xref: /linux/mm/memcontrol.c (revision f6154d8babbb8a98f0d3ea325aafae2e33bfd8be)
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/memremap.h>
57 #include <linux/mm_inline.h>
58 #include <linux/swap_cgroup.h>
59 #include <linux/cpu.h>
60 #include <linux/oom.h>
61 #include <linux/lockdep.h>
62 #include <linux/file.h>
63 #include <linux/resume_user_mode.h>
64 #include <linux/psi.h>
65 #include <linux/seq_buf.h>
66 #include <linux/sched/isolation.h>
67 #include "internal.h"
68 #include <net/sock.h>
69 #include <net/ip.h>
70 #include "slab.h"
71 #include "swap.h"
72 
73 #include <linux/uaccess.h>
74 
75 #include <trace/events/vmscan.h>
76 
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
79 
80 struct mem_cgroup *root_mem_cgroup __read_mostly;
81 
82 /* Active memory cgroup to use from an interrupt context */
83 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
84 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
85 
86 /* Socket memory accounting disabled? */
87 static bool cgroup_memory_nosocket __ro_after_init;
88 
89 /* Kernel memory accounting disabled? */
90 static bool cgroup_memory_nokmem __ro_after_init;
91 
92 /* BPF memory accounting disabled? */
93 static bool cgroup_memory_nobpf __ro_after_init;
94 
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
97 #endif
98 
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
101 {
102 	return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
103 }
104 
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
107 
108 /*
109  * Cgroups above their limits are maintained in a RB-Tree, independent of
110  * their hierarchy representation
111  */
112 
113 struct mem_cgroup_tree_per_node {
114 	struct rb_root rb_root;
115 	struct rb_node *rb_rightmost;
116 	spinlock_t lock;
117 };
118 
119 struct mem_cgroup_tree {
120 	struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
121 };
122 
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
124 
125 /* for OOM */
126 struct mem_cgroup_eventfd_list {
127 	struct list_head list;
128 	struct eventfd_ctx *eventfd;
129 };
130 
131 /*
132  * cgroup_event represents events which userspace want to receive.
133  */
134 struct mem_cgroup_event {
135 	/*
136 	 * memcg which the event belongs to.
137 	 */
138 	struct mem_cgroup *memcg;
139 	/*
140 	 * eventfd to signal userspace about the event.
141 	 */
142 	struct eventfd_ctx *eventfd;
143 	/*
144 	 * Each of these stored in a list by the cgroup.
145 	 */
146 	struct list_head list;
147 	/*
148 	 * register_event() callback will be used to add new userspace
149 	 * waiter for changes related to this event.  Use eventfd_signal()
150 	 * on eventfd to send notification to userspace.
151 	 */
152 	int (*register_event)(struct mem_cgroup *memcg,
153 			      struct eventfd_ctx *eventfd, const char *args);
154 	/*
155 	 * unregister_event() callback will be called when userspace closes
156 	 * the eventfd or on cgroup removing.  This callback must be set,
157 	 * if you want provide notification functionality.
158 	 */
159 	void (*unregister_event)(struct mem_cgroup *memcg,
160 				 struct eventfd_ctx *eventfd);
161 	/*
162 	 * All fields below needed to unregister event when
163 	 * userspace closes eventfd.
164 	 */
165 	poll_table pt;
166 	wait_queue_head_t *wqh;
167 	wait_queue_entry_t wait;
168 	struct work_struct remove;
169 };
170 
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
173 
174 /* Stuffs for move charges at task migration. */
175 /*
176  * Types of charges to be moved.
177  */
178 #define MOVE_ANON	0x1U
179 #define MOVE_FILE	0x2U
180 #define MOVE_MASK	(MOVE_ANON | MOVE_FILE)
181 
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 	spinlock_t	  lock; /* for from, to */
185 	struct mm_struct  *mm;
186 	struct mem_cgroup *from;
187 	struct mem_cgroup *to;
188 	unsigned long flags;
189 	unsigned long precharge;
190 	unsigned long moved_charge;
191 	unsigned long moved_swap;
192 	struct task_struct *moving_task;	/* a task moving charges */
193 	wait_queue_head_t waitq;		/* a waitq for other context */
194 } mc = {
195 	.lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 	.waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 };
198 
199 /*
200  * Maximum loops in mem_cgroup_soft_reclaim(), used for soft
201  * limit reclaim to prevent infinite loops, if they ever occur.
202  */
203 #define	MEM_CGROUP_MAX_RECLAIM_LOOPS		100
204 #define	MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS	2
205 
206 /* for encoding cft->private value on file */
207 enum res_type {
208 	_MEM,
209 	_MEMSWAP,
210 	_KMEM,
211 	_TCP,
212 };
213 
214 #define MEMFILE_PRIVATE(x, val)	((x) << 16 | (val))
215 #define MEMFILE_TYPE(val)	((val) >> 16 & 0xffff)
216 #define MEMFILE_ATTR(val)	((val) & 0xffff)
217 
218 /*
219  * Iteration constructs for visiting all cgroups (under a tree).  If
220  * loops are exited prematurely (break), mem_cgroup_iter_break() must
221  * be used for reference counting.
222  */
223 #define for_each_mem_cgroup_tree(iter, root)		\
224 	for (iter = mem_cgroup_iter(root, NULL, NULL);	\
225 	     iter != NULL;				\
226 	     iter = mem_cgroup_iter(root, iter, NULL))
227 
228 #define for_each_mem_cgroup(iter)			\
229 	for (iter = mem_cgroup_iter(NULL, NULL, NULL);	\
230 	     iter != NULL;				\
231 	     iter = mem_cgroup_iter(NULL, iter, NULL))
232 
233 static inline bool task_is_dying(void)
234 {
235 	return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 		(current->flags & PF_EXITING);
237 }
238 
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
241 {
242 	if (!memcg)
243 		memcg = root_mem_cgroup;
244 	return &memcg->vmpressure;
245 }
246 
247 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
248 {
249 	return container_of(vmpr, struct mem_cgroup, vmpressure);
250 }
251 
252 #define CURRENT_OBJCG_UPDATE_BIT 0
253 #define CURRENT_OBJCG_UPDATE_FLAG (1UL << CURRENT_OBJCG_UPDATE_BIT)
254 
255 #ifdef CONFIG_MEMCG_KMEM
256 static DEFINE_SPINLOCK(objcg_lock);
257 
258 bool mem_cgroup_kmem_disabled(void)
259 {
260 	return cgroup_memory_nokmem;
261 }
262 
263 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
264 				      unsigned int nr_pages);
265 
266 static void obj_cgroup_release(struct percpu_ref *ref)
267 {
268 	struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
269 	unsigned int nr_bytes;
270 	unsigned int nr_pages;
271 	unsigned long flags;
272 
273 	/*
274 	 * At this point all allocated objects are freed, and
275 	 * objcg->nr_charged_bytes can't have an arbitrary byte value.
276 	 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
277 	 *
278 	 * The following sequence can lead to it:
279 	 * 1) CPU0: objcg == stock->cached_objcg
280 	 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
281 	 *          PAGE_SIZE bytes are charged
282 	 * 3) CPU1: a process from another memcg is allocating something,
283 	 *          the stock if flushed,
284 	 *          objcg->nr_charged_bytes = PAGE_SIZE - 92
285 	 * 5) CPU0: we do release this object,
286 	 *          92 bytes are added to stock->nr_bytes
287 	 * 6) CPU0: stock is flushed,
288 	 *          92 bytes are added to objcg->nr_charged_bytes
289 	 *
290 	 * In the result, nr_charged_bytes == PAGE_SIZE.
291 	 * This page will be uncharged in obj_cgroup_release().
292 	 */
293 	nr_bytes = atomic_read(&objcg->nr_charged_bytes);
294 	WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
295 	nr_pages = nr_bytes >> PAGE_SHIFT;
296 
297 	if (nr_pages)
298 		obj_cgroup_uncharge_pages(objcg, nr_pages);
299 
300 	spin_lock_irqsave(&objcg_lock, flags);
301 	list_del(&objcg->list);
302 	spin_unlock_irqrestore(&objcg_lock, flags);
303 
304 	percpu_ref_exit(ref);
305 	kfree_rcu(objcg, rcu);
306 }
307 
308 static struct obj_cgroup *obj_cgroup_alloc(void)
309 {
310 	struct obj_cgroup *objcg;
311 	int ret;
312 
313 	objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
314 	if (!objcg)
315 		return NULL;
316 
317 	ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
318 			      GFP_KERNEL);
319 	if (ret) {
320 		kfree(objcg);
321 		return NULL;
322 	}
323 	INIT_LIST_HEAD(&objcg->list);
324 	return objcg;
325 }
326 
327 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
328 				  struct mem_cgroup *parent)
329 {
330 	struct obj_cgroup *objcg, *iter;
331 
332 	objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
333 
334 	spin_lock_irq(&objcg_lock);
335 
336 	/* 1) Ready to reparent active objcg. */
337 	list_add(&objcg->list, &memcg->objcg_list);
338 	/* 2) Reparent active objcg and already reparented objcgs to parent. */
339 	list_for_each_entry(iter, &memcg->objcg_list, list)
340 		WRITE_ONCE(iter->memcg, parent);
341 	/* 3) Move already reparented objcgs to the parent's list */
342 	list_splice(&memcg->objcg_list, &parent->objcg_list);
343 
344 	spin_unlock_irq(&objcg_lock);
345 
346 	percpu_ref_kill(&objcg->refcnt);
347 }
348 
349 /*
350  * A lot of the calls to the cache allocation functions are expected to be
351  * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
352  * conditional to this static branch, we'll have to allow modules that does
353  * kmem_cache_alloc and the such to see this symbol as well
354  */
355 DEFINE_STATIC_KEY_FALSE(memcg_kmem_online_key);
356 EXPORT_SYMBOL(memcg_kmem_online_key);
357 
358 DEFINE_STATIC_KEY_FALSE(memcg_bpf_enabled_key);
359 EXPORT_SYMBOL(memcg_bpf_enabled_key);
360 #endif
361 
362 /**
363  * mem_cgroup_css_from_folio - css of the memcg associated with a folio
364  * @folio: folio of interest
365  *
366  * If memcg is bound to the default hierarchy, css of the memcg associated
367  * with @folio is returned.  The returned css remains associated with @folio
368  * until it is released.
369  *
370  * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
371  * is returned.
372  */
373 struct cgroup_subsys_state *mem_cgroup_css_from_folio(struct folio *folio)
374 {
375 	struct mem_cgroup *memcg = folio_memcg(folio);
376 
377 	if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
378 		memcg = root_mem_cgroup;
379 
380 	return &memcg->css;
381 }
382 
383 /**
384  * page_cgroup_ino - return inode number of the memcg a page is charged to
385  * @page: the page
386  *
387  * Look up the closest online ancestor of the memory cgroup @page is charged to
388  * and return its inode number or 0 if @page is not charged to any cgroup. It
389  * is safe to call this function without holding a reference to @page.
390  *
391  * Note, this function is inherently racy, because there is nothing to prevent
392  * the cgroup inode from getting torn down and potentially reallocated a moment
393  * after page_cgroup_ino() returns, so it only should be used by callers that
394  * do not care (such as procfs interfaces).
395  */
396 ino_t page_cgroup_ino(struct page *page)
397 {
398 	struct mem_cgroup *memcg;
399 	unsigned long ino = 0;
400 
401 	rcu_read_lock();
402 	/* page_folio() is racy here, but the entire function is racy anyway */
403 	memcg = folio_memcg_check(page_folio(page));
404 
405 	while (memcg && !(memcg->css.flags & CSS_ONLINE))
406 		memcg = parent_mem_cgroup(memcg);
407 	if (memcg)
408 		ino = cgroup_ino(memcg->css.cgroup);
409 	rcu_read_unlock();
410 	return ino;
411 }
412 
413 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
414 					 struct mem_cgroup_tree_per_node *mctz,
415 					 unsigned long new_usage_in_excess)
416 {
417 	struct rb_node **p = &mctz->rb_root.rb_node;
418 	struct rb_node *parent = NULL;
419 	struct mem_cgroup_per_node *mz_node;
420 	bool rightmost = true;
421 
422 	if (mz->on_tree)
423 		return;
424 
425 	mz->usage_in_excess = new_usage_in_excess;
426 	if (!mz->usage_in_excess)
427 		return;
428 	while (*p) {
429 		parent = *p;
430 		mz_node = rb_entry(parent, struct mem_cgroup_per_node,
431 					tree_node);
432 		if (mz->usage_in_excess < mz_node->usage_in_excess) {
433 			p = &(*p)->rb_left;
434 			rightmost = false;
435 		} else {
436 			p = &(*p)->rb_right;
437 		}
438 	}
439 
440 	if (rightmost)
441 		mctz->rb_rightmost = &mz->tree_node;
442 
443 	rb_link_node(&mz->tree_node, parent, p);
444 	rb_insert_color(&mz->tree_node, &mctz->rb_root);
445 	mz->on_tree = true;
446 }
447 
448 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
449 					 struct mem_cgroup_tree_per_node *mctz)
450 {
451 	if (!mz->on_tree)
452 		return;
453 
454 	if (&mz->tree_node == mctz->rb_rightmost)
455 		mctz->rb_rightmost = rb_prev(&mz->tree_node);
456 
457 	rb_erase(&mz->tree_node, &mctz->rb_root);
458 	mz->on_tree = false;
459 }
460 
461 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
462 				       struct mem_cgroup_tree_per_node *mctz)
463 {
464 	unsigned long flags;
465 
466 	spin_lock_irqsave(&mctz->lock, flags);
467 	__mem_cgroup_remove_exceeded(mz, mctz);
468 	spin_unlock_irqrestore(&mctz->lock, flags);
469 }
470 
471 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
472 {
473 	unsigned long nr_pages = page_counter_read(&memcg->memory);
474 	unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
475 	unsigned long excess = 0;
476 
477 	if (nr_pages > soft_limit)
478 		excess = nr_pages - soft_limit;
479 
480 	return excess;
481 }
482 
483 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
484 {
485 	unsigned long excess;
486 	struct mem_cgroup_per_node *mz;
487 	struct mem_cgroup_tree_per_node *mctz;
488 
489 	if (lru_gen_enabled()) {
490 		if (soft_limit_excess(memcg))
491 			lru_gen_soft_reclaim(memcg, nid);
492 		return;
493 	}
494 
495 	mctz = soft_limit_tree.rb_tree_per_node[nid];
496 	if (!mctz)
497 		return;
498 	/*
499 	 * Necessary to update all ancestors when hierarchy is used.
500 	 * because their event counter is not touched.
501 	 */
502 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
503 		mz = memcg->nodeinfo[nid];
504 		excess = soft_limit_excess(memcg);
505 		/*
506 		 * We have to update the tree if mz is on RB-tree or
507 		 * mem is over its softlimit.
508 		 */
509 		if (excess || mz->on_tree) {
510 			unsigned long flags;
511 
512 			spin_lock_irqsave(&mctz->lock, flags);
513 			/* if on-tree, remove it */
514 			if (mz->on_tree)
515 				__mem_cgroup_remove_exceeded(mz, mctz);
516 			/*
517 			 * Insert again. mz->usage_in_excess will be updated.
518 			 * If excess is 0, no tree ops.
519 			 */
520 			__mem_cgroup_insert_exceeded(mz, mctz, excess);
521 			spin_unlock_irqrestore(&mctz->lock, flags);
522 		}
523 	}
524 }
525 
526 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
527 {
528 	struct mem_cgroup_tree_per_node *mctz;
529 	struct mem_cgroup_per_node *mz;
530 	int nid;
531 
532 	for_each_node(nid) {
533 		mz = memcg->nodeinfo[nid];
534 		mctz = soft_limit_tree.rb_tree_per_node[nid];
535 		if (mctz)
536 			mem_cgroup_remove_exceeded(mz, mctz);
537 	}
538 }
539 
540 static struct mem_cgroup_per_node *
541 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
542 {
543 	struct mem_cgroup_per_node *mz;
544 
545 retry:
546 	mz = NULL;
547 	if (!mctz->rb_rightmost)
548 		goto done;		/* Nothing to reclaim from */
549 
550 	mz = rb_entry(mctz->rb_rightmost,
551 		      struct mem_cgroup_per_node, tree_node);
552 	/*
553 	 * Remove the node now but someone else can add it back,
554 	 * we will to add it back at the end of reclaim to its correct
555 	 * position in the tree.
556 	 */
557 	__mem_cgroup_remove_exceeded(mz, mctz);
558 	if (!soft_limit_excess(mz->memcg) ||
559 	    !css_tryget(&mz->memcg->css))
560 		goto retry;
561 done:
562 	return mz;
563 }
564 
565 static struct mem_cgroup_per_node *
566 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
567 {
568 	struct mem_cgroup_per_node *mz;
569 
570 	spin_lock_irq(&mctz->lock);
571 	mz = __mem_cgroup_largest_soft_limit_node(mctz);
572 	spin_unlock_irq(&mctz->lock);
573 	return mz;
574 }
575 
576 /*
577  * memcg and lruvec stats flushing
578  *
579  * Many codepaths leading to stats update or read are performance sensitive and
580  * adding stats flushing in such codepaths is not desirable. So, to optimize the
581  * flushing the kernel does:
582  *
583  * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
584  *    rstat update tree grow unbounded.
585  *
586  * 2) Flush the stats synchronously on reader side only when there are more than
587  *    (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
588  *    will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
589  *    only for 2 seconds due to (1).
590  */
591 static void flush_memcg_stats_dwork(struct work_struct *w);
592 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
593 static DEFINE_PER_CPU(unsigned int, stats_updates);
594 static atomic_t stats_flush_ongoing = ATOMIC_INIT(0);
595 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
596 static u64 flush_next_time;
597 
598 #define FLUSH_TIME (2UL*HZ)
599 
600 /*
601  * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
602  * not rely on this as part of an acquired spinlock_t lock. These functions are
603  * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
604  * is sufficient.
605  */
606 static void memcg_stats_lock(void)
607 {
608 	preempt_disable_nested();
609 	VM_WARN_ON_IRQS_ENABLED();
610 }
611 
612 static void __memcg_stats_lock(void)
613 {
614 	preempt_disable_nested();
615 }
616 
617 static void memcg_stats_unlock(void)
618 {
619 	preempt_enable_nested();
620 }
621 
622 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
623 {
624 	unsigned int x;
625 
626 	if (!val)
627 		return;
628 
629 	cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
630 
631 	x = __this_cpu_add_return(stats_updates, abs(val));
632 	if (x > MEMCG_CHARGE_BATCH) {
633 		/*
634 		 * If stats_flush_threshold exceeds the threshold
635 		 * (>num_online_cpus()), cgroup stats update will be triggered
636 		 * in __mem_cgroup_flush_stats(). Increasing this var further
637 		 * is redundant and simply adds overhead in atomic update.
638 		 */
639 		if (atomic_read(&stats_flush_threshold) <= num_online_cpus())
640 			atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
641 		__this_cpu_write(stats_updates, 0);
642 	}
643 }
644 
645 static void do_flush_stats(void)
646 {
647 	/*
648 	 * We always flush the entire tree, so concurrent flushers can just
649 	 * skip. This avoids a thundering herd problem on the rstat global lock
650 	 * from memcg flushers (e.g. reclaim, refault, etc).
651 	 */
652 	if (atomic_read(&stats_flush_ongoing) ||
653 	    atomic_xchg(&stats_flush_ongoing, 1))
654 		return;
655 
656 	WRITE_ONCE(flush_next_time, jiffies_64 + 2*FLUSH_TIME);
657 
658 	cgroup_rstat_flush(root_mem_cgroup->css.cgroup);
659 
660 	atomic_set(&stats_flush_threshold, 0);
661 	atomic_set(&stats_flush_ongoing, 0);
662 }
663 
664 void mem_cgroup_flush_stats(void)
665 {
666 	if (atomic_read(&stats_flush_threshold) > num_online_cpus())
667 		do_flush_stats();
668 }
669 
670 void mem_cgroup_flush_stats_ratelimited(void)
671 {
672 	if (time_after64(jiffies_64, READ_ONCE(flush_next_time)))
673 		mem_cgroup_flush_stats();
674 }
675 
676 static void flush_memcg_stats_dwork(struct work_struct *w)
677 {
678 	/*
679 	 * Always flush here so that flushing in latency-sensitive paths is
680 	 * as cheap as possible.
681 	 */
682 	do_flush_stats();
683 	queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
684 }
685 
686 /* Subset of vm_event_item to report for memcg event stats */
687 static const unsigned int memcg_vm_event_stat[] = {
688 	PGPGIN,
689 	PGPGOUT,
690 	PGSCAN_KSWAPD,
691 	PGSCAN_DIRECT,
692 	PGSCAN_KHUGEPAGED,
693 	PGSTEAL_KSWAPD,
694 	PGSTEAL_DIRECT,
695 	PGSTEAL_KHUGEPAGED,
696 	PGFAULT,
697 	PGMAJFAULT,
698 	PGREFILL,
699 	PGACTIVATE,
700 	PGDEACTIVATE,
701 	PGLAZYFREE,
702 	PGLAZYFREED,
703 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
704 	ZSWPIN,
705 	ZSWPOUT,
706 #endif
707 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
708 	THP_FAULT_ALLOC,
709 	THP_COLLAPSE_ALLOC,
710 	THP_SWPOUT,
711 	THP_SWPOUT_FALLBACK,
712 #endif
713 };
714 
715 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
716 static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
717 
718 static void init_memcg_events(void)
719 {
720 	int i;
721 
722 	for (i = 0; i < NR_MEMCG_EVENTS; ++i)
723 		mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
724 }
725 
726 static inline int memcg_events_index(enum vm_event_item idx)
727 {
728 	return mem_cgroup_events_index[idx] - 1;
729 }
730 
731 struct memcg_vmstats_percpu {
732 	/* Local (CPU and cgroup) page state & events */
733 	long			state[MEMCG_NR_STAT];
734 	unsigned long		events[NR_MEMCG_EVENTS];
735 
736 	/* Delta calculation for lockless upward propagation */
737 	long			state_prev[MEMCG_NR_STAT];
738 	unsigned long		events_prev[NR_MEMCG_EVENTS];
739 
740 	/* Cgroup1: threshold notifications & softlimit tree updates */
741 	unsigned long		nr_page_events;
742 	unsigned long		targets[MEM_CGROUP_NTARGETS];
743 };
744 
745 struct memcg_vmstats {
746 	/* Aggregated (CPU and subtree) page state & events */
747 	long			state[MEMCG_NR_STAT];
748 	unsigned long		events[NR_MEMCG_EVENTS];
749 
750 	/* Non-hierarchical (CPU aggregated) page state & events */
751 	long			state_local[MEMCG_NR_STAT];
752 	unsigned long		events_local[NR_MEMCG_EVENTS];
753 
754 	/* Pending child counts during tree propagation */
755 	long			state_pending[MEMCG_NR_STAT];
756 	unsigned long		events_pending[NR_MEMCG_EVENTS];
757 };
758 
759 unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
760 {
761 	long x = READ_ONCE(memcg->vmstats->state[idx]);
762 #ifdef CONFIG_SMP
763 	if (x < 0)
764 		x = 0;
765 #endif
766 	return x;
767 }
768 
769 static int memcg_page_state_unit(int item);
770 
771 /*
772  * Normalize the value passed into memcg_rstat_updated() to be in pages. Round
773  * up non-zero sub-page updates to 1 page as zero page updates are ignored.
774  */
775 static int memcg_state_val_in_pages(int idx, int val)
776 {
777 	int unit = memcg_page_state_unit(idx);
778 
779 	if (!val || unit == PAGE_SIZE)
780 		return val;
781 	else
782 		return max(val * unit / PAGE_SIZE, 1UL);
783 }
784 
785 /**
786  * __mod_memcg_state - update cgroup memory statistics
787  * @memcg: the memory cgroup
788  * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
789  * @val: delta to add to the counter, can be negative
790  */
791 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
792 {
793 	if (mem_cgroup_disabled())
794 		return;
795 
796 	__this_cpu_add(memcg->vmstats_percpu->state[idx], val);
797 	memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
798 }
799 
800 /* idx can be of type enum memcg_stat_item or node_stat_item. */
801 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
802 {
803 	long x = READ_ONCE(memcg->vmstats->state_local[idx]);
804 
805 #ifdef CONFIG_SMP
806 	if (x < 0)
807 		x = 0;
808 #endif
809 	return x;
810 }
811 
812 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
813 			      int val)
814 {
815 	struct mem_cgroup_per_node *pn;
816 	struct mem_cgroup *memcg;
817 
818 	pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
819 	memcg = pn->memcg;
820 
821 	/*
822 	 * The caller from rmap relies on disabled preemption because they never
823 	 * update their counter from in-interrupt context. For these two
824 	 * counters we check that the update is never performed from an
825 	 * interrupt context while other caller need to have disabled interrupt.
826 	 */
827 	__memcg_stats_lock();
828 	if (IS_ENABLED(CONFIG_DEBUG_VM)) {
829 		switch (idx) {
830 		case NR_ANON_MAPPED:
831 		case NR_FILE_MAPPED:
832 		case NR_ANON_THPS:
833 		case NR_SHMEM_PMDMAPPED:
834 		case NR_FILE_PMDMAPPED:
835 			WARN_ON_ONCE(!in_task());
836 			break;
837 		default:
838 			VM_WARN_ON_IRQS_ENABLED();
839 		}
840 	}
841 
842 	/* Update memcg */
843 	__this_cpu_add(memcg->vmstats_percpu->state[idx], val);
844 
845 	/* Update lruvec */
846 	__this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
847 
848 	memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
849 	memcg_stats_unlock();
850 }
851 
852 /**
853  * __mod_lruvec_state - update lruvec memory statistics
854  * @lruvec: the lruvec
855  * @idx: the stat item
856  * @val: delta to add to the counter, can be negative
857  *
858  * The lruvec is the intersection of the NUMA node and a cgroup. This
859  * function updates the all three counters that are affected by a
860  * change of state at this level: per-node, per-cgroup, per-lruvec.
861  */
862 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
863 			int val)
864 {
865 	/* Update node */
866 	__mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
867 
868 	/* Update memcg and lruvec */
869 	if (!mem_cgroup_disabled())
870 		__mod_memcg_lruvec_state(lruvec, idx, val);
871 }
872 
873 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
874 			     int val)
875 {
876 	struct page *head = compound_head(page); /* rmap on tail pages */
877 	struct mem_cgroup *memcg;
878 	pg_data_t *pgdat = page_pgdat(page);
879 	struct lruvec *lruvec;
880 
881 	rcu_read_lock();
882 	memcg = page_memcg(head);
883 	/* Untracked pages have no memcg, no lruvec. Update only the node */
884 	if (!memcg) {
885 		rcu_read_unlock();
886 		__mod_node_page_state(pgdat, idx, val);
887 		return;
888 	}
889 
890 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
891 	__mod_lruvec_state(lruvec, idx, val);
892 	rcu_read_unlock();
893 }
894 EXPORT_SYMBOL(__mod_lruvec_page_state);
895 
896 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
897 {
898 	pg_data_t *pgdat = page_pgdat(virt_to_page(p));
899 	struct mem_cgroup *memcg;
900 	struct lruvec *lruvec;
901 
902 	rcu_read_lock();
903 	memcg = mem_cgroup_from_slab_obj(p);
904 
905 	/*
906 	 * Untracked pages have no memcg, no lruvec. Update only the
907 	 * node. If we reparent the slab objects to the root memcg,
908 	 * when we free the slab object, we need to update the per-memcg
909 	 * vmstats to keep it correct for the root memcg.
910 	 */
911 	if (!memcg) {
912 		__mod_node_page_state(pgdat, idx, val);
913 	} else {
914 		lruvec = mem_cgroup_lruvec(memcg, pgdat);
915 		__mod_lruvec_state(lruvec, idx, val);
916 	}
917 	rcu_read_unlock();
918 }
919 
920 /**
921  * __count_memcg_events - account VM events in a cgroup
922  * @memcg: the memory cgroup
923  * @idx: the event item
924  * @count: the number of events that occurred
925  */
926 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
927 			  unsigned long count)
928 {
929 	int index = memcg_events_index(idx);
930 
931 	if (mem_cgroup_disabled() || index < 0)
932 		return;
933 
934 	memcg_stats_lock();
935 	__this_cpu_add(memcg->vmstats_percpu->events[index], count);
936 	memcg_rstat_updated(memcg, count);
937 	memcg_stats_unlock();
938 }
939 
940 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
941 {
942 	int index = memcg_events_index(event);
943 
944 	if (index < 0)
945 		return 0;
946 	return READ_ONCE(memcg->vmstats->events[index]);
947 }
948 
949 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
950 {
951 	int index = memcg_events_index(event);
952 
953 	if (index < 0)
954 		return 0;
955 
956 	return READ_ONCE(memcg->vmstats->events_local[index]);
957 }
958 
959 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
960 					 int nr_pages)
961 {
962 	/* pagein of a big page is an event. So, ignore page size */
963 	if (nr_pages > 0)
964 		__count_memcg_events(memcg, PGPGIN, 1);
965 	else {
966 		__count_memcg_events(memcg, PGPGOUT, 1);
967 		nr_pages = -nr_pages; /* for event */
968 	}
969 
970 	__this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
971 }
972 
973 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
974 				       enum mem_cgroup_events_target target)
975 {
976 	unsigned long val, next;
977 
978 	val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
979 	next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
980 	/* from time_after() in jiffies.h */
981 	if ((long)(next - val) < 0) {
982 		switch (target) {
983 		case MEM_CGROUP_TARGET_THRESH:
984 			next = val + THRESHOLDS_EVENTS_TARGET;
985 			break;
986 		case MEM_CGROUP_TARGET_SOFTLIMIT:
987 			next = val + SOFTLIMIT_EVENTS_TARGET;
988 			break;
989 		default:
990 			break;
991 		}
992 		__this_cpu_write(memcg->vmstats_percpu->targets[target], next);
993 		return true;
994 	}
995 	return false;
996 }
997 
998 /*
999  * Check events in order.
1000  *
1001  */
1002 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
1003 {
1004 	if (IS_ENABLED(CONFIG_PREEMPT_RT))
1005 		return;
1006 
1007 	/* threshold event is triggered in finer grain than soft limit */
1008 	if (unlikely(mem_cgroup_event_ratelimit(memcg,
1009 						MEM_CGROUP_TARGET_THRESH))) {
1010 		bool do_softlimit;
1011 
1012 		do_softlimit = mem_cgroup_event_ratelimit(memcg,
1013 						MEM_CGROUP_TARGET_SOFTLIMIT);
1014 		mem_cgroup_threshold(memcg);
1015 		if (unlikely(do_softlimit))
1016 			mem_cgroup_update_tree(memcg, nid);
1017 	}
1018 }
1019 
1020 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1021 {
1022 	/*
1023 	 * mm_update_next_owner() may clear mm->owner to NULL
1024 	 * if it races with swapoff, page migration, etc.
1025 	 * So this can be called with p == NULL.
1026 	 */
1027 	if (unlikely(!p))
1028 		return NULL;
1029 
1030 	return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1031 }
1032 EXPORT_SYMBOL(mem_cgroup_from_task);
1033 
1034 static __always_inline struct mem_cgroup *active_memcg(void)
1035 {
1036 	if (!in_task())
1037 		return this_cpu_read(int_active_memcg);
1038 	else
1039 		return current->active_memcg;
1040 }
1041 
1042 /**
1043  * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1044  * @mm: mm from which memcg should be extracted. It can be NULL.
1045  *
1046  * Obtain a reference on mm->memcg and returns it if successful. If mm
1047  * is NULL, then the memcg is chosen as follows:
1048  * 1) The active memcg, if set.
1049  * 2) current->mm->memcg, if available
1050  * 3) root memcg
1051  * If mem_cgroup is disabled, NULL is returned.
1052  */
1053 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1054 {
1055 	struct mem_cgroup *memcg;
1056 
1057 	if (mem_cgroup_disabled())
1058 		return NULL;
1059 
1060 	/*
1061 	 * Page cache insertions can happen without an
1062 	 * actual mm context, e.g. during disk probing
1063 	 * on boot, loopback IO, acct() writes etc.
1064 	 *
1065 	 * No need to css_get on root memcg as the reference
1066 	 * counting is disabled on the root level in the
1067 	 * cgroup core. See CSS_NO_REF.
1068 	 */
1069 	if (unlikely(!mm)) {
1070 		memcg = active_memcg();
1071 		if (unlikely(memcg)) {
1072 			/* remote memcg must hold a ref */
1073 			css_get(&memcg->css);
1074 			return memcg;
1075 		}
1076 		mm = current->mm;
1077 		if (unlikely(!mm))
1078 			return root_mem_cgroup;
1079 	}
1080 
1081 	rcu_read_lock();
1082 	do {
1083 		memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1084 		if (unlikely(!memcg))
1085 			memcg = root_mem_cgroup;
1086 	} while (!css_tryget(&memcg->css));
1087 	rcu_read_unlock();
1088 	return memcg;
1089 }
1090 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1091 
1092 /**
1093  * get_mem_cgroup_from_current - Obtain a reference on current task's memcg.
1094  */
1095 struct mem_cgroup *get_mem_cgroup_from_current(void)
1096 {
1097 	struct mem_cgroup *memcg;
1098 
1099 	if (mem_cgroup_disabled())
1100 		return NULL;
1101 
1102 again:
1103 	rcu_read_lock();
1104 	memcg = mem_cgroup_from_task(current);
1105 	if (!css_tryget(&memcg->css)) {
1106 		rcu_read_unlock();
1107 		goto again;
1108 	}
1109 	rcu_read_unlock();
1110 	return memcg;
1111 }
1112 
1113 /**
1114  * mem_cgroup_iter - iterate over memory cgroup hierarchy
1115  * @root: hierarchy root
1116  * @prev: previously returned memcg, NULL on first invocation
1117  * @reclaim: cookie for shared reclaim walks, NULL for full walks
1118  *
1119  * Returns references to children of the hierarchy below @root, or
1120  * @root itself, or %NULL after a full round-trip.
1121  *
1122  * Caller must pass the return value in @prev on subsequent
1123  * invocations for reference counting, or use mem_cgroup_iter_break()
1124  * to cancel a hierarchy walk before the round-trip is complete.
1125  *
1126  * Reclaimers can specify a node in @reclaim to divide up the memcgs
1127  * in the hierarchy among all concurrent reclaimers operating on the
1128  * same node.
1129  */
1130 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1131 				   struct mem_cgroup *prev,
1132 				   struct mem_cgroup_reclaim_cookie *reclaim)
1133 {
1134 	struct mem_cgroup_reclaim_iter *iter;
1135 	struct cgroup_subsys_state *css = NULL;
1136 	struct mem_cgroup *memcg = NULL;
1137 	struct mem_cgroup *pos = NULL;
1138 
1139 	if (mem_cgroup_disabled())
1140 		return NULL;
1141 
1142 	if (!root)
1143 		root = root_mem_cgroup;
1144 
1145 	rcu_read_lock();
1146 
1147 	if (reclaim) {
1148 		struct mem_cgroup_per_node *mz;
1149 
1150 		mz = root->nodeinfo[reclaim->pgdat->node_id];
1151 		iter = &mz->iter;
1152 
1153 		/*
1154 		 * On start, join the current reclaim iteration cycle.
1155 		 * Exit when a concurrent walker completes it.
1156 		 */
1157 		if (!prev)
1158 			reclaim->generation = iter->generation;
1159 		else if (reclaim->generation != iter->generation)
1160 			goto out_unlock;
1161 
1162 		while (1) {
1163 			pos = READ_ONCE(iter->position);
1164 			if (!pos || css_tryget(&pos->css))
1165 				break;
1166 			/*
1167 			 * css reference reached zero, so iter->position will
1168 			 * be cleared by ->css_released. However, we should not
1169 			 * rely on this happening soon, because ->css_released
1170 			 * is called from a work queue, and by busy-waiting we
1171 			 * might block it. So we clear iter->position right
1172 			 * away.
1173 			 */
1174 			(void)cmpxchg(&iter->position, pos, NULL);
1175 		}
1176 	} else if (prev) {
1177 		pos = prev;
1178 	}
1179 
1180 	if (pos)
1181 		css = &pos->css;
1182 
1183 	for (;;) {
1184 		css = css_next_descendant_pre(css, &root->css);
1185 		if (!css) {
1186 			/*
1187 			 * Reclaimers share the hierarchy walk, and a
1188 			 * new one might jump in right at the end of
1189 			 * the hierarchy - make sure they see at least
1190 			 * one group and restart from the beginning.
1191 			 */
1192 			if (!prev)
1193 				continue;
1194 			break;
1195 		}
1196 
1197 		/*
1198 		 * Verify the css and acquire a reference.  The root
1199 		 * is provided by the caller, so we know it's alive
1200 		 * and kicking, and don't take an extra reference.
1201 		 */
1202 		if (css == &root->css || css_tryget(css)) {
1203 			memcg = mem_cgroup_from_css(css);
1204 			break;
1205 		}
1206 	}
1207 
1208 	if (reclaim) {
1209 		/*
1210 		 * The position could have already been updated by a competing
1211 		 * thread, so check that the value hasn't changed since we read
1212 		 * it to avoid reclaiming from the same cgroup twice.
1213 		 */
1214 		(void)cmpxchg(&iter->position, pos, memcg);
1215 
1216 		if (pos)
1217 			css_put(&pos->css);
1218 
1219 		if (!memcg)
1220 			iter->generation++;
1221 	}
1222 
1223 out_unlock:
1224 	rcu_read_unlock();
1225 	if (prev && prev != root)
1226 		css_put(&prev->css);
1227 
1228 	return memcg;
1229 }
1230 
1231 /**
1232  * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1233  * @root: hierarchy root
1234  * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1235  */
1236 void mem_cgroup_iter_break(struct mem_cgroup *root,
1237 			   struct mem_cgroup *prev)
1238 {
1239 	if (!root)
1240 		root = root_mem_cgroup;
1241 	if (prev && prev != root)
1242 		css_put(&prev->css);
1243 }
1244 
1245 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1246 					struct mem_cgroup *dead_memcg)
1247 {
1248 	struct mem_cgroup_reclaim_iter *iter;
1249 	struct mem_cgroup_per_node *mz;
1250 	int nid;
1251 
1252 	for_each_node(nid) {
1253 		mz = from->nodeinfo[nid];
1254 		iter = &mz->iter;
1255 		cmpxchg(&iter->position, dead_memcg, NULL);
1256 	}
1257 }
1258 
1259 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1260 {
1261 	struct mem_cgroup *memcg = dead_memcg;
1262 	struct mem_cgroup *last;
1263 
1264 	do {
1265 		__invalidate_reclaim_iterators(memcg, dead_memcg);
1266 		last = memcg;
1267 	} while ((memcg = parent_mem_cgroup(memcg)));
1268 
1269 	/*
1270 	 * When cgroup1 non-hierarchy mode is used,
1271 	 * parent_mem_cgroup() does not walk all the way up to the
1272 	 * cgroup root (root_mem_cgroup). So we have to handle
1273 	 * dead_memcg from cgroup root separately.
1274 	 */
1275 	if (!mem_cgroup_is_root(last))
1276 		__invalidate_reclaim_iterators(root_mem_cgroup,
1277 						dead_memcg);
1278 }
1279 
1280 /**
1281  * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1282  * @memcg: hierarchy root
1283  * @fn: function to call for each task
1284  * @arg: argument passed to @fn
1285  *
1286  * This function iterates over tasks attached to @memcg or to any of its
1287  * descendants and calls @fn for each task. If @fn returns a non-zero
1288  * value, the function breaks the iteration loop. Otherwise, it will iterate
1289  * over all tasks and return 0.
1290  *
1291  * This function must not be called for the root memory cgroup.
1292  */
1293 void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1294 			   int (*fn)(struct task_struct *, void *), void *arg)
1295 {
1296 	struct mem_cgroup *iter;
1297 	int ret = 0;
1298 
1299 	BUG_ON(mem_cgroup_is_root(memcg));
1300 
1301 	for_each_mem_cgroup_tree(iter, memcg) {
1302 		struct css_task_iter it;
1303 		struct task_struct *task;
1304 
1305 		css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1306 		while (!ret && (task = css_task_iter_next(&it)))
1307 			ret = fn(task, arg);
1308 		css_task_iter_end(&it);
1309 		if (ret) {
1310 			mem_cgroup_iter_break(memcg, iter);
1311 			break;
1312 		}
1313 	}
1314 }
1315 
1316 #ifdef CONFIG_DEBUG_VM
1317 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1318 {
1319 	struct mem_cgroup *memcg;
1320 
1321 	if (mem_cgroup_disabled())
1322 		return;
1323 
1324 	memcg = folio_memcg(folio);
1325 
1326 	if (!memcg)
1327 		VM_BUG_ON_FOLIO(!mem_cgroup_is_root(lruvec_memcg(lruvec)), folio);
1328 	else
1329 		VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1330 }
1331 #endif
1332 
1333 /**
1334  * folio_lruvec_lock - Lock the lruvec for a folio.
1335  * @folio: Pointer to the folio.
1336  *
1337  * These functions are safe to use under any of the following conditions:
1338  * - folio locked
1339  * - folio_test_lru false
1340  * - folio_memcg_lock()
1341  * - folio frozen (refcount of 0)
1342  *
1343  * Return: The lruvec this folio is on with its lock held.
1344  */
1345 struct lruvec *folio_lruvec_lock(struct folio *folio)
1346 {
1347 	struct lruvec *lruvec = folio_lruvec(folio);
1348 
1349 	spin_lock(&lruvec->lru_lock);
1350 	lruvec_memcg_debug(lruvec, folio);
1351 
1352 	return lruvec;
1353 }
1354 
1355 /**
1356  * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1357  * @folio: Pointer to the folio.
1358  *
1359  * These functions are safe to use under any of the following conditions:
1360  * - folio locked
1361  * - folio_test_lru false
1362  * - folio_memcg_lock()
1363  * - folio frozen (refcount of 0)
1364  *
1365  * Return: The lruvec this folio is on with its lock held and interrupts
1366  * disabled.
1367  */
1368 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1369 {
1370 	struct lruvec *lruvec = folio_lruvec(folio);
1371 
1372 	spin_lock_irq(&lruvec->lru_lock);
1373 	lruvec_memcg_debug(lruvec, folio);
1374 
1375 	return lruvec;
1376 }
1377 
1378 /**
1379  * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1380  * @folio: Pointer to the folio.
1381  * @flags: Pointer to irqsave flags.
1382  *
1383  * These functions are safe to use under any of the following conditions:
1384  * - folio locked
1385  * - folio_test_lru false
1386  * - folio_memcg_lock()
1387  * - folio frozen (refcount of 0)
1388  *
1389  * Return: The lruvec this folio is on with its lock held and interrupts
1390  * disabled.
1391  */
1392 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1393 		unsigned long *flags)
1394 {
1395 	struct lruvec *lruvec = folio_lruvec(folio);
1396 
1397 	spin_lock_irqsave(&lruvec->lru_lock, *flags);
1398 	lruvec_memcg_debug(lruvec, folio);
1399 
1400 	return lruvec;
1401 }
1402 
1403 /**
1404  * mem_cgroup_update_lru_size - account for adding or removing an lru page
1405  * @lruvec: mem_cgroup per zone lru vector
1406  * @lru: index of lru list the page is sitting on
1407  * @zid: zone id of the accounted pages
1408  * @nr_pages: positive when adding or negative when removing
1409  *
1410  * This function must be called under lru_lock, just before a page is added
1411  * to or just after a page is removed from an lru list.
1412  */
1413 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1414 				int zid, int nr_pages)
1415 {
1416 	struct mem_cgroup_per_node *mz;
1417 	unsigned long *lru_size;
1418 	long size;
1419 
1420 	if (mem_cgroup_disabled())
1421 		return;
1422 
1423 	mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1424 	lru_size = &mz->lru_zone_size[zid][lru];
1425 
1426 	if (nr_pages < 0)
1427 		*lru_size += nr_pages;
1428 
1429 	size = *lru_size;
1430 	if (WARN_ONCE(size < 0,
1431 		"%s(%p, %d, %d): lru_size %ld\n",
1432 		__func__, lruvec, lru, nr_pages, size)) {
1433 		VM_BUG_ON(1);
1434 		*lru_size = 0;
1435 	}
1436 
1437 	if (nr_pages > 0)
1438 		*lru_size += nr_pages;
1439 }
1440 
1441 /**
1442  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1443  * @memcg: the memory cgroup
1444  *
1445  * Returns the maximum amount of memory @mem can be charged with, in
1446  * pages.
1447  */
1448 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1449 {
1450 	unsigned long margin = 0;
1451 	unsigned long count;
1452 	unsigned long limit;
1453 
1454 	count = page_counter_read(&memcg->memory);
1455 	limit = READ_ONCE(memcg->memory.max);
1456 	if (count < limit)
1457 		margin = limit - count;
1458 
1459 	if (do_memsw_account()) {
1460 		count = page_counter_read(&memcg->memsw);
1461 		limit = READ_ONCE(memcg->memsw.max);
1462 		if (count < limit)
1463 			margin = min(margin, limit - count);
1464 		else
1465 			margin = 0;
1466 	}
1467 
1468 	return margin;
1469 }
1470 
1471 /*
1472  * A routine for checking "mem" is under move_account() or not.
1473  *
1474  * Checking a cgroup is mc.from or mc.to or under hierarchy of
1475  * moving cgroups. This is for waiting at high-memory pressure
1476  * caused by "move".
1477  */
1478 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1479 {
1480 	struct mem_cgroup *from;
1481 	struct mem_cgroup *to;
1482 	bool ret = false;
1483 	/*
1484 	 * Unlike task_move routines, we access mc.to, mc.from not under
1485 	 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1486 	 */
1487 	spin_lock(&mc.lock);
1488 	from = mc.from;
1489 	to = mc.to;
1490 	if (!from)
1491 		goto unlock;
1492 
1493 	ret = mem_cgroup_is_descendant(from, memcg) ||
1494 		mem_cgroup_is_descendant(to, memcg);
1495 unlock:
1496 	spin_unlock(&mc.lock);
1497 	return ret;
1498 }
1499 
1500 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1501 {
1502 	if (mc.moving_task && current != mc.moving_task) {
1503 		if (mem_cgroup_under_move(memcg)) {
1504 			DEFINE_WAIT(wait);
1505 			prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1506 			/* moving charge context might have finished. */
1507 			if (mc.moving_task)
1508 				schedule();
1509 			finish_wait(&mc.waitq, &wait);
1510 			return true;
1511 		}
1512 	}
1513 	return false;
1514 }
1515 
1516 struct memory_stat {
1517 	const char *name;
1518 	unsigned int idx;
1519 };
1520 
1521 static const struct memory_stat memory_stats[] = {
1522 	{ "anon",			NR_ANON_MAPPED			},
1523 	{ "file",			NR_FILE_PAGES			},
1524 	{ "kernel",			MEMCG_KMEM			},
1525 	{ "kernel_stack",		NR_KERNEL_STACK_KB		},
1526 	{ "pagetables",			NR_PAGETABLE			},
1527 	{ "sec_pagetables",		NR_SECONDARY_PAGETABLE		},
1528 	{ "percpu",			MEMCG_PERCPU_B			},
1529 	{ "sock",			MEMCG_SOCK			},
1530 	{ "vmalloc",			MEMCG_VMALLOC			},
1531 	{ "shmem",			NR_SHMEM			},
1532 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1533 	{ "zswap",			MEMCG_ZSWAP_B			},
1534 	{ "zswapped",			MEMCG_ZSWAPPED			},
1535 #endif
1536 	{ "file_mapped",		NR_FILE_MAPPED			},
1537 	{ "file_dirty",			NR_FILE_DIRTY			},
1538 	{ "file_writeback",		NR_WRITEBACK			},
1539 #ifdef CONFIG_SWAP
1540 	{ "swapcached",			NR_SWAPCACHE			},
1541 #endif
1542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1543 	{ "anon_thp",			NR_ANON_THPS			},
1544 	{ "file_thp",			NR_FILE_THPS			},
1545 	{ "shmem_thp",			NR_SHMEM_THPS			},
1546 #endif
1547 	{ "inactive_anon",		NR_INACTIVE_ANON		},
1548 	{ "active_anon",		NR_ACTIVE_ANON			},
1549 	{ "inactive_file",		NR_INACTIVE_FILE		},
1550 	{ "active_file",		NR_ACTIVE_FILE			},
1551 	{ "unevictable",		NR_UNEVICTABLE			},
1552 	{ "slab_reclaimable",		NR_SLAB_RECLAIMABLE_B		},
1553 	{ "slab_unreclaimable",		NR_SLAB_UNRECLAIMABLE_B		},
1554 
1555 	/* The memory events */
1556 	{ "workingset_refault_anon",	WORKINGSET_REFAULT_ANON		},
1557 	{ "workingset_refault_file",	WORKINGSET_REFAULT_FILE		},
1558 	{ "workingset_activate_anon",	WORKINGSET_ACTIVATE_ANON	},
1559 	{ "workingset_activate_file",	WORKINGSET_ACTIVATE_FILE	},
1560 	{ "workingset_restore_anon",	WORKINGSET_RESTORE_ANON		},
1561 	{ "workingset_restore_file",	WORKINGSET_RESTORE_FILE		},
1562 	{ "workingset_nodereclaim",	WORKINGSET_NODERECLAIM		},
1563 };
1564 
1565 /* The actual unit of the state item, not the same as the output unit */
1566 static int memcg_page_state_unit(int item)
1567 {
1568 	switch (item) {
1569 	case MEMCG_PERCPU_B:
1570 	case MEMCG_ZSWAP_B:
1571 	case NR_SLAB_RECLAIMABLE_B:
1572 	case NR_SLAB_UNRECLAIMABLE_B:
1573 		return 1;
1574 	case NR_KERNEL_STACK_KB:
1575 		return SZ_1K;
1576 	default:
1577 		return PAGE_SIZE;
1578 	}
1579 }
1580 
1581 /* Translate stat items to the correct unit for memory.stat output */
1582 static int memcg_page_state_output_unit(int item)
1583 {
1584 	/*
1585 	 * Workingset state is actually in pages, but we export it to userspace
1586 	 * as a scalar count of events, so special case it here.
1587 	 */
1588 	switch (item) {
1589 	case WORKINGSET_REFAULT_ANON:
1590 	case WORKINGSET_REFAULT_FILE:
1591 	case WORKINGSET_ACTIVATE_ANON:
1592 	case WORKINGSET_ACTIVATE_FILE:
1593 	case WORKINGSET_RESTORE_ANON:
1594 	case WORKINGSET_RESTORE_FILE:
1595 	case WORKINGSET_NODERECLAIM:
1596 		return 1;
1597 	default:
1598 		return memcg_page_state_unit(item);
1599 	}
1600 }
1601 
1602 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1603 						    int item)
1604 {
1605 	return memcg_page_state(memcg, item) *
1606 		memcg_page_state_output_unit(item);
1607 }
1608 
1609 static inline unsigned long memcg_page_state_local_output(
1610 		struct mem_cgroup *memcg, int item)
1611 {
1612 	return memcg_page_state_local(memcg, item) *
1613 		memcg_page_state_output_unit(item);
1614 }
1615 
1616 static void memcg_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1617 {
1618 	int i;
1619 
1620 	/*
1621 	 * Provide statistics on the state of the memory subsystem as
1622 	 * well as cumulative event counters that show past behavior.
1623 	 *
1624 	 * This list is ordered following a combination of these gradients:
1625 	 * 1) generic big picture -> specifics and details
1626 	 * 2) reflecting userspace activity -> reflecting kernel heuristics
1627 	 *
1628 	 * Current memory state:
1629 	 */
1630 	mem_cgroup_flush_stats();
1631 
1632 	for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1633 		u64 size;
1634 
1635 		size = memcg_page_state_output(memcg, memory_stats[i].idx);
1636 		seq_buf_printf(s, "%s %llu\n", memory_stats[i].name, size);
1637 
1638 		if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1639 			size += memcg_page_state_output(memcg,
1640 							NR_SLAB_RECLAIMABLE_B);
1641 			seq_buf_printf(s, "slab %llu\n", size);
1642 		}
1643 	}
1644 
1645 	/* Accumulated memory events */
1646 	seq_buf_printf(s, "pgscan %lu\n",
1647 		       memcg_events(memcg, PGSCAN_KSWAPD) +
1648 		       memcg_events(memcg, PGSCAN_DIRECT) +
1649 		       memcg_events(memcg, PGSCAN_KHUGEPAGED));
1650 	seq_buf_printf(s, "pgsteal %lu\n",
1651 		       memcg_events(memcg, PGSTEAL_KSWAPD) +
1652 		       memcg_events(memcg, PGSTEAL_DIRECT) +
1653 		       memcg_events(memcg, PGSTEAL_KHUGEPAGED));
1654 
1655 	for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) {
1656 		if (memcg_vm_event_stat[i] == PGPGIN ||
1657 		    memcg_vm_event_stat[i] == PGPGOUT)
1658 			continue;
1659 
1660 		seq_buf_printf(s, "%s %lu\n",
1661 			       vm_event_name(memcg_vm_event_stat[i]),
1662 			       memcg_events(memcg, memcg_vm_event_stat[i]));
1663 	}
1664 
1665 	/* The above should easily fit into one page */
1666 	WARN_ON_ONCE(seq_buf_has_overflowed(s));
1667 }
1668 
1669 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s);
1670 
1671 static void memory_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1672 {
1673 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1674 		memcg_stat_format(memcg, s);
1675 	else
1676 		memcg1_stat_format(memcg, s);
1677 	WARN_ON_ONCE(seq_buf_has_overflowed(s));
1678 }
1679 
1680 /**
1681  * mem_cgroup_print_oom_context: Print OOM information relevant to
1682  * memory controller.
1683  * @memcg: The memory cgroup that went over limit
1684  * @p: Task that is going to be killed
1685  *
1686  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1687  * enabled
1688  */
1689 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1690 {
1691 	rcu_read_lock();
1692 
1693 	if (memcg) {
1694 		pr_cont(",oom_memcg=");
1695 		pr_cont_cgroup_path(memcg->css.cgroup);
1696 	} else
1697 		pr_cont(",global_oom");
1698 	if (p) {
1699 		pr_cont(",task_memcg=");
1700 		pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1701 	}
1702 	rcu_read_unlock();
1703 }
1704 
1705 /**
1706  * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1707  * memory controller.
1708  * @memcg: The memory cgroup that went over limit
1709  */
1710 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1711 {
1712 	/* Use static buffer, for the caller is holding oom_lock. */
1713 	static char buf[PAGE_SIZE];
1714 	struct seq_buf s;
1715 
1716 	lockdep_assert_held(&oom_lock);
1717 
1718 	pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1719 		K((u64)page_counter_read(&memcg->memory)),
1720 		K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1721 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1722 		pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1723 			K((u64)page_counter_read(&memcg->swap)),
1724 			K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1725 	else {
1726 		pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1727 			K((u64)page_counter_read(&memcg->memsw)),
1728 			K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1729 		pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1730 			K((u64)page_counter_read(&memcg->kmem)),
1731 			K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1732 	}
1733 
1734 	pr_info("Memory cgroup stats for ");
1735 	pr_cont_cgroup_path(memcg->css.cgroup);
1736 	pr_cont(":");
1737 	seq_buf_init(&s, buf, sizeof(buf));
1738 	memory_stat_format(memcg, &s);
1739 	seq_buf_do_printk(&s, KERN_INFO);
1740 }
1741 
1742 /*
1743  * Return the memory (and swap, if configured) limit for a memcg.
1744  */
1745 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1746 {
1747 	unsigned long max = READ_ONCE(memcg->memory.max);
1748 
1749 	if (do_memsw_account()) {
1750 		if (mem_cgroup_swappiness(memcg)) {
1751 			/* Calculate swap excess capacity from memsw limit */
1752 			unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1753 
1754 			max += min(swap, (unsigned long)total_swap_pages);
1755 		}
1756 	} else {
1757 		if (mem_cgroup_swappiness(memcg))
1758 			max += min(READ_ONCE(memcg->swap.max),
1759 				   (unsigned long)total_swap_pages);
1760 	}
1761 	return max;
1762 }
1763 
1764 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1765 {
1766 	return page_counter_read(&memcg->memory);
1767 }
1768 
1769 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1770 				     int order)
1771 {
1772 	struct oom_control oc = {
1773 		.zonelist = NULL,
1774 		.nodemask = NULL,
1775 		.memcg = memcg,
1776 		.gfp_mask = gfp_mask,
1777 		.order = order,
1778 	};
1779 	bool ret = true;
1780 
1781 	if (mutex_lock_killable(&oom_lock))
1782 		return true;
1783 
1784 	if (mem_cgroup_margin(memcg) >= (1 << order))
1785 		goto unlock;
1786 
1787 	/*
1788 	 * A few threads which were not waiting at mutex_lock_killable() can
1789 	 * fail to bail out. Therefore, check again after holding oom_lock.
1790 	 */
1791 	ret = task_is_dying() || out_of_memory(&oc);
1792 
1793 unlock:
1794 	mutex_unlock(&oom_lock);
1795 	return ret;
1796 }
1797 
1798 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1799 				   pg_data_t *pgdat,
1800 				   gfp_t gfp_mask,
1801 				   unsigned long *total_scanned)
1802 {
1803 	struct mem_cgroup *victim = NULL;
1804 	int total = 0;
1805 	int loop = 0;
1806 	unsigned long excess;
1807 	unsigned long nr_scanned;
1808 	struct mem_cgroup_reclaim_cookie reclaim = {
1809 		.pgdat = pgdat,
1810 	};
1811 
1812 	excess = soft_limit_excess(root_memcg);
1813 
1814 	while (1) {
1815 		victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1816 		if (!victim) {
1817 			loop++;
1818 			if (loop >= 2) {
1819 				/*
1820 				 * If we have not been able to reclaim
1821 				 * anything, it might because there are
1822 				 * no reclaimable pages under this hierarchy
1823 				 */
1824 				if (!total)
1825 					break;
1826 				/*
1827 				 * We want to do more targeted reclaim.
1828 				 * excess >> 2 is not to excessive so as to
1829 				 * reclaim too much, nor too less that we keep
1830 				 * coming back to reclaim from this cgroup
1831 				 */
1832 				if (total >= (excess >> 2) ||
1833 					(loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1834 					break;
1835 			}
1836 			continue;
1837 		}
1838 		total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1839 					pgdat, &nr_scanned);
1840 		*total_scanned += nr_scanned;
1841 		if (!soft_limit_excess(root_memcg))
1842 			break;
1843 	}
1844 	mem_cgroup_iter_break(root_memcg, victim);
1845 	return total;
1846 }
1847 
1848 #ifdef CONFIG_LOCKDEP
1849 static struct lockdep_map memcg_oom_lock_dep_map = {
1850 	.name = "memcg_oom_lock",
1851 };
1852 #endif
1853 
1854 static DEFINE_SPINLOCK(memcg_oom_lock);
1855 
1856 /*
1857  * Check OOM-Killer is already running under our hierarchy.
1858  * If someone is running, return false.
1859  */
1860 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1861 {
1862 	struct mem_cgroup *iter, *failed = NULL;
1863 
1864 	spin_lock(&memcg_oom_lock);
1865 
1866 	for_each_mem_cgroup_tree(iter, memcg) {
1867 		if (iter->oom_lock) {
1868 			/*
1869 			 * this subtree of our hierarchy is already locked
1870 			 * so we cannot give a lock.
1871 			 */
1872 			failed = iter;
1873 			mem_cgroup_iter_break(memcg, iter);
1874 			break;
1875 		} else
1876 			iter->oom_lock = true;
1877 	}
1878 
1879 	if (failed) {
1880 		/*
1881 		 * OK, we failed to lock the whole subtree so we have
1882 		 * to clean up what we set up to the failing subtree
1883 		 */
1884 		for_each_mem_cgroup_tree(iter, memcg) {
1885 			if (iter == failed) {
1886 				mem_cgroup_iter_break(memcg, iter);
1887 				break;
1888 			}
1889 			iter->oom_lock = false;
1890 		}
1891 	} else
1892 		mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1893 
1894 	spin_unlock(&memcg_oom_lock);
1895 
1896 	return !failed;
1897 }
1898 
1899 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1900 {
1901 	struct mem_cgroup *iter;
1902 
1903 	spin_lock(&memcg_oom_lock);
1904 	mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1905 	for_each_mem_cgroup_tree(iter, memcg)
1906 		iter->oom_lock = false;
1907 	spin_unlock(&memcg_oom_lock);
1908 }
1909 
1910 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1911 {
1912 	struct mem_cgroup *iter;
1913 
1914 	spin_lock(&memcg_oom_lock);
1915 	for_each_mem_cgroup_tree(iter, memcg)
1916 		iter->under_oom++;
1917 	spin_unlock(&memcg_oom_lock);
1918 }
1919 
1920 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1921 {
1922 	struct mem_cgroup *iter;
1923 
1924 	/*
1925 	 * Be careful about under_oom underflows because a child memcg
1926 	 * could have been added after mem_cgroup_mark_under_oom.
1927 	 */
1928 	spin_lock(&memcg_oom_lock);
1929 	for_each_mem_cgroup_tree(iter, memcg)
1930 		if (iter->under_oom > 0)
1931 			iter->under_oom--;
1932 	spin_unlock(&memcg_oom_lock);
1933 }
1934 
1935 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1936 
1937 struct oom_wait_info {
1938 	struct mem_cgroup *memcg;
1939 	wait_queue_entry_t	wait;
1940 };
1941 
1942 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1943 	unsigned mode, int sync, void *arg)
1944 {
1945 	struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1946 	struct mem_cgroup *oom_wait_memcg;
1947 	struct oom_wait_info *oom_wait_info;
1948 
1949 	oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1950 	oom_wait_memcg = oom_wait_info->memcg;
1951 
1952 	if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1953 	    !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1954 		return 0;
1955 	return autoremove_wake_function(wait, mode, sync, arg);
1956 }
1957 
1958 static void memcg_oom_recover(struct mem_cgroup *memcg)
1959 {
1960 	/*
1961 	 * For the following lockless ->under_oom test, the only required
1962 	 * guarantee is that it must see the state asserted by an OOM when
1963 	 * this function is called as a result of userland actions
1964 	 * triggered by the notification of the OOM.  This is trivially
1965 	 * achieved by invoking mem_cgroup_mark_under_oom() before
1966 	 * triggering notification.
1967 	 */
1968 	if (memcg && memcg->under_oom)
1969 		__wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1970 }
1971 
1972 /*
1973  * Returns true if successfully killed one or more processes. Though in some
1974  * corner cases it can return true even without killing any process.
1975  */
1976 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1977 {
1978 	bool locked, ret;
1979 
1980 	if (order > PAGE_ALLOC_COSTLY_ORDER)
1981 		return false;
1982 
1983 	memcg_memory_event(memcg, MEMCG_OOM);
1984 
1985 	/*
1986 	 * We are in the middle of the charge context here, so we
1987 	 * don't want to block when potentially sitting on a callstack
1988 	 * that holds all kinds of filesystem and mm locks.
1989 	 *
1990 	 * cgroup1 allows disabling the OOM killer and waiting for outside
1991 	 * handling until the charge can succeed; remember the context and put
1992 	 * the task to sleep at the end of the page fault when all locks are
1993 	 * released.
1994 	 *
1995 	 * On the other hand, in-kernel OOM killer allows for an async victim
1996 	 * memory reclaim (oom_reaper) and that means that we are not solely
1997 	 * relying on the oom victim to make a forward progress and we can
1998 	 * invoke the oom killer here.
1999 	 *
2000 	 * Please note that mem_cgroup_out_of_memory might fail to find a
2001 	 * victim and then we have to bail out from the charge path.
2002 	 */
2003 	if (READ_ONCE(memcg->oom_kill_disable)) {
2004 		if (current->in_user_fault) {
2005 			css_get(&memcg->css);
2006 			current->memcg_in_oom = memcg;
2007 			current->memcg_oom_gfp_mask = mask;
2008 			current->memcg_oom_order = order;
2009 		}
2010 		return false;
2011 	}
2012 
2013 	mem_cgroup_mark_under_oom(memcg);
2014 
2015 	locked = mem_cgroup_oom_trylock(memcg);
2016 
2017 	if (locked)
2018 		mem_cgroup_oom_notify(memcg);
2019 
2020 	mem_cgroup_unmark_under_oom(memcg);
2021 	ret = mem_cgroup_out_of_memory(memcg, mask, order);
2022 
2023 	if (locked)
2024 		mem_cgroup_oom_unlock(memcg);
2025 
2026 	return ret;
2027 }
2028 
2029 /**
2030  * mem_cgroup_oom_synchronize - complete memcg OOM handling
2031  * @handle: actually kill/wait or just clean up the OOM state
2032  *
2033  * This has to be called at the end of a page fault if the memcg OOM
2034  * handler was enabled.
2035  *
2036  * Memcg supports userspace OOM handling where failed allocations must
2037  * sleep on a waitqueue until the userspace task resolves the
2038  * situation.  Sleeping directly in the charge context with all kinds
2039  * of locks held is not a good idea, instead we remember an OOM state
2040  * in the task and mem_cgroup_oom_synchronize() has to be called at
2041  * the end of the page fault to complete the OOM handling.
2042  *
2043  * Returns %true if an ongoing memcg OOM situation was detected and
2044  * completed, %false otherwise.
2045  */
2046 bool mem_cgroup_oom_synchronize(bool handle)
2047 {
2048 	struct mem_cgroup *memcg = current->memcg_in_oom;
2049 	struct oom_wait_info owait;
2050 	bool locked;
2051 
2052 	/* OOM is global, do not handle */
2053 	if (!memcg)
2054 		return false;
2055 
2056 	if (!handle)
2057 		goto cleanup;
2058 
2059 	owait.memcg = memcg;
2060 	owait.wait.flags = 0;
2061 	owait.wait.func = memcg_oom_wake_function;
2062 	owait.wait.private = current;
2063 	INIT_LIST_HEAD(&owait.wait.entry);
2064 
2065 	prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2066 	mem_cgroup_mark_under_oom(memcg);
2067 
2068 	locked = mem_cgroup_oom_trylock(memcg);
2069 
2070 	if (locked)
2071 		mem_cgroup_oom_notify(memcg);
2072 
2073 	schedule();
2074 	mem_cgroup_unmark_under_oom(memcg);
2075 	finish_wait(&memcg_oom_waitq, &owait.wait);
2076 
2077 	if (locked)
2078 		mem_cgroup_oom_unlock(memcg);
2079 cleanup:
2080 	current->memcg_in_oom = NULL;
2081 	css_put(&memcg->css);
2082 	return true;
2083 }
2084 
2085 /**
2086  * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2087  * @victim: task to be killed by the OOM killer
2088  * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2089  *
2090  * Returns a pointer to a memory cgroup, which has to be cleaned up
2091  * by killing all belonging OOM-killable tasks.
2092  *
2093  * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2094  */
2095 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2096 					    struct mem_cgroup *oom_domain)
2097 {
2098 	struct mem_cgroup *oom_group = NULL;
2099 	struct mem_cgroup *memcg;
2100 
2101 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2102 		return NULL;
2103 
2104 	if (!oom_domain)
2105 		oom_domain = root_mem_cgroup;
2106 
2107 	rcu_read_lock();
2108 
2109 	memcg = mem_cgroup_from_task(victim);
2110 	if (mem_cgroup_is_root(memcg))
2111 		goto out;
2112 
2113 	/*
2114 	 * If the victim task has been asynchronously moved to a different
2115 	 * memory cgroup, we might end up killing tasks outside oom_domain.
2116 	 * In this case it's better to ignore memory.group.oom.
2117 	 */
2118 	if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2119 		goto out;
2120 
2121 	/*
2122 	 * Traverse the memory cgroup hierarchy from the victim task's
2123 	 * cgroup up to the OOMing cgroup (or root) to find the
2124 	 * highest-level memory cgroup with oom.group set.
2125 	 */
2126 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2127 		if (READ_ONCE(memcg->oom_group))
2128 			oom_group = memcg;
2129 
2130 		if (memcg == oom_domain)
2131 			break;
2132 	}
2133 
2134 	if (oom_group)
2135 		css_get(&oom_group->css);
2136 out:
2137 	rcu_read_unlock();
2138 
2139 	return oom_group;
2140 }
2141 
2142 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2143 {
2144 	pr_info("Tasks in ");
2145 	pr_cont_cgroup_path(memcg->css.cgroup);
2146 	pr_cont(" are going to be killed due to memory.oom.group set\n");
2147 }
2148 
2149 /**
2150  * folio_memcg_lock - Bind a folio to its memcg.
2151  * @folio: The folio.
2152  *
2153  * This function prevents unlocked LRU folios from being moved to
2154  * another cgroup.
2155  *
2156  * It ensures lifetime of the bound memcg.  The caller is responsible
2157  * for the lifetime of the folio.
2158  */
2159 void folio_memcg_lock(struct folio *folio)
2160 {
2161 	struct mem_cgroup *memcg;
2162 	unsigned long flags;
2163 
2164 	/*
2165 	 * The RCU lock is held throughout the transaction.  The fast
2166 	 * path can get away without acquiring the memcg->move_lock
2167 	 * because page moving starts with an RCU grace period.
2168          */
2169 	rcu_read_lock();
2170 
2171 	if (mem_cgroup_disabled())
2172 		return;
2173 again:
2174 	memcg = folio_memcg(folio);
2175 	if (unlikely(!memcg))
2176 		return;
2177 
2178 #ifdef CONFIG_PROVE_LOCKING
2179 	local_irq_save(flags);
2180 	might_lock(&memcg->move_lock);
2181 	local_irq_restore(flags);
2182 #endif
2183 
2184 	if (atomic_read(&memcg->moving_account) <= 0)
2185 		return;
2186 
2187 	spin_lock_irqsave(&memcg->move_lock, flags);
2188 	if (memcg != folio_memcg(folio)) {
2189 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2190 		goto again;
2191 	}
2192 
2193 	/*
2194 	 * When charge migration first begins, we can have multiple
2195 	 * critical sections holding the fast-path RCU lock and one
2196 	 * holding the slowpath move_lock. Track the task who has the
2197 	 * move_lock for folio_memcg_unlock().
2198 	 */
2199 	memcg->move_lock_task = current;
2200 	memcg->move_lock_flags = flags;
2201 }
2202 
2203 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2204 {
2205 	if (memcg && memcg->move_lock_task == current) {
2206 		unsigned long flags = memcg->move_lock_flags;
2207 
2208 		memcg->move_lock_task = NULL;
2209 		memcg->move_lock_flags = 0;
2210 
2211 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2212 	}
2213 
2214 	rcu_read_unlock();
2215 }
2216 
2217 /**
2218  * folio_memcg_unlock - Release the binding between a folio and its memcg.
2219  * @folio: The folio.
2220  *
2221  * This releases the binding created by folio_memcg_lock().  This does
2222  * not change the accounting of this folio to its memcg, but it does
2223  * permit others to change it.
2224  */
2225 void folio_memcg_unlock(struct folio *folio)
2226 {
2227 	__folio_memcg_unlock(folio_memcg(folio));
2228 }
2229 
2230 struct memcg_stock_pcp {
2231 	local_lock_t stock_lock;
2232 	struct mem_cgroup *cached; /* this never be root cgroup */
2233 	unsigned int nr_pages;
2234 
2235 #ifdef CONFIG_MEMCG_KMEM
2236 	struct obj_cgroup *cached_objcg;
2237 	struct pglist_data *cached_pgdat;
2238 	unsigned int nr_bytes;
2239 	int nr_slab_reclaimable_b;
2240 	int nr_slab_unreclaimable_b;
2241 #endif
2242 
2243 	struct work_struct work;
2244 	unsigned long flags;
2245 #define FLUSHING_CACHED_CHARGE	0
2246 };
2247 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2248 	.stock_lock = INIT_LOCAL_LOCK(stock_lock),
2249 };
2250 static DEFINE_MUTEX(percpu_charge_mutex);
2251 
2252 #ifdef CONFIG_MEMCG_KMEM
2253 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2254 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2255 				     struct mem_cgroup *root_memcg);
2256 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2257 
2258 #else
2259 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2260 {
2261 	return NULL;
2262 }
2263 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2264 				     struct mem_cgroup *root_memcg)
2265 {
2266 	return false;
2267 }
2268 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2269 {
2270 }
2271 #endif
2272 
2273 /**
2274  * consume_stock: Try to consume stocked charge on this cpu.
2275  * @memcg: memcg to consume from.
2276  * @nr_pages: how many pages to charge.
2277  *
2278  * The charges will only happen if @memcg matches the current cpu's memcg
2279  * stock, and at least @nr_pages are available in that stock.  Failure to
2280  * service an allocation will refill the stock.
2281  *
2282  * returns true if successful, false otherwise.
2283  */
2284 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2285 {
2286 	struct memcg_stock_pcp *stock;
2287 	unsigned long flags;
2288 	bool ret = false;
2289 
2290 	if (nr_pages > MEMCG_CHARGE_BATCH)
2291 		return ret;
2292 
2293 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
2294 
2295 	stock = this_cpu_ptr(&memcg_stock);
2296 	if (memcg == READ_ONCE(stock->cached) && stock->nr_pages >= nr_pages) {
2297 		stock->nr_pages -= nr_pages;
2298 		ret = true;
2299 	}
2300 
2301 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2302 
2303 	return ret;
2304 }
2305 
2306 /*
2307  * Returns stocks cached in percpu and reset cached information.
2308  */
2309 static void drain_stock(struct memcg_stock_pcp *stock)
2310 {
2311 	struct mem_cgroup *old = READ_ONCE(stock->cached);
2312 
2313 	if (!old)
2314 		return;
2315 
2316 	if (stock->nr_pages) {
2317 		page_counter_uncharge(&old->memory, stock->nr_pages);
2318 		if (do_memsw_account())
2319 			page_counter_uncharge(&old->memsw, stock->nr_pages);
2320 		stock->nr_pages = 0;
2321 	}
2322 
2323 	css_put(&old->css);
2324 	WRITE_ONCE(stock->cached, NULL);
2325 }
2326 
2327 static void drain_local_stock(struct work_struct *dummy)
2328 {
2329 	struct memcg_stock_pcp *stock;
2330 	struct obj_cgroup *old = NULL;
2331 	unsigned long flags;
2332 
2333 	/*
2334 	 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2335 	 * drain_stock races is that we always operate on local CPU stock
2336 	 * here with IRQ disabled
2337 	 */
2338 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
2339 
2340 	stock = this_cpu_ptr(&memcg_stock);
2341 	old = drain_obj_stock(stock);
2342 	drain_stock(stock);
2343 	clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2344 
2345 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2346 	if (old)
2347 		obj_cgroup_put(old);
2348 }
2349 
2350 /*
2351  * Cache charges(val) to local per_cpu area.
2352  * This will be consumed by consume_stock() function, later.
2353  */
2354 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2355 {
2356 	struct memcg_stock_pcp *stock;
2357 
2358 	stock = this_cpu_ptr(&memcg_stock);
2359 	if (READ_ONCE(stock->cached) != memcg) { /* reset if necessary */
2360 		drain_stock(stock);
2361 		css_get(&memcg->css);
2362 		WRITE_ONCE(stock->cached, memcg);
2363 	}
2364 	stock->nr_pages += nr_pages;
2365 
2366 	if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2367 		drain_stock(stock);
2368 }
2369 
2370 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2371 {
2372 	unsigned long flags;
2373 
2374 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
2375 	__refill_stock(memcg, nr_pages);
2376 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2377 }
2378 
2379 /*
2380  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2381  * of the hierarchy under it.
2382  */
2383 static void drain_all_stock(struct mem_cgroup *root_memcg)
2384 {
2385 	int cpu, curcpu;
2386 
2387 	/* If someone's already draining, avoid adding running more workers. */
2388 	if (!mutex_trylock(&percpu_charge_mutex))
2389 		return;
2390 	/*
2391 	 * Notify other cpus that system-wide "drain" is running
2392 	 * We do not care about races with the cpu hotplug because cpu down
2393 	 * as well as workers from this path always operate on the local
2394 	 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2395 	 */
2396 	migrate_disable();
2397 	curcpu = smp_processor_id();
2398 	for_each_online_cpu(cpu) {
2399 		struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2400 		struct mem_cgroup *memcg;
2401 		bool flush = false;
2402 
2403 		rcu_read_lock();
2404 		memcg = READ_ONCE(stock->cached);
2405 		if (memcg && stock->nr_pages &&
2406 		    mem_cgroup_is_descendant(memcg, root_memcg))
2407 			flush = true;
2408 		else if (obj_stock_flush_required(stock, root_memcg))
2409 			flush = true;
2410 		rcu_read_unlock();
2411 
2412 		if (flush &&
2413 		    !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2414 			if (cpu == curcpu)
2415 				drain_local_stock(&stock->work);
2416 			else if (!cpu_is_isolated(cpu))
2417 				schedule_work_on(cpu, &stock->work);
2418 		}
2419 	}
2420 	migrate_enable();
2421 	mutex_unlock(&percpu_charge_mutex);
2422 }
2423 
2424 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2425 {
2426 	struct memcg_stock_pcp *stock;
2427 
2428 	stock = &per_cpu(memcg_stock, cpu);
2429 	drain_stock(stock);
2430 
2431 	return 0;
2432 }
2433 
2434 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2435 				  unsigned int nr_pages,
2436 				  gfp_t gfp_mask)
2437 {
2438 	unsigned long nr_reclaimed = 0;
2439 
2440 	do {
2441 		unsigned long pflags;
2442 
2443 		if (page_counter_read(&memcg->memory) <=
2444 		    READ_ONCE(memcg->memory.high))
2445 			continue;
2446 
2447 		memcg_memory_event(memcg, MEMCG_HIGH);
2448 
2449 		psi_memstall_enter(&pflags);
2450 		nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2451 							gfp_mask,
2452 							MEMCG_RECLAIM_MAY_SWAP);
2453 		psi_memstall_leave(&pflags);
2454 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2455 		 !mem_cgroup_is_root(memcg));
2456 
2457 	return nr_reclaimed;
2458 }
2459 
2460 static void high_work_func(struct work_struct *work)
2461 {
2462 	struct mem_cgroup *memcg;
2463 
2464 	memcg = container_of(work, struct mem_cgroup, high_work);
2465 	reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2466 }
2467 
2468 /*
2469  * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2470  * enough to still cause a significant slowdown in most cases, while still
2471  * allowing diagnostics and tracing to proceed without becoming stuck.
2472  */
2473 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2474 
2475 /*
2476  * When calculating the delay, we use these either side of the exponentiation to
2477  * maintain precision and scale to a reasonable number of jiffies (see the table
2478  * below.
2479  *
2480  * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2481  *   overage ratio to a delay.
2482  * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2483  *   proposed penalty in order to reduce to a reasonable number of jiffies, and
2484  *   to produce a reasonable delay curve.
2485  *
2486  * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2487  * reasonable delay curve compared to precision-adjusted overage, not
2488  * penalising heavily at first, but still making sure that growth beyond the
2489  * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2490  * example, with a high of 100 megabytes:
2491  *
2492  *  +-------+------------------------+
2493  *  | usage | time to allocate in ms |
2494  *  +-------+------------------------+
2495  *  | 100M  |                      0 |
2496  *  | 101M  |                      6 |
2497  *  | 102M  |                     25 |
2498  *  | 103M  |                     57 |
2499  *  | 104M  |                    102 |
2500  *  | 105M  |                    159 |
2501  *  | 106M  |                    230 |
2502  *  | 107M  |                    313 |
2503  *  | 108M  |                    409 |
2504  *  | 109M  |                    518 |
2505  *  | 110M  |                    639 |
2506  *  | 111M  |                    774 |
2507  *  | 112M  |                    921 |
2508  *  | 113M  |                   1081 |
2509  *  | 114M  |                   1254 |
2510  *  | 115M  |                   1439 |
2511  *  | 116M  |                   1638 |
2512  *  | 117M  |                   1849 |
2513  *  | 118M  |                   2000 |
2514  *  | 119M  |                   2000 |
2515  *  | 120M  |                   2000 |
2516  *  +-------+------------------------+
2517  */
2518  #define MEMCG_DELAY_PRECISION_SHIFT 20
2519  #define MEMCG_DELAY_SCALING_SHIFT 14
2520 
2521 static u64 calculate_overage(unsigned long usage, unsigned long high)
2522 {
2523 	u64 overage;
2524 
2525 	if (usage <= high)
2526 		return 0;
2527 
2528 	/*
2529 	 * Prevent division by 0 in overage calculation by acting as if
2530 	 * it was a threshold of 1 page
2531 	 */
2532 	high = max(high, 1UL);
2533 
2534 	overage = usage - high;
2535 	overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2536 	return div64_u64(overage, high);
2537 }
2538 
2539 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2540 {
2541 	u64 overage, max_overage = 0;
2542 
2543 	do {
2544 		overage = calculate_overage(page_counter_read(&memcg->memory),
2545 					    READ_ONCE(memcg->memory.high));
2546 		max_overage = max(overage, max_overage);
2547 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2548 		 !mem_cgroup_is_root(memcg));
2549 
2550 	return max_overage;
2551 }
2552 
2553 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2554 {
2555 	u64 overage, max_overage = 0;
2556 
2557 	do {
2558 		overage = calculate_overage(page_counter_read(&memcg->swap),
2559 					    READ_ONCE(memcg->swap.high));
2560 		if (overage)
2561 			memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2562 		max_overage = max(overage, max_overage);
2563 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2564 		 !mem_cgroup_is_root(memcg));
2565 
2566 	return max_overage;
2567 }
2568 
2569 /*
2570  * Get the number of jiffies that we should penalise a mischievous cgroup which
2571  * is exceeding its memory.high by checking both it and its ancestors.
2572  */
2573 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2574 					  unsigned int nr_pages,
2575 					  u64 max_overage)
2576 {
2577 	unsigned long penalty_jiffies;
2578 
2579 	if (!max_overage)
2580 		return 0;
2581 
2582 	/*
2583 	 * We use overage compared to memory.high to calculate the number of
2584 	 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2585 	 * fairly lenient on small overages, and increasingly harsh when the
2586 	 * memcg in question makes it clear that it has no intention of stopping
2587 	 * its crazy behaviour, so we exponentially increase the delay based on
2588 	 * overage amount.
2589 	 */
2590 	penalty_jiffies = max_overage * max_overage * HZ;
2591 	penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2592 	penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2593 
2594 	/*
2595 	 * Factor in the task's own contribution to the overage, such that four
2596 	 * N-sized allocations are throttled approximately the same as one
2597 	 * 4N-sized allocation.
2598 	 *
2599 	 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2600 	 * larger the current charge patch is than that.
2601 	 */
2602 	return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2603 }
2604 
2605 /*
2606  * Scheduled by try_charge() to be executed from the userland return path
2607  * and reclaims memory over the high limit.
2608  */
2609 void mem_cgroup_handle_over_high(gfp_t gfp_mask)
2610 {
2611 	unsigned long penalty_jiffies;
2612 	unsigned long pflags;
2613 	unsigned long nr_reclaimed;
2614 	unsigned int nr_pages = current->memcg_nr_pages_over_high;
2615 	int nr_retries = MAX_RECLAIM_RETRIES;
2616 	struct mem_cgroup *memcg;
2617 	bool in_retry = false;
2618 
2619 	if (likely(!nr_pages))
2620 		return;
2621 
2622 	memcg = get_mem_cgroup_from_mm(current->mm);
2623 	current->memcg_nr_pages_over_high = 0;
2624 
2625 retry_reclaim:
2626 	/*
2627 	 * The allocating task should reclaim at least the batch size, but for
2628 	 * subsequent retries we only want to do what's necessary to prevent oom
2629 	 * or breaching resource isolation.
2630 	 *
2631 	 * This is distinct from memory.max or page allocator behaviour because
2632 	 * memory.high is currently batched, whereas memory.max and the page
2633 	 * allocator run every time an allocation is made.
2634 	 */
2635 	nr_reclaimed = reclaim_high(memcg,
2636 				    in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2637 				    gfp_mask);
2638 
2639 	/*
2640 	 * memory.high is breached and reclaim is unable to keep up. Throttle
2641 	 * allocators proactively to slow down excessive growth.
2642 	 */
2643 	penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2644 					       mem_find_max_overage(memcg));
2645 
2646 	penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2647 						swap_find_max_overage(memcg));
2648 
2649 	/*
2650 	 * Clamp the max delay per usermode return so as to still keep the
2651 	 * application moving forwards and also permit diagnostics, albeit
2652 	 * extremely slowly.
2653 	 */
2654 	penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2655 
2656 	/*
2657 	 * Don't sleep if the amount of jiffies this memcg owes us is so low
2658 	 * that it's not even worth doing, in an attempt to be nice to those who
2659 	 * go only a small amount over their memory.high value and maybe haven't
2660 	 * been aggressively reclaimed enough yet.
2661 	 */
2662 	if (penalty_jiffies <= HZ / 100)
2663 		goto out;
2664 
2665 	/*
2666 	 * If reclaim is making forward progress but we're still over
2667 	 * memory.high, we want to encourage that rather than doing allocator
2668 	 * throttling.
2669 	 */
2670 	if (nr_reclaimed || nr_retries--) {
2671 		in_retry = true;
2672 		goto retry_reclaim;
2673 	}
2674 
2675 	/*
2676 	 * If we exit early, we're guaranteed to die (since
2677 	 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2678 	 * need to account for any ill-begotten jiffies to pay them off later.
2679 	 */
2680 	psi_memstall_enter(&pflags);
2681 	schedule_timeout_killable(penalty_jiffies);
2682 	psi_memstall_leave(&pflags);
2683 
2684 out:
2685 	css_put(&memcg->css);
2686 }
2687 
2688 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2689 			unsigned int nr_pages)
2690 {
2691 	unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2692 	int nr_retries = MAX_RECLAIM_RETRIES;
2693 	struct mem_cgroup *mem_over_limit;
2694 	struct page_counter *counter;
2695 	unsigned long nr_reclaimed;
2696 	bool passed_oom = false;
2697 	unsigned int reclaim_options = MEMCG_RECLAIM_MAY_SWAP;
2698 	bool drained = false;
2699 	bool raised_max_event = false;
2700 	unsigned long pflags;
2701 
2702 retry:
2703 	if (consume_stock(memcg, nr_pages))
2704 		return 0;
2705 
2706 	if (!do_memsw_account() ||
2707 	    page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2708 		if (page_counter_try_charge(&memcg->memory, batch, &counter))
2709 			goto done_restock;
2710 		if (do_memsw_account())
2711 			page_counter_uncharge(&memcg->memsw, batch);
2712 		mem_over_limit = mem_cgroup_from_counter(counter, memory);
2713 	} else {
2714 		mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2715 		reclaim_options &= ~MEMCG_RECLAIM_MAY_SWAP;
2716 	}
2717 
2718 	if (batch > nr_pages) {
2719 		batch = nr_pages;
2720 		goto retry;
2721 	}
2722 
2723 	/*
2724 	 * Prevent unbounded recursion when reclaim operations need to
2725 	 * allocate memory. This might exceed the limits temporarily,
2726 	 * but we prefer facilitating memory reclaim and getting back
2727 	 * under the limit over triggering OOM kills in these cases.
2728 	 */
2729 	if (unlikely(current->flags & PF_MEMALLOC))
2730 		goto force;
2731 
2732 	if (unlikely(task_in_memcg_oom(current)))
2733 		goto nomem;
2734 
2735 	if (!gfpflags_allow_blocking(gfp_mask))
2736 		goto nomem;
2737 
2738 	memcg_memory_event(mem_over_limit, MEMCG_MAX);
2739 	raised_max_event = true;
2740 
2741 	psi_memstall_enter(&pflags);
2742 	nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2743 						    gfp_mask, reclaim_options);
2744 	psi_memstall_leave(&pflags);
2745 
2746 	if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2747 		goto retry;
2748 
2749 	if (!drained) {
2750 		drain_all_stock(mem_over_limit);
2751 		drained = true;
2752 		goto retry;
2753 	}
2754 
2755 	if (gfp_mask & __GFP_NORETRY)
2756 		goto nomem;
2757 	/*
2758 	 * Even though the limit is exceeded at this point, reclaim
2759 	 * may have been able to free some pages.  Retry the charge
2760 	 * before killing the task.
2761 	 *
2762 	 * Only for regular pages, though: huge pages are rather
2763 	 * unlikely to succeed so close to the limit, and we fall back
2764 	 * to regular pages anyway in case of failure.
2765 	 */
2766 	if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2767 		goto retry;
2768 	/*
2769 	 * At task move, charge accounts can be doubly counted. So, it's
2770 	 * better to wait until the end of task_move if something is going on.
2771 	 */
2772 	if (mem_cgroup_wait_acct_move(mem_over_limit))
2773 		goto retry;
2774 
2775 	if (nr_retries--)
2776 		goto retry;
2777 
2778 	if (gfp_mask & __GFP_RETRY_MAYFAIL)
2779 		goto nomem;
2780 
2781 	/* Avoid endless loop for tasks bypassed by the oom killer */
2782 	if (passed_oom && task_is_dying())
2783 		goto nomem;
2784 
2785 	/*
2786 	 * keep retrying as long as the memcg oom killer is able to make
2787 	 * a forward progress or bypass the charge if the oom killer
2788 	 * couldn't make any progress.
2789 	 */
2790 	if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2791 			   get_order(nr_pages * PAGE_SIZE))) {
2792 		passed_oom = true;
2793 		nr_retries = MAX_RECLAIM_RETRIES;
2794 		goto retry;
2795 	}
2796 nomem:
2797 	/*
2798 	 * Memcg doesn't have a dedicated reserve for atomic
2799 	 * allocations. But like the global atomic pool, we need to
2800 	 * put the burden of reclaim on regular allocation requests
2801 	 * and let these go through as privileged allocations.
2802 	 */
2803 	if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2804 		return -ENOMEM;
2805 force:
2806 	/*
2807 	 * If the allocation has to be enforced, don't forget to raise
2808 	 * a MEMCG_MAX event.
2809 	 */
2810 	if (!raised_max_event)
2811 		memcg_memory_event(mem_over_limit, MEMCG_MAX);
2812 
2813 	/*
2814 	 * The allocation either can't fail or will lead to more memory
2815 	 * being freed very soon.  Allow memory usage go over the limit
2816 	 * temporarily by force charging it.
2817 	 */
2818 	page_counter_charge(&memcg->memory, nr_pages);
2819 	if (do_memsw_account())
2820 		page_counter_charge(&memcg->memsw, nr_pages);
2821 
2822 	return 0;
2823 
2824 done_restock:
2825 	if (batch > nr_pages)
2826 		refill_stock(memcg, batch - nr_pages);
2827 
2828 	/*
2829 	 * If the hierarchy is above the normal consumption range, schedule
2830 	 * reclaim on returning to userland.  We can perform reclaim here
2831 	 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2832 	 * GFP_KERNEL can consistently be used during reclaim.  @memcg is
2833 	 * not recorded as it most likely matches current's and won't
2834 	 * change in the meantime.  As high limit is checked again before
2835 	 * reclaim, the cost of mismatch is negligible.
2836 	 */
2837 	do {
2838 		bool mem_high, swap_high;
2839 
2840 		mem_high = page_counter_read(&memcg->memory) >
2841 			READ_ONCE(memcg->memory.high);
2842 		swap_high = page_counter_read(&memcg->swap) >
2843 			READ_ONCE(memcg->swap.high);
2844 
2845 		/* Don't bother a random interrupted task */
2846 		if (!in_task()) {
2847 			if (mem_high) {
2848 				schedule_work(&memcg->high_work);
2849 				break;
2850 			}
2851 			continue;
2852 		}
2853 
2854 		if (mem_high || swap_high) {
2855 			/*
2856 			 * The allocating tasks in this cgroup will need to do
2857 			 * reclaim or be throttled to prevent further growth
2858 			 * of the memory or swap footprints.
2859 			 *
2860 			 * Target some best-effort fairness between the tasks,
2861 			 * and distribute reclaim work and delay penalties
2862 			 * based on how much each task is actually allocating.
2863 			 */
2864 			current->memcg_nr_pages_over_high += batch;
2865 			set_notify_resume(current);
2866 			break;
2867 		}
2868 	} while ((memcg = parent_mem_cgroup(memcg)));
2869 
2870 	if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2871 	    !(current->flags & PF_MEMALLOC) &&
2872 	    gfpflags_allow_blocking(gfp_mask)) {
2873 		mem_cgroup_handle_over_high(gfp_mask);
2874 	}
2875 	return 0;
2876 }
2877 
2878 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2879 			     unsigned int nr_pages)
2880 {
2881 	if (mem_cgroup_is_root(memcg))
2882 		return 0;
2883 
2884 	return try_charge_memcg(memcg, gfp_mask, nr_pages);
2885 }
2886 
2887 /**
2888  * mem_cgroup_cancel_charge() - cancel an uncommitted try_charge() call.
2889  * @memcg: memcg previously charged.
2890  * @nr_pages: number of pages previously charged.
2891  */
2892 void mem_cgroup_cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2893 {
2894 	if (mem_cgroup_is_root(memcg))
2895 		return;
2896 
2897 	page_counter_uncharge(&memcg->memory, nr_pages);
2898 	if (do_memsw_account())
2899 		page_counter_uncharge(&memcg->memsw, nr_pages);
2900 }
2901 
2902 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2903 {
2904 	VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2905 	/*
2906 	 * Any of the following ensures page's memcg stability:
2907 	 *
2908 	 * - the page lock
2909 	 * - LRU isolation
2910 	 * - folio_memcg_lock()
2911 	 * - exclusive reference
2912 	 * - mem_cgroup_trylock_pages()
2913 	 */
2914 	folio->memcg_data = (unsigned long)memcg;
2915 }
2916 
2917 /**
2918  * mem_cgroup_commit_charge - commit a previously successful try_charge().
2919  * @folio: folio to commit the charge to.
2920  * @memcg: memcg previously charged.
2921  */
2922 void mem_cgroup_commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2923 {
2924 	css_get(&memcg->css);
2925 	commit_charge(folio, memcg);
2926 
2927 	local_irq_disable();
2928 	mem_cgroup_charge_statistics(memcg, folio_nr_pages(folio));
2929 	memcg_check_events(memcg, folio_nid(folio));
2930 	local_irq_enable();
2931 }
2932 
2933 #ifdef CONFIG_MEMCG_KMEM
2934 /*
2935  * The allocated objcg pointers array is not accounted directly.
2936  * Moreover, it should not come from DMA buffer and is not readily
2937  * reclaimable. So those GFP bits should be masked off.
2938  */
2939 #define OBJCGS_CLEAR_MASK	(__GFP_DMA | __GFP_RECLAIMABLE | \
2940 				 __GFP_ACCOUNT | __GFP_NOFAIL)
2941 
2942 /*
2943  * mod_objcg_mlstate() may be called with irq enabled, so
2944  * mod_memcg_lruvec_state() should be used.
2945  */
2946 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2947 				     struct pglist_data *pgdat,
2948 				     enum node_stat_item idx, int nr)
2949 {
2950 	struct mem_cgroup *memcg;
2951 	struct lruvec *lruvec;
2952 
2953 	rcu_read_lock();
2954 	memcg = obj_cgroup_memcg(objcg);
2955 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
2956 	mod_memcg_lruvec_state(lruvec, idx, nr);
2957 	rcu_read_unlock();
2958 }
2959 
2960 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2961 				 gfp_t gfp, bool new_slab)
2962 {
2963 	unsigned int objects = objs_per_slab(s, slab);
2964 	unsigned long memcg_data;
2965 	void *vec;
2966 
2967 	gfp &= ~OBJCGS_CLEAR_MASK;
2968 	vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2969 			   slab_nid(slab));
2970 	if (!vec)
2971 		return -ENOMEM;
2972 
2973 	memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2974 	if (new_slab) {
2975 		/*
2976 		 * If the slab is brand new and nobody can yet access its
2977 		 * memcg_data, no synchronization is required and memcg_data can
2978 		 * be simply assigned.
2979 		 */
2980 		slab->memcg_data = memcg_data;
2981 	} else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2982 		/*
2983 		 * If the slab is already in use, somebody can allocate and
2984 		 * assign obj_cgroups in parallel. In this case the existing
2985 		 * objcg vector should be reused.
2986 		 */
2987 		kfree(vec);
2988 		return 0;
2989 	}
2990 
2991 	kmemleak_not_leak(vec);
2992 	return 0;
2993 }
2994 
2995 static __always_inline
2996 struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
2997 {
2998 	/*
2999 	 * Slab objects are accounted individually, not per-page.
3000 	 * Memcg membership data for each individual object is saved in
3001 	 * slab->memcg_data.
3002 	 */
3003 	if (folio_test_slab(folio)) {
3004 		struct obj_cgroup **objcgs;
3005 		struct slab *slab;
3006 		unsigned int off;
3007 
3008 		slab = folio_slab(folio);
3009 		objcgs = slab_objcgs(slab);
3010 		if (!objcgs)
3011 			return NULL;
3012 
3013 		off = obj_to_index(slab->slab_cache, slab, p);
3014 		if (objcgs[off])
3015 			return obj_cgroup_memcg(objcgs[off]);
3016 
3017 		return NULL;
3018 	}
3019 
3020 	/*
3021 	 * folio_memcg_check() is used here, because in theory we can encounter
3022 	 * a folio where the slab flag has been cleared already, but
3023 	 * slab->memcg_data has not been freed yet
3024 	 * folio_memcg_check() will guarantee that a proper memory
3025 	 * cgroup pointer or NULL will be returned.
3026 	 */
3027 	return folio_memcg_check(folio);
3028 }
3029 
3030 /*
3031  * Returns a pointer to the memory cgroup to which the kernel object is charged.
3032  *
3033  * A passed kernel object can be a slab object, vmalloc object or a generic
3034  * kernel page, so different mechanisms for getting the memory cgroup pointer
3035  * should be used.
3036  *
3037  * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
3038  * can not know for sure how the kernel object is implemented.
3039  * mem_cgroup_from_obj() can be safely used in such cases.
3040  *
3041  * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3042  * cgroup_mutex, etc.
3043  */
3044 struct mem_cgroup *mem_cgroup_from_obj(void *p)
3045 {
3046 	struct folio *folio;
3047 
3048 	if (mem_cgroup_disabled())
3049 		return NULL;
3050 
3051 	if (unlikely(is_vmalloc_addr(p)))
3052 		folio = page_folio(vmalloc_to_page(p));
3053 	else
3054 		folio = virt_to_folio(p);
3055 
3056 	return mem_cgroup_from_obj_folio(folio, p);
3057 }
3058 
3059 /*
3060  * Returns a pointer to the memory cgroup to which the kernel object is charged.
3061  * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
3062  * allocated using vmalloc().
3063  *
3064  * A passed kernel object must be a slab object or a generic kernel page.
3065  *
3066  * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3067  * cgroup_mutex, etc.
3068  */
3069 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
3070 {
3071 	if (mem_cgroup_disabled())
3072 		return NULL;
3073 
3074 	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
3075 }
3076 
3077 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
3078 {
3079 	struct obj_cgroup *objcg = NULL;
3080 
3081 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3082 		objcg = rcu_dereference(memcg->objcg);
3083 		if (likely(objcg && obj_cgroup_tryget(objcg)))
3084 			break;
3085 		objcg = NULL;
3086 	}
3087 	return objcg;
3088 }
3089 
3090 static struct obj_cgroup *current_objcg_update(void)
3091 {
3092 	struct mem_cgroup *memcg;
3093 	struct obj_cgroup *old, *objcg = NULL;
3094 
3095 	do {
3096 		/* Atomically drop the update bit. */
3097 		old = xchg(&current->objcg, NULL);
3098 		if (old) {
3099 			old = (struct obj_cgroup *)
3100 				((unsigned long)old & ~CURRENT_OBJCG_UPDATE_FLAG);
3101 			if (old)
3102 				obj_cgroup_put(old);
3103 
3104 			old = NULL;
3105 		}
3106 
3107 		/* If new objcg is NULL, no reason for the second atomic update. */
3108 		if (!current->mm || (current->flags & PF_KTHREAD))
3109 			return NULL;
3110 
3111 		/*
3112 		 * Release the objcg pointer from the previous iteration,
3113 		 * if try_cmpxcg() below fails.
3114 		 */
3115 		if (unlikely(objcg)) {
3116 			obj_cgroup_put(objcg);
3117 			objcg = NULL;
3118 		}
3119 
3120 		/*
3121 		 * Obtain the new objcg pointer. The current task can be
3122 		 * asynchronously moved to another memcg and the previous
3123 		 * memcg can be offlined. So let's get the memcg pointer
3124 		 * and try get a reference to objcg under a rcu read lock.
3125 		 */
3126 
3127 		rcu_read_lock();
3128 		memcg = mem_cgroup_from_task(current);
3129 		objcg = __get_obj_cgroup_from_memcg(memcg);
3130 		rcu_read_unlock();
3131 
3132 		/*
3133 		 * Try set up a new objcg pointer atomically. If it
3134 		 * fails, it means the update flag was set concurrently, so
3135 		 * the whole procedure should be repeated.
3136 		 */
3137 	} while (!try_cmpxchg(&current->objcg, &old, objcg));
3138 
3139 	return objcg;
3140 }
3141 
3142 __always_inline struct obj_cgroup *current_obj_cgroup(void)
3143 {
3144 	struct mem_cgroup *memcg;
3145 	struct obj_cgroup *objcg;
3146 
3147 	if (in_task()) {
3148 		memcg = current->active_memcg;
3149 		if (unlikely(memcg))
3150 			goto from_memcg;
3151 
3152 		objcg = READ_ONCE(current->objcg);
3153 		if (unlikely((unsigned long)objcg & CURRENT_OBJCG_UPDATE_FLAG))
3154 			objcg = current_objcg_update();
3155 		/*
3156 		 * Objcg reference is kept by the task, so it's safe
3157 		 * to use the objcg by the current task.
3158 		 */
3159 		return objcg;
3160 	}
3161 
3162 	memcg = this_cpu_read(int_active_memcg);
3163 	if (unlikely(memcg))
3164 		goto from_memcg;
3165 
3166 	return NULL;
3167 
3168 from_memcg:
3169 	objcg = NULL;
3170 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3171 		/*
3172 		 * Memcg pointer is protected by scope (see set_active_memcg())
3173 		 * and is pinning the corresponding objcg, so objcg can't go
3174 		 * away and can be used within the scope without any additional
3175 		 * protection.
3176 		 */
3177 		objcg = rcu_dereference_check(memcg->objcg, 1);
3178 		if (likely(objcg))
3179 			break;
3180 	}
3181 
3182 	return objcg;
3183 }
3184 
3185 struct obj_cgroup *get_obj_cgroup_from_folio(struct folio *folio)
3186 {
3187 	struct obj_cgroup *objcg;
3188 
3189 	if (!memcg_kmem_online())
3190 		return NULL;
3191 
3192 	if (folio_memcg_kmem(folio)) {
3193 		objcg = __folio_objcg(folio);
3194 		obj_cgroup_get(objcg);
3195 	} else {
3196 		struct mem_cgroup *memcg;
3197 
3198 		rcu_read_lock();
3199 		memcg = __folio_memcg(folio);
3200 		if (memcg)
3201 			objcg = __get_obj_cgroup_from_memcg(memcg);
3202 		else
3203 			objcg = NULL;
3204 		rcu_read_unlock();
3205 	}
3206 	return objcg;
3207 }
3208 
3209 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
3210 {
3211 	mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
3212 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
3213 		if (nr_pages > 0)
3214 			page_counter_charge(&memcg->kmem, nr_pages);
3215 		else
3216 			page_counter_uncharge(&memcg->kmem, -nr_pages);
3217 	}
3218 }
3219 
3220 
3221 /*
3222  * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3223  * @objcg: object cgroup to uncharge
3224  * @nr_pages: number of pages to uncharge
3225  */
3226 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3227 				      unsigned int nr_pages)
3228 {
3229 	struct mem_cgroup *memcg;
3230 
3231 	memcg = get_mem_cgroup_from_objcg(objcg);
3232 
3233 	memcg_account_kmem(memcg, -nr_pages);
3234 	refill_stock(memcg, nr_pages);
3235 
3236 	css_put(&memcg->css);
3237 }
3238 
3239 /*
3240  * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3241  * @objcg: object cgroup to charge
3242  * @gfp: reclaim mode
3243  * @nr_pages: number of pages to charge
3244  *
3245  * Returns 0 on success, an error code on failure.
3246  */
3247 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3248 				   unsigned int nr_pages)
3249 {
3250 	struct mem_cgroup *memcg;
3251 	int ret;
3252 
3253 	memcg = get_mem_cgroup_from_objcg(objcg);
3254 
3255 	ret = try_charge_memcg(memcg, gfp, nr_pages);
3256 	if (ret)
3257 		goto out;
3258 
3259 	memcg_account_kmem(memcg, nr_pages);
3260 out:
3261 	css_put(&memcg->css);
3262 
3263 	return ret;
3264 }
3265 
3266 /**
3267  * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3268  * @page: page to charge
3269  * @gfp: reclaim mode
3270  * @order: allocation order
3271  *
3272  * Returns 0 on success, an error code on failure.
3273  */
3274 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3275 {
3276 	struct obj_cgroup *objcg;
3277 	int ret = 0;
3278 
3279 	objcg = current_obj_cgroup();
3280 	if (objcg) {
3281 		ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3282 		if (!ret) {
3283 			obj_cgroup_get(objcg);
3284 			page->memcg_data = (unsigned long)objcg |
3285 				MEMCG_DATA_KMEM;
3286 			return 0;
3287 		}
3288 	}
3289 	return ret;
3290 }
3291 
3292 /**
3293  * __memcg_kmem_uncharge_page: uncharge a kmem page
3294  * @page: page to uncharge
3295  * @order: allocation order
3296  */
3297 void __memcg_kmem_uncharge_page(struct page *page, int order)
3298 {
3299 	struct folio *folio = page_folio(page);
3300 	struct obj_cgroup *objcg;
3301 	unsigned int nr_pages = 1 << order;
3302 
3303 	if (!folio_memcg_kmem(folio))
3304 		return;
3305 
3306 	objcg = __folio_objcg(folio);
3307 	obj_cgroup_uncharge_pages(objcg, nr_pages);
3308 	folio->memcg_data = 0;
3309 	obj_cgroup_put(objcg);
3310 }
3311 
3312 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3313 		     enum node_stat_item idx, int nr)
3314 {
3315 	struct memcg_stock_pcp *stock;
3316 	struct obj_cgroup *old = NULL;
3317 	unsigned long flags;
3318 	int *bytes;
3319 
3320 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
3321 	stock = this_cpu_ptr(&memcg_stock);
3322 
3323 	/*
3324 	 * Save vmstat data in stock and skip vmstat array update unless
3325 	 * accumulating over a page of vmstat data or when pgdat or idx
3326 	 * changes.
3327 	 */
3328 	if (READ_ONCE(stock->cached_objcg) != objcg) {
3329 		old = drain_obj_stock(stock);
3330 		obj_cgroup_get(objcg);
3331 		stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3332 				? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3333 		WRITE_ONCE(stock->cached_objcg, objcg);
3334 		stock->cached_pgdat = pgdat;
3335 	} else if (stock->cached_pgdat != pgdat) {
3336 		/* Flush the existing cached vmstat data */
3337 		struct pglist_data *oldpg = stock->cached_pgdat;
3338 
3339 		if (stock->nr_slab_reclaimable_b) {
3340 			mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3341 					  stock->nr_slab_reclaimable_b);
3342 			stock->nr_slab_reclaimable_b = 0;
3343 		}
3344 		if (stock->nr_slab_unreclaimable_b) {
3345 			mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3346 					  stock->nr_slab_unreclaimable_b);
3347 			stock->nr_slab_unreclaimable_b = 0;
3348 		}
3349 		stock->cached_pgdat = pgdat;
3350 	}
3351 
3352 	bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3353 					       : &stock->nr_slab_unreclaimable_b;
3354 	/*
3355 	 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3356 	 * cached locally at least once before pushing it out.
3357 	 */
3358 	if (!*bytes) {
3359 		*bytes = nr;
3360 		nr = 0;
3361 	} else {
3362 		*bytes += nr;
3363 		if (abs(*bytes) > PAGE_SIZE) {
3364 			nr = *bytes;
3365 			*bytes = 0;
3366 		} else {
3367 			nr = 0;
3368 		}
3369 	}
3370 	if (nr)
3371 		mod_objcg_mlstate(objcg, pgdat, idx, nr);
3372 
3373 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3374 	if (old)
3375 		obj_cgroup_put(old);
3376 }
3377 
3378 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3379 {
3380 	struct memcg_stock_pcp *stock;
3381 	unsigned long flags;
3382 	bool ret = false;
3383 
3384 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
3385 
3386 	stock = this_cpu_ptr(&memcg_stock);
3387 	if (objcg == READ_ONCE(stock->cached_objcg) && stock->nr_bytes >= nr_bytes) {
3388 		stock->nr_bytes -= nr_bytes;
3389 		ret = true;
3390 	}
3391 
3392 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3393 
3394 	return ret;
3395 }
3396 
3397 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3398 {
3399 	struct obj_cgroup *old = READ_ONCE(stock->cached_objcg);
3400 
3401 	if (!old)
3402 		return NULL;
3403 
3404 	if (stock->nr_bytes) {
3405 		unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3406 		unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3407 
3408 		if (nr_pages) {
3409 			struct mem_cgroup *memcg;
3410 
3411 			memcg = get_mem_cgroup_from_objcg(old);
3412 
3413 			memcg_account_kmem(memcg, -nr_pages);
3414 			__refill_stock(memcg, nr_pages);
3415 
3416 			css_put(&memcg->css);
3417 		}
3418 
3419 		/*
3420 		 * The leftover is flushed to the centralized per-memcg value.
3421 		 * On the next attempt to refill obj stock it will be moved
3422 		 * to a per-cpu stock (probably, on an other CPU), see
3423 		 * refill_obj_stock().
3424 		 *
3425 		 * How often it's flushed is a trade-off between the memory
3426 		 * limit enforcement accuracy and potential CPU contention,
3427 		 * so it might be changed in the future.
3428 		 */
3429 		atomic_add(nr_bytes, &old->nr_charged_bytes);
3430 		stock->nr_bytes = 0;
3431 	}
3432 
3433 	/*
3434 	 * Flush the vmstat data in current stock
3435 	 */
3436 	if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3437 		if (stock->nr_slab_reclaimable_b) {
3438 			mod_objcg_mlstate(old, stock->cached_pgdat,
3439 					  NR_SLAB_RECLAIMABLE_B,
3440 					  stock->nr_slab_reclaimable_b);
3441 			stock->nr_slab_reclaimable_b = 0;
3442 		}
3443 		if (stock->nr_slab_unreclaimable_b) {
3444 			mod_objcg_mlstate(old, stock->cached_pgdat,
3445 					  NR_SLAB_UNRECLAIMABLE_B,
3446 					  stock->nr_slab_unreclaimable_b);
3447 			stock->nr_slab_unreclaimable_b = 0;
3448 		}
3449 		stock->cached_pgdat = NULL;
3450 	}
3451 
3452 	WRITE_ONCE(stock->cached_objcg, NULL);
3453 	/*
3454 	 * The `old' objects needs to be released by the caller via
3455 	 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3456 	 */
3457 	return old;
3458 }
3459 
3460 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3461 				     struct mem_cgroup *root_memcg)
3462 {
3463 	struct obj_cgroup *objcg = READ_ONCE(stock->cached_objcg);
3464 	struct mem_cgroup *memcg;
3465 
3466 	if (objcg) {
3467 		memcg = obj_cgroup_memcg(objcg);
3468 		if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3469 			return true;
3470 	}
3471 
3472 	return false;
3473 }
3474 
3475 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3476 			     bool allow_uncharge)
3477 {
3478 	struct memcg_stock_pcp *stock;
3479 	struct obj_cgroup *old = NULL;
3480 	unsigned long flags;
3481 	unsigned int nr_pages = 0;
3482 
3483 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
3484 
3485 	stock = this_cpu_ptr(&memcg_stock);
3486 	if (READ_ONCE(stock->cached_objcg) != objcg) { /* reset if necessary */
3487 		old = drain_obj_stock(stock);
3488 		obj_cgroup_get(objcg);
3489 		WRITE_ONCE(stock->cached_objcg, objcg);
3490 		stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3491 				? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3492 		allow_uncharge = true;	/* Allow uncharge when objcg changes */
3493 	}
3494 	stock->nr_bytes += nr_bytes;
3495 
3496 	if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3497 		nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3498 		stock->nr_bytes &= (PAGE_SIZE - 1);
3499 	}
3500 
3501 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3502 	if (old)
3503 		obj_cgroup_put(old);
3504 
3505 	if (nr_pages)
3506 		obj_cgroup_uncharge_pages(objcg, nr_pages);
3507 }
3508 
3509 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3510 {
3511 	unsigned int nr_pages, nr_bytes;
3512 	int ret;
3513 
3514 	if (consume_obj_stock(objcg, size))
3515 		return 0;
3516 
3517 	/*
3518 	 * In theory, objcg->nr_charged_bytes can have enough
3519 	 * pre-charged bytes to satisfy the allocation. However,
3520 	 * flushing objcg->nr_charged_bytes requires two atomic
3521 	 * operations, and objcg->nr_charged_bytes can't be big.
3522 	 * The shared objcg->nr_charged_bytes can also become a
3523 	 * performance bottleneck if all tasks of the same memcg are
3524 	 * trying to update it. So it's better to ignore it and try
3525 	 * grab some new pages. The stock's nr_bytes will be flushed to
3526 	 * objcg->nr_charged_bytes later on when objcg changes.
3527 	 *
3528 	 * The stock's nr_bytes may contain enough pre-charged bytes
3529 	 * to allow one less page from being charged, but we can't rely
3530 	 * on the pre-charged bytes not being changed outside of
3531 	 * consume_obj_stock() or refill_obj_stock(). So ignore those
3532 	 * pre-charged bytes as well when charging pages. To avoid a
3533 	 * page uncharge right after a page charge, we set the
3534 	 * allow_uncharge flag to false when calling refill_obj_stock()
3535 	 * to temporarily allow the pre-charged bytes to exceed the page
3536 	 * size limit. The maximum reachable value of the pre-charged
3537 	 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3538 	 * race.
3539 	 */
3540 	nr_pages = size >> PAGE_SHIFT;
3541 	nr_bytes = size & (PAGE_SIZE - 1);
3542 
3543 	if (nr_bytes)
3544 		nr_pages += 1;
3545 
3546 	ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3547 	if (!ret && nr_bytes)
3548 		refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3549 
3550 	return ret;
3551 }
3552 
3553 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3554 {
3555 	refill_obj_stock(objcg, size, true);
3556 }
3557 
3558 #endif /* CONFIG_MEMCG_KMEM */
3559 
3560 /*
3561  * Because page_memcg(head) is not set on tails, set it now.
3562  */
3563 void split_page_memcg(struct page *head, unsigned int nr)
3564 {
3565 	struct folio *folio = page_folio(head);
3566 	struct mem_cgroup *memcg = folio_memcg(folio);
3567 	int i;
3568 
3569 	if (mem_cgroup_disabled() || !memcg)
3570 		return;
3571 
3572 	for (i = 1; i < nr; i++)
3573 		folio_page(folio, i)->memcg_data = folio->memcg_data;
3574 
3575 	if (folio_memcg_kmem(folio))
3576 		obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3577 	else
3578 		css_get_many(&memcg->css, nr - 1);
3579 }
3580 
3581 #ifdef CONFIG_SWAP
3582 /**
3583  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3584  * @entry: swap entry to be moved
3585  * @from:  mem_cgroup which the entry is moved from
3586  * @to:  mem_cgroup which the entry is moved to
3587  *
3588  * It succeeds only when the swap_cgroup's record for this entry is the same
3589  * as the mem_cgroup's id of @from.
3590  *
3591  * Returns 0 on success, -EINVAL on failure.
3592  *
3593  * The caller must have charged to @to, IOW, called page_counter_charge() about
3594  * both res and memsw, and called css_get().
3595  */
3596 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3597 				struct mem_cgroup *from, struct mem_cgroup *to)
3598 {
3599 	unsigned short old_id, new_id;
3600 
3601 	old_id = mem_cgroup_id(from);
3602 	new_id = mem_cgroup_id(to);
3603 
3604 	if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3605 		mod_memcg_state(from, MEMCG_SWAP, -1);
3606 		mod_memcg_state(to, MEMCG_SWAP, 1);
3607 		return 0;
3608 	}
3609 	return -EINVAL;
3610 }
3611 #else
3612 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3613 				struct mem_cgroup *from, struct mem_cgroup *to)
3614 {
3615 	return -EINVAL;
3616 }
3617 #endif
3618 
3619 static DEFINE_MUTEX(memcg_max_mutex);
3620 
3621 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3622 				 unsigned long max, bool memsw)
3623 {
3624 	bool enlarge = false;
3625 	bool drained = false;
3626 	int ret;
3627 	bool limits_invariant;
3628 	struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3629 
3630 	do {
3631 		if (signal_pending(current)) {
3632 			ret = -EINTR;
3633 			break;
3634 		}
3635 
3636 		mutex_lock(&memcg_max_mutex);
3637 		/*
3638 		 * Make sure that the new limit (memsw or memory limit) doesn't
3639 		 * break our basic invariant rule memory.max <= memsw.max.
3640 		 */
3641 		limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3642 					   max <= memcg->memsw.max;
3643 		if (!limits_invariant) {
3644 			mutex_unlock(&memcg_max_mutex);
3645 			ret = -EINVAL;
3646 			break;
3647 		}
3648 		if (max > counter->max)
3649 			enlarge = true;
3650 		ret = page_counter_set_max(counter, max);
3651 		mutex_unlock(&memcg_max_mutex);
3652 
3653 		if (!ret)
3654 			break;
3655 
3656 		if (!drained) {
3657 			drain_all_stock(memcg);
3658 			drained = true;
3659 			continue;
3660 		}
3661 
3662 		if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3663 					memsw ? 0 : MEMCG_RECLAIM_MAY_SWAP)) {
3664 			ret = -EBUSY;
3665 			break;
3666 		}
3667 	} while (true);
3668 
3669 	if (!ret && enlarge)
3670 		memcg_oom_recover(memcg);
3671 
3672 	return ret;
3673 }
3674 
3675 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3676 					    gfp_t gfp_mask,
3677 					    unsigned long *total_scanned)
3678 {
3679 	unsigned long nr_reclaimed = 0;
3680 	struct mem_cgroup_per_node *mz, *next_mz = NULL;
3681 	unsigned long reclaimed;
3682 	int loop = 0;
3683 	struct mem_cgroup_tree_per_node *mctz;
3684 	unsigned long excess;
3685 
3686 	if (lru_gen_enabled())
3687 		return 0;
3688 
3689 	if (order > 0)
3690 		return 0;
3691 
3692 	mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3693 
3694 	/*
3695 	 * Do not even bother to check the largest node if the root
3696 	 * is empty. Do it lockless to prevent lock bouncing. Races
3697 	 * are acceptable as soft limit is best effort anyway.
3698 	 */
3699 	if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3700 		return 0;
3701 
3702 	/*
3703 	 * This loop can run a while, specially if mem_cgroup's continuously
3704 	 * keep exceeding their soft limit and putting the system under
3705 	 * pressure
3706 	 */
3707 	do {
3708 		if (next_mz)
3709 			mz = next_mz;
3710 		else
3711 			mz = mem_cgroup_largest_soft_limit_node(mctz);
3712 		if (!mz)
3713 			break;
3714 
3715 		reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3716 						    gfp_mask, total_scanned);
3717 		nr_reclaimed += reclaimed;
3718 		spin_lock_irq(&mctz->lock);
3719 
3720 		/*
3721 		 * If we failed to reclaim anything from this memory cgroup
3722 		 * it is time to move on to the next cgroup
3723 		 */
3724 		next_mz = NULL;
3725 		if (!reclaimed)
3726 			next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3727 
3728 		excess = soft_limit_excess(mz->memcg);
3729 		/*
3730 		 * One school of thought says that we should not add
3731 		 * back the node to the tree if reclaim returns 0.
3732 		 * But our reclaim could return 0, simply because due
3733 		 * to priority we are exposing a smaller subset of
3734 		 * memory to reclaim from. Consider this as a longer
3735 		 * term TODO.
3736 		 */
3737 		/* If excess == 0, no tree ops */
3738 		__mem_cgroup_insert_exceeded(mz, mctz, excess);
3739 		spin_unlock_irq(&mctz->lock);
3740 		css_put(&mz->memcg->css);
3741 		loop++;
3742 		/*
3743 		 * Could not reclaim anything and there are no more
3744 		 * mem cgroups to try or we seem to be looping without
3745 		 * reclaiming anything.
3746 		 */
3747 		if (!nr_reclaimed &&
3748 			(next_mz == NULL ||
3749 			loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3750 			break;
3751 	} while (!nr_reclaimed);
3752 	if (next_mz)
3753 		css_put(&next_mz->memcg->css);
3754 	return nr_reclaimed;
3755 }
3756 
3757 /*
3758  * Reclaims as many pages from the given memcg as possible.
3759  *
3760  * Caller is responsible for holding css reference for memcg.
3761  */
3762 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3763 {
3764 	int nr_retries = MAX_RECLAIM_RETRIES;
3765 
3766 	/* we call try-to-free pages for make this cgroup empty */
3767 	lru_add_drain_all();
3768 
3769 	drain_all_stock(memcg);
3770 
3771 	/* try to free all pages in this cgroup */
3772 	while (nr_retries && page_counter_read(&memcg->memory)) {
3773 		if (signal_pending(current))
3774 			return -EINTR;
3775 
3776 		if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3777 						  MEMCG_RECLAIM_MAY_SWAP))
3778 			nr_retries--;
3779 	}
3780 
3781 	return 0;
3782 }
3783 
3784 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3785 					    char *buf, size_t nbytes,
3786 					    loff_t off)
3787 {
3788 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3789 
3790 	if (mem_cgroup_is_root(memcg))
3791 		return -EINVAL;
3792 	return mem_cgroup_force_empty(memcg) ?: nbytes;
3793 }
3794 
3795 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3796 				     struct cftype *cft)
3797 {
3798 	return 1;
3799 }
3800 
3801 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3802 				      struct cftype *cft, u64 val)
3803 {
3804 	if (val == 1)
3805 		return 0;
3806 
3807 	pr_warn_once("Non-hierarchical mode is deprecated. "
3808 		     "Please report your usecase to linux-mm@kvack.org if you "
3809 		     "depend on this functionality.\n");
3810 
3811 	return -EINVAL;
3812 }
3813 
3814 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3815 {
3816 	unsigned long val;
3817 
3818 	if (mem_cgroup_is_root(memcg)) {
3819 		/*
3820 		 * Approximate root's usage from global state. This isn't
3821 		 * perfect, but the root usage was always an approximation.
3822 		 */
3823 		val = global_node_page_state(NR_FILE_PAGES) +
3824 			global_node_page_state(NR_ANON_MAPPED);
3825 		if (swap)
3826 			val += total_swap_pages - get_nr_swap_pages();
3827 	} else {
3828 		if (!swap)
3829 			val = page_counter_read(&memcg->memory);
3830 		else
3831 			val = page_counter_read(&memcg->memsw);
3832 	}
3833 	return val;
3834 }
3835 
3836 enum {
3837 	RES_USAGE,
3838 	RES_LIMIT,
3839 	RES_MAX_USAGE,
3840 	RES_FAILCNT,
3841 	RES_SOFT_LIMIT,
3842 };
3843 
3844 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3845 			       struct cftype *cft)
3846 {
3847 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3848 	struct page_counter *counter;
3849 
3850 	switch (MEMFILE_TYPE(cft->private)) {
3851 	case _MEM:
3852 		counter = &memcg->memory;
3853 		break;
3854 	case _MEMSWAP:
3855 		counter = &memcg->memsw;
3856 		break;
3857 	case _KMEM:
3858 		counter = &memcg->kmem;
3859 		break;
3860 	case _TCP:
3861 		counter = &memcg->tcpmem;
3862 		break;
3863 	default:
3864 		BUG();
3865 	}
3866 
3867 	switch (MEMFILE_ATTR(cft->private)) {
3868 	case RES_USAGE:
3869 		if (counter == &memcg->memory)
3870 			return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3871 		if (counter == &memcg->memsw)
3872 			return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3873 		return (u64)page_counter_read(counter) * PAGE_SIZE;
3874 	case RES_LIMIT:
3875 		return (u64)counter->max * PAGE_SIZE;
3876 	case RES_MAX_USAGE:
3877 		return (u64)counter->watermark * PAGE_SIZE;
3878 	case RES_FAILCNT:
3879 		return counter->failcnt;
3880 	case RES_SOFT_LIMIT:
3881 		return (u64)READ_ONCE(memcg->soft_limit) * PAGE_SIZE;
3882 	default:
3883 		BUG();
3884 	}
3885 }
3886 
3887 /*
3888  * This function doesn't do anything useful. Its only job is to provide a read
3889  * handler for a file so that cgroup_file_mode() will add read permissions.
3890  */
3891 static int mem_cgroup_dummy_seq_show(__always_unused struct seq_file *m,
3892 				     __always_unused void *v)
3893 {
3894 	return -EINVAL;
3895 }
3896 
3897 #ifdef CONFIG_MEMCG_KMEM
3898 static int memcg_online_kmem(struct mem_cgroup *memcg)
3899 {
3900 	struct obj_cgroup *objcg;
3901 
3902 	if (mem_cgroup_kmem_disabled())
3903 		return 0;
3904 
3905 	if (unlikely(mem_cgroup_is_root(memcg)))
3906 		return 0;
3907 
3908 	objcg = obj_cgroup_alloc();
3909 	if (!objcg)
3910 		return -ENOMEM;
3911 
3912 	objcg->memcg = memcg;
3913 	rcu_assign_pointer(memcg->objcg, objcg);
3914 	obj_cgroup_get(objcg);
3915 	memcg->orig_objcg = objcg;
3916 
3917 	static_branch_enable(&memcg_kmem_online_key);
3918 
3919 	memcg->kmemcg_id = memcg->id.id;
3920 
3921 	return 0;
3922 }
3923 
3924 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3925 {
3926 	struct mem_cgroup *parent;
3927 
3928 	if (mem_cgroup_kmem_disabled())
3929 		return;
3930 
3931 	if (unlikely(mem_cgroup_is_root(memcg)))
3932 		return;
3933 
3934 	parent = parent_mem_cgroup(memcg);
3935 	if (!parent)
3936 		parent = root_mem_cgroup;
3937 
3938 	memcg_reparent_objcgs(memcg, parent);
3939 
3940 	/*
3941 	 * After we have finished memcg_reparent_objcgs(), all list_lrus
3942 	 * corresponding to this cgroup are guaranteed to remain empty.
3943 	 * The ordering is imposed by list_lru_node->lock taken by
3944 	 * memcg_reparent_list_lrus().
3945 	 */
3946 	memcg_reparent_list_lrus(memcg, parent);
3947 }
3948 #else
3949 static int memcg_online_kmem(struct mem_cgroup *memcg)
3950 {
3951 	return 0;
3952 }
3953 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3954 {
3955 }
3956 #endif /* CONFIG_MEMCG_KMEM */
3957 
3958 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3959 {
3960 	int ret;
3961 
3962 	mutex_lock(&memcg_max_mutex);
3963 
3964 	ret = page_counter_set_max(&memcg->tcpmem, max);
3965 	if (ret)
3966 		goto out;
3967 
3968 	if (!memcg->tcpmem_active) {
3969 		/*
3970 		 * The active flag needs to be written after the static_key
3971 		 * update. This is what guarantees that the socket activation
3972 		 * function is the last one to run. See mem_cgroup_sk_alloc()
3973 		 * for details, and note that we don't mark any socket as
3974 		 * belonging to this memcg until that flag is up.
3975 		 *
3976 		 * We need to do this, because static_keys will span multiple
3977 		 * sites, but we can't control their order. If we mark a socket
3978 		 * as accounted, but the accounting functions are not patched in
3979 		 * yet, we'll lose accounting.
3980 		 *
3981 		 * We never race with the readers in mem_cgroup_sk_alloc(),
3982 		 * because when this value change, the code to process it is not
3983 		 * patched in yet.
3984 		 */
3985 		static_branch_inc(&memcg_sockets_enabled_key);
3986 		memcg->tcpmem_active = true;
3987 	}
3988 out:
3989 	mutex_unlock(&memcg_max_mutex);
3990 	return ret;
3991 }
3992 
3993 /*
3994  * The user of this function is...
3995  * RES_LIMIT.
3996  */
3997 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3998 				char *buf, size_t nbytes, loff_t off)
3999 {
4000 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4001 	unsigned long nr_pages;
4002 	int ret;
4003 
4004 	buf = strstrip(buf);
4005 	ret = page_counter_memparse(buf, "-1", &nr_pages);
4006 	if (ret)
4007 		return ret;
4008 
4009 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
4010 	case RES_LIMIT:
4011 		if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4012 			ret = -EINVAL;
4013 			break;
4014 		}
4015 		switch (MEMFILE_TYPE(of_cft(of)->private)) {
4016 		case _MEM:
4017 			ret = mem_cgroup_resize_max(memcg, nr_pages, false);
4018 			break;
4019 		case _MEMSWAP:
4020 			ret = mem_cgroup_resize_max(memcg, nr_pages, true);
4021 			break;
4022 		case _KMEM:
4023 			pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
4024 				     "Writing any value to this file has no effect. "
4025 				     "Please report your usecase to linux-mm@kvack.org if you "
4026 				     "depend on this functionality.\n");
4027 			ret = 0;
4028 			break;
4029 		case _TCP:
4030 			ret = memcg_update_tcp_max(memcg, nr_pages);
4031 			break;
4032 		}
4033 		break;
4034 	case RES_SOFT_LIMIT:
4035 		if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4036 			ret = -EOPNOTSUPP;
4037 		} else {
4038 			WRITE_ONCE(memcg->soft_limit, nr_pages);
4039 			ret = 0;
4040 		}
4041 		break;
4042 	}
4043 	return ret ?: nbytes;
4044 }
4045 
4046 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4047 				size_t nbytes, loff_t off)
4048 {
4049 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4050 	struct page_counter *counter;
4051 
4052 	switch (MEMFILE_TYPE(of_cft(of)->private)) {
4053 	case _MEM:
4054 		counter = &memcg->memory;
4055 		break;
4056 	case _MEMSWAP:
4057 		counter = &memcg->memsw;
4058 		break;
4059 	case _KMEM:
4060 		counter = &memcg->kmem;
4061 		break;
4062 	case _TCP:
4063 		counter = &memcg->tcpmem;
4064 		break;
4065 	default:
4066 		BUG();
4067 	}
4068 
4069 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
4070 	case RES_MAX_USAGE:
4071 		page_counter_reset_watermark(counter);
4072 		break;
4073 	case RES_FAILCNT:
4074 		counter->failcnt = 0;
4075 		break;
4076 	default:
4077 		BUG();
4078 	}
4079 
4080 	return nbytes;
4081 }
4082 
4083 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4084 					struct cftype *cft)
4085 {
4086 	return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4087 }
4088 
4089 #ifdef CONFIG_MMU
4090 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4091 					struct cftype *cft, u64 val)
4092 {
4093 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4094 
4095 	pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
4096 		     "Please report your usecase to linux-mm@kvack.org if you "
4097 		     "depend on this functionality.\n");
4098 
4099 	if (val & ~MOVE_MASK)
4100 		return -EINVAL;
4101 
4102 	/*
4103 	 * No kind of locking is needed in here, because ->can_attach() will
4104 	 * check this value once in the beginning of the process, and then carry
4105 	 * on with stale data. This means that changes to this value will only
4106 	 * affect task migrations starting after the change.
4107 	 */
4108 	memcg->move_charge_at_immigrate = val;
4109 	return 0;
4110 }
4111 #else
4112 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4113 					struct cftype *cft, u64 val)
4114 {
4115 	return -ENOSYS;
4116 }
4117 #endif
4118 
4119 #ifdef CONFIG_NUMA
4120 
4121 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
4122 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
4123 #define LRU_ALL	     ((1 << NR_LRU_LISTS) - 1)
4124 
4125 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
4126 				int nid, unsigned int lru_mask, bool tree)
4127 {
4128 	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
4129 	unsigned long nr = 0;
4130 	enum lru_list lru;
4131 
4132 	VM_BUG_ON((unsigned)nid >= nr_node_ids);
4133 
4134 	for_each_lru(lru) {
4135 		if (!(BIT(lru) & lru_mask))
4136 			continue;
4137 		if (tree)
4138 			nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
4139 		else
4140 			nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4141 	}
4142 	return nr;
4143 }
4144 
4145 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4146 					     unsigned int lru_mask,
4147 					     bool tree)
4148 {
4149 	unsigned long nr = 0;
4150 	enum lru_list lru;
4151 
4152 	for_each_lru(lru) {
4153 		if (!(BIT(lru) & lru_mask))
4154 			continue;
4155 		if (tree)
4156 			nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4157 		else
4158 			nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4159 	}
4160 	return nr;
4161 }
4162 
4163 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4164 {
4165 	struct numa_stat {
4166 		const char *name;
4167 		unsigned int lru_mask;
4168 	};
4169 
4170 	static const struct numa_stat stats[] = {
4171 		{ "total", LRU_ALL },
4172 		{ "file", LRU_ALL_FILE },
4173 		{ "anon", LRU_ALL_ANON },
4174 		{ "unevictable", BIT(LRU_UNEVICTABLE) },
4175 	};
4176 	const struct numa_stat *stat;
4177 	int nid;
4178 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4179 
4180 	mem_cgroup_flush_stats();
4181 
4182 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4183 		seq_printf(m, "%s=%lu", stat->name,
4184 			   mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4185 						   false));
4186 		for_each_node_state(nid, N_MEMORY)
4187 			seq_printf(m, " N%d=%lu", nid,
4188 				   mem_cgroup_node_nr_lru_pages(memcg, nid,
4189 							stat->lru_mask, false));
4190 		seq_putc(m, '\n');
4191 	}
4192 
4193 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4194 
4195 		seq_printf(m, "hierarchical_%s=%lu", stat->name,
4196 			   mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4197 						   true));
4198 		for_each_node_state(nid, N_MEMORY)
4199 			seq_printf(m, " N%d=%lu", nid,
4200 				   mem_cgroup_node_nr_lru_pages(memcg, nid,
4201 							stat->lru_mask, true));
4202 		seq_putc(m, '\n');
4203 	}
4204 
4205 	return 0;
4206 }
4207 #endif /* CONFIG_NUMA */
4208 
4209 static const unsigned int memcg1_stats[] = {
4210 	NR_FILE_PAGES,
4211 	NR_ANON_MAPPED,
4212 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4213 	NR_ANON_THPS,
4214 #endif
4215 	NR_SHMEM,
4216 	NR_FILE_MAPPED,
4217 	NR_FILE_DIRTY,
4218 	NR_WRITEBACK,
4219 	WORKINGSET_REFAULT_ANON,
4220 	WORKINGSET_REFAULT_FILE,
4221 #ifdef CONFIG_SWAP
4222 	MEMCG_SWAP,
4223 	NR_SWAPCACHE,
4224 #endif
4225 };
4226 
4227 static const char *const memcg1_stat_names[] = {
4228 	"cache",
4229 	"rss",
4230 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4231 	"rss_huge",
4232 #endif
4233 	"shmem",
4234 	"mapped_file",
4235 	"dirty",
4236 	"writeback",
4237 	"workingset_refault_anon",
4238 	"workingset_refault_file",
4239 #ifdef CONFIG_SWAP
4240 	"swap",
4241 	"swapcached",
4242 #endif
4243 };
4244 
4245 /* Universal VM events cgroup1 shows, original sort order */
4246 static const unsigned int memcg1_events[] = {
4247 	PGPGIN,
4248 	PGPGOUT,
4249 	PGFAULT,
4250 	PGMAJFAULT,
4251 };
4252 
4253 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
4254 {
4255 	unsigned long memory, memsw;
4256 	struct mem_cgroup *mi;
4257 	unsigned int i;
4258 
4259 	BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4260 
4261 	mem_cgroup_flush_stats();
4262 
4263 	for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4264 		unsigned long nr;
4265 
4266 		nr = memcg_page_state_local_output(memcg, memcg1_stats[i]);
4267 		seq_buf_printf(s, "%s %lu\n", memcg1_stat_names[i], nr);
4268 	}
4269 
4270 	for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4271 		seq_buf_printf(s, "%s %lu\n", vm_event_name(memcg1_events[i]),
4272 			       memcg_events_local(memcg, memcg1_events[i]));
4273 
4274 	for (i = 0; i < NR_LRU_LISTS; i++)
4275 		seq_buf_printf(s, "%s %lu\n", lru_list_name(i),
4276 			       memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4277 			       PAGE_SIZE);
4278 
4279 	/* Hierarchical information */
4280 	memory = memsw = PAGE_COUNTER_MAX;
4281 	for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4282 		memory = min(memory, READ_ONCE(mi->memory.max));
4283 		memsw = min(memsw, READ_ONCE(mi->memsw.max));
4284 	}
4285 	seq_buf_printf(s, "hierarchical_memory_limit %llu\n",
4286 		       (u64)memory * PAGE_SIZE);
4287 	seq_buf_printf(s, "hierarchical_memsw_limit %llu\n",
4288 		       (u64)memsw * PAGE_SIZE);
4289 
4290 	for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4291 		unsigned long nr;
4292 
4293 		nr = memcg_page_state_output(memcg, memcg1_stats[i]);
4294 		seq_buf_printf(s, "total_%s %llu\n", memcg1_stat_names[i],
4295 			       (u64)nr);
4296 	}
4297 
4298 	for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4299 		seq_buf_printf(s, "total_%s %llu\n",
4300 			       vm_event_name(memcg1_events[i]),
4301 			       (u64)memcg_events(memcg, memcg1_events[i]));
4302 
4303 	for (i = 0; i < NR_LRU_LISTS; i++)
4304 		seq_buf_printf(s, "total_%s %llu\n", lru_list_name(i),
4305 			       (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4306 			       PAGE_SIZE);
4307 
4308 #ifdef CONFIG_DEBUG_VM
4309 	{
4310 		pg_data_t *pgdat;
4311 		struct mem_cgroup_per_node *mz;
4312 		unsigned long anon_cost = 0;
4313 		unsigned long file_cost = 0;
4314 
4315 		for_each_online_pgdat(pgdat) {
4316 			mz = memcg->nodeinfo[pgdat->node_id];
4317 
4318 			anon_cost += mz->lruvec.anon_cost;
4319 			file_cost += mz->lruvec.file_cost;
4320 		}
4321 		seq_buf_printf(s, "anon_cost %lu\n", anon_cost);
4322 		seq_buf_printf(s, "file_cost %lu\n", file_cost);
4323 	}
4324 #endif
4325 }
4326 
4327 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4328 				      struct cftype *cft)
4329 {
4330 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4331 
4332 	return mem_cgroup_swappiness(memcg);
4333 }
4334 
4335 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4336 				       struct cftype *cft, u64 val)
4337 {
4338 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4339 
4340 	if (val > 200)
4341 		return -EINVAL;
4342 
4343 	if (!mem_cgroup_is_root(memcg))
4344 		WRITE_ONCE(memcg->swappiness, val);
4345 	else
4346 		WRITE_ONCE(vm_swappiness, val);
4347 
4348 	return 0;
4349 }
4350 
4351 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4352 {
4353 	struct mem_cgroup_threshold_ary *t;
4354 	unsigned long usage;
4355 	int i;
4356 
4357 	rcu_read_lock();
4358 	if (!swap)
4359 		t = rcu_dereference(memcg->thresholds.primary);
4360 	else
4361 		t = rcu_dereference(memcg->memsw_thresholds.primary);
4362 
4363 	if (!t)
4364 		goto unlock;
4365 
4366 	usage = mem_cgroup_usage(memcg, swap);
4367 
4368 	/*
4369 	 * current_threshold points to threshold just below or equal to usage.
4370 	 * If it's not true, a threshold was crossed after last
4371 	 * call of __mem_cgroup_threshold().
4372 	 */
4373 	i = t->current_threshold;
4374 
4375 	/*
4376 	 * Iterate backward over array of thresholds starting from
4377 	 * current_threshold and check if a threshold is crossed.
4378 	 * If none of thresholds below usage is crossed, we read
4379 	 * only one element of the array here.
4380 	 */
4381 	for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4382 		eventfd_signal(t->entries[i].eventfd, 1);
4383 
4384 	/* i = current_threshold + 1 */
4385 	i++;
4386 
4387 	/*
4388 	 * Iterate forward over array of thresholds starting from
4389 	 * current_threshold+1 and check if a threshold is crossed.
4390 	 * If none of thresholds above usage is crossed, we read
4391 	 * only one element of the array here.
4392 	 */
4393 	for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4394 		eventfd_signal(t->entries[i].eventfd, 1);
4395 
4396 	/* Update current_threshold */
4397 	t->current_threshold = i - 1;
4398 unlock:
4399 	rcu_read_unlock();
4400 }
4401 
4402 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4403 {
4404 	while (memcg) {
4405 		__mem_cgroup_threshold(memcg, false);
4406 		if (do_memsw_account())
4407 			__mem_cgroup_threshold(memcg, true);
4408 
4409 		memcg = parent_mem_cgroup(memcg);
4410 	}
4411 }
4412 
4413 static int compare_thresholds(const void *a, const void *b)
4414 {
4415 	const struct mem_cgroup_threshold *_a = a;
4416 	const struct mem_cgroup_threshold *_b = b;
4417 
4418 	if (_a->threshold > _b->threshold)
4419 		return 1;
4420 
4421 	if (_a->threshold < _b->threshold)
4422 		return -1;
4423 
4424 	return 0;
4425 }
4426 
4427 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4428 {
4429 	struct mem_cgroup_eventfd_list *ev;
4430 
4431 	spin_lock(&memcg_oom_lock);
4432 
4433 	list_for_each_entry(ev, &memcg->oom_notify, list)
4434 		eventfd_signal(ev->eventfd, 1);
4435 
4436 	spin_unlock(&memcg_oom_lock);
4437 	return 0;
4438 }
4439 
4440 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4441 {
4442 	struct mem_cgroup *iter;
4443 
4444 	for_each_mem_cgroup_tree(iter, memcg)
4445 		mem_cgroup_oom_notify_cb(iter);
4446 }
4447 
4448 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4449 	struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4450 {
4451 	struct mem_cgroup_thresholds *thresholds;
4452 	struct mem_cgroup_threshold_ary *new;
4453 	unsigned long threshold;
4454 	unsigned long usage;
4455 	int i, size, ret;
4456 
4457 	ret = page_counter_memparse(args, "-1", &threshold);
4458 	if (ret)
4459 		return ret;
4460 
4461 	mutex_lock(&memcg->thresholds_lock);
4462 
4463 	if (type == _MEM) {
4464 		thresholds = &memcg->thresholds;
4465 		usage = mem_cgroup_usage(memcg, false);
4466 	} else if (type == _MEMSWAP) {
4467 		thresholds = &memcg->memsw_thresholds;
4468 		usage = mem_cgroup_usage(memcg, true);
4469 	} else
4470 		BUG();
4471 
4472 	/* Check if a threshold crossed before adding a new one */
4473 	if (thresholds->primary)
4474 		__mem_cgroup_threshold(memcg, type == _MEMSWAP);
4475 
4476 	size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4477 
4478 	/* Allocate memory for new array of thresholds */
4479 	new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4480 	if (!new) {
4481 		ret = -ENOMEM;
4482 		goto unlock;
4483 	}
4484 	new->size = size;
4485 
4486 	/* Copy thresholds (if any) to new array */
4487 	if (thresholds->primary)
4488 		memcpy(new->entries, thresholds->primary->entries,
4489 		       flex_array_size(new, entries, size - 1));
4490 
4491 	/* Add new threshold */
4492 	new->entries[size - 1].eventfd = eventfd;
4493 	new->entries[size - 1].threshold = threshold;
4494 
4495 	/* Sort thresholds. Registering of new threshold isn't time-critical */
4496 	sort(new->entries, size, sizeof(*new->entries),
4497 			compare_thresholds, NULL);
4498 
4499 	/* Find current threshold */
4500 	new->current_threshold = -1;
4501 	for (i = 0; i < size; i++) {
4502 		if (new->entries[i].threshold <= usage) {
4503 			/*
4504 			 * new->current_threshold will not be used until
4505 			 * rcu_assign_pointer(), so it's safe to increment
4506 			 * it here.
4507 			 */
4508 			++new->current_threshold;
4509 		} else
4510 			break;
4511 	}
4512 
4513 	/* Free old spare buffer and save old primary buffer as spare */
4514 	kfree(thresholds->spare);
4515 	thresholds->spare = thresholds->primary;
4516 
4517 	rcu_assign_pointer(thresholds->primary, new);
4518 
4519 	/* To be sure that nobody uses thresholds */
4520 	synchronize_rcu();
4521 
4522 unlock:
4523 	mutex_unlock(&memcg->thresholds_lock);
4524 
4525 	return ret;
4526 }
4527 
4528 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4529 	struct eventfd_ctx *eventfd, const char *args)
4530 {
4531 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4532 }
4533 
4534 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4535 	struct eventfd_ctx *eventfd, const char *args)
4536 {
4537 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4538 }
4539 
4540 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4541 	struct eventfd_ctx *eventfd, enum res_type type)
4542 {
4543 	struct mem_cgroup_thresholds *thresholds;
4544 	struct mem_cgroup_threshold_ary *new;
4545 	unsigned long usage;
4546 	int i, j, size, entries;
4547 
4548 	mutex_lock(&memcg->thresholds_lock);
4549 
4550 	if (type == _MEM) {
4551 		thresholds = &memcg->thresholds;
4552 		usage = mem_cgroup_usage(memcg, false);
4553 	} else if (type == _MEMSWAP) {
4554 		thresholds = &memcg->memsw_thresholds;
4555 		usage = mem_cgroup_usage(memcg, true);
4556 	} else
4557 		BUG();
4558 
4559 	if (!thresholds->primary)
4560 		goto unlock;
4561 
4562 	/* Check if a threshold crossed before removing */
4563 	__mem_cgroup_threshold(memcg, type == _MEMSWAP);
4564 
4565 	/* Calculate new number of threshold */
4566 	size = entries = 0;
4567 	for (i = 0; i < thresholds->primary->size; i++) {
4568 		if (thresholds->primary->entries[i].eventfd != eventfd)
4569 			size++;
4570 		else
4571 			entries++;
4572 	}
4573 
4574 	new = thresholds->spare;
4575 
4576 	/* If no items related to eventfd have been cleared, nothing to do */
4577 	if (!entries)
4578 		goto unlock;
4579 
4580 	/* Set thresholds array to NULL if we don't have thresholds */
4581 	if (!size) {
4582 		kfree(new);
4583 		new = NULL;
4584 		goto swap_buffers;
4585 	}
4586 
4587 	new->size = size;
4588 
4589 	/* Copy thresholds and find current threshold */
4590 	new->current_threshold = -1;
4591 	for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4592 		if (thresholds->primary->entries[i].eventfd == eventfd)
4593 			continue;
4594 
4595 		new->entries[j] = thresholds->primary->entries[i];
4596 		if (new->entries[j].threshold <= usage) {
4597 			/*
4598 			 * new->current_threshold will not be used
4599 			 * until rcu_assign_pointer(), so it's safe to increment
4600 			 * it here.
4601 			 */
4602 			++new->current_threshold;
4603 		}
4604 		j++;
4605 	}
4606 
4607 swap_buffers:
4608 	/* Swap primary and spare array */
4609 	thresholds->spare = thresholds->primary;
4610 
4611 	rcu_assign_pointer(thresholds->primary, new);
4612 
4613 	/* To be sure that nobody uses thresholds */
4614 	synchronize_rcu();
4615 
4616 	/* If all events are unregistered, free the spare array */
4617 	if (!new) {
4618 		kfree(thresholds->spare);
4619 		thresholds->spare = NULL;
4620 	}
4621 unlock:
4622 	mutex_unlock(&memcg->thresholds_lock);
4623 }
4624 
4625 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4626 	struct eventfd_ctx *eventfd)
4627 {
4628 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4629 }
4630 
4631 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4632 	struct eventfd_ctx *eventfd)
4633 {
4634 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4635 }
4636 
4637 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4638 	struct eventfd_ctx *eventfd, const char *args)
4639 {
4640 	struct mem_cgroup_eventfd_list *event;
4641 
4642 	event = kmalloc(sizeof(*event),	GFP_KERNEL);
4643 	if (!event)
4644 		return -ENOMEM;
4645 
4646 	spin_lock(&memcg_oom_lock);
4647 
4648 	event->eventfd = eventfd;
4649 	list_add(&event->list, &memcg->oom_notify);
4650 
4651 	/* already in OOM ? */
4652 	if (memcg->under_oom)
4653 		eventfd_signal(eventfd, 1);
4654 	spin_unlock(&memcg_oom_lock);
4655 
4656 	return 0;
4657 }
4658 
4659 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4660 	struct eventfd_ctx *eventfd)
4661 {
4662 	struct mem_cgroup_eventfd_list *ev, *tmp;
4663 
4664 	spin_lock(&memcg_oom_lock);
4665 
4666 	list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4667 		if (ev->eventfd == eventfd) {
4668 			list_del(&ev->list);
4669 			kfree(ev);
4670 		}
4671 	}
4672 
4673 	spin_unlock(&memcg_oom_lock);
4674 }
4675 
4676 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4677 {
4678 	struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4679 
4680 	seq_printf(sf, "oom_kill_disable %d\n", READ_ONCE(memcg->oom_kill_disable));
4681 	seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4682 	seq_printf(sf, "oom_kill %lu\n",
4683 		   atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4684 	return 0;
4685 }
4686 
4687 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4688 	struct cftype *cft, u64 val)
4689 {
4690 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4691 
4692 	/* cannot set to root cgroup and only 0 and 1 are allowed */
4693 	if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4694 		return -EINVAL;
4695 
4696 	WRITE_ONCE(memcg->oom_kill_disable, val);
4697 	if (!val)
4698 		memcg_oom_recover(memcg);
4699 
4700 	return 0;
4701 }
4702 
4703 #ifdef CONFIG_CGROUP_WRITEBACK
4704 
4705 #include <trace/events/writeback.h>
4706 
4707 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4708 {
4709 	return wb_domain_init(&memcg->cgwb_domain, gfp);
4710 }
4711 
4712 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4713 {
4714 	wb_domain_exit(&memcg->cgwb_domain);
4715 }
4716 
4717 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4718 {
4719 	wb_domain_size_changed(&memcg->cgwb_domain);
4720 }
4721 
4722 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4723 {
4724 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4725 
4726 	if (!memcg->css.parent)
4727 		return NULL;
4728 
4729 	return &memcg->cgwb_domain;
4730 }
4731 
4732 /**
4733  * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4734  * @wb: bdi_writeback in question
4735  * @pfilepages: out parameter for number of file pages
4736  * @pheadroom: out parameter for number of allocatable pages according to memcg
4737  * @pdirty: out parameter for number of dirty pages
4738  * @pwriteback: out parameter for number of pages under writeback
4739  *
4740  * Determine the numbers of file, headroom, dirty, and writeback pages in
4741  * @wb's memcg.  File, dirty and writeback are self-explanatory.  Headroom
4742  * is a bit more involved.
4743  *
4744  * A memcg's headroom is "min(max, high) - used".  In the hierarchy, the
4745  * headroom is calculated as the lowest headroom of itself and the
4746  * ancestors.  Note that this doesn't consider the actual amount of
4747  * available memory in the system.  The caller should further cap
4748  * *@pheadroom accordingly.
4749  */
4750 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4751 			 unsigned long *pheadroom, unsigned long *pdirty,
4752 			 unsigned long *pwriteback)
4753 {
4754 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4755 	struct mem_cgroup *parent;
4756 
4757 	mem_cgroup_flush_stats();
4758 
4759 	*pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4760 	*pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4761 	*pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4762 			memcg_page_state(memcg, NR_ACTIVE_FILE);
4763 
4764 	*pheadroom = PAGE_COUNTER_MAX;
4765 	while ((parent = parent_mem_cgroup(memcg))) {
4766 		unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4767 					    READ_ONCE(memcg->memory.high));
4768 		unsigned long used = page_counter_read(&memcg->memory);
4769 
4770 		*pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4771 		memcg = parent;
4772 	}
4773 }
4774 
4775 /*
4776  * Foreign dirty flushing
4777  *
4778  * There's an inherent mismatch between memcg and writeback.  The former
4779  * tracks ownership per-page while the latter per-inode.  This was a
4780  * deliberate design decision because honoring per-page ownership in the
4781  * writeback path is complicated, may lead to higher CPU and IO overheads
4782  * and deemed unnecessary given that write-sharing an inode across
4783  * different cgroups isn't a common use-case.
4784  *
4785  * Combined with inode majority-writer ownership switching, this works well
4786  * enough in most cases but there are some pathological cases.  For
4787  * example, let's say there are two cgroups A and B which keep writing to
4788  * different but confined parts of the same inode.  B owns the inode and
4789  * A's memory is limited far below B's.  A's dirty ratio can rise enough to
4790  * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4791  * triggering background writeback.  A will be slowed down without a way to
4792  * make writeback of the dirty pages happen.
4793  *
4794  * Conditions like the above can lead to a cgroup getting repeatedly and
4795  * severely throttled after making some progress after each
4796  * dirty_expire_interval while the underlying IO device is almost
4797  * completely idle.
4798  *
4799  * Solving this problem completely requires matching the ownership tracking
4800  * granularities between memcg and writeback in either direction.  However,
4801  * the more egregious behaviors can be avoided by simply remembering the
4802  * most recent foreign dirtying events and initiating remote flushes on
4803  * them when local writeback isn't enough to keep the memory clean enough.
4804  *
4805  * The following two functions implement such mechanism.  When a foreign
4806  * page - a page whose memcg and writeback ownerships don't match - is
4807  * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4808  * bdi_writeback on the page owning memcg.  When balance_dirty_pages()
4809  * decides that the memcg needs to sleep due to high dirty ratio, it calls
4810  * mem_cgroup_flush_foreign() which queues writeback on the recorded
4811  * foreign bdi_writebacks which haven't expired.  Both the numbers of
4812  * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4813  * limited to MEMCG_CGWB_FRN_CNT.
4814  *
4815  * The mechanism only remembers IDs and doesn't hold any object references.
4816  * As being wrong occasionally doesn't matter, updates and accesses to the
4817  * records are lockless and racy.
4818  */
4819 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4820 					     struct bdi_writeback *wb)
4821 {
4822 	struct mem_cgroup *memcg = folio_memcg(folio);
4823 	struct memcg_cgwb_frn *frn;
4824 	u64 now = get_jiffies_64();
4825 	u64 oldest_at = now;
4826 	int oldest = -1;
4827 	int i;
4828 
4829 	trace_track_foreign_dirty(folio, wb);
4830 
4831 	/*
4832 	 * Pick the slot to use.  If there is already a slot for @wb, keep
4833 	 * using it.  If not replace the oldest one which isn't being
4834 	 * written out.
4835 	 */
4836 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4837 		frn = &memcg->cgwb_frn[i];
4838 		if (frn->bdi_id == wb->bdi->id &&
4839 		    frn->memcg_id == wb->memcg_css->id)
4840 			break;
4841 		if (time_before64(frn->at, oldest_at) &&
4842 		    atomic_read(&frn->done.cnt) == 1) {
4843 			oldest = i;
4844 			oldest_at = frn->at;
4845 		}
4846 	}
4847 
4848 	if (i < MEMCG_CGWB_FRN_CNT) {
4849 		/*
4850 		 * Re-using an existing one.  Update timestamp lazily to
4851 		 * avoid making the cacheline hot.  We want them to be
4852 		 * reasonably up-to-date and significantly shorter than
4853 		 * dirty_expire_interval as that's what expires the record.
4854 		 * Use the shorter of 1s and dirty_expire_interval / 8.
4855 		 */
4856 		unsigned long update_intv =
4857 			min_t(unsigned long, HZ,
4858 			      msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4859 
4860 		if (time_before64(frn->at, now - update_intv))
4861 			frn->at = now;
4862 	} else if (oldest >= 0) {
4863 		/* replace the oldest free one */
4864 		frn = &memcg->cgwb_frn[oldest];
4865 		frn->bdi_id = wb->bdi->id;
4866 		frn->memcg_id = wb->memcg_css->id;
4867 		frn->at = now;
4868 	}
4869 }
4870 
4871 /* issue foreign writeback flushes for recorded foreign dirtying events */
4872 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4873 {
4874 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4875 	unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4876 	u64 now = jiffies_64;
4877 	int i;
4878 
4879 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4880 		struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4881 
4882 		/*
4883 		 * If the record is older than dirty_expire_interval,
4884 		 * writeback on it has already started.  No need to kick it
4885 		 * off again.  Also, don't start a new one if there's
4886 		 * already one in flight.
4887 		 */
4888 		if (time_after64(frn->at, now - intv) &&
4889 		    atomic_read(&frn->done.cnt) == 1) {
4890 			frn->at = 0;
4891 			trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4892 			cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4893 					       WB_REASON_FOREIGN_FLUSH,
4894 					       &frn->done);
4895 		}
4896 	}
4897 }
4898 
4899 #else	/* CONFIG_CGROUP_WRITEBACK */
4900 
4901 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4902 {
4903 	return 0;
4904 }
4905 
4906 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4907 {
4908 }
4909 
4910 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4911 {
4912 }
4913 
4914 #endif	/* CONFIG_CGROUP_WRITEBACK */
4915 
4916 /*
4917  * DO NOT USE IN NEW FILES.
4918  *
4919  * "cgroup.event_control" implementation.
4920  *
4921  * This is way over-engineered.  It tries to support fully configurable
4922  * events for each user.  Such level of flexibility is completely
4923  * unnecessary especially in the light of the planned unified hierarchy.
4924  *
4925  * Please deprecate this and replace with something simpler if at all
4926  * possible.
4927  */
4928 
4929 /*
4930  * Unregister event and free resources.
4931  *
4932  * Gets called from workqueue.
4933  */
4934 static void memcg_event_remove(struct work_struct *work)
4935 {
4936 	struct mem_cgroup_event *event =
4937 		container_of(work, struct mem_cgroup_event, remove);
4938 	struct mem_cgroup *memcg = event->memcg;
4939 
4940 	remove_wait_queue(event->wqh, &event->wait);
4941 
4942 	event->unregister_event(memcg, event->eventfd);
4943 
4944 	/* Notify userspace the event is going away. */
4945 	eventfd_signal(event->eventfd, 1);
4946 
4947 	eventfd_ctx_put(event->eventfd);
4948 	kfree(event);
4949 	css_put(&memcg->css);
4950 }
4951 
4952 /*
4953  * Gets called on EPOLLHUP on eventfd when user closes it.
4954  *
4955  * Called with wqh->lock held and interrupts disabled.
4956  */
4957 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4958 			    int sync, void *key)
4959 {
4960 	struct mem_cgroup_event *event =
4961 		container_of(wait, struct mem_cgroup_event, wait);
4962 	struct mem_cgroup *memcg = event->memcg;
4963 	__poll_t flags = key_to_poll(key);
4964 
4965 	if (flags & EPOLLHUP) {
4966 		/*
4967 		 * If the event has been detached at cgroup removal, we
4968 		 * can simply return knowing the other side will cleanup
4969 		 * for us.
4970 		 *
4971 		 * We can't race against event freeing since the other
4972 		 * side will require wqh->lock via remove_wait_queue(),
4973 		 * which we hold.
4974 		 */
4975 		spin_lock(&memcg->event_list_lock);
4976 		if (!list_empty(&event->list)) {
4977 			list_del_init(&event->list);
4978 			/*
4979 			 * We are in atomic context, but cgroup_event_remove()
4980 			 * may sleep, so we have to call it in workqueue.
4981 			 */
4982 			schedule_work(&event->remove);
4983 		}
4984 		spin_unlock(&memcg->event_list_lock);
4985 	}
4986 
4987 	return 0;
4988 }
4989 
4990 static void memcg_event_ptable_queue_proc(struct file *file,
4991 		wait_queue_head_t *wqh, poll_table *pt)
4992 {
4993 	struct mem_cgroup_event *event =
4994 		container_of(pt, struct mem_cgroup_event, pt);
4995 
4996 	event->wqh = wqh;
4997 	add_wait_queue(wqh, &event->wait);
4998 }
4999 
5000 /*
5001  * DO NOT USE IN NEW FILES.
5002  *
5003  * Parse input and register new cgroup event handler.
5004  *
5005  * Input must be in format '<event_fd> <control_fd> <args>'.
5006  * Interpretation of args is defined by control file implementation.
5007  */
5008 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5009 					 char *buf, size_t nbytes, loff_t off)
5010 {
5011 	struct cgroup_subsys_state *css = of_css(of);
5012 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5013 	struct mem_cgroup_event *event;
5014 	struct cgroup_subsys_state *cfile_css;
5015 	unsigned int efd, cfd;
5016 	struct fd efile;
5017 	struct fd cfile;
5018 	struct dentry *cdentry;
5019 	const char *name;
5020 	char *endp;
5021 	int ret;
5022 
5023 	if (IS_ENABLED(CONFIG_PREEMPT_RT))
5024 		return -EOPNOTSUPP;
5025 
5026 	buf = strstrip(buf);
5027 
5028 	efd = simple_strtoul(buf, &endp, 10);
5029 	if (*endp != ' ')
5030 		return -EINVAL;
5031 	buf = endp + 1;
5032 
5033 	cfd = simple_strtoul(buf, &endp, 10);
5034 	if ((*endp != ' ') && (*endp != '\0'))
5035 		return -EINVAL;
5036 	buf = endp + 1;
5037 
5038 	event = kzalloc(sizeof(*event), GFP_KERNEL);
5039 	if (!event)
5040 		return -ENOMEM;
5041 
5042 	event->memcg = memcg;
5043 	INIT_LIST_HEAD(&event->list);
5044 	init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5045 	init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5046 	INIT_WORK(&event->remove, memcg_event_remove);
5047 
5048 	efile = fdget(efd);
5049 	if (!efile.file) {
5050 		ret = -EBADF;
5051 		goto out_kfree;
5052 	}
5053 
5054 	event->eventfd = eventfd_ctx_fileget(efile.file);
5055 	if (IS_ERR(event->eventfd)) {
5056 		ret = PTR_ERR(event->eventfd);
5057 		goto out_put_efile;
5058 	}
5059 
5060 	cfile = fdget(cfd);
5061 	if (!cfile.file) {
5062 		ret = -EBADF;
5063 		goto out_put_eventfd;
5064 	}
5065 
5066 	/* the process need read permission on control file */
5067 	/* AV: shouldn't we check that it's been opened for read instead? */
5068 	ret = file_permission(cfile.file, MAY_READ);
5069 	if (ret < 0)
5070 		goto out_put_cfile;
5071 
5072 	/*
5073 	 * The control file must be a regular cgroup1 file. As a regular cgroup
5074 	 * file can't be renamed, it's safe to access its name afterwards.
5075 	 */
5076 	cdentry = cfile.file->f_path.dentry;
5077 	if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
5078 		ret = -EINVAL;
5079 		goto out_put_cfile;
5080 	}
5081 
5082 	/*
5083 	 * Determine the event callbacks and set them in @event.  This used
5084 	 * to be done via struct cftype but cgroup core no longer knows
5085 	 * about these events.  The following is crude but the whole thing
5086 	 * is for compatibility anyway.
5087 	 *
5088 	 * DO NOT ADD NEW FILES.
5089 	 */
5090 	name = cdentry->d_name.name;
5091 
5092 	if (!strcmp(name, "memory.usage_in_bytes")) {
5093 		event->register_event = mem_cgroup_usage_register_event;
5094 		event->unregister_event = mem_cgroup_usage_unregister_event;
5095 	} else if (!strcmp(name, "memory.oom_control")) {
5096 		event->register_event = mem_cgroup_oom_register_event;
5097 		event->unregister_event = mem_cgroup_oom_unregister_event;
5098 	} else if (!strcmp(name, "memory.pressure_level")) {
5099 		event->register_event = vmpressure_register_event;
5100 		event->unregister_event = vmpressure_unregister_event;
5101 	} else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5102 		event->register_event = memsw_cgroup_usage_register_event;
5103 		event->unregister_event = memsw_cgroup_usage_unregister_event;
5104 	} else {
5105 		ret = -EINVAL;
5106 		goto out_put_cfile;
5107 	}
5108 
5109 	/*
5110 	 * Verify @cfile should belong to @css.  Also, remaining events are
5111 	 * automatically removed on cgroup destruction but the removal is
5112 	 * asynchronous, so take an extra ref on @css.
5113 	 */
5114 	cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
5115 					       &memory_cgrp_subsys);
5116 	ret = -EINVAL;
5117 	if (IS_ERR(cfile_css))
5118 		goto out_put_cfile;
5119 	if (cfile_css != css) {
5120 		css_put(cfile_css);
5121 		goto out_put_cfile;
5122 	}
5123 
5124 	ret = event->register_event(memcg, event->eventfd, buf);
5125 	if (ret)
5126 		goto out_put_css;
5127 
5128 	vfs_poll(efile.file, &event->pt);
5129 
5130 	spin_lock_irq(&memcg->event_list_lock);
5131 	list_add(&event->list, &memcg->event_list);
5132 	spin_unlock_irq(&memcg->event_list_lock);
5133 
5134 	fdput(cfile);
5135 	fdput(efile);
5136 
5137 	return nbytes;
5138 
5139 out_put_css:
5140 	css_put(css);
5141 out_put_cfile:
5142 	fdput(cfile);
5143 out_put_eventfd:
5144 	eventfd_ctx_put(event->eventfd);
5145 out_put_efile:
5146 	fdput(efile);
5147 out_kfree:
5148 	kfree(event);
5149 
5150 	return ret;
5151 }
5152 
5153 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5154 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
5155 {
5156 	/*
5157 	 * Deprecated.
5158 	 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
5159 	 */
5160 	return 0;
5161 }
5162 #endif
5163 
5164 static int memory_stat_show(struct seq_file *m, void *v);
5165 
5166 static struct cftype mem_cgroup_legacy_files[] = {
5167 	{
5168 		.name = "usage_in_bytes",
5169 		.private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5170 		.read_u64 = mem_cgroup_read_u64,
5171 	},
5172 	{
5173 		.name = "max_usage_in_bytes",
5174 		.private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5175 		.write = mem_cgroup_reset,
5176 		.read_u64 = mem_cgroup_read_u64,
5177 	},
5178 	{
5179 		.name = "limit_in_bytes",
5180 		.private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5181 		.write = mem_cgroup_write,
5182 		.read_u64 = mem_cgroup_read_u64,
5183 	},
5184 	{
5185 		.name = "soft_limit_in_bytes",
5186 		.private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5187 		.write = mem_cgroup_write,
5188 		.read_u64 = mem_cgroup_read_u64,
5189 	},
5190 	{
5191 		.name = "failcnt",
5192 		.private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5193 		.write = mem_cgroup_reset,
5194 		.read_u64 = mem_cgroup_read_u64,
5195 	},
5196 	{
5197 		.name = "stat",
5198 		.seq_show = memory_stat_show,
5199 	},
5200 	{
5201 		.name = "force_empty",
5202 		.write = mem_cgroup_force_empty_write,
5203 	},
5204 	{
5205 		.name = "use_hierarchy",
5206 		.write_u64 = mem_cgroup_hierarchy_write,
5207 		.read_u64 = mem_cgroup_hierarchy_read,
5208 	},
5209 	{
5210 		.name = "cgroup.event_control",		/* XXX: for compat */
5211 		.write = memcg_write_event_control,
5212 		.flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5213 	},
5214 	{
5215 		.name = "swappiness",
5216 		.read_u64 = mem_cgroup_swappiness_read,
5217 		.write_u64 = mem_cgroup_swappiness_write,
5218 	},
5219 	{
5220 		.name = "move_charge_at_immigrate",
5221 		.read_u64 = mem_cgroup_move_charge_read,
5222 		.write_u64 = mem_cgroup_move_charge_write,
5223 	},
5224 	{
5225 		.name = "oom_control",
5226 		.seq_show = mem_cgroup_oom_control_read,
5227 		.write_u64 = mem_cgroup_oom_control_write,
5228 	},
5229 	{
5230 		.name = "pressure_level",
5231 		.seq_show = mem_cgroup_dummy_seq_show,
5232 	},
5233 #ifdef CONFIG_NUMA
5234 	{
5235 		.name = "numa_stat",
5236 		.seq_show = memcg_numa_stat_show,
5237 	},
5238 #endif
5239 	{
5240 		.name = "kmem.limit_in_bytes",
5241 		.private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5242 		.write = mem_cgroup_write,
5243 		.read_u64 = mem_cgroup_read_u64,
5244 	},
5245 	{
5246 		.name = "kmem.usage_in_bytes",
5247 		.private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5248 		.read_u64 = mem_cgroup_read_u64,
5249 	},
5250 	{
5251 		.name = "kmem.failcnt",
5252 		.private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5253 		.write = mem_cgroup_reset,
5254 		.read_u64 = mem_cgroup_read_u64,
5255 	},
5256 	{
5257 		.name = "kmem.max_usage_in_bytes",
5258 		.private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5259 		.write = mem_cgroup_reset,
5260 		.read_u64 = mem_cgroup_read_u64,
5261 	},
5262 #if defined(CONFIG_MEMCG_KMEM) && \
5263 	(defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5264 	{
5265 		.name = "kmem.slabinfo",
5266 		.seq_show = mem_cgroup_slab_show,
5267 	},
5268 #endif
5269 	{
5270 		.name = "kmem.tcp.limit_in_bytes",
5271 		.private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5272 		.write = mem_cgroup_write,
5273 		.read_u64 = mem_cgroup_read_u64,
5274 	},
5275 	{
5276 		.name = "kmem.tcp.usage_in_bytes",
5277 		.private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5278 		.read_u64 = mem_cgroup_read_u64,
5279 	},
5280 	{
5281 		.name = "kmem.tcp.failcnt",
5282 		.private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5283 		.write = mem_cgroup_reset,
5284 		.read_u64 = mem_cgroup_read_u64,
5285 	},
5286 	{
5287 		.name = "kmem.tcp.max_usage_in_bytes",
5288 		.private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5289 		.write = mem_cgroup_reset,
5290 		.read_u64 = mem_cgroup_read_u64,
5291 	},
5292 	{ },	/* terminate */
5293 };
5294 
5295 /*
5296  * Private memory cgroup IDR
5297  *
5298  * Swap-out records and page cache shadow entries need to store memcg
5299  * references in constrained space, so we maintain an ID space that is
5300  * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5301  * memory-controlled cgroups to 64k.
5302  *
5303  * However, there usually are many references to the offline CSS after
5304  * the cgroup has been destroyed, such as page cache or reclaimable
5305  * slab objects, that don't need to hang on to the ID. We want to keep
5306  * those dead CSS from occupying IDs, or we might quickly exhaust the
5307  * relatively small ID space and prevent the creation of new cgroups
5308  * even when there are much fewer than 64k cgroups - possibly none.
5309  *
5310  * Maintain a private 16-bit ID space for memcg, and allow the ID to
5311  * be freed and recycled when it's no longer needed, which is usually
5312  * when the CSS is offlined.
5313  *
5314  * The only exception to that are records of swapped out tmpfs/shmem
5315  * pages that need to be attributed to live ancestors on swapin. But
5316  * those references are manageable from userspace.
5317  */
5318 
5319 #define MEM_CGROUP_ID_MAX	((1UL << MEM_CGROUP_ID_SHIFT) - 1)
5320 static DEFINE_IDR(mem_cgroup_idr);
5321 
5322 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5323 {
5324 	if (memcg->id.id > 0) {
5325 		idr_remove(&mem_cgroup_idr, memcg->id.id);
5326 		memcg->id.id = 0;
5327 	}
5328 }
5329 
5330 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5331 						  unsigned int n)
5332 {
5333 	refcount_add(n, &memcg->id.ref);
5334 }
5335 
5336 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5337 {
5338 	if (refcount_sub_and_test(n, &memcg->id.ref)) {
5339 		mem_cgroup_id_remove(memcg);
5340 
5341 		/* Memcg ID pins CSS */
5342 		css_put(&memcg->css);
5343 	}
5344 }
5345 
5346 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5347 {
5348 	mem_cgroup_id_put_many(memcg, 1);
5349 }
5350 
5351 /**
5352  * mem_cgroup_from_id - look up a memcg from a memcg id
5353  * @id: the memcg id to look up
5354  *
5355  * Caller must hold rcu_read_lock().
5356  */
5357 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5358 {
5359 	WARN_ON_ONCE(!rcu_read_lock_held());
5360 	return idr_find(&mem_cgroup_idr, id);
5361 }
5362 
5363 #ifdef CONFIG_SHRINKER_DEBUG
5364 struct mem_cgroup *mem_cgroup_get_from_ino(unsigned long ino)
5365 {
5366 	struct cgroup *cgrp;
5367 	struct cgroup_subsys_state *css;
5368 	struct mem_cgroup *memcg;
5369 
5370 	cgrp = cgroup_get_from_id(ino);
5371 	if (IS_ERR(cgrp))
5372 		return ERR_CAST(cgrp);
5373 
5374 	css = cgroup_get_e_css(cgrp, &memory_cgrp_subsys);
5375 	if (css)
5376 		memcg = container_of(css, struct mem_cgroup, css);
5377 	else
5378 		memcg = ERR_PTR(-ENOENT);
5379 
5380 	cgroup_put(cgrp);
5381 
5382 	return memcg;
5383 }
5384 #endif
5385 
5386 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5387 {
5388 	struct mem_cgroup_per_node *pn;
5389 
5390 	pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5391 	if (!pn)
5392 		return 1;
5393 
5394 	pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5395 						   GFP_KERNEL_ACCOUNT);
5396 	if (!pn->lruvec_stats_percpu) {
5397 		kfree(pn);
5398 		return 1;
5399 	}
5400 
5401 	lruvec_init(&pn->lruvec);
5402 	pn->memcg = memcg;
5403 
5404 	memcg->nodeinfo[node] = pn;
5405 	return 0;
5406 }
5407 
5408 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5409 {
5410 	struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5411 
5412 	if (!pn)
5413 		return;
5414 
5415 	free_percpu(pn->lruvec_stats_percpu);
5416 	kfree(pn);
5417 }
5418 
5419 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5420 {
5421 	int node;
5422 
5423 	if (memcg->orig_objcg)
5424 		obj_cgroup_put(memcg->orig_objcg);
5425 
5426 	for_each_node(node)
5427 		free_mem_cgroup_per_node_info(memcg, node);
5428 	kfree(memcg->vmstats);
5429 	free_percpu(memcg->vmstats_percpu);
5430 	kfree(memcg);
5431 }
5432 
5433 static void mem_cgroup_free(struct mem_cgroup *memcg)
5434 {
5435 	lru_gen_exit_memcg(memcg);
5436 	memcg_wb_domain_exit(memcg);
5437 	__mem_cgroup_free(memcg);
5438 }
5439 
5440 static struct mem_cgroup *mem_cgroup_alloc(void)
5441 {
5442 	struct mem_cgroup *memcg;
5443 	int node;
5444 	int __maybe_unused i;
5445 	long error = -ENOMEM;
5446 
5447 	memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5448 	if (!memcg)
5449 		return ERR_PTR(error);
5450 
5451 	memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5452 				 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
5453 	if (memcg->id.id < 0) {
5454 		error = memcg->id.id;
5455 		goto fail;
5456 	}
5457 
5458 	memcg->vmstats = kzalloc(sizeof(struct memcg_vmstats), GFP_KERNEL);
5459 	if (!memcg->vmstats)
5460 		goto fail;
5461 
5462 	memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5463 						 GFP_KERNEL_ACCOUNT);
5464 	if (!memcg->vmstats_percpu)
5465 		goto fail;
5466 
5467 	for_each_node(node)
5468 		if (alloc_mem_cgroup_per_node_info(memcg, node))
5469 			goto fail;
5470 
5471 	if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5472 		goto fail;
5473 
5474 	INIT_WORK(&memcg->high_work, high_work_func);
5475 	INIT_LIST_HEAD(&memcg->oom_notify);
5476 	mutex_init(&memcg->thresholds_lock);
5477 	spin_lock_init(&memcg->move_lock);
5478 	vmpressure_init(&memcg->vmpressure);
5479 	INIT_LIST_HEAD(&memcg->event_list);
5480 	spin_lock_init(&memcg->event_list_lock);
5481 	memcg->socket_pressure = jiffies;
5482 #ifdef CONFIG_MEMCG_KMEM
5483 	memcg->kmemcg_id = -1;
5484 	INIT_LIST_HEAD(&memcg->objcg_list);
5485 #endif
5486 #ifdef CONFIG_CGROUP_WRITEBACK
5487 	INIT_LIST_HEAD(&memcg->cgwb_list);
5488 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5489 		memcg->cgwb_frn[i].done =
5490 			__WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5491 #endif
5492 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5493 	spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5494 	INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5495 	memcg->deferred_split_queue.split_queue_len = 0;
5496 #endif
5497 	lru_gen_init_memcg(memcg);
5498 	return memcg;
5499 fail:
5500 	mem_cgroup_id_remove(memcg);
5501 	__mem_cgroup_free(memcg);
5502 	return ERR_PTR(error);
5503 }
5504 
5505 static struct cgroup_subsys_state * __ref
5506 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5507 {
5508 	struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5509 	struct mem_cgroup *memcg, *old_memcg;
5510 
5511 	old_memcg = set_active_memcg(parent);
5512 	memcg = mem_cgroup_alloc();
5513 	set_active_memcg(old_memcg);
5514 	if (IS_ERR(memcg))
5515 		return ERR_CAST(memcg);
5516 
5517 	page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5518 	WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5519 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5520 	memcg->zswap_max = PAGE_COUNTER_MAX;
5521 #endif
5522 	page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5523 	if (parent) {
5524 		WRITE_ONCE(memcg->swappiness, mem_cgroup_swappiness(parent));
5525 		WRITE_ONCE(memcg->oom_kill_disable, READ_ONCE(parent->oom_kill_disable));
5526 
5527 		page_counter_init(&memcg->memory, &parent->memory);
5528 		page_counter_init(&memcg->swap, &parent->swap);
5529 		page_counter_init(&memcg->kmem, &parent->kmem);
5530 		page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5531 	} else {
5532 		init_memcg_events();
5533 		page_counter_init(&memcg->memory, NULL);
5534 		page_counter_init(&memcg->swap, NULL);
5535 		page_counter_init(&memcg->kmem, NULL);
5536 		page_counter_init(&memcg->tcpmem, NULL);
5537 
5538 		root_mem_cgroup = memcg;
5539 		return &memcg->css;
5540 	}
5541 
5542 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5543 		static_branch_inc(&memcg_sockets_enabled_key);
5544 
5545 #if defined(CONFIG_MEMCG_KMEM)
5546 	if (!cgroup_memory_nobpf)
5547 		static_branch_inc(&memcg_bpf_enabled_key);
5548 #endif
5549 
5550 	return &memcg->css;
5551 }
5552 
5553 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5554 {
5555 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5556 
5557 	if (memcg_online_kmem(memcg))
5558 		goto remove_id;
5559 
5560 	/*
5561 	 * A memcg must be visible for expand_shrinker_info()
5562 	 * by the time the maps are allocated. So, we allocate maps
5563 	 * here, when for_each_mem_cgroup() can't skip it.
5564 	 */
5565 	if (alloc_shrinker_info(memcg))
5566 		goto offline_kmem;
5567 
5568 	if (unlikely(mem_cgroup_is_root(memcg)))
5569 		queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5570 				   FLUSH_TIME);
5571 	lru_gen_online_memcg(memcg);
5572 
5573 	/* Online state pins memcg ID, memcg ID pins CSS */
5574 	refcount_set(&memcg->id.ref, 1);
5575 	css_get(css);
5576 
5577 	/*
5578 	 * Ensure mem_cgroup_from_id() works once we're fully online.
5579 	 *
5580 	 * We could do this earlier and require callers to filter with
5581 	 * css_tryget_online(). But right now there are no users that
5582 	 * need earlier access, and the workingset code relies on the
5583 	 * cgroup tree linkage (mem_cgroup_get_nr_swap_pages()). So
5584 	 * publish it here at the end of onlining. This matches the
5585 	 * regular ID destruction during offlining.
5586 	 */
5587 	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5588 
5589 	return 0;
5590 offline_kmem:
5591 	memcg_offline_kmem(memcg);
5592 remove_id:
5593 	mem_cgroup_id_remove(memcg);
5594 	return -ENOMEM;
5595 }
5596 
5597 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5598 {
5599 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5600 	struct mem_cgroup_event *event, *tmp;
5601 
5602 	/*
5603 	 * Unregister events and notify userspace.
5604 	 * Notify userspace about cgroup removing only after rmdir of cgroup
5605 	 * directory to avoid race between userspace and kernelspace.
5606 	 */
5607 	spin_lock_irq(&memcg->event_list_lock);
5608 	list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5609 		list_del_init(&event->list);
5610 		schedule_work(&event->remove);
5611 	}
5612 	spin_unlock_irq(&memcg->event_list_lock);
5613 
5614 	page_counter_set_min(&memcg->memory, 0);
5615 	page_counter_set_low(&memcg->memory, 0);
5616 
5617 	memcg_offline_kmem(memcg);
5618 	reparent_shrinker_deferred(memcg);
5619 	wb_memcg_offline(memcg);
5620 	lru_gen_offline_memcg(memcg);
5621 
5622 	drain_all_stock(memcg);
5623 
5624 	mem_cgroup_id_put(memcg);
5625 }
5626 
5627 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5628 {
5629 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5630 
5631 	invalidate_reclaim_iterators(memcg);
5632 	lru_gen_release_memcg(memcg);
5633 }
5634 
5635 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5636 {
5637 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5638 	int __maybe_unused i;
5639 
5640 #ifdef CONFIG_CGROUP_WRITEBACK
5641 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5642 		wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5643 #endif
5644 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5645 		static_branch_dec(&memcg_sockets_enabled_key);
5646 
5647 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5648 		static_branch_dec(&memcg_sockets_enabled_key);
5649 
5650 #if defined(CONFIG_MEMCG_KMEM)
5651 	if (!cgroup_memory_nobpf)
5652 		static_branch_dec(&memcg_bpf_enabled_key);
5653 #endif
5654 
5655 	vmpressure_cleanup(&memcg->vmpressure);
5656 	cancel_work_sync(&memcg->high_work);
5657 	mem_cgroup_remove_from_trees(memcg);
5658 	free_shrinker_info(memcg);
5659 	mem_cgroup_free(memcg);
5660 }
5661 
5662 /**
5663  * mem_cgroup_css_reset - reset the states of a mem_cgroup
5664  * @css: the target css
5665  *
5666  * Reset the states of the mem_cgroup associated with @css.  This is
5667  * invoked when the userland requests disabling on the default hierarchy
5668  * but the memcg is pinned through dependency.  The memcg should stop
5669  * applying policies and should revert to the vanilla state as it may be
5670  * made visible again.
5671  *
5672  * The current implementation only resets the essential configurations.
5673  * This needs to be expanded to cover all the visible parts.
5674  */
5675 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5676 {
5677 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5678 
5679 	page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5680 	page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5681 	page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5682 	page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5683 	page_counter_set_min(&memcg->memory, 0);
5684 	page_counter_set_low(&memcg->memory, 0);
5685 	page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5686 	WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5687 	page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5688 	memcg_wb_domain_size_changed(memcg);
5689 }
5690 
5691 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5692 {
5693 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5694 	struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5695 	struct memcg_vmstats_percpu *statc;
5696 	long delta, delta_cpu, v;
5697 	int i, nid;
5698 
5699 	statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5700 
5701 	for (i = 0; i < MEMCG_NR_STAT; i++) {
5702 		/*
5703 		 * Collect the aggregated propagation counts of groups
5704 		 * below us. We're in a per-cpu loop here and this is
5705 		 * a global counter, so the first cycle will get them.
5706 		 */
5707 		delta = memcg->vmstats->state_pending[i];
5708 		if (delta)
5709 			memcg->vmstats->state_pending[i] = 0;
5710 
5711 		/* Add CPU changes on this level since the last flush */
5712 		delta_cpu = 0;
5713 		v = READ_ONCE(statc->state[i]);
5714 		if (v != statc->state_prev[i]) {
5715 			delta_cpu = v - statc->state_prev[i];
5716 			delta += delta_cpu;
5717 			statc->state_prev[i] = v;
5718 		}
5719 
5720 		/* Aggregate counts on this level and propagate upwards */
5721 		if (delta_cpu)
5722 			memcg->vmstats->state_local[i] += delta_cpu;
5723 
5724 		if (delta) {
5725 			memcg->vmstats->state[i] += delta;
5726 			if (parent)
5727 				parent->vmstats->state_pending[i] += delta;
5728 		}
5729 	}
5730 
5731 	for (i = 0; i < NR_MEMCG_EVENTS; i++) {
5732 		delta = memcg->vmstats->events_pending[i];
5733 		if (delta)
5734 			memcg->vmstats->events_pending[i] = 0;
5735 
5736 		delta_cpu = 0;
5737 		v = READ_ONCE(statc->events[i]);
5738 		if (v != statc->events_prev[i]) {
5739 			delta_cpu = v - statc->events_prev[i];
5740 			delta += delta_cpu;
5741 			statc->events_prev[i] = v;
5742 		}
5743 
5744 		if (delta_cpu)
5745 			memcg->vmstats->events_local[i] += delta_cpu;
5746 
5747 		if (delta) {
5748 			memcg->vmstats->events[i] += delta;
5749 			if (parent)
5750 				parent->vmstats->events_pending[i] += delta;
5751 		}
5752 	}
5753 
5754 	for_each_node_state(nid, N_MEMORY) {
5755 		struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5756 		struct mem_cgroup_per_node *ppn = NULL;
5757 		struct lruvec_stats_percpu *lstatc;
5758 
5759 		if (parent)
5760 			ppn = parent->nodeinfo[nid];
5761 
5762 		lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5763 
5764 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5765 			delta = pn->lruvec_stats.state_pending[i];
5766 			if (delta)
5767 				pn->lruvec_stats.state_pending[i] = 0;
5768 
5769 			delta_cpu = 0;
5770 			v = READ_ONCE(lstatc->state[i]);
5771 			if (v != lstatc->state_prev[i]) {
5772 				delta_cpu = v - lstatc->state_prev[i];
5773 				delta += delta_cpu;
5774 				lstatc->state_prev[i] = v;
5775 			}
5776 
5777 			if (delta_cpu)
5778 				pn->lruvec_stats.state_local[i] += delta_cpu;
5779 
5780 			if (delta) {
5781 				pn->lruvec_stats.state[i] += delta;
5782 				if (ppn)
5783 					ppn->lruvec_stats.state_pending[i] += delta;
5784 			}
5785 		}
5786 	}
5787 }
5788 
5789 #ifdef CONFIG_MMU
5790 /* Handlers for move charge at task migration. */
5791 static int mem_cgroup_do_precharge(unsigned long count)
5792 {
5793 	int ret;
5794 
5795 	/* Try a single bulk charge without reclaim first, kswapd may wake */
5796 	ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5797 	if (!ret) {
5798 		mc.precharge += count;
5799 		return ret;
5800 	}
5801 
5802 	/* Try charges one by one with reclaim, but do not retry */
5803 	while (count--) {
5804 		ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5805 		if (ret)
5806 			return ret;
5807 		mc.precharge++;
5808 		cond_resched();
5809 	}
5810 	return 0;
5811 }
5812 
5813 union mc_target {
5814 	struct page	*page;
5815 	swp_entry_t	ent;
5816 };
5817 
5818 enum mc_target_type {
5819 	MC_TARGET_NONE = 0,
5820 	MC_TARGET_PAGE,
5821 	MC_TARGET_SWAP,
5822 	MC_TARGET_DEVICE,
5823 };
5824 
5825 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5826 						unsigned long addr, pte_t ptent)
5827 {
5828 	struct page *page = vm_normal_page(vma, addr, ptent);
5829 
5830 	if (!page)
5831 		return NULL;
5832 	if (PageAnon(page)) {
5833 		if (!(mc.flags & MOVE_ANON))
5834 			return NULL;
5835 	} else {
5836 		if (!(mc.flags & MOVE_FILE))
5837 			return NULL;
5838 	}
5839 	get_page(page);
5840 
5841 	return page;
5842 }
5843 
5844 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5845 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5846 			pte_t ptent, swp_entry_t *entry)
5847 {
5848 	struct page *page = NULL;
5849 	swp_entry_t ent = pte_to_swp_entry(ptent);
5850 
5851 	if (!(mc.flags & MOVE_ANON))
5852 		return NULL;
5853 
5854 	/*
5855 	 * Handle device private pages that are not accessible by the CPU, but
5856 	 * stored as special swap entries in the page table.
5857 	 */
5858 	if (is_device_private_entry(ent)) {
5859 		page = pfn_swap_entry_to_page(ent);
5860 		if (!get_page_unless_zero(page))
5861 			return NULL;
5862 		return page;
5863 	}
5864 
5865 	if (non_swap_entry(ent))
5866 		return NULL;
5867 
5868 	/*
5869 	 * Because swap_cache_get_folio() updates some statistics counter,
5870 	 * we call find_get_page() with swapper_space directly.
5871 	 */
5872 	page = find_get_page(swap_address_space(ent), swp_offset(ent));
5873 	entry->val = ent.val;
5874 
5875 	return page;
5876 }
5877 #else
5878 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5879 			pte_t ptent, swp_entry_t *entry)
5880 {
5881 	return NULL;
5882 }
5883 #endif
5884 
5885 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5886 			unsigned long addr, pte_t ptent)
5887 {
5888 	unsigned long index;
5889 	struct folio *folio;
5890 
5891 	if (!vma->vm_file) /* anonymous vma */
5892 		return NULL;
5893 	if (!(mc.flags & MOVE_FILE))
5894 		return NULL;
5895 
5896 	/* folio is moved even if it's not RSS of this task(page-faulted). */
5897 	/* shmem/tmpfs may report page out on swap: account for that too. */
5898 	index = linear_page_index(vma, addr);
5899 	folio = filemap_get_incore_folio(vma->vm_file->f_mapping, index);
5900 	if (IS_ERR(folio))
5901 		return NULL;
5902 	return folio_file_page(folio, index);
5903 }
5904 
5905 /**
5906  * mem_cgroup_move_account - move account of the page
5907  * @page: the page
5908  * @compound: charge the page as compound or small page
5909  * @from: mem_cgroup which the page is moved from.
5910  * @to:	mem_cgroup which the page is moved to. @from != @to.
5911  *
5912  * The page must be locked and not on the LRU.
5913  *
5914  * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5915  * from old cgroup.
5916  */
5917 static int mem_cgroup_move_account(struct page *page,
5918 				   bool compound,
5919 				   struct mem_cgroup *from,
5920 				   struct mem_cgroup *to)
5921 {
5922 	struct folio *folio = page_folio(page);
5923 	struct lruvec *from_vec, *to_vec;
5924 	struct pglist_data *pgdat;
5925 	unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5926 	int nid, ret;
5927 
5928 	VM_BUG_ON(from == to);
5929 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
5930 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5931 	VM_BUG_ON(compound && !folio_test_large(folio));
5932 
5933 	ret = -EINVAL;
5934 	if (folio_memcg(folio) != from)
5935 		goto out;
5936 
5937 	pgdat = folio_pgdat(folio);
5938 	from_vec = mem_cgroup_lruvec(from, pgdat);
5939 	to_vec = mem_cgroup_lruvec(to, pgdat);
5940 
5941 	folio_memcg_lock(folio);
5942 
5943 	if (folio_test_anon(folio)) {
5944 		if (folio_mapped(folio)) {
5945 			__mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5946 			__mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5947 			if (folio_test_pmd_mappable(folio)) {
5948 				__mod_lruvec_state(from_vec, NR_ANON_THPS,
5949 						   -nr_pages);
5950 				__mod_lruvec_state(to_vec, NR_ANON_THPS,
5951 						   nr_pages);
5952 			}
5953 		}
5954 	} else {
5955 		__mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5956 		__mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5957 
5958 		if (folio_test_swapbacked(folio)) {
5959 			__mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5960 			__mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5961 		}
5962 
5963 		if (folio_mapped(folio)) {
5964 			__mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5965 			__mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5966 		}
5967 
5968 		if (folio_test_dirty(folio)) {
5969 			struct address_space *mapping = folio_mapping(folio);
5970 
5971 			if (mapping_can_writeback(mapping)) {
5972 				__mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5973 						   -nr_pages);
5974 				__mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5975 						   nr_pages);
5976 			}
5977 		}
5978 	}
5979 
5980 #ifdef CONFIG_SWAP
5981 	if (folio_test_swapcache(folio)) {
5982 		__mod_lruvec_state(from_vec, NR_SWAPCACHE, -nr_pages);
5983 		__mod_lruvec_state(to_vec, NR_SWAPCACHE, nr_pages);
5984 	}
5985 #endif
5986 	if (folio_test_writeback(folio)) {
5987 		__mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5988 		__mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5989 	}
5990 
5991 	/*
5992 	 * All state has been migrated, let's switch to the new memcg.
5993 	 *
5994 	 * It is safe to change page's memcg here because the page
5995 	 * is referenced, charged, isolated, and locked: we can't race
5996 	 * with (un)charging, migration, LRU putback, or anything else
5997 	 * that would rely on a stable page's memory cgroup.
5998 	 *
5999 	 * Note that folio_memcg_lock is a memcg lock, not a page lock,
6000 	 * to save space. As soon as we switch page's memory cgroup to a
6001 	 * new memcg that isn't locked, the above state can change
6002 	 * concurrently again. Make sure we're truly done with it.
6003 	 */
6004 	smp_mb();
6005 
6006 	css_get(&to->css);
6007 	css_put(&from->css);
6008 
6009 	folio->memcg_data = (unsigned long)to;
6010 
6011 	__folio_memcg_unlock(from);
6012 
6013 	ret = 0;
6014 	nid = folio_nid(folio);
6015 
6016 	local_irq_disable();
6017 	mem_cgroup_charge_statistics(to, nr_pages);
6018 	memcg_check_events(to, nid);
6019 	mem_cgroup_charge_statistics(from, -nr_pages);
6020 	memcg_check_events(from, nid);
6021 	local_irq_enable();
6022 out:
6023 	return ret;
6024 }
6025 
6026 /**
6027  * get_mctgt_type - get target type of moving charge
6028  * @vma: the vma the pte to be checked belongs
6029  * @addr: the address corresponding to the pte to be checked
6030  * @ptent: the pte to be checked
6031  * @target: the pointer the target page or swap ent will be stored(can be NULL)
6032  *
6033  * Context: Called with pte lock held.
6034  * Return:
6035  * * MC_TARGET_NONE - If the pte is not a target for move charge.
6036  * * MC_TARGET_PAGE - If the page corresponding to this pte is a target for
6037  *   move charge. If @target is not NULL, the page is stored in target->page
6038  *   with extra refcnt taken (Caller should release it).
6039  * * MC_TARGET_SWAP - If the swap entry corresponding to this pte is a
6040  *   target for charge migration.  If @target is not NULL, the entry is
6041  *   stored in target->ent.
6042  * * MC_TARGET_DEVICE - Like MC_TARGET_PAGE but page is device memory and
6043  *   thus not on the lru.  For now such page is charged like a regular page
6044  *   would be as it is just special memory taking the place of a regular page.
6045  *   See Documentations/vm/hmm.txt and include/linux/hmm.h
6046  */
6047 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6048 		unsigned long addr, pte_t ptent, union mc_target *target)
6049 {
6050 	struct page *page = NULL;
6051 	enum mc_target_type ret = MC_TARGET_NONE;
6052 	swp_entry_t ent = { .val = 0 };
6053 
6054 	if (pte_present(ptent))
6055 		page = mc_handle_present_pte(vma, addr, ptent);
6056 	else if (pte_none_mostly(ptent))
6057 		/*
6058 		 * PTE markers should be treated as a none pte here, separated
6059 		 * from other swap handling below.
6060 		 */
6061 		page = mc_handle_file_pte(vma, addr, ptent);
6062 	else if (is_swap_pte(ptent))
6063 		page = mc_handle_swap_pte(vma, ptent, &ent);
6064 
6065 	if (target && page) {
6066 		if (!trylock_page(page)) {
6067 			put_page(page);
6068 			return ret;
6069 		}
6070 		/*
6071 		 * page_mapped() must be stable during the move. This
6072 		 * pte is locked, so if it's present, the page cannot
6073 		 * become unmapped. If it isn't, we have only partial
6074 		 * control over the mapped state: the page lock will
6075 		 * prevent new faults against pagecache and swapcache,
6076 		 * so an unmapped page cannot become mapped. However,
6077 		 * if the page is already mapped elsewhere, it can
6078 		 * unmap, and there is nothing we can do about it.
6079 		 * Alas, skip moving the page in this case.
6080 		 */
6081 		if (!pte_present(ptent) && page_mapped(page)) {
6082 			unlock_page(page);
6083 			put_page(page);
6084 			return ret;
6085 		}
6086 	}
6087 
6088 	if (!page && !ent.val)
6089 		return ret;
6090 	if (page) {
6091 		/*
6092 		 * Do only loose check w/o serialization.
6093 		 * mem_cgroup_move_account() checks the page is valid or
6094 		 * not under LRU exclusion.
6095 		 */
6096 		if (page_memcg(page) == mc.from) {
6097 			ret = MC_TARGET_PAGE;
6098 			if (is_device_private_page(page) ||
6099 			    is_device_coherent_page(page))
6100 				ret = MC_TARGET_DEVICE;
6101 			if (target)
6102 				target->page = page;
6103 		}
6104 		if (!ret || !target) {
6105 			if (target)
6106 				unlock_page(page);
6107 			put_page(page);
6108 		}
6109 	}
6110 	/*
6111 	 * There is a swap entry and a page doesn't exist or isn't charged.
6112 	 * But we cannot move a tail-page in a THP.
6113 	 */
6114 	if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
6115 	    mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6116 		ret = MC_TARGET_SWAP;
6117 		if (target)
6118 			target->ent = ent;
6119 	}
6120 	return ret;
6121 }
6122 
6123 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6124 /*
6125  * We don't consider PMD mapped swapping or file mapped pages because THP does
6126  * not support them for now.
6127  * Caller should make sure that pmd_trans_huge(pmd) is true.
6128  */
6129 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6130 		unsigned long addr, pmd_t pmd, union mc_target *target)
6131 {
6132 	struct page *page = NULL;
6133 	enum mc_target_type ret = MC_TARGET_NONE;
6134 
6135 	if (unlikely(is_swap_pmd(pmd))) {
6136 		VM_BUG_ON(thp_migration_supported() &&
6137 				  !is_pmd_migration_entry(pmd));
6138 		return ret;
6139 	}
6140 	page = pmd_page(pmd);
6141 	VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6142 	if (!(mc.flags & MOVE_ANON))
6143 		return ret;
6144 	if (page_memcg(page) == mc.from) {
6145 		ret = MC_TARGET_PAGE;
6146 		if (target) {
6147 			get_page(page);
6148 			if (!trylock_page(page)) {
6149 				put_page(page);
6150 				return MC_TARGET_NONE;
6151 			}
6152 			target->page = page;
6153 		}
6154 	}
6155 	return ret;
6156 }
6157 #else
6158 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6159 		unsigned long addr, pmd_t pmd, union mc_target *target)
6160 {
6161 	return MC_TARGET_NONE;
6162 }
6163 #endif
6164 
6165 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6166 					unsigned long addr, unsigned long end,
6167 					struct mm_walk *walk)
6168 {
6169 	struct vm_area_struct *vma = walk->vma;
6170 	pte_t *pte;
6171 	spinlock_t *ptl;
6172 
6173 	ptl = pmd_trans_huge_lock(pmd, vma);
6174 	if (ptl) {
6175 		/*
6176 		 * Note their can not be MC_TARGET_DEVICE for now as we do not
6177 		 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
6178 		 * this might change.
6179 		 */
6180 		if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6181 			mc.precharge += HPAGE_PMD_NR;
6182 		spin_unlock(ptl);
6183 		return 0;
6184 	}
6185 
6186 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6187 	if (!pte)
6188 		return 0;
6189 	for (; addr != end; pte++, addr += PAGE_SIZE)
6190 		if (get_mctgt_type(vma, addr, ptep_get(pte), NULL))
6191 			mc.precharge++;	/* increment precharge temporarily */
6192 	pte_unmap_unlock(pte - 1, ptl);
6193 	cond_resched();
6194 
6195 	return 0;
6196 }
6197 
6198 static const struct mm_walk_ops precharge_walk_ops = {
6199 	.pmd_entry	= mem_cgroup_count_precharge_pte_range,
6200 	.walk_lock	= PGWALK_RDLOCK,
6201 };
6202 
6203 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6204 {
6205 	unsigned long precharge;
6206 
6207 	mmap_read_lock(mm);
6208 	walk_page_range(mm, 0, ULONG_MAX, &precharge_walk_ops, NULL);
6209 	mmap_read_unlock(mm);
6210 
6211 	precharge = mc.precharge;
6212 	mc.precharge = 0;
6213 
6214 	return precharge;
6215 }
6216 
6217 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6218 {
6219 	unsigned long precharge = mem_cgroup_count_precharge(mm);
6220 
6221 	VM_BUG_ON(mc.moving_task);
6222 	mc.moving_task = current;
6223 	return mem_cgroup_do_precharge(precharge);
6224 }
6225 
6226 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6227 static void __mem_cgroup_clear_mc(void)
6228 {
6229 	struct mem_cgroup *from = mc.from;
6230 	struct mem_cgroup *to = mc.to;
6231 
6232 	/* we must uncharge all the leftover precharges from mc.to */
6233 	if (mc.precharge) {
6234 		mem_cgroup_cancel_charge(mc.to, mc.precharge);
6235 		mc.precharge = 0;
6236 	}
6237 	/*
6238 	 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6239 	 * we must uncharge here.
6240 	 */
6241 	if (mc.moved_charge) {
6242 		mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6243 		mc.moved_charge = 0;
6244 	}
6245 	/* we must fixup refcnts and charges */
6246 	if (mc.moved_swap) {
6247 		/* uncharge swap account from the old cgroup */
6248 		if (!mem_cgroup_is_root(mc.from))
6249 			page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6250 
6251 		mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6252 
6253 		/*
6254 		 * we charged both to->memory and to->memsw, so we
6255 		 * should uncharge to->memory.
6256 		 */
6257 		if (!mem_cgroup_is_root(mc.to))
6258 			page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6259 
6260 		mc.moved_swap = 0;
6261 	}
6262 	memcg_oom_recover(from);
6263 	memcg_oom_recover(to);
6264 	wake_up_all(&mc.waitq);
6265 }
6266 
6267 static void mem_cgroup_clear_mc(void)
6268 {
6269 	struct mm_struct *mm = mc.mm;
6270 
6271 	/*
6272 	 * we must clear moving_task before waking up waiters at the end of
6273 	 * task migration.
6274 	 */
6275 	mc.moving_task = NULL;
6276 	__mem_cgroup_clear_mc();
6277 	spin_lock(&mc.lock);
6278 	mc.from = NULL;
6279 	mc.to = NULL;
6280 	mc.mm = NULL;
6281 	spin_unlock(&mc.lock);
6282 
6283 	mmput(mm);
6284 }
6285 
6286 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6287 {
6288 	struct cgroup_subsys_state *css;
6289 	struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6290 	struct mem_cgroup *from;
6291 	struct task_struct *leader, *p;
6292 	struct mm_struct *mm;
6293 	unsigned long move_flags;
6294 	int ret = 0;
6295 
6296 	/* charge immigration isn't supported on the default hierarchy */
6297 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6298 		return 0;
6299 
6300 	/*
6301 	 * Multi-process migrations only happen on the default hierarchy
6302 	 * where charge immigration is not used.  Perform charge
6303 	 * immigration if @tset contains a leader and whine if there are
6304 	 * multiple.
6305 	 */
6306 	p = NULL;
6307 	cgroup_taskset_for_each_leader(leader, css, tset) {
6308 		WARN_ON_ONCE(p);
6309 		p = leader;
6310 		memcg = mem_cgroup_from_css(css);
6311 	}
6312 	if (!p)
6313 		return 0;
6314 
6315 	/*
6316 	 * We are now committed to this value whatever it is. Changes in this
6317 	 * tunable will only affect upcoming migrations, not the current one.
6318 	 * So we need to save it, and keep it going.
6319 	 */
6320 	move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6321 	if (!move_flags)
6322 		return 0;
6323 
6324 	from = mem_cgroup_from_task(p);
6325 
6326 	VM_BUG_ON(from == memcg);
6327 
6328 	mm = get_task_mm(p);
6329 	if (!mm)
6330 		return 0;
6331 	/* We move charges only when we move a owner of the mm */
6332 	if (mm->owner == p) {
6333 		VM_BUG_ON(mc.from);
6334 		VM_BUG_ON(mc.to);
6335 		VM_BUG_ON(mc.precharge);
6336 		VM_BUG_ON(mc.moved_charge);
6337 		VM_BUG_ON(mc.moved_swap);
6338 
6339 		spin_lock(&mc.lock);
6340 		mc.mm = mm;
6341 		mc.from = from;
6342 		mc.to = memcg;
6343 		mc.flags = move_flags;
6344 		spin_unlock(&mc.lock);
6345 		/* We set mc.moving_task later */
6346 
6347 		ret = mem_cgroup_precharge_mc(mm);
6348 		if (ret)
6349 			mem_cgroup_clear_mc();
6350 	} else {
6351 		mmput(mm);
6352 	}
6353 	return ret;
6354 }
6355 
6356 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6357 {
6358 	if (mc.to)
6359 		mem_cgroup_clear_mc();
6360 }
6361 
6362 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6363 				unsigned long addr, unsigned long end,
6364 				struct mm_walk *walk)
6365 {
6366 	int ret = 0;
6367 	struct vm_area_struct *vma = walk->vma;
6368 	pte_t *pte;
6369 	spinlock_t *ptl;
6370 	enum mc_target_type target_type;
6371 	union mc_target target;
6372 	struct page *page;
6373 
6374 	ptl = pmd_trans_huge_lock(pmd, vma);
6375 	if (ptl) {
6376 		if (mc.precharge < HPAGE_PMD_NR) {
6377 			spin_unlock(ptl);
6378 			return 0;
6379 		}
6380 		target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6381 		if (target_type == MC_TARGET_PAGE) {
6382 			page = target.page;
6383 			if (isolate_lru_page(page)) {
6384 				if (!mem_cgroup_move_account(page, true,
6385 							     mc.from, mc.to)) {
6386 					mc.precharge -= HPAGE_PMD_NR;
6387 					mc.moved_charge += HPAGE_PMD_NR;
6388 				}
6389 				putback_lru_page(page);
6390 			}
6391 			unlock_page(page);
6392 			put_page(page);
6393 		} else if (target_type == MC_TARGET_DEVICE) {
6394 			page = target.page;
6395 			if (!mem_cgroup_move_account(page, true,
6396 						     mc.from, mc.to)) {
6397 				mc.precharge -= HPAGE_PMD_NR;
6398 				mc.moved_charge += HPAGE_PMD_NR;
6399 			}
6400 			unlock_page(page);
6401 			put_page(page);
6402 		}
6403 		spin_unlock(ptl);
6404 		return 0;
6405 	}
6406 
6407 retry:
6408 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6409 	if (!pte)
6410 		return 0;
6411 	for (; addr != end; addr += PAGE_SIZE) {
6412 		pte_t ptent = ptep_get(pte++);
6413 		bool device = false;
6414 		swp_entry_t ent;
6415 
6416 		if (!mc.precharge)
6417 			break;
6418 
6419 		switch (get_mctgt_type(vma, addr, ptent, &target)) {
6420 		case MC_TARGET_DEVICE:
6421 			device = true;
6422 			fallthrough;
6423 		case MC_TARGET_PAGE:
6424 			page = target.page;
6425 			/*
6426 			 * We can have a part of the split pmd here. Moving it
6427 			 * can be done but it would be too convoluted so simply
6428 			 * ignore such a partial THP and keep it in original
6429 			 * memcg. There should be somebody mapping the head.
6430 			 */
6431 			if (PageTransCompound(page))
6432 				goto put;
6433 			if (!device && !isolate_lru_page(page))
6434 				goto put;
6435 			if (!mem_cgroup_move_account(page, false,
6436 						mc.from, mc.to)) {
6437 				mc.precharge--;
6438 				/* we uncharge from mc.from later. */
6439 				mc.moved_charge++;
6440 			}
6441 			if (!device)
6442 				putback_lru_page(page);
6443 put:			/* get_mctgt_type() gets & locks the page */
6444 			unlock_page(page);
6445 			put_page(page);
6446 			break;
6447 		case MC_TARGET_SWAP:
6448 			ent = target.ent;
6449 			if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6450 				mc.precharge--;
6451 				mem_cgroup_id_get_many(mc.to, 1);
6452 				/* we fixup other refcnts and charges later. */
6453 				mc.moved_swap++;
6454 			}
6455 			break;
6456 		default:
6457 			break;
6458 		}
6459 	}
6460 	pte_unmap_unlock(pte - 1, ptl);
6461 	cond_resched();
6462 
6463 	if (addr != end) {
6464 		/*
6465 		 * We have consumed all precharges we got in can_attach().
6466 		 * We try charge one by one, but don't do any additional
6467 		 * charges to mc.to if we have failed in charge once in attach()
6468 		 * phase.
6469 		 */
6470 		ret = mem_cgroup_do_precharge(1);
6471 		if (!ret)
6472 			goto retry;
6473 	}
6474 
6475 	return ret;
6476 }
6477 
6478 static const struct mm_walk_ops charge_walk_ops = {
6479 	.pmd_entry	= mem_cgroup_move_charge_pte_range,
6480 	.walk_lock	= PGWALK_RDLOCK,
6481 };
6482 
6483 static void mem_cgroup_move_charge(void)
6484 {
6485 	lru_add_drain_all();
6486 	/*
6487 	 * Signal folio_memcg_lock() to take the memcg's move_lock
6488 	 * while we're moving its pages to another memcg. Then wait
6489 	 * for already started RCU-only updates to finish.
6490 	 */
6491 	atomic_inc(&mc.from->moving_account);
6492 	synchronize_rcu();
6493 retry:
6494 	if (unlikely(!mmap_read_trylock(mc.mm))) {
6495 		/*
6496 		 * Someone who are holding the mmap_lock might be waiting in
6497 		 * waitq. So we cancel all extra charges, wake up all waiters,
6498 		 * and retry. Because we cancel precharges, we might not be able
6499 		 * to move enough charges, but moving charge is a best-effort
6500 		 * feature anyway, so it wouldn't be a big problem.
6501 		 */
6502 		__mem_cgroup_clear_mc();
6503 		cond_resched();
6504 		goto retry;
6505 	}
6506 	/*
6507 	 * When we have consumed all precharges and failed in doing
6508 	 * additional charge, the page walk just aborts.
6509 	 */
6510 	walk_page_range(mc.mm, 0, ULONG_MAX, &charge_walk_ops, NULL);
6511 	mmap_read_unlock(mc.mm);
6512 	atomic_dec(&mc.from->moving_account);
6513 }
6514 
6515 static void mem_cgroup_move_task(void)
6516 {
6517 	if (mc.to) {
6518 		mem_cgroup_move_charge();
6519 		mem_cgroup_clear_mc();
6520 	}
6521 }
6522 
6523 #else	/* !CONFIG_MMU */
6524 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6525 {
6526 	return 0;
6527 }
6528 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6529 {
6530 }
6531 static void mem_cgroup_move_task(void)
6532 {
6533 }
6534 #endif
6535 
6536 #ifdef CONFIG_MEMCG_KMEM
6537 static void mem_cgroup_fork(struct task_struct *task)
6538 {
6539 	/*
6540 	 * Set the update flag to cause task->objcg to be initialized lazily
6541 	 * on the first allocation. It can be done without any synchronization
6542 	 * because it's always performed on the current task, so does
6543 	 * current_objcg_update().
6544 	 */
6545 	task->objcg = (struct obj_cgroup *)CURRENT_OBJCG_UPDATE_FLAG;
6546 }
6547 
6548 static void mem_cgroup_exit(struct task_struct *task)
6549 {
6550 	struct obj_cgroup *objcg = task->objcg;
6551 
6552 	objcg = (struct obj_cgroup *)
6553 		((unsigned long)objcg & ~CURRENT_OBJCG_UPDATE_FLAG);
6554 	if (objcg)
6555 		obj_cgroup_put(objcg);
6556 
6557 	/*
6558 	 * Some kernel allocations can happen after this point,
6559 	 * but let's ignore them. It can be done without any synchronization
6560 	 * because it's always performed on the current task, so does
6561 	 * current_objcg_update().
6562 	 */
6563 	task->objcg = NULL;
6564 }
6565 #endif
6566 
6567 #ifdef CONFIG_LRU_GEN
6568 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset)
6569 {
6570 	struct task_struct *task;
6571 	struct cgroup_subsys_state *css;
6572 
6573 	/* find the first leader if there is any */
6574 	cgroup_taskset_for_each_leader(task, css, tset)
6575 		break;
6576 
6577 	if (!task)
6578 		return;
6579 
6580 	task_lock(task);
6581 	if (task->mm && READ_ONCE(task->mm->owner) == task)
6582 		lru_gen_migrate_mm(task->mm);
6583 	task_unlock(task);
6584 }
6585 #else
6586 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset) {}
6587 #endif /* CONFIG_LRU_GEN */
6588 
6589 #ifdef CONFIG_MEMCG_KMEM
6590 static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset)
6591 {
6592 	struct task_struct *task;
6593 	struct cgroup_subsys_state *css;
6594 
6595 	cgroup_taskset_for_each(task, css, tset) {
6596 		/* atomically set the update bit */
6597 		set_bit(CURRENT_OBJCG_UPDATE_BIT, (unsigned long *)&task->objcg);
6598 	}
6599 }
6600 #else
6601 static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset) {}
6602 #endif /* CONFIG_MEMCG_KMEM */
6603 
6604 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
6605 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6606 {
6607 	mem_cgroup_lru_gen_attach(tset);
6608 	mem_cgroup_kmem_attach(tset);
6609 }
6610 #endif
6611 
6612 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6613 {
6614 	if (value == PAGE_COUNTER_MAX)
6615 		seq_puts(m, "max\n");
6616 	else
6617 		seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6618 
6619 	return 0;
6620 }
6621 
6622 static u64 memory_current_read(struct cgroup_subsys_state *css,
6623 			       struct cftype *cft)
6624 {
6625 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6626 
6627 	return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6628 }
6629 
6630 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6631 			    struct cftype *cft)
6632 {
6633 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6634 
6635 	return (u64)memcg->memory.watermark * PAGE_SIZE;
6636 }
6637 
6638 static int memory_min_show(struct seq_file *m, void *v)
6639 {
6640 	return seq_puts_memcg_tunable(m,
6641 		READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6642 }
6643 
6644 static ssize_t memory_min_write(struct kernfs_open_file *of,
6645 				char *buf, size_t nbytes, loff_t off)
6646 {
6647 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6648 	unsigned long min;
6649 	int err;
6650 
6651 	buf = strstrip(buf);
6652 	err = page_counter_memparse(buf, "max", &min);
6653 	if (err)
6654 		return err;
6655 
6656 	page_counter_set_min(&memcg->memory, min);
6657 
6658 	return nbytes;
6659 }
6660 
6661 static int memory_low_show(struct seq_file *m, void *v)
6662 {
6663 	return seq_puts_memcg_tunable(m,
6664 		READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6665 }
6666 
6667 static ssize_t memory_low_write(struct kernfs_open_file *of,
6668 				char *buf, size_t nbytes, loff_t off)
6669 {
6670 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6671 	unsigned long low;
6672 	int err;
6673 
6674 	buf = strstrip(buf);
6675 	err = page_counter_memparse(buf, "max", &low);
6676 	if (err)
6677 		return err;
6678 
6679 	page_counter_set_low(&memcg->memory, low);
6680 
6681 	return nbytes;
6682 }
6683 
6684 static int memory_high_show(struct seq_file *m, void *v)
6685 {
6686 	return seq_puts_memcg_tunable(m,
6687 		READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6688 }
6689 
6690 static ssize_t memory_high_write(struct kernfs_open_file *of,
6691 				 char *buf, size_t nbytes, loff_t off)
6692 {
6693 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6694 	unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6695 	bool drained = false;
6696 	unsigned long high;
6697 	int err;
6698 
6699 	buf = strstrip(buf);
6700 	err = page_counter_memparse(buf, "max", &high);
6701 	if (err)
6702 		return err;
6703 
6704 	page_counter_set_high(&memcg->memory, high);
6705 
6706 	for (;;) {
6707 		unsigned long nr_pages = page_counter_read(&memcg->memory);
6708 		unsigned long reclaimed;
6709 
6710 		if (nr_pages <= high)
6711 			break;
6712 
6713 		if (signal_pending(current))
6714 			break;
6715 
6716 		if (!drained) {
6717 			drain_all_stock(memcg);
6718 			drained = true;
6719 			continue;
6720 		}
6721 
6722 		reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6723 					GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP);
6724 
6725 		if (!reclaimed && !nr_retries--)
6726 			break;
6727 	}
6728 
6729 	memcg_wb_domain_size_changed(memcg);
6730 	return nbytes;
6731 }
6732 
6733 static int memory_max_show(struct seq_file *m, void *v)
6734 {
6735 	return seq_puts_memcg_tunable(m,
6736 		READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6737 }
6738 
6739 static ssize_t memory_max_write(struct kernfs_open_file *of,
6740 				char *buf, size_t nbytes, loff_t off)
6741 {
6742 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6743 	unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6744 	bool drained = false;
6745 	unsigned long max;
6746 	int err;
6747 
6748 	buf = strstrip(buf);
6749 	err = page_counter_memparse(buf, "max", &max);
6750 	if (err)
6751 		return err;
6752 
6753 	xchg(&memcg->memory.max, max);
6754 
6755 	for (;;) {
6756 		unsigned long nr_pages = page_counter_read(&memcg->memory);
6757 
6758 		if (nr_pages <= max)
6759 			break;
6760 
6761 		if (signal_pending(current))
6762 			break;
6763 
6764 		if (!drained) {
6765 			drain_all_stock(memcg);
6766 			drained = true;
6767 			continue;
6768 		}
6769 
6770 		if (nr_reclaims) {
6771 			if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6772 					GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP))
6773 				nr_reclaims--;
6774 			continue;
6775 		}
6776 
6777 		memcg_memory_event(memcg, MEMCG_OOM);
6778 		if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6779 			break;
6780 	}
6781 
6782 	memcg_wb_domain_size_changed(memcg);
6783 	return nbytes;
6784 }
6785 
6786 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6787 {
6788 	seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6789 	seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6790 	seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6791 	seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6792 	seq_printf(m, "oom_kill %lu\n",
6793 		   atomic_long_read(&events[MEMCG_OOM_KILL]));
6794 	seq_printf(m, "oom_group_kill %lu\n",
6795 		   atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6796 }
6797 
6798 static int memory_events_show(struct seq_file *m, void *v)
6799 {
6800 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6801 
6802 	__memory_events_show(m, memcg->memory_events);
6803 	return 0;
6804 }
6805 
6806 static int memory_events_local_show(struct seq_file *m, void *v)
6807 {
6808 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6809 
6810 	__memory_events_show(m, memcg->memory_events_local);
6811 	return 0;
6812 }
6813 
6814 static int memory_stat_show(struct seq_file *m, void *v)
6815 {
6816 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6817 	char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
6818 	struct seq_buf s;
6819 
6820 	if (!buf)
6821 		return -ENOMEM;
6822 	seq_buf_init(&s, buf, PAGE_SIZE);
6823 	memory_stat_format(memcg, &s);
6824 	seq_puts(m, buf);
6825 	kfree(buf);
6826 	return 0;
6827 }
6828 
6829 #ifdef CONFIG_NUMA
6830 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6831 						     int item)
6832 {
6833 	return lruvec_page_state(lruvec, item) *
6834 		memcg_page_state_output_unit(item);
6835 }
6836 
6837 static int memory_numa_stat_show(struct seq_file *m, void *v)
6838 {
6839 	int i;
6840 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6841 
6842 	mem_cgroup_flush_stats();
6843 
6844 	for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6845 		int nid;
6846 
6847 		if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6848 			continue;
6849 
6850 		seq_printf(m, "%s", memory_stats[i].name);
6851 		for_each_node_state(nid, N_MEMORY) {
6852 			u64 size;
6853 			struct lruvec *lruvec;
6854 
6855 			lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6856 			size = lruvec_page_state_output(lruvec,
6857 							memory_stats[i].idx);
6858 			seq_printf(m, " N%d=%llu", nid, size);
6859 		}
6860 		seq_putc(m, '\n');
6861 	}
6862 
6863 	return 0;
6864 }
6865 #endif
6866 
6867 static int memory_oom_group_show(struct seq_file *m, void *v)
6868 {
6869 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6870 
6871 	seq_printf(m, "%d\n", READ_ONCE(memcg->oom_group));
6872 
6873 	return 0;
6874 }
6875 
6876 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6877 				      char *buf, size_t nbytes, loff_t off)
6878 {
6879 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6880 	int ret, oom_group;
6881 
6882 	buf = strstrip(buf);
6883 	if (!buf)
6884 		return -EINVAL;
6885 
6886 	ret = kstrtoint(buf, 0, &oom_group);
6887 	if (ret)
6888 		return ret;
6889 
6890 	if (oom_group != 0 && oom_group != 1)
6891 		return -EINVAL;
6892 
6893 	WRITE_ONCE(memcg->oom_group, oom_group);
6894 
6895 	return nbytes;
6896 }
6897 
6898 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6899 			      size_t nbytes, loff_t off)
6900 {
6901 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6902 	unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6903 	unsigned long nr_to_reclaim, nr_reclaimed = 0;
6904 	unsigned int reclaim_options;
6905 	int err;
6906 
6907 	buf = strstrip(buf);
6908 	err = page_counter_memparse(buf, "", &nr_to_reclaim);
6909 	if (err)
6910 		return err;
6911 
6912 	reclaim_options	= MEMCG_RECLAIM_MAY_SWAP | MEMCG_RECLAIM_PROACTIVE;
6913 	while (nr_reclaimed < nr_to_reclaim) {
6914 		unsigned long reclaimed;
6915 
6916 		if (signal_pending(current))
6917 			return -EINTR;
6918 
6919 		/*
6920 		 * This is the final attempt, drain percpu lru caches in the
6921 		 * hope of introducing more evictable pages for
6922 		 * try_to_free_mem_cgroup_pages().
6923 		 */
6924 		if (!nr_retries)
6925 			lru_add_drain_all();
6926 
6927 		reclaimed = try_to_free_mem_cgroup_pages(memcg,
6928 					min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
6929 					GFP_KERNEL, reclaim_options);
6930 
6931 		if (!reclaimed && !nr_retries--)
6932 			return -EAGAIN;
6933 
6934 		nr_reclaimed += reclaimed;
6935 	}
6936 
6937 	return nbytes;
6938 }
6939 
6940 static struct cftype memory_files[] = {
6941 	{
6942 		.name = "current",
6943 		.flags = CFTYPE_NOT_ON_ROOT,
6944 		.read_u64 = memory_current_read,
6945 	},
6946 	{
6947 		.name = "peak",
6948 		.flags = CFTYPE_NOT_ON_ROOT,
6949 		.read_u64 = memory_peak_read,
6950 	},
6951 	{
6952 		.name = "min",
6953 		.flags = CFTYPE_NOT_ON_ROOT,
6954 		.seq_show = memory_min_show,
6955 		.write = memory_min_write,
6956 	},
6957 	{
6958 		.name = "low",
6959 		.flags = CFTYPE_NOT_ON_ROOT,
6960 		.seq_show = memory_low_show,
6961 		.write = memory_low_write,
6962 	},
6963 	{
6964 		.name = "high",
6965 		.flags = CFTYPE_NOT_ON_ROOT,
6966 		.seq_show = memory_high_show,
6967 		.write = memory_high_write,
6968 	},
6969 	{
6970 		.name = "max",
6971 		.flags = CFTYPE_NOT_ON_ROOT,
6972 		.seq_show = memory_max_show,
6973 		.write = memory_max_write,
6974 	},
6975 	{
6976 		.name = "events",
6977 		.flags = CFTYPE_NOT_ON_ROOT,
6978 		.file_offset = offsetof(struct mem_cgroup, events_file),
6979 		.seq_show = memory_events_show,
6980 	},
6981 	{
6982 		.name = "events.local",
6983 		.flags = CFTYPE_NOT_ON_ROOT,
6984 		.file_offset = offsetof(struct mem_cgroup, events_local_file),
6985 		.seq_show = memory_events_local_show,
6986 	},
6987 	{
6988 		.name = "stat",
6989 		.seq_show = memory_stat_show,
6990 	},
6991 #ifdef CONFIG_NUMA
6992 	{
6993 		.name = "numa_stat",
6994 		.seq_show = memory_numa_stat_show,
6995 	},
6996 #endif
6997 	{
6998 		.name = "oom.group",
6999 		.flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
7000 		.seq_show = memory_oom_group_show,
7001 		.write = memory_oom_group_write,
7002 	},
7003 	{
7004 		.name = "reclaim",
7005 		.flags = CFTYPE_NS_DELEGATABLE,
7006 		.write = memory_reclaim,
7007 	},
7008 	{ }	/* terminate */
7009 };
7010 
7011 struct cgroup_subsys memory_cgrp_subsys = {
7012 	.css_alloc = mem_cgroup_css_alloc,
7013 	.css_online = mem_cgroup_css_online,
7014 	.css_offline = mem_cgroup_css_offline,
7015 	.css_released = mem_cgroup_css_released,
7016 	.css_free = mem_cgroup_css_free,
7017 	.css_reset = mem_cgroup_css_reset,
7018 	.css_rstat_flush = mem_cgroup_css_rstat_flush,
7019 	.can_attach = mem_cgroup_can_attach,
7020 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
7021 	.attach = mem_cgroup_attach,
7022 #endif
7023 	.cancel_attach = mem_cgroup_cancel_attach,
7024 	.post_attach = mem_cgroup_move_task,
7025 #ifdef CONFIG_MEMCG_KMEM
7026 	.fork = mem_cgroup_fork,
7027 	.exit = mem_cgroup_exit,
7028 #endif
7029 	.dfl_cftypes = memory_files,
7030 	.legacy_cftypes = mem_cgroup_legacy_files,
7031 	.early_init = 0,
7032 };
7033 
7034 /*
7035  * This function calculates an individual cgroup's effective
7036  * protection which is derived from its own memory.min/low, its
7037  * parent's and siblings' settings, as well as the actual memory
7038  * distribution in the tree.
7039  *
7040  * The following rules apply to the effective protection values:
7041  *
7042  * 1. At the first level of reclaim, effective protection is equal to
7043  *    the declared protection in memory.min and memory.low.
7044  *
7045  * 2. To enable safe delegation of the protection configuration, at
7046  *    subsequent levels the effective protection is capped to the
7047  *    parent's effective protection.
7048  *
7049  * 3. To make complex and dynamic subtrees easier to configure, the
7050  *    user is allowed to overcommit the declared protection at a given
7051  *    level. If that is the case, the parent's effective protection is
7052  *    distributed to the children in proportion to how much protection
7053  *    they have declared and how much of it they are utilizing.
7054  *
7055  *    This makes distribution proportional, but also work-conserving:
7056  *    if one cgroup claims much more protection than it uses memory,
7057  *    the unused remainder is available to its siblings.
7058  *
7059  * 4. Conversely, when the declared protection is undercommitted at a
7060  *    given level, the distribution of the larger parental protection
7061  *    budget is NOT proportional. A cgroup's protection from a sibling
7062  *    is capped to its own memory.min/low setting.
7063  *
7064  * 5. However, to allow protecting recursive subtrees from each other
7065  *    without having to declare each individual cgroup's fixed share
7066  *    of the ancestor's claim to protection, any unutilized -
7067  *    "floating" - protection from up the tree is distributed in
7068  *    proportion to each cgroup's *usage*. This makes the protection
7069  *    neutral wrt sibling cgroups and lets them compete freely over
7070  *    the shared parental protection budget, but it protects the
7071  *    subtree as a whole from neighboring subtrees.
7072  *
7073  * Note that 4. and 5. are not in conflict: 4. is about protecting
7074  * against immediate siblings whereas 5. is about protecting against
7075  * neighboring subtrees.
7076  */
7077 static unsigned long effective_protection(unsigned long usage,
7078 					  unsigned long parent_usage,
7079 					  unsigned long setting,
7080 					  unsigned long parent_effective,
7081 					  unsigned long siblings_protected)
7082 {
7083 	unsigned long protected;
7084 	unsigned long ep;
7085 
7086 	protected = min(usage, setting);
7087 	/*
7088 	 * If all cgroups at this level combined claim and use more
7089 	 * protection than what the parent affords them, distribute
7090 	 * shares in proportion to utilization.
7091 	 *
7092 	 * We are using actual utilization rather than the statically
7093 	 * claimed protection in order to be work-conserving: claimed
7094 	 * but unused protection is available to siblings that would
7095 	 * otherwise get a smaller chunk than what they claimed.
7096 	 */
7097 	if (siblings_protected > parent_effective)
7098 		return protected * parent_effective / siblings_protected;
7099 
7100 	/*
7101 	 * Ok, utilized protection of all children is within what the
7102 	 * parent affords them, so we know whatever this child claims
7103 	 * and utilizes is effectively protected.
7104 	 *
7105 	 * If there is unprotected usage beyond this value, reclaim
7106 	 * will apply pressure in proportion to that amount.
7107 	 *
7108 	 * If there is unutilized protection, the cgroup will be fully
7109 	 * shielded from reclaim, but we do return a smaller value for
7110 	 * protection than what the group could enjoy in theory. This
7111 	 * is okay. With the overcommit distribution above, effective
7112 	 * protection is always dependent on how memory is actually
7113 	 * consumed among the siblings anyway.
7114 	 */
7115 	ep = protected;
7116 
7117 	/*
7118 	 * If the children aren't claiming (all of) the protection
7119 	 * afforded to them by the parent, distribute the remainder in
7120 	 * proportion to the (unprotected) memory of each cgroup. That
7121 	 * way, cgroups that aren't explicitly prioritized wrt each
7122 	 * other compete freely over the allowance, but they are
7123 	 * collectively protected from neighboring trees.
7124 	 *
7125 	 * We're using unprotected memory for the weight so that if
7126 	 * some cgroups DO claim explicit protection, we don't protect
7127 	 * the same bytes twice.
7128 	 *
7129 	 * Check both usage and parent_usage against the respective
7130 	 * protected values. One should imply the other, but they
7131 	 * aren't read atomically - make sure the division is sane.
7132 	 */
7133 	if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
7134 		return ep;
7135 	if (parent_effective > siblings_protected &&
7136 	    parent_usage > siblings_protected &&
7137 	    usage > protected) {
7138 		unsigned long unclaimed;
7139 
7140 		unclaimed = parent_effective - siblings_protected;
7141 		unclaimed *= usage - protected;
7142 		unclaimed /= parent_usage - siblings_protected;
7143 
7144 		ep += unclaimed;
7145 	}
7146 
7147 	return ep;
7148 }
7149 
7150 /**
7151  * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
7152  * @root: the top ancestor of the sub-tree being checked
7153  * @memcg: the memory cgroup to check
7154  *
7155  * WARNING: This function is not stateless! It can only be used as part
7156  *          of a top-down tree iteration, not for isolated queries.
7157  */
7158 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
7159 				     struct mem_cgroup *memcg)
7160 {
7161 	unsigned long usage, parent_usage;
7162 	struct mem_cgroup *parent;
7163 
7164 	if (mem_cgroup_disabled())
7165 		return;
7166 
7167 	if (!root)
7168 		root = root_mem_cgroup;
7169 
7170 	/*
7171 	 * Effective values of the reclaim targets are ignored so they
7172 	 * can be stale. Have a look at mem_cgroup_protection for more
7173 	 * details.
7174 	 * TODO: calculation should be more robust so that we do not need
7175 	 * that special casing.
7176 	 */
7177 	if (memcg == root)
7178 		return;
7179 
7180 	usage = page_counter_read(&memcg->memory);
7181 	if (!usage)
7182 		return;
7183 
7184 	parent = parent_mem_cgroup(memcg);
7185 
7186 	if (parent == root) {
7187 		memcg->memory.emin = READ_ONCE(memcg->memory.min);
7188 		memcg->memory.elow = READ_ONCE(memcg->memory.low);
7189 		return;
7190 	}
7191 
7192 	parent_usage = page_counter_read(&parent->memory);
7193 
7194 	WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
7195 			READ_ONCE(memcg->memory.min),
7196 			READ_ONCE(parent->memory.emin),
7197 			atomic_long_read(&parent->memory.children_min_usage)));
7198 
7199 	WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
7200 			READ_ONCE(memcg->memory.low),
7201 			READ_ONCE(parent->memory.elow),
7202 			atomic_long_read(&parent->memory.children_low_usage)));
7203 }
7204 
7205 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
7206 			gfp_t gfp)
7207 {
7208 	int ret;
7209 
7210 	ret = try_charge(memcg, gfp, folio_nr_pages(folio));
7211 	if (ret)
7212 		goto out;
7213 
7214 	mem_cgroup_commit_charge(folio, memcg);
7215 out:
7216 	return ret;
7217 }
7218 
7219 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
7220 {
7221 	struct mem_cgroup *memcg;
7222 	int ret;
7223 
7224 	memcg = get_mem_cgroup_from_mm(mm);
7225 	ret = charge_memcg(folio, memcg, gfp);
7226 	css_put(&memcg->css);
7227 
7228 	return ret;
7229 }
7230 
7231 /**
7232  * mem_cgroup_hugetlb_try_charge - try to charge the memcg for a hugetlb folio
7233  * @memcg: memcg to charge.
7234  * @gfp: reclaim mode.
7235  * @nr_pages: number of pages to charge.
7236  *
7237  * This function is called when allocating a huge page folio to determine if
7238  * the memcg has the capacity for it. It does not commit the charge yet,
7239  * as the hugetlb folio itself has not been obtained from the hugetlb pool.
7240  *
7241  * Once we have obtained the hugetlb folio, we can call
7242  * mem_cgroup_commit_charge() to commit the charge. If we fail to obtain the
7243  * folio, we should instead call mem_cgroup_cancel_charge() to undo the effect
7244  * of try_charge().
7245  *
7246  * Returns 0 on success. Otherwise, an error code is returned.
7247  */
7248 int mem_cgroup_hugetlb_try_charge(struct mem_cgroup *memcg, gfp_t gfp,
7249 			long nr_pages)
7250 {
7251 	/*
7252 	 * If hugetlb memcg charging is not enabled, do not fail hugetlb allocation,
7253 	 * but do not attempt to commit charge later (or cancel on error) either.
7254 	 */
7255 	if (mem_cgroup_disabled() || !memcg ||
7256 		!cgroup_subsys_on_dfl(memory_cgrp_subsys) ||
7257 		!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING))
7258 		return -EOPNOTSUPP;
7259 
7260 	if (try_charge(memcg, gfp, nr_pages))
7261 		return -ENOMEM;
7262 
7263 	return 0;
7264 }
7265 
7266 /**
7267  * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
7268  * @folio: folio to charge.
7269  * @mm: mm context of the victim
7270  * @gfp: reclaim mode
7271  * @entry: swap entry for which the folio is allocated
7272  *
7273  * This function charges a folio allocated for swapin. Please call this before
7274  * adding the folio to the swapcache.
7275  *
7276  * Returns 0 on success. Otherwise, an error code is returned.
7277  */
7278 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
7279 				  gfp_t gfp, swp_entry_t entry)
7280 {
7281 	struct mem_cgroup *memcg;
7282 	unsigned short id;
7283 	int ret;
7284 
7285 	if (mem_cgroup_disabled())
7286 		return 0;
7287 
7288 	id = lookup_swap_cgroup_id(entry);
7289 	rcu_read_lock();
7290 	memcg = mem_cgroup_from_id(id);
7291 	if (!memcg || !css_tryget_online(&memcg->css))
7292 		memcg = get_mem_cgroup_from_mm(mm);
7293 	rcu_read_unlock();
7294 
7295 	ret = charge_memcg(folio, memcg, gfp);
7296 
7297 	css_put(&memcg->css);
7298 	return ret;
7299 }
7300 
7301 /*
7302  * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
7303  * @entry: swap entry for which the page is charged
7304  *
7305  * Call this function after successfully adding the charged page to swapcache.
7306  *
7307  * Note: This function assumes the page for which swap slot is being uncharged
7308  * is order 0 page.
7309  */
7310 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
7311 {
7312 	/*
7313 	 * Cgroup1's unified memory+swap counter has been charged with the
7314 	 * new swapcache page, finish the transfer by uncharging the swap
7315 	 * slot. The swap slot would also get uncharged when it dies, but
7316 	 * it can stick around indefinitely and we'd count the page twice
7317 	 * the entire time.
7318 	 *
7319 	 * Cgroup2 has separate resource counters for memory and swap,
7320 	 * so this is a non-issue here. Memory and swap charge lifetimes
7321 	 * correspond 1:1 to page and swap slot lifetimes: we charge the
7322 	 * page to memory here, and uncharge swap when the slot is freed.
7323 	 */
7324 	if (!mem_cgroup_disabled() && do_memsw_account()) {
7325 		/*
7326 		 * The swap entry might not get freed for a long time,
7327 		 * let's not wait for it.  The page already received a
7328 		 * memory+swap charge, drop the swap entry duplicate.
7329 		 */
7330 		mem_cgroup_uncharge_swap(entry, 1);
7331 	}
7332 }
7333 
7334 struct uncharge_gather {
7335 	struct mem_cgroup *memcg;
7336 	unsigned long nr_memory;
7337 	unsigned long pgpgout;
7338 	unsigned long nr_kmem;
7339 	int nid;
7340 };
7341 
7342 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
7343 {
7344 	memset(ug, 0, sizeof(*ug));
7345 }
7346 
7347 static void uncharge_batch(const struct uncharge_gather *ug)
7348 {
7349 	unsigned long flags;
7350 
7351 	if (ug->nr_memory) {
7352 		page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
7353 		if (do_memsw_account())
7354 			page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
7355 		if (ug->nr_kmem)
7356 			memcg_account_kmem(ug->memcg, -ug->nr_kmem);
7357 		memcg_oom_recover(ug->memcg);
7358 	}
7359 
7360 	local_irq_save(flags);
7361 	__count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
7362 	__this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
7363 	memcg_check_events(ug->memcg, ug->nid);
7364 	local_irq_restore(flags);
7365 
7366 	/* drop reference from uncharge_folio */
7367 	css_put(&ug->memcg->css);
7368 }
7369 
7370 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
7371 {
7372 	long nr_pages;
7373 	struct mem_cgroup *memcg;
7374 	struct obj_cgroup *objcg;
7375 
7376 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7377 
7378 	/*
7379 	 * Nobody should be changing or seriously looking at
7380 	 * folio memcg or objcg at this point, we have fully
7381 	 * exclusive access to the folio.
7382 	 */
7383 	if (folio_memcg_kmem(folio)) {
7384 		objcg = __folio_objcg(folio);
7385 		/*
7386 		 * This get matches the put at the end of the function and
7387 		 * kmem pages do not hold memcg references anymore.
7388 		 */
7389 		memcg = get_mem_cgroup_from_objcg(objcg);
7390 	} else {
7391 		memcg = __folio_memcg(folio);
7392 	}
7393 
7394 	if (!memcg)
7395 		return;
7396 
7397 	if (ug->memcg != memcg) {
7398 		if (ug->memcg) {
7399 			uncharge_batch(ug);
7400 			uncharge_gather_clear(ug);
7401 		}
7402 		ug->memcg = memcg;
7403 		ug->nid = folio_nid(folio);
7404 
7405 		/* pairs with css_put in uncharge_batch */
7406 		css_get(&memcg->css);
7407 	}
7408 
7409 	nr_pages = folio_nr_pages(folio);
7410 
7411 	if (folio_memcg_kmem(folio)) {
7412 		ug->nr_memory += nr_pages;
7413 		ug->nr_kmem += nr_pages;
7414 
7415 		folio->memcg_data = 0;
7416 		obj_cgroup_put(objcg);
7417 	} else {
7418 		/* LRU pages aren't accounted at the root level */
7419 		if (!mem_cgroup_is_root(memcg))
7420 			ug->nr_memory += nr_pages;
7421 		ug->pgpgout++;
7422 
7423 		folio->memcg_data = 0;
7424 	}
7425 
7426 	css_put(&memcg->css);
7427 }
7428 
7429 void __mem_cgroup_uncharge(struct folio *folio)
7430 {
7431 	struct uncharge_gather ug;
7432 
7433 	/* Don't touch folio->lru of any random page, pre-check: */
7434 	if (!folio_memcg(folio))
7435 		return;
7436 
7437 	uncharge_gather_clear(&ug);
7438 	uncharge_folio(folio, &ug);
7439 	uncharge_batch(&ug);
7440 }
7441 
7442 /**
7443  * __mem_cgroup_uncharge_list - uncharge a list of page
7444  * @page_list: list of pages to uncharge
7445  *
7446  * Uncharge a list of pages previously charged with
7447  * __mem_cgroup_charge().
7448  */
7449 void __mem_cgroup_uncharge_list(struct list_head *page_list)
7450 {
7451 	struct uncharge_gather ug;
7452 	struct folio *folio;
7453 
7454 	uncharge_gather_clear(&ug);
7455 	list_for_each_entry(folio, page_list, lru)
7456 		uncharge_folio(folio, &ug);
7457 	if (ug.memcg)
7458 		uncharge_batch(&ug);
7459 }
7460 
7461 /**
7462  * mem_cgroup_replace_folio - Charge a folio's replacement.
7463  * @old: Currently circulating folio.
7464  * @new: Replacement folio.
7465  *
7466  * Charge @new as a replacement folio for @old. @old will
7467  * be uncharged upon free. This is only used by the page cache
7468  * (in replace_page_cache_folio()).
7469  *
7470  * Both folios must be locked, @new->mapping must be set up.
7471  */
7472 void mem_cgroup_replace_folio(struct folio *old, struct folio *new)
7473 {
7474 	struct mem_cgroup *memcg;
7475 	long nr_pages = folio_nr_pages(new);
7476 	unsigned long flags;
7477 
7478 	VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7479 	VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7480 	VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7481 	VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
7482 
7483 	if (mem_cgroup_disabled())
7484 		return;
7485 
7486 	/* Page cache replacement: new folio already charged? */
7487 	if (folio_memcg(new))
7488 		return;
7489 
7490 	memcg = folio_memcg(old);
7491 	VM_WARN_ON_ONCE_FOLIO(!memcg, old);
7492 	if (!memcg)
7493 		return;
7494 
7495 	/* Force-charge the new page. The old one will be freed soon */
7496 	if (!mem_cgroup_is_root(memcg)) {
7497 		page_counter_charge(&memcg->memory, nr_pages);
7498 		if (do_memsw_account())
7499 			page_counter_charge(&memcg->memsw, nr_pages);
7500 	}
7501 
7502 	css_get(&memcg->css);
7503 	commit_charge(new, memcg);
7504 
7505 	local_irq_save(flags);
7506 	mem_cgroup_charge_statistics(memcg, nr_pages);
7507 	memcg_check_events(memcg, folio_nid(new));
7508 	local_irq_restore(flags);
7509 }
7510 
7511 /**
7512  * mem_cgroup_migrate - Transfer the memcg data from the old to the new folio.
7513  * @old: Currently circulating folio.
7514  * @new: Replacement folio.
7515  *
7516  * Transfer the memcg data from the old folio to the new folio for migration.
7517  * The old folio's data info will be cleared. Note that the memory counters
7518  * will remain unchanged throughout the process.
7519  *
7520  * Both folios must be locked, @new->mapping must be set up.
7521  */
7522 void mem_cgroup_migrate(struct folio *old, struct folio *new)
7523 {
7524 	struct mem_cgroup *memcg;
7525 
7526 	VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7527 	VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7528 	VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7529 	VM_BUG_ON_FOLIO(folio_nr_pages(old) != folio_nr_pages(new), new);
7530 
7531 	if (mem_cgroup_disabled())
7532 		return;
7533 
7534 	memcg = folio_memcg(old);
7535 	/*
7536 	 * Note that it is normal to see !memcg for a hugetlb folio.
7537 	 * For e.g, itt could have been allocated when memory_hugetlb_accounting
7538 	 * was not selected.
7539 	 */
7540 	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(old) && !memcg, old);
7541 	if (!memcg)
7542 		return;
7543 
7544 	/* Transfer the charge and the css ref */
7545 	commit_charge(new, memcg);
7546 	old->memcg_data = 0;
7547 }
7548 
7549 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7550 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7551 
7552 void mem_cgroup_sk_alloc(struct sock *sk)
7553 {
7554 	struct mem_cgroup *memcg;
7555 
7556 	if (!mem_cgroup_sockets_enabled)
7557 		return;
7558 
7559 	/* Do not associate the sock with unrelated interrupted task's memcg. */
7560 	if (!in_task())
7561 		return;
7562 
7563 	rcu_read_lock();
7564 	memcg = mem_cgroup_from_task(current);
7565 	if (mem_cgroup_is_root(memcg))
7566 		goto out;
7567 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7568 		goto out;
7569 	if (css_tryget(&memcg->css))
7570 		sk->sk_memcg = memcg;
7571 out:
7572 	rcu_read_unlock();
7573 }
7574 
7575 void mem_cgroup_sk_free(struct sock *sk)
7576 {
7577 	if (sk->sk_memcg)
7578 		css_put(&sk->sk_memcg->css);
7579 }
7580 
7581 /**
7582  * mem_cgroup_charge_skmem - charge socket memory
7583  * @memcg: memcg to charge
7584  * @nr_pages: number of pages to charge
7585  * @gfp_mask: reclaim mode
7586  *
7587  * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7588  * @memcg's configured limit, %false if it doesn't.
7589  */
7590 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7591 			     gfp_t gfp_mask)
7592 {
7593 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7594 		struct page_counter *fail;
7595 
7596 		if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7597 			memcg->tcpmem_pressure = 0;
7598 			return true;
7599 		}
7600 		memcg->tcpmem_pressure = 1;
7601 		if (gfp_mask & __GFP_NOFAIL) {
7602 			page_counter_charge(&memcg->tcpmem, nr_pages);
7603 			return true;
7604 		}
7605 		return false;
7606 	}
7607 
7608 	if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7609 		mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7610 		return true;
7611 	}
7612 
7613 	return false;
7614 }
7615 
7616 /**
7617  * mem_cgroup_uncharge_skmem - uncharge socket memory
7618  * @memcg: memcg to uncharge
7619  * @nr_pages: number of pages to uncharge
7620  */
7621 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7622 {
7623 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7624 		page_counter_uncharge(&memcg->tcpmem, nr_pages);
7625 		return;
7626 	}
7627 
7628 	mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7629 
7630 	refill_stock(memcg, nr_pages);
7631 }
7632 
7633 static int __init cgroup_memory(char *s)
7634 {
7635 	char *token;
7636 
7637 	while ((token = strsep(&s, ",")) != NULL) {
7638 		if (!*token)
7639 			continue;
7640 		if (!strcmp(token, "nosocket"))
7641 			cgroup_memory_nosocket = true;
7642 		if (!strcmp(token, "nokmem"))
7643 			cgroup_memory_nokmem = true;
7644 		if (!strcmp(token, "nobpf"))
7645 			cgroup_memory_nobpf = true;
7646 	}
7647 	return 1;
7648 }
7649 __setup("cgroup.memory=", cgroup_memory);
7650 
7651 /*
7652  * subsys_initcall() for memory controller.
7653  *
7654  * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7655  * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7656  * basically everything that doesn't depend on a specific mem_cgroup structure
7657  * should be initialized from here.
7658  */
7659 static int __init mem_cgroup_init(void)
7660 {
7661 	int cpu, node;
7662 
7663 	/*
7664 	 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7665 	 * used for per-memcg-per-cpu caching of per-node statistics. In order
7666 	 * to work fine, we should make sure that the overfill threshold can't
7667 	 * exceed S32_MAX / PAGE_SIZE.
7668 	 */
7669 	BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7670 
7671 	cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7672 				  memcg_hotplug_cpu_dead);
7673 
7674 	for_each_possible_cpu(cpu)
7675 		INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7676 			  drain_local_stock);
7677 
7678 	for_each_node(node) {
7679 		struct mem_cgroup_tree_per_node *rtpn;
7680 
7681 		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, node);
7682 
7683 		rtpn->rb_root = RB_ROOT;
7684 		rtpn->rb_rightmost = NULL;
7685 		spin_lock_init(&rtpn->lock);
7686 		soft_limit_tree.rb_tree_per_node[node] = rtpn;
7687 	}
7688 
7689 	return 0;
7690 }
7691 subsys_initcall(mem_cgroup_init);
7692 
7693 #ifdef CONFIG_SWAP
7694 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7695 {
7696 	while (!refcount_inc_not_zero(&memcg->id.ref)) {
7697 		/*
7698 		 * The root cgroup cannot be destroyed, so it's refcount must
7699 		 * always be >= 1.
7700 		 */
7701 		if (WARN_ON_ONCE(mem_cgroup_is_root(memcg))) {
7702 			VM_BUG_ON(1);
7703 			break;
7704 		}
7705 		memcg = parent_mem_cgroup(memcg);
7706 		if (!memcg)
7707 			memcg = root_mem_cgroup;
7708 	}
7709 	return memcg;
7710 }
7711 
7712 /**
7713  * mem_cgroup_swapout - transfer a memsw charge to swap
7714  * @folio: folio whose memsw charge to transfer
7715  * @entry: swap entry to move the charge to
7716  *
7717  * Transfer the memsw charge of @folio to @entry.
7718  */
7719 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7720 {
7721 	struct mem_cgroup *memcg, *swap_memcg;
7722 	unsigned int nr_entries;
7723 	unsigned short oldid;
7724 
7725 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7726 	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7727 
7728 	if (mem_cgroup_disabled())
7729 		return;
7730 
7731 	if (!do_memsw_account())
7732 		return;
7733 
7734 	memcg = folio_memcg(folio);
7735 
7736 	VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7737 	if (!memcg)
7738 		return;
7739 
7740 	/*
7741 	 * In case the memcg owning these pages has been offlined and doesn't
7742 	 * have an ID allocated to it anymore, charge the closest online
7743 	 * ancestor for the swap instead and transfer the memory+swap charge.
7744 	 */
7745 	swap_memcg = mem_cgroup_id_get_online(memcg);
7746 	nr_entries = folio_nr_pages(folio);
7747 	/* Get references for the tail pages, too */
7748 	if (nr_entries > 1)
7749 		mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7750 	oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7751 				   nr_entries);
7752 	VM_BUG_ON_FOLIO(oldid, folio);
7753 	mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7754 
7755 	folio->memcg_data = 0;
7756 
7757 	if (!mem_cgroup_is_root(memcg))
7758 		page_counter_uncharge(&memcg->memory, nr_entries);
7759 
7760 	if (memcg != swap_memcg) {
7761 		if (!mem_cgroup_is_root(swap_memcg))
7762 			page_counter_charge(&swap_memcg->memsw, nr_entries);
7763 		page_counter_uncharge(&memcg->memsw, nr_entries);
7764 	}
7765 
7766 	/*
7767 	 * Interrupts should be disabled here because the caller holds the
7768 	 * i_pages lock which is taken with interrupts-off. It is
7769 	 * important here to have the interrupts disabled because it is the
7770 	 * only synchronisation we have for updating the per-CPU variables.
7771 	 */
7772 	memcg_stats_lock();
7773 	mem_cgroup_charge_statistics(memcg, -nr_entries);
7774 	memcg_stats_unlock();
7775 	memcg_check_events(memcg, folio_nid(folio));
7776 
7777 	css_put(&memcg->css);
7778 }
7779 
7780 /**
7781  * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7782  * @folio: folio being added to swap
7783  * @entry: swap entry to charge
7784  *
7785  * Try to charge @folio's memcg for the swap space at @entry.
7786  *
7787  * Returns 0 on success, -ENOMEM on failure.
7788  */
7789 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7790 {
7791 	unsigned int nr_pages = folio_nr_pages(folio);
7792 	struct page_counter *counter;
7793 	struct mem_cgroup *memcg;
7794 	unsigned short oldid;
7795 
7796 	if (do_memsw_account())
7797 		return 0;
7798 
7799 	memcg = folio_memcg(folio);
7800 
7801 	VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7802 	if (!memcg)
7803 		return 0;
7804 
7805 	if (!entry.val) {
7806 		memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7807 		return 0;
7808 	}
7809 
7810 	memcg = mem_cgroup_id_get_online(memcg);
7811 
7812 	if (!mem_cgroup_is_root(memcg) &&
7813 	    !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7814 		memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7815 		memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7816 		mem_cgroup_id_put(memcg);
7817 		return -ENOMEM;
7818 	}
7819 
7820 	/* Get references for the tail pages, too */
7821 	if (nr_pages > 1)
7822 		mem_cgroup_id_get_many(memcg, nr_pages - 1);
7823 	oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7824 	VM_BUG_ON_FOLIO(oldid, folio);
7825 	mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7826 
7827 	return 0;
7828 }
7829 
7830 /**
7831  * __mem_cgroup_uncharge_swap - uncharge swap space
7832  * @entry: swap entry to uncharge
7833  * @nr_pages: the amount of swap space to uncharge
7834  */
7835 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7836 {
7837 	struct mem_cgroup *memcg;
7838 	unsigned short id;
7839 
7840 	id = swap_cgroup_record(entry, 0, nr_pages);
7841 	rcu_read_lock();
7842 	memcg = mem_cgroup_from_id(id);
7843 	if (memcg) {
7844 		if (!mem_cgroup_is_root(memcg)) {
7845 			if (do_memsw_account())
7846 				page_counter_uncharge(&memcg->memsw, nr_pages);
7847 			else
7848 				page_counter_uncharge(&memcg->swap, nr_pages);
7849 		}
7850 		mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7851 		mem_cgroup_id_put_many(memcg, nr_pages);
7852 	}
7853 	rcu_read_unlock();
7854 }
7855 
7856 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7857 {
7858 	long nr_swap_pages = get_nr_swap_pages();
7859 
7860 	if (mem_cgroup_disabled() || do_memsw_account())
7861 		return nr_swap_pages;
7862 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg))
7863 		nr_swap_pages = min_t(long, nr_swap_pages,
7864 				      READ_ONCE(memcg->swap.max) -
7865 				      page_counter_read(&memcg->swap));
7866 	return nr_swap_pages;
7867 }
7868 
7869 bool mem_cgroup_swap_full(struct folio *folio)
7870 {
7871 	struct mem_cgroup *memcg;
7872 
7873 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
7874 
7875 	if (vm_swap_full())
7876 		return true;
7877 	if (do_memsw_account())
7878 		return false;
7879 
7880 	memcg = folio_memcg(folio);
7881 	if (!memcg)
7882 		return false;
7883 
7884 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
7885 		unsigned long usage = page_counter_read(&memcg->swap);
7886 
7887 		if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7888 		    usage * 2 >= READ_ONCE(memcg->swap.max))
7889 			return true;
7890 	}
7891 
7892 	return false;
7893 }
7894 
7895 static int __init setup_swap_account(char *s)
7896 {
7897 	pr_warn_once("The swapaccount= commandline option is deprecated. "
7898 		     "Please report your usecase to linux-mm@kvack.org if you "
7899 		     "depend on this functionality.\n");
7900 	return 1;
7901 }
7902 __setup("swapaccount=", setup_swap_account);
7903 
7904 static u64 swap_current_read(struct cgroup_subsys_state *css,
7905 			     struct cftype *cft)
7906 {
7907 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7908 
7909 	return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7910 }
7911 
7912 static u64 swap_peak_read(struct cgroup_subsys_state *css,
7913 			  struct cftype *cft)
7914 {
7915 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7916 
7917 	return (u64)memcg->swap.watermark * PAGE_SIZE;
7918 }
7919 
7920 static int swap_high_show(struct seq_file *m, void *v)
7921 {
7922 	return seq_puts_memcg_tunable(m,
7923 		READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7924 }
7925 
7926 static ssize_t swap_high_write(struct kernfs_open_file *of,
7927 			       char *buf, size_t nbytes, loff_t off)
7928 {
7929 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7930 	unsigned long high;
7931 	int err;
7932 
7933 	buf = strstrip(buf);
7934 	err = page_counter_memparse(buf, "max", &high);
7935 	if (err)
7936 		return err;
7937 
7938 	page_counter_set_high(&memcg->swap, high);
7939 
7940 	return nbytes;
7941 }
7942 
7943 static int swap_max_show(struct seq_file *m, void *v)
7944 {
7945 	return seq_puts_memcg_tunable(m,
7946 		READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7947 }
7948 
7949 static ssize_t swap_max_write(struct kernfs_open_file *of,
7950 			      char *buf, size_t nbytes, loff_t off)
7951 {
7952 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7953 	unsigned long max;
7954 	int err;
7955 
7956 	buf = strstrip(buf);
7957 	err = page_counter_memparse(buf, "max", &max);
7958 	if (err)
7959 		return err;
7960 
7961 	xchg(&memcg->swap.max, max);
7962 
7963 	return nbytes;
7964 }
7965 
7966 static int swap_events_show(struct seq_file *m, void *v)
7967 {
7968 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7969 
7970 	seq_printf(m, "high %lu\n",
7971 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7972 	seq_printf(m, "max %lu\n",
7973 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7974 	seq_printf(m, "fail %lu\n",
7975 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7976 
7977 	return 0;
7978 }
7979 
7980 static struct cftype swap_files[] = {
7981 	{
7982 		.name = "swap.current",
7983 		.flags = CFTYPE_NOT_ON_ROOT,
7984 		.read_u64 = swap_current_read,
7985 	},
7986 	{
7987 		.name = "swap.high",
7988 		.flags = CFTYPE_NOT_ON_ROOT,
7989 		.seq_show = swap_high_show,
7990 		.write = swap_high_write,
7991 	},
7992 	{
7993 		.name = "swap.max",
7994 		.flags = CFTYPE_NOT_ON_ROOT,
7995 		.seq_show = swap_max_show,
7996 		.write = swap_max_write,
7997 	},
7998 	{
7999 		.name = "swap.peak",
8000 		.flags = CFTYPE_NOT_ON_ROOT,
8001 		.read_u64 = swap_peak_read,
8002 	},
8003 	{
8004 		.name = "swap.events",
8005 		.flags = CFTYPE_NOT_ON_ROOT,
8006 		.file_offset = offsetof(struct mem_cgroup, swap_events_file),
8007 		.seq_show = swap_events_show,
8008 	},
8009 	{ }	/* terminate */
8010 };
8011 
8012 static struct cftype memsw_files[] = {
8013 	{
8014 		.name = "memsw.usage_in_bytes",
8015 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
8016 		.read_u64 = mem_cgroup_read_u64,
8017 	},
8018 	{
8019 		.name = "memsw.max_usage_in_bytes",
8020 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
8021 		.write = mem_cgroup_reset,
8022 		.read_u64 = mem_cgroup_read_u64,
8023 	},
8024 	{
8025 		.name = "memsw.limit_in_bytes",
8026 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
8027 		.write = mem_cgroup_write,
8028 		.read_u64 = mem_cgroup_read_u64,
8029 	},
8030 	{
8031 		.name = "memsw.failcnt",
8032 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
8033 		.write = mem_cgroup_reset,
8034 		.read_u64 = mem_cgroup_read_u64,
8035 	},
8036 	{ },	/* terminate */
8037 };
8038 
8039 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8040 /**
8041  * obj_cgroup_may_zswap - check if this cgroup can zswap
8042  * @objcg: the object cgroup
8043  *
8044  * Check if the hierarchical zswap limit has been reached.
8045  *
8046  * This doesn't check for specific headroom, and it is not atomic
8047  * either. But with zswap, the size of the allocation is only known
8048  * once compression has occurred, and this optimistic pre-check avoids
8049  * spending cycles on compression when there is already no room left
8050  * or zswap is disabled altogether somewhere in the hierarchy.
8051  */
8052 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
8053 {
8054 	struct mem_cgroup *memcg, *original_memcg;
8055 	bool ret = true;
8056 
8057 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8058 		return true;
8059 
8060 	original_memcg = get_mem_cgroup_from_objcg(objcg);
8061 	for (memcg = original_memcg; !mem_cgroup_is_root(memcg);
8062 	     memcg = parent_mem_cgroup(memcg)) {
8063 		unsigned long max = READ_ONCE(memcg->zswap_max);
8064 		unsigned long pages;
8065 
8066 		if (max == PAGE_COUNTER_MAX)
8067 			continue;
8068 		if (max == 0) {
8069 			ret = false;
8070 			break;
8071 		}
8072 
8073 		cgroup_rstat_flush(memcg->css.cgroup);
8074 		pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
8075 		if (pages < max)
8076 			continue;
8077 		ret = false;
8078 		break;
8079 	}
8080 	mem_cgroup_put(original_memcg);
8081 	return ret;
8082 }
8083 
8084 /**
8085  * obj_cgroup_charge_zswap - charge compression backend memory
8086  * @objcg: the object cgroup
8087  * @size: size of compressed object
8088  *
8089  * This forces the charge after obj_cgroup_may_zswap() allowed
8090  * compression and storage in zwap for this cgroup to go ahead.
8091  */
8092 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
8093 {
8094 	struct mem_cgroup *memcg;
8095 
8096 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8097 		return;
8098 
8099 	VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
8100 
8101 	/* PF_MEMALLOC context, charging must succeed */
8102 	if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
8103 		VM_WARN_ON_ONCE(1);
8104 
8105 	rcu_read_lock();
8106 	memcg = obj_cgroup_memcg(objcg);
8107 	mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
8108 	mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
8109 	rcu_read_unlock();
8110 }
8111 
8112 /**
8113  * obj_cgroup_uncharge_zswap - uncharge compression backend memory
8114  * @objcg: the object cgroup
8115  * @size: size of compressed object
8116  *
8117  * Uncharges zswap memory on page in.
8118  */
8119 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
8120 {
8121 	struct mem_cgroup *memcg;
8122 
8123 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8124 		return;
8125 
8126 	obj_cgroup_uncharge(objcg, size);
8127 
8128 	rcu_read_lock();
8129 	memcg = obj_cgroup_memcg(objcg);
8130 	mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
8131 	mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
8132 	rcu_read_unlock();
8133 }
8134 
8135 static u64 zswap_current_read(struct cgroup_subsys_state *css,
8136 			      struct cftype *cft)
8137 {
8138 	cgroup_rstat_flush(css->cgroup);
8139 	return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
8140 }
8141 
8142 static int zswap_max_show(struct seq_file *m, void *v)
8143 {
8144 	return seq_puts_memcg_tunable(m,
8145 		READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
8146 }
8147 
8148 static ssize_t zswap_max_write(struct kernfs_open_file *of,
8149 			       char *buf, size_t nbytes, loff_t off)
8150 {
8151 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8152 	unsigned long max;
8153 	int err;
8154 
8155 	buf = strstrip(buf);
8156 	err = page_counter_memparse(buf, "max", &max);
8157 	if (err)
8158 		return err;
8159 
8160 	xchg(&memcg->zswap_max, max);
8161 
8162 	return nbytes;
8163 }
8164 
8165 static struct cftype zswap_files[] = {
8166 	{
8167 		.name = "zswap.current",
8168 		.flags = CFTYPE_NOT_ON_ROOT,
8169 		.read_u64 = zswap_current_read,
8170 	},
8171 	{
8172 		.name = "zswap.max",
8173 		.flags = CFTYPE_NOT_ON_ROOT,
8174 		.seq_show = zswap_max_show,
8175 		.write = zswap_max_write,
8176 	},
8177 	{ }	/* terminate */
8178 };
8179 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
8180 
8181 static int __init mem_cgroup_swap_init(void)
8182 {
8183 	if (mem_cgroup_disabled())
8184 		return 0;
8185 
8186 	WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
8187 	WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
8188 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8189 	WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
8190 #endif
8191 	return 0;
8192 }
8193 subsys_initcall(mem_cgroup_swap_init);
8194 
8195 #endif /* CONFIG_SWAP */
8196