ug;
6923 	struct page *page;
6924 
6925 	uncharge_gather_clear(&ug);
6926 	list_for_each_entry(page, page_list, lru)
6927 		uncharge_page(page, &ug);
6928 	if (ug.memcg)
6929 		uncharge_batch(&ug);
6930 }
6931 
6932 /**
6933  * mem_cgroup_migrate - charge a page's replacement
6934  * @oldpage: currently circulating page
6935  * @newpage: replacement page
6936  *
6937  * Charge @newpage as a replacement page for @oldpage. @oldpage will
6938  * be uncharged upon free.
6939  *
6940  * Both pages must be locked, @newpage->mapping must be set up.
6941  */
6942 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6943 {
6944 	struct mem_cgroup *memcg;
6945 	unsigned int nr_pages;
6946 	unsigned long flags;
6947 
6948 	VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6949 	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6950 	VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6951 	VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6952 		       newpage);
6953 
6954 	if (mem_cgroup_disabled())
6955 		return;
6956 
6957 	/* Page cache replacement: new page already charged? */
6958 	if (page_memcg(newpage))
6959 		return;
6960 
6961 	memcg = page_memcg(oldpage);
6962 	VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6963 	if (!memcg)
6964 		return;
6965 
6966 	/* Force-charge the new page. The old one will be freed soon */
6967 	nr_pages = thp_nr_pages(newpage);
6968 
6969 	if (!mem_cgroup_is_root(memcg)) {
6970 		page_counter_charge(&memcg->memory, nr_pages);
6971 		if (do_memsw_account())
6972 			page_counter_charge(&memcg->memsw, nr_pages);
6973 	}
6974 
6975 	css_get(&memcg->css);
6976 	commit_charge(newpage, memcg);
6977 
6978 	local_irq_save(flags);
6979 	mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6980 	memcg_check_events(memcg, newpage);
6981 	local_irq_restore(flags);
6982 }
6983 
6984 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6985 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6986 
6987 void mem_cgroup_sk_alloc(struct sock *sk)
6988 {
6989 	struct mem_cgroup *memcg;
6990 
6991 	if (!mem_cgroup_sockets_enabled)
6992 		return;
6993 
6994 	/* Do not associate the sock with unrelated interrupted task's memcg. */
6995 	if (in_interrupt())
6996 		return;
6997 
6998 	rcu_read_lock();
6999 	memcg = mem_cgroup_from_task(current);
7000 	if (memcg == root_mem_cgroup)
7001 		goto out;
7002 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7003 		goto out;
7004 	if (css_tryget(&memcg->css))
7005 		sk->sk_memcg = memcg;
7006 out:
7007 	rcu_read_unlock();
7008 }
7009 
7010 void mem_cgroup_sk_free(struct sock *sk)
7011 {
7012 	if (sk->sk_memcg)
7013 		css_put(&sk->sk_memcg->css);
7014 }
7015 
7016 /**
7017  * mem_cgroup_charge_skmem - charge socket memory
7018  * @memcg: memcg to charge
7019  * @nr_pages: number of pages to charge
7020  * @gfp_mask: reclaim mode
7021  *
7022  * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7023  * @memcg's configured limit, %false if it doesn't.
7024  */
7025 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7026 			     gfp_t gfp_mask)
7027 {
7028 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7029 		struct page_counter *fail;
7030 
7031 		if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7032 			memcg->tcpmem_pressure = 0;
7033 			return true;
7034 		}
7035 		memcg->tcpmem_pressure = 1;
7036 		if (gfp_mask & __GFP_NOFAIL) {
7037 			page_counter_charge(&memcg->tcpmem, nr_pages);
7038 			return true;
7039 		}
7040 		return false;
7041 	}
7042 
7043 	if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7044 		mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7045 		return true;
7046 	}
7047 
7048 	return false;
7049 }
7050 
7051 /**
7052  * mem_cgroup_uncharge_skmem - uncharge socket memory
7053  * @memcg: memcg to uncharge
7054  * @nr_pages: number of pages to uncharge
7055  */
7056 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7057 {
7058 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7059 		page_counter_uncharge(&memcg->tcpmem, nr_pages);
7060 		return;
7061 	}
7062 
7063 	mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7064 
7065 	refill_stock(memcg, nr_pages);
7066 }
7067 
7068 static int __init cgroup_memory(char *s)
7069 {
7070 	char *token;
7071 
7072 	while ((token = strsep(&s, ",")) != NULL) {
7073 		if (!*token)
7074 			continue;
7075 		if (!strcmp(token, "nosocket"))
7076 			cgroup_memory_nosocket = true;
7077 		if (!strcmp(token, "nokmem"))
7078 			cgroup_memory_nokmem = true;
7079 	}
7080 	return 0;
7081 }
7082 __setup("cgroup.memory=", cgroup_memory);
7083 
7084 /*
7085  * subsys_initcall() for memory controller.
7086  *
7087  * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7088  * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7089  * basically everything that doesn't depend on a specific mem_cgroup structure
7090  * should be initialized from here.
7091  */
7092 static int __init mem_cgroup_init(void)
7093 {
7094 	int cpu, node;
7095 
7096 	/*
7097 	 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7098 	 * used for per-memcg-per-cpu caching of per-node statistics. In order
7099 	 * to work fine, we should make sure that the overfill threshold can't
7100 	 * exceed S32_MAX / PAGE_SIZE.
7101 	 */
7102 	BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7103 
7104 	cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7105 				  memcg_hotplug_cpu_dead);
7106 
7107 	for_each_possible_cpu(cpu)
7108 		INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7109 			  drain_local_stock);
7110 
7111 	for_each_node(node) {
7112 		struct mem_cgroup_tree_per_node *rtpn;
7113 
7114 		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7115 				    node_online(node) ? node : NUMA_NO_NODE);
7116 
7117 		rtpn->rb_root = RB_ROOT;
7118 		rtpn->rb_rightmost = NULL;
7119 		spin_lock_init(&rtpn->lock);
7120 		soft_limit_tree.rb_tree_per_node[node] = rtpn;
7121 	}
7122 
7123 	return 0;
7124 }
7125 subsys_initcall(mem_cgroup_init);
7126 
7127 #ifdef CONFIG_MEMCG_SWAP
7128 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7129 {
7130 	while (!refcount_inc_not_zero(&memcg->id.ref)) {
7131 		/*
7132 		 * The root cgroup cannot be destroyed, so it's refcount must
7133 		 * always be >= 1.
7134 		 */
7135 		if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7136 			VM_BUG_ON(1);
7137 			break;
7138 		}
7139 		memcg = parent_mem_cgroup(memcg);
7140 		if (!memcg)
7141 			memcg = root_mem_cgroup;
7142 	}
7143 	return memcg;
7144 }
7145 
7146 /**
7147  * mem_cgroup_swapout - transfer a memsw charge to swap
7148  * @page: page whose memsw charge to transfer
7149  * @entry: swap entry to move the charge to
7150  *
7151  * Transfer the memsw charge of @page to @entry.
7152  */
7153 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7154 {
7155 	struct mem_cgroup *memcg, *swap_memcg;
7156 	unsigned int nr_entries;
7157 	unsigned short oldid;
7158 
7159 	VM_BUG_ON_PAGE(PageLRU(page), page);
7160 	VM_BUG_ON_PAGE(page_count(page), page);
7161 
7162 	if (mem_cgroup_disabled())
7163 		return;
7164 
7165 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7166 		return;
7167 
7168 	memcg = page_memcg(page);
7169 
7170 	VM_WARN_ON_ONCE_PAGE(!memcg, page);
7171 	if (!memcg)
7172 		return;
7173 
7174 	/*
7175 	 * In case the memcg owning these pages has been offlined and doesn't
7176 	 * have an ID allocated to it anymore, charge the closest online
7177 	 * ancestor for the swap instead and transfer the memory+swap charge.
7178 	 */
7179 	swap_memcg = mem_cgroup_id_get_online(memcg);
7180 	nr_entries = thp_nr_pages(page);
7181 	/* Get references for the tail pages, too */
7182 	if (nr_entries > 1)
7183 		mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7184 	oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7185 				   nr_entries);
7186 	VM_BUG_ON_PAGE(oldid, page);
7187 	mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7188 
7189 	page->memcg_data = 0;
7190 
7191 	if (!mem_cgroup_is_root(memcg))
7192 		page_counter_uncharge(&memcg->memory, nr_entries);
7193 
7194 	if (!cgroup_memory_noswap && memcg != swap_memcg) {
7195 		if (!mem_cgroup_is_root(swap_memcg))
7196 			page_counter_charge(&swap_memcg->memsw, nr_entries);
7197 		page_counter_uncharge(&memcg->memsw, nr_entries);
7198 	}
7199 
7200 	/*
7201 	 * Interrupts should be disabled here because the caller holds the
7202 	 * i_pages lock which is taken with interrupts-off. It is
7203 	 * important here to have the interrupts disabled because it is the
7204 	 * only synchronisation we have for updating the per-CPU variables.
7205 	 */
7206 	VM_BUG_ON(!irqs_disabled());
7207 	mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7208 	memcg_check_events(memcg, page);
7209 
7210 	css_put(&memcg->css);
7211 }
7212 
7213 /**
7214  * __mem_cgroup_try_charge_swap - try charging swap space for a page
7215  * @page: page being added to swap
7216  * @entry: swap entry to charge
7217  *
7218  * Try to charge @page's memcg for the swap space at @entry.
7219  *
7220  * Returns 0 on success, -ENOMEM on failure.
7221  */
7222 int __mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7223 {
7224 	unsigned int nr_pages = thp_nr_pages(page);
7225 	struct page_counter *counter;
7226 	struct mem_cgroup *memcg;
7227 	unsigned short oldid;
7228 
7229 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7230 		return 0;
7231 
7232 	memcg = page_memcg(page);
7233 
7234 	VM_WARN_ON_ONCE_PAGE(!memcg, page);
7235 	if (!memcg)
7236 		return 0;
7237 
7238 	if (!entry.val) {
7239 		memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7240 		return 0;
7241 	}
7242 
7243 	memcg = mem_cgroup_id_get_online(memcg);
7244 
7245 	if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7246 	    !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7247 		memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7248 		memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7249 		mem_cgroup_id_put(memcg);
7250 		return -ENOMEM;
7251 	}
7252 
7253 	/* Get references for the tail pages, too */
7254 	if (nr_pages > 1)
7255 		mem_cgroup_id_get_many(memcg, nr_pages - 1);
7256 	oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7257 	VM_BUG_ON_PAGE(oldid, page);
7258 	mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7259 
7260 	return 0;
7261 }
7262 
7263 /**
7264  * __mem_cgroup_uncharge_swap - uncharge swap space
7265  * @entry: swap entry to uncharge
7266  * @nr_pages: the amount of swap space to uncharge
7267  */
7268 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7269 {
7270 	struct mem_cgroup *memcg;
7271 	unsigned short id;
7272 
7273 	id = swap_cgroup_record(entry, 0, nr_pages);
7274 	rcu_read_lock();
7275 	memcg = mem_cgroup_from_id(id);
7276 	if (memcg) {
7277 		if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7278 			if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7279 				page_counter_uncharge(&memcg->swap, nr_pages);
7280 			else
7281 				page_counter_uncharge(&memcg->memsw, nr_pages);
7282 		}
7283 		mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7284 		mem_cgroup_id_put_many(memcg, nr_pages);
7285 	}
7286 	rcu_read_unlock();
7287 }
7288 
7289 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7290 {
7291 	long nr_swap_pages = get_nr_swap_pages();
7292 
7293 	if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7294 		return nr_swap_pages;
7295 	for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7296 		nr_swap_pages = min_t(long, nr_swap_pages,
7297 				      READ_ONCE(memcg->swap.max) -
7298 				      page_counter_read(&memcg->swap));
7299 	return nr_swap_pages;
7300 }
7301 
7302 bool mem_cgroup_swap_full(struct page *page)
7303 {
7304 	struct mem_cgroup *memcg;
7305 
7306 	VM_BUG_ON_PAGE(!PageLocked(page), page);
7307 
7308 	if (vm_swap_full())
7309 		return true;
7310 	if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7311 		return false;
7312 
7313 	memcg = page_memcg(page);
7314 	if (!memcg)
7315 		return false;
7316 
7317 	for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7318 		unsigned long usage = page_counter_read(&memcg->swap);
7319 
7320 		if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7321 		    usage * 2 >= READ_ONCE(memcg->swap.max))
7322 			return true;
7323 	}
7324 
7325 	return false;
7326 }
7327 
7328 static int __init setup_swap_account(char *s)
7329 {
7330 	if (!strcmp(s, "1"))
7331 		cgroup_memory_noswap = false;
7332 	else if (!strcmp(s, "0"))
7333 		cgroup_memory_noswap = true;
7334 	return 1;
7335 }
7336 __setup("swapaccount=", setup_swap_account);
7337 
7338 static u64 swap_current_read(struct cgroup_subsys_state *css,
7339 			     struct cftype *cft)
7340 {
7341 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7342 
7343 	return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7344 }
7345 
7346 static int swap_high_show(struct seq_file *m, void *v)
7347 {
7348 	return seq_puts_memcg_tunable(m,
7349 		READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7350 }
7351 
7352 static ssize_t swap_high_write(struct kernfs_open_file *of,
7353 			       char *buf, size_t nbytes, loff_t off)
7354 {
7355 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7356 	unsigned long high;
7357 	int err;
7358 
7359 	buf = strstrip(buf);
7360 	err = page_counter_memparse(buf, "max", &high);
7361 	if (err)
7362 		return err;
7363 
7364 	page_counter_set_high(&memcg->swap, high);
7365 
7366 	return nbytes;
7367 }
7368 
7369 static int swap_max_show(struct seq_file *m, void *v)
7370 {
7371 	return seq_puts_memcg_tunable(m,
7372 		READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7373 }
7374 
7375 static ssize_t swap_max_write(struct kernfs_open_file *of,
7376 			      char *buf, size_t nbytes, loff_t off)
7377 {
7378 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7379 	unsigned long max;
7380 	int err;
7381 
7382 	buf = strstrip(buf);
7383 	err = page_counter_memparse(buf, "max", &max);
7384 	if (err)
7385 		return err;
7386 
7387 	xchg(&memcg->swap.max, max);
7388 
7389 	return nbytes;
7390 }
7391 
7392 static int swap_events_show(struct seq_file *m, void *v)
7393 {
7394 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7395 
7396 	seq_printf(m, "high %lu\n",
7397 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7398 	seq_printf(m, "max %lu\n",
7399 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7400 	seq_printf(m, "fail %lu\n",
7401 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7402 
7403 	return 0;
7404 }
7405 
7406 static struct cftype swap_files[] = {
7407 	{
7408 		.name = "swap.current",
7409 		.flags = CFTYPE_NOT_ON_ROOT,
7410 		.read_u64 = swap_current_read,
7411 	},
7412 	{
7413 		.name = "swap.high",
7414 		.flags = CFTYPE_NOT_ON_ROOT,
7415 		.seq_show = swap_high_show,
7416 		.write = swap_high_write,
7417 	},
7418 	{
7419 		.name = "swap.max",
7420 		.flags = CFTYPE_NOT_ON_ROOT,
7421 		.seq_show = swap_max_show,
7422 		.write = swap_max_write,
7423 	},
7424 	{
7425 		.name = "swap.events",
7426 		.flags = CFTYPE_NOT_ON_ROOT,
7427 		.file_offset = offsetof(struct mem_cgroup, swap_events_file),
7428 		.seq_show = swap_events_show,
7429 	},
7430 	{ }	/* terminate */
7431 };
7432 
7433 static struct cftype memsw_files[] = {
7434 	{
7435 		.name = "memsw.usage_in_bytes",
7436 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7437 		.read_u64 = mem_cgroup_read_u64,
7438 	},
7439 	{
7440 		.name = "memsw.max_usage_in_bytes",
7441 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7442 		.write = mem_cgroup_reset,
7443 		.read_u64 = mem_cgroup_read_u64,
7444 	},
7445 	{
7446 		.name = "memsw.limit_in_bytes",
7447 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7448 		.write = mem_cgroup_write,
7449 		.read_u64 = mem_cgroup_read_u64,
7450 	},
7451 	{
7452 		.name = "memsw.failcnt",
7453 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7454 		.write = mem_cgroup_reset,
7455 		.read_u64 = mem_cgroup_read_u64,
7456 	},
7457 	{ },	/* terminate */
7458 };
7459 
7460 /*
7461  * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7462  * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7463  * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7464  * boot parameter. This may result in premature OOPS inside
7465  * mem_cgroup_get_nr_swap_pages() function in corner cases.
7466  */
7467 static int __init mem_cgroup_swap_init(void)
7468 {
7469 	/* No memory control -> no swap control */
7470 	if (mem_cgroup_disabled())
7471 		cgroup_memory_noswap = true;
7472 
7473 	if (cgroup_memory_noswap)
7474 		return 0;
7475 
7476 	WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7477 	WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7478 
7479 	return 0;
7480 }
7481 core_initcall(mem_cgroup_swap_init);
7482 
7483 #endif /* CONFIG_MEMCG_SWAP */
7484