xref: /linux/mm/memcontrol.c (revision ea518afc992032f7570c0a89ac9240b387dc0faf)
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 <linux/kmemleak.h>
68 #include "internal.h"
69 #include <net/sock.h>
70 #include <net/ip.h>
71 #include "slab.h"
72 #include "swap.h"
73 
74 #include <linux/uaccess.h>
75 
76 #include <trace/events/vmscan.h>
77 
78 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
79 EXPORT_SYMBOL(memory_cgrp_subsys);
80 
81 struct mem_cgroup *root_mem_cgroup __read_mostly;
82 
83 /* Active memory cgroup to use from an interrupt context */
84 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
85 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
86 
87 /* Socket memory accounting disabled? */
88 static bool cgroup_memory_nosocket __ro_after_init;
89 
90 /* Kernel memory accounting disabled? */
91 static bool cgroup_memory_nokmem __ro_after_init;
92 
93 /* BPF memory accounting disabled? */
94 static bool cgroup_memory_nobpf __ro_after_init;
95 
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
98 #endif
99 
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
102 {
103 	return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
104 }
105 
106 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
108 
109 /*
110  * Cgroups above their limits are maintained in a RB-Tree, independent of
111  * their hierarchy representation
112  */
113 
114 struct mem_cgroup_tree_per_node {
115 	struct rb_root rb_root;
116 	struct rb_node *rb_rightmost;
117 	spinlock_t lock;
118 };
119 
120 struct mem_cgroup_tree {
121 	struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
122 };
123 
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
125 
126 /* for OOM */
127 struct mem_cgroup_eventfd_list {
128 	struct list_head list;
129 	struct eventfd_ctx *eventfd;
130 };
131 
132 /*
133  * cgroup_event represents events which userspace want to receive.
134  */
135 struct mem_cgroup_event {
136 	/*
137 	 * memcg which the event belongs to.
138 	 */
139 	struct mem_cgroup *memcg;
140 	/*
141 	 * eventfd to signal userspace about the event.
142 	 */
143 	struct eventfd_ctx *eventfd;
144 	/*
145 	 * Each of these stored in a list by the cgroup.
146 	 */
147 	struct list_head list;
148 	/*
149 	 * register_event() callback will be used to add new userspace
150 	 * waiter for changes related to this event.  Use eventfd_signal()
151 	 * on eventfd to send notification to userspace.
152 	 */
153 	int (*register_event)(struct mem_cgroup *memcg,
154 			      struct eventfd_ctx *eventfd, const char *args);
155 	/*
156 	 * unregister_event() callback will be called when userspace closes
157 	 * the eventfd or on cgroup removing.  This callback must be set,
158 	 * if you want provide notification functionality.
159 	 */
160 	void (*unregister_event)(struct mem_cgroup *memcg,
161 				 struct eventfd_ctx *eventfd);
162 	/*
163 	 * All fields below needed to unregister event when
164 	 * userspace closes eventfd.
165 	 */
166 	poll_table pt;
167 	wait_queue_head_t *wqh;
168 	wait_queue_entry_t wait;
169 	struct work_struct remove;
170 };
171 
172 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
173 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 
175 /* Stuffs for move charges at task migration. */
176 /*
177  * Types of charges to be moved.
178  */
179 #define MOVE_ANON	0x1U
180 #define MOVE_FILE	0x2U
181 #define MOVE_MASK	(MOVE_ANON | MOVE_FILE)
182 
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct {
185 	spinlock_t	  lock; /* for from, to */
186 	struct mm_struct  *mm;
187 	struct mem_cgroup *from;
188 	struct mem_cgroup *to;
189 	unsigned long flags;
190 	unsigned long precharge;
191 	unsigned long moved_charge;
192 	unsigned long moved_swap;
193 	struct task_struct *moving_task;	/* a task moving charges */
194 	wait_queue_head_t waitq;		/* a waitq for other context */
195 } mc = {
196 	.lock = __SPIN_LOCK_UNLOCKED(mc.lock),
197 	.waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
198 };
199 
200 /*
201  * Maximum loops in mem_cgroup_soft_reclaim(), used for soft
202  * limit reclaim to prevent infinite loops, if they ever occur.
203  */
204 #define	MEM_CGROUP_MAX_RECLAIM_LOOPS		100
205 #define	MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS	2
206 
207 /* for encoding cft->private value on file */
208 enum res_type {
209 	_MEM,
210 	_MEMSWAP,
211 	_KMEM,
212 	_TCP,
213 };
214 
215 #define MEMFILE_PRIVATE(x, val)	((x) << 16 | (val))
216 #define MEMFILE_TYPE(val)	((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val)	((val) & 0xffff)
218 
219 /*
220  * Iteration constructs for visiting all cgroups (under a tree).  If
221  * loops are exited prematurely (break), mem_cgroup_iter_break() must
222  * be used for reference counting.
223  */
224 #define for_each_mem_cgroup_tree(iter, root)		\
225 	for (iter = mem_cgroup_iter(root, NULL, NULL);	\
226 	     iter != NULL;				\
227 	     iter = mem_cgroup_iter(root, iter, NULL))
228 
229 #define for_each_mem_cgroup(iter)			\
230 	for (iter = mem_cgroup_iter(NULL, NULL, NULL);	\
231 	     iter != NULL;				\
232 	     iter = mem_cgroup_iter(NULL, iter, NULL))
233 
234 static inline bool task_is_dying(void)
235 {
236 	return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
237 		(current->flags & PF_EXITING);
238 }
239 
240 /* Some nice accessors for the vmpressure. */
241 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
242 {
243 	if (!memcg)
244 		memcg = root_mem_cgroup;
245 	return &memcg->vmpressure;
246 }
247 
248 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
249 {
250 	return container_of(vmpr, struct mem_cgroup, vmpressure);
251 }
252 
253 #define CURRENT_OBJCG_UPDATE_BIT 0
254 #define CURRENT_OBJCG_UPDATE_FLAG (1UL << CURRENT_OBJCG_UPDATE_BIT)
255 
256 #ifdef CONFIG_MEMCG_KMEM
257 static DEFINE_SPINLOCK(objcg_lock);
258 
259 bool mem_cgroup_kmem_disabled(void)
260 {
261 	return cgroup_memory_nokmem;
262 }
263 
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
265 				      unsigned int nr_pages);
266 
267 static void obj_cgroup_release(struct percpu_ref *ref)
268 {
269 	struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
270 	unsigned int nr_bytes;
271 	unsigned int nr_pages;
272 	unsigned long flags;
273 
274 	/*
275 	 * At this point all allocated objects are freed, and
276 	 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 	 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
278 	 *
279 	 * The following sequence can lead to it:
280 	 * 1) CPU0: objcg == stock->cached_objcg
281 	 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 	 *          PAGE_SIZE bytes are charged
283 	 * 3) CPU1: a process from another memcg is allocating something,
284 	 *          the stock if flushed,
285 	 *          objcg->nr_charged_bytes = PAGE_SIZE - 92
286 	 * 5) CPU0: we do release this object,
287 	 *          92 bytes are added to stock->nr_bytes
288 	 * 6) CPU0: stock is flushed,
289 	 *          92 bytes are added to objcg->nr_charged_bytes
290 	 *
291 	 * In the result, nr_charged_bytes == PAGE_SIZE.
292 	 * This page will be uncharged in obj_cgroup_release().
293 	 */
294 	nr_bytes = atomic_read(&objcg->nr_charged_bytes);
295 	WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
296 	nr_pages = nr_bytes >> PAGE_SHIFT;
297 
298 	if (nr_pages)
299 		obj_cgroup_uncharge_pages(objcg, nr_pages);
300 
301 	spin_lock_irqsave(&objcg_lock, flags);
302 	list_del(&objcg->list);
303 	spin_unlock_irqrestore(&objcg_lock, flags);
304 
305 	percpu_ref_exit(ref);
306 	kfree_rcu(objcg, rcu);
307 }
308 
309 static struct obj_cgroup *obj_cgroup_alloc(void)
310 {
311 	struct obj_cgroup *objcg;
312 	int ret;
313 
314 	objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
315 	if (!objcg)
316 		return NULL;
317 
318 	ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
319 			      GFP_KERNEL);
320 	if (ret) {
321 		kfree(objcg);
322 		return NULL;
323 	}
324 	INIT_LIST_HEAD(&objcg->list);
325 	return objcg;
326 }
327 
328 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
329 				  struct mem_cgroup *parent)
330 {
331 	struct obj_cgroup *objcg, *iter;
332 
333 	objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
334 
335 	spin_lock_irq(&objcg_lock);
336 
337 	/* 1) Ready to reparent active objcg. */
338 	list_add(&objcg->list, &memcg->objcg_list);
339 	/* 2) Reparent active objcg and already reparented objcgs to parent. */
340 	list_for_each_entry(iter, &memcg->objcg_list, list)
341 		WRITE_ONCE(iter->memcg, parent);
342 	/* 3) Move already reparented objcgs to the parent's list */
343 	list_splice(&memcg->objcg_list, &parent->objcg_list);
344 
345 	spin_unlock_irq(&objcg_lock);
346 
347 	percpu_ref_kill(&objcg->refcnt);
348 }
349 
350 /*
351  * A lot of the calls to the cache allocation functions are expected to be
352  * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
353  * conditional to this static branch, we'll have to allow modules that does
354  * kmem_cache_alloc and the such to see this symbol as well
355  */
356 DEFINE_STATIC_KEY_FALSE(memcg_kmem_online_key);
357 EXPORT_SYMBOL(memcg_kmem_online_key);
358 
359 DEFINE_STATIC_KEY_FALSE(memcg_bpf_enabled_key);
360 EXPORT_SYMBOL(memcg_bpf_enabled_key);
361 #endif
362 
363 /**
364  * mem_cgroup_css_from_folio - css of the memcg associated with a folio
365  * @folio: folio of interest
366  *
367  * If memcg is bound to the default hierarchy, css of the memcg associated
368  * with @folio is returned.  The returned css remains associated with @folio
369  * until it is released.
370  *
371  * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
372  * is returned.
373  */
374 struct cgroup_subsys_state *mem_cgroup_css_from_folio(struct folio *folio)
375 {
376 	struct mem_cgroup *memcg = folio_memcg(folio);
377 
378 	if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
379 		memcg = root_mem_cgroup;
380 
381 	return &memcg->css;
382 }
383 
384 /**
385  * page_cgroup_ino - return inode number of the memcg a page is charged to
386  * @page: the page
387  *
388  * Look up the closest online ancestor of the memory cgroup @page is charged to
389  * and return its inode number or 0 if @page is not charged to any cgroup. It
390  * is safe to call this function without holding a reference to @page.
391  *
392  * Note, this function is inherently racy, because there is nothing to prevent
393  * the cgroup inode from getting torn down and potentially reallocated a moment
394  * after page_cgroup_ino() returns, so it only should be used by callers that
395  * do not care (such as procfs interfaces).
396  */
397 ino_t page_cgroup_ino(struct page *page)
398 {
399 	struct mem_cgroup *memcg;
400 	unsigned long ino = 0;
401 
402 	rcu_read_lock();
403 	/* page_folio() is racy here, but the entire function is racy anyway */
404 	memcg = folio_memcg_check(page_folio(page));
405 
406 	while (memcg && !(memcg->css.flags & CSS_ONLINE))
407 		memcg = parent_mem_cgroup(memcg);
408 	if (memcg)
409 		ino = cgroup_ino(memcg->css.cgroup);
410 	rcu_read_unlock();
411 	return ino;
412 }
413 
414 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
415 					 struct mem_cgroup_tree_per_node *mctz,
416 					 unsigned long new_usage_in_excess)
417 {
418 	struct rb_node **p = &mctz->rb_root.rb_node;
419 	struct rb_node *parent = NULL;
420 	struct mem_cgroup_per_node *mz_node;
421 	bool rightmost = true;
422 
423 	if (mz->on_tree)
424 		return;
425 
426 	mz->usage_in_excess = new_usage_in_excess;
427 	if (!mz->usage_in_excess)
428 		return;
429 	while (*p) {
430 		parent = *p;
431 		mz_node = rb_entry(parent, struct mem_cgroup_per_node,
432 					tree_node);
433 		if (mz->usage_in_excess < mz_node->usage_in_excess) {
434 			p = &(*p)->rb_left;
435 			rightmost = false;
436 		} else {
437 			p = &(*p)->rb_right;
438 		}
439 	}
440 
441 	if (rightmost)
442 		mctz->rb_rightmost = &mz->tree_node;
443 
444 	rb_link_node(&mz->tree_node, parent, p);
445 	rb_insert_color(&mz->tree_node, &mctz->rb_root);
446 	mz->on_tree = true;
447 }
448 
449 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
450 					 struct mem_cgroup_tree_per_node *mctz)
451 {
452 	if (!mz->on_tree)
453 		return;
454 
455 	if (&mz->tree_node == mctz->rb_rightmost)
456 		mctz->rb_rightmost = rb_prev(&mz->tree_node);
457 
458 	rb_erase(&mz->tree_node, &mctz->rb_root);
459 	mz->on_tree = false;
460 }
461 
462 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
463 				       struct mem_cgroup_tree_per_node *mctz)
464 {
465 	unsigned long flags;
466 
467 	spin_lock_irqsave(&mctz->lock, flags);
468 	__mem_cgroup_remove_exceeded(mz, mctz);
469 	spin_unlock_irqrestore(&mctz->lock, flags);
470 }
471 
472 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
473 {
474 	unsigned long nr_pages = page_counter_read(&memcg->memory);
475 	unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
476 	unsigned long excess = 0;
477 
478 	if (nr_pages > soft_limit)
479 		excess = nr_pages - soft_limit;
480 
481 	return excess;
482 }
483 
484 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
485 {
486 	unsigned long excess;
487 	struct mem_cgroup_per_node *mz;
488 	struct mem_cgroup_tree_per_node *mctz;
489 
490 	if (lru_gen_enabled()) {
491 		if (soft_limit_excess(memcg))
492 			lru_gen_soft_reclaim(memcg, nid);
493 		return;
494 	}
495 
496 	mctz = soft_limit_tree.rb_tree_per_node[nid];
497 	if (!mctz)
498 		return;
499 	/*
500 	 * Necessary to update all ancestors when hierarchy is used.
501 	 * because their event counter is not touched.
502 	 */
503 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
504 		mz = memcg->nodeinfo[nid];
505 		excess = soft_limit_excess(memcg);
506 		/*
507 		 * We have to update the tree if mz is on RB-tree or
508 		 * mem is over its softlimit.
509 		 */
510 		if (excess || mz->on_tree) {
511 			unsigned long flags;
512 
513 			spin_lock_irqsave(&mctz->lock, flags);
514 			/* if on-tree, remove it */
515 			if (mz->on_tree)
516 				__mem_cgroup_remove_exceeded(mz, mctz);
517 			/*
518 			 * Insert again. mz->usage_in_excess will be updated.
519 			 * If excess is 0, no tree ops.
520 			 */
521 			__mem_cgroup_insert_exceeded(mz, mctz, excess);
522 			spin_unlock_irqrestore(&mctz->lock, flags);
523 		}
524 	}
525 }
526 
527 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
528 {
529 	struct mem_cgroup_tree_per_node *mctz;
530 	struct mem_cgroup_per_node *mz;
531 	int nid;
532 
533 	for_each_node(nid) {
534 		mz = memcg->nodeinfo[nid];
535 		mctz = soft_limit_tree.rb_tree_per_node[nid];
536 		if (mctz)
537 			mem_cgroup_remove_exceeded(mz, mctz);
538 	}
539 }
540 
541 static struct mem_cgroup_per_node *
542 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
543 {
544 	struct mem_cgroup_per_node *mz;
545 
546 retry:
547 	mz = NULL;
548 	if (!mctz->rb_rightmost)
549 		goto done;		/* Nothing to reclaim from */
550 
551 	mz = rb_entry(mctz->rb_rightmost,
552 		      struct mem_cgroup_per_node, tree_node);
553 	/*
554 	 * Remove the node now but someone else can add it back,
555 	 * we will to add it back at the end of reclaim to its correct
556 	 * position in the tree.
557 	 */
558 	__mem_cgroup_remove_exceeded(mz, mctz);
559 	if (!soft_limit_excess(mz->memcg) ||
560 	    !css_tryget(&mz->memcg->css))
561 		goto retry;
562 done:
563 	return mz;
564 }
565 
566 static struct mem_cgroup_per_node *
567 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
568 {
569 	struct mem_cgroup_per_node *mz;
570 
571 	spin_lock_irq(&mctz->lock);
572 	mz = __mem_cgroup_largest_soft_limit_node(mctz);
573 	spin_unlock_irq(&mctz->lock);
574 	return mz;
575 }
576 
577 /* Subset of vm_event_item to report for memcg event stats */
578 static const unsigned int memcg_vm_event_stat[] = {
579 	PGPGIN,
580 	PGPGOUT,
581 	PGSCAN_KSWAPD,
582 	PGSCAN_DIRECT,
583 	PGSCAN_KHUGEPAGED,
584 	PGSTEAL_KSWAPD,
585 	PGSTEAL_DIRECT,
586 	PGSTEAL_KHUGEPAGED,
587 	PGFAULT,
588 	PGMAJFAULT,
589 	PGREFILL,
590 	PGACTIVATE,
591 	PGDEACTIVATE,
592 	PGLAZYFREE,
593 	PGLAZYFREED,
594 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
595 	ZSWPIN,
596 	ZSWPOUT,
597 	ZSWPWB,
598 #endif
599 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
600 	THP_FAULT_ALLOC,
601 	THP_COLLAPSE_ALLOC,
602 	THP_SWPOUT,
603 	THP_SWPOUT_FALLBACK,
604 #endif
605 };
606 
607 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
608 static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
609 
610 static void init_memcg_events(void)
611 {
612 	int i;
613 
614 	for (i = 0; i < NR_MEMCG_EVENTS; ++i)
615 		mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
616 }
617 
618 static inline int memcg_events_index(enum vm_event_item idx)
619 {
620 	return mem_cgroup_events_index[idx] - 1;
621 }
622 
623 struct memcg_vmstats_percpu {
624 	/* Local (CPU and cgroup) page state & events */
625 	long			state[MEMCG_NR_STAT];
626 	unsigned long		events[NR_MEMCG_EVENTS];
627 
628 	/* Delta calculation for lockless upward propagation */
629 	long			state_prev[MEMCG_NR_STAT];
630 	unsigned long		events_prev[NR_MEMCG_EVENTS];
631 
632 	/* Cgroup1: threshold notifications & softlimit tree updates */
633 	unsigned long		nr_page_events;
634 	unsigned long		targets[MEM_CGROUP_NTARGETS];
635 
636 	/* Stats updates since the last flush */
637 	unsigned int		stats_updates;
638 };
639 
640 struct memcg_vmstats {
641 	/* Aggregated (CPU and subtree) page state & events */
642 	long			state[MEMCG_NR_STAT];
643 	unsigned long		events[NR_MEMCG_EVENTS];
644 
645 	/* Non-hierarchical (CPU aggregated) page state & events */
646 	long			state_local[MEMCG_NR_STAT];
647 	unsigned long		events_local[NR_MEMCG_EVENTS];
648 
649 	/* Pending child counts during tree propagation */
650 	long			state_pending[MEMCG_NR_STAT];
651 	unsigned long		events_pending[NR_MEMCG_EVENTS];
652 
653 	/* Stats updates since the last flush */
654 	atomic64_t		stats_updates;
655 };
656 
657 /*
658  * memcg and lruvec stats flushing
659  *
660  * Many codepaths leading to stats update or read are performance sensitive and
661  * adding stats flushing in such codepaths is not desirable. So, to optimize the
662  * flushing the kernel does:
663  *
664  * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
665  *    rstat update tree grow unbounded.
666  *
667  * 2) Flush the stats synchronously on reader side only when there are more than
668  *    (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
669  *    will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
670  *    only for 2 seconds due to (1).
671  */
672 static void flush_memcg_stats_dwork(struct work_struct *w);
673 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
674 static u64 flush_last_time;
675 
676 #define FLUSH_TIME (2UL*HZ)
677 
678 /*
679  * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
680  * not rely on this as part of an acquired spinlock_t lock. These functions are
681  * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
682  * is sufficient.
683  */
684 static void memcg_stats_lock(void)
685 {
686 	preempt_disable_nested();
687 	VM_WARN_ON_IRQS_ENABLED();
688 }
689 
690 static void __memcg_stats_lock(void)
691 {
692 	preempt_disable_nested();
693 }
694 
695 static void memcg_stats_unlock(void)
696 {
697 	preempt_enable_nested();
698 }
699 
700 
701 static bool memcg_should_flush_stats(struct mem_cgroup *memcg)
702 {
703 	return atomic64_read(&memcg->vmstats->stats_updates) >
704 		MEMCG_CHARGE_BATCH * num_online_cpus();
705 }
706 
707 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
708 {
709 	int cpu = smp_processor_id();
710 	unsigned int x;
711 
712 	if (!val)
713 		return;
714 
715 	cgroup_rstat_updated(memcg->css.cgroup, cpu);
716 
717 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
718 		x = __this_cpu_add_return(memcg->vmstats_percpu->stats_updates,
719 					  abs(val));
720 
721 		if (x < MEMCG_CHARGE_BATCH)
722 			continue;
723 
724 		/*
725 		 * If @memcg is already flush-able, increasing stats_updates is
726 		 * redundant. Avoid the overhead of the atomic update.
727 		 */
728 		if (!memcg_should_flush_stats(memcg))
729 			atomic64_add(x, &memcg->vmstats->stats_updates);
730 		__this_cpu_write(memcg->vmstats_percpu->stats_updates, 0);
731 	}
732 }
733 
734 static void do_flush_stats(struct mem_cgroup *memcg)
735 {
736 	if (mem_cgroup_is_root(memcg))
737 		WRITE_ONCE(flush_last_time, jiffies_64);
738 
739 	cgroup_rstat_flush(memcg->css.cgroup);
740 }
741 
742 /*
743  * mem_cgroup_flush_stats - flush the stats of a memory cgroup subtree
744  * @memcg: root of the subtree to flush
745  *
746  * Flushing is serialized by the underlying global rstat lock. There is also a
747  * minimum amount of work to be done even if there are no stat updates to flush.
748  * Hence, we only flush the stats if the updates delta exceeds a threshold. This
749  * avoids unnecessary work and contention on the underlying lock.
750  */
751 void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
752 {
753 	if (mem_cgroup_disabled())
754 		return;
755 
756 	if (!memcg)
757 		memcg = root_mem_cgroup;
758 
759 	if (memcg_should_flush_stats(memcg))
760 		do_flush_stats(memcg);
761 }
762 
763 void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
764 {
765 	/* Only flush if the periodic flusher is one full cycle late */
766 	if (time_after64(jiffies_64, READ_ONCE(flush_last_time) + 2*FLUSH_TIME))
767 		mem_cgroup_flush_stats(memcg);
768 }
769 
770 static void flush_memcg_stats_dwork(struct work_struct *w)
771 {
772 	/*
773 	 * Deliberately ignore memcg_should_flush_stats() here so that flushing
774 	 * in latency-sensitive paths is as cheap as possible.
775 	 */
776 	do_flush_stats(root_mem_cgroup);
777 	queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
778 }
779 
780 unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
781 {
782 	long x = READ_ONCE(memcg->vmstats->state[idx]);
783 #ifdef CONFIG_SMP
784 	if (x < 0)
785 		x = 0;
786 #endif
787 	return x;
788 }
789 
790 static int memcg_page_state_unit(int item);
791 
792 /*
793  * Normalize the value passed into memcg_rstat_updated() to be in pages. Round
794  * up non-zero sub-page updates to 1 page as zero page updates are ignored.
795  */
796 static int memcg_state_val_in_pages(int idx, int val)
797 {
798 	int unit = memcg_page_state_unit(idx);
799 
800 	if (!val || unit == PAGE_SIZE)
801 		return val;
802 	else
803 		return max(val * unit / PAGE_SIZE, 1UL);
804 }
805 
806 /**
807  * __mod_memcg_state - update cgroup memory statistics
808  * @memcg: the memory cgroup
809  * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
810  * @val: delta to add to the counter, can be negative
811  */
812 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
813 {
814 	if (mem_cgroup_disabled())
815 		return;
816 
817 	__this_cpu_add(memcg->vmstats_percpu->state[idx], val);
818 	memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
819 }
820 
821 /* idx can be of type enum memcg_stat_item or node_stat_item. */
822 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
823 {
824 	long x = READ_ONCE(memcg->vmstats->state_local[idx]);
825 
826 #ifdef CONFIG_SMP
827 	if (x < 0)
828 		x = 0;
829 #endif
830 	return x;
831 }
832 
833 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
834 			      int val)
835 {
836 	struct mem_cgroup_per_node *pn;
837 	struct mem_cgroup *memcg;
838 
839 	pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
840 	memcg = pn->memcg;
841 
842 	/*
843 	 * The caller from rmap relies on disabled preemption because they never
844 	 * update their counter from in-interrupt context. For these two
845 	 * counters we check that the update is never performed from an
846 	 * interrupt context while other caller need to have disabled interrupt.
847 	 */
848 	__memcg_stats_lock();
849 	if (IS_ENABLED(CONFIG_DEBUG_VM)) {
850 		switch (idx) {
851 		case NR_ANON_MAPPED:
852 		case NR_FILE_MAPPED:
853 		case NR_ANON_THPS:
854 		case NR_SHMEM_PMDMAPPED:
855 		case NR_FILE_PMDMAPPED:
856 			WARN_ON_ONCE(!in_task());
857 			break;
858 		default:
859 			VM_WARN_ON_IRQS_ENABLED();
860 		}
861 	}
862 
863 	/* Update memcg */
864 	__this_cpu_add(memcg->vmstats_percpu->state[idx], val);
865 
866 	/* Update lruvec */
867 	__this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
868 
869 	memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
870 	memcg_stats_unlock();
871 }
872 
873 /**
874  * __mod_lruvec_state - update lruvec memory statistics
875  * @lruvec: the lruvec
876  * @idx: the stat item
877  * @val: delta to add to the counter, can be negative
878  *
879  * The lruvec is the intersection of the NUMA node and a cgroup. This
880  * function updates the all three counters that are affected by a
881  * change of state at this level: per-node, per-cgroup, per-lruvec.
882  */
883 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
884 			int val)
885 {
886 	/* Update node */
887 	__mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
888 
889 	/* Update memcg and lruvec */
890 	if (!mem_cgroup_disabled())
891 		__mod_memcg_lruvec_state(lruvec, idx, val);
892 }
893 
894 void __lruvec_stat_mod_folio(struct folio *folio, enum node_stat_item idx,
895 			     int val)
896 {
897 	struct mem_cgroup *memcg;
898 	pg_data_t *pgdat = folio_pgdat(folio);
899 	struct lruvec *lruvec;
900 
901 	rcu_read_lock();
902 	memcg = folio_memcg(folio);
903 	/* Untracked pages have no memcg, no lruvec. Update only the node */
904 	if (!memcg) {
905 		rcu_read_unlock();
906 		__mod_node_page_state(pgdat, idx, val);
907 		return;
908 	}
909 
910 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
911 	__mod_lruvec_state(lruvec, idx, val);
912 	rcu_read_unlock();
913 }
914 EXPORT_SYMBOL(__lruvec_stat_mod_folio);
915 
916 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
917 {
918 	pg_data_t *pgdat = page_pgdat(virt_to_page(p));
919 	struct mem_cgroup *memcg;
920 	struct lruvec *lruvec;
921 
922 	rcu_read_lock();
923 	memcg = mem_cgroup_from_slab_obj(p);
924 
925 	/*
926 	 * Untracked pages have no memcg, no lruvec. Update only the
927 	 * node. If we reparent the slab objects to the root memcg,
928 	 * when we free the slab object, we need to update the per-memcg
929 	 * vmstats to keep it correct for the root memcg.
930 	 */
931 	if (!memcg) {
932 		__mod_node_page_state(pgdat, idx, val);
933 	} else {
934 		lruvec = mem_cgroup_lruvec(memcg, pgdat);
935 		__mod_lruvec_state(lruvec, idx, val);
936 	}
937 	rcu_read_unlock();
938 }
939 
940 /**
941  * __count_memcg_events - account VM events in a cgroup
942  * @memcg: the memory cgroup
943  * @idx: the event item
944  * @count: the number of events that occurred
945  */
946 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
947 			  unsigned long count)
948 {
949 	int index = memcg_events_index(idx);
950 
951 	if (mem_cgroup_disabled() || index < 0)
952 		return;
953 
954 	memcg_stats_lock();
955 	__this_cpu_add(memcg->vmstats_percpu->events[index], count);
956 	memcg_rstat_updated(memcg, count);
957 	memcg_stats_unlock();
958 }
959 
960 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
961 {
962 	int index = memcg_events_index(event);
963 
964 	if (index < 0)
965 		return 0;
966 	return READ_ONCE(memcg->vmstats->events[index]);
967 }
968 
969 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
970 {
971 	int index = memcg_events_index(event);
972 
973 	if (index < 0)
974 		return 0;
975 
976 	return READ_ONCE(memcg->vmstats->events_local[index]);
977 }
978 
979 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
980 					 int nr_pages)
981 {
982 	/* pagein of a big page is an event. So, ignore page size */
983 	if (nr_pages > 0)
984 		__count_memcg_events(memcg, PGPGIN, 1);
985 	else {
986 		__count_memcg_events(memcg, PGPGOUT, 1);
987 		nr_pages = -nr_pages; /* for event */
988 	}
989 
990 	__this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
991 }
992 
993 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
994 				       enum mem_cgroup_events_target target)
995 {
996 	unsigned long val, next;
997 
998 	val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
999 	next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
1000 	/* from time_after() in jiffies.h */
1001 	if ((long)(next - val) < 0) {
1002 		switch (target) {
1003 		case MEM_CGROUP_TARGET_THRESH:
1004 			next = val + THRESHOLDS_EVENTS_TARGET;
1005 			break;
1006 		case MEM_CGROUP_TARGET_SOFTLIMIT:
1007 			next = val + SOFTLIMIT_EVENTS_TARGET;
1008 			break;
1009 		default:
1010 			break;
1011 		}
1012 		__this_cpu_write(memcg->vmstats_percpu->targets[target], next);
1013 		return true;
1014 	}
1015 	return false;
1016 }
1017 
1018 /*
1019  * Check events in order.
1020  *
1021  */
1022 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
1023 {
1024 	if (IS_ENABLED(CONFIG_PREEMPT_RT))
1025 		return;
1026 
1027 	/* threshold event is triggered in finer grain than soft limit */
1028 	if (unlikely(mem_cgroup_event_ratelimit(memcg,
1029 						MEM_CGROUP_TARGET_THRESH))) {
1030 		bool do_softlimit;
1031 
1032 		do_softlimit = mem_cgroup_event_ratelimit(memcg,
1033 						MEM_CGROUP_TARGET_SOFTLIMIT);
1034 		mem_cgroup_threshold(memcg);
1035 		if (unlikely(do_softlimit))
1036 			mem_cgroup_update_tree(memcg, nid);
1037 	}
1038 }
1039 
1040 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1041 {
1042 	/*
1043 	 * mm_update_next_owner() may clear mm->owner to NULL
1044 	 * if it races with swapoff, page migration, etc.
1045 	 * So this can be called with p == NULL.
1046 	 */
1047 	if (unlikely(!p))
1048 		return NULL;
1049 
1050 	return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1051 }
1052 EXPORT_SYMBOL(mem_cgroup_from_task);
1053 
1054 static __always_inline struct mem_cgroup *active_memcg(void)
1055 {
1056 	if (!in_task())
1057 		return this_cpu_read(int_active_memcg);
1058 	else
1059 		return current->active_memcg;
1060 }
1061 
1062 /**
1063  * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1064  * @mm: mm from which memcg should be extracted. It can be NULL.
1065  *
1066  * Obtain a reference on mm->memcg and returns it if successful. If mm
1067  * is NULL, then the memcg is chosen as follows:
1068  * 1) The active memcg, if set.
1069  * 2) current->mm->memcg, if available
1070  * 3) root memcg
1071  * If mem_cgroup is disabled, NULL is returned.
1072  */
1073 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1074 {
1075 	struct mem_cgroup *memcg;
1076 
1077 	if (mem_cgroup_disabled())
1078 		return NULL;
1079 
1080 	/*
1081 	 * Page cache insertions can happen without an
1082 	 * actual mm context, e.g. during disk probing
1083 	 * on boot, loopback IO, acct() writes etc.
1084 	 *
1085 	 * No need to css_get on root memcg as the reference
1086 	 * counting is disabled on the root level in the
1087 	 * cgroup core. See CSS_NO_REF.
1088 	 */
1089 	if (unlikely(!mm)) {
1090 		memcg = active_memcg();
1091 		if (unlikely(memcg)) {
1092 			/* remote memcg must hold a ref */
1093 			css_get(&memcg->css);
1094 			return memcg;
1095 		}
1096 		mm = current->mm;
1097 		if (unlikely(!mm))
1098 			return root_mem_cgroup;
1099 	}
1100 
1101 	rcu_read_lock();
1102 	do {
1103 		memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1104 		if (unlikely(!memcg))
1105 			memcg = root_mem_cgroup;
1106 	} while (!css_tryget(&memcg->css));
1107 	rcu_read_unlock();
1108 	return memcg;
1109 }
1110 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1111 
1112 /**
1113  * get_mem_cgroup_from_current - Obtain a reference on current task's memcg.
1114  */
1115 struct mem_cgroup *get_mem_cgroup_from_current(void)
1116 {
1117 	struct mem_cgroup *memcg;
1118 
1119 	if (mem_cgroup_disabled())
1120 		return NULL;
1121 
1122 again:
1123 	rcu_read_lock();
1124 	memcg = mem_cgroup_from_task(current);
1125 	if (!css_tryget(&memcg->css)) {
1126 		rcu_read_unlock();
1127 		goto again;
1128 	}
1129 	rcu_read_unlock();
1130 	return memcg;
1131 }
1132 
1133 /**
1134  * mem_cgroup_iter - iterate over memory cgroup hierarchy
1135  * @root: hierarchy root
1136  * @prev: previously returned memcg, NULL on first invocation
1137  * @reclaim: cookie for shared reclaim walks, NULL for full walks
1138  *
1139  * Returns references to children of the hierarchy below @root, or
1140  * @root itself, or %NULL after a full round-trip.
1141  *
1142  * Caller must pass the return value in @prev on subsequent
1143  * invocations for reference counting, or use mem_cgroup_iter_break()
1144  * to cancel a hierarchy walk before the round-trip is complete.
1145  *
1146  * Reclaimers can specify a node in @reclaim to divide up the memcgs
1147  * in the hierarchy among all concurrent reclaimers operating on the
1148  * same node.
1149  */
1150 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1151 				   struct mem_cgroup *prev,
1152 				   struct mem_cgroup_reclaim_cookie *reclaim)
1153 {
1154 	struct mem_cgroup_reclaim_iter *iter;
1155 	struct cgroup_subsys_state *css = NULL;
1156 	struct mem_cgroup *memcg = NULL;
1157 	struct mem_cgroup *pos = NULL;
1158 
1159 	if (mem_cgroup_disabled())
1160 		return NULL;
1161 
1162 	if (!root)
1163 		root = root_mem_cgroup;
1164 
1165 	rcu_read_lock();
1166 
1167 	if (reclaim) {
1168 		struct mem_cgroup_per_node *mz;
1169 
1170 		mz = root->nodeinfo[reclaim->pgdat->node_id];
1171 		iter = &mz->iter;
1172 
1173 		/*
1174 		 * On start, join the current reclaim iteration cycle.
1175 		 * Exit when a concurrent walker completes it.
1176 		 */
1177 		if (!prev)
1178 			reclaim->generation = iter->generation;
1179 		else if (reclaim->generation != iter->generation)
1180 			goto out_unlock;
1181 
1182 		while (1) {
1183 			pos = READ_ONCE(iter->position);
1184 			if (!pos || css_tryget(&pos->css))
1185 				break;
1186 			/*
1187 			 * css reference reached zero, so iter->position will
1188 			 * be cleared by ->css_released. However, we should not
1189 			 * rely on this happening soon, because ->css_released
1190 			 * is called from a work queue, and by busy-waiting we
1191 			 * might block it. So we clear iter->position right
1192 			 * away.
1193 			 */
1194 			(void)cmpxchg(&iter->position, pos, NULL);
1195 		}
1196 	} else if (prev) {
1197 		pos = prev;
1198 	}
1199 
1200 	if (pos)
1201 		css = &pos->css;
1202 
1203 	for (;;) {
1204 		css = css_next_descendant_pre(css, &root->css);
1205 		if (!css) {
1206 			/*
1207 			 * Reclaimers share the hierarchy walk, and a
1208 			 * new one might jump in right at the end of
1209 			 * the hierarchy - make sure they see at least
1210 			 * one group and restart from the beginning.
1211 			 */
1212 			if (!prev)
1213 				continue;
1214 			break;
1215 		}
1216 
1217 		/*
1218 		 * Verify the css and acquire a reference.  The root
1219 		 * is provided by the caller, so we know it's alive
1220 		 * and kicking, and don't take an extra reference.
1221 		 */
1222 		if (css == &root->css || css_tryget(css)) {
1223 			memcg = mem_cgroup_from_css(css);
1224 			break;
1225 		}
1226 	}
1227 
1228 	if (reclaim) {
1229 		/*
1230 		 * The position could have already been updated by a competing
1231 		 * thread, so check that the value hasn't changed since we read
1232 		 * it to avoid reclaiming from the same cgroup twice.
1233 		 */
1234 		(void)cmpxchg(&iter->position, pos, memcg);
1235 
1236 		if (pos)
1237 			css_put(&pos->css);
1238 
1239 		if (!memcg)
1240 			iter->generation++;
1241 	}
1242 
1243 out_unlock:
1244 	rcu_read_unlock();
1245 	if (prev && prev != root)
1246 		css_put(&prev->css);
1247 
1248 	return memcg;
1249 }
1250 
1251 /**
1252  * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1253  * @root: hierarchy root
1254  * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1255  */
1256 void mem_cgroup_iter_break(struct mem_cgroup *root,
1257 			   struct mem_cgroup *prev)
1258 {
1259 	if (!root)
1260 		root = root_mem_cgroup;
1261 	if (prev && prev != root)
1262 		css_put(&prev->css);
1263 }
1264 
1265 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1266 					struct mem_cgroup *dead_memcg)
1267 {
1268 	struct mem_cgroup_reclaim_iter *iter;
1269 	struct mem_cgroup_per_node *mz;
1270 	int nid;
1271 
1272 	for_each_node(nid) {
1273 		mz = from->nodeinfo[nid];
1274 		iter = &mz->iter;
1275 		cmpxchg(&iter->position, dead_memcg, NULL);
1276 	}
1277 }
1278 
1279 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1280 {
1281 	struct mem_cgroup *memcg = dead_memcg;
1282 	struct mem_cgroup *last;
1283 
1284 	do {
1285 		__invalidate_reclaim_iterators(memcg, dead_memcg);
1286 		last = memcg;
1287 	} while ((memcg = parent_mem_cgroup(memcg)));
1288 
1289 	/*
1290 	 * When cgroup1 non-hierarchy mode is used,
1291 	 * parent_mem_cgroup() does not walk all the way up to the
1292 	 * cgroup root (root_mem_cgroup). So we have to handle
1293 	 * dead_memcg from cgroup root separately.
1294 	 */
1295 	if (!mem_cgroup_is_root(last))
1296 		__invalidate_reclaim_iterators(root_mem_cgroup,
1297 						dead_memcg);
1298 }
1299 
1300 /**
1301  * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1302  * @memcg: hierarchy root
1303  * @fn: function to call for each task
1304  * @arg: argument passed to @fn
1305  *
1306  * This function iterates over tasks attached to @memcg or to any of its
1307  * descendants and calls @fn for each task. If @fn returns a non-zero
1308  * value, the function breaks the iteration loop. Otherwise, it will iterate
1309  * over all tasks and return 0.
1310  *
1311  * This function must not be called for the root memory cgroup.
1312  */
1313 void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1314 			   int (*fn)(struct task_struct *, void *), void *arg)
1315 {
1316 	struct mem_cgroup *iter;
1317 	int ret = 0;
1318 
1319 	BUG_ON(mem_cgroup_is_root(memcg));
1320 
1321 	for_each_mem_cgroup_tree(iter, memcg) {
1322 		struct css_task_iter it;
1323 		struct task_struct *task;
1324 
1325 		css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1326 		while (!ret && (task = css_task_iter_next(&it)))
1327 			ret = fn(task, arg);
1328 		css_task_iter_end(&it);
1329 		if (ret) {
1330 			mem_cgroup_iter_break(memcg, iter);
1331 			break;
1332 		}
1333 	}
1334 }
1335 
1336 #ifdef CONFIG_DEBUG_VM
1337 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1338 {
1339 	struct mem_cgroup *memcg;
1340 
1341 	if (mem_cgroup_disabled())
1342 		return;
1343 
1344 	memcg = folio_memcg(folio);
1345 
1346 	if (!memcg)
1347 		VM_BUG_ON_FOLIO(!mem_cgroup_is_root(lruvec_memcg(lruvec)), folio);
1348 	else
1349 		VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1350 }
1351 #endif
1352 
1353 /**
1354  * folio_lruvec_lock - Lock the lruvec for a folio.
1355  * @folio: Pointer to the folio.
1356  *
1357  * These functions are safe to use under any of the following conditions:
1358  * - folio locked
1359  * - folio_test_lru false
1360  * - folio_memcg_lock()
1361  * - folio frozen (refcount of 0)
1362  *
1363  * Return: The lruvec this folio is on with its lock held.
1364  */
1365 struct lruvec *folio_lruvec_lock(struct folio *folio)
1366 {
1367 	struct lruvec *lruvec = folio_lruvec(folio);
1368 
1369 	spin_lock(&lruvec->lru_lock);
1370 	lruvec_memcg_debug(lruvec, folio);
1371 
1372 	return lruvec;
1373 }
1374 
1375 /**
1376  * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1377  * @folio: Pointer to the folio.
1378  *
1379  * These functions are safe to use under any of the following conditions:
1380  * - folio locked
1381  * - folio_test_lru false
1382  * - folio_memcg_lock()
1383  * - folio frozen (refcount of 0)
1384  *
1385  * Return: The lruvec this folio is on with its lock held and interrupts
1386  * disabled.
1387  */
1388 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1389 {
1390 	struct lruvec *lruvec = folio_lruvec(folio);
1391 
1392 	spin_lock_irq(&lruvec->lru_lock);
1393 	lruvec_memcg_debug(lruvec, folio);
1394 
1395 	return lruvec;
1396 }
1397 
1398 /**
1399  * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1400  * @folio: Pointer to the folio.
1401  * @flags: Pointer to irqsave flags.
1402  *
1403  * These functions are safe to use under any of the following conditions:
1404  * - folio locked
1405  * - folio_test_lru false
1406  * - folio_memcg_lock()
1407  * - folio frozen (refcount of 0)
1408  *
1409  * Return: The lruvec this folio is on with its lock held and interrupts
1410  * disabled.
1411  */
1412 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1413 		unsigned long *flags)
1414 {
1415 	struct lruvec *lruvec = folio_lruvec(folio);
1416 
1417 	spin_lock_irqsave(&lruvec->lru_lock, *flags);
1418 	lruvec_memcg_debug(lruvec, folio);
1419 
1420 	return lruvec;
1421 }
1422 
1423 /**
1424  * mem_cgroup_update_lru_size - account for adding or removing an lru page
1425  * @lruvec: mem_cgroup per zone lru vector
1426  * @lru: index of lru list the page is sitting on
1427  * @zid: zone id of the accounted pages
1428  * @nr_pages: positive when adding or negative when removing
1429  *
1430  * This function must be called under lru_lock, just before a page is added
1431  * to or just after a page is removed from an lru list.
1432  */
1433 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1434 				int zid, int nr_pages)
1435 {
1436 	struct mem_cgroup_per_node *mz;
1437 	unsigned long *lru_size;
1438 	long size;
1439 
1440 	if (mem_cgroup_disabled())
1441 		return;
1442 
1443 	mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1444 	lru_size = &mz->lru_zone_size[zid][lru];
1445 
1446 	if (nr_pages < 0)
1447 		*lru_size += nr_pages;
1448 
1449 	size = *lru_size;
1450 	if (WARN_ONCE(size < 0,
1451 		"%s(%p, %d, %d): lru_size %ld\n",
1452 		__func__, lruvec, lru, nr_pages, size)) {
1453 		VM_BUG_ON(1);
1454 		*lru_size = 0;
1455 	}
1456 
1457 	if (nr_pages > 0)
1458 		*lru_size += nr_pages;
1459 }
1460 
1461 /**
1462  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1463  * @memcg: the memory cgroup
1464  *
1465  * Returns the maximum amount of memory @mem can be charged with, in
1466  * pages.
1467  */
1468 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1469 {
1470 	unsigned long margin = 0;
1471 	unsigned long count;
1472 	unsigned long limit;
1473 
1474 	count = page_counter_read(&memcg->memory);
1475 	limit = READ_ONCE(memcg->memory.max);
1476 	if (count < limit)
1477 		margin = limit - count;
1478 
1479 	if (do_memsw_account()) {
1480 		count = page_counter_read(&memcg->memsw);
1481 		limit = READ_ONCE(memcg->memsw.max);
1482 		if (count < limit)
1483 			margin = min(margin, limit - count);
1484 		else
1485 			margin = 0;
1486 	}
1487 
1488 	return margin;
1489 }
1490 
1491 /*
1492  * A routine for checking "mem" is under move_account() or not.
1493  *
1494  * Checking a cgroup is mc.from or mc.to or under hierarchy of
1495  * moving cgroups. This is for waiting at high-memory pressure
1496  * caused by "move".
1497  */
1498 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1499 {
1500 	struct mem_cgroup *from;
1501 	struct mem_cgroup *to;
1502 	bool ret = false;
1503 	/*
1504 	 * Unlike task_move routines, we access mc.to, mc.from not under
1505 	 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1506 	 */
1507 	spin_lock(&mc.lock);
1508 	from = mc.from;
1509 	to = mc.to;
1510 	if (!from)
1511 		goto unlock;
1512 
1513 	ret = mem_cgroup_is_descendant(from, memcg) ||
1514 		mem_cgroup_is_descendant(to, memcg);
1515 unlock:
1516 	spin_unlock(&mc.lock);
1517 	return ret;
1518 }
1519 
1520 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1521 {
1522 	if (mc.moving_task && current != mc.moving_task) {
1523 		if (mem_cgroup_under_move(memcg)) {
1524 			DEFINE_WAIT(wait);
1525 			prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1526 			/* moving charge context might have finished. */
1527 			if (mc.moving_task)
1528 				schedule();
1529 			finish_wait(&mc.waitq, &wait);
1530 			return true;
1531 		}
1532 	}
1533 	return false;
1534 }
1535 
1536 struct memory_stat {
1537 	const char *name;
1538 	unsigned int idx;
1539 };
1540 
1541 static const struct memory_stat memory_stats[] = {
1542 	{ "anon",			NR_ANON_MAPPED			},
1543 	{ "file",			NR_FILE_PAGES			},
1544 	{ "kernel",			MEMCG_KMEM			},
1545 	{ "kernel_stack",		NR_KERNEL_STACK_KB		},
1546 	{ "pagetables",			NR_PAGETABLE			},
1547 	{ "sec_pagetables",		NR_SECONDARY_PAGETABLE		},
1548 	{ "percpu",			MEMCG_PERCPU_B			},
1549 	{ "sock",			MEMCG_SOCK			},
1550 	{ "vmalloc",			MEMCG_VMALLOC			},
1551 	{ "shmem",			NR_SHMEM			},
1552 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1553 	{ "zswap",			MEMCG_ZSWAP_B			},
1554 	{ "zswapped",			MEMCG_ZSWAPPED			},
1555 #endif
1556 	{ "file_mapped",		NR_FILE_MAPPED			},
1557 	{ "file_dirty",			NR_FILE_DIRTY			},
1558 	{ "file_writeback",		NR_WRITEBACK			},
1559 #ifdef CONFIG_SWAP
1560 	{ "swapcached",			NR_SWAPCACHE			},
1561 #endif
1562 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1563 	{ "anon_thp",			NR_ANON_THPS			},
1564 	{ "file_thp",			NR_FILE_THPS			},
1565 	{ "shmem_thp",			NR_SHMEM_THPS			},
1566 #endif
1567 	{ "inactive_anon",		NR_INACTIVE_ANON		},
1568 	{ "active_anon",		NR_ACTIVE_ANON			},
1569 	{ "inactive_file",		NR_INACTIVE_FILE		},
1570 	{ "active_file",		NR_ACTIVE_FILE			},
1571 	{ "unevictable",		NR_UNEVICTABLE			},
1572 	{ "slab_reclaimable",		NR_SLAB_RECLAIMABLE_B		},
1573 	{ "slab_unreclaimable",		NR_SLAB_UNRECLAIMABLE_B		},
1574 
1575 	/* The memory events */
1576 	{ "workingset_refault_anon",	WORKINGSET_REFAULT_ANON		},
1577 	{ "workingset_refault_file",	WORKINGSET_REFAULT_FILE		},
1578 	{ "workingset_activate_anon",	WORKINGSET_ACTIVATE_ANON	},
1579 	{ "workingset_activate_file",	WORKINGSET_ACTIVATE_FILE	},
1580 	{ "workingset_restore_anon",	WORKINGSET_RESTORE_ANON		},
1581 	{ "workingset_restore_file",	WORKINGSET_RESTORE_FILE		},
1582 	{ "workingset_nodereclaim",	WORKINGSET_NODERECLAIM		},
1583 };
1584 
1585 /* The actual unit of the state item, not the same as the output unit */
1586 static int memcg_page_state_unit(int item)
1587 {
1588 	switch (item) {
1589 	case MEMCG_PERCPU_B:
1590 	case MEMCG_ZSWAP_B:
1591 	case NR_SLAB_RECLAIMABLE_B:
1592 	case NR_SLAB_UNRECLAIMABLE_B:
1593 		return 1;
1594 	case NR_KERNEL_STACK_KB:
1595 		return SZ_1K;
1596 	default:
1597 		return PAGE_SIZE;
1598 	}
1599 }
1600 
1601 /* Translate stat items to the correct unit for memory.stat output */
1602 static int memcg_page_state_output_unit(int item)
1603 {
1604 	/*
1605 	 * Workingset state is actually in pages, but we export it to userspace
1606 	 * as a scalar count of events, so special case it here.
1607 	 */
1608 	switch (item) {
1609 	case WORKINGSET_REFAULT_ANON:
1610 	case WORKINGSET_REFAULT_FILE:
1611 	case WORKINGSET_ACTIVATE_ANON:
1612 	case WORKINGSET_ACTIVATE_FILE:
1613 	case WORKINGSET_RESTORE_ANON:
1614 	case WORKINGSET_RESTORE_FILE:
1615 	case WORKINGSET_NODERECLAIM:
1616 		return 1;
1617 	default:
1618 		return memcg_page_state_unit(item);
1619 	}
1620 }
1621 
1622 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1623 						    int item)
1624 {
1625 	return memcg_page_state(memcg, item) *
1626 		memcg_page_state_output_unit(item);
1627 }
1628 
1629 static inline unsigned long memcg_page_state_local_output(
1630 		struct mem_cgroup *memcg, int item)
1631 {
1632 	return memcg_page_state_local(memcg, item) *
1633 		memcg_page_state_output_unit(item);
1634 }
1635 
1636 static void memcg_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1637 {
1638 	int i;
1639 
1640 	/*
1641 	 * Provide statistics on the state of the memory subsystem as
1642 	 * well as cumulative event counters that show past behavior.
1643 	 *
1644 	 * This list is ordered following a combination of these gradients:
1645 	 * 1) generic big picture -> specifics and details
1646 	 * 2) reflecting userspace activity -> reflecting kernel heuristics
1647 	 *
1648 	 * Current memory state:
1649 	 */
1650 	mem_cgroup_flush_stats(memcg);
1651 
1652 	for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1653 		u64 size;
1654 
1655 		size = memcg_page_state_output(memcg, memory_stats[i].idx);
1656 		seq_buf_printf(s, "%s %llu\n", memory_stats[i].name, size);
1657 
1658 		if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1659 			size += memcg_page_state_output(memcg,
1660 							NR_SLAB_RECLAIMABLE_B);
1661 			seq_buf_printf(s, "slab %llu\n", size);
1662 		}
1663 	}
1664 
1665 	/* Accumulated memory events */
1666 	seq_buf_printf(s, "pgscan %lu\n",
1667 		       memcg_events(memcg, PGSCAN_KSWAPD) +
1668 		       memcg_events(memcg, PGSCAN_DIRECT) +
1669 		       memcg_events(memcg, PGSCAN_KHUGEPAGED));
1670 	seq_buf_printf(s, "pgsteal %lu\n",
1671 		       memcg_events(memcg, PGSTEAL_KSWAPD) +
1672 		       memcg_events(memcg, PGSTEAL_DIRECT) +
1673 		       memcg_events(memcg, PGSTEAL_KHUGEPAGED));
1674 
1675 	for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) {
1676 		if (memcg_vm_event_stat[i] == PGPGIN ||
1677 		    memcg_vm_event_stat[i] == PGPGOUT)
1678 			continue;
1679 
1680 		seq_buf_printf(s, "%s %lu\n",
1681 			       vm_event_name(memcg_vm_event_stat[i]),
1682 			       memcg_events(memcg, memcg_vm_event_stat[i]));
1683 	}
1684 
1685 	/* The above should easily fit into one page */
1686 	WARN_ON_ONCE(seq_buf_has_overflowed(s));
1687 }
1688 
1689 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s);
1690 
1691 static void memory_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1692 {
1693 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1694 		memcg_stat_format(memcg, s);
1695 	else
1696 		memcg1_stat_format(memcg, s);
1697 	WARN_ON_ONCE(seq_buf_has_overflowed(s));
1698 }
1699 
1700 /**
1701  * mem_cgroup_print_oom_context: Print OOM information relevant to
1702  * memory controller.
1703  * @memcg: The memory cgroup that went over limit
1704  * @p: Task that is going to be killed
1705  *
1706  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1707  * enabled
1708  */
1709 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1710 {
1711 	rcu_read_lock();
1712 
1713 	if (memcg) {
1714 		pr_cont(",oom_memcg=");
1715 		pr_cont_cgroup_path(memcg->css.cgroup);
1716 	} else
1717 		pr_cont(",global_oom");
1718 	if (p) {
1719 		pr_cont(",task_memcg=");
1720 		pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1721 	}
1722 	rcu_read_unlock();
1723 }
1724 
1725 /**
1726  * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1727  * memory controller.
1728  * @memcg: The memory cgroup that went over limit
1729  */
1730 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1731 {
1732 	/* Use static buffer, for the caller is holding oom_lock. */
1733 	static char buf[PAGE_SIZE];
1734 	struct seq_buf s;
1735 
1736 	lockdep_assert_held(&oom_lock);
1737 
1738 	pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1739 		K((u64)page_counter_read(&memcg->memory)),
1740 		K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1741 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1742 		pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1743 			K((u64)page_counter_read(&memcg->swap)),
1744 			K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1745 	else {
1746 		pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1747 			K((u64)page_counter_read(&memcg->memsw)),
1748 			K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1749 		pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1750 			K((u64)page_counter_read(&memcg->kmem)),
1751 			K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1752 	}
1753 
1754 	pr_info("Memory cgroup stats for ");
1755 	pr_cont_cgroup_path(memcg->css.cgroup);
1756 	pr_cont(":");
1757 	seq_buf_init(&s, buf, sizeof(buf));
1758 	memory_stat_format(memcg, &s);
1759 	seq_buf_do_printk(&s, KERN_INFO);
1760 }
1761 
1762 /*
1763  * Return the memory (and swap, if configured) limit for a memcg.
1764  */
1765 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1766 {
1767 	unsigned long max = READ_ONCE(memcg->memory.max);
1768 
1769 	if (do_memsw_account()) {
1770 		if (mem_cgroup_swappiness(memcg)) {
1771 			/* Calculate swap excess capacity from memsw limit */
1772 			unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1773 
1774 			max += min(swap, (unsigned long)total_swap_pages);
1775 		}
1776 	} else {
1777 		if (mem_cgroup_swappiness(memcg))
1778 			max += min(READ_ONCE(memcg->swap.max),
1779 				   (unsigned long)total_swap_pages);
1780 	}
1781 	return max;
1782 }
1783 
1784 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1785 {
1786 	return page_counter_read(&memcg->memory);
1787 }
1788 
1789 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1790 				     int order)
1791 {
1792 	struct oom_control oc = {
1793 		.zonelist = NULL,
1794 		.nodemask = NULL,
1795 		.memcg = memcg,
1796 		.gfp_mask = gfp_mask,
1797 		.order = order,
1798 	};
1799 	bool ret = true;
1800 
1801 	if (mutex_lock_killable(&oom_lock))
1802 		return true;
1803 
1804 	if (mem_cgroup_margin(memcg) >= (1 << order))
1805 		goto unlock;
1806 
1807 	/*
1808 	 * A few threads which were not waiting at mutex_lock_killable() can
1809 	 * fail to bail out. Therefore, check again after holding oom_lock.
1810 	 */
1811 	ret = task_is_dying() || out_of_memory(&oc);
1812 
1813 unlock:
1814 	mutex_unlock(&oom_lock);
1815 	return ret;
1816 }
1817 
1818 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1819 				   pg_data_t *pgdat,
1820 				   gfp_t gfp_mask,
1821 				   unsigned long *total_scanned)
1822 {
1823 	struct mem_cgroup *victim = NULL;
1824 	int total = 0;
1825 	int loop = 0;
1826 	unsigned long excess;
1827 	unsigned long nr_scanned;
1828 	struct mem_cgroup_reclaim_cookie reclaim = {
1829 		.pgdat = pgdat,
1830 	};
1831 
1832 	excess = soft_limit_excess(root_memcg);
1833 
1834 	while (1) {
1835 		victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1836 		if (!victim) {
1837 			loop++;
1838 			if (loop >= 2) {
1839 				/*
1840 				 * If we have not been able to reclaim
1841 				 * anything, it might because there are
1842 				 * no reclaimable pages under this hierarchy
1843 				 */
1844 				if (!total)
1845 					break;
1846 				/*
1847 				 * We want to do more targeted reclaim.
1848 				 * excess >> 2 is not to excessive so as to
1849 				 * reclaim too much, nor too less that we keep
1850 				 * coming back to reclaim from this cgroup
1851 				 */
1852 				if (total >= (excess >> 2) ||
1853 					(loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1854 					break;
1855 			}
1856 			continue;
1857 		}
1858 		total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1859 					pgdat, &nr_scanned);
1860 		*total_scanned += nr_scanned;
1861 		if (!soft_limit_excess(root_memcg))
1862 			break;
1863 	}
1864 	mem_cgroup_iter_break(root_memcg, victim);
1865 	return total;
1866 }
1867 
1868 #ifdef CONFIG_LOCKDEP
1869 static struct lockdep_map memcg_oom_lock_dep_map = {
1870 	.name = "memcg_oom_lock",
1871 };
1872 #endif
1873 
1874 static DEFINE_SPINLOCK(memcg_oom_lock);
1875 
1876 /*
1877  * Check OOM-Killer is already running under our hierarchy.
1878  * If someone is running, return false.
1879  */
1880 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1881 {
1882 	struct mem_cgroup *iter, *failed = NULL;
1883 
1884 	spin_lock(&memcg_oom_lock);
1885 
1886 	for_each_mem_cgroup_tree(iter, memcg) {
1887 		if (iter->oom_lock) {
1888 			/*
1889 			 * this subtree of our hierarchy is already locked
1890 			 * so we cannot give a lock.
1891 			 */
1892 			failed = iter;
1893 			mem_cgroup_iter_break(memcg, iter);
1894 			break;
1895 		} else
1896 			iter->oom_lock = true;
1897 	}
1898 
1899 	if (failed) {
1900 		/*
1901 		 * OK, we failed to lock the whole subtree so we have
1902 		 * to clean up what we set up to the failing subtree
1903 		 */
1904 		for_each_mem_cgroup_tree(iter, memcg) {
1905 			if (iter == failed) {
1906 				mem_cgroup_iter_break(memcg, iter);
1907 				break;
1908 			}
1909 			iter->oom_lock = false;
1910 		}
1911 	} else
1912 		mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1913 
1914 	spin_unlock(&memcg_oom_lock);
1915 
1916 	return !failed;
1917 }
1918 
1919 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1920 {
1921 	struct mem_cgroup *iter;
1922 
1923 	spin_lock(&memcg_oom_lock);
1924 	mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1925 	for_each_mem_cgroup_tree(iter, memcg)
1926 		iter->oom_lock = false;
1927 	spin_unlock(&memcg_oom_lock);
1928 }
1929 
1930 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1931 {
1932 	struct mem_cgroup *iter;
1933 
1934 	spin_lock(&memcg_oom_lock);
1935 	for_each_mem_cgroup_tree(iter, memcg)
1936 		iter->under_oom++;
1937 	spin_unlock(&memcg_oom_lock);
1938 }
1939 
1940 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1941 {
1942 	struct mem_cgroup *iter;
1943 
1944 	/*
1945 	 * Be careful about under_oom underflows because a child memcg
1946 	 * could have been added after mem_cgroup_mark_under_oom.
1947 	 */
1948 	spin_lock(&memcg_oom_lock);
1949 	for_each_mem_cgroup_tree(iter, memcg)
1950 		if (iter->under_oom > 0)
1951 			iter->under_oom--;
1952 	spin_unlock(&memcg_oom_lock);
1953 }
1954 
1955 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1956 
1957 struct oom_wait_info {
1958 	struct mem_cgroup *memcg;
1959 	wait_queue_entry_t	wait;
1960 };
1961 
1962 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1963 	unsigned mode, int sync, void *arg)
1964 {
1965 	struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1966 	struct mem_cgroup *oom_wait_memcg;
1967 	struct oom_wait_info *oom_wait_info;
1968 
1969 	oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1970 	oom_wait_memcg = oom_wait_info->memcg;
1971 
1972 	if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1973 	    !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1974 		return 0;
1975 	return autoremove_wake_function(wait, mode, sync, arg);
1976 }
1977 
1978 static void memcg_oom_recover(struct mem_cgroup *memcg)
1979 {
1980 	/*
1981 	 * For the following lockless ->under_oom test, the only required
1982 	 * guarantee is that it must see the state asserted by an OOM when
1983 	 * this function is called as a result of userland actions
1984 	 * triggered by the notification of the OOM.  This is trivially
1985 	 * achieved by invoking mem_cgroup_mark_under_oom() before
1986 	 * triggering notification.
1987 	 */
1988 	if (memcg && memcg->under_oom)
1989 		__wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1990 }
1991 
1992 /*
1993  * Returns true if successfully killed one or more processes. Though in some
1994  * corner cases it can return true even without killing any process.
1995  */
1996 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1997 {
1998 	bool locked, ret;
1999 
2000 	if (order > PAGE_ALLOC_COSTLY_ORDER)
2001 		return false;
2002 
2003 	memcg_memory_event(memcg, MEMCG_OOM);
2004 
2005 	/*
2006 	 * We are in the middle of the charge context here, so we
2007 	 * don't want to block when potentially sitting on a callstack
2008 	 * that holds all kinds of filesystem and mm locks.
2009 	 *
2010 	 * cgroup1 allows disabling the OOM killer and waiting for outside
2011 	 * handling until the charge can succeed; remember the context and put
2012 	 * the task to sleep at the end of the page fault when all locks are
2013 	 * released.
2014 	 *
2015 	 * On the other hand, in-kernel OOM killer allows for an async victim
2016 	 * memory reclaim (oom_reaper) and that means that we are not solely
2017 	 * relying on the oom victim to make a forward progress and we can
2018 	 * invoke the oom killer here.
2019 	 *
2020 	 * Please note that mem_cgroup_out_of_memory might fail to find a
2021 	 * victim and then we have to bail out from the charge path.
2022 	 */
2023 	if (READ_ONCE(memcg->oom_kill_disable)) {
2024 		if (current->in_user_fault) {
2025 			css_get(&memcg->css);
2026 			current->memcg_in_oom = memcg;
2027 			current->memcg_oom_gfp_mask = mask;
2028 			current->memcg_oom_order = order;
2029 		}
2030 		return false;
2031 	}
2032 
2033 	mem_cgroup_mark_under_oom(memcg);
2034 
2035 	locked = mem_cgroup_oom_trylock(memcg);
2036 
2037 	if (locked)
2038 		mem_cgroup_oom_notify(memcg);
2039 
2040 	mem_cgroup_unmark_under_oom(memcg);
2041 	ret = mem_cgroup_out_of_memory(memcg, mask, order);
2042 
2043 	if (locked)
2044 		mem_cgroup_oom_unlock(memcg);
2045 
2046 	return ret;
2047 }
2048 
2049 /**
2050  * mem_cgroup_oom_synchronize - complete memcg OOM handling
2051  * @handle: actually kill/wait or just clean up the OOM state
2052  *
2053  * This has to be called at the end of a page fault if the memcg OOM
2054  * handler was enabled.
2055  *
2056  * Memcg supports userspace OOM handling where failed allocations must
2057  * sleep on a waitqueue until the userspace task resolves the
2058  * situation.  Sleeping directly in the charge context with all kinds
2059  * of locks held is not a good idea, instead we remember an OOM state
2060  * in the task and mem_cgroup_oom_synchronize() has to be called at
2061  * the end of the page fault to complete the OOM handling.
2062  *
2063  * Returns %true if an ongoing memcg OOM situation was detected and
2064  * completed, %false otherwise.
2065  */
2066 bool mem_cgroup_oom_synchronize(bool handle)
2067 {
2068 	struct mem_cgroup *memcg = current->memcg_in_oom;
2069 	struct oom_wait_info owait;
2070 	bool locked;
2071 
2072 	/* OOM is global, do not handle */
2073 	if (!memcg)
2074 		return false;
2075 
2076 	if (!handle)
2077 		goto cleanup;
2078 
2079 	owait.memcg = memcg;
2080 	owait.wait.flags = 0;
2081 	owait.wait.func = memcg_oom_wake_function;
2082 	owait.wait.private = current;
2083 	INIT_LIST_HEAD(&owait.wait.entry);
2084 
2085 	prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2086 	mem_cgroup_mark_under_oom(memcg);
2087 
2088 	locked = mem_cgroup_oom_trylock(memcg);
2089 
2090 	if (locked)
2091 		mem_cgroup_oom_notify(memcg);
2092 
2093 	schedule();
2094 	mem_cgroup_unmark_under_oom(memcg);
2095 	finish_wait(&memcg_oom_waitq, &owait.wait);
2096 
2097 	if (locked)
2098 		mem_cgroup_oom_unlock(memcg);
2099 cleanup:
2100 	current->memcg_in_oom = NULL;
2101 	css_put(&memcg->css);
2102 	return true;
2103 }
2104 
2105 /**
2106  * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2107  * @victim: task to be killed by the OOM killer
2108  * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2109  *
2110  * Returns a pointer to a memory cgroup, which has to be cleaned up
2111  * by killing all belonging OOM-killable tasks.
2112  *
2113  * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2114  */
2115 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2116 					    struct mem_cgroup *oom_domain)
2117 {
2118 	struct mem_cgroup *oom_group = NULL;
2119 	struct mem_cgroup *memcg;
2120 
2121 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2122 		return NULL;
2123 
2124 	if (!oom_domain)
2125 		oom_domain = root_mem_cgroup;
2126 
2127 	rcu_read_lock();
2128 
2129 	memcg = mem_cgroup_from_task(victim);
2130 	if (mem_cgroup_is_root(memcg))
2131 		goto out;
2132 
2133 	/*
2134 	 * If the victim task has been asynchronously moved to a different
2135 	 * memory cgroup, we might end up killing tasks outside oom_domain.
2136 	 * In this case it's better to ignore memory.group.oom.
2137 	 */
2138 	if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2139 		goto out;
2140 
2141 	/*
2142 	 * Traverse the memory cgroup hierarchy from the victim task's
2143 	 * cgroup up to the OOMing cgroup (or root) to find the
2144 	 * highest-level memory cgroup with oom.group set.
2145 	 */
2146 	for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2147 		if (READ_ONCE(memcg->oom_group))
2148 			oom_group = memcg;
2149 
2150 		if (memcg == oom_domain)
2151 			break;
2152 	}
2153 
2154 	if (oom_group)
2155 		css_get(&oom_group->css);
2156 out:
2157 	rcu_read_unlock();
2158 
2159 	return oom_group;
2160 }
2161 
2162 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2163 {
2164 	pr_info("Tasks in ");
2165 	pr_cont_cgroup_path(memcg->css.cgroup);
2166 	pr_cont(" are going to be killed due to memory.oom.group set\n");
2167 }
2168 
2169 /**
2170  * folio_memcg_lock - Bind a folio to its memcg.
2171  * @folio: The folio.
2172  *
2173  * This function prevents unlocked LRU folios from being moved to
2174  * another cgroup.
2175  *
2176  * It ensures lifetime of the bound memcg.  The caller is responsible
2177  * for the lifetime of the folio.
2178  */
2179 void folio_memcg_lock(struct folio *folio)
2180 {
2181 	struct mem_cgroup *memcg;
2182 	unsigned long flags;
2183 
2184 	/*
2185 	 * The RCU lock is held throughout the transaction.  The fast
2186 	 * path can get away without acquiring the memcg->move_lock
2187 	 * because page moving starts with an RCU grace period.
2188          */
2189 	rcu_read_lock();
2190 
2191 	if (mem_cgroup_disabled())
2192 		return;
2193 again:
2194 	memcg = folio_memcg(folio);
2195 	if (unlikely(!memcg))
2196 		return;
2197 
2198 #ifdef CONFIG_PROVE_LOCKING
2199 	local_irq_save(flags);
2200 	might_lock(&memcg->move_lock);
2201 	local_irq_restore(flags);
2202 #endif
2203 
2204 	if (atomic_read(&memcg->moving_account) <= 0)
2205 		return;
2206 
2207 	spin_lock_irqsave(&memcg->move_lock, flags);
2208 	if (memcg != folio_memcg(folio)) {
2209 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2210 		goto again;
2211 	}
2212 
2213 	/*
2214 	 * When charge migration first begins, we can have multiple
2215 	 * critical sections holding the fast-path RCU lock and one
2216 	 * holding the slowpath move_lock. Track the task who has the
2217 	 * move_lock for folio_memcg_unlock().
2218 	 */
2219 	memcg->move_lock_task = current;
2220 	memcg->move_lock_flags = flags;
2221 }
2222 
2223 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2224 {
2225 	if (memcg && memcg->move_lock_task == current) {
2226 		unsigned long flags = memcg->move_lock_flags;
2227 
2228 		memcg->move_lock_task = NULL;
2229 		memcg->move_lock_flags = 0;
2230 
2231 		spin_unlock_irqrestore(&memcg->move_lock, flags);
2232 	}
2233 
2234 	rcu_read_unlock();
2235 }
2236 
2237 /**
2238  * folio_memcg_unlock - Release the binding between a folio and its memcg.
2239  * @folio: The folio.
2240  *
2241  * This releases the binding created by folio_memcg_lock().  This does
2242  * not change the accounting of this folio to its memcg, but it does
2243  * permit others to change it.
2244  */
2245 void folio_memcg_unlock(struct folio *folio)
2246 {
2247 	__folio_memcg_unlock(folio_memcg(folio));
2248 }
2249 
2250 struct memcg_stock_pcp {
2251 	local_lock_t stock_lock;
2252 	struct mem_cgroup *cached; /* this never be root cgroup */
2253 	unsigned int nr_pages;
2254 
2255 #ifdef CONFIG_MEMCG_KMEM
2256 	struct obj_cgroup *cached_objcg;
2257 	struct pglist_data *cached_pgdat;
2258 	unsigned int nr_bytes;
2259 	int nr_slab_reclaimable_b;
2260 	int nr_slab_unreclaimable_b;
2261 #endif
2262 
2263 	struct work_struct work;
2264 	unsigned long flags;
2265 #define FLUSHING_CACHED_CHARGE	0
2266 };
2267 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2268 	.stock_lock = INIT_LOCAL_LOCK(stock_lock),
2269 };
2270 static DEFINE_MUTEX(percpu_charge_mutex);
2271 
2272 #ifdef CONFIG_MEMCG_KMEM
2273 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2274 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2275 				     struct mem_cgroup *root_memcg);
2276 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2277 
2278 #else
2279 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2280 {
2281 	return NULL;
2282 }
2283 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2284 				     struct mem_cgroup *root_memcg)
2285 {
2286 	return false;
2287 }
2288 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2289 {
2290 }
2291 #endif
2292 
2293 /**
2294  * consume_stock: Try to consume stocked charge on this cpu.
2295  * @memcg: memcg to consume from.
2296  * @nr_pages: how many pages to charge.
2297  *
2298  * The charges will only happen if @memcg matches the current cpu's memcg
2299  * stock, and at least @nr_pages are available in that stock.  Failure to
2300  * service an allocation will refill the stock.
2301  *
2302  * returns true if successful, false otherwise.
2303  */
2304 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2305 {
2306 	struct memcg_stock_pcp *stock;
2307 	unsigned long flags;
2308 	bool ret = false;
2309 
2310 	if (nr_pages > MEMCG_CHARGE_BATCH)
2311 		return ret;
2312 
2313 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
2314 
2315 	stock = this_cpu_ptr(&memcg_stock);
2316 	if (memcg == READ_ONCE(stock->cached) && stock->nr_pages >= nr_pages) {
2317 		stock->nr_pages -= nr_pages;
2318 		ret = true;
2319 	}
2320 
2321 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2322 
2323 	return ret;
2324 }
2325 
2326 /*
2327  * Returns stocks cached in percpu and reset cached information.
2328  */
2329 static void drain_stock(struct memcg_stock_pcp *stock)
2330 {
2331 	struct mem_cgroup *old = READ_ONCE(stock->cached);
2332 
2333 	if (!old)
2334 		return;
2335 
2336 	if (stock->nr_pages) {
2337 		page_counter_uncharge(&old->memory, stock->nr_pages);
2338 		if (do_memsw_account())
2339 			page_counter_uncharge(&old->memsw, stock->nr_pages);
2340 		stock->nr_pages = 0;
2341 	}
2342 
2343 	css_put(&old->css);
2344 	WRITE_ONCE(stock->cached, NULL);
2345 }
2346 
2347 static void drain_local_stock(struct work_struct *dummy)
2348 {
2349 	struct memcg_stock_pcp *stock;
2350 	struct obj_cgroup *old = NULL;
2351 	unsigned long flags;
2352 
2353 	/*
2354 	 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2355 	 * drain_stock races is that we always operate on local CPU stock
2356 	 * here with IRQ disabled
2357 	 */
2358 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
2359 
2360 	stock = this_cpu_ptr(&memcg_stock);
2361 	old = drain_obj_stock(stock);
2362 	drain_stock(stock);
2363 	clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2364 
2365 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2366 	if (old)
2367 		obj_cgroup_put(old);
2368 }
2369 
2370 /*
2371  * Cache charges(val) to local per_cpu area.
2372  * This will be consumed by consume_stock() function, later.
2373  */
2374 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2375 {
2376 	struct memcg_stock_pcp *stock;
2377 
2378 	stock = this_cpu_ptr(&memcg_stock);
2379 	if (READ_ONCE(stock->cached) != memcg) { /* reset if necessary */
2380 		drain_stock(stock);
2381 		css_get(&memcg->css);
2382 		WRITE_ONCE(stock->cached, memcg);
2383 	}
2384 	stock->nr_pages += nr_pages;
2385 
2386 	if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2387 		drain_stock(stock);
2388 }
2389 
2390 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2391 {
2392 	unsigned long flags;
2393 
2394 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
2395 	__refill_stock(memcg, nr_pages);
2396 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2397 }
2398 
2399 /*
2400  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2401  * of the hierarchy under it.
2402  */
2403 static void drain_all_stock(struct mem_cgroup *root_memcg)
2404 {
2405 	int cpu, curcpu;
2406 
2407 	/* If someone's already draining, avoid adding running more workers. */
2408 	if (!mutex_trylock(&percpu_charge_mutex))
2409 		return;
2410 	/*
2411 	 * Notify other cpus that system-wide "drain" is running
2412 	 * We do not care about races with the cpu hotplug because cpu down
2413 	 * as well as workers from this path always operate on the local
2414 	 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2415 	 */
2416 	migrate_disable();
2417 	curcpu = smp_processor_id();
2418 	for_each_online_cpu(cpu) {
2419 		struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2420 		struct mem_cgroup *memcg;
2421 		bool flush = false;
2422 
2423 		rcu_read_lock();
2424 		memcg = READ_ONCE(stock->cached);
2425 		if (memcg && stock->nr_pages &&
2426 		    mem_cgroup_is_descendant(memcg, root_memcg))
2427 			flush = true;
2428 		else if (obj_stock_flush_required(stock, root_memcg))
2429 			flush = true;
2430 		rcu_read_unlock();
2431 
2432 		if (flush &&
2433 		    !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2434 			if (cpu == curcpu)
2435 				drain_local_stock(&stock->work);
2436 			else if (!cpu_is_isolated(cpu))
2437 				schedule_work_on(cpu, &stock->work);
2438 		}
2439 	}
2440 	migrate_enable();
2441 	mutex_unlock(&percpu_charge_mutex);
2442 }
2443 
2444 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2445 {
2446 	struct memcg_stock_pcp *stock;
2447 
2448 	stock = &per_cpu(memcg_stock, cpu);
2449 	drain_stock(stock);
2450 
2451 	return 0;
2452 }
2453 
2454 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2455 				  unsigned int nr_pages,
2456 				  gfp_t gfp_mask)
2457 {
2458 	unsigned long nr_reclaimed = 0;
2459 
2460 	do {
2461 		unsigned long pflags;
2462 
2463 		if (page_counter_read(&memcg->memory) <=
2464 		    READ_ONCE(memcg->memory.high))
2465 			continue;
2466 
2467 		memcg_memory_event(memcg, MEMCG_HIGH);
2468 
2469 		psi_memstall_enter(&pflags);
2470 		nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2471 							gfp_mask,
2472 							MEMCG_RECLAIM_MAY_SWAP);
2473 		psi_memstall_leave(&pflags);
2474 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2475 		 !mem_cgroup_is_root(memcg));
2476 
2477 	return nr_reclaimed;
2478 }
2479 
2480 static void high_work_func(struct work_struct *work)
2481 {
2482 	struct mem_cgroup *memcg;
2483 
2484 	memcg = container_of(work, struct mem_cgroup, high_work);
2485 	reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2486 }
2487 
2488 /*
2489  * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2490  * enough to still cause a significant slowdown in most cases, while still
2491  * allowing diagnostics and tracing to proceed without becoming stuck.
2492  */
2493 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2494 
2495 /*
2496  * When calculating the delay, we use these either side of the exponentiation to
2497  * maintain precision and scale to a reasonable number of jiffies (see the table
2498  * below.
2499  *
2500  * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2501  *   overage ratio to a delay.
2502  * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2503  *   proposed penalty in order to reduce to a reasonable number of jiffies, and
2504  *   to produce a reasonable delay curve.
2505  *
2506  * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2507  * reasonable delay curve compared to precision-adjusted overage, not
2508  * penalising heavily at first, but still making sure that growth beyond the
2509  * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2510  * example, with a high of 100 megabytes:
2511  *
2512  *  +-------+------------------------+
2513  *  | usage | time to allocate in ms |
2514  *  +-------+------------------------+
2515  *  | 100M  |                      0 |
2516  *  | 101M  |                      6 |
2517  *  | 102M  |                     25 |
2518  *  | 103M  |                     57 |
2519  *  | 104M  |                    102 |
2520  *  | 105M  |                    159 |
2521  *  | 106M  |                    230 |
2522  *  | 107M  |                    313 |
2523  *  | 108M  |                    409 |
2524  *  | 109M  |                    518 |
2525  *  | 110M  |                    639 |
2526  *  | 111M  |                    774 |
2527  *  | 112M  |                    921 |
2528  *  | 113M  |                   1081 |
2529  *  | 114M  |                   1254 |
2530  *  | 115M  |                   1439 |
2531  *  | 116M  |                   1638 |
2532  *  | 117M  |                   1849 |
2533  *  | 118M  |                   2000 |
2534  *  | 119M  |                   2000 |
2535  *  | 120M  |                   2000 |
2536  *  +-------+------------------------+
2537  */
2538  #define MEMCG_DELAY_PRECISION_SHIFT 20
2539  #define MEMCG_DELAY_SCALING_SHIFT 14
2540 
2541 static u64 calculate_overage(unsigned long usage, unsigned long high)
2542 {
2543 	u64 overage;
2544 
2545 	if (usage <= high)
2546 		return 0;
2547 
2548 	/*
2549 	 * Prevent division by 0 in overage calculation by acting as if
2550 	 * it was a threshold of 1 page
2551 	 */
2552 	high = max(high, 1UL);
2553 
2554 	overage = usage - high;
2555 	overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2556 	return div64_u64(overage, high);
2557 }
2558 
2559 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2560 {
2561 	u64 overage, max_overage = 0;
2562 
2563 	do {
2564 		overage = calculate_overage(page_counter_read(&memcg->memory),
2565 					    READ_ONCE(memcg->memory.high));
2566 		max_overage = max(overage, max_overage);
2567 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2568 		 !mem_cgroup_is_root(memcg));
2569 
2570 	return max_overage;
2571 }
2572 
2573 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2574 {
2575 	u64 overage, max_overage = 0;
2576 
2577 	do {
2578 		overage = calculate_overage(page_counter_read(&memcg->swap),
2579 					    READ_ONCE(memcg->swap.high));
2580 		if (overage)
2581 			memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2582 		max_overage = max(overage, max_overage);
2583 	} while ((memcg = parent_mem_cgroup(memcg)) &&
2584 		 !mem_cgroup_is_root(memcg));
2585 
2586 	return max_overage;
2587 }
2588 
2589 /*
2590  * Get the number of jiffies that we should penalise a mischievous cgroup which
2591  * is exceeding its memory.high by checking both it and its ancestors.
2592  */
2593 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2594 					  unsigned int nr_pages,
2595 					  u64 max_overage)
2596 {
2597 	unsigned long penalty_jiffies;
2598 
2599 	if (!max_overage)
2600 		return 0;
2601 
2602 	/*
2603 	 * We use overage compared to memory.high to calculate the number of
2604 	 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2605 	 * fairly lenient on small overages, and increasingly harsh when the
2606 	 * memcg in question makes it clear that it has no intention of stopping
2607 	 * its crazy behaviour, so we exponentially increase the delay based on
2608 	 * overage amount.
2609 	 */
2610 	penalty_jiffies = max_overage * max_overage * HZ;
2611 	penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2612 	penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2613 
2614 	/*
2615 	 * Factor in the task's own contribution to the overage, such that four
2616 	 * N-sized allocations are throttled approximately the same as one
2617 	 * 4N-sized allocation.
2618 	 *
2619 	 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2620 	 * larger the current charge patch is than that.
2621 	 */
2622 	return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2623 }
2624 
2625 /*
2626  * Scheduled by try_charge() to be executed from the userland return path
2627  * and reclaims memory over the high limit.
2628  */
2629 void mem_cgroup_handle_over_high(gfp_t gfp_mask)
2630 {
2631 	unsigned long penalty_jiffies;
2632 	unsigned long pflags;
2633 	unsigned long nr_reclaimed;
2634 	unsigned int nr_pages = current->memcg_nr_pages_over_high;
2635 	int nr_retries = MAX_RECLAIM_RETRIES;
2636 	struct mem_cgroup *memcg;
2637 	bool in_retry = false;
2638 
2639 	if (likely(!nr_pages))
2640 		return;
2641 
2642 	memcg = get_mem_cgroup_from_mm(current->mm);
2643 	current->memcg_nr_pages_over_high = 0;
2644 
2645 retry_reclaim:
2646 	/*
2647 	 * The allocating task should reclaim at least the batch size, but for
2648 	 * subsequent retries we only want to do what's necessary to prevent oom
2649 	 * or breaching resource isolation.
2650 	 *
2651 	 * This is distinct from memory.max or page allocator behaviour because
2652 	 * memory.high is currently batched, whereas memory.max and the page
2653 	 * allocator run every time an allocation is made.
2654 	 */
2655 	nr_reclaimed = reclaim_high(memcg,
2656 				    in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2657 				    gfp_mask);
2658 
2659 	/*
2660 	 * memory.high is breached and reclaim is unable to keep up. Throttle
2661 	 * allocators proactively to slow down excessive growth.
2662 	 */
2663 	penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2664 					       mem_find_max_overage(memcg));
2665 
2666 	penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2667 						swap_find_max_overage(memcg));
2668 
2669 	/*
2670 	 * Clamp the max delay per usermode return so as to still keep the
2671 	 * application moving forwards and also permit diagnostics, albeit
2672 	 * extremely slowly.
2673 	 */
2674 	penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2675 
2676 	/*
2677 	 * Don't sleep if the amount of jiffies this memcg owes us is so low
2678 	 * that it's not even worth doing, in an attempt to be nice to those who
2679 	 * go only a small amount over their memory.high value and maybe haven't
2680 	 * been aggressively reclaimed enough yet.
2681 	 */
2682 	if (penalty_jiffies <= HZ / 100)
2683 		goto out;
2684 
2685 	/*
2686 	 * If reclaim is making forward progress but we're still over
2687 	 * memory.high, we want to encourage that rather than doing allocator
2688 	 * throttling.
2689 	 */
2690 	if (nr_reclaimed || nr_retries--) {
2691 		in_retry = true;
2692 		goto retry_reclaim;
2693 	}
2694 
2695 	/*
2696 	 * If we exit early, we're guaranteed to die (since
2697 	 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2698 	 * need to account for any ill-begotten jiffies to pay them off later.
2699 	 */
2700 	psi_memstall_enter(&pflags);
2701 	schedule_timeout_killable(penalty_jiffies);
2702 	psi_memstall_leave(&pflags);
2703 
2704 out:
2705 	css_put(&memcg->css);
2706 }
2707 
2708 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2709 			unsigned int nr_pages)
2710 {
2711 	unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2712 	int nr_retries = MAX_RECLAIM_RETRIES;
2713 	struct mem_cgroup *mem_over_limit;
2714 	struct page_counter *counter;
2715 	unsigned long nr_reclaimed;
2716 	bool passed_oom = false;
2717 	unsigned int reclaim_options = MEMCG_RECLAIM_MAY_SWAP;
2718 	bool drained = false;
2719 	bool raised_max_event = false;
2720 	unsigned long pflags;
2721 
2722 retry:
2723 	if (consume_stock(memcg, nr_pages))
2724 		return 0;
2725 
2726 	if (!do_memsw_account() ||
2727 	    page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2728 		if (page_counter_try_charge(&memcg->memory, batch, &counter))
2729 			goto done_restock;
2730 		if (do_memsw_account())
2731 			page_counter_uncharge(&memcg->memsw, batch);
2732 		mem_over_limit = mem_cgroup_from_counter(counter, memory);
2733 	} else {
2734 		mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2735 		reclaim_options &= ~MEMCG_RECLAIM_MAY_SWAP;
2736 	}
2737 
2738 	if (batch > nr_pages) {
2739 		batch = nr_pages;
2740 		goto retry;
2741 	}
2742 
2743 	/*
2744 	 * Prevent unbounded recursion when reclaim operations need to
2745 	 * allocate memory. This might exceed the limits temporarily,
2746 	 * but we prefer facilitating memory reclaim and getting back
2747 	 * under the limit over triggering OOM kills in these cases.
2748 	 */
2749 	if (unlikely(current->flags & PF_MEMALLOC))
2750 		goto force;
2751 
2752 	if (unlikely(task_in_memcg_oom(current)))
2753 		goto nomem;
2754 
2755 	if (!gfpflags_allow_blocking(gfp_mask))
2756 		goto nomem;
2757 
2758 	memcg_memory_event(mem_over_limit, MEMCG_MAX);
2759 	raised_max_event = true;
2760 
2761 	psi_memstall_enter(&pflags);
2762 	nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2763 						    gfp_mask, reclaim_options);
2764 	psi_memstall_leave(&pflags);
2765 
2766 	if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2767 		goto retry;
2768 
2769 	if (!drained) {
2770 		drain_all_stock(mem_over_limit);
2771 		drained = true;
2772 		goto retry;
2773 	}
2774 
2775 	if (gfp_mask & __GFP_NORETRY)
2776 		goto nomem;
2777 	/*
2778 	 * Even though the limit is exceeded at this point, reclaim
2779 	 * may have been able to free some pages.  Retry the charge
2780 	 * before killing the task.
2781 	 *
2782 	 * Only for regular pages, though: huge pages are rather
2783 	 * unlikely to succeed so close to the limit, and we fall back
2784 	 * to regular pages anyway in case of failure.
2785 	 */
2786 	if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2787 		goto retry;
2788 	/*
2789 	 * At task move, charge accounts can be doubly counted. So, it's
2790 	 * better to wait until the end of task_move if something is going on.
2791 	 */
2792 	if (mem_cgroup_wait_acct_move(mem_over_limit))
2793 		goto retry;
2794 
2795 	if (nr_retries--)
2796 		goto retry;
2797 
2798 	if (gfp_mask & __GFP_RETRY_MAYFAIL)
2799 		goto nomem;
2800 
2801 	/* Avoid endless loop for tasks bypassed by the oom killer */
2802 	if (passed_oom && task_is_dying())
2803 		goto nomem;
2804 
2805 	/*
2806 	 * keep retrying as long as the memcg oom killer is able to make
2807 	 * a forward progress or bypass the charge if the oom killer
2808 	 * couldn't make any progress.
2809 	 */
2810 	if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2811 			   get_order(nr_pages * PAGE_SIZE))) {
2812 		passed_oom = true;
2813 		nr_retries = MAX_RECLAIM_RETRIES;
2814 		goto retry;
2815 	}
2816 nomem:
2817 	/*
2818 	 * Memcg doesn't have a dedicated reserve for atomic
2819 	 * allocations. But like the global atomic pool, we need to
2820 	 * put the burden of reclaim on regular allocation requests
2821 	 * and let these go through as privileged allocations.
2822 	 */
2823 	if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2824 		return -ENOMEM;
2825 force:
2826 	/*
2827 	 * If the allocation has to be enforced, don't forget to raise
2828 	 * a MEMCG_MAX event.
2829 	 */
2830 	if (!raised_max_event)
2831 		memcg_memory_event(mem_over_limit, MEMCG_MAX);
2832 
2833 	/*
2834 	 * The allocation either can't fail or will lead to more memory
2835 	 * being freed very soon.  Allow memory usage go over the limit
2836 	 * temporarily by force charging it.
2837 	 */
2838 	page_counter_charge(&memcg->memory, nr_pages);
2839 	if (do_memsw_account())
2840 		page_counter_charge(&memcg->memsw, nr_pages);
2841 
2842 	return 0;
2843 
2844 done_restock:
2845 	if (batch > nr_pages)
2846 		refill_stock(memcg, batch - nr_pages);
2847 
2848 	/*
2849 	 * If the hierarchy is above the normal consumption range, schedule
2850 	 * reclaim on returning to userland.  We can perform reclaim here
2851 	 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2852 	 * GFP_KERNEL can consistently be used during reclaim.  @memcg is
2853 	 * not recorded as it most likely matches current's and won't
2854 	 * change in the meantime.  As high limit is checked again before
2855 	 * reclaim, the cost of mismatch is negligible.
2856 	 */
2857 	do {
2858 		bool mem_high, swap_high;
2859 
2860 		mem_high = page_counter_read(&memcg->memory) >
2861 			READ_ONCE(memcg->memory.high);
2862 		swap_high = page_counter_read(&memcg->swap) >
2863 			READ_ONCE(memcg->swap.high);
2864 
2865 		/* Don't bother a random interrupted task */
2866 		if (!in_task()) {
2867 			if (mem_high) {
2868 				schedule_work(&memcg->high_work);
2869 				break;
2870 			}
2871 			continue;
2872 		}
2873 
2874 		if (mem_high || swap_high) {
2875 			/*
2876 			 * The allocating tasks in this cgroup will need to do
2877 			 * reclaim or be throttled to prevent further growth
2878 			 * of the memory or swap footprints.
2879 			 *
2880 			 * Target some best-effort fairness between the tasks,
2881 			 * and distribute reclaim work and delay penalties
2882 			 * based on how much each task is actually allocating.
2883 			 */
2884 			current->memcg_nr_pages_over_high += batch;
2885 			set_notify_resume(current);
2886 			break;
2887 		}
2888 	} while ((memcg = parent_mem_cgroup(memcg)));
2889 
2890 	if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2891 	    !(current->flags & PF_MEMALLOC) &&
2892 	    gfpflags_allow_blocking(gfp_mask)) {
2893 		mem_cgroup_handle_over_high(gfp_mask);
2894 	}
2895 	return 0;
2896 }
2897 
2898 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2899 			     unsigned int nr_pages)
2900 {
2901 	if (mem_cgroup_is_root(memcg))
2902 		return 0;
2903 
2904 	return try_charge_memcg(memcg, gfp_mask, nr_pages);
2905 }
2906 
2907 /**
2908  * mem_cgroup_cancel_charge() - cancel an uncommitted try_charge() call.
2909  * @memcg: memcg previously charged.
2910  * @nr_pages: number of pages previously charged.
2911  */
2912 void mem_cgroup_cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2913 {
2914 	if (mem_cgroup_is_root(memcg))
2915 		return;
2916 
2917 	page_counter_uncharge(&memcg->memory, nr_pages);
2918 	if (do_memsw_account())
2919 		page_counter_uncharge(&memcg->memsw, nr_pages);
2920 }
2921 
2922 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2923 {
2924 	VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2925 	/*
2926 	 * Any of the following ensures page's memcg stability:
2927 	 *
2928 	 * - the page lock
2929 	 * - LRU isolation
2930 	 * - folio_memcg_lock()
2931 	 * - exclusive reference
2932 	 * - mem_cgroup_trylock_pages()
2933 	 */
2934 	folio->memcg_data = (unsigned long)memcg;
2935 }
2936 
2937 /**
2938  * mem_cgroup_commit_charge - commit a previously successful try_charge().
2939  * @folio: folio to commit the charge to.
2940  * @memcg: memcg previously charged.
2941  */
2942 void mem_cgroup_commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2943 {
2944 	css_get(&memcg->css);
2945 	commit_charge(folio, memcg);
2946 
2947 	local_irq_disable();
2948 	mem_cgroup_charge_statistics(memcg, folio_nr_pages(folio));
2949 	memcg_check_events(memcg, folio_nid(folio));
2950 	local_irq_enable();
2951 }
2952 
2953 #ifdef CONFIG_MEMCG_KMEM
2954 /*
2955  * The allocated objcg pointers array is not accounted directly.
2956  * Moreover, it should not come from DMA buffer and is not readily
2957  * reclaimable. So those GFP bits should be masked off.
2958  */
2959 #define OBJCGS_CLEAR_MASK	(__GFP_DMA | __GFP_RECLAIMABLE | \
2960 				 __GFP_ACCOUNT | __GFP_NOFAIL)
2961 
2962 /*
2963  * mod_objcg_mlstate() may be called with irq enabled, so
2964  * mod_memcg_lruvec_state() should be used.
2965  */
2966 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2967 				     struct pglist_data *pgdat,
2968 				     enum node_stat_item idx, int nr)
2969 {
2970 	struct mem_cgroup *memcg;
2971 	struct lruvec *lruvec;
2972 
2973 	rcu_read_lock();
2974 	memcg = obj_cgroup_memcg(objcg);
2975 	lruvec = mem_cgroup_lruvec(memcg, pgdat);
2976 	mod_memcg_lruvec_state(lruvec, idx, nr);
2977 	rcu_read_unlock();
2978 }
2979 
2980 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2981 				 gfp_t gfp, bool new_slab)
2982 {
2983 	unsigned int objects = objs_per_slab(s, slab);
2984 	unsigned long memcg_data;
2985 	void *vec;
2986 
2987 	gfp &= ~OBJCGS_CLEAR_MASK;
2988 	vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2989 			   slab_nid(slab));
2990 	if (!vec)
2991 		return -ENOMEM;
2992 
2993 	memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2994 	if (new_slab) {
2995 		/*
2996 		 * If the slab is brand new and nobody can yet access its
2997 		 * memcg_data, no synchronization is required and memcg_data can
2998 		 * be simply assigned.
2999 		 */
3000 		slab->memcg_data = memcg_data;
3001 	} else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
3002 		/*
3003 		 * If the slab is already in use, somebody can allocate and
3004 		 * assign obj_cgroups in parallel. In this case the existing
3005 		 * objcg vector should be reused.
3006 		 */
3007 		kfree(vec);
3008 		return 0;
3009 	}
3010 
3011 	kmemleak_not_leak(vec);
3012 	return 0;
3013 }
3014 
3015 static __always_inline
3016 struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
3017 {
3018 	/*
3019 	 * Slab objects are accounted individually, not per-page.
3020 	 * Memcg membership data for each individual object is saved in
3021 	 * slab->memcg_data.
3022 	 */
3023 	if (folio_test_slab(folio)) {
3024 		struct obj_cgroup **objcgs;
3025 		struct slab *slab;
3026 		unsigned int off;
3027 
3028 		slab = folio_slab(folio);
3029 		objcgs = slab_objcgs(slab);
3030 		if (!objcgs)
3031 			return NULL;
3032 
3033 		off = obj_to_index(slab->slab_cache, slab, p);
3034 		if (objcgs[off])
3035 			return obj_cgroup_memcg(objcgs[off]);
3036 
3037 		return NULL;
3038 	}
3039 
3040 	/*
3041 	 * folio_memcg_check() is used here, because in theory we can encounter
3042 	 * a folio where the slab flag has been cleared already, but
3043 	 * slab->memcg_data has not been freed yet
3044 	 * folio_memcg_check() will guarantee that a proper memory
3045 	 * cgroup pointer or NULL will be returned.
3046 	 */
3047 	return folio_memcg_check(folio);
3048 }
3049 
3050 /*
3051  * Returns a pointer to the memory cgroup to which the kernel object is charged.
3052  *
3053  * A passed kernel object can be a slab object, vmalloc object or a generic
3054  * kernel page, so different mechanisms for getting the memory cgroup pointer
3055  * should be used.
3056  *
3057  * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
3058  * can not know for sure how the kernel object is implemented.
3059  * mem_cgroup_from_obj() can be safely used in such cases.
3060  *
3061  * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3062  * cgroup_mutex, etc.
3063  */
3064 struct mem_cgroup *mem_cgroup_from_obj(void *p)
3065 {
3066 	struct folio *folio;
3067 
3068 	if (mem_cgroup_disabled())
3069 		return NULL;
3070 
3071 	if (unlikely(is_vmalloc_addr(p)))
3072 		folio = page_folio(vmalloc_to_page(p));
3073 	else
3074 		folio = virt_to_folio(p);
3075 
3076 	return mem_cgroup_from_obj_folio(folio, p);
3077 }
3078 
3079 /*
3080  * Returns a pointer to the memory cgroup to which the kernel object is charged.
3081  * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
3082  * allocated using vmalloc().
3083  *
3084  * A passed kernel object must be a slab object or a generic kernel page.
3085  *
3086  * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3087  * cgroup_mutex, etc.
3088  */
3089 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
3090 {
3091 	if (mem_cgroup_disabled())
3092 		return NULL;
3093 
3094 	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
3095 }
3096 
3097 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
3098 {
3099 	struct obj_cgroup *objcg = NULL;
3100 
3101 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3102 		objcg = rcu_dereference(memcg->objcg);
3103 		if (likely(objcg && obj_cgroup_tryget(objcg)))
3104 			break;
3105 		objcg = NULL;
3106 	}
3107 	return objcg;
3108 }
3109 
3110 static struct obj_cgroup *current_objcg_update(void)
3111 {
3112 	struct mem_cgroup *memcg;
3113 	struct obj_cgroup *old, *objcg = NULL;
3114 
3115 	do {
3116 		/* Atomically drop the update bit. */
3117 		old = xchg(&current->objcg, NULL);
3118 		if (old) {
3119 			old = (struct obj_cgroup *)
3120 				((unsigned long)old & ~CURRENT_OBJCG_UPDATE_FLAG);
3121 			if (old)
3122 				obj_cgroup_put(old);
3123 
3124 			old = NULL;
3125 		}
3126 
3127 		/* If new objcg is NULL, no reason for the second atomic update. */
3128 		if (!current->mm || (current->flags & PF_KTHREAD))
3129 			return NULL;
3130 
3131 		/*
3132 		 * Release the objcg pointer from the previous iteration,
3133 		 * if try_cmpxcg() below fails.
3134 		 */
3135 		if (unlikely(objcg)) {
3136 			obj_cgroup_put(objcg);
3137 			objcg = NULL;
3138 		}
3139 
3140 		/*
3141 		 * Obtain the new objcg pointer. The current task can be
3142 		 * asynchronously moved to another memcg and the previous
3143 		 * memcg can be offlined. So let's get the memcg pointer
3144 		 * and try get a reference to objcg under a rcu read lock.
3145 		 */
3146 
3147 		rcu_read_lock();
3148 		memcg = mem_cgroup_from_task(current);
3149 		objcg = __get_obj_cgroup_from_memcg(memcg);
3150 		rcu_read_unlock();
3151 
3152 		/*
3153 		 * Try set up a new objcg pointer atomically. If it
3154 		 * fails, it means the update flag was set concurrently, so
3155 		 * the whole procedure should be repeated.
3156 		 */
3157 	} while (!try_cmpxchg(&current->objcg, &old, objcg));
3158 
3159 	return objcg;
3160 }
3161 
3162 __always_inline struct obj_cgroup *current_obj_cgroup(void)
3163 {
3164 	struct mem_cgroup *memcg;
3165 	struct obj_cgroup *objcg;
3166 
3167 	if (in_task()) {
3168 		memcg = current->active_memcg;
3169 		if (unlikely(memcg))
3170 			goto from_memcg;
3171 
3172 		objcg = READ_ONCE(current->objcg);
3173 		if (unlikely((unsigned long)objcg & CURRENT_OBJCG_UPDATE_FLAG))
3174 			objcg = current_objcg_update();
3175 		/*
3176 		 * Objcg reference is kept by the task, so it's safe
3177 		 * to use the objcg by the current task.
3178 		 */
3179 		return objcg;
3180 	}
3181 
3182 	memcg = this_cpu_read(int_active_memcg);
3183 	if (unlikely(memcg))
3184 		goto from_memcg;
3185 
3186 	return NULL;
3187 
3188 from_memcg:
3189 	objcg = NULL;
3190 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3191 		/*
3192 		 * Memcg pointer is protected by scope (see set_active_memcg())
3193 		 * and is pinning the corresponding objcg, so objcg can't go
3194 		 * away and can be used within the scope without any additional
3195 		 * protection.
3196 		 */
3197 		objcg = rcu_dereference_check(memcg->objcg, 1);
3198 		if (likely(objcg))
3199 			break;
3200 	}
3201 
3202 	return objcg;
3203 }
3204 
3205 struct obj_cgroup *get_obj_cgroup_from_folio(struct folio *folio)
3206 {
3207 	struct obj_cgroup *objcg;
3208 
3209 	if (!memcg_kmem_online())
3210 		return NULL;
3211 
3212 	if (folio_memcg_kmem(folio)) {
3213 		objcg = __folio_objcg(folio);
3214 		obj_cgroup_get(objcg);
3215 	} else {
3216 		struct mem_cgroup *memcg;
3217 
3218 		rcu_read_lock();
3219 		memcg = __folio_memcg(folio);
3220 		if (memcg)
3221 			objcg = __get_obj_cgroup_from_memcg(memcg);
3222 		else
3223 			objcg = NULL;
3224 		rcu_read_unlock();
3225 	}
3226 	return objcg;
3227 }
3228 
3229 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
3230 {
3231 	mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
3232 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
3233 		if (nr_pages > 0)
3234 			page_counter_charge(&memcg->kmem, nr_pages);
3235 		else
3236 			page_counter_uncharge(&memcg->kmem, -nr_pages);
3237 	}
3238 }
3239 
3240 
3241 /*
3242  * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3243  * @objcg: object cgroup to uncharge
3244  * @nr_pages: number of pages to uncharge
3245  */
3246 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3247 				      unsigned int nr_pages)
3248 {
3249 	struct mem_cgroup *memcg;
3250 
3251 	memcg = get_mem_cgroup_from_objcg(objcg);
3252 
3253 	memcg_account_kmem(memcg, -nr_pages);
3254 	refill_stock(memcg, nr_pages);
3255 
3256 	css_put(&memcg->css);
3257 }
3258 
3259 /*
3260  * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3261  * @objcg: object cgroup to charge
3262  * @gfp: reclaim mode
3263  * @nr_pages: number of pages to charge
3264  *
3265  * Returns 0 on success, an error code on failure.
3266  */
3267 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3268 				   unsigned int nr_pages)
3269 {
3270 	struct mem_cgroup *memcg;
3271 	int ret;
3272 
3273 	memcg = get_mem_cgroup_from_objcg(objcg);
3274 
3275 	ret = try_charge_memcg(memcg, gfp, nr_pages);
3276 	if (ret)
3277 		goto out;
3278 
3279 	memcg_account_kmem(memcg, nr_pages);
3280 out:
3281 	css_put(&memcg->css);
3282 
3283 	return ret;
3284 }
3285 
3286 /**
3287  * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3288  * @page: page to charge
3289  * @gfp: reclaim mode
3290  * @order: allocation order
3291  *
3292  * Returns 0 on success, an error code on failure.
3293  */
3294 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3295 {
3296 	struct obj_cgroup *objcg;
3297 	int ret = 0;
3298 
3299 	objcg = current_obj_cgroup();
3300 	if (objcg) {
3301 		ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3302 		if (!ret) {
3303 			obj_cgroup_get(objcg);
3304 			page->memcg_data = (unsigned long)objcg |
3305 				MEMCG_DATA_KMEM;
3306 			return 0;
3307 		}
3308 	}
3309 	return ret;
3310 }
3311 
3312 /**
3313  * __memcg_kmem_uncharge_page: uncharge a kmem page
3314  * @page: page to uncharge
3315  * @order: allocation order
3316  */
3317 void __memcg_kmem_uncharge_page(struct page *page, int order)
3318 {
3319 	struct folio *folio = page_folio(page);
3320 	struct obj_cgroup *objcg;
3321 	unsigned int nr_pages = 1 << order;
3322 
3323 	if (!folio_memcg_kmem(folio))
3324 		return;
3325 
3326 	objcg = __folio_objcg(folio);
3327 	obj_cgroup_uncharge_pages(objcg, nr_pages);
3328 	folio->memcg_data = 0;
3329 	obj_cgroup_put(objcg);
3330 }
3331 
3332 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3333 		     enum node_stat_item idx, int nr)
3334 {
3335 	struct memcg_stock_pcp *stock;
3336 	struct obj_cgroup *old = NULL;
3337 	unsigned long flags;
3338 	int *bytes;
3339 
3340 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
3341 	stock = this_cpu_ptr(&memcg_stock);
3342 
3343 	/*
3344 	 * Save vmstat data in stock and skip vmstat array update unless
3345 	 * accumulating over a page of vmstat data or when pgdat or idx
3346 	 * changes.
3347 	 */
3348 	if (READ_ONCE(stock->cached_objcg) != objcg) {
3349 		old = drain_obj_stock(stock);
3350 		obj_cgroup_get(objcg);
3351 		stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3352 				? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3353 		WRITE_ONCE(stock->cached_objcg, objcg);
3354 		stock->cached_pgdat = pgdat;
3355 	} else if (stock->cached_pgdat != pgdat) {
3356 		/* Flush the existing cached vmstat data */
3357 		struct pglist_data *oldpg = stock->cached_pgdat;
3358 
3359 		if (stock->nr_slab_reclaimable_b) {
3360 			mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3361 					  stock->nr_slab_reclaimable_b);
3362 			stock->nr_slab_reclaimable_b = 0;
3363 		}
3364 		if (stock->nr_slab_unreclaimable_b) {
3365 			mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3366 					  stock->nr_slab_unreclaimable_b);
3367 			stock->nr_slab_unreclaimable_b = 0;
3368 		}
3369 		stock->cached_pgdat = pgdat;
3370 	}
3371 
3372 	bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3373 					       : &stock->nr_slab_unreclaimable_b;
3374 	/*
3375 	 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3376 	 * cached locally at least once before pushing it out.
3377 	 */
3378 	if (!*bytes) {
3379 		*bytes = nr;
3380 		nr = 0;
3381 	} else {
3382 		*bytes += nr;
3383 		if (abs(*bytes) > PAGE_SIZE) {
3384 			nr = *bytes;
3385 			*bytes = 0;
3386 		} else {
3387 			nr = 0;
3388 		}
3389 	}
3390 	if (nr)
3391 		mod_objcg_mlstate(objcg, pgdat, idx, nr);
3392 
3393 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3394 	if (old)
3395 		obj_cgroup_put(old);
3396 }
3397 
3398 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3399 {
3400 	struct memcg_stock_pcp *stock;
3401 	unsigned long flags;
3402 	bool ret = false;
3403 
3404 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
3405 
3406 	stock = this_cpu_ptr(&memcg_stock);
3407 	if (objcg == READ_ONCE(stock->cached_objcg) && stock->nr_bytes >= nr_bytes) {
3408 		stock->nr_bytes -= nr_bytes;
3409 		ret = true;
3410 	}
3411 
3412 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3413 
3414 	return ret;
3415 }
3416 
3417 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3418 {
3419 	struct obj_cgroup *old = READ_ONCE(stock->cached_objcg);
3420 
3421 	if (!old)
3422 		return NULL;
3423 
3424 	if (stock->nr_bytes) {
3425 		unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3426 		unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3427 
3428 		if (nr_pages) {
3429 			struct mem_cgroup *memcg;
3430 
3431 			memcg = get_mem_cgroup_from_objcg(old);
3432 
3433 			memcg_account_kmem(memcg, -nr_pages);
3434 			__refill_stock(memcg, nr_pages);
3435 
3436 			css_put(&memcg->css);
3437 		}
3438 
3439 		/*
3440 		 * The leftover is flushed to the centralized per-memcg value.
3441 		 * On the next attempt to refill obj stock it will be moved
3442 		 * to a per-cpu stock (probably, on an other CPU), see
3443 		 * refill_obj_stock().
3444 		 *
3445 		 * How often it's flushed is a trade-off between the memory
3446 		 * limit enforcement accuracy and potential CPU contention,
3447 		 * so it might be changed in the future.
3448 		 */
3449 		atomic_add(nr_bytes, &old->nr_charged_bytes);
3450 		stock->nr_bytes = 0;
3451 	}
3452 
3453 	/*
3454 	 * Flush the vmstat data in current stock
3455 	 */
3456 	if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3457 		if (stock->nr_slab_reclaimable_b) {
3458 			mod_objcg_mlstate(old, stock->cached_pgdat,
3459 					  NR_SLAB_RECLAIMABLE_B,
3460 					  stock->nr_slab_reclaimable_b);
3461 			stock->nr_slab_reclaimable_b = 0;
3462 		}
3463 		if (stock->nr_slab_unreclaimable_b) {
3464 			mod_objcg_mlstate(old, stock->cached_pgdat,
3465 					  NR_SLAB_UNRECLAIMABLE_B,
3466 					  stock->nr_slab_unreclaimable_b);
3467 			stock->nr_slab_unreclaimable_b = 0;
3468 		}
3469 		stock->cached_pgdat = NULL;
3470 	}
3471 
3472 	WRITE_ONCE(stock->cached_objcg, NULL);
3473 	/*
3474 	 * The `old' objects needs to be released by the caller via
3475 	 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3476 	 */
3477 	return old;
3478 }
3479 
3480 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3481 				     struct mem_cgroup *root_memcg)
3482 {
3483 	struct obj_cgroup *objcg = READ_ONCE(stock->cached_objcg);
3484 	struct mem_cgroup *memcg;
3485 
3486 	if (objcg) {
3487 		memcg = obj_cgroup_memcg(objcg);
3488 		if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3489 			return true;
3490 	}
3491 
3492 	return false;
3493 }
3494 
3495 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3496 			     bool allow_uncharge)
3497 {
3498 	struct memcg_stock_pcp *stock;
3499 	struct obj_cgroup *old = NULL;
3500 	unsigned long flags;
3501 	unsigned int nr_pages = 0;
3502 
3503 	local_lock_irqsave(&memcg_stock.stock_lock, flags);
3504 
3505 	stock = this_cpu_ptr(&memcg_stock);
3506 	if (READ_ONCE(stock->cached_objcg) != objcg) { /* reset if necessary */
3507 		old = drain_obj_stock(stock);
3508 		obj_cgroup_get(objcg);
3509 		WRITE_ONCE(stock->cached_objcg, objcg);
3510 		stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3511 				? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3512 		allow_uncharge = true;	/* Allow uncharge when objcg changes */
3513 	}
3514 	stock->nr_bytes += nr_bytes;
3515 
3516 	if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3517 		nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3518 		stock->nr_bytes &= (PAGE_SIZE - 1);
3519 	}
3520 
3521 	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3522 	if (old)
3523 		obj_cgroup_put(old);
3524 
3525 	if (nr_pages)
3526 		obj_cgroup_uncharge_pages(objcg, nr_pages);
3527 }
3528 
3529 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3530 {
3531 	unsigned int nr_pages, nr_bytes;
3532 	int ret;
3533 
3534 	if (consume_obj_stock(objcg, size))
3535 		return 0;
3536 
3537 	/*
3538 	 * In theory, objcg->nr_charged_bytes can have enough
3539 	 * pre-charged bytes to satisfy the allocation. However,
3540 	 * flushing objcg->nr_charged_bytes requires two atomic
3541 	 * operations, and objcg->nr_charged_bytes can't be big.
3542 	 * The shared objcg->nr_charged_bytes can also become a
3543 	 * performance bottleneck if all tasks of the same memcg are
3544 	 * trying to update it. So it's better to ignore it and try
3545 	 * grab some new pages. The stock's nr_bytes will be flushed to
3546 	 * objcg->nr_charged_bytes later on when objcg changes.
3547 	 *
3548 	 * The stock's nr_bytes may contain enough pre-charged bytes
3549 	 * to allow one less page from being charged, but we can't rely
3550 	 * on the pre-charged bytes not being changed outside of
3551 	 * consume_obj_stock() or refill_obj_stock(). So ignore those
3552 	 * pre-charged bytes as well when charging pages. To avoid a
3553 	 * page uncharge right after a page charge, we set the
3554 	 * allow_uncharge flag to false when calling refill_obj_stock()
3555 	 * to temporarily allow the pre-charged bytes to exceed the page
3556 	 * size limit. The maximum reachable value of the pre-charged
3557 	 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3558 	 * race.
3559 	 */
3560 	nr_pages = size >> PAGE_SHIFT;
3561 	nr_bytes = size & (PAGE_SIZE - 1);
3562 
3563 	if (nr_bytes)
3564 		nr_pages += 1;
3565 
3566 	ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3567 	if (!ret && nr_bytes)
3568 		refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3569 
3570 	return ret;
3571 }
3572 
3573 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3574 {
3575 	refill_obj_stock(objcg, size, true);
3576 }
3577 
3578 #endif /* CONFIG_MEMCG_KMEM */
3579 
3580 /*
3581  * Because page_memcg(head) is not set on tails, set it now.
3582  */
3583 void split_page_memcg(struct page *head, unsigned int nr)
3584 {
3585 	struct folio *folio = page_folio(head);
3586 	struct mem_cgroup *memcg = folio_memcg(folio);
3587 	int i;
3588 
3589 	if (mem_cgroup_disabled() || !memcg)
3590 		return;
3591 
3592 	for (i = 1; i < nr; i++)
3593 		folio_page(folio, i)->memcg_data = folio->memcg_data;
3594 
3595 	if (folio_memcg_kmem(folio))
3596 		obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3597 	else
3598 		css_get_many(&memcg->css, nr - 1);
3599 }
3600 
3601 #ifdef CONFIG_SWAP
3602 /**
3603  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3604  * @entry: swap entry to be moved
3605  * @from:  mem_cgroup which the entry is moved from
3606  * @to:  mem_cgroup which the entry is moved to
3607  *
3608  * It succeeds only when the swap_cgroup's record for this entry is the same
3609  * as the mem_cgroup's id of @from.
3610  *
3611  * Returns 0 on success, -EINVAL on failure.
3612  *
3613  * The caller must have charged to @to, IOW, called page_counter_charge() about
3614  * both res and memsw, and called css_get().
3615  */
3616 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3617 				struct mem_cgroup *from, struct mem_cgroup *to)
3618 {
3619 	unsigned short old_id, new_id;
3620 
3621 	old_id = mem_cgroup_id(from);
3622 	new_id = mem_cgroup_id(to);
3623 
3624 	if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3625 		mod_memcg_state(from, MEMCG_SWAP, -1);
3626 		mod_memcg_state(to, MEMCG_SWAP, 1);
3627 		return 0;
3628 	}
3629 	return -EINVAL;
3630 }
3631 #else
3632 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3633 				struct mem_cgroup *from, struct mem_cgroup *to)
3634 {
3635 	return -EINVAL;
3636 }
3637 #endif
3638 
3639 static DEFINE_MUTEX(memcg_max_mutex);
3640 
3641 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3642 				 unsigned long max, bool memsw)
3643 {
3644 	bool enlarge = false;
3645 	bool drained = false;
3646 	int ret;
3647 	bool limits_invariant;
3648 	struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3649 
3650 	do {
3651 		if (signal_pending(current)) {
3652 			ret = -EINTR;
3653 			break;
3654 		}
3655 
3656 		mutex_lock(&memcg_max_mutex);
3657 		/*
3658 		 * Make sure that the new limit (memsw or memory limit) doesn't
3659 		 * break our basic invariant rule memory.max <= memsw.max.
3660 		 */
3661 		limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3662 					   max <= memcg->memsw.max;
3663 		if (!limits_invariant) {
3664 			mutex_unlock(&memcg_max_mutex);
3665 			ret = -EINVAL;
3666 			break;
3667 		}
3668 		if (max > counter->max)
3669 			enlarge = true;
3670 		ret = page_counter_set_max(counter, max);
3671 		mutex_unlock(&memcg_max_mutex);
3672 
3673 		if (!ret)
3674 			break;
3675 
3676 		if (!drained) {
3677 			drain_all_stock(memcg);
3678 			drained = true;
3679 			continue;
3680 		}
3681 
3682 		if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3683 					memsw ? 0 : MEMCG_RECLAIM_MAY_SWAP)) {
3684 			ret = -EBUSY;
3685 			break;
3686 		}
3687 	} while (true);
3688 
3689 	if (!ret && enlarge)
3690 		memcg_oom_recover(memcg);
3691 
3692 	return ret;
3693 }
3694 
3695 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3696 					    gfp_t gfp_mask,
3697 					    unsigned long *total_scanned)
3698 {
3699 	unsigned long nr_reclaimed = 0;
3700 	struct mem_cgroup_per_node *mz, *next_mz = NULL;
3701 	unsigned long reclaimed;
3702 	int loop = 0;
3703 	struct mem_cgroup_tree_per_node *mctz;
3704 	unsigned long excess;
3705 
3706 	if (lru_gen_enabled())
3707 		return 0;
3708 
3709 	if (order > 0)
3710 		return 0;
3711 
3712 	mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3713 
3714 	/*
3715 	 * Do not even bother to check the largest node if the root
3716 	 * is empty. Do it lockless to prevent lock bouncing. Races
3717 	 * are acceptable as soft limit is best effort anyway.
3718 	 */
3719 	if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3720 		return 0;
3721 
3722 	/*
3723 	 * This loop can run a while, specially if mem_cgroup's continuously
3724 	 * keep exceeding their soft limit and putting the system under
3725 	 * pressure
3726 	 */
3727 	do {
3728 		if (next_mz)
3729 			mz = next_mz;
3730 		else
3731 			mz = mem_cgroup_largest_soft_limit_node(mctz);
3732 		if (!mz)
3733 			break;
3734 
3735 		reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3736 						    gfp_mask, total_scanned);
3737 		nr_reclaimed += reclaimed;
3738 		spin_lock_irq(&mctz->lock);
3739 
3740 		/*
3741 		 * If we failed to reclaim anything from this memory cgroup
3742 		 * it is time to move on to the next cgroup
3743 		 */
3744 		next_mz = NULL;
3745 		if (!reclaimed)
3746 			next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3747 
3748 		excess = soft_limit_excess(mz->memcg);
3749 		/*
3750 		 * One school of thought says that we should not add
3751 		 * back the node to the tree if reclaim returns 0.
3752 		 * But our reclaim could return 0, simply because due
3753 		 * to priority we are exposing a smaller subset of
3754 		 * memory to reclaim from. Consider this as a longer
3755 		 * term TODO.
3756 		 */
3757 		/* If excess == 0, no tree ops */
3758 		__mem_cgroup_insert_exceeded(mz, mctz, excess);
3759 		spin_unlock_irq(&mctz->lock);
3760 		css_put(&mz->memcg->css);
3761 		loop++;
3762 		/*
3763 		 * Could not reclaim anything and there are no more
3764 		 * mem cgroups to try or we seem to be looping without
3765 		 * reclaiming anything.
3766 		 */
3767 		if (!nr_reclaimed &&
3768 			(next_mz == NULL ||
3769 			loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3770 			break;
3771 	} while (!nr_reclaimed);
3772 	if (next_mz)
3773 		css_put(&next_mz->memcg->css);
3774 	return nr_reclaimed;
3775 }
3776 
3777 /*
3778  * Reclaims as many pages from the given memcg as possible.
3779  *
3780  * Caller is responsible for holding css reference for memcg.
3781  */
3782 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3783 {
3784 	int nr_retries = MAX_RECLAIM_RETRIES;
3785 
3786 	/* we call try-to-free pages for make this cgroup empty */
3787 	lru_add_drain_all();
3788 
3789 	drain_all_stock(memcg);
3790 
3791 	/* try to free all pages in this cgroup */
3792 	while (nr_retries && page_counter_read(&memcg->memory)) {
3793 		if (signal_pending(current))
3794 			return -EINTR;
3795 
3796 		if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3797 						  MEMCG_RECLAIM_MAY_SWAP))
3798 			nr_retries--;
3799 	}
3800 
3801 	return 0;
3802 }
3803 
3804 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3805 					    char *buf, size_t nbytes,
3806 					    loff_t off)
3807 {
3808 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3809 
3810 	if (mem_cgroup_is_root(memcg))
3811 		return -EINVAL;
3812 	return mem_cgroup_force_empty(memcg) ?: nbytes;
3813 }
3814 
3815 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3816 				     struct cftype *cft)
3817 {
3818 	return 1;
3819 }
3820 
3821 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3822 				      struct cftype *cft, u64 val)
3823 {
3824 	if (val == 1)
3825 		return 0;
3826 
3827 	pr_warn_once("Non-hierarchical mode is deprecated. "
3828 		     "Please report your usecase to linux-mm@kvack.org if you "
3829 		     "depend on this functionality.\n");
3830 
3831 	return -EINVAL;
3832 }
3833 
3834 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3835 {
3836 	unsigned long val;
3837 
3838 	if (mem_cgroup_is_root(memcg)) {
3839 		/*
3840 		 * Approximate root's usage from global state. This isn't
3841 		 * perfect, but the root usage was always an approximation.
3842 		 */
3843 		val = global_node_page_state(NR_FILE_PAGES) +
3844 			global_node_page_state(NR_ANON_MAPPED);
3845 		if (swap)
3846 			val += total_swap_pages - get_nr_swap_pages();
3847 	} else {
3848 		if (!swap)
3849 			val = page_counter_read(&memcg->memory);
3850 		else
3851 			val = page_counter_read(&memcg->memsw);
3852 	}
3853 	return val;
3854 }
3855 
3856 enum {
3857 	RES_USAGE,
3858 	RES_LIMIT,
3859 	RES_MAX_USAGE,
3860 	RES_FAILCNT,
3861 	RES_SOFT_LIMIT,
3862 };
3863 
3864 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3865 			       struct cftype *cft)
3866 {
3867 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3868 	struct page_counter *counter;
3869 
3870 	switch (MEMFILE_TYPE(cft->private)) {
3871 	case _MEM:
3872 		counter = &memcg->memory;
3873 		break;
3874 	case _MEMSWAP:
3875 		counter = &memcg->memsw;
3876 		break;
3877 	case _KMEM:
3878 		counter = &memcg->kmem;
3879 		break;
3880 	case _TCP:
3881 		counter = &memcg->tcpmem;
3882 		break;
3883 	default:
3884 		BUG();
3885 	}
3886 
3887 	switch (MEMFILE_ATTR(cft->private)) {
3888 	case RES_USAGE:
3889 		if (counter == &memcg->memory)
3890 			return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3891 		if (counter == &memcg->memsw)
3892 			return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3893 		return (u64)page_counter_read(counter) * PAGE_SIZE;
3894 	case RES_LIMIT:
3895 		return (u64)counter->max * PAGE_SIZE;
3896 	case RES_MAX_USAGE:
3897 		return (u64)counter->watermark * PAGE_SIZE;
3898 	case RES_FAILCNT:
3899 		return counter->failcnt;
3900 	case RES_SOFT_LIMIT:
3901 		return (u64)READ_ONCE(memcg->soft_limit) * PAGE_SIZE;
3902 	default:
3903 		BUG();
3904 	}
3905 }
3906 
3907 /*
3908  * This function doesn't do anything useful. Its only job is to provide a read
3909  * handler for a file so that cgroup_file_mode() will add read permissions.
3910  */
3911 static int mem_cgroup_dummy_seq_show(__always_unused struct seq_file *m,
3912 				     __always_unused void *v)
3913 {
3914 	return -EINVAL;
3915 }
3916 
3917 #ifdef CONFIG_MEMCG_KMEM
3918 static int memcg_online_kmem(struct mem_cgroup *memcg)
3919 {
3920 	struct obj_cgroup *objcg;
3921 
3922 	if (mem_cgroup_kmem_disabled())
3923 		return 0;
3924 
3925 	if (unlikely(mem_cgroup_is_root(memcg)))
3926 		return 0;
3927 
3928 	objcg = obj_cgroup_alloc();
3929 	if (!objcg)
3930 		return -ENOMEM;
3931 
3932 	objcg->memcg = memcg;
3933 	rcu_assign_pointer(memcg->objcg, objcg);
3934 	obj_cgroup_get(objcg);
3935 	memcg->orig_objcg = objcg;
3936 
3937 	static_branch_enable(&memcg_kmem_online_key);
3938 
3939 	memcg->kmemcg_id = memcg->id.id;
3940 
3941 	return 0;
3942 }
3943 
3944 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3945 {
3946 	struct mem_cgroup *parent;
3947 
3948 	if (mem_cgroup_kmem_disabled())
3949 		return;
3950 
3951 	if (unlikely(mem_cgroup_is_root(memcg)))
3952 		return;
3953 
3954 	parent = parent_mem_cgroup(memcg);
3955 	if (!parent)
3956 		parent = root_mem_cgroup;
3957 
3958 	memcg_reparent_objcgs(memcg, parent);
3959 
3960 	/*
3961 	 * After we have finished memcg_reparent_objcgs(), all list_lrus
3962 	 * corresponding to this cgroup are guaranteed to remain empty.
3963 	 * The ordering is imposed by list_lru_node->lock taken by
3964 	 * memcg_reparent_list_lrus().
3965 	 */
3966 	memcg_reparent_list_lrus(memcg, parent);
3967 }
3968 #else
3969 static int memcg_online_kmem(struct mem_cgroup *memcg)
3970 {
3971 	return 0;
3972 }
3973 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3974 {
3975 }
3976 #endif /* CONFIG_MEMCG_KMEM */
3977 
3978 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3979 {
3980 	int ret;
3981 
3982 	mutex_lock(&memcg_max_mutex);
3983 
3984 	ret = page_counter_set_max(&memcg->tcpmem, max);
3985 	if (ret)
3986 		goto out;
3987 
3988 	if (!memcg->tcpmem_active) {
3989 		/*
3990 		 * The active flag needs to be written after the static_key
3991 		 * update. This is what guarantees that the socket activation
3992 		 * function is the last one to run. See mem_cgroup_sk_alloc()
3993 		 * for details, and note that we don't mark any socket as
3994 		 * belonging to this memcg until that flag is up.
3995 		 *
3996 		 * We need to do this, because static_keys will span multiple
3997 		 * sites, but we can't control their order. If we mark a socket
3998 		 * as accounted, but the accounting functions are not patched in
3999 		 * yet, we'll lose accounting.
4000 		 *
4001 		 * We never race with the readers in mem_cgroup_sk_alloc(),
4002 		 * because when this value change, the code to process it is not
4003 		 * patched in yet.
4004 		 */
4005 		static_branch_inc(&memcg_sockets_enabled_key);
4006 		memcg->tcpmem_active = true;
4007 	}
4008 out:
4009 	mutex_unlock(&memcg_max_mutex);
4010 	return ret;
4011 }
4012 
4013 /*
4014  * The user of this function is...
4015  * RES_LIMIT.
4016  */
4017 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4018 				char *buf, size_t nbytes, loff_t off)
4019 {
4020 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4021 	unsigned long nr_pages;
4022 	int ret;
4023 
4024 	buf = strstrip(buf);
4025 	ret = page_counter_memparse(buf, "-1", &nr_pages);
4026 	if (ret)
4027 		return ret;
4028 
4029 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
4030 	case RES_LIMIT:
4031 		if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4032 			ret = -EINVAL;
4033 			break;
4034 		}
4035 		switch (MEMFILE_TYPE(of_cft(of)->private)) {
4036 		case _MEM:
4037 			ret = mem_cgroup_resize_max(memcg, nr_pages, false);
4038 			break;
4039 		case _MEMSWAP:
4040 			ret = mem_cgroup_resize_max(memcg, nr_pages, true);
4041 			break;
4042 		case _KMEM:
4043 			pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
4044 				     "Writing any value to this file has no effect. "
4045 				     "Please report your usecase to linux-mm@kvack.org if you "
4046 				     "depend on this functionality.\n");
4047 			ret = 0;
4048 			break;
4049 		case _TCP:
4050 			ret = memcg_update_tcp_max(memcg, nr_pages);
4051 			break;
4052 		}
4053 		break;
4054 	case RES_SOFT_LIMIT:
4055 		if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4056 			ret = -EOPNOTSUPP;
4057 		} else {
4058 			WRITE_ONCE(memcg->soft_limit, nr_pages);
4059 			ret = 0;
4060 		}
4061 		break;
4062 	}
4063 	return ret ?: nbytes;
4064 }
4065 
4066 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4067 				size_t nbytes, loff_t off)
4068 {
4069 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4070 	struct page_counter *counter;
4071 
4072 	switch (MEMFILE_TYPE(of_cft(of)->private)) {
4073 	case _MEM:
4074 		counter = &memcg->memory;
4075 		break;
4076 	case _MEMSWAP:
4077 		counter = &memcg->memsw;
4078 		break;
4079 	case _KMEM:
4080 		counter = &memcg->kmem;
4081 		break;
4082 	case _TCP:
4083 		counter = &memcg->tcpmem;
4084 		break;
4085 	default:
4086 		BUG();
4087 	}
4088 
4089 	switch (MEMFILE_ATTR(of_cft(of)->private)) {
4090 	case RES_MAX_USAGE:
4091 		page_counter_reset_watermark(counter);
4092 		break;
4093 	case RES_FAILCNT:
4094 		counter->failcnt = 0;
4095 		break;
4096 	default:
4097 		BUG();
4098 	}
4099 
4100 	return nbytes;
4101 }
4102 
4103 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4104 					struct cftype *cft)
4105 {
4106 	return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4107 }
4108 
4109 #ifdef CONFIG_MMU
4110 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4111 					struct cftype *cft, u64 val)
4112 {
4113 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4114 
4115 	pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
4116 		     "Please report your usecase to linux-mm@kvack.org if you "
4117 		     "depend on this functionality.\n");
4118 
4119 	if (val & ~MOVE_MASK)
4120 		return -EINVAL;
4121 
4122 	/*
4123 	 * No kind of locking is needed in here, because ->can_attach() will
4124 	 * check this value once in the beginning of the process, and then carry
4125 	 * on with stale data. This means that changes to this value will only
4126 	 * affect task migrations starting after the change.
4127 	 */
4128 	memcg->move_charge_at_immigrate = val;
4129 	return 0;
4130 }
4131 #else
4132 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4133 					struct cftype *cft, u64 val)
4134 {
4135 	return -ENOSYS;
4136 }
4137 #endif
4138 
4139 #ifdef CONFIG_NUMA
4140 
4141 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
4142 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
4143 #define LRU_ALL	     ((1 << NR_LRU_LISTS) - 1)
4144 
4145 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
4146 				int nid, unsigned int lru_mask, bool tree)
4147 {
4148 	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
4149 	unsigned long nr = 0;
4150 	enum lru_list lru;
4151 
4152 	VM_BUG_ON((unsigned)nid >= nr_node_ids);
4153 
4154 	for_each_lru(lru) {
4155 		if (!(BIT(lru) & lru_mask))
4156 			continue;
4157 		if (tree)
4158 			nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
4159 		else
4160 			nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4161 	}
4162 	return nr;
4163 }
4164 
4165 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4166 					     unsigned int lru_mask,
4167 					     bool tree)
4168 {
4169 	unsigned long nr = 0;
4170 	enum lru_list lru;
4171 
4172 	for_each_lru(lru) {
4173 		if (!(BIT(lru) & lru_mask))
4174 			continue;
4175 		if (tree)
4176 			nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4177 		else
4178 			nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4179 	}
4180 	return nr;
4181 }
4182 
4183 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4184 {
4185 	struct numa_stat {
4186 		const char *name;
4187 		unsigned int lru_mask;
4188 	};
4189 
4190 	static const struct numa_stat stats[] = {
4191 		{ "total", LRU_ALL },
4192 		{ "file", LRU_ALL_FILE },
4193 		{ "anon", LRU_ALL_ANON },
4194 		{ "unevictable", BIT(LRU_UNEVICTABLE) },
4195 	};
4196 	const struct numa_stat *stat;
4197 	int nid;
4198 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4199 
4200 	mem_cgroup_flush_stats(memcg);
4201 
4202 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4203 		seq_printf(m, "%s=%lu", stat->name,
4204 			   mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4205 						   false));
4206 		for_each_node_state(nid, N_MEMORY)
4207 			seq_printf(m, " N%d=%lu", nid,
4208 				   mem_cgroup_node_nr_lru_pages(memcg, nid,
4209 							stat->lru_mask, false));
4210 		seq_putc(m, '\n');
4211 	}
4212 
4213 	for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4214 
4215 		seq_printf(m, "hierarchical_%s=%lu", stat->name,
4216 			   mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4217 						   true));
4218 		for_each_node_state(nid, N_MEMORY)
4219 			seq_printf(m, " N%d=%lu", nid,
4220 				   mem_cgroup_node_nr_lru_pages(memcg, nid,
4221 							stat->lru_mask, true));
4222 		seq_putc(m, '\n');
4223 	}
4224 
4225 	return 0;
4226 }
4227 #endif /* CONFIG_NUMA */
4228 
4229 static const unsigned int memcg1_stats[] = {
4230 	NR_FILE_PAGES,
4231 	NR_ANON_MAPPED,
4232 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4233 	NR_ANON_THPS,
4234 #endif
4235 	NR_SHMEM,
4236 	NR_FILE_MAPPED,
4237 	NR_FILE_DIRTY,
4238 	NR_WRITEBACK,
4239 	WORKINGSET_REFAULT_ANON,
4240 	WORKINGSET_REFAULT_FILE,
4241 #ifdef CONFIG_SWAP
4242 	MEMCG_SWAP,
4243 	NR_SWAPCACHE,
4244 #endif
4245 };
4246 
4247 static const char *const memcg1_stat_names[] = {
4248 	"cache",
4249 	"rss",
4250 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4251 	"rss_huge",
4252 #endif
4253 	"shmem",
4254 	"mapped_file",
4255 	"dirty",
4256 	"writeback",
4257 	"workingset_refault_anon",
4258 	"workingset_refault_file",
4259 #ifdef CONFIG_SWAP
4260 	"swap",
4261 	"swapcached",
4262 #endif
4263 };
4264 
4265 /* Universal VM events cgroup1 shows, original sort order */
4266 static const unsigned int memcg1_events[] = {
4267 	PGPGIN,
4268 	PGPGOUT,
4269 	PGFAULT,
4270 	PGMAJFAULT,
4271 };
4272 
4273 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
4274 {
4275 	unsigned long memory, memsw;
4276 	struct mem_cgroup *mi;
4277 	unsigned int i;
4278 
4279 	BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4280 
4281 	mem_cgroup_flush_stats(memcg);
4282 
4283 	for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4284 		unsigned long nr;
4285 
4286 		nr = memcg_page_state_local_output(memcg, memcg1_stats[i]);
4287 		seq_buf_printf(s, "%s %lu\n", memcg1_stat_names[i], nr);
4288 	}
4289 
4290 	for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4291 		seq_buf_printf(s, "%s %lu\n", vm_event_name(memcg1_events[i]),
4292 			       memcg_events_local(memcg, memcg1_events[i]));
4293 
4294 	for (i = 0; i < NR_LRU_LISTS; i++)
4295 		seq_buf_printf(s, "%s %lu\n", lru_list_name(i),
4296 			       memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4297 			       PAGE_SIZE);
4298 
4299 	/* Hierarchical information */
4300 	memory = memsw = PAGE_COUNTER_MAX;
4301 	for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4302 		memory = min(memory, READ_ONCE(mi->memory.max));
4303 		memsw = min(memsw, READ_ONCE(mi->memsw.max));
4304 	}
4305 	seq_buf_printf(s, "hierarchical_memory_limit %llu\n",
4306 		       (u64)memory * PAGE_SIZE);
4307 	seq_buf_printf(s, "hierarchical_memsw_limit %llu\n",
4308 		       (u64)memsw * PAGE_SIZE);
4309 
4310 	for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4311 		unsigned long nr;
4312 
4313 		nr = memcg_page_state_output(memcg, memcg1_stats[i]);
4314 		seq_buf_printf(s, "total_%s %llu\n", memcg1_stat_names[i],
4315 			       (u64)nr);
4316 	}
4317 
4318 	for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4319 		seq_buf_printf(s, "total_%s %llu\n",
4320 			       vm_event_name(memcg1_events[i]),
4321 			       (u64)memcg_events(memcg, memcg1_events[i]));
4322 
4323 	for (i = 0; i < NR_LRU_LISTS; i++)
4324 		seq_buf_printf(s, "total_%s %llu\n", lru_list_name(i),
4325 			       (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4326 			       PAGE_SIZE);
4327 
4328 #ifdef CONFIG_DEBUG_VM
4329 	{
4330 		pg_data_t *pgdat;
4331 		struct mem_cgroup_per_node *mz;
4332 		unsigned long anon_cost = 0;
4333 		unsigned long file_cost = 0;
4334 
4335 		for_each_online_pgdat(pgdat) {
4336 			mz = memcg->nodeinfo[pgdat->node_id];
4337 
4338 			anon_cost += mz->lruvec.anon_cost;
4339 			file_cost += mz->lruvec.file_cost;
4340 		}
4341 		seq_buf_printf(s, "anon_cost %lu\n", anon_cost);
4342 		seq_buf_printf(s, "file_cost %lu\n", file_cost);
4343 	}
4344 #endif
4345 }
4346 
4347 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4348 				      struct cftype *cft)
4349 {
4350 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4351 
4352 	return mem_cgroup_swappiness(memcg);
4353 }
4354 
4355 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4356 				       struct cftype *cft, u64 val)
4357 {
4358 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4359 
4360 	if (val > 200)
4361 		return -EINVAL;
4362 
4363 	if (!mem_cgroup_is_root(memcg))
4364 		WRITE_ONCE(memcg->swappiness, val);
4365 	else
4366 		WRITE_ONCE(vm_swappiness, val);
4367 
4368 	return 0;
4369 }
4370 
4371 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4372 {
4373 	struct mem_cgroup_threshold_ary *t;
4374 	unsigned long usage;
4375 	int i;
4376 
4377 	rcu_read_lock();
4378 	if (!swap)
4379 		t = rcu_dereference(memcg->thresholds.primary);
4380 	else
4381 		t = rcu_dereference(memcg->memsw_thresholds.primary);
4382 
4383 	if (!t)
4384 		goto unlock;
4385 
4386 	usage = mem_cgroup_usage(memcg, swap);
4387 
4388 	/*
4389 	 * current_threshold points to threshold just below or equal to usage.
4390 	 * If it's not true, a threshold was crossed after last
4391 	 * call of __mem_cgroup_threshold().
4392 	 */
4393 	i = t->current_threshold;
4394 
4395 	/*
4396 	 * Iterate backward over array of thresholds starting from
4397 	 * current_threshold and check if a threshold is crossed.
4398 	 * If none of thresholds below usage is crossed, we read
4399 	 * only one element of the array here.
4400 	 */
4401 	for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4402 		eventfd_signal(t->entries[i].eventfd);
4403 
4404 	/* i = current_threshold + 1 */
4405 	i++;
4406 
4407 	/*
4408 	 * Iterate forward over array of thresholds starting from
4409 	 * current_threshold+1 and check if a threshold is crossed.
4410 	 * If none of thresholds above usage is crossed, we read
4411 	 * only one element of the array here.
4412 	 */
4413 	for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4414 		eventfd_signal(t->entries[i].eventfd);
4415 
4416 	/* Update current_threshold */
4417 	t->current_threshold = i - 1;
4418 unlock:
4419 	rcu_read_unlock();
4420 }
4421 
4422 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4423 {
4424 	while (memcg) {
4425 		__mem_cgroup_threshold(memcg, false);
4426 		if (do_memsw_account())
4427 			__mem_cgroup_threshold(memcg, true);
4428 
4429 		memcg = parent_mem_cgroup(memcg);
4430 	}
4431 }
4432 
4433 static int compare_thresholds(const void *a, const void *b)
4434 {
4435 	const struct mem_cgroup_threshold *_a = a;
4436 	const struct mem_cgroup_threshold *_b = b;
4437 
4438 	if (_a->threshold > _b->threshold)
4439 		return 1;
4440 
4441 	if (_a->threshold < _b->threshold)
4442 		return -1;
4443 
4444 	return 0;
4445 }
4446 
4447 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4448 {
4449 	struct mem_cgroup_eventfd_list *ev;
4450 
4451 	spin_lock(&memcg_oom_lock);
4452 
4453 	list_for_each_entry(ev, &memcg->oom_notify, list)
4454 		eventfd_signal(ev->eventfd);
4455 
4456 	spin_unlock(&memcg_oom_lock);
4457 	return 0;
4458 }
4459 
4460 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4461 {
4462 	struct mem_cgroup *iter;
4463 
4464 	for_each_mem_cgroup_tree(iter, memcg)
4465 		mem_cgroup_oom_notify_cb(iter);
4466 }
4467 
4468 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4469 	struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4470 {
4471 	struct mem_cgroup_thresholds *thresholds;
4472 	struct mem_cgroup_threshold_ary *new;
4473 	unsigned long threshold;
4474 	unsigned long usage;
4475 	int i, size, ret;
4476 
4477 	ret = page_counter_memparse(args, "-1", &threshold);
4478 	if (ret)
4479 		return ret;
4480 
4481 	mutex_lock(&memcg->thresholds_lock);
4482 
4483 	if (type == _MEM) {
4484 		thresholds = &memcg->thresholds;
4485 		usage = mem_cgroup_usage(memcg, false);
4486 	} else if (type == _MEMSWAP) {
4487 		thresholds = &memcg->memsw_thresholds;
4488 		usage = mem_cgroup_usage(memcg, true);
4489 	} else
4490 		BUG();
4491 
4492 	/* Check if a threshold crossed before adding a new one */
4493 	if (thresholds->primary)
4494 		__mem_cgroup_threshold(memcg, type == _MEMSWAP);
4495 
4496 	size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4497 
4498 	/* Allocate memory for new array of thresholds */
4499 	new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4500 	if (!new) {
4501 		ret = -ENOMEM;
4502 		goto unlock;
4503 	}
4504 	new->size = size;
4505 
4506 	/* Copy thresholds (if any) to new array */
4507 	if (thresholds->primary)
4508 		memcpy(new->entries, thresholds->primary->entries,
4509 		       flex_array_size(new, entries, size - 1));
4510 
4511 	/* Add new threshold */
4512 	new->entries[size - 1].eventfd = eventfd;
4513 	new->entries[size - 1].threshold = threshold;
4514 
4515 	/* Sort thresholds. Registering of new threshold isn't time-critical */
4516 	sort(new->entries, size, sizeof(*new->entries),
4517 			compare_thresholds, NULL);
4518 
4519 	/* Find current threshold */
4520 	new->current_threshold = -1;
4521 	for (i = 0; i < size; i++) {
4522 		if (new->entries[i].threshold <= usage) {
4523 			/*
4524 			 * new->current_threshold will not be used until
4525 			 * rcu_assign_pointer(), so it's safe to increment
4526 			 * it here.
4527 			 */
4528 			++new->current_threshold;
4529 		} else
4530 			break;
4531 	}
4532 
4533 	/* Free old spare buffer and save old primary buffer as spare */
4534 	kfree(thresholds->spare);
4535 	thresholds->spare = thresholds->primary;
4536 
4537 	rcu_assign_pointer(thresholds->primary, new);
4538 
4539 	/* To be sure that nobody uses thresholds */
4540 	synchronize_rcu();
4541 
4542 unlock:
4543 	mutex_unlock(&memcg->thresholds_lock);
4544 
4545 	return ret;
4546 }
4547 
4548 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4549 	struct eventfd_ctx *eventfd, const char *args)
4550 {
4551 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4552 }
4553 
4554 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4555 	struct eventfd_ctx *eventfd, const char *args)
4556 {
4557 	return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4558 }
4559 
4560 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4561 	struct eventfd_ctx *eventfd, enum res_type type)
4562 {
4563 	struct mem_cgroup_thresholds *thresholds;
4564 	struct mem_cgroup_threshold_ary *new;
4565 	unsigned long usage;
4566 	int i, j, size, entries;
4567 
4568 	mutex_lock(&memcg->thresholds_lock);
4569 
4570 	if (type == _MEM) {
4571 		thresholds = &memcg->thresholds;
4572 		usage = mem_cgroup_usage(memcg, false);
4573 	} else if (type == _MEMSWAP) {
4574 		thresholds = &memcg->memsw_thresholds;
4575 		usage = mem_cgroup_usage(memcg, true);
4576 	} else
4577 		BUG();
4578 
4579 	if (!thresholds->primary)
4580 		goto unlock;
4581 
4582 	/* Check if a threshold crossed before removing */
4583 	__mem_cgroup_threshold(memcg, type == _MEMSWAP);
4584 
4585 	/* Calculate new number of threshold */
4586 	size = entries = 0;
4587 	for (i = 0; i < thresholds->primary->size; i++) {
4588 		if (thresholds->primary->entries[i].eventfd != eventfd)
4589 			size++;
4590 		else
4591 			entries++;
4592 	}
4593 
4594 	new = thresholds->spare;
4595 
4596 	/* If no items related to eventfd have been cleared, nothing to do */
4597 	if (!entries)
4598 		goto unlock;
4599 
4600 	/* Set thresholds array to NULL if we don't have thresholds */
4601 	if (!size) {
4602 		kfree(new);
4603 		new = NULL;
4604 		goto swap_buffers;
4605 	}
4606 
4607 	new->size = size;
4608 
4609 	/* Copy thresholds and find current threshold */
4610 	new->current_threshold = -1;
4611 	for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4612 		if (thresholds->primary->entries[i].eventfd == eventfd)
4613 			continue;
4614 
4615 		new->entries[j] = thresholds->primary->entries[i];
4616 		if (new->entries[j].threshold <= usage) {
4617 			/*
4618 			 * new->current_threshold will not be used
4619 			 * until rcu_assign_pointer(), so it's safe to increment
4620 			 * it here.
4621 			 */
4622 			++new->current_threshold;
4623 		}
4624 		j++;
4625 	}
4626 
4627 swap_buffers:
4628 	/* Swap primary and spare array */
4629 	thresholds->spare = thresholds->primary;
4630 
4631 	rcu_assign_pointer(thresholds->primary, new);
4632 
4633 	/* To be sure that nobody uses thresholds */
4634 	synchronize_rcu();
4635 
4636 	/* If all events are unregistered, free the spare array */
4637 	if (!new) {
4638 		kfree(thresholds->spare);
4639 		thresholds->spare = NULL;
4640 	}
4641 unlock:
4642 	mutex_unlock(&memcg->thresholds_lock);
4643 }
4644 
4645 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4646 	struct eventfd_ctx *eventfd)
4647 {
4648 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4649 }
4650 
4651 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4652 	struct eventfd_ctx *eventfd)
4653 {
4654 	return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4655 }
4656 
4657 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4658 	struct eventfd_ctx *eventfd, const char *args)
4659 {
4660 	struct mem_cgroup_eventfd_list *event;
4661 
4662 	event = kmalloc(sizeof(*event),	GFP_KERNEL);
4663 	if (!event)
4664 		return -ENOMEM;
4665 
4666 	spin_lock(&memcg_oom_lock);
4667 
4668 	event->eventfd = eventfd;
4669 	list_add(&event->list, &memcg->oom_notify);
4670 
4671 	/* already in OOM ? */
4672 	if (memcg->under_oom)
4673 		eventfd_signal(eventfd);
4674 	spin_unlock(&memcg_oom_lock);
4675 
4676 	return 0;
4677 }
4678 
4679 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4680 	struct eventfd_ctx *eventfd)
4681 {
4682 	struct mem_cgroup_eventfd_list *ev, *tmp;
4683 
4684 	spin_lock(&memcg_oom_lock);
4685 
4686 	list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4687 		if (ev->eventfd == eventfd) {
4688 			list_del(&ev->list);
4689 			kfree(ev);
4690 		}
4691 	}
4692 
4693 	spin_unlock(&memcg_oom_lock);
4694 }
4695 
4696 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4697 {
4698 	struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4699 
4700 	seq_printf(sf, "oom_kill_disable %d\n", READ_ONCE(memcg->oom_kill_disable));
4701 	seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4702 	seq_printf(sf, "oom_kill %lu\n",
4703 		   atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4704 	return 0;
4705 }
4706 
4707 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4708 	struct cftype *cft, u64 val)
4709 {
4710 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4711 
4712 	/* cannot set to root cgroup and only 0 and 1 are allowed */
4713 	if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4714 		return -EINVAL;
4715 
4716 	WRITE_ONCE(memcg->oom_kill_disable, val);
4717 	if (!val)
4718 		memcg_oom_recover(memcg);
4719 
4720 	return 0;
4721 }
4722 
4723 #ifdef CONFIG_CGROUP_WRITEBACK
4724 
4725 #include <trace/events/writeback.h>
4726 
4727 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4728 {
4729 	return wb_domain_init(&memcg->cgwb_domain, gfp);
4730 }
4731 
4732 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4733 {
4734 	wb_domain_exit(&memcg->cgwb_domain);
4735 }
4736 
4737 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4738 {
4739 	wb_domain_size_changed(&memcg->cgwb_domain);
4740 }
4741 
4742 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4743 {
4744 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4745 
4746 	if (!memcg->css.parent)
4747 		return NULL;
4748 
4749 	return &memcg->cgwb_domain;
4750 }
4751 
4752 /**
4753  * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4754  * @wb: bdi_writeback in question
4755  * @pfilepages: out parameter for number of file pages
4756  * @pheadroom: out parameter for number of allocatable pages according to memcg
4757  * @pdirty: out parameter for number of dirty pages
4758  * @pwriteback: out parameter for number of pages under writeback
4759  *
4760  * Determine the numbers of file, headroom, dirty, and writeback pages in
4761  * @wb's memcg.  File, dirty and writeback are self-explanatory.  Headroom
4762  * is a bit more involved.
4763  *
4764  * A memcg's headroom is "min(max, high) - used".  In the hierarchy, the
4765  * headroom is calculated as the lowest headroom of itself and the
4766  * ancestors.  Note that this doesn't consider the actual amount of
4767  * available memory in the system.  The caller should further cap
4768  * *@pheadroom accordingly.
4769  */
4770 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4771 			 unsigned long *pheadroom, unsigned long *pdirty,
4772 			 unsigned long *pwriteback)
4773 {
4774 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4775 	struct mem_cgroup *parent;
4776 
4777 	mem_cgroup_flush_stats(memcg);
4778 
4779 	*pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4780 	*pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4781 	*pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4782 			memcg_page_state(memcg, NR_ACTIVE_FILE);
4783 
4784 	*pheadroom = PAGE_COUNTER_MAX;
4785 	while ((parent = parent_mem_cgroup(memcg))) {
4786 		unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4787 					    READ_ONCE(memcg->memory.high));
4788 		unsigned long used = page_counter_read(&memcg->memory);
4789 
4790 		*pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4791 		memcg = parent;
4792 	}
4793 }
4794 
4795 /*
4796  * Foreign dirty flushing
4797  *
4798  * There's an inherent mismatch between memcg and writeback.  The former
4799  * tracks ownership per-page while the latter per-inode.  This was a
4800  * deliberate design decision because honoring per-page ownership in the
4801  * writeback path is complicated, may lead to higher CPU and IO overheads
4802  * and deemed unnecessary given that write-sharing an inode across
4803  * different cgroups isn't a common use-case.
4804  *
4805  * Combined with inode majority-writer ownership switching, this works well
4806  * enough in most cases but there are some pathological cases.  For
4807  * example, let's say there are two cgroups A and B which keep writing to
4808  * different but confined parts of the same inode.  B owns the inode and
4809  * A's memory is limited far below B's.  A's dirty ratio can rise enough to
4810  * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4811  * triggering background writeback.  A will be slowed down without a way to
4812  * make writeback of the dirty pages happen.
4813  *
4814  * Conditions like the above can lead to a cgroup getting repeatedly and
4815  * severely throttled after making some progress after each
4816  * dirty_expire_interval while the underlying IO device is almost
4817  * completely idle.
4818  *
4819  * Solving this problem completely requires matching the ownership tracking
4820  * granularities between memcg and writeback in either direction.  However,
4821  * the more egregious behaviors can be avoided by simply remembering the
4822  * most recent foreign dirtying events and initiating remote flushes on
4823  * them when local writeback isn't enough to keep the memory clean enough.
4824  *
4825  * The following two functions implement such mechanism.  When a foreign
4826  * page - a page whose memcg and writeback ownerships don't match - is
4827  * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4828  * bdi_writeback on the page owning memcg.  When balance_dirty_pages()
4829  * decides that the memcg needs to sleep due to high dirty ratio, it calls
4830  * mem_cgroup_flush_foreign() which queues writeback on the recorded
4831  * foreign bdi_writebacks which haven't expired.  Both the numbers of
4832  * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4833  * limited to MEMCG_CGWB_FRN_CNT.
4834  *
4835  * The mechanism only remembers IDs and doesn't hold any object references.
4836  * As being wrong occasionally doesn't matter, updates and accesses to the
4837  * records are lockless and racy.
4838  */
4839 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4840 					     struct bdi_writeback *wb)
4841 {
4842 	struct mem_cgroup *memcg = folio_memcg(folio);
4843 	struct memcg_cgwb_frn *frn;
4844 	u64 now = get_jiffies_64();
4845 	u64 oldest_at = now;
4846 	int oldest = -1;
4847 	int i;
4848 
4849 	trace_track_foreign_dirty(folio, wb);
4850 
4851 	/*
4852 	 * Pick the slot to use.  If there is already a slot for @wb, keep
4853 	 * using it.  If not replace the oldest one which isn't being
4854 	 * written out.
4855 	 */
4856 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4857 		frn = &memcg->cgwb_frn[i];
4858 		if (frn->bdi_id == wb->bdi->id &&
4859 		    frn->memcg_id == wb->memcg_css->id)
4860 			break;
4861 		if (time_before64(frn->at, oldest_at) &&
4862 		    atomic_read(&frn->done.cnt) == 1) {
4863 			oldest = i;
4864 			oldest_at = frn->at;
4865 		}
4866 	}
4867 
4868 	if (i < MEMCG_CGWB_FRN_CNT) {
4869 		/*
4870 		 * Re-using an existing one.  Update timestamp lazily to
4871 		 * avoid making the cacheline hot.  We want them to be
4872 		 * reasonably up-to-date and significantly shorter than
4873 		 * dirty_expire_interval as that's what expires the record.
4874 		 * Use the shorter of 1s and dirty_expire_interval / 8.
4875 		 */
4876 		unsigned long update_intv =
4877 			min_t(unsigned long, HZ,
4878 			      msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4879 
4880 		if (time_before64(frn->at, now - update_intv))
4881 			frn->at = now;
4882 	} else if (oldest >= 0) {
4883 		/* replace the oldest free one */
4884 		frn = &memcg->cgwb_frn[oldest];
4885 		frn->bdi_id = wb->bdi->id;
4886 		frn->memcg_id = wb->memcg_css->id;
4887 		frn->at = now;
4888 	}
4889 }
4890 
4891 /* issue foreign writeback flushes for recorded foreign dirtying events */
4892 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4893 {
4894 	struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4895 	unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4896 	u64 now = jiffies_64;
4897 	int i;
4898 
4899 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4900 		struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4901 
4902 		/*
4903 		 * If the record is older than dirty_expire_interval,
4904 		 * writeback on it has already started.  No need to kick it
4905 		 * off again.  Also, don't start a new one if there's
4906 		 * already one in flight.
4907 		 */
4908 		if (time_after64(frn->at, now - intv) &&
4909 		    atomic_read(&frn->done.cnt) == 1) {
4910 			frn->at = 0;
4911 			trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4912 			cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4913 					       WB_REASON_FOREIGN_FLUSH,
4914 					       &frn->done);
4915 		}
4916 	}
4917 }
4918 
4919 #else	/* CONFIG_CGROUP_WRITEBACK */
4920 
4921 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4922 {
4923 	return 0;
4924 }
4925 
4926 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4927 {
4928 }
4929 
4930 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4931 {
4932 }
4933 
4934 #endif	/* CONFIG_CGROUP_WRITEBACK */
4935 
4936 /*
4937  * DO NOT USE IN NEW FILES.
4938  *
4939  * "cgroup.event_control" implementation.
4940  *
4941  * This is way over-engineered.  It tries to support fully configurable
4942  * events for each user.  Such level of flexibility is completely
4943  * unnecessary especially in the light of the planned unified hierarchy.
4944  *
4945  * Please deprecate this and replace with something simpler if at all
4946  * possible.
4947  */
4948 
4949 /*
4950  * Unregister event and free resources.
4951  *
4952  * Gets called from workqueue.
4953  */
4954 static void memcg_event_remove(struct work_struct *work)
4955 {
4956 	struct mem_cgroup_event *event =
4957 		container_of(work, struct mem_cgroup_event, remove);
4958 	struct mem_cgroup *memcg = event->memcg;
4959 
4960 	remove_wait_queue(event->wqh, &event->wait);
4961 
4962 	event->unregister_event(memcg, event->eventfd);
4963 
4964 	/* Notify userspace the event is going away. */
4965 	eventfd_signal(event->eventfd);
4966 
4967 	eventfd_ctx_put(event->eventfd);
4968 	kfree(event);
4969 	css_put(&memcg->css);
4970 }
4971 
4972 /*
4973  * Gets called on EPOLLHUP on eventfd when user closes it.
4974  *
4975  * Called with wqh->lock held and interrupts disabled.
4976  */
4977 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4978 			    int sync, void *key)
4979 {
4980 	struct mem_cgroup_event *event =
4981 		container_of(wait, struct mem_cgroup_event, wait);
4982 	struct mem_cgroup *memcg = event->memcg;
4983 	__poll_t flags = key_to_poll(key);
4984 
4985 	if (flags & EPOLLHUP) {
4986 		/*
4987 		 * If the event has been detached at cgroup removal, we
4988 		 * can simply return knowing the other side will cleanup
4989 		 * for us.
4990 		 *
4991 		 * We can't race against event freeing since the other
4992 		 * side will require wqh->lock via remove_wait_queue(),
4993 		 * which we hold.
4994 		 */
4995 		spin_lock(&memcg->event_list_lock);
4996 		if (!list_empty(&event->list)) {
4997 			list_del_init(&event->list);
4998 			/*
4999 			 * We are in atomic context, but cgroup_event_remove()
5000 			 * may sleep, so we have to call it in workqueue.
5001 			 */
5002 			schedule_work(&event->remove);
5003 		}
5004 		spin_unlock(&memcg->event_list_lock);
5005 	}
5006 
5007 	return 0;
5008 }
5009 
5010 static void memcg_event_ptable_queue_proc(struct file *file,
5011 		wait_queue_head_t *wqh, poll_table *pt)
5012 {
5013 	struct mem_cgroup_event *event =
5014 		container_of(pt, struct mem_cgroup_event, pt);
5015 
5016 	event->wqh = wqh;
5017 	add_wait_queue(wqh, &event->wait);
5018 }
5019 
5020 /*
5021  * DO NOT USE IN NEW FILES.
5022  *
5023  * Parse input and register new cgroup event handler.
5024  *
5025  * Input must be in format '<event_fd> <control_fd> <args>'.
5026  * Interpretation of args is defined by control file implementation.
5027  */
5028 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5029 					 char *buf, size_t nbytes, loff_t off)
5030 {
5031 	struct cgroup_subsys_state *css = of_css(of);
5032 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5033 	struct mem_cgroup_event *event;
5034 	struct cgroup_subsys_state *cfile_css;
5035 	unsigned int efd, cfd;
5036 	struct fd efile;
5037 	struct fd cfile;
5038 	struct dentry *cdentry;
5039 	const char *name;
5040 	char *endp;
5041 	int ret;
5042 
5043 	if (IS_ENABLED(CONFIG_PREEMPT_RT))
5044 		return -EOPNOTSUPP;
5045 
5046 	buf = strstrip(buf);
5047 
5048 	efd = simple_strtoul(buf, &endp, 10);
5049 	if (*endp != ' ')
5050 		return -EINVAL;
5051 	buf = endp + 1;
5052 
5053 	cfd = simple_strtoul(buf, &endp, 10);
5054 	if ((*endp != ' ') && (*endp != '\0'))
5055 		return -EINVAL;
5056 	buf = endp + 1;
5057 
5058 	event = kzalloc(sizeof(*event), GFP_KERNEL);
5059 	if (!event)
5060 		return -ENOMEM;
5061 
5062 	event->memcg = memcg;
5063 	INIT_LIST_HEAD(&event->list);
5064 	init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5065 	init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5066 	INIT_WORK(&event->remove, memcg_event_remove);
5067 
5068 	efile = fdget(efd);
5069 	if (!efile.file) {
5070 		ret = -EBADF;
5071 		goto out_kfree;
5072 	}
5073 
5074 	event->eventfd = eventfd_ctx_fileget(efile.file);
5075 	if (IS_ERR(event->eventfd)) {
5076 		ret = PTR_ERR(event->eventfd);
5077 		goto out_put_efile;
5078 	}
5079 
5080 	cfile = fdget(cfd);
5081 	if (!cfile.file) {
5082 		ret = -EBADF;
5083 		goto out_put_eventfd;
5084 	}
5085 
5086 	/* the process need read permission on control file */
5087 	/* AV: shouldn't we check that it's been opened for read instead? */
5088 	ret = file_permission(cfile.file, MAY_READ);
5089 	if (ret < 0)
5090 		goto out_put_cfile;
5091 
5092 	/*
5093 	 * The control file must be a regular cgroup1 file. As a regular cgroup
5094 	 * file can't be renamed, it's safe to access its name afterwards.
5095 	 */
5096 	cdentry = cfile.file->f_path.dentry;
5097 	if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
5098 		ret = -EINVAL;
5099 		goto out_put_cfile;
5100 	}
5101 
5102 	/*
5103 	 * Determine the event callbacks and set them in @event.  This used
5104 	 * to be done via struct cftype but cgroup core no longer knows
5105 	 * about these events.  The following is crude but the whole thing
5106 	 * is for compatibility anyway.
5107 	 *
5108 	 * DO NOT ADD NEW FILES.
5109 	 */
5110 	name = cdentry->d_name.name;
5111 
5112 	if (!strcmp(name, "memory.usage_in_bytes")) {
5113 		event->register_event = mem_cgroup_usage_register_event;
5114 		event->unregister_event = mem_cgroup_usage_unregister_event;
5115 	} else if (!strcmp(name, "memory.oom_control")) {
5116 		event->register_event = mem_cgroup_oom_register_event;
5117 		event->unregister_event = mem_cgroup_oom_unregister_event;
5118 	} else if (!strcmp(name, "memory.pressure_level")) {
5119 		event->register_event = vmpressure_register_event;
5120 		event->unregister_event = vmpressure_unregister_event;
5121 	} else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5122 		event->register_event = memsw_cgroup_usage_register_event;
5123 		event->unregister_event = memsw_cgroup_usage_unregister_event;
5124 	} else {
5125 		ret = -EINVAL;
5126 		goto out_put_cfile;
5127 	}
5128 
5129 	/*
5130 	 * Verify @cfile should belong to @css.  Also, remaining events are
5131 	 * automatically removed on cgroup destruction but the removal is
5132 	 * asynchronous, so take an extra ref on @css.
5133 	 */
5134 	cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
5135 					       &memory_cgrp_subsys);
5136 	ret = -EINVAL;
5137 	if (IS_ERR(cfile_css))
5138 		goto out_put_cfile;
5139 	if (cfile_css != css) {
5140 		css_put(cfile_css);
5141 		goto out_put_cfile;
5142 	}
5143 
5144 	ret = event->register_event(memcg, event->eventfd, buf);
5145 	if (ret)
5146 		goto out_put_css;
5147 
5148 	vfs_poll(efile.file, &event->pt);
5149 
5150 	spin_lock_irq(&memcg->event_list_lock);
5151 	list_add(&event->list, &memcg->event_list);
5152 	spin_unlock_irq(&memcg->event_list_lock);
5153 
5154 	fdput(cfile);
5155 	fdput(efile);
5156 
5157 	return nbytes;
5158 
5159 out_put_css:
5160 	css_put(css);
5161 out_put_cfile:
5162 	fdput(cfile);
5163 out_put_eventfd:
5164 	eventfd_ctx_put(event->eventfd);
5165 out_put_efile:
5166 	fdput(efile);
5167 out_kfree:
5168 	kfree(event);
5169 
5170 	return ret;
5171 }
5172 
5173 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_SLUB_DEBUG)
5174 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
5175 {
5176 	/*
5177 	 * Deprecated.
5178 	 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
5179 	 */
5180 	return 0;
5181 }
5182 #endif
5183 
5184 static int memory_stat_show(struct seq_file *m, void *v);
5185 
5186 static struct cftype mem_cgroup_legacy_files[] = {
5187 	{
5188 		.name = "usage_in_bytes",
5189 		.private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5190 		.read_u64 = mem_cgroup_read_u64,
5191 	},
5192 	{
5193 		.name = "max_usage_in_bytes",
5194 		.private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5195 		.write = mem_cgroup_reset,
5196 		.read_u64 = mem_cgroup_read_u64,
5197 	},
5198 	{
5199 		.name = "limit_in_bytes",
5200 		.private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5201 		.write = mem_cgroup_write,
5202 		.read_u64 = mem_cgroup_read_u64,
5203 	},
5204 	{
5205 		.name = "soft_limit_in_bytes",
5206 		.private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5207 		.write = mem_cgroup_write,
5208 		.read_u64 = mem_cgroup_read_u64,
5209 	},
5210 	{
5211 		.name = "failcnt",
5212 		.private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5213 		.write = mem_cgroup_reset,
5214 		.read_u64 = mem_cgroup_read_u64,
5215 	},
5216 	{
5217 		.name = "stat",
5218 		.seq_show = memory_stat_show,
5219 	},
5220 	{
5221 		.name = "force_empty",
5222 		.write = mem_cgroup_force_empty_write,
5223 	},
5224 	{
5225 		.name = "use_hierarchy",
5226 		.write_u64 = mem_cgroup_hierarchy_write,
5227 		.read_u64 = mem_cgroup_hierarchy_read,
5228 	},
5229 	{
5230 		.name = "cgroup.event_control",		/* XXX: for compat */
5231 		.write = memcg_write_event_control,
5232 		.flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5233 	},
5234 	{
5235 		.name = "swappiness",
5236 		.read_u64 = mem_cgroup_swappiness_read,
5237 		.write_u64 = mem_cgroup_swappiness_write,
5238 	},
5239 	{
5240 		.name = "move_charge_at_immigrate",
5241 		.read_u64 = mem_cgroup_move_charge_read,
5242 		.write_u64 = mem_cgroup_move_charge_write,
5243 	},
5244 	{
5245 		.name = "oom_control",
5246 		.seq_show = mem_cgroup_oom_control_read,
5247 		.write_u64 = mem_cgroup_oom_control_write,
5248 	},
5249 	{
5250 		.name = "pressure_level",
5251 		.seq_show = mem_cgroup_dummy_seq_show,
5252 	},
5253 #ifdef CONFIG_NUMA
5254 	{
5255 		.name = "numa_stat",
5256 		.seq_show = memcg_numa_stat_show,
5257 	},
5258 #endif
5259 	{
5260 		.name = "kmem.limit_in_bytes",
5261 		.private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5262 		.write = mem_cgroup_write,
5263 		.read_u64 = mem_cgroup_read_u64,
5264 	},
5265 	{
5266 		.name = "kmem.usage_in_bytes",
5267 		.private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5268 		.read_u64 = mem_cgroup_read_u64,
5269 	},
5270 	{
5271 		.name = "kmem.failcnt",
5272 		.private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5273 		.write = mem_cgroup_reset,
5274 		.read_u64 = mem_cgroup_read_u64,
5275 	},
5276 	{
5277 		.name = "kmem.max_usage_in_bytes",
5278 		.private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5279 		.write = mem_cgroup_reset,
5280 		.read_u64 = mem_cgroup_read_u64,
5281 	},
5282 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_SLUB_DEBUG)
5283 	{
5284 		.name = "kmem.slabinfo",
5285 		.seq_show = mem_cgroup_slab_show,
5286 	},
5287 #endif
5288 	{
5289 		.name = "kmem.tcp.limit_in_bytes",
5290 		.private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5291 		.write = mem_cgroup_write,
5292 		.read_u64 = mem_cgroup_read_u64,
5293 	},
5294 	{
5295 		.name = "kmem.tcp.usage_in_bytes",
5296 		.private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5297 		.read_u64 = mem_cgroup_read_u64,
5298 	},
5299 	{
5300 		.name = "kmem.tcp.failcnt",
5301 		.private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5302 		.write = mem_cgroup_reset,
5303 		.read_u64 = mem_cgroup_read_u64,
5304 	},
5305 	{
5306 		.name = "kmem.tcp.max_usage_in_bytes",
5307 		.private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5308 		.write = mem_cgroup_reset,
5309 		.read_u64 = mem_cgroup_read_u64,
5310 	},
5311 	{ },	/* terminate */
5312 };
5313 
5314 /*
5315  * Private memory cgroup IDR
5316  *
5317  * Swap-out records and page cache shadow entries need to store memcg
5318  * references in constrained space, so we maintain an ID space that is
5319  * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5320  * memory-controlled cgroups to 64k.
5321  *
5322  * However, there usually are many references to the offline CSS after
5323  * the cgroup has been destroyed, such as page cache or reclaimable
5324  * slab objects, that don't need to hang on to the ID. We want to keep
5325  * those dead CSS from occupying IDs, or we might quickly exhaust the
5326  * relatively small ID space and prevent the creation of new cgroups
5327  * even when there are much fewer than 64k cgroups - possibly none.
5328  *
5329  * Maintain a private 16-bit ID space for memcg, and allow the ID to
5330  * be freed and recycled when it's no longer needed, which is usually
5331  * when the CSS is offlined.
5332  *
5333  * The only exception to that are records of swapped out tmpfs/shmem
5334  * pages that need to be attributed to live ancestors on swapin. But
5335  * those references are manageable from userspace.
5336  */
5337 
5338 #define MEM_CGROUP_ID_MAX	((1UL << MEM_CGROUP_ID_SHIFT) - 1)
5339 static DEFINE_IDR(mem_cgroup_idr);
5340 
5341 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5342 {
5343 	if (memcg->id.id > 0) {
5344 		idr_remove(&mem_cgroup_idr, memcg->id.id);
5345 		memcg->id.id = 0;
5346 	}
5347 }
5348 
5349 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5350 						  unsigned int n)
5351 {
5352 	refcount_add(n, &memcg->id.ref);
5353 }
5354 
5355 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5356 {
5357 	if (refcount_sub_and_test(n, &memcg->id.ref)) {
5358 		mem_cgroup_id_remove(memcg);
5359 
5360 		/* Memcg ID pins CSS */
5361 		css_put(&memcg->css);
5362 	}
5363 }
5364 
5365 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5366 {
5367 	mem_cgroup_id_put_many(memcg, 1);
5368 }
5369 
5370 /**
5371  * mem_cgroup_from_id - look up a memcg from a memcg id
5372  * @id: the memcg id to look up
5373  *
5374  * Caller must hold rcu_read_lock().
5375  */
5376 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5377 {
5378 	WARN_ON_ONCE(!rcu_read_lock_held());
5379 	return idr_find(&mem_cgroup_idr, id);
5380 }
5381 
5382 #ifdef CONFIG_SHRINKER_DEBUG
5383 struct mem_cgroup *mem_cgroup_get_from_ino(unsigned long ino)
5384 {
5385 	struct cgroup *cgrp;
5386 	struct cgroup_subsys_state *css;
5387 	struct mem_cgroup *memcg;
5388 
5389 	cgrp = cgroup_get_from_id(ino);
5390 	if (IS_ERR(cgrp))
5391 		return ERR_CAST(cgrp);
5392 
5393 	css = cgroup_get_e_css(cgrp, &memory_cgrp_subsys);
5394 	if (css)
5395 		memcg = container_of(css, struct mem_cgroup, css);
5396 	else
5397 		memcg = ERR_PTR(-ENOENT);
5398 
5399 	cgroup_put(cgrp);
5400 
5401 	return memcg;
5402 }
5403 #endif
5404 
5405 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5406 {
5407 	struct mem_cgroup_per_node *pn;
5408 
5409 	pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5410 	if (!pn)
5411 		return 1;
5412 
5413 	pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5414 						   GFP_KERNEL_ACCOUNT);
5415 	if (!pn->lruvec_stats_percpu) {
5416 		kfree(pn);
5417 		return 1;
5418 	}
5419 
5420 	lruvec_init(&pn->lruvec);
5421 	pn->memcg = memcg;
5422 
5423 	memcg->nodeinfo[node] = pn;
5424 	return 0;
5425 }
5426 
5427 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5428 {
5429 	struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5430 
5431 	if (!pn)
5432 		return;
5433 
5434 	free_percpu(pn->lruvec_stats_percpu);
5435 	kfree(pn);
5436 }
5437 
5438 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5439 {
5440 	int node;
5441 
5442 	if (memcg->orig_objcg)
5443 		obj_cgroup_put(memcg->orig_objcg);
5444 
5445 	for_each_node(node)
5446 		free_mem_cgroup_per_node_info(memcg, node);
5447 	kfree(memcg->vmstats);
5448 	free_percpu(memcg->vmstats_percpu);
5449 	kfree(memcg);
5450 }
5451 
5452 static void mem_cgroup_free(struct mem_cgroup *memcg)
5453 {
5454 	lru_gen_exit_memcg(memcg);
5455 	memcg_wb_domain_exit(memcg);
5456 	__mem_cgroup_free(memcg);
5457 }
5458 
5459 static struct mem_cgroup *mem_cgroup_alloc(void)
5460 {
5461 	struct mem_cgroup *memcg;
5462 	int node;
5463 	int __maybe_unused i;
5464 	long error = -ENOMEM;
5465 
5466 	memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5467 	if (!memcg)
5468 		return ERR_PTR(error);
5469 
5470 	memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5471 				 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
5472 	if (memcg->id.id < 0) {
5473 		error = memcg->id.id;
5474 		goto fail;
5475 	}
5476 
5477 	memcg->vmstats = kzalloc(sizeof(struct memcg_vmstats), GFP_KERNEL);
5478 	if (!memcg->vmstats)
5479 		goto fail;
5480 
5481 	memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5482 						 GFP_KERNEL_ACCOUNT);
5483 	if (!memcg->vmstats_percpu)
5484 		goto fail;
5485 
5486 	for_each_node(node)
5487 		if (alloc_mem_cgroup_per_node_info(memcg, node))
5488 			goto fail;
5489 
5490 	if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5491 		goto fail;
5492 
5493 	INIT_WORK(&memcg->high_work, high_work_func);
5494 	INIT_LIST_HEAD(&memcg->oom_notify);
5495 	mutex_init(&memcg->thresholds_lock);
5496 	spin_lock_init(&memcg->move_lock);
5497 	vmpressure_init(&memcg->vmpressure);
5498 	INIT_LIST_HEAD(&memcg->event_list);
5499 	spin_lock_init(&memcg->event_list_lock);
5500 	memcg->socket_pressure = jiffies;
5501 #ifdef CONFIG_MEMCG_KMEM
5502 	memcg->kmemcg_id = -1;
5503 	INIT_LIST_HEAD(&memcg->objcg_list);
5504 #endif
5505 #ifdef CONFIG_CGROUP_WRITEBACK
5506 	INIT_LIST_HEAD(&memcg->cgwb_list);
5507 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5508 		memcg->cgwb_frn[i].done =
5509 			__WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5510 #endif
5511 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5512 	spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5513 	INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5514 	memcg->deferred_split_queue.split_queue_len = 0;
5515 #endif
5516 	lru_gen_init_memcg(memcg);
5517 	return memcg;
5518 fail:
5519 	mem_cgroup_id_remove(memcg);
5520 	__mem_cgroup_free(memcg);
5521 	return ERR_PTR(error);
5522 }
5523 
5524 static struct cgroup_subsys_state * __ref
5525 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5526 {
5527 	struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5528 	struct mem_cgroup *memcg, *old_memcg;
5529 
5530 	old_memcg = set_active_memcg(parent);
5531 	memcg = mem_cgroup_alloc();
5532 	set_active_memcg(old_memcg);
5533 	if (IS_ERR(memcg))
5534 		return ERR_CAST(memcg);
5535 
5536 	page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5537 	WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5538 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5539 	memcg->zswap_max = PAGE_COUNTER_MAX;
5540 	WRITE_ONCE(memcg->zswap_writeback,
5541 		!parent || READ_ONCE(parent->zswap_writeback));
5542 #endif
5543 	page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5544 	if (parent) {
5545 		WRITE_ONCE(memcg->swappiness, mem_cgroup_swappiness(parent));
5546 		WRITE_ONCE(memcg->oom_kill_disable, READ_ONCE(parent->oom_kill_disable));
5547 
5548 		page_counter_init(&memcg->memory, &parent->memory);
5549 		page_counter_init(&memcg->swap, &parent->swap);
5550 		page_counter_init(&memcg->kmem, &parent->kmem);
5551 		page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5552 	} else {
5553 		init_memcg_events();
5554 		page_counter_init(&memcg->memory, NULL);
5555 		page_counter_init(&memcg->swap, NULL);
5556 		page_counter_init(&memcg->kmem, NULL);
5557 		page_counter_init(&memcg->tcpmem, NULL);
5558 
5559 		root_mem_cgroup = memcg;
5560 		return &memcg->css;
5561 	}
5562 
5563 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5564 		static_branch_inc(&memcg_sockets_enabled_key);
5565 
5566 #if defined(CONFIG_MEMCG_KMEM)
5567 	if (!cgroup_memory_nobpf)
5568 		static_branch_inc(&memcg_bpf_enabled_key);
5569 #endif
5570 
5571 	return &memcg->css;
5572 }
5573 
5574 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5575 {
5576 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5577 
5578 	if (memcg_online_kmem(memcg))
5579 		goto remove_id;
5580 
5581 	/*
5582 	 * A memcg must be visible for expand_shrinker_info()
5583 	 * by the time the maps are allocated. So, we allocate maps
5584 	 * here, when for_each_mem_cgroup() can't skip it.
5585 	 */
5586 	if (alloc_shrinker_info(memcg))
5587 		goto offline_kmem;
5588 
5589 	if (unlikely(mem_cgroup_is_root(memcg)))
5590 		queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5591 				   FLUSH_TIME);
5592 	lru_gen_online_memcg(memcg);
5593 
5594 	/* Online state pins memcg ID, memcg ID pins CSS */
5595 	refcount_set(&memcg->id.ref, 1);
5596 	css_get(css);
5597 
5598 	/*
5599 	 * Ensure mem_cgroup_from_id() works once we're fully online.
5600 	 *
5601 	 * We could do this earlier and require callers to filter with
5602 	 * css_tryget_online(). But right now there are no users that
5603 	 * need earlier access, and the workingset code relies on the
5604 	 * cgroup tree linkage (mem_cgroup_get_nr_swap_pages()). So
5605 	 * publish it here at the end of onlining. This matches the
5606 	 * regular ID destruction during offlining.
5607 	 */
5608 	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5609 
5610 	return 0;
5611 offline_kmem:
5612 	memcg_offline_kmem(memcg);
5613 remove_id:
5614 	mem_cgroup_id_remove(memcg);
5615 	return -ENOMEM;
5616 }
5617 
5618 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5619 {
5620 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5621 	struct mem_cgroup_event *event, *tmp;
5622 
5623 	/*
5624 	 * Unregister events and notify userspace.
5625 	 * Notify userspace about cgroup removing only after rmdir of cgroup
5626 	 * directory to avoid race between userspace and kernelspace.
5627 	 */
5628 	spin_lock_irq(&memcg->event_list_lock);
5629 	list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5630 		list_del_init(&event->list);
5631 		schedule_work(&event->remove);
5632 	}
5633 	spin_unlock_irq(&memcg->event_list_lock);
5634 
5635 	page_counter_set_min(&memcg->memory, 0);
5636 	page_counter_set_low(&memcg->memory, 0);
5637 
5638 	zswap_memcg_offline_cleanup(memcg);
5639 
5640 	memcg_offline_kmem(memcg);
5641 	reparent_shrinker_deferred(memcg);
5642 	wb_memcg_offline(memcg);
5643 	lru_gen_offline_memcg(memcg);
5644 
5645 	drain_all_stock(memcg);
5646 
5647 	mem_cgroup_id_put(memcg);
5648 }
5649 
5650 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5651 {
5652 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5653 
5654 	invalidate_reclaim_iterators(memcg);
5655 	lru_gen_release_memcg(memcg);
5656 }
5657 
5658 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5659 {
5660 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5661 	int __maybe_unused i;
5662 
5663 #ifdef CONFIG_CGROUP_WRITEBACK
5664 	for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5665 		wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5666 #endif
5667 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5668 		static_branch_dec(&memcg_sockets_enabled_key);
5669 
5670 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5671 		static_branch_dec(&memcg_sockets_enabled_key);
5672 
5673 #if defined(CONFIG_MEMCG_KMEM)
5674 	if (!cgroup_memory_nobpf)
5675 		static_branch_dec(&memcg_bpf_enabled_key);
5676 #endif
5677 
5678 	vmpressure_cleanup(&memcg->vmpressure);
5679 	cancel_work_sync(&memcg->high_work);
5680 	mem_cgroup_remove_from_trees(memcg);
5681 	free_shrinker_info(memcg);
5682 	mem_cgroup_free(memcg);
5683 }
5684 
5685 /**
5686  * mem_cgroup_css_reset - reset the states of a mem_cgroup
5687  * @css: the target css
5688  *
5689  * Reset the states of the mem_cgroup associated with @css.  This is
5690  * invoked when the userland requests disabling on the default hierarchy
5691  * but the memcg is pinned through dependency.  The memcg should stop
5692  * applying policies and should revert to the vanilla state as it may be
5693  * made visible again.
5694  *
5695  * The current implementation only resets the essential configurations.
5696  * This needs to be expanded to cover all the visible parts.
5697  */
5698 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5699 {
5700 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5701 
5702 	page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5703 	page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5704 	page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5705 	page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5706 	page_counter_set_min(&memcg->memory, 0);
5707 	page_counter_set_low(&memcg->memory, 0);
5708 	page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5709 	WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5710 	page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5711 	memcg_wb_domain_size_changed(memcg);
5712 }
5713 
5714 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5715 {
5716 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5717 	struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5718 	struct memcg_vmstats_percpu *statc;
5719 	long delta, delta_cpu, v;
5720 	int i, nid;
5721 
5722 	statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5723 
5724 	for (i = 0; i < MEMCG_NR_STAT; i++) {
5725 		/*
5726 		 * Collect the aggregated propagation counts of groups
5727 		 * below us. We're in a per-cpu loop here and this is
5728 		 * a global counter, so the first cycle will get them.
5729 		 */
5730 		delta = memcg->vmstats->state_pending[i];
5731 		if (delta)
5732 			memcg->vmstats->state_pending[i] = 0;
5733 
5734 		/* Add CPU changes on this level since the last flush */
5735 		delta_cpu = 0;
5736 		v = READ_ONCE(statc->state[i]);
5737 		if (v != statc->state_prev[i]) {
5738 			delta_cpu = v - statc->state_prev[i];
5739 			delta += delta_cpu;
5740 			statc->state_prev[i] = v;
5741 		}
5742 
5743 		/* Aggregate counts on this level and propagate upwards */
5744 		if (delta_cpu)
5745 			memcg->vmstats->state_local[i] += delta_cpu;
5746 
5747 		if (delta) {
5748 			memcg->vmstats->state[i] += delta;
5749 			if (parent)
5750 				parent->vmstats->state_pending[i] += delta;
5751 		}
5752 	}
5753 
5754 	for (i = 0; i < NR_MEMCG_EVENTS; i++) {
5755 		delta = memcg->vmstats->events_pending[i];
5756 		if (delta)
5757 			memcg->vmstats->events_pending[i] = 0;
5758 
5759 		delta_cpu = 0;
5760 		v = READ_ONCE(statc->events[i]);
5761 		if (v != statc->events_prev[i]) {
5762 			delta_cpu = v - statc->events_prev[i];
5763 			delta += delta_cpu;
5764 			statc->events_prev[i] = v;
5765 		}
5766 
5767 		if (delta_cpu)
5768 			memcg->vmstats->events_local[i] += delta_cpu;
5769 
5770 		if (delta) {
5771 			memcg->vmstats->events[i] += delta;
5772 			if (parent)
5773 				parent->vmstats->events_pending[i] += delta;
5774 		}
5775 	}
5776 
5777 	for_each_node_state(nid, N_MEMORY) {
5778 		struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5779 		struct mem_cgroup_per_node *ppn = NULL;
5780 		struct lruvec_stats_percpu *lstatc;
5781 
5782 		if (parent)
5783 			ppn = parent->nodeinfo[nid];
5784 
5785 		lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5786 
5787 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5788 			delta = pn->lruvec_stats.state_pending[i];
5789 			if (delta)
5790 				pn->lruvec_stats.state_pending[i] = 0;
5791 
5792 			delta_cpu = 0;
5793 			v = READ_ONCE(lstatc->state[i]);
5794 			if (v != lstatc->state_prev[i]) {
5795 				delta_cpu = v - lstatc->state_prev[i];
5796 				delta += delta_cpu;
5797 				lstatc->state_prev[i] = v;
5798 			}
5799 
5800 			if (delta_cpu)
5801 				pn->lruvec_stats.state_local[i] += delta_cpu;
5802 
5803 			if (delta) {
5804 				pn->lruvec_stats.state[i] += delta;
5805 				if (ppn)
5806 					ppn->lruvec_stats.state_pending[i] += delta;
5807 			}
5808 		}
5809 	}
5810 	statc->stats_updates = 0;
5811 	/* We are in a per-cpu loop here, only do the atomic write once */
5812 	if (atomic64_read(&memcg->vmstats->stats_updates))
5813 		atomic64_set(&memcg->vmstats->stats_updates, 0);
5814 }
5815 
5816 #ifdef CONFIG_MMU
5817 /* Handlers for move charge at task migration. */
5818 static int mem_cgroup_do_precharge(unsigned long count)
5819 {
5820 	int ret;
5821 
5822 	/* Try a single bulk charge without reclaim first, kswapd may wake */
5823 	ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5824 	if (!ret) {
5825 		mc.precharge += count;
5826 		return ret;
5827 	}
5828 
5829 	/* Try charges one by one with reclaim, but do not retry */
5830 	while (count--) {
5831 		ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5832 		if (ret)
5833 			return ret;
5834 		mc.precharge++;
5835 		cond_resched();
5836 	}
5837 	return 0;
5838 }
5839 
5840 union mc_target {
5841 	struct page	*page;
5842 	swp_entry_t	ent;
5843 };
5844 
5845 enum mc_target_type {
5846 	MC_TARGET_NONE = 0,
5847 	MC_TARGET_PAGE,
5848 	MC_TARGET_SWAP,
5849 	MC_TARGET_DEVICE,
5850 };
5851 
5852 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5853 						unsigned long addr, pte_t ptent)
5854 {
5855 	struct page *page = vm_normal_page(vma, addr, ptent);
5856 
5857 	if (!page)
5858 		return NULL;
5859 	if (PageAnon(page)) {
5860 		if (!(mc.flags & MOVE_ANON))
5861 			return NULL;
5862 	} else {
5863 		if (!(mc.flags & MOVE_FILE))
5864 			return NULL;
5865 	}
5866 	get_page(page);
5867 
5868 	return page;
5869 }
5870 
5871 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5872 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5873 			pte_t ptent, swp_entry_t *entry)
5874 {
5875 	struct page *page = NULL;
5876 	swp_entry_t ent = pte_to_swp_entry(ptent);
5877 
5878 	if (!(mc.flags & MOVE_ANON))
5879 		return NULL;
5880 
5881 	/*
5882 	 * Handle device private pages that are not accessible by the CPU, but
5883 	 * stored as special swap entries in the page table.
5884 	 */
5885 	if (is_device_private_entry(ent)) {
5886 		page = pfn_swap_entry_to_page(ent);
5887 		if (!get_page_unless_zero(page))
5888 			return NULL;
5889 		return page;
5890 	}
5891 
5892 	if (non_swap_entry(ent))
5893 		return NULL;
5894 
5895 	/*
5896 	 * Because swap_cache_get_folio() updates some statistics counter,
5897 	 * we call find_get_page() with swapper_space directly.
5898 	 */
5899 	page = find_get_page(swap_address_space(ent), swp_offset(ent));
5900 	entry->val = ent.val;
5901 
5902 	return page;
5903 }
5904 #else
5905 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5906 			pte_t ptent, swp_entry_t *entry)
5907 {
5908 	return NULL;
5909 }
5910 #endif
5911 
5912 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5913 			unsigned long addr, pte_t ptent)
5914 {
5915 	unsigned long index;
5916 	struct folio *folio;
5917 
5918 	if (!vma->vm_file) /* anonymous vma */
5919 		return NULL;
5920 	if (!(mc.flags & MOVE_FILE))
5921 		return NULL;
5922 
5923 	/* folio is moved even if it's not RSS of this task(page-faulted). */
5924 	/* shmem/tmpfs may report page out on swap: account for that too. */
5925 	index = linear_page_index(vma, addr);
5926 	folio = filemap_get_incore_folio(vma->vm_file->f_mapping, index);
5927 	if (IS_ERR(folio))
5928 		return NULL;
5929 	return folio_file_page(folio, index);
5930 }
5931 
5932 /**
5933  * mem_cgroup_move_account - move account of the page
5934  * @page: the page
5935  * @compound: charge the page as compound or small page
5936  * @from: mem_cgroup which the page is moved from.
5937  * @to:	mem_cgroup which the page is moved to. @from != @to.
5938  *
5939  * The page must be locked and not on the LRU.
5940  *
5941  * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5942  * from old cgroup.
5943  */
5944 static int mem_cgroup_move_account(struct page *page,
5945 				   bool compound,
5946 				   struct mem_cgroup *from,
5947 				   struct mem_cgroup *to)
5948 {
5949 	struct folio *folio = page_folio(page);
5950 	struct lruvec *from_vec, *to_vec;
5951 	struct pglist_data *pgdat;
5952 	unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5953 	int nid, ret;
5954 
5955 	VM_BUG_ON(from == to);
5956 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
5957 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5958 	VM_BUG_ON(compound && !folio_test_large(folio));
5959 
5960 	ret = -EINVAL;
5961 	if (folio_memcg(folio) != from)
5962 		goto out;
5963 
5964 	pgdat = folio_pgdat(folio);
5965 	from_vec = mem_cgroup_lruvec(from, pgdat);
5966 	to_vec = mem_cgroup_lruvec(to, pgdat);
5967 
5968 	folio_memcg_lock(folio);
5969 
5970 	if (folio_test_anon(folio)) {
5971 		if (folio_mapped(folio)) {
5972 			__mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5973 			__mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5974 			if (folio_test_pmd_mappable(folio)) {
5975 				__mod_lruvec_state(from_vec, NR_ANON_THPS,
5976 						   -nr_pages);
5977 				__mod_lruvec_state(to_vec, NR_ANON_THPS,
5978 						   nr_pages);
5979 			}
5980 		}
5981 	} else {
5982 		__mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5983 		__mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5984 
5985 		if (folio_test_swapbacked(folio)) {
5986 			__mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5987 			__mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5988 		}
5989 
5990 		if (folio_mapped(folio)) {
5991 			__mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5992 			__mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5993 		}
5994 
5995 		if (folio_test_dirty(folio)) {
5996 			struct address_space *mapping = folio_mapping(folio);
5997 
5998 			if (mapping_can_writeback(mapping)) {
5999 				__mod_lruvec_state(from_vec, NR_FILE_DIRTY,
6000 						   -nr_pages);
6001 				__mod_lruvec_state(to_vec, NR_FILE_DIRTY,
6002 						   nr_pages);
6003 			}
6004 		}
6005 	}
6006 
6007 #ifdef CONFIG_SWAP
6008 	if (folio_test_swapcache(folio)) {
6009 		__mod_lruvec_state(from_vec, NR_SWAPCACHE, -nr_pages);
6010 		__mod_lruvec_state(to_vec, NR_SWAPCACHE, nr_pages);
6011 	}
6012 #endif
6013 	if (folio_test_writeback(folio)) {
6014 		__mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
6015 		__mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
6016 	}
6017 
6018 	/*
6019 	 * All state has been migrated, let's switch to the new memcg.
6020 	 *
6021 	 * It is safe to change page's memcg here because the page
6022 	 * is referenced, charged, isolated, and locked: we can't race
6023 	 * with (un)charging, migration, LRU putback, or anything else
6024 	 * that would rely on a stable page's memory cgroup.
6025 	 *
6026 	 * Note that folio_memcg_lock is a memcg lock, not a page lock,
6027 	 * to save space. As soon as we switch page's memory cgroup to a
6028 	 * new memcg that isn't locked, the above state can change
6029 	 * concurrently again. Make sure we're truly done with it.
6030 	 */
6031 	smp_mb();
6032 
6033 	css_get(&to->css);
6034 	css_put(&from->css);
6035 
6036 	folio->memcg_data = (unsigned long)to;
6037 
6038 	__folio_memcg_unlock(from);
6039 
6040 	ret = 0;
6041 	nid = folio_nid(folio);
6042 
6043 	local_irq_disable();
6044 	mem_cgroup_charge_statistics(to, nr_pages);
6045 	memcg_check_events(to, nid);
6046 	mem_cgroup_charge_statistics(from, -nr_pages);
6047 	memcg_check_events(from, nid);
6048 	local_irq_enable();
6049 out:
6050 	return ret;
6051 }
6052 
6053 /**
6054  * get_mctgt_type - get target type of moving charge
6055  * @vma: the vma the pte to be checked belongs
6056  * @addr: the address corresponding to the pte to be checked
6057  * @ptent: the pte to be checked
6058  * @target: the pointer the target page or swap ent will be stored(can be NULL)
6059  *
6060  * Context: Called with pte lock held.
6061  * Return:
6062  * * MC_TARGET_NONE - If the pte is not a target for move charge.
6063  * * MC_TARGET_PAGE - If the page corresponding to this pte is a target for
6064  *   move charge. If @target is not NULL, the page is stored in target->page
6065  *   with extra refcnt taken (Caller should release it).
6066  * * MC_TARGET_SWAP - If the swap entry corresponding to this pte is a
6067  *   target for charge migration.  If @target is not NULL, the entry is
6068  *   stored in target->ent.
6069  * * MC_TARGET_DEVICE - Like MC_TARGET_PAGE but page is device memory and
6070  *   thus not on the lru.  For now such page is charged like a regular page
6071  *   would be as it is just special memory taking the place of a regular page.
6072  *   See Documentations/vm/hmm.txt and include/linux/hmm.h
6073  */
6074 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6075 		unsigned long addr, pte_t ptent, union mc_target *target)
6076 {
6077 	struct page *page = NULL;
6078 	enum mc_target_type ret = MC_TARGET_NONE;
6079 	swp_entry_t ent = { .val = 0 };
6080 
6081 	if (pte_present(ptent))
6082 		page = mc_handle_present_pte(vma, addr, ptent);
6083 	else if (pte_none_mostly(ptent))
6084 		/*
6085 		 * PTE markers should be treated as a none pte here, separated
6086 		 * from other swap handling below.
6087 		 */
6088 		page = mc_handle_file_pte(vma, addr, ptent);
6089 	else if (is_swap_pte(ptent))
6090 		page = mc_handle_swap_pte(vma, ptent, &ent);
6091 
6092 	if (target && page) {
6093 		if (!trylock_page(page)) {
6094 			put_page(page);
6095 			return ret;
6096 		}
6097 		/*
6098 		 * page_mapped() must be stable during the move. This
6099 		 * pte is locked, so if it's present, the page cannot
6100 		 * become unmapped. If it isn't, we have only partial
6101 		 * control over the mapped state: the page lock will
6102 		 * prevent new faults against pagecache and swapcache,
6103 		 * so an unmapped page cannot become mapped. However,
6104 		 * if the page is already mapped elsewhere, it can
6105 		 * unmap, and there is nothing we can do about it.
6106 		 * Alas, skip moving the page in this case.
6107 		 */
6108 		if (!pte_present(ptent) && page_mapped(page)) {
6109 			unlock_page(page);
6110 			put_page(page);
6111 			return ret;
6112 		}
6113 	}
6114 
6115 	if (!page && !ent.val)
6116 		return ret;
6117 	if (page) {
6118 		/*
6119 		 * Do only loose check w/o serialization.
6120 		 * mem_cgroup_move_account() checks the page is valid or
6121 		 * not under LRU exclusion.
6122 		 */
6123 		if (page_memcg(page) == mc.from) {
6124 			ret = MC_TARGET_PAGE;
6125 			if (is_device_private_page(page) ||
6126 			    is_device_coherent_page(page))
6127 				ret = MC_TARGET_DEVICE;
6128 			if (target)
6129 				target->page = page;
6130 		}
6131 		if (!ret || !target) {
6132 			if (target)
6133 				unlock_page(page);
6134 			put_page(page);
6135 		}
6136 	}
6137 	/*
6138 	 * There is a swap entry and a page doesn't exist or isn't charged.
6139 	 * But we cannot move a tail-page in a THP.
6140 	 */
6141 	if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
6142 	    mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6143 		ret = MC_TARGET_SWAP;
6144 		if (target)
6145 			target->ent = ent;
6146 	}
6147 	return ret;
6148 }
6149 
6150 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6151 /*
6152  * We don't consider PMD mapped swapping or file mapped pages because THP does
6153  * not support them for now.
6154  * Caller should make sure that pmd_trans_huge(pmd) is true.
6155  */
6156 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6157 		unsigned long addr, pmd_t pmd, union mc_target *target)
6158 {
6159 	struct page *page = NULL;
6160 	enum mc_target_type ret = MC_TARGET_NONE;
6161 
6162 	if (unlikely(is_swap_pmd(pmd))) {
6163 		VM_BUG_ON(thp_migration_supported() &&
6164 				  !is_pmd_migration_entry(pmd));
6165 		return ret;
6166 	}
6167 	page = pmd_page(pmd);
6168 	VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6169 	if (!(mc.flags & MOVE_ANON))
6170 		return ret;
6171 	if (page_memcg(page) == mc.from) {
6172 		ret = MC_TARGET_PAGE;
6173 		if (target) {
6174 			get_page(page);
6175 			if (!trylock_page(page)) {
6176 				put_page(page);
6177 				return MC_TARGET_NONE;
6178 			}
6179 			target->page = page;
6180 		}
6181 	}
6182 	return ret;
6183 }
6184 #else
6185 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6186 		unsigned long addr, pmd_t pmd, union mc_target *target)
6187 {
6188 	return MC_TARGET_NONE;
6189 }
6190 #endif
6191 
6192 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6193 					unsigned long addr, unsigned long end,
6194 					struct mm_walk *walk)
6195 {
6196 	struct vm_area_struct *vma = walk->vma;
6197 	pte_t *pte;
6198 	spinlock_t *ptl;
6199 
6200 	ptl = pmd_trans_huge_lock(pmd, vma);
6201 	if (ptl) {
6202 		/*
6203 		 * Note their can not be MC_TARGET_DEVICE for now as we do not
6204 		 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
6205 		 * this might change.
6206 		 */
6207 		if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6208 			mc.precharge += HPAGE_PMD_NR;
6209 		spin_unlock(ptl);
6210 		return 0;
6211 	}
6212 
6213 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6214 	if (!pte)
6215 		return 0;
6216 	for (; addr != end; pte++, addr += PAGE_SIZE)
6217 		if (get_mctgt_type(vma, addr, ptep_get(pte), NULL))
6218 			mc.precharge++;	/* increment precharge temporarily */
6219 	pte_unmap_unlock(pte - 1, ptl);
6220 	cond_resched();
6221 
6222 	return 0;
6223 }
6224 
6225 static const struct mm_walk_ops precharge_walk_ops = {
6226 	.pmd_entry	= mem_cgroup_count_precharge_pte_range,
6227 	.walk_lock	= PGWALK_RDLOCK,
6228 };
6229 
6230 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6231 {
6232 	unsigned long precharge;
6233 
6234 	mmap_read_lock(mm);
6235 	walk_page_range(mm, 0, ULONG_MAX, &precharge_walk_ops, NULL);
6236 	mmap_read_unlock(mm);
6237 
6238 	precharge = mc.precharge;
6239 	mc.precharge = 0;
6240 
6241 	return precharge;
6242 }
6243 
6244 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6245 {
6246 	unsigned long precharge = mem_cgroup_count_precharge(mm);
6247 
6248 	VM_BUG_ON(mc.moving_task);
6249 	mc.moving_task = current;
6250 	return mem_cgroup_do_precharge(precharge);
6251 }
6252 
6253 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6254 static void __mem_cgroup_clear_mc(void)
6255 {
6256 	struct mem_cgroup *from = mc.from;
6257 	struct mem_cgroup *to = mc.to;
6258 
6259 	/* we must uncharge all the leftover precharges from mc.to */
6260 	if (mc.precharge) {
6261 		mem_cgroup_cancel_charge(mc.to, mc.precharge);
6262 		mc.precharge = 0;
6263 	}
6264 	/*
6265 	 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6266 	 * we must uncharge here.
6267 	 */
6268 	if (mc.moved_charge) {
6269 		mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6270 		mc.moved_charge = 0;
6271 	}
6272 	/* we must fixup refcnts and charges */
6273 	if (mc.moved_swap) {
6274 		/* uncharge swap account from the old cgroup */
6275 		if (!mem_cgroup_is_root(mc.from))
6276 			page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6277 
6278 		mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6279 
6280 		/*
6281 		 * we charged both to->memory and to->memsw, so we
6282 		 * should uncharge to->memory.
6283 		 */
6284 		if (!mem_cgroup_is_root(mc.to))
6285 			page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6286 
6287 		mc.moved_swap = 0;
6288 	}
6289 	memcg_oom_recover(from);
6290 	memcg_oom_recover(to);
6291 	wake_up_all(&mc.waitq);
6292 }
6293 
6294 static void mem_cgroup_clear_mc(void)
6295 {
6296 	struct mm_struct *mm = mc.mm;
6297 
6298 	/*
6299 	 * we must clear moving_task before waking up waiters at the end of
6300 	 * task migration.
6301 	 */
6302 	mc.moving_task = NULL;
6303 	__mem_cgroup_clear_mc();
6304 	spin_lock(&mc.lock);
6305 	mc.from = NULL;
6306 	mc.to = NULL;
6307 	mc.mm = NULL;
6308 	spin_unlock(&mc.lock);
6309 
6310 	mmput(mm);
6311 }
6312 
6313 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6314 {
6315 	struct cgroup_subsys_state *css;
6316 	struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6317 	struct mem_cgroup *from;
6318 	struct task_struct *leader, *p;
6319 	struct mm_struct *mm;
6320 	unsigned long move_flags;
6321 	int ret = 0;
6322 
6323 	/* charge immigration isn't supported on the default hierarchy */
6324 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6325 		return 0;
6326 
6327 	/*
6328 	 * Multi-process migrations only happen on the default hierarchy
6329 	 * where charge immigration is not used.  Perform charge
6330 	 * immigration if @tset contains a leader and whine if there are
6331 	 * multiple.
6332 	 */
6333 	p = NULL;
6334 	cgroup_taskset_for_each_leader(leader, css, tset) {
6335 		WARN_ON_ONCE(p);
6336 		p = leader;
6337 		memcg = mem_cgroup_from_css(css);
6338 	}
6339 	if (!p)
6340 		return 0;
6341 
6342 	/*
6343 	 * We are now committed to this value whatever it is. Changes in this
6344 	 * tunable will only affect upcoming migrations, not the current one.
6345 	 * So we need to save it, and keep it going.
6346 	 */
6347 	move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6348 	if (!move_flags)
6349 		return 0;
6350 
6351 	from = mem_cgroup_from_task(p);
6352 
6353 	VM_BUG_ON(from == memcg);
6354 
6355 	mm = get_task_mm(p);
6356 	if (!mm)
6357 		return 0;
6358 	/* We move charges only when we move a owner of the mm */
6359 	if (mm->owner == p) {
6360 		VM_BUG_ON(mc.from);
6361 		VM_BUG_ON(mc.to);
6362 		VM_BUG_ON(mc.precharge);
6363 		VM_BUG_ON(mc.moved_charge);
6364 		VM_BUG_ON(mc.moved_swap);
6365 
6366 		spin_lock(&mc.lock);
6367 		mc.mm = mm;
6368 		mc.from = from;
6369 		mc.to = memcg;
6370 		mc.flags = move_flags;
6371 		spin_unlock(&mc.lock);
6372 		/* We set mc.moving_task later */
6373 
6374 		ret = mem_cgroup_precharge_mc(mm);
6375 		if (ret)
6376 			mem_cgroup_clear_mc();
6377 	} else {
6378 		mmput(mm);
6379 	}
6380 	return ret;
6381 }
6382 
6383 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6384 {
6385 	if (mc.to)
6386 		mem_cgroup_clear_mc();
6387 }
6388 
6389 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6390 				unsigned long addr, unsigned long end,
6391 				struct mm_walk *walk)
6392 {
6393 	int ret = 0;
6394 	struct vm_area_struct *vma = walk->vma;
6395 	pte_t *pte;
6396 	spinlock_t *ptl;
6397 	enum mc_target_type target_type;
6398 	union mc_target target;
6399 	struct page *page;
6400 
6401 	ptl = pmd_trans_huge_lock(pmd, vma);
6402 	if (ptl) {
6403 		if (mc.precharge < HPAGE_PMD_NR) {
6404 			spin_unlock(ptl);
6405 			return 0;
6406 		}
6407 		target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6408 		if (target_type == MC_TARGET_PAGE) {
6409 			page = target.page;
6410 			if (isolate_lru_page(page)) {
6411 				if (!mem_cgroup_move_account(page, true,
6412 							     mc.from, mc.to)) {
6413 					mc.precharge -= HPAGE_PMD_NR;
6414 					mc.moved_charge += HPAGE_PMD_NR;
6415 				}
6416 				putback_lru_page(page);
6417 			}
6418 			unlock_page(page);
6419 			put_page(page);
6420 		} else if (target_type == MC_TARGET_DEVICE) {
6421 			page = target.page;
6422 			if (!mem_cgroup_move_account(page, true,
6423 						     mc.from, mc.to)) {
6424 				mc.precharge -= HPAGE_PMD_NR;
6425 				mc.moved_charge += HPAGE_PMD_NR;
6426 			}
6427 			unlock_page(page);
6428 			put_page(page);
6429 		}
6430 		spin_unlock(ptl);
6431 		return 0;
6432 	}
6433 
6434 retry:
6435 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6436 	if (!pte)
6437 		return 0;
6438 	for (; addr != end; addr += PAGE_SIZE) {
6439 		pte_t ptent = ptep_get(pte++);
6440 		bool device = false;
6441 		swp_entry_t ent;
6442 
6443 		if (!mc.precharge)
6444 			break;
6445 
6446 		switch (get_mctgt_type(vma, addr, ptent, &target)) {
6447 		case MC_TARGET_DEVICE:
6448 			device = true;
6449 			fallthrough;
6450 		case MC_TARGET_PAGE:
6451 			page = target.page;
6452 			/*
6453 			 * We can have a part of the split pmd here. Moving it
6454 			 * can be done but it would be too convoluted so simply
6455 			 * ignore such a partial THP and keep it in original
6456 			 * memcg. There should be somebody mapping the head.
6457 			 */
6458 			if (PageTransCompound(page))
6459 				goto put;
6460 			if (!device && !isolate_lru_page(page))
6461 				goto put;
6462 			if (!mem_cgroup_move_account(page, false,
6463 						mc.from, mc.to)) {
6464 				mc.precharge--;
6465 				/* we uncharge from mc.from later. */
6466 				mc.moved_charge++;
6467 			}
6468 			if (!device)
6469 				putback_lru_page(page);
6470 put:			/* get_mctgt_type() gets & locks the page */
6471 			unlock_page(page);
6472 			put_page(page);
6473 			break;
6474 		case MC_TARGET_SWAP:
6475 			ent = target.ent;
6476 			if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6477 				mc.precharge--;
6478 				mem_cgroup_id_get_many(mc.to, 1);
6479 				/* we fixup other refcnts and charges later. */
6480 				mc.moved_swap++;
6481 			}
6482 			break;
6483 		default:
6484 			break;
6485 		}
6486 	}
6487 	pte_unmap_unlock(pte - 1, ptl);
6488 	cond_resched();
6489 
6490 	if (addr != end) {
6491 		/*
6492 		 * We have consumed all precharges we got in can_attach().
6493 		 * We try charge one by one, but don't do any additional
6494 		 * charges to mc.to if we have failed in charge once in attach()
6495 		 * phase.
6496 		 */
6497 		ret = mem_cgroup_do_precharge(1);
6498 		if (!ret)
6499 			goto retry;
6500 	}
6501 
6502 	return ret;
6503 }
6504 
6505 static const struct mm_walk_ops charge_walk_ops = {
6506 	.pmd_entry	= mem_cgroup_move_charge_pte_range,
6507 	.walk_lock	= PGWALK_RDLOCK,
6508 };
6509 
6510 static void mem_cgroup_move_charge(void)
6511 {
6512 	lru_add_drain_all();
6513 	/*
6514 	 * Signal folio_memcg_lock() to take the memcg's move_lock
6515 	 * while we're moving its pages to another memcg. Then wait
6516 	 * for already started RCU-only updates to finish.
6517 	 */
6518 	atomic_inc(&mc.from->moving_account);
6519 	synchronize_rcu();
6520 retry:
6521 	if (unlikely(!mmap_read_trylock(mc.mm))) {
6522 		/*
6523 		 * Someone who are holding the mmap_lock might be waiting in
6524 		 * waitq. So we cancel all extra charges, wake up all waiters,
6525 		 * and retry. Because we cancel precharges, we might not be able
6526 		 * to move enough charges, but moving charge is a best-effort
6527 		 * feature anyway, so it wouldn't be a big problem.
6528 		 */
6529 		__mem_cgroup_clear_mc();
6530 		cond_resched();
6531 		goto retry;
6532 	}
6533 	/*
6534 	 * When we have consumed all precharges and failed in doing
6535 	 * additional charge, the page walk just aborts.
6536 	 */
6537 	walk_page_range(mc.mm, 0, ULONG_MAX, &charge_walk_ops, NULL);
6538 	mmap_read_unlock(mc.mm);
6539 	atomic_dec(&mc.from->moving_account);
6540 }
6541 
6542 static void mem_cgroup_move_task(void)
6543 {
6544 	if (mc.to) {
6545 		mem_cgroup_move_charge();
6546 		mem_cgroup_clear_mc();
6547 	}
6548 }
6549 
6550 #else	/* !CONFIG_MMU */
6551 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6552 {
6553 	return 0;
6554 }
6555 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6556 {
6557 }
6558 static void mem_cgroup_move_task(void)
6559 {
6560 }
6561 #endif
6562 
6563 #ifdef CONFIG_MEMCG_KMEM
6564 static void mem_cgroup_fork(struct task_struct *task)
6565 {
6566 	/*
6567 	 * Set the update flag to cause task->objcg to be initialized lazily
6568 	 * on the first allocation. It can be done without any synchronization
6569 	 * because it's always performed on the current task, so does
6570 	 * current_objcg_update().
6571 	 */
6572 	task->objcg = (struct obj_cgroup *)CURRENT_OBJCG_UPDATE_FLAG;
6573 }
6574 
6575 static void mem_cgroup_exit(struct task_struct *task)
6576 {
6577 	struct obj_cgroup *objcg = task->objcg;
6578 
6579 	objcg = (struct obj_cgroup *)
6580 		((unsigned long)objcg & ~CURRENT_OBJCG_UPDATE_FLAG);
6581 	if (objcg)
6582 		obj_cgroup_put(objcg);
6583 
6584 	/*
6585 	 * Some kernel allocations can happen after this point,
6586 	 * but let's ignore them. It can be done without any synchronization
6587 	 * because it's always performed on the current task, so does
6588 	 * current_objcg_update().
6589 	 */
6590 	task->objcg = NULL;
6591 }
6592 #endif
6593 
6594 #ifdef CONFIG_LRU_GEN
6595 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset)
6596 {
6597 	struct task_struct *task;
6598 	struct cgroup_subsys_state *css;
6599 
6600 	/* find the first leader if there is any */
6601 	cgroup_taskset_for_each_leader(task, css, tset)
6602 		break;
6603 
6604 	if (!task)
6605 		return;
6606 
6607 	task_lock(task);
6608 	if (task->mm && READ_ONCE(task->mm->owner) == task)
6609 		lru_gen_migrate_mm(task->mm);
6610 	task_unlock(task);
6611 }
6612 #else
6613 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset) {}
6614 #endif /* CONFIG_LRU_GEN */
6615 
6616 #ifdef CONFIG_MEMCG_KMEM
6617 static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset)
6618 {
6619 	struct task_struct *task;
6620 	struct cgroup_subsys_state *css;
6621 
6622 	cgroup_taskset_for_each(task, css, tset) {
6623 		/* atomically set the update bit */
6624 		set_bit(CURRENT_OBJCG_UPDATE_BIT, (unsigned long *)&task->objcg);
6625 	}
6626 }
6627 #else
6628 static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset) {}
6629 #endif /* CONFIG_MEMCG_KMEM */
6630 
6631 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
6632 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6633 {
6634 	mem_cgroup_lru_gen_attach(tset);
6635 	mem_cgroup_kmem_attach(tset);
6636 }
6637 #endif
6638 
6639 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6640 {
6641 	if (value == PAGE_COUNTER_MAX)
6642 		seq_puts(m, "max\n");
6643 	else
6644 		seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6645 
6646 	return 0;
6647 }
6648 
6649 static u64 memory_current_read(struct cgroup_subsys_state *css,
6650 			       struct cftype *cft)
6651 {
6652 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6653 
6654 	return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6655 }
6656 
6657 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6658 			    struct cftype *cft)
6659 {
6660 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6661 
6662 	return (u64)memcg->memory.watermark * PAGE_SIZE;
6663 }
6664 
6665 static int memory_min_show(struct seq_file *m, void *v)
6666 {
6667 	return seq_puts_memcg_tunable(m,
6668 		READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6669 }
6670 
6671 static ssize_t memory_min_write(struct kernfs_open_file *of,
6672 				char *buf, size_t nbytes, loff_t off)
6673 {
6674 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6675 	unsigned long min;
6676 	int err;
6677 
6678 	buf = strstrip(buf);
6679 	err = page_counter_memparse(buf, "max", &min);
6680 	if (err)
6681 		return err;
6682 
6683 	page_counter_set_min(&memcg->memory, min);
6684 
6685 	return nbytes;
6686 }
6687 
6688 static int memory_low_show(struct seq_file *m, void *v)
6689 {
6690 	return seq_puts_memcg_tunable(m,
6691 		READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6692 }
6693 
6694 static ssize_t memory_low_write(struct kernfs_open_file *of,
6695 				char *buf, size_t nbytes, loff_t off)
6696 {
6697 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6698 	unsigned long low;
6699 	int err;
6700 
6701 	buf = strstrip(buf);
6702 	err = page_counter_memparse(buf, "max", &low);
6703 	if (err)
6704 		return err;
6705 
6706 	page_counter_set_low(&memcg->memory, low);
6707 
6708 	return nbytes;
6709 }
6710 
6711 static int memory_high_show(struct seq_file *m, void *v)
6712 {
6713 	return seq_puts_memcg_tunable(m,
6714 		READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6715 }
6716 
6717 static ssize_t memory_high_write(struct kernfs_open_file *of,
6718 				 char *buf, size_t nbytes, loff_t off)
6719 {
6720 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6721 	unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6722 	bool drained = false;
6723 	unsigned long high;
6724 	int err;
6725 
6726 	buf = strstrip(buf);
6727 	err = page_counter_memparse(buf, "max", &high);
6728 	if (err)
6729 		return err;
6730 
6731 	page_counter_set_high(&memcg->memory, high);
6732 
6733 	for (;;) {
6734 		unsigned long nr_pages = page_counter_read(&memcg->memory);
6735 		unsigned long reclaimed;
6736 
6737 		if (nr_pages <= high)
6738 			break;
6739 
6740 		if (signal_pending(current))
6741 			break;
6742 
6743 		if (!drained) {
6744 			drain_all_stock(memcg);
6745 			drained = true;
6746 			continue;
6747 		}
6748 
6749 		reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6750 					GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP);
6751 
6752 		if (!reclaimed && !nr_retries--)
6753 			break;
6754 	}
6755 
6756 	memcg_wb_domain_size_changed(memcg);
6757 	return nbytes;
6758 }
6759 
6760 static int memory_max_show(struct seq_file *m, void *v)
6761 {
6762 	return seq_puts_memcg_tunable(m,
6763 		READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6764 }
6765 
6766 static ssize_t memory_max_write(struct kernfs_open_file *of,
6767 				char *buf, size_t nbytes, loff_t off)
6768 {
6769 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6770 	unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6771 	bool drained = false;
6772 	unsigned long max;
6773 	int err;
6774 
6775 	buf = strstrip(buf);
6776 	err = page_counter_memparse(buf, "max", &max);
6777 	if (err)
6778 		return err;
6779 
6780 	xchg(&memcg->memory.max, max);
6781 
6782 	for (;;) {
6783 		unsigned long nr_pages = page_counter_read(&memcg->memory);
6784 
6785 		if (nr_pages <= max)
6786 			break;
6787 
6788 		if (signal_pending(current))
6789 			break;
6790 
6791 		if (!drained) {
6792 			drain_all_stock(memcg);
6793 			drained = true;
6794 			continue;
6795 		}
6796 
6797 		if (nr_reclaims) {
6798 			if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6799 					GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP))
6800 				nr_reclaims--;
6801 			continue;
6802 		}
6803 
6804 		memcg_memory_event(memcg, MEMCG_OOM);
6805 		if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6806 			break;
6807 	}
6808 
6809 	memcg_wb_domain_size_changed(memcg);
6810 	return nbytes;
6811 }
6812 
6813 /*
6814  * Note: don't forget to update the 'samples/cgroup/memcg_event_listener'
6815  * if any new events become available.
6816  */
6817 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6818 {
6819 	seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6820 	seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6821 	seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6822 	seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6823 	seq_printf(m, "oom_kill %lu\n",
6824 		   atomic_long_read(&events[MEMCG_OOM_KILL]));
6825 	seq_printf(m, "oom_group_kill %lu\n",
6826 		   atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6827 }
6828 
6829 static int memory_events_show(struct seq_file *m, void *v)
6830 {
6831 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6832 
6833 	__memory_events_show(m, memcg->memory_events);
6834 	return 0;
6835 }
6836 
6837 static int memory_events_local_show(struct seq_file *m, void *v)
6838 {
6839 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6840 
6841 	__memory_events_show(m, memcg->memory_events_local);
6842 	return 0;
6843 }
6844 
6845 static int memory_stat_show(struct seq_file *m, void *v)
6846 {
6847 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6848 	char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
6849 	struct seq_buf s;
6850 
6851 	if (!buf)
6852 		return -ENOMEM;
6853 	seq_buf_init(&s, buf, PAGE_SIZE);
6854 	memory_stat_format(memcg, &s);
6855 	seq_puts(m, buf);
6856 	kfree(buf);
6857 	return 0;
6858 }
6859 
6860 #ifdef CONFIG_NUMA
6861 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6862 						     int item)
6863 {
6864 	return lruvec_page_state(lruvec, item) *
6865 		memcg_page_state_output_unit(item);
6866 }
6867 
6868 static int memory_numa_stat_show(struct seq_file *m, void *v)
6869 {
6870 	int i;
6871 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6872 
6873 	mem_cgroup_flush_stats(memcg);
6874 
6875 	for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6876 		int nid;
6877 
6878 		if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6879 			continue;
6880 
6881 		seq_printf(m, "%s", memory_stats[i].name);
6882 		for_each_node_state(nid, N_MEMORY) {
6883 			u64 size;
6884 			struct lruvec *lruvec;
6885 
6886 			lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6887 			size = lruvec_page_state_output(lruvec,
6888 							memory_stats[i].idx);
6889 			seq_printf(m, " N%d=%llu", nid, size);
6890 		}
6891 		seq_putc(m, '\n');
6892 	}
6893 
6894 	return 0;
6895 }
6896 #endif
6897 
6898 static int memory_oom_group_show(struct seq_file *m, void *v)
6899 {
6900 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6901 
6902 	seq_printf(m, "%d\n", READ_ONCE(memcg->oom_group));
6903 
6904 	return 0;
6905 }
6906 
6907 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6908 				      char *buf, size_t nbytes, loff_t off)
6909 {
6910 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6911 	int ret, oom_group;
6912 
6913 	buf = strstrip(buf);
6914 	if (!buf)
6915 		return -EINVAL;
6916 
6917 	ret = kstrtoint(buf, 0, &oom_group);
6918 	if (ret)
6919 		return ret;
6920 
6921 	if (oom_group != 0 && oom_group != 1)
6922 		return -EINVAL;
6923 
6924 	WRITE_ONCE(memcg->oom_group, oom_group);
6925 
6926 	return nbytes;
6927 }
6928 
6929 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6930 			      size_t nbytes, loff_t off)
6931 {
6932 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6933 	unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6934 	unsigned long nr_to_reclaim, nr_reclaimed = 0;
6935 	unsigned int reclaim_options;
6936 	int err;
6937 
6938 	buf = strstrip(buf);
6939 	err = page_counter_memparse(buf, "", &nr_to_reclaim);
6940 	if (err)
6941 		return err;
6942 
6943 	reclaim_options	= MEMCG_RECLAIM_MAY_SWAP | MEMCG_RECLAIM_PROACTIVE;
6944 	while (nr_reclaimed < nr_to_reclaim) {
6945 		unsigned long reclaimed;
6946 
6947 		if (signal_pending(current))
6948 			return -EINTR;
6949 
6950 		/*
6951 		 * This is the final attempt, drain percpu lru caches in the
6952 		 * hope of introducing more evictable pages for
6953 		 * try_to_free_mem_cgroup_pages().
6954 		 */
6955 		if (!nr_retries)
6956 			lru_add_drain_all();
6957 
6958 		reclaimed = try_to_free_mem_cgroup_pages(memcg,
6959 					min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
6960 					GFP_KERNEL, reclaim_options);
6961 
6962 		if (!reclaimed && !nr_retries--)
6963 			return -EAGAIN;
6964 
6965 		nr_reclaimed += reclaimed;
6966 	}
6967 
6968 	return nbytes;
6969 }
6970 
6971 static struct cftype memory_files[] = {
6972 	{
6973 		.name = "current",
6974 		.flags = CFTYPE_NOT_ON_ROOT,
6975 		.read_u64 = memory_current_read,
6976 	},
6977 	{
6978 		.name = "peak",
6979 		.flags = CFTYPE_NOT_ON_ROOT,
6980 		.read_u64 = memory_peak_read,
6981 	},
6982 	{
6983 		.name = "min",
6984 		.flags = CFTYPE_NOT_ON_ROOT,
6985 		.seq_show = memory_min_show,
6986 		.write = memory_min_write,
6987 	},
6988 	{
6989 		.name = "low",
6990 		.flags = CFTYPE_NOT_ON_ROOT,
6991 		.seq_show = memory_low_show,
6992 		.write = memory_low_write,
6993 	},
6994 	{
6995 		.name = "high",
6996 		.flags = CFTYPE_NOT_ON_ROOT,
6997 		.seq_show = memory_high_show,
6998 		.write = memory_high_write,
6999 	},
7000 	{
7001 		.name = "max",
7002 		.flags = CFTYPE_NOT_ON_ROOT,
7003 		.seq_show = memory_max_show,
7004 		.write = memory_max_write,
7005 	},
7006 	{
7007 		.name = "events",
7008 		.flags = CFTYPE_NOT_ON_ROOT,
7009 		.file_offset = offsetof(struct mem_cgroup, events_file),
7010 		.seq_show = memory_events_show,
7011 	},
7012 	{
7013 		.name = "events.local",
7014 		.flags = CFTYPE_NOT_ON_ROOT,
7015 		.file_offset = offsetof(struct mem_cgroup, events_local_file),
7016 		.seq_show = memory_events_local_show,
7017 	},
7018 	{
7019 		.name = "stat",
7020 		.seq_show = memory_stat_show,
7021 	},
7022 #ifdef CONFIG_NUMA
7023 	{
7024 		.name = "numa_stat",
7025 		.seq_show = memory_numa_stat_show,
7026 	},
7027 #endif
7028 	{
7029 		.name = "oom.group",
7030 		.flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
7031 		.seq_show = memory_oom_group_show,
7032 		.write = memory_oom_group_write,
7033 	},
7034 	{
7035 		.name = "reclaim",
7036 		.flags = CFTYPE_NS_DELEGATABLE,
7037 		.write = memory_reclaim,
7038 	},
7039 	{ }	/* terminate */
7040 };
7041 
7042 struct cgroup_subsys memory_cgrp_subsys = {
7043 	.css_alloc = mem_cgroup_css_alloc,
7044 	.css_online = mem_cgroup_css_online,
7045 	.css_offline = mem_cgroup_css_offline,
7046 	.css_released = mem_cgroup_css_released,
7047 	.css_free = mem_cgroup_css_free,
7048 	.css_reset = mem_cgroup_css_reset,
7049 	.css_rstat_flush = mem_cgroup_css_rstat_flush,
7050 	.can_attach = mem_cgroup_can_attach,
7051 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
7052 	.attach = mem_cgroup_attach,
7053 #endif
7054 	.cancel_attach = mem_cgroup_cancel_attach,
7055 	.post_attach = mem_cgroup_move_task,
7056 #ifdef CONFIG_MEMCG_KMEM
7057 	.fork = mem_cgroup_fork,
7058 	.exit = mem_cgroup_exit,
7059 #endif
7060 	.dfl_cftypes = memory_files,
7061 	.legacy_cftypes = mem_cgroup_legacy_files,
7062 	.early_init = 0,
7063 };
7064 
7065 /*
7066  * This function calculates an individual cgroup's effective
7067  * protection which is derived from its own memory.min/low, its
7068  * parent's and siblings' settings, as well as the actual memory
7069  * distribution in the tree.
7070  *
7071  * The following rules apply to the effective protection values:
7072  *
7073  * 1. At the first level of reclaim, effective protection is equal to
7074  *    the declared protection in memory.min and memory.low.
7075  *
7076  * 2. To enable safe delegation of the protection configuration, at
7077  *    subsequent levels the effective protection is capped to the
7078  *    parent's effective protection.
7079  *
7080  * 3. To make complex and dynamic subtrees easier to configure, the
7081  *    user is allowed to overcommit the declared protection at a given
7082  *    level. If that is the case, the parent's effective protection is
7083  *    distributed to the children in proportion to how much protection
7084  *    they have declared and how much of it they are utilizing.
7085  *
7086  *    This makes distribution proportional, but also work-conserving:
7087  *    if one cgroup claims much more protection than it uses memory,
7088  *    the unused remainder is available to its siblings.
7089  *
7090  * 4. Conversely, when the declared protection is undercommitted at a
7091  *    given level, the distribution of the larger parental protection
7092  *    budget is NOT proportional. A cgroup's protection from a sibling
7093  *    is capped to its own memory.min/low setting.
7094  *
7095  * 5. However, to allow protecting recursive subtrees from each other
7096  *    without having to declare each individual cgroup's fixed share
7097  *    of the ancestor's claim to protection, any unutilized -
7098  *    "floating" - protection from up the tree is distributed in
7099  *    proportion to each cgroup's *usage*. This makes the protection
7100  *    neutral wrt sibling cgroups and lets them compete freely over
7101  *    the shared parental protection budget, but it protects the
7102  *    subtree as a whole from neighboring subtrees.
7103  *
7104  * Note that 4. and 5. are not in conflict: 4. is about protecting
7105  * against immediate siblings whereas 5. is about protecting against
7106  * neighboring subtrees.
7107  */
7108 static unsigned long effective_protection(unsigned long usage,
7109 					  unsigned long parent_usage,
7110 					  unsigned long setting,
7111 					  unsigned long parent_effective,
7112 					  unsigned long siblings_protected)
7113 {
7114 	unsigned long protected;
7115 	unsigned long ep;
7116 
7117 	protected = min(usage, setting);
7118 	/*
7119 	 * If all cgroups at this level combined claim and use more
7120 	 * protection than what the parent affords them, distribute
7121 	 * shares in proportion to utilization.
7122 	 *
7123 	 * We are using actual utilization rather than the statically
7124 	 * claimed protection in order to be work-conserving: claimed
7125 	 * but unused protection is available to siblings that would
7126 	 * otherwise get a smaller chunk than what they claimed.
7127 	 */
7128 	if (siblings_protected > parent_effective)
7129 		return protected * parent_effective / siblings_protected;
7130 
7131 	/*
7132 	 * Ok, utilized protection of all children is within what the
7133 	 * parent affords them, so we know whatever this child claims
7134 	 * and utilizes is effectively protected.
7135 	 *
7136 	 * If there is unprotected usage beyond this value, reclaim
7137 	 * will apply pressure in proportion to that amount.
7138 	 *
7139 	 * If there is unutilized protection, the cgroup will be fully
7140 	 * shielded from reclaim, but we do return a smaller value for
7141 	 * protection than what the group could enjoy in theory. This
7142 	 * is okay. With the overcommit distribution above, effective
7143 	 * protection is always dependent on how memory is actually
7144 	 * consumed among the siblings anyway.
7145 	 */
7146 	ep = protected;
7147 
7148 	/*
7149 	 * If the children aren't claiming (all of) the protection
7150 	 * afforded to them by the parent, distribute the remainder in
7151 	 * proportion to the (unprotected) memory of each cgroup. That
7152 	 * way, cgroups that aren't explicitly prioritized wrt each
7153 	 * other compete freely over the allowance, but they are
7154 	 * collectively protected from neighboring trees.
7155 	 *
7156 	 * We're using unprotected memory for the weight so that if
7157 	 * some cgroups DO claim explicit protection, we don't protect
7158 	 * the same bytes twice.
7159 	 *
7160 	 * Check both usage and parent_usage against the respective
7161 	 * protected values. One should imply the other, but they
7162 	 * aren't read atomically - make sure the division is sane.
7163 	 */
7164 	if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
7165 		return ep;
7166 	if (parent_effective > siblings_protected &&
7167 	    parent_usage > siblings_protected &&
7168 	    usage > protected) {
7169 		unsigned long unclaimed;
7170 
7171 		unclaimed = parent_effective - siblings_protected;
7172 		unclaimed *= usage - protected;
7173 		unclaimed /= parent_usage - siblings_protected;
7174 
7175 		ep += unclaimed;
7176 	}
7177 
7178 	return ep;
7179 }
7180 
7181 /**
7182  * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
7183  * @root: the top ancestor of the sub-tree being checked
7184  * @memcg: the memory cgroup to check
7185  *
7186  * WARNING: This function is not stateless! It can only be used as part
7187  *          of a top-down tree iteration, not for isolated queries.
7188  */
7189 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
7190 				     struct mem_cgroup *memcg)
7191 {
7192 	unsigned long usage, parent_usage;
7193 	struct mem_cgroup *parent;
7194 
7195 	if (mem_cgroup_disabled())
7196 		return;
7197 
7198 	if (!root)
7199 		root = root_mem_cgroup;
7200 
7201 	/*
7202 	 * Effective values of the reclaim targets are ignored so they
7203 	 * can be stale. Have a look at mem_cgroup_protection for more
7204 	 * details.
7205 	 * TODO: calculation should be more robust so that we do not need
7206 	 * that special casing.
7207 	 */
7208 	if (memcg == root)
7209 		return;
7210 
7211 	usage = page_counter_read(&memcg->memory);
7212 	if (!usage)
7213 		return;
7214 
7215 	parent = parent_mem_cgroup(memcg);
7216 
7217 	if (parent == root) {
7218 		memcg->memory.emin = READ_ONCE(memcg->memory.min);
7219 		memcg->memory.elow = READ_ONCE(memcg->memory.low);
7220 		return;
7221 	}
7222 
7223 	parent_usage = page_counter_read(&parent->memory);
7224 
7225 	WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
7226 			READ_ONCE(memcg->memory.min),
7227 			READ_ONCE(parent->memory.emin),
7228 			atomic_long_read(&parent->memory.children_min_usage)));
7229 
7230 	WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
7231 			READ_ONCE(memcg->memory.low),
7232 			READ_ONCE(parent->memory.elow),
7233 			atomic_long_read(&parent->memory.children_low_usage)));
7234 }
7235 
7236 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
7237 			gfp_t gfp)
7238 {
7239 	int ret;
7240 
7241 	ret = try_charge(memcg, gfp, folio_nr_pages(folio));
7242 	if (ret)
7243 		goto out;
7244 
7245 	mem_cgroup_commit_charge(folio, memcg);
7246 out:
7247 	return ret;
7248 }
7249 
7250 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
7251 {
7252 	struct mem_cgroup *memcg;
7253 	int ret;
7254 
7255 	memcg = get_mem_cgroup_from_mm(mm);
7256 	ret = charge_memcg(folio, memcg, gfp);
7257 	css_put(&memcg->css);
7258 
7259 	return ret;
7260 }
7261 
7262 /**
7263  * mem_cgroup_hugetlb_try_charge - try to charge the memcg for a hugetlb folio
7264  * @memcg: memcg to charge.
7265  * @gfp: reclaim mode.
7266  * @nr_pages: number of pages to charge.
7267  *
7268  * This function is called when allocating a huge page folio to determine if
7269  * the memcg has the capacity for it. It does not commit the charge yet,
7270  * as the hugetlb folio itself has not been obtained from the hugetlb pool.
7271  *
7272  * Once we have obtained the hugetlb folio, we can call
7273  * mem_cgroup_commit_charge() to commit the charge. If we fail to obtain the
7274  * folio, we should instead call mem_cgroup_cancel_charge() to undo the effect
7275  * of try_charge().
7276  *
7277  * Returns 0 on success. Otherwise, an error code is returned.
7278  */
7279 int mem_cgroup_hugetlb_try_charge(struct mem_cgroup *memcg, gfp_t gfp,
7280 			long nr_pages)
7281 {
7282 	/*
7283 	 * If hugetlb memcg charging is not enabled, do not fail hugetlb allocation,
7284 	 * but do not attempt to commit charge later (or cancel on error) either.
7285 	 */
7286 	if (mem_cgroup_disabled() || !memcg ||
7287 		!cgroup_subsys_on_dfl(memory_cgrp_subsys) ||
7288 		!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING))
7289 		return -EOPNOTSUPP;
7290 
7291 	if (try_charge(memcg, gfp, nr_pages))
7292 		return -ENOMEM;
7293 
7294 	return 0;
7295 }
7296 
7297 /**
7298  * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
7299  * @folio: folio to charge.
7300  * @mm: mm context of the victim
7301  * @gfp: reclaim mode
7302  * @entry: swap entry for which the folio is allocated
7303  *
7304  * This function charges a folio allocated for swapin. Please call this before
7305  * adding the folio to the swapcache.
7306  *
7307  * Returns 0 on success. Otherwise, an error code is returned.
7308  */
7309 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
7310 				  gfp_t gfp, swp_entry_t entry)
7311 {
7312 	struct mem_cgroup *memcg;
7313 	unsigned short id;
7314 	int ret;
7315 
7316 	if (mem_cgroup_disabled())
7317 		return 0;
7318 
7319 	id = lookup_swap_cgroup_id(entry);
7320 	rcu_read_lock();
7321 	memcg = mem_cgroup_from_id(id);
7322 	if (!memcg || !css_tryget_online(&memcg->css))
7323 		memcg = get_mem_cgroup_from_mm(mm);
7324 	rcu_read_unlock();
7325 
7326 	ret = charge_memcg(folio, memcg, gfp);
7327 
7328 	css_put(&memcg->css);
7329 	return ret;
7330 }
7331 
7332 /*
7333  * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
7334  * @entry: swap entry for which the page is charged
7335  *
7336  * Call this function after successfully adding the charged page to swapcache.
7337  *
7338  * Note: This function assumes the page for which swap slot is being uncharged
7339  * is order 0 page.
7340  */
7341 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
7342 {
7343 	/*
7344 	 * Cgroup1's unified memory+swap counter has been charged with the
7345 	 * new swapcache page, finish the transfer by uncharging the swap
7346 	 * slot. The swap slot would also get uncharged when it dies, but
7347 	 * it can stick around indefinitely and we'd count the page twice
7348 	 * the entire time.
7349 	 *
7350 	 * Cgroup2 has separate resource counters for memory and swap,
7351 	 * so this is a non-issue here. Memory and swap charge lifetimes
7352 	 * correspond 1:1 to page and swap slot lifetimes: we charge the
7353 	 * page to memory here, and uncharge swap when the slot is freed.
7354 	 */
7355 	if (!mem_cgroup_disabled() && do_memsw_account()) {
7356 		/*
7357 		 * The swap entry might not get freed for a long time,
7358 		 * let's not wait for it.  The page already received a
7359 		 * memory+swap charge, drop the swap entry duplicate.
7360 		 */
7361 		mem_cgroup_uncharge_swap(entry, 1);
7362 	}
7363 }
7364 
7365 struct uncharge_gather {
7366 	struct mem_cgroup *memcg;
7367 	unsigned long nr_memory;
7368 	unsigned long pgpgout;
7369 	unsigned long nr_kmem;
7370 	int nid;
7371 };
7372 
7373 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
7374 {
7375 	memset(ug, 0, sizeof(*ug));
7376 }
7377 
7378 static void uncharge_batch(const struct uncharge_gather *ug)
7379 {
7380 	unsigned long flags;
7381 
7382 	if (ug->nr_memory) {
7383 		page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
7384 		if (do_memsw_account())
7385 			page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
7386 		if (ug->nr_kmem)
7387 			memcg_account_kmem(ug->memcg, -ug->nr_kmem);
7388 		memcg_oom_recover(ug->memcg);
7389 	}
7390 
7391 	local_irq_save(flags);
7392 	__count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
7393 	__this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
7394 	memcg_check_events(ug->memcg, ug->nid);
7395 	local_irq_restore(flags);
7396 
7397 	/* drop reference from uncharge_folio */
7398 	css_put(&ug->memcg->css);
7399 }
7400 
7401 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
7402 {
7403 	long nr_pages;
7404 	struct mem_cgroup *memcg;
7405 	struct obj_cgroup *objcg;
7406 
7407 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7408 
7409 	/*
7410 	 * Nobody should be changing or seriously looking at
7411 	 * folio memcg or objcg at this point, we have fully
7412 	 * exclusive access to the folio.
7413 	 */
7414 	if (folio_memcg_kmem(folio)) {
7415 		objcg = __folio_objcg(folio);
7416 		/*
7417 		 * This get matches the put at the end of the function and
7418 		 * kmem pages do not hold memcg references anymore.
7419 		 */
7420 		memcg = get_mem_cgroup_from_objcg(objcg);
7421 	} else {
7422 		memcg = __folio_memcg(folio);
7423 	}
7424 
7425 	if (!memcg)
7426 		return;
7427 
7428 	if (ug->memcg != memcg) {
7429 		if (ug->memcg) {
7430 			uncharge_batch(ug);
7431 			uncharge_gather_clear(ug);
7432 		}
7433 		ug->memcg = memcg;
7434 		ug->nid = folio_nid(folio);
7435 
7436 		/* pairs with css_put in uncharge_batch */
7437 		css_get(&memcg->css);
7438 	}
7439 
7440 	nr_pages = folio_nr_pages(folio);
7441 
7442 	if (folio_memcg_kmem(folio)) {
7443 		ug->nr_memory += nr_pages;
7444 		ug->nr_kmem += nr_pages;
7445 
7446 		folio->memcg_data = 0;
7447 		obj_cgroup_put(objcg);
7448 	} else {
7449 		/* LRU pages aren't accounted at the root level */
7450 		if (!mem_cgroup_is_root(memcg))
7451 			ug->nr_memory += nr_pages;
7452 		ug->pgpgout++;
7453 
7454 		folio->memcg_data = 0;
7455 	}
7456 
7457 	css_put(&memcg->css);
7458 }
7459 
7460 void __mem_cgroup_uncharge(struct folio *folio)
7461 {
7462 	struct uncharge_gather ug;
7463 
7464 	/* Don't touch folio->lru of any random page, pre-check: */
7465 	if (!folio_memcg(folio))
7466 		return;
7467 
7468 	uncharge_gather_clear(&ug);
7469 	uncharge_folio(folio, &ug);
7470 	uncharge_batch(&ug);
7471 }
7472 
7473 /**
7474  * __mem_cgroup_uncharge_list - uncharge a list of page
7475  * @page_list: list of pages to uncharge
7476  *
7477  * Uncharge a list of pages previously charged with
7478  * __mem_cgroup_charge().
7479  */
7480 void __mem_cgroup_uncharge_list(struct list_head *page_list)
7481 {
7482 	struct uncharge_gather ug;
7483 	struct folio *folio;
7484 
7485 	uncharge_gather_clear(&ug);
7486 	list_for_each_entry(folio, page_list, lru)
7487 		uncharge_folio(folio, &ug);
7488 	if (ug.memcg)
7489 		uncharge_batch(&ug);
7490 }
7491 
7492 /**
7493  * mem_cgroup_replace_folio - Charge a folio's replacement.
7494  * @old: Currently circulating folio.
7495  * @new: Replacement folio.
7496  *
7497  * Charge @new as a replacement folio for @old. @old will
7498  * be uncharged upon free. This is only used by the page cache
7499  * (in replace_page_cache_folio()).
7500  *
7501  * Both folios must be locked, @new->mapping must be set up.
7502  */
7503 void mem_cgroup_replace_folio(struct folio *old, struct folio *new)
7504 {
7505 	struct mem_cgroup *memcg;
7506 	long nr_pages = folio_nr_pages(new);
7507 	unsigned long flags;
7508 
7509 	VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7510 	VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7511 	VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7512 	VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
7513 
7514 	if (mem_cgroup_disabled())
7515 		return;
7516 
7517 	/* Page cache replacement: new folio already charged? */
7518 	if (folio_memcg(new))
7519 		return;
7520 
7521 	memcg = folio_memcg(old);
7522 	VM_WARN_ON_ONCE_FOLIO(!memcg, old);
7523 	if (!memcg)
7524 		return;
7525 
7526 	/* Force-charge the new page. The old one will be freed soon */
7527 	if (!mem_cgroup_is_root(memcg)) {
7528 		page_counter_charge(&memcg->memory, nr_pages);
7529 		if (do_memsw_account())
7530 			page_counter_charge(&memcg->memsw, nr_pages);
7531 	}
7532 
7533 	css_get(&memcg->css);
7534 	commit_charge(new, memcg);
7535 
7536 	local_irq_save(flags);
7537 	mem_cgroup_charge_statistics(memcg, nr_pages);
7538 	memcg_check_events(memcg, folio_nid(new));
7539 	local_irq_restore(flags);
7540 }
7541 
7542 /**
7543  * mem_cgroup_migrate - Transfer the memcg data from the old to the new folio.
7544  * @old: Currently circulating folio.
7545  * @new: Replacement folio.
7546  *
7547  * Transfer the memcg data from the old folio to the new folio for migration.
7548  * The old folio's data info will be cleared. Note that the memory counters
7549  * will remain unchanged throughout the process.
7550  *
7551  * Both folios must be locked, @new->mapping must be set up.
7552  */
7553 void mem_cgroup_migrate(struct folio *old, struct folio *new)
7554 {
7555 	struct mem_cgroup *memcg;
7556 
7557 	VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7558 	VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7559 	VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7560 	VM_BUG_ON_FOLIO(folio_nr_pages(old) != folio_nr_pages(new), new);
7561 
7562 	if (mem_cgroup_disabled())
7563 		return;
7564 
7565 	memcg = folio_memcg(old);
7566 	/*
7567 	 * Note that it is normal to see !memcg for a hugetlb folio.
7568 	 * For e.g, itt could have been allocated when memory_hugetlb_accounting
7569 	 * was not selected.
7570 	 */
7571 	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(old) && !memcg, old);
7572 	if (!memcg)
7573 		return;
7574 
7575 	/* Transfer the charge and the css ref */
7576 	commit_charge(new, memcg);
7577 	/*
7578 	 * If the old folio is a large folio and is in the split queue, it needs
7579 	 * to be removed from the split queue now, in case getting an incorrect
7580 	 * split queue in destroy_large_folio() after the memcg of the old folio
7581 	 * is cleared.
7582 	 *
7583 	 * In addition, the old folio is about to be freed after migration, so
7584 	 * removing from the split queue a bit earlier seems reasonable.
7585 	 */
7586 	if (folio_test_large(old) && folio_test_large_rmappable(old))
7587 		folio_undo_large_rmappable(old);
7588 	old->memcg_data = 0;
7589 }
7590 
7591 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7592 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7593 
7594 void mem_cgroup_sk_alloc(struct sock *sk)
7595 {
7596 	struct mem_cgroup *memcg;
7597 
7598 	if (!mem_cgroup_sockets_enabled)
7599 		return;
7600 
7601 	/* Do not associate the sock with unrelated interrupted task's memcg. */
7602 	if (!in_task())
7603 		return;
7604 
7605 	rcu_read_lock();
7606 	memcg = mem_cgroup_from_task(current);
7607 	if (mem_cgroup_is_root(memcg))
7608 		goto out;
7609 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7610 		goto out;
7611 	if (css_tryget(&memcg->css))
7612 		sk->sk_memcg = memcg;
7613 out:
7614 	rcu_read_unlock();
7615 }
7616 
7617 void mem_cgroup_sk_free(struct sock *sk)
7618 {
7619 	if (sk->sk_memcg)
7620 		css_put(&sk->sk_memcg->css);
7621 }
7622 
7623 /**
7624  * mem_cgroup_charge_skmem - charge socket memory
7625  * @memcg: memcg to charge
7626  * @nr_pages: number of pages to charge
7627  * @gfp_mask: reclaim mode
7628  *
7629  * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7630  * @memcg's configured limit, %false if it doesn't.
7631  */
7632 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7633 			     gfp_t gfp_mask)
7634 {
7635 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7636 		struct page_counter *fail;
7637 
7638 		if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7639 			memcg->tcpmem_pressure = 0;
7640 			return true;
7641 		}
7642 		memcg->tcpmem_pressure = 1;
7643 		if (gfp_mask & __GFP_NOFAIL) {
7644 			page_counter_charge(&memcg->tcpmem, nr_pages);
7645 			return true;
7646 		}
7647 		return false;
7648 	}
7649 
7650 	if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7651 		mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7652 		return true;
7653 	}
7654 
7655 	return false;
7656 }
7657 
7658 /**
7659  * mem_cgroup_uncharge_skmem - uncharge socket memory
7660  * @memcg: memcg to uncharge
7661  * @nr_pages: number of pages to uncharge
7662  */
7663 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7664 {
7665 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7666 		page_counter_uncharge(&memcg->tcpmem, nr_pages);
7667 		return;
7668 	}
7669 
7670 	mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7671 
7672 	refill_stock(memcg, nr_pages);
7673 }
7674 
7675 static int __init cgroup_memory(char *s)
7676 {
7677 	char *token;
7678 
7679 	while ((token = strsep(&s, ",")) != NULL) {
7680 		if (!*token)
7681 			continue;
7682 		if (!strcmp(token, "nosocket"))
7683 			cgroup_memory_nosocket = true;
7684 		if (!strcmp(token, "nokmem"))
7685 			cgroup_memory_nokmem = true;
7686 		if (!strcmp(token, "nobpf"))
7687 			cgroup_memory_nobpf = true;
7688 	}
7689 	return 1;
7690 }
7691 __setup("cgroup.memory=", cgroup_memory);
7692 
7693 /*
7694  * subsys_initcall() for memory controller.
7695  *
7696  * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7697  * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7698  * basically everything that doesn't depend on a specific mem_cgroup structure
7699  * should be initialized from here.
7700  */
7701 static int __init mem_cgroup_init(void)
7702 {
7703 	int cpu, node;
7704 
7705 	/*
7706 	 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7707 	 * used for per-memcg-per-cpu caching of per-node statistics. In order
7708 	 * to work fine, we should make sure that the overfill threshold can't
7709 	 * exceed S32_MAX / PAGE_SIZE.
7710 	 */
7711 	BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7712 
7713 	cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7714 				  memcg_hotplug_cpu_dead);
7715 
7716 	for_each_possible_cpu(cpu)
7717 		INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7718 			  drain_local_stock);
7719 
7720 	for_each_node(node) {
7721 		struct mem_cgroup_tree_per_node *rtpn;
7722 
7723 		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, node);
7724 
7725 		rtpn->rb_root = RB_ROOT;
7726 		rtpn->rb_rightmost = NULL;
7727 		spin_lock_init(&rtpn->lock);
7728 		soft_limit_tree.rb_tree_per_node[node] = rtpn;
7729 	}
7730 
7731 	return 0;
7732 }
7733 subsys_initcall(mem_cgroup_init);
7734 
7735 #ifdef CONFIG_SWAP
7736 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7737 {
7738 	while (!refcount_inc_not_zero(&memcg->id.ref)) {
7739 		/*
7740 		 * The root cgroup cannot be destroyed, so it's refcount must
7741 		 * always be >= 1.
7742 		 */
7743 		if (WARN_ON_ONCE(mem_cgroup_is_root(memcg))) {
7744 			VM_BUG_ON(1);
7745 			break;
7746 		}
7747 		memcg = parent_mem_cgroup(memcg);
7748 		if (!memcg)
7749 			memcg = root_mem_cgroup;
7750 	}
7751 	return memcg;
7752 }
7753 
7754 /**
7755  * mem_cgroup_swapout - transfer a memsw charge to swap
7756  * @folio: folio whose memsw charge to transfer
7757  * @entry: swap entry to move the charge to
7758  *
7759  * Transfer the memsw charge of @folio to @entry.
7760  */
7761 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7762 {
7763 	struct mem_cgroup *memcg, *swap_memcg;
7764 	unsigned int nr_entries;
7765 	unsigned short oldid;
7766 
7767 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7768 	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7769 
7770 	if (mem_cgroup_disabled())
7771 		return;
7772 
7773 	if (!do_memsw_account())
7774 		return;
7775 
7776 	memcg = folio_memcg(folio);
7777 
7778 	VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7779 	if (!memcg)
7780 		return;
7781 
7782 	/*
7783 	 * In case the memcg owning these pages has been offlined and doesn't
7784 	 * have an ID allocated to it anymore, charge the closest online
7785 	 * ancestor for the swap instead and transfer the memory+swap charge.
7786 	 */
7787 	swap_memcg = mem_cgroup_id_get_online(memcg);
7788 	nr_entries = folio_nr_pages(folio);
7789 	/* Get references for the tail pages, too */
7790 	if (nr_entries > 1)
7791 		mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7792 	oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7793 				   nr_entries);
7794 	VM_BUG_ON_FOLIO(oldid, folio);
7795 	mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7796 
7797 	folio->memcg_data = 0;
7798 
7799 	if (!mem_cgroup_is_root(memcg))
7800 		page_counter_uncharge(&memcg->memory, nr_entries);
7801 
7802 	if (memcg != swap_memcg) {
7803 		if (!mem_cgroup_is_root(swap_memcg))
7804 			page_counter_charge(&swap_memcg->memsw, nr_entries);
7805 		page_counter_uncharge(&memcg->memsw, nr_entries);
7806 	}
7807 
7808 	/*
7809 	 * Interrupts should be disabled here because the caller holds the
7810 	 * i_pages lock which is taken with interrupts-off. It is
7811 	 * important here to have the interrupts disabled because it is the
7812 	 * only synchronisation we have for updating the per-CPU variables.
7813 	 */
7814 	memcg_stats_lock();
7815 	mem_cgroup_charge_statistics(memcg, -nr_entries);
7816 	memcg_stats_unlock();
7817 	memcg_check_events(memcg, folio_nid(folio));
7818 
7819 	css_put(&memcg->css);
7820 }
7821 
7822 /**
7823  * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7824  * @folio: folio being added to swap
7825  * @entry: swap entry to charge
7826  *
7827  * Try to charge @folio's memcg for the swap space at @entry.
7828  *
7829  * Returns 0 on success, -ENOMEM on failure.
7830  */
7831 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7832 {
7833 	unsigned int nr_pages = folio_nr_pages(folio);
7834 	struct page_counter *counter;
7835 	struct mem_cgroup *memcg;
7836 	unsigned short oldid;
7837 
7838 	if (do_memsw_account())
7839 		return 0;
7840 
7841 	memcg = folio_memcg(folio);
7842 
7843 	VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7844 	if (!memcg)
7845 		return 0;
7846 
7847 	if (!entry.val) {
7848 		memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7849 		return 0;
7850 	}
7851 
7852 	memcg = mem_cgroup_id_get_online(memcg);
7853 
7854 	if (!mem_cgroup_is_root(memcg) &&
7855 	    !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7856 		memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7857 		memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7858 		mem_cgroup_id_put(memcg);
7859 		return -ENOMEM;
7860 	}
7861 
7862 	/* Get references for the tail pages, too */
7863 	if (nr_pages > 1)
7864 		mem_cgroup_id_get_many(memcg, nr_pages - 1);
7865 	oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7866 	VM_BUG_ON_FOLIO(oldid, folio);
7867 	mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7868 
7869 	return 0;
7870 }
7871 
7872 /**
7873  * __mem_cgroup_uncharge_swap - uncharge swap space
7874  * @entry: swap entry to uncharge
7875  * @nr_pages: the amount of swap space to uncharge
7876  */
7877 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7878 {
7879 	struct mem_cgroup *memcg;
7880 	unsigned short id;
7881 
7882 	id = swap_cgroup_record(entry, 0, nr_pages);
7883 	rcu_read_lock();
7884 	memcg = mem_cgroup_from_id(id);
7885 	if (memcg) {
7886 		if (!mem_cgroup_is_root(memcg)) {
7887 			if (do_memsw_account())
7888 				page_counter_uncharge(&memcg->memsw, nr_pages);
7889 			else
7890 				page_counter_uncharge(&memcg->swap, nr_pages);
7891 		}
7892 		mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7893 		mem_cgroup_id_put_many(memcg, nr_pages);
7894 	}
7895 	rcu_read_unlock();
7896 }
7897 
7898 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7899 {
7900 	long nr_swap_pages = get_nr_swap_pages();
7901 
7902 	if (mem_cgroup_disabled() || do_memsw_account())
7903 		return nr_swap_pages;
7904 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg))
7905 		nr_swap_pages = min_t(long, nr_swap_pages,
7906 				      READ_ONCE(memcg->swap.max) -
7907 				      page_counter_read(&memcg->swap));
7908 	return nr_swap_pages;
7909 }
7910 
7911 bool mem_cgroup_swap_full(struct folio *folio)
7912 {
7913 	struct mem_cgroup *memcg;
7914 
7915 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
7916 
7917 	if (vm_swap_full())
7918 		return true;
7919 	if (do_memsw_account())
7920 		return false;
7921 
7922 	memcg = folio_memcg(folio);
7923 	if (!memcg)
7924 		return false;
7925 
7926 	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
7927 		unsigned long usage = page_counter_read(&memcg->swap);
7928 
7929 		if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7930 		    usage * 2 >= READ_ONCE(memcg->swap.max))
7931 			return true;
7932 	}
7933 
7934 	return false;
7935 }
7936 
7937 static int __init setup_swap_account(char *s)
7938 {
7939 	pr_warn_once("The swapaccount= commandline option is deprecated. "
7940 		     "Please report your usecase to linux-mm@kvack.org if you "
7941 		     "depend on this functionality.\n");
7942 	return 1;
7943 }
7944 __setup("swapaccount=", setup_swap_account);
7945 
7946 static u64 swap_current_read(struct cgroup_subsys_state *css,
7947 			     struct cftype *cft)
7948 {
7949 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7950 
7951 	return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7952 }
7953 
7954 static u64 swap_peak_read(struct cgroup_subsys_state *css,
7955 			  struct cftype *cft)
7956 {
7957 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7958 
7959 	return (u64)memcg->swap.watermark * PAGE_SIZE;
7960 }
7961 
7962 static int swap_high_show(struct seq_file *m, void *v)
7963 {
7964 	return seq_puts_memcg_tunable(m,
7965 		READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7966 }
7967 
7968 static ssize_t swap_high_write(struct kernfs_open_file *of,
7969 			       char *buf, size_t nbytes, loff_t off)
7970 {
7971 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7972 	unsigned long high;
7973 	int err;
7974 
7975 	buf = strstrip(buf);
7976 	err = page_counter_memparse(buf, "max", &high);
7977 	if (err)
7978 		return err;
7979 
7980 	page_counter_set_high(&memcg->swap, high);
7981 
7982 	return nbytes;
7983 }
7984 
7985 static int swap_max_show(struct seq_file *m, void *v)
7986 {
7987 	return seq_puts_memcg_tunable(m,
7988 		READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7989 }
7990 
7991 static ssize_t swap_max_write(struct kernfs_open_file *of,
7992 			      char *buf, size_t nbytes, loff_t off)
7993 {
7994 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7995 	unsigned long max;
7996 	int err;
7997 
7998 	buf = strstrip(buf);
7999 	err = page_counter_memparse(buf, "max", &max);
8000 	if (err)
8001 		return err;
8002 
8003 	xchg(&memcg->swap.max, max);
8004 
8005 	return nbytes;
8006 }
8007 
8008 static int swap_events_show(struct seq_file *m, void *v)
8009 {
8010 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
8011 
8012 	seq_printf(m, "high %lu\n",
8013 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
8014 	seq_printf(m, "max %lu\n",
8015 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
8016 	seq_printf(m, "fail %lu\n",
8017 		   atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
8018 
8019 	return 0;
8020 }
8021 
8022 static struct cftype swap_files[] = {
8023 	{
8024 		.name = "swap.current",
8025 		.flags = CFTYPE_NOT_ON_ROOT,
8026 		.read_u64 = swap_current_read,
8027 	},
8028 	{
8029 		.name = "swap.high",
8030 		.flags = CFTYPE_NOT_ON_ROOT,
8031 		.seq_show = swap_high_show,
8032 		.write = swap_high_write,
8033 	},
8034 	{
8035 		.name = "swap.max",
8036 		.flags = CFTYPE_NOT_ON_ROOT,
8037 		.seq_show = swap_max_show,
8038 		.write = swap_max_write,
8039 	},
8040 	{
8041 		.name = "swap.peak",
8042 		.flags = CFTYPE_NOT_ON_ROOT,
8043 		.read_u64 = swap_peak_read,
8044 	},
8045 	{
8046 		.name = "swap.events",
8047 		.flags = CFTYPE_NOT_ON_ROOT,
8048 		.file_offset = offsetof(struct mem_cgroup, swap_events_file),
8049 		.seq_show = swap_events_show,
8050 	},
8051 	{ }	/* terminate */
8052 };
8053 
8054 static struct cftype memsw_files[] = {
8055 	{
8056 		.name = "memsw.usage_in_bytes",
8057 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
8058 		.read_u64 = mem_cgroup_read_u64,
8059 	},
8060 	{
8061 		.name = "memsw.max_usage_in_bytes",
8062 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
8063 		.write = mem_cgroup_reset,
8064 		.read_u64 = mem_cgroup_read_u64,
8065 	},
8066 	{
8067 		.name = "memsw.limit_in_bytes",
8068 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
8069 		.write = mem_cgroup_write,
8070 		.read_u64 = mem_cgroup_read_u64,
8071 	},
8072 	{
8073 		.name = "memsw.failcnt",
8074 		.private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
8075 		.write = mem_cgroup_reset,
8076 		.read_u64 = mem_cgroup_read_u64,
8077 	},
8078 	{ },	/* terminate */
8079 };
8080 
8081 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8082 /**
8083  * obj_cgroup_may_zswap - check if this cgroup can zswap
8084  * @objcg: the object cgroup
8085  *
8086  * Check if the hierarchical zswap limit has been reached.
8087  *
8088  * This doesn't check for specific headroom, and it is not atomic
8089  * either. But with zswap, the size of the allocation is only known
8090  * once compression has occurred, and this optimistic pre-check avoids
8091  * spending cycles on compression when there is already no room left
8092  * or zswap is disabled altogether somewhere in the hierarchy.
8093  */
8094 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
8095 {
8096 	struct mem_cgroup *memcg, *original_memcg;
8097 	bool ret = true;
8098 
8099 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8100 		return true;
8101 
8102 	original_memcg = get_mem_cgroup_from_objcg(objcg);
8103 	for (memcg = original_memcg; !mem_cgroup_is_root(memcg);
8104 	     memcg = parent_mem_cgroup(memcg)) {
8105 		unsigned long max = READ_ONCE(memcg->zswap_max);
8106 		unsigned long pages;
8107 
8108 		if (max == PAGE_COUNTER_MAX)
8109 			continue;
8110 		if (max == 0) {
8111 			ret = false;
8112 			break;
8113 		}
8114 
8115 		/*
8116 		 * mem_cgroup_flush_stats() ignores small changes. Use
8117 		 * do_flush_stats() directly to get accurate stats for charging.
8118 		 */
8119 		do_flush_stats(memcg);
8120 		pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
8121 		if (pages < max)
8122 			continue;
8123 		ret = false;
8124 		break;
8125 	}
8126 	mem_cgroup_put(original_memcg);
8127 	return ret;
8128 }
8129 
8130 /**
8131  * obj_cgroup_charge_zswap - charge compression backend memory
8132  * @objcg: the object cgroup
8133  * @size: size of compressed object
8134  *
8135  * This forces the charge after obj_cgroup_may_zswap() allowed
8136  * compression and storage in zwap for this cgroup to go ahead.
8137  */
8138 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
8139 {
8140 	struct mem_cgroup *memcg;
8141 
8142 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8143 		return;
8144 
8145 	VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
8146 
8147 	/* PF_MEMALLOC context, charging must succeed */
8148 	if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
8149 		VM_WARN_ON_ONCE(1);
8150 
8151 	rcu_read_lock();
8152 	memcg = obj_cgroup_memcg(objcg);
8153 	mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
8154 	mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
8155 	rcu_read_unlock();
8156 }
8157 
8158 /**
8159  * obj_cgroup_uncharge_zswap - uncharge compression backend memory
8160  * @objcg: the object cgroup
8161  * @size: size of compressed object
8162  *
8163  * Uncharges zswap memory on page in.
8164  */
8165 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
8166 {
8167 	struct mem_cgroup *memcg;
8168 
8169 	if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8170 		return;
8171 
8172 	obj_cgroup_uncharge(objcg, size);
8173 
8174 	rcu_read_lock();
8175 	memcg = obj_cgroup_memcg(objcg);
8176 	mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
8177 	mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
8178 	rcu_read_unlock();
8179 }
8180 
8181 bool mem_cgroup_zswap_writeback_enabled(struct mem_cgroup *memcg)
8182 {
8183 	/* if zswap is disabled, do not block pages going to the swapping device */
8184 	return !is_zswap_enabled() || !memcg || READ_ONCE(memcg->zswap_writeback);
8185 }
8186 
8187 static u64 zswap_current_read(struct cgroup_subsys_state *css,
8188 			      struct cftype *cft)
8189 {
8190 	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
8191 
8192 	mem_cgroup_flush_stats(memcg);
8193 	return memcg_page_state(memcg, MEMCG_ZSWAP_B);
8194 }
8195 
8196 static int zswap_max_show(struct seq_file *m, void *v)
8197 {
8198 	return seq_puts_memcg_tunable(m,
8199 		READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
8200 }
8201 
8202 static ssize_t zswap_max_write(struct kernfs_open_file *of,
8203 			       char *buf, size_t nbytes, loff_t off)
8204 {
8205 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8206 	unsigned long max;
8207 	int err;
8208 
8209 	buf = strstrip(buf);
8210 	err = page_counter_memparse(buf, "max", &max);
8211 	if (err)
8212 		return err;
8213 
8214 	xchg(&memcg->zswap_max, max);
8215 
8216 	return nbytes;
8217 }
8218 
8219 static int zswap_writeback_show(struct seq_file *m, void *v)
8220 {
8221 	struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
8222 
8223 	seq_printf(m, "%d\n", READ_ONCE(memcg->zswap_writeback));
8224 	return 0;
8225 }
8226 
8227 static ssize_t zswap_writeback_write(struct kernfs_open_file *of,
8228 				char *buf, size_t nbytes, loff_t off)
8229 {
8230 	struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8231 	int zswap_writeback;
8232 	ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_writeback);
8233 
8234 	if (parse_ret)
8235 		return parse_ret;
8236 
8237 	if (zswap_writeback != 0 && zswap_writeback != 1)
8238 		return -EINVAL;
8239 
8240 	WRITE_ONCE(memcg->zswap_writeback, zswap_writeback);
8241 	return nbytes;
8242 }
8243 
8244 static struct cftype zswap_files[] = {
8245 	{
8246 		.name = "zswap.current",
8247 		.flags = CFTYPE_NOT_ON_ROOT,
8248 		.read_u64 = zswap_current_read,
8249 	},
8250 	{
8251 		.name = "zswap.max",
8252 		.flags = CFTYPE_NOT_ON_ROOT,
8253 		.seq_show = zswap_max_show,
8254 		.write = zswap_max_write,
8255 	},
8256 	{
8257 		.name = "zswap.writeback",
8258 		.seq_show = zswap_writeback_show,
8259 		.write = zswap_writeback_write,
8260 	},
8261 	{ }	/* terminate */
8262 };
8263 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
8264 
8265 static int __init mem_cgroup_swap_init(void)
8266 {
8267 	if (mem_cgroup_disabled())
8268 		return 0;
8269 
8270 	WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
8271 	WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
8272 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8273 	WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
8274 #endif
8275 	return 0;
8276 }
8277 subsys_initcall(mem_cgroup_swap_init);
8278 
8279 #endif /* CONFIG_SWAP */
8280