xref: /linux/mm/ksm.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *	Izik Eidus
10  *	Andrea Arcangeli
11  *	Chris Wright
12  *	Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16 
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/jhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
42 
43 #include <asm/tlbflush.h>
44 #include "internal.h"
45 
46 #ifdef CONFIG_NUMA
47 #define NUMA(x)		(x)
48 #define DO_NUMA(x)	do { (x); } while (0)
49 #else
50 #define NUMA(x)		(0)
51 #define DO_NUMA(x)	do { } while (0)
52 #endif
53 
54 /**
55  * DOC: Overview
56  *
57  * A few notes about the KSM scanning process,
58  * to make it easier to understand the data structures below:
59  *
60  * In order to reduce excessive scanning, KSM sorts the memory pages by their
61  * contents into a data structure that holds pointers to the pages' locations.
62  *
63  * Since the contents of the pages may change at any moment, KSM cannot just
64  * insert the pages into a normal sorted tree and expect it to find anything.
65  * Therefore KSM uses two data structures - the stable and the unstable tree.
66  *
67  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68  * by their contents.  Because each such page is write-protected, searching on
69  * this tree is fully assured to be working (except when pages are unmapped),
70  * and therefore this tree is called the stable tree.
71  *
72  * The stable tree node includes information required for reverse
73  * mapping from a KSM page to virtual addresses that map this page.
74  *
75  * In order to avoid large latencies of the rmap walks on KSM pages,
76  * KSM maintains two types of nodes in the stable tree:
77  *
78  * * the regular nodes that keep the reverse mapping structures in a
79  *   linked list
80  * * the "chains" that link nodes ("dups") that represent the same
81  *   write protected memory content, but each "dup" corresponds to a
82  *   different KSM page copy of that content
83  *
84  * Internally, the regular nodes, "dups" and "chains" are represented
85  * using the same :c:type:`struct stable_node` structure.
86  *
87  * In addition to the stable tree, KSM uses a second data structure called the
88  * unstable tree: this tree holds pointers to pages which have been found to
89  * be "unchanged for a period of time".  The unstable tree sorts these pages
90  * by their contents, but since they are not write-protected, KSM cannot rely
91  * upon the unstable tree to work correctly - the unstable tree is liable to
92  * be corrupted as its contents are modified, and so it is called unstable.
93  *
94  * KSM solves this problem by several techniques:
95  *
96  * 1) The unstable tree is flushed every time KSM completes scanning all
97  *    memory areas, and then the tree is rebuilt again from the beginning.
98  * 2) KSM will only insert into the unstable tree, pages whose hash value
99  *    has not changed since the previous scan of all memory areas.
100  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
101  *    colors of the nodes and not on their contents, assuring that even when
102  *    the tree gets "corrupted" it won't get out of balance, so scanning time
103  *    remains the same (also, searching and inserting nodes in an rbtree uses
104  *    the same algorithm, so we have no overhead when we flush and rebuild).
105  * 4) KSM never flushes the stable tree, which means that even if it were to
106  *    take 10 attempts to find a page in the unstable tree, once it is found,
107  *    it is secured in the stable tree.  (When we scan a new page, we first
108  *    compare it against the stable tree, and then against the unstable tree.)
109  *
110  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
111  * stable trees and multiple unstable trees: one of each for each NUMA node.
112  */
113 
114 /**
115  * struct mm_slot - ksm information per mm that is being scanned
116  * @link: link to the mm_slots hash list
117  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
118  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
119  * @mm: the mm that this information is valid for
120  */
121 struct mm_slot {
122 	struct hlist_node link;
123 	struct list_head mm_list;
124 	struct rmap_item *rmap_list;
125 	struct mm_struct *mm;
126 };
127 
128 /**
129  * struct ksm_scan - cursor for scanning
130  * @mm_slot: the current mm_slot we are scanning
131  * @address: the next address inside that to be scanned
132  * @rmap_list: link to the next rmap to be scanned in the rmap_list
133  * @seqnr: count of completed full scans (needed when removing unstable node)
134  *
135  * There is only the one ksm_scan instance of this cursor structure.
136  */
137 struct ksm_scan {
138 	struct mm_slot *mm_slot;
139 	unsigned long address;
140 	struct rmap_item **rmap_list;
141 	unsigned long seqnr;
142 };
143 
144 /**
145  * struct stable_node - node of the stable rbtree
146  * @node: rb node of this ksm page in the stable tree
147  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
148  * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
149  * @list: linked into migrate_nodes, pending placement in the proper node tree
150  * @hlist: hlist head of rmap_items using this ksm page
151  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
152  * @chain_prune_time: time of the last full garbage collection
153  * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
154  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
155  */
156 struct stable_node {
157 	union {
158 		struct rb_node node;	/* when node of stable tree */
159 		struct {		/* when listed for migration */
160 			struct list_head *head;
161 			struct {
162 				struct hlist_node hlist_dup;
163 				struct list_head list;
164 			};
165 		};
166 	};
167 	struct hlist_head hlist;
168 	union {
169 		unsigned long kpfn;
170 		unsigned long chain_prune_time;
171 	};
172 	/*
173 	 * STABLE_NODE_CHAIN can be any negative number in
174 	 * rmap_hlist_len negative range, but better not -1 to be able
175 	 * to reliably detect underflows.
176 	 */
177 #define STABLE_NODE_CHAIN -1024
178 	int rmap_hlist_len;
179 #ifdef CONFIG_NUMA
180 	int nid;
181 #endif
182 };
183 
184 /**
185  * struct rmap_item - reverse mapping item for virtual addresses
186  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
187  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
188  * @nid: NUMA node id of unstable tree in which linked (may not match page)
189  * @mm: the memory structure this rmap_item is pointing into
190  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
191  * @oldchecksum: previous checksum of the page at that virtual address
192  * @node: rb node of this rmap_item in the unstable tree
193  * @head: pointer to stable_node heading this list in the stable tree
194  * @hlist: link into hlist of rmap_items hanging off that stable_node
195  */
196 struct rmap_item {
197 	struct rmap_item *rmap_list;
198 	union {
199 		struct anon_vma *anon_vma;	/* when stable */
200 #ifdef CONFIG_NUMA
201 		int nid;		/* when node of unstable tree */
202 #endif
203 	};
204 	struct mm_struct *mm;
205 	unsigned long address;		/* + low bits used for flags below */
206 	unsigned int oldchecksum;	/* when unstable */
207 	union {
208 		struct rb_node node;	/* when node of unstable tree */
209 		struct {		/* when listed from stable tree */
210 			struct stable_node *head;
211 			struct hlist_node hlist;
212 		};
213 	};
214 };
215 
216 #define SEQNR_MASK	0x0ff	/* low bits of unstable tree seqnr */
217 #define UNSTABLE_FLAG	0x100	/* is a node of the unstable tree */
218 #define STABLE_FLAG	0x200	/* is listed from the stable tree */
219 #define KSM_FLAG_MASK	(SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
220 				/* to mask all the flags */
221 
222 /* The stable and unstable tree heads */
223 static struct rb_root one_stable_tree[1] = { RB_ROOT };
224 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
225 static struct rb_root *root_stable_tree = one_stable_tree;
226 static struct rb_root *root_unstable_tree = one_unstable_tree;
227 
228 /* Recently migrated nodes of stable tree, pending proper placement */
229 static LIST_HEAD(migrate_nodes);
230 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
231 
232 #define MM_SLOTS_HASH_BITS 10
233 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
234 
235 static struct mm_slot ksm_mm_head = {
236 	.mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
237 };
238 static struct ksm_scan ksm_scan = {
239 	.mm_slot = &ksm_mm_head,
240 };
241 
242 static struct kmem_cache *rmap_item_cache;
243 static struct kmem_cache *stable_node_cache;
244 static struct kmem_cache *mm_slot_cache;
245 
246 /* The number of nodes in the stable tree */
247 static unsigned long ksm_pages_shared;
248 
249 /* The number of page slots additionally sharing those nodes */
250 static unsigned long ksm_pages_sharing;
251 
252 /* The number of nodes in the unstable tree */
253 static unsigned long ksm_pages_unshared;
254 
255 /* The number of rmap_items in use: to calculate pages_volatile */
256 static unsigned long ksm_rmap_items;
257 
258 /* The number of stable_node chains */
259 static unsigned long ksm_stable_node_chains;
260 
261 /* The number of stable_node dups linked to the stable_node chains */
262 static unsigned long ksm_stable_node_dups;
263 
264 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
265 static int ksm_stable_node_chains_prune_millisecs = 2000;
266 
267 /* Maximum number of page slots sharing a stable node */
268 static int ksm_max_page_sharing = 256;
269 
270 /* Number of pages ksmd should scan in one batch */
271 static unsigned int ksm_thread_pages_to_scan = 100;
272 
273 /* Milliseconds ksmd should sleep between batches */
274 static unsigned int ksm_thread_sleep_millisecs = 20;
275 
276 /* Checksum of an empty (zeroed) page */
277 static unsigned int zero_checksum __read_mostly;
278 
279 /* Whether to merge empty (zeroed) pages with actual zero pages */
280 static bool ksm_use_zero_pages __read_mostly;
281 
282 #ifdef CONFIG_NUMA
283 /* Zeroed when merging across nodes is not allowed */
284 static unsigned int ksm_merge_across_nodes = 1;
285 static int ksm_nr_node_ids = 1;
286 #else
287 #define ksm_merge_across_nodes	1U
288 #define ksm_nr_node_ids		1
289 #endif
290 
291 #define KSM_RUN_STOP	0
292 #define KSM_RUN_MERGE	1
293 #define KSM_RUN_UNMERGE	2
294 #define KSM_RUN_OFFLINE	4
295 static unsigned long ksm_run = KSM_RUN_STOP;
296 static void wait_while_offlining(void);
297 
298 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
299 static DEFINE_MUTEX(ksm_thread_mutex);
300 static DEFINE_SPINLOCK(ksm_mmlist_lock);
301 
302 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
303 		sizeof(struct __struct), __alignof__(struct __struct),\
304 		(__flags), NULL)
305 
306 static int __init ksm_slab_init(void)
307 {
308 	rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
309 	if (!rmap_item_cache)
310 		goto out;
311 
312 	stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
313 	if (!stable_node_cache)
314 		goto out_free1;
315 
316 	mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
317 	if (!mm_slot_cache)
318 		goto out_free2;
319 
320 	return 0;
321 
322 out_free2:
323 	kmem_cache_destroy(stable_node_cache);
324 out_free1:
325 	kmem_cache_destroy(rmap_item_cache);
326 out:
327 	return -ENOMEM;
328 }
329 
330 static void __init ksm_slab_free(void)
331 {
332 	kmem_cache_destroy(mm_slot_cache);
333 	kmem_cache_destroy(stable_node_cache);
334 	kmem_cache_destroy(rmap_item_cache);
335 	mm_slot_cache = NULL;
336 }
337 
338 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
339 {
340 	return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
341 }
342 
343 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
344 {
345 	return dup->head == STABLE_NODE_DUP_HEAD;
346 }
347 
348 static inline void stable_node_chain_add_dup(struct stable_node *dup,
349 					     struct stable_node *chain)
350 {
351 	VM_BUG_ON(is_stable_node_dup(dup));
352 	dup->head = STABLE_NODE_DUP_HEAD;
353 	VM_BUG_ON(!is_stable_node_chain(chain));
354 	hlist_add_head(&dup->hlist_dup, &chain->hlist);
355 	ksm_stable_node_dups++;
356 }
357 
358 static inline void __stable_node_dup_del(struct stable_node *dup)
359 {
360 	VM_BUG_ON(!is_stable_node_dup(dup));
361 	hlist_del(&dup->hlist_dup);
362 	ksm_stable_node_dups--;
363 }
364 
365 static inline void stable_node_dup_del(struct stable_node *dup)
366 {
367 	VM_BUG_ON(is_stable_node_chain(dup));
368 	if (is_stable_node_dup(dup))
369 		__stable_node_dup_del(dup);
370 	else
371 		rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
372 #ifdef CONFIG_DEBUG_VM
373 	dup->head = NULL;
374 #endif
375 }
376 
377 static inline struct rmap_item *alloc_rmap_item(void)
378 {
379 	struct rmap_item *rmap_item;
380 
381 	rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
382 						__GFP_NORETRY | __GFP_NOWARN);
383 	if (rmap_item)
384 		ksm_rmap_items++;
385 	return rmap_item;
386 }
387 
388 static inline void free_rmap_item(struct rmap_item *rmap_item)
389 {
390 	ksm_rmap_items--;
391 	rmap_item->mm = NULL;	/* debug safety */
392 	kmem_cache_free(rmap_item_cache, rmap_item);
393 }
394 
395 static inline struct stable_node *alloc_stable_node(void)
396 {
397 	/*
398 	 * The allocation can take too long with GFP_KERNEL when memory is under
399 	 * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
400 	 * grants access to memory reserves, helping to avoid this problem.
401 	 */
402 	return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
403 }
404 
405 static inline void free_stable_node(struct stable_node *stable_node)
406 {
407 	VM_BUG_ON(stable_node->rmap_hlist_len &&
408 		  !is_stable_node_chain(stable_node));
409 	kmem_cache_free(stable_node_cache, stable_node);
410 }
411 
412 static inline struct mm_slot *alloc_mm_slot(void)
413 {
414 	if (!mm_slot_cache)	/* initialization failed */
415 		return NULL;
416 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
417 }
418 
419 static inline void free_mm_slot(struct mm_slot *mm_slot)
420 {
421 	kmem_cache_free(mm_slot_cache, mm_slot);
422 }
423 
424 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
425 {
426 	struct mm_slot *slot;
427 
428 	hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
429 		if (slot->mm == mm)
430 			return slot;
431 
432 	return NULL;
433 }
434 
435 static void insert_to_mm_slots_hash(struct mm_struct *mm,
436 				    struct mm_slot *mm_slot)
437 {
438 	mm_slot->mm = mm;
439 	hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
440 }
441 
442 /*
443  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
444  * page tables after it has passed through ksm_exit() - which, if necessary,
445  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
446  * a special flag: they can just back out as soon as mm_users goes to zero.
447  * ksm_test_exit() is used throughout to make this test for exit: in some
448  * places for correctness, in some places just to avoid unnecessary work.
449  */
450 static inline bool ksm_test_exit(struct mm_struct *mm)
451 {
452 	return atomic_read(&mm->mm_users) == 0;
453 }
454 
455 /*
456  * We use break_ksm to break COW on a ksm page: it's a stripped down
457  *
458  *	if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
459  *		put_page(page);
460  *
461  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
462  * in case the application has unmapped and remapped mm,addr meanwhile.
463  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
464  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
465  *
466  * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
467  * of the process that owns 'vma'.  We also do not want to enforce
468  * protection keys here anyway.
469  */
470 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
471 {
472 	struct page *page;
473 	vm_fault_t ret = 0;
474 
475 	do {
476 		cond_resched();
477 		page = follow_page(vma, addr,
478 				FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
479 		if (IS_ERR_OR_NULL(page))
480 			break;
481 		if (PageKsm(page))
482 			ret = handle_mm_fault(vma, addr,
483 					FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
484 		else
485 			ret = VM_FAULT_WRITE;
486 		put_page(page);
487 	} while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
488 	/*
489 	 * We must loop because handle_mm_fault() may back out if there's
490 	 * any difficulty e.g. if pte accessed bit gets updated concurrently.
491 	 *
492 	 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
493 	 * COW has been broken, even if the vma does not permit VM_WRITE;
494 	 * but note that a concurrent fault might break PageKsm for us.
495 	 *
496 	 * VM_FAULT_SIGBUS could occur if we race with truncation of the
497 	 * backing file, which also invalidates anonymous pages: that's
498 	 * okay, that truncation will have unmapped the PageKsm for us.
499 	 *
500 	 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
501 	 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
502 	 * current task has TIF_MEMDIE set, and will be OOM killed on return
503 	 * to user; and ksmd, having no mm, would never be chosen for that.
504 	 *
505 	 * But if the mm is in a limited mem_cgroup, then the fault may fail
506 	 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
507 	 * even ksmd can fail in this way - though it's usually breaking ksm
508 	 * just to undo a merge it made a moment before, so unlikely to oom.
509 	 *
510 	 * That's a pity: we might therefore have more kernel pages allocated
511 	 * than we're counting as nodes in the stable tree; but ksm_do_scan
512 	 * will retry to break_cow on each pass, so should recover the page
513 	 * in due course.  The important thing is to not let VM_MERGEABLE
514 	 * be cleared while any such pages might remain in the area.
515 	 */
516 	return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
517 }
518 
519 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
520 		unsigned long addr)
521 {
522 	struct vm_area_struct *vma;
523 	if (ksm_test_exit(mm))
524 		return NULL;
525 	vma = find_vma(mm, addr);
526 	if (!vma || vma->vm_start > addr)
527 		return NULL;
528 	if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
529 		return NULL;
530 	return vma;
531 }
532 
533 static void break_cow(struct rmap_item *rmap_item)
534 {
535 	struct mm_struct *mm = rmap_item->mm;
536 	unsigned long addr = rmap_item->address;
537 	struct vm_area_struct *vma;
538 
539 	/*
540 	 * It is not an accident that whenever we want to break COW
541 	 * to undo, we also need to drop a reference to the anon_vma.
542 	 */
543 	put_anon_vma(rmap_item->anon_vma);
544 
545 	down_read(&mm->mmap_sem);
546 	vma = find_mergeable_vma(mm, addr);
547 	if (vma)
548 		break_ksm(vma, addr);
549 	up_read(&mm->mmap_sem);
550 }
551 
552 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
553 {
554 	struct mm_struct *mm = rmap_item->mm;
555 	unsigned long addr = rmap_item->address;
556 	struct vm_area_struct *vma;
557 	struct page *page;
558 
559 	down_read(&mm->mmap_sem);
560 	vma = find_mergeable_vma(mm, addr);
561 	if (!vma)
562 		goto out;
563 
564 	page = follow_page(vma, addr, FOLL_GET);
565 	if (IS_ERR_OR_NULL(page))
566 		goto out;
567 	if (PageAnon(page)) {
568 		flush_anon_page(vma, page, addr);
569 		flush_dcache_page(page);
570 	} else {
571 		put_page(page);
572 out:
573 		page = NULL;
574 	}
575 	up_read(&mm->mmap_sem);
576 	return page;
577 }
578 
579 /*
580  * This helper is used for getting right index into array of tree roots.
581  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
582  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
583  * every node has its own stable and unstable tree.
584  */
585 static inline int get_kpfn_nid(unsigned long kpfn)
586 {
587 	return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
588 }
589 
590 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
591 						   struct rb_root *root)
592 {
593 	struct stable_node *chain = alloc_stable_node();
594 	VM_BUG_ON(is_stable_node_chain(dup));
595 	if (likely(chain)) {
596 		INIT_HLIST_HEAD(&chain->hlist);
597 		chain->chain_prune_time = jiffies;
598 		chain->rmap_hlist_len = STABLE_NODE_CHAIN;
599 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
600 		chain->nid = -1; /* debug */
601 #endif
602 		ksm_stable_node_chains++;
603 
604 		/*
605 		 * Put the stable node chain in the first dimension of
606 		 * the stable tree and at the same time remove the old
607 		 * stable node.
608 		 */
609 		rb_replace_node(&dup->node, &chain->node, root);
610 
611 		/*
612 		 * Move the old stable node to the second dimension
613 		 * queued in the hlist_dup. The invariant is that all
614 		 * dup stable_nodes in the chain->hlist point to pages
615 		 * that are wrprotected and have the exact same
616 		 * content.
617 		 */
618 		stable_node_chain_add_dup(dup, chain);
619 	}
620 	return chain;
621 }
622 
623 static inline void free_stable_node_chain(struct stable_node *chain,
624 					  struct rb_root *root)
625 {
626 	rb_erase(&chain->node, root);
627 	free_stable_node(chain);
628 	ksm_stable_node_chains--;
629 }
630 
631 static void remove_node_from_stable_tree(struct stable_node *stable_node)
632 {
633 	struct rmap_item *rmap_item;
634 
635 	/* check it's not STABLE_NODE_CHAIN or negative */
636 	BUG_ON(stable_node->rmap_hlist_len < 0);
637 
638 	hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
639 		if (rmap_item->hlist.next)
640 			ksm_pages_sharing--;
641 		else
642 			ksm_pages_shared--;
643 		VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
644 		stable_node->rmap_hlist_len--;
645 		put_anon_vma(rmap_item->anon_vma);
646 		rmap_item->address &= PAGE_MASK;
647 		cond_resched();
648 	}
649 
650 	/*
651 	 * We need the second aligned pointer of the migrate_nodes
652 	 * list_head to stay clear from the rb_parent_color union
653 	 * (aligned and different than any node) and also different
654 	 * from &migrate_nodes. This will verify that future list.h changes
655 	 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
656 	 */
657 #if defined(GCC_VERSION) && GCC_VERSION >= 40903
658 	BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
659 	BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
660 #endif
661 
662 	if (stable_node->head == &migrate_nodes)
663 		list_del(&stable_node->list);
664 	else
665 		stable_node_dup_del(stable_node);
666 	free_stable_node(stable_node);
667 }
668 
669 /*
670  * get_ksm_page: checks if the page indicated by the stable node
671  * is still its ksm page, despite having held no reference to it.
672  * In which case we can trust the content of the page, and it
673  * returns the gotten page; but if the page has now been zapped,
674  * remove the stale node from the stable tree and return NULL.
675  * But beware, the stable node's page might be being migrated.
676  *
677  * You would expect the stable_node to hold a reference to the ksm page.
678  * But if it increments the page's count, swapping out has to wait for
679  * ksmd to come around again before it can free the page, which may take
680  * seconds or even minutes: much too unresponsive.  So instead we use a
681  * "keyhole reference": access to the ksm page from the stable node peeps
682  * out through its keyhole to see if that page still holds the right key,
683  * pointing back to this stable node.  This relies on freeing a PageAnon
684  * page to reset its page->mapping to NULL, and relies on no other use of
685  * a page to put something that might look like our key in page->mapping.
686  * is on its way to being freed; but it is an anomaly to bear in mind.
687  */
688 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
689 {
690 	struct page *page;
691 	void *expected_mapping;
692 	unsigned long kpfn;
693 
694 	expected_mapping = (void *)((unsigned long)stable_node |
695 					PAGE_MAPPING_KSM);
696 again:
697 	kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
698 	page = pfn_to_page(kpfn);
699 	if (READ_ONCE(page->mapping) != expected_mapping)
700 		goto stale;
701 
702 	/*
703 	 * We cannot do anything with the page while its refcount is 0.
704 	 * Usually 0 means free, or tail of a higher-order page: in which
705 	 * case this node is no longer referenced, and should be freed;
706 	 * however, it might mean that the page is under page_ref_freeze().
707 	 * The __remove_mapping() case is easy, again the node is now stale;
708 	 * but if page is swapcache in migrate_page_move_mapping(), it might
709 	 * still be our page, in which case it's essential to keep the node.
710 	 */
711 	while (!get_page_unless_zero(page)) {
712 		/*
713 		 * Another check for page->mapping != expected_mapping would
714 		 * work here too.  We have chosen the !PageSwapCache test to
715 		 * optimize the common case, when the page is or is about to
716 		 * be freed: PageSwapCache is cleared (under spin_lock_irq)
717 		 * in the ref_freeze section of __remove_mapping(); but Anon
718 		 * page->mapping reset to NULL later, in free_pages_prepare().
719 		 */
720 		if (!PageSwapCache(page))
721 			goto stale;
722 		cpu_relax();
723 	}
724 
725 	if (READ_ONCE(page->mapping) != expected_mapping) {
726 		put_page(page);
727 		goto stale;
728 	}
729 
730 	if (lock_it) {
731 		lock_page(page);
732 		if (READ_ONCE(page->mapping) != expected_mapping) {
733 			unlock_page(page);
734 			put_page(page);
735 			goto stale;
736 		}
737 	}
738 	return page;
739 
740 stale:
741 	/*
742 	 * We come here from above when page->mapping or !PageSwapCache
743 	 * suggests that the node is stale; but it might be under migration.
744 	 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
745 	 * before checking whether node->kpfn has been changed.
746 	 */
747 	smp_rmb();
748 	if (READ_ONCE(stable_node->kpfn) != kpfn)
749 		goto again;
750 	remove_node_from_stable_tree(stable_node);
751 	return NULL;
752 }
753 
754 /*
755  * Removing rmap_item from stable or unstable tree.
756  * This function will clean the information from the stable/unstable tree.
757  */
758 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
759 {
760 	if (rmap_item->address & STABLE_FLAG) {
761 		struct stable_node *stable_node;
762 		struct page *page;
763 
764 		stable_node = rmap_item->head;
765 		page = get_ksm_page(stable_node, true);
766 		if (!page)
767 			goto out;
768 
769 		hlist_del(&rmap_item->hlist);
770 		unlock_page(page);
771 		put_page(page);
772 
773 		if (!hlist_empty(&stable_node->hlist))
774 			ksm_pages_sharing--;
775 		else
776 			ksm_pages_shared--;
777 		VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
778 		stable_node->rmap_hlist_len--;
779 
780 		put_anon_vma(rmap_item->anon_vma);
781 		rmap_item->address &= PAGE_MASK;
782 
783 	} else if (rmap_item->address & UNSTABLE_FLAG) {
784 		unsigned char age;
785 		/*
786 		 * Usually ksmd can and must skip the rb_erase, because
787 		 * root_unstable_tree was already reset to RB_ROOT.
788 		 * But be careful when an mm is exiting: do the rb_erase
789 		 * if this rmap_item was inserted by this scan, rather
790 		 * than left over from before.
791 		 */
792 		age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
793 		BUG_ON(age > 1);
794 		if (!age)
795 			rb_erase(&rmap_item->node,
796 				 root_unstable_tree + NUMA(rmap_item->nid));
797 		ksm_pages_unshared--;
798 		rmap_item->address &= PAGE_MASK;
799 	}
800 out:
801 	cond_resched();		/* we're called from many long loops */
802 }
803 
804 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
805 				       struct rmap_item **rmap_list)
806 {
807 	while (*rmap_list) {
808 		struct rmap_item *rmap_item = *rmap_list;
809 		*rmap_list = rmap_item->rmap_list;
810 		remove_rmap_item_from_tree(rmap_item);
811 		free_rmap_item(rmap_item);
812 	}
813 }
814 
815 /*
816  * Though it's very tempting to unmerge rmap_items from stable tree rather
817  * than check every pte of a given vma, the locking doesn't quite work for
818  * that - an rmap_item is assigned to the stable tree after inserting ksm
819  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
820  * rmap_items from parent to child at fork time (so as not to waste time
821  * if exit comes before the next scan reaches it).
822  *
823  * Similarly, although we'd like to remove rmap_items (so updating counts
824  * and freeing memory) when unmerging an area, it's easier to leave that
825  * to the next pass of ksmd - consider, for example, how ksmd might be
826  * in cmp_and_merge_page on one of the rmap_items we would be removing.
827  */
828 static int unmerge_ksm_pages(struct vm_area_struct *vma,
829 			     unsigned long start, unsigned long end)
830 {
831 	unsigned long addr;
832 	int err = 0;
833 
834 	for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
835 		if (ksm_test_exit(vma->vm_mm))
836 			break;
837 		if (signal_pending(current))
838 			err = -ERESTARTSYS;
839 		else
840 			err = break_ksm(vma, addr);
841 	}
842 	return err;
843 }
844 
845 static inline struct stable_node *page_stable_node(struct page *page)
846 {
847 	return PageKsm(page) ? page_rmapping(page) : NULL;
848 }
849 
850 static inline void set_page_stable_node(struct page *page,
851 					struct stable_node *stable_node)
852 {
853 	page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
854 }
855 
856 #ifdef CONFIG_SYSFS
857 /*
858  * Only called through the sysfs control interface:
859  */
860 static int remove_stable_node(struct stable_node *stable_node)
861 {
862 	struct page *page;
863 	int err;
864 
865 	page = get_ksm_page(stable_node, true);
866 	if (!page) {
867 		/*
868 		 * get_ksm_page did remove_node_from_stable_tree itself.
869 		 */
870 		return 0;
871 	}
872 
873 	if (WARN_ON_ONCE(page_mapped(page))) {
874 		/*
875 		 * This should not happen: but if it does, just refuse to let
876 		 * merge_across_nodes be switched - there is no need to panic.
877 		 */
878 		err = -EBUSY;
879 	} else {
880 		/*
881 		 * The stable node did not yet appear stale to get_ksm_page(),
882 		 * since that allows for an unmapped ksm page to be recognized
883 		 * right up until it is freed; but the node is safe to remove.
884 		 * This page might be in a pagevec waiting to be freed,
885 		 * or it might be PageSwapCache (perhaps under writeback),
886 		 * or it might have been removed from swapcache a moment ago.
887 		 */
888 		set_page_stable_node(page, NULL);
889 		remove_node_from_stable_tree(stable_node);
890 		err = 0;
891 	}
892 
893 	unlock_page(page);
894 	put_page(page);
895 	return err;
896 }
897 
898 static int remove_stable_node_chain(struct stable_node *stable_node,
899 				    struct rb_root *root)
900 {
901 	struct stable_node *dup;
902 	struct hlist_node *hlist_safe;
903 
904 	if (!is_stable_node_chain(stable_node)) {
905 		VM_BUG_ON(is_stable_node_dup(stable_node));
906 		if (remove_stable_node(stable_node))
907 			return true;
908 		else
909 			return false;
910 	}
911 
912 	hlist_for_each_entry_safe(dup, hlist_safe,
913 				  &stable_node->hlist, hlist_dup) {
914 		VM_BUG_ON(!is_stable_node_dup(dup));
915 		if (remove_stable_node(dup))
916 			return true;
917 	}
918 	BUG_ON(!hlist_empty(&stable_node->hlist));
919 	free_stable_node_chain(stable_node, root);
920 	return false;
921 }
922 
923 static int remove_all_stable_nodes(void)
924 {
925 	struct stable_node *stable_node, *next;
926 	int nid;
927 	int err = 0;
928 
929 	for (nid = 0; nid < ksm_nr_node_ids; nid++) {
930 		while (root_stable_tree[nid].rb_node) {
931 			stable_node = rb_entry(root_stable_tree[nid].rb_node,
932 						struct stable_node, node);
933 			if (remove_stable_node_chain(stable_node,
934 						     root_stable_tree + nid)) {
935 				err = -EBUSY;
936 				break;	/* proceed to next nid */
937 			}
938 			cond_resched();
939 		}
940 	}
941 	list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
942 		if (remove_stable_node(stable_node))
943 			err = -EBUSY;
944 		cond_resched();
945 	}
946 	return err;
947 }
948 
949 static int unmerge_and_remove_all_rmap_items(void)
950 {
951 	struct mm_slot *mm_slot;
952 	struct mm_struct *mm;
953 	struct vm_area_struct *vma;
954 	int err = 0;
955 
956 	spin_lock(&ksm_mmlist_lock);
957 	ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
958 						struct mm_slot, mm_list);
959 	spin_unlock(&ksm_mmlist_lock);
960 
961 	for (mm_slot = ksm_scan.mm_slot;
962 			mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
963 		mm = mm_slot->mm;
964 		down_read(&mm->mmap_sem);
965 		for (vma = mm->mmap; vma; vma = vma->vm_next) {
966 			if (ksm_test_exit(mm))
967 				break;
968 			if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
969 				continue;
970 			err = unmerge_ksm_pages(vma,
971 						vma->vm_start, vma->vm_end);
972 			if (err)
973 				goto error;
974 		}
975 
976 		remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
977 		up_read(&mm->mmap_sem);
978 
979 		spin_lock(&ksm_mmlist_lock);
980 		ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
981 						struct mm_slot, mm_list);
982 		if (ksm_test_exit(mm)) {
983 			hash_del(&mm_slot->link);
984 			list_del(&mm_slot->mm_list);
985 			spin_unlock(&ksm_mmlist_lock);
986 
987 			free_mm_slot(mm_slot);
988 			clear_bit(MMF_VM_MERGEABLE, &mm->flags);
989 			mmdrop(mm);
990 		} else
991 			spin_unlock(&ksm_mmlist_lock);
992 	}
993 
994 	/* Clean up stable nodes, but don't worry if some are still busy */
995 	remove_all_stable_nodes();
996 	ksm_scan.seqnr = 0;
997 	return 0;
998 
999 error:
1000 	up_read(&mm->mmap_sem);
1001 	spin_lock(&ksm_mmlist_lock);
1002 	ksm_scan.mm_slot = &ksm_mm_head;
1003 	spin_unlock(&ksm_mmlist_lock);
1004 	return err;
1005 }
1006 #endif /* CONFIG_SYSFS */
1007 
1008 static u32 calc_checksum(struct page *page)
1009 {
1010 	u32 checksum;
1011 	void *addr = kmap_atomic(page);
1012 	checksum = jhash2(addr, PAGE_SIZE / 4, 17);
1013 	kunmap_atomic(addr);
1014 	return checksum;
1015 }
1016 
1017 static int memcmp_pages(struct page *page1, struct page *page2)
1018 {
1019 	char *addr1, *addr2;
1020 	int ret;
1021 
1022 	addr1 = kmap_atomic(page1);
1023 	addr2 = kmap_atomic(page2);
1024 	ret = memcmp(addr1, addr2, PAGE_SIZE);
1025 	kunmap_atomic(addr2);
1026 	kunmap_atomic(addr1);
1027 	return ret;
1028 }
1029 
1030 static inline int pages_identical(struct page *page1, struct page *page2)
1031 {
1032 	return !memcmp_pages(page1, page2);
1033 }
1034 
1035 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1036 			      pte_t *orig_pte)
1037 {
1038 	struct mm_struct *mm = vma->vm_mm;
1039 	struct page_vma_mapped_walk pvmw = {
1040 		.page = page,
1041 		.vma = vma,
1042 	};
1043 	int swapped;
1044 	int err = -EFAULT;
1045 	unsigned long mmun_start;	/* For mmu_notifiers */
1046 	unsigned long mmun_end;		/* For mmu_notifiers */
1047 
1048 	pvmw.address = page_address_in_vma(page, vma);
1049 	if (pvmw.address == -EFAULT)
1050 		goto out;
1051 
1052 	BUG_ON(PageTransCompound(page));
1053 
1054 	mmun_start = pvmw.address;
1055 	mmun_end   = pvmw.address + PAGE_SIZE;
1056 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1057 
1058 	if (!page_vma_mapped_walk(&pvmw))
1059 		goto out_mn;
1060 	if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1061 		goto out_unlock;
1062 
1063 	if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1064 	    (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1065 						mm_tlb_flush_pending(mm)) {
1066 		pte_t entry;
1067 
1068 		swapped = PageSwapCache(page);
1069 		flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1070 		/*
1071 		 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1072 		 * take any lock, therefore the check that we are going to make
1073 		 * with the pagecount against the mapcount is racey and
1074 		 * O_DIRECT can happen right after the check.
1075 		 * So we clear the pte and flush the tlb before the check
1076 		 * this assure us that no O_DIRECT can happen after the check
1077 		 * or in the middle of the check.
1078 		 *
1079 		 * No need to notify as we are downgrading page table to read
1080 		 * only not changing it to point to a new page.
1081 		 *
1082 		 * See Documentation/vm/mmu_notifier.rst
1083 		 */
1084 		entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1085 		/*
1086 		 * Check that no O_DIRECT or similar I/O is in progress on the
1087 		 * page
1088 		 */
1089 		if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1090 			set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1091 			goto out_unlock;
1092 		}
1093 		if (pte_dirty(entry))
1094 			set_page_dirty(page);
1095 
1096 		if (pte_protnone(entry))
1097 			entry = pte_mkclean(pte_clear_savedwrite(entry));
1098 		else
1099 			entry = pte_mkclean(pte_wrprotect(entry));
1100 		set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1101 	}
1102 	*orig_pte = *pvmw.pte;
1103 	err = 0;
1104 
1105 out_unlock:
1106 	page_vma_mapped_walk_done(&pvmw);
1107 out_mn:
1108 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1109 out:
1110 	return err;
1111 }
1112 
1113 /**
1114  * replace_page - replace page in vma by new ksm page
1115  * @vma:      vma that holds the pte pointing to page
1116  * @page:     the page we are replacing by kpage
1117  * @kpage:    the ksm page we replace page by
1118  * @orig_pte: the original value of the pte
1119  *
1120  * Returns 0 on success, -EFAULT on failure.
1121  */
1122 static int replace_page(struct vm_area_struct *vma, struct page *page,
1123 			struct page *kpage, pte_t orig_pte)
1124 {
1125 	struct mm_struct *mm = vma->vm_mm;
1126 	pmd_t *pmd;
1127 	pte_t *ptep;
1128 	pte_t newpte;
1129 	spinlock_t *ptl;
1130 	unsigned long addr;
1131 	int err = -EFAULT;
1132 	unsigned long mmun_start;	/* For mmu_notifiers */
1133 	unsigned long mmun_end;		/* For mmu_notifiers */
1134 
1135 	addr = page_address_in_vma(page, vma);
1136 	if (addr == -EFAULT)
1137 		goto out;
1138 
1139 	pmd = mm_find_pmd(mm, addr);
1140 	if (!pmd)
1141 		goto out;
1142 
1143 	mmun_start = addr;
1144 	mmun_end   = addr + PAGE_SIZE;
1145 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1146 
1147 	ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1148 	if (!pte_same(*ptep, orig_pte)) {
1149 		pte_unmap_unlock(ptep, ptl);
1150 		goto out_mn;
1151 	}
1152 
1153 	/*
1154 	 * No need to check ksm_use_zero_pages here: we can only have a
1155 	 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1156 	 */
1157 	if (!is_zero_pfn(page_to_pfn(kpage))) {
1158 		get_page(kpage);
1159 		page_add_anon_rmap(kpage, vma, addr, false);
1160 		newpte = mk_pte(kpage, vma->vm_page_prot);
1161 	} else {
1162 		newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1163 					       vma->vm_page_prot));
1164 		/*
1165 		 * We're replacing an anonymous page with a zero page, which is
1166 		 * not anonymous. We need to do proper accounting otherwise we
1167 		 * will get wrong values in /proc, and a BUG message in dmesg
1168 		 * when tearing down the mm.
1169 		 */
1170 		dec_mm_counter(mm, MM_ANONPAGES);
1171 	}
1172 
1173 	flush_cache_page(vma, addr, pte_pfn(*ptep));
1174 	/*
1175 	 * No need to notify as we are replacing a read only page with another
1176 	 * read only page with the same content.
1177 	 *
1178 	 * See Documentation/vm/mmu_notifier.rst
1179 	 */
1180 	ptep_clear_flush(vma, addr, ptep);
1181 	set_pte_at_notify(mm, addr, ptep, newpte);
1182 
1183 	page_remove_rmap(page, false);
1184 	if (!page_mapped(page))
1185 		try_to_free_swap(page);
1186 	put_page(page);
1187 
1188 	pte_unmap_unlock(ptep, ptl);
1189 	err = 0;
1190 out_mn:
1191 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1192 out:
1193 	return err;
1194 }
1195 
1196 /*
1197  * try_to_merge_one_page - take two pages and merge them into one
1198  * @vma: the vma that holds the pte pointing to page
1199  * @page: the PageAnon page that we want to replace with kpage
1200  * @kpage: the PageKsm page that we want to map instead of page,
1201  *         or NULL the first time when we want to use page as kpage.
1202  *
1203  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1204  */
1205 static int try_to_merge_one_page(struct vm_area_struct *vma,
1206 				 struct page *page, struct page *kpage)
1207 {
1208 	pte_t orig_pte = __pte(0);
1209 	int err = -EFAULT;
1210 
1211 	if (page == kpage)			/* ksm page forked */
1212 		return 0;
1213 
1214 	if (!PageAnon(page))
1215 		goto out;
1216 
1217 	/*
1218 	 * We need the page lock to read a stable PageSwapCache in
1219 	 * write_protect_page().  We use trylock_page() instead of
1220 	 * lock_page() because we don't want to wait here - we
1221 	 * prefer to continue scanning and merging different pages,
1222 	 * then come back to this page when it is unlocked.
1223 	 */
1224 	if (!trylock_page(page))
1225 		goto out;
1226 
1227 	if (PageTransCompound(page)) {
1228 		if (split_huge_page(page))
1229 			goto out_unlock;
1230 	}
1231 
1232 	/*
1233 	 * If this anonymous page is mapped only here, its pte may need
1234 	 * to be write-protected.  If it's mapped elsewhere, all of its
1235 	 * ptes are necessarily already write-protected.  But in either
1236 	 * case, we need to lock and check page_count is not raised.
1237 	 */
1238 	if (write_protect_page(vma, page, &orig_pte) == 0) {
1239 		if (!kpage) {
1240 			/*
1241 			 * While we hold page lock, upgrade page from
1242 			 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1243 			 * stable_tree_insert() will update stable_node.
1244 			 */
1245 			set_page_stable_node(page, NULL);
1246 			mark_page_accessed(page);
1247 			/*
1248 			 * Page reclaim just frees a clean page with no dirty
1249 			 * ptes: make sure that the ksm page would be swapped.
1250 			 */
1251 			if (!PageDirty(page))
1252 				SetPageDirty(page);
1253 			err = 0;
1254 		} else if (pages_identical(page, kpage))
1255 			err = replace_page(vma, page, kpage, orig_pte);
1256 	}
1257 
1258 	if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1259 		munlock_vma_page(page);
1260 		if (!PageMlocked(kpage)) {
1261 			unlock_page(page);
1262 			lock_page(kpage);
1263 			mlock_vma_page(kpage);
1264 			page = kpage;		/* for final unlock */
1265 		}
1266 	}
1267 
1268 out_unlock:
1269 	unlock_page(page);
1270 out:
1271 	return err;
1272 }
1273 
1274 /*
1275  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1276  * but no new kernel page is allocated: kpage must already be a ksm page.
1277  *
1278  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1279  */
1280 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1281 				      struct page *page, struct page *kpage)
1282 {
1283 	struct mm_struct *mm = rmap_item->mm;
1284 	struct vm_area_struct *vma;
1285 	int err = -EFAULT;
1286 
1287 	down_read(&mm->mmap_sem);
1288 	vma = find_mergeable_vma(mm, rmap_item->address);
1289 	if (!vma)
1290 		goto out;
1291 
1292 	err = try_to_merge_one_page(vma, page, kpage);
1293 	if (err)
1294 		goto out;
1295 
1296 	/* Unstable nid is in union with stable anon_vma: remove first */
1297 	remove_rmap_item_from_tree(rmap_item);
1298 
1299 	/* Must get reference to anon_vma while still holding mmap_sem */
1300 	rmap_item->anon_vma = vma->anon_vma;
1301 	get_anon_vma(vma->anon_vma);
1302 out:
1303 	up_read(&mm->mmap_sem);
1304 	return err;
1305 }
1306 
1307 /*
1308  * try_to_merge_two_pages - take two identical pages and prepare them
1309  * to be merged into one page.
1310  *
1311  * This function returns the kpage if we successfully merged two identical
1312  * pages into one ksm page, NULL otherwise.
1313  *
1314  * Note that this function upgrades page to ksm page: if one of the pages
1315  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1316  */
1317 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1318 					   struct page *page,
1319 					   struct rmap_item *tree_rmap_item,
1320 					   struct page *tree_page)
1321 {
1322 	int err;
1323 
1324 	err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1325 	if (!err) {
1326 		err = try_to_merge_with_ksm_page(tree_rmap_item,
1327 							tree_page, page);
1328 		/*
1329 		 * If that fails, we have a ksm page with only one pte
1330 		 * pointing to it: so break it.
1331 		 */
1332 		if (err)
1333 			break_cow(rmap_item);
1334 	}
1335 	return err ? NULL : page;
1336 }
1337 
1338 static __always_inline
1339 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1340 {
1341 	VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1342 	/*
1343 	 * Check that at least one mapping still exists, otherwise
1344 	 * there's no much point to merge and share with this
1345 	 * stable_node, as the underlying tree_page of the other
1346 	 * sharer is going to be freed soon.
1347 	 */
1348 	return stable_node->rmap_hlist_len &&
1349 		stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1350 }
1351 
1352 static __always_inline
1353 bool is_page_sharing_candidate(struct stable_node *stable_node)
1354 {
1355 	return __is_page_sharing_candidate(stable_node, 0);
1356 }
1357 
1358 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1359 				    struct stable_node **_stable_node,
1360 				    struct rb_root *root,
1361 				    bool prune_stale_stable_nodes)
1362 {
1363 	struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1364 	struct hlist_node *hlist_safe;
1365 	struct page *_tree_page, *tree_page = NULL;
1366 	int nr = 0;
1367 	int found_rmap_hlist_len;
1368 
1369 	if (!prune_stale_stable_nodes ||
1370 	    time_before(jiffies, stable_node->chain_prune_time +
1371 			msecs_to_jiffies(
1372 				ksm_stable_node_chains_prune_millisecs)))
1373 		prune_stale_stable_nodes = false;
1374 	else
1375 		stable_node->chain_prune_time = jiffies;
1376 
1377 	hlist_for_each_entry_safe(dup, hlist_safe,
1378 				  &stable_node->hlist, hlist_dup) {
1379 		cond_resched();
1380 		/*
1381 		 * We must walk all stable_node_dup to prune the stale
1382 		 * stable nodes during lookup.
1383 		 *
1384 		 * get_ksm_page can drop the nodes from the
1385 		 * stable_node->hlist if they point to freed pages
1386 		 * (that's why we do a _safe walk). The "dup"
1387 		 * stable_node parameter itself will be freed from
1388 		 * under us if it returns NULL.
1389 		 */
1390 		_tree_page = get_ksm_page(dup, false);
1391 		if (!_tree_page)
1392 			continue;
1393 		nr += 1;
1394 		if (is_page_sharing_candidate(dup)) {
1395 			if (!found ||
1396 			    dup->rmap_hlist_len > found_rmap_hlist_len) {
1397 				if (found)
1398 					put_page(tree_page);
1399 				found = dup;
1400 				found_rmap_hlist_len = found->rmap_hlist_len;
1401 				tree_page = _tree_page;
1402 
1403 				/* skip put_page for found dup */
1404 				if (!prune_stale_stable_nodes)
1405 					break;
1406 				continue;
1407 			}
1408 		}
1409 		put_page(_tree_page);
1410 	}
1411 
1412 	if (found) {
1413 		/*
1414 		 * nr is counting all dups in the chain only if
1415 		 * prune_stale_stable_nodes is true, otherwise we may
1416 		 * break the loop at nr == 1 even if there are
1417 		 * multiple entries.
1418 		 */
1419 		if (prune_stale_stable_nodes && nr == 1) {
1420 			/*
1421 			 * If there's not just one entry it would
1422 			 * corrupt memory, better BUG_ON. In KSM
1423 			 * context with no lock held it's not even
1424 			 * fatal.
1425 			 */
1426 			BUG_ON(stable_node->hlist.first->next);
1427 
1428 			/*
1429 			 * There's just one entry and it is below the
1430 			 * deduplication limit so drop the chain.
1431 			 */
1432 			rb_replace_node(&stable_node->node, &found->node,
1433 					root);
1434 			free_stable_node(stable_node);
1435 			ksm_stable_node_chains--;
1436 			ksm_stable_node_dups--;
1437 			/*
1438 			 * NOTE: the caller depends on the stable_node
1439 			 * to be equal to stable_node_dup if the chain
1440 			 * was collapsed.
1441 			 */
1442 			*_stable_node = found;
1443 			/*
1444 			 * Just for robustneess as stable_node is
1445 			 * otherwise left as a stable pointer, the
1446 			 * compiler shall optimize it away at build
1447 			 * time.
1448 			 */
1449 			stable_node = NULL;
1450 		} else if (stable_node->hlist.first != &found->hlist_dup &&
1451 			   __is_page_sharing_candidate(found, 1)) {
1452 			/*
1453 			 * If the found stable_node dup can accept one
1454 			 * more future merge (in addition to the one
1455 			 * that is underway) and is not at the head of
1456 			 * the chain, put it there so next search will
1457 			 * be quicker in the !prune_stale_stable_nodes
1458 			 * case.
1459 			 *
1460 			 * NOTE: it would be inaccurate to use nr > 1
1461 			 * instead of checking the hlist.first pointer
1462 			 * directly, because in the
1463 			 * prune_stale_stable_nodes case "nr" isn't
1464 			 * the position of the found dup in the chain,
1465 			 * but the total number of dups in the chain.
1466 			 */
1467 			hlist_del(&found->hlist_dup);
1468 			hlist_add_head(&found->hlist_dup,
1469 				       &stable_node->hlist);
1470 		}
1471 	}
1472 
1473 	*_stable_node_dup = found;
1474 	return tree_page;
1475 }
1476 
1477 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1478 					       struct rb_root *root)
1479 {
1480 	if (!is_stable_node_chain(stable_node))
1481 		return stable_node;
1482 	if (hlist_empty(&stable_node->hlist)) {
1483 		free_stable_node_chain(stable_node, root);
1484 		return NULL;
1485 	}
1486 	return hlist_entry(stable_node->hlist.first,
1487 			   typeof(*stable_node), hlist_dup);
1488 }
1489 
1490 /*
1491  * Like for get_ksm_page, this function can free the *_stable_node and
1492  * *_stable_node_dup if the returned tree_page is NULL.
1493  *
1494  * It can also free and overwrite *_stable_node with the found
1495  * stable_node_dup if the chain is collapsed (in which case
1496  * *_stable_node will be equal to *_stable_node_dup like if the chain
1497  * never existed). It's up to the caller to verify tree_page is not
1498  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1499  *
1500  * *_stable_node_dup is really a second output parameter of this
1501  * function and will be overwritten in all cases, the caller doesn't
1502  * need to initialize it.
1503  */
1504 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1505 					struct stable_node **_stable_node,
1506 					struct rb_root *root,
1507 					bool prune_stale_stable_nodes)
1508 {
1509 	struct stable_node *stable_node = *_stable_node;
1510 	if (!is_stable_node_chain(stable_node)) {
1511 		if (is_page_sharing_candidate(stable_node)) {
1512 			*_stable_node_dup = stable_node;
1513 			return get_ksm_page(stable_node, false);
1514 		}
1515 		/*
1516 		 * _stable_node_dup set to NULL means the stable_node
1517 		 * reached the ksm_max_page_sharing limit.
1518 		 */
1519 		*_stable_node_dup = NULL;
1520 		return NULL;
1521 	}
1522 	return stable_node_dup(_stable_node_dup, _stable_node, root,
1523 			       prune_stale_stable_nodes);
1524 }
1525 
1526 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1527 						struct stable_node **s_n,
1528 						struct rb_root *root)
1529 {
1530 	return __stable_node_chain(s_n_d, s_n, root, true);
1531 }
1532 
1533 static __always_inline struct page *chain(struct stable_node **s_n_d,
1534 					  struct stable_node *s_n,
1535 					  struct rb_root *root)
1536 {
1537 	struct stable_node *old_stable_node = s_n;
1538 	struct page *tree_page;
1539 
1540 	tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1541 	/* not pruning dups so s_n cannot have changed */
1542 	VM_BUG_ON(s_n != old_stable_node);
1543 	return tree_page;
1544 }
1545 
1546 /*
1547  * stable_tree_search - search for page inside the stable tree
1548  *
1549  * This function checks if there is a page inside the stable tree
1550  * with identical content to the page that we are scanning right now.
1551  *
1552  * This function returns the stable tree node of identical content if found,
1553  * NULL otherwise.
1554  */
1555 static struct page *stable_tree_search(struct page *page)
1556 {
1557 	int nid;
1558 	struct rb_root *root;
1559 	struct rb_node **new;
1560 	struct rb_node *parent;
1561 	struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1562 	struct stable_node *page_node;
1563 
1564 	page_node = page_stable_node(page);
1565 	if (page_node && page_node->head != &migrate_nodes) {
1566 		/* ksm page forked */
1567 		get_page(page);
1568 		return page;
1569 	}
1570 
1571 	nid = get_kpfn_nid(page_to_pfn(page));
1572 	root = root_stable_tree + nid;
1573 again:
1574 	new = &root->rb_node;
1575 	parent = NULL;
1576 
1577 	while (*new) {
1578 		struct page *tree_page;
1579 		int ret;
1580 
1581 		cond_resched();
1582 		stable_node = rb_entry(*new, struct stable_node, node);
1583 		stable_node_any = NULL;
1584 		tree_page = chain_prune(&stable_node_dup, &stable_node,	root);
1585 		/*
1586 		 * NOTE: stable_node may have been freed by
1587 		 * chain_prune() if the returned stable_node_dup is
1588 		 * not NULL. stable_node_dup may have been inserted in
1589 		 * the rbtree instead as a regular stable_node (in
1590 		 * order to collapse the stable_node chain if a single
1591 		 * stable_node dup was found in it). In such case the
1592 		 * stable_node is overwritten by the calleee to point
1593 		 * to the stable_node_dup that was collapsed in the
1594 		 * stable rbtree and stable_node will be equal to
1595 		 * stable_node_dup like if the chain never existed.
1596 		 */
1597 		if (!stable_node_dup) {
1598 			/*
1599 			 * Either all stable_node dups were full in
1600 			 * this stable_node chain, or this chain was
1601 			 * empty and should be rb_erased.
1602 			 */
1603 			stable_node_any = stable_node_dup_any(stable_node,
1604 							      root);
1605 			if (!stable_node_any) {
1606 				/* rb_erase just run */
1607 				goto again;
1608 			}
1609 			/*
1610 			 * Take any of the stable_node dups page of
1611 			 * this stable_node chain to let the tree walk
1612 			 * continue. All KSM pages belonging to the
1613 			 * stable_node dups in a stable_node chain
1614 			 * have the same content and they're
1615 			 * wrprotected at all times. Any will work
1616 			 * fine to continue the walk.
1617 			 */
1618 			tree_page = get_ksm_page(stable_node_any, false);
1619 		}
1620 		VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1621 		if (!tree_page) {
1622 			/*
1623 			 * If we walked over a stale stable_node,
1624 			 * get_ksm_page() will call rb_erase() and it
1625 			 * may rebalance the tree from under us. So
1626 			 * restart the search from scratch. Returning
1627 			 * NULL would be safe too, but we'd generate
1628 			 * false negative insertions just because some
1629 			 * stable_node was stale.
1630 			 */
1631 			goto again;
1632 		}
1633 
1634 		ret = memcmp_pages(page, tree_page);
1635 		put_page(tree_page);
1636 
1637 		parent = *new;
1638 		if (ret < 0)
1639 			new = &parent->rb_left;
1640 		else if (ret > 0)
1641 			new = &parent->rb_right;
1642 		else {
1643 			if (page_node) {
1644 				VM_BUG_ON(page_node->head != &migrate_nodes);
1645 				/*
1646 				 * Test if the migrated page should be merged
1647 				 * into a stable node dup. If the mapcount is
1648 				 * 1 we can migrate it with another KSM page
1649 				 * without adding it to the chain.
1650 				 */
1651 				if (page_mapcount(page) > 1)
1652 					goto chain_append;
1653 			}
1654 
1655 			if (!stable_node_dup) {
1656 				/*
1657 				 * If the stable_node is a chain and
1658 				 * we got a payload match in memcmp
1659 				 * but we cannot merge the scanned
1660 				 * page in any of the existing
1661 				 * stable_node dups because they're
1662 				 * all full, we need to wait the
1663 				 * scanned page to find itself a match
1664 				 * in the unstable tree to create a
1665 				 * brand new KSM page to add later to
1666 				 * the dups of this stable_node.
1667 				 */
1668 				return NULL;
1669 			}
1670 
1671 			/*
1672 			 * Lock and unlock the stable_node's page (which
1673 			 * might already have been migrated) so that page
1674 			 * migration is sure to notice its raised count.
1675 			 * It would be more elegant to return stable_node
1676 			 * than kpage, but that involves more changes.
1677 			 */
1678 			tree_page = get_ksm_page(stable_node_dup, true);
1679 			if (unlikely(!tree_page))
1680 				/*
1681 				 * The tree may have been rebalanced,
1682 				 * so re-evaluate parent and new.
1683 				 */
1684 				goto again;
1685 			unlock_page(tree_page);
1686 
1687 			if (get_kpfn_nid(stable_node_dup->kpfn) !=
1688 			    NUMA(stable_node_dup->nid)) {
1689 				put_page(tree_page);
1690 				goto replace;
1691 			}
1692 			return tree_page;
1693 		}
1694 	}
1695 
1696 	if (!page_node)
1697 		return NULL;
1698 
1699 	list_del(&page_node->list);
1700 	DO_NUMA(page_node->nid = nid);
1701 	rb_link_node(&page_node->node, parent, new);
1702 	rb_insert_color(&page_node->node, root);
1703 out:
1704 	if (is_page_sharing_candidate(page_node)) {
1705 		get_page(page);
1706 		return page;
1707 	} else
1708 		return NULL;
1709 
1710 replace:
1711 	/*
1712 	 * If stable_node was a chain and chain_prune collapsed it,
1713 	 * stable_node has been updated to be the new regular
1714 	 * stable_node. A collapse of the chain is indistinguishable
1715 	 * from the case there was no chain in the stable
1716 	 * rbtree. Otherwise stable_node is the chain and
1717 	 * stable_node_dup is the dup to replace.
1718 	 */
1719 	if (stable_node_dup == stable_node) {
1720 		VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1721 		VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1722 		/* there is no chain */
1723 		if (page_node) {
1724 			VM_BUG_ON(page_node->head != &migrate_nodes);
1725 			list_del(&page_node->list);
1726 			DO_NUMA(page_node->nid = nid);
1727 			rb_replace_node(&stable_node_dup->node,
1728 					&page_node->node,
1729 					root);
1730 			if (is_page_sharing_candidate(page_node))
1731 				get_page(page);
1732 			else
1733 				page = NULL;
1734 		} else {
1735 			rb_erase(&stable_node_dup->node, root);
1736 			page = NULL;
1737 		}
1738 	} else {
1739 		VM_BUG_ON(!is_stable_node_chain(stable_node));
1740 		__stable_node_dup_del(stable_node_dup);
1741 		if (page_node) {
1742 			VM_BUG_ON(page_node->head != &migrate_nodes);
1743 			list_del(&page_node->list);
1744 			DO_NUMA(page_node->nid = nid);
1745 			stable_node_chain_add_dup(page_node, stable_node);
1746 			if (is_page_sharing_candidate(page_node))
1747 				get_page(page);
1748 			else
1749 				page = NULL;
1750 		} else {
1751 			page = NULL;
1752 		}
1753 	}
1754 	stable_node_dup->head = &migrate_nodes;
1755 	list_add(&stable_node_dup->list, stable_node_dup->head);
1756 	return page;
1757 
1758 chain_append:
1759 	/* stable_node_dup could be null if it reached the limit */
1760 	if (!stable_node_dup)
1761 		stable_node_dup = stable_node_any;
1762 	/*
1763 	 * If stable_node was a chain and chain_prune collapsed it,
1764 	 * stable_node has been updated to be the new regular
1765 	 * stable_node. A collapse of the chain is indistinguishable
1766 	 * from the case there was no chain in the stable
1767 	 * rbtree. Otherwise stable_node is the chain and
1768 	 * stable_node_dup is the dup to replace.
1769 	 */
1770 	if (stable_node_dup == stable_node) {
1771 		VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1772 		VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1773 		/* chain is missing so create it */
1774 		stable_node = alloc_stable_node_chain(stable_node_dup,
1775 						      root);
1776 		if (!stable_node)
1777 			return NULL;
1778 	}
1779 	/*
1780 	 * Add this stable_node dup that was
1781 	 * migrated to the stable_node chain
1782 	 * of the current nid for this page
1783 	 * content.
1784 	 */
1785 	VM_BUG_ON(!is_stable_node_chain(stable_node));
1786 	VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1787 	VM_BUG_ON(page_node->head != &migrate_nodes);
1788 	list_del(&page_node->list);
1789 	DO_NUMA(page_node->nid = nid);
1790 	stable_node_chain_add_dup(page_node, stable_node);
1791 	goto out;
1792 }
1793 
1794 /*
1795  * stable_tree_insert - insert stable tree node pointing to new ksm page
1796  * into the stable tree.
1797  *
1798  * This function returns the stable tree node just allocated on success,
1799  * NULL otherwise.
1800  */
1801 static struct stable_node *stable_tree_insert(struct page *kpage)
1802 {
1803 	int nid;
1804 	unsigned long kpfn;
1805 	struct rb_root *root;
1806 	struct rb_node **new;
1807 	struct rb_node *parent;
1808 	struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1809 	bool need_chain = false;
1810 
1811 	kpfn = page_to_pfn(kpage);
1812 	nid = get_kpfn_nid(kpfn);
1813 	root = root_stable_tree + nid;
1814 again:
1815 	parent = NULL;
1816 	new = &root->rb_node;
1817 
1818 	while (*new) {
1819 		struct page *tree_page;
1820 		int ret;
1821 
1822 		cond_resched();
1823 		stable_node = rb_entry(*new, struct stable_node, node);
1824 		stable_node_any = NULL;
1825 		tree_page = chain(&stable_node_dup, stable_node, root);
1826 		if (!stable_node_dup) {
1827 			/*
1828 			 * Either all stable_node dups were full in
1829 			 * this stable_node chain, or this chain was
1830 			 * empty and should be rb_erased.
1831 			 */
1832 			stable_node_any = stable_node_dup_any(stable_node,
1833 							      root);
1834 			if (!stable_node_any) {
1835 				/* rb_erase just run */
1836 				goto again;
1837 			}
1838 			/*
1839 			 * Take any of the stable_node dups page of
1840 			 * this stable_node chain to let the tree walk
1841 			 * continue. All KSM pages belonging to the
1842 			 * stable_node dups in a stable_node chain
1843 			 * have the same content and they're
1844 			 * wrprotected at all times. Any will work
1845 			 * fine to continue the walk.
1846 			 */
1847 			tree_page = get_ksm_page(stable_node_any, false);
1848 		}
1849 		VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1850 		if (!tree_page) {
1851 			/*
1852 			 * If we walked over a stale stable_node,
1853 			 * get_ksm_page() will call rb_erase() and it
1854 			 * may rebalance the tree from under us. So
1855 			 * restart the search from scratch. Returning
1856 			 * NULL would be safe too, but we'd generate
1857 			 * false negative insertions just because some
1858 			 * stable_node was stale.
1859 			 */
1860 			goto again;
1861 		}
1862 
1863 		ret = memcmp_pages(kpage, tree_page);
1864 		put_page(tree_page);
1865 
1866 		parent = *new;
1867 		if (ret < 0)
1868 			new = &parent->rb_left;
1869 		else if (ret > 0)
1870 			new = &parent->rb_right;
1871 		else {
1872 			need_chain = true;
1873 			break;
1874 		}
1875 	}
1876 
1877 	stable_node_dup = alloc_stable_node();
1878 	if (!stable_node_dup)
1879 		return NULL;
1880 
1881 	INIT_HLIST_HEAD(&stable_node_dup->hlist);
1882 	stable_node_dup->kpfn = kpfn;
1883 	set_page_stable_node(kpage, stable_node_dup);
1884 	stable_node_dup->rmap_hlist_len = 0;
1885 	DO_NUMA(stable_node_dup->nid = nid);
1886 	if (!need_chain) {
1887 		rb_link_node(&stable_node_dup->node, parent, new);
1888 		rb_insert_color(&stable_node_dup->node, root);
1889 	} else {
1890 		if (!is_stable_node_chain(stable_node)) {
1891 			struct stable_node *orig = stable_node;
1892 			/* chain is missing so create it */
1893 			stable_node = alloc_stable_node_chain(orig, root);
1894 			if (!stable_node) {
1895 				free_stable_node(stable_node_dup);
1896 				return NULL;
1897 			}
1898 		}
1899 		stable_node_chain_add_dup(stable_node_dup, stable_node);
1900 	}
1901 
1902 	return stable_node_dup;
1903 }
1904 
1905 /*
1906  * unstable_tree_search_insert - search for identical page,
1907  * else insert rmap_item into the unstable tree.
1908  *
1909  * This function searches for a page in the unstable tree identical to the
1910  * page currently being scanned; and if no identical page is found in the
1911  * tree, we insert rmap_item as a new object into the unstable tree.
1912  *
1913  * This function returns pointer to rmap_item found to be identical
1914  * to the currently scanned page, NULL otherwise.
1915  *
1916  * This function does both searching and inserting, because they share
1917  * the same walking algorithm in an rbtree.
1918  */
1919 static
1920 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1921 					      struct page *page,
1922 					      struct page **tree_pagep)
1923 {
1924 	struct rb_node **new;
1925 	struct rb_root *root;
1926 	struct rb_node *parent = NULL;
1927 	int nid;
1928 
1929 	nid = get_kpfn_nid(page_to_pfn(page));
1930 	root = root_unstable_tree + nid;
1931 	new = &root->rb_node;
1932 
1933 	while (*new) {
1934 		struct rmap_item *tree_rmap_item;
1935 		struct page *tree_page;
1936 		int ret;
1937 
1938 		cond_resched();
1939 		tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1940 		tree_page = get_mergeable_page(tree_rmap_item);
1941 		if (!tree_page)
1942 			return NULL;
1943 
1944 		/*
1945 		 * Don't substitute a ksm page for a forked page.
1946 		 */
1947 		if (page == tree_page) {
1948 			put_page(tree_page);
1949 			return NULL;
1950 		}
1951 
1952 		ret = memcmp_pages(page, tree_page);
1953 
1954 		parent = *new;
1955 		if (ret < 0) {
1956 			put_page(tree_page);
1957 			new = &parent->rb_left;
1958 		} else if (ret > 0) {
1959 			put_page(tree_page);
1960 			new = &parent->rb_right;
1961 		} else if (!ksm_merge_across_nodes &&
1962 			   page_to_nid(tree_page) != nid) {
1963 			/*
1964 			 * If tree_page has been migrated to another NUMA node,
1965 			 * it will be flushed out and put in the right unstable
1966 			 * tree next time: only merge with it when across_nodes.
1967 			 */
1968 			put_page(tree_page);
1969 			return NULL;
1970 		} else {
1971 			*tree_pagep = tree_page;
1972 			return tree_rmap_item;
1973 		}
1974 	}
1975 
1976 	rmap_item->address |= UNSTABLE_FLAG;
1977 	rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1978 	DO_NUMA(rmap_item->nid = nid);
1979 	rb_link_node(&rmap_item->node, parent, new);
1980 	rb_insert_color(&rmap_item->node, root);
1981 
1982 	ksm_pages_unshared++;
1983 	return NULL;
1984 }
1985 
1986 /*
1987  * stable_tree_append - add another rmap_item to the linked list of
1988  * rmap_items hanging off a given node of the stable tree, all sharing
1989  * the same ksm page.
1990  */
1991 static void stable_tree_append(struct rmap_item *rmap_item,
1992 			       struct stable_node *stable_node,
1993 			       bool max_page_sharing_bypass)
1994 {
1995 	/*
1996 	 * rmap won't find this mapping if we don't insert the
1997 	 * rmap_item in the right stable_node
1998 	 * duplicate. page_migration could break later if rmap breaks,
1999 	 * so we can as well crash here. We really need to check for
2000 	 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2001 	 * for other negative values as an undeflow if detected here
2002 	 * for the first time (and not when decreasing rmap_hlist_len)
2003 	 * would be sign of memory corruption in the stable_node.
2004 	 */
2005 	BUG_ON(stable_node->rmap_hlist_len < 0);
2006 
2007 	stable_node->rmap_hlist_len++;
2008 	if (!max_page_sharing_bypass)
2009 		/* possibly non fatal but unexpected overflow, only warn */
2010 		WARN_ON_ONCE(stable_node->rmap_hlist_len >
2011 			     ksm_max_page_sharing);
2012 
2013 	rmap_item->head = stable_node;
2014 	rmap_item->address |= STABLE_FLAG;
2015 	hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2016 
2017 	if (rmap_item->hlist.next)
2018 		ksm_pages_sharing++;
2019 	else
2020 		ksm_pages_shared++;
2021 }
2022 
2023 /*
2024  * cmp_and_merge_page - first see if page can be merged into the stable tree;
2025  * if not, compare checksum to previous and if it's the same, see if page can
2026  * be inserted into the unstable tree, or merged with a page already there and
2027  * both transferred to the stable tree.
2028  *
2029  * @page: the page that we are searching identical page to.
2030  * @rmap_item: the reverse mapping into the virtual address of this page
2031  */
2032 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2033 {
2034 	struct mm_struct *mm = rmap_item->mm;
2035 	struct rmap_item *tree_rmap_item;
2036 	struct page *tree_page = NULL;
2037 	struct stable_node *stable_node;
2038 	struct page *kpage;
2039 	unsigned int checksum;
2040 	int err;
2041 	bool max_page_sharing_bypass = false;
2042 
2043 	stable_node = page_stable_node(page);
2044 	if (stable_node) {
2045 		if (stable_node->head != &migrate_nodes &&
2046 		    get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2047 		    NUMA(stable_node->nid)) {
2048 			stable_node_dup_del(stable_node);
2049 			stable_node->head = &migrate_nodes;
2050 			list_add(&stable_node->list, stable_node->head);
2051 		}
2052 		if (stable_node->head != &migrate_nodes &&
2053 		    rmap_item->head == stable_node)
2054 			return;
2055 		/*
2056 		 * If it's a KSM fork, allow it to go over the sharing limit
2057 		 * without warnings.
2058 		 */
2059 		if (!is_page_sharing_candidate(stable_node))
2060 			max_page_sharing_bypass = true;
2061 	}
2062 
2063 	/* We first start with searching the page inside the stable tree */
2064 	kpage = stable_tree_search(page);
2065 	if (kpage == page && rmap_item->head == stable_node) {
2066 		put_page(kpage);
2067 		return;
2068 	}
2069 
2070 	remove_rmap_item_from_tree(rmap_item);
2071 
2072 	if (kpage) {
2073 		err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2074 		if (!err) {
2075 			/*
2076 			 * The page was successfully merged:
2077 			 * add its rmap_item to the stable tree.
2078 			 */
2079 			lock_page(kpage);
2080 			stable_tree_append(rmap_item, page_stable_node(kpage),
2081 					   max_page_sharing_bypass);
2082 			unlock_page(kpage);
2083 		}
2084 		put_page(kpage);
2085 		return;
2086 	}
2087 
2088 	/*
2089 	 * If the hash value of the page has changed from the last time
2090 	 * we calculated it, this page is changing frequently: therefore we
2091 	 * don't want to insert it in the unstable tree, and we don't want
2092 	 * to waste our time searching for something identical to it there.
2093 	 */
2094 	checksum = calc_checksum(page);
2095 	if (rmap_item->oldchecksum != checksum) {
2096 		rmap_item->oldchecksum = checksum;
2097 		return;
2098 	}
2099 
2100 	/*
2101 	 * Same checksum as an empty page. We attempt to merge it with the
2102 	 * appropriate zero page if the user enabled this via sysfs.
2103 	 */
2104 	if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2105 		struct vm_area_struct *vma;
2106 
2107 		down_read(&mm->mmap_sem);
2108 		vma = find_mergeable_vma(mm, rmap_item->address);
2109 		err = try_to_merge_one_page(vma, page,
2110 					    ZERO_PAGE(rmap_item->address));
2111 		up_read(&mm->mmap_sem);
2112 		/*
2113 		 * In case of failure, the page was not really empty, so we
2114 		 * need to continue. Otherwise we're done.
2115 		 */
2116 		if (!err)
2117 			return;
2118 	}
2119 	tree_rmap_item =
2120 		unstable_tree_search_insert(rmap_item, page, &tree_page);
2121 	if (tree_rmap_item) {
2122 		bool split;
2123 
2124 		kpage = try_to_merge_two_pages(rmap_item, page,
2125 						tree_rmap_item, tree_page);
2126 		/*
2127 		 * If both pages we tried to merge belong to the same compound
2128 		 * page, then we actually ended up increasing the reference
2129 		 * count of the same compound page twice, and split_huge_page
2130 		 * failed.
2131 		 * Here we set a flag if that happened, and we use it later to
2132 		 * try split_huge_page again. Since we call put_page right
2133 		 * afterwards, the reference count will be correct and
2134 		 * split_huge_page should succeed.
2135 		 */
2136 		split = PageTransCompound(page)
2137 			&& compound_head(page) == compound_head(tree_page);
2138 		put_page(tree_page);
2139 		if (kpage) {
2140 			/*
2141 			 * The pages were successfully merged: insert new
2142 			 * node in the stable tree and add both rmap_items.
2143 			 */
2144 			lock_page(kpage);
2145 			stable_node = stable_tree_insert(kpage);
2146 			if (stable_node) {
2147 				stable_tree_append(tree_rmap_item, stable_node,
2148 						   false);
2149 				stable_tree_append(rmap_item, stable_node,
2150 						   false);
2151 			}
2152 			unlock_page(kpage);
2153 
2154 			/*
2155 			 * If we fail to insert the page into the stable tree,
2156 			 * we will have 2 virtual addresses that are pointing
2157 			 * to a ksm page left outside the stable tree,
2158 			 * in which case we need to break_cow on both.
2159 			 */
2160 			if (!stable_node) {
2161 				break_cow(tree_rmap_item);
2162 				break_cow(rmap_item);
2163 			}
2164 		} else if (split) {
2165 			/*
2166 			 * We are here if we tried to merge two pages and
2167 			 * failed because they both belonged to the same
2168 			 * compound page. We will split the page now, but no
2169 			 * merging will take place.
2170 			 * We do not want to add the cost of a full lock; if
2171 			 * the page is locked, it is better to skip it and
2172 			 * perhaps try again later.
2173 			 */
2174 			if (!trylock_page(page))
2175 				return;
2176 			split_huge_page(page);
2177 			unlock_page(page);
2178 		}
2179 	}
2180 }
2181 
2182 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2183 					    struct rmap_item **rmap_list,
2184 					    unsigned long addr)
2185 {
2186 	struct rmap_item *rmap_item;
2187 
2188 	while (*rmap_list) {
2189 		rmap_item = *rmap_list;
2190 		if ((rmap_item->address & PAGE_MASK) == addr)
2191 			return rmap_item;
2192 		if (rmap_item->address > addr)
2193 			break;
2194 		*rmap_list = rmap_item->rmap_list;
2195 		remove_rmap_item_from_tree(rmap_item);
2196 		free_rmap_item(rmap_item);
2197 	}
2198 
2199 	rmap_item = alloc_rmap_item();
2200 	if (rmap_item) {
2201 		/* It has already been zeroed */
2202 		rmap_item->mm = mm_slot->mm;
2203 		rmap_item->address = addr;
2204 		rmap_item->rmap_list = *rmap_list;
2205 		*rmap_list = rmap_item;
2206 	}
2207 	return rmap_item;
2208 }
2209 
2210 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2211 {
2212 	struct mm_struct *mm;
2213 	struct mm_slot *slot;
2214 	struct vm_area_struct *vma;
2215 	struct rmap_item *rmap_item;
2216 	int nid;
2217 
2218 	if (list_empty(&ksm_mm_head.mm_list))
2219 		return NULL;
2220 
2221 	slot = ksm_scan.mm_slot;
2222 	if (slot == &ksm_mm_head) {
2223 		/*
2224 		 * A number of pages can hang around indefinitely on per-cpu
2225 		 * pagevecs, raised page count preventing write_protect_page
2226 		 * from merging them.  Though it doesn't really matter much,
2227 		 * it is puzzling to see some stuck in pages_volatile until
2228 		 * other activity jostles them out, and they also prevented
2229 		 * LTP's KSM test from succeeding deterministically; so drain
2230 		 * them here (here rather than on entry to ksm_do_scan(),
2231 		 * so we don't IPI too often when pages_to_scan is set low).
2232 		 */
2233 		lru_add_drain_all();
2234 
2235 		/*
2236 		 * Whereas stale stable_nodes on the stable_tree itself
2237 		 * get pruned in the regular course of stable_tree_search(),
2238 		 * those moved out to the migrate_nodes list can accumulate:
2239 		 * so prune them once before each full scan.
2240 		 */
2241 		if (!ksm_merge_across_nodes) {
2242 			struct stable_node *stable_node, *next;
2243 			struct page *page;
2244 
2245 			list_for_each_entry_safe(stable_node, next,
2246 						 &migrate_nodes, list) {
2247 				page = get_ksm_page(stable_node, false);
2248 				if (page)
2249 					put_page(page);
2250 				cond_resched();
2251 			}
2252 		}
2253 
2254 		for (nid = 0; nid < ksm_nr_node_ids; nid++)
2255 			root_unstable_tree[nid] = RB_ROOT;
2256 
2257 		spin_lock(&ksm_mmlist_lock);
2258 		slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2259 		ksm_scan.mm_slot = slot;
2260 		spin_unlock(&ksm_mmlist_lock);
2261 		/*
2262 		 * Although we tested list_empty() above, a racing __ksm_exit
2263 		 * of the last mm on the list may have removed it since then.
2264 		 */
2265 		if (slot == &ksm_mm_head)
2266 			return NULL;
2267 next_mm:
2268 		ksm_scan.address = 0;
2269 		ksm_scan.rmap_list = &slot->rmap_list;
2270 	}
2271 
2272 	mm = slot->mm;
2273 	down_read(&mm->mmap_sem);
2274 	if (ksm_test_exit(mm))
2275 		vma = NULL;
2276 	else
2277 		vma = find_vma(mm, ksm_scan.address);
2278 
2279 	for (; vma; vma = vma->vm_next) {
2280 		if (!(vma->vm_flags & VM_MERGEABLE))
2281 			continue;
2282 		if (ksm_scan.address < vma->vm_start)
2283 			ksm_scan.address = vma->vm_start;
2284 		if (!vma->anon_vma)
2285 			ksm_scan.address = vma->vm_end;
2286 
2287 		while (ksm_scan.address < vma->vm_end) {
2288 			if (ksm_test_exit(mm))
2289 				break;
2290 			*page = follow_page(vma, ksm_scan.address, FOLL_GET);
2291 			if (IS_ERR_OR_NULL(*page)) {
2292 				ksm_scan.address += PAGE_SIZE;
2293 				cond_resched();
2294 				continue;
2295 			}
2296 			if (PageAnon(*page)) {
2297 				flush_anon_page(vma, *page, ksm_scan.address);
2298 				flush_dcache_page(*page);
2299 				rmap_item = get_next_rmap_item(slot,
2300 					ksm_scan.rmap_list, ksm_scan.address);
2301 				if (rmap_item) {
2302 					ksm_scan.rmap_list =
2303 							&rmap_item->rmap_list;
2304 					ksm_scan.address += PAGE_SIZE;
2305 				} else
2306 					put_page(*page);
2307 				up_read(&mm->mmap_sem);
2308 				return rmap_item;
2309 			}
2310 			put_page(*page);
2311 			ksm_scan.address += PAGE_SIZE;
2312 			cond_resched();
2313 		}
2314 	}
2315 
2316 	if (ksm_test_exit(mm)) {
2317 		ksm_scan.address = 0;
2318 		ksm_scan.rmap_list = &slot->rmap_list;
2319 	}
2320 	/*
2321 	 * Nuke all the rmap_items that are above this current rmap:
2322 	 * because there were no VM_MERGEABLE vmas with such addresses.
2323 	 */
2324 	remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2325 
2326 	spin_lock(&ksm_mmlist_lock);
2327 	ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2328 						struct mm_slot, mm_list);
2329 	if (ksm_scan.address == 0) {
2330 		/*
2331 		 * We've completed a full scan of all vmas, holding mmap_sem
2332 		 * throughout, and found no VM_MERGEABLE: so do the same as
2333 		 * __ksm_exit does to remove this mm from all our lists now.
2334 		 * This applies either when cleaning up after __ksm_exit
2335 		 * (but beware: we can reach here even before __ksm_exit),
2336 		 * or when all VM_MERGEABLE areas have been unmapped (and
2337 		 * mmap_sem then protects against race with MADV_MERGEABLE).
2338 		 */
2339 		hash_del(&slot->link);
2340 		list_del(&slot->mm_list);
2341 		spin_unlock(&ksm_mmlist_lock);
2342 
2343 		free_mm_slot(slot);
2344 		clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2345 		up_read(&mm->mmap_sem);
2346 		mmdrop(mm);
2347 	} else {
2348 		up_read(&mm->mmap_sem);
2349 		/*
2350 		 * up_read(&mm->mmap_sem) first because after
2351 		 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2352 		 * already have been freed under us by __ksm_exit()
2353 		 * because the "mm_slot" is still hashed and
2354 		 * ksm_scan.mm_slot doesn't point to it anymore.
2355 		 */
2356 		spin_unlock(&ksm_mmlist_lock);
2357 	}
2358 
2359 	/* Repeat until we've completed scanning the whole list */
2360 	slot = ksm_scan.mm_slot;
2361 	if (slot != &ksm_mm_head)
2362 		goto next_mm;
2363 
2364 	ksm_scan.seqnr++;
2365 	return NULL;
2366 }
2367 
2368 /**
2369  * ksm_do_scan  - the ksm scanner main worker function.
2370  * @scan_npages:  number of pages we want to scan before we return.
2371  */
2372 static void ksm_do_scan(unsigned int scan_npages)
2373 {
2374 	struct rmap_item *rmap_item;
2375 	struct page *uninitialized_var(page);
2376 
2377 	while (scan_npages-- && likely(!freezing(current))) {
2378 		cond_resched();
2379 		rmap_item = scan_get_next_rmap_item(&page);
2380 		if (!rmap_item)
2381 			return;
2382 		cmp_and_merge_page(page, rmap_item);
2383 		put_page(page);
2384 	}
2385 }
2386 
2387 static int ksmd_should_run(void)
2388 {
2389 	return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2390 }
2391 
2392 static int ksm_scan_thread(void *nothing)
2393 {
2394 	set_freezable();
2395 	set_user_nice(current, 5);
2396 
2397 	while (!kthread_should_stop()) {
2398 		mutex_lock(&ksm_thread_mutex);
2399 		wait_while_offlining();
2400 		if (ksmd_should_run())
2401 			ksm_do_scan(ksm_thread_pages_to_scan);
2402 		mutex_unlock(&ksm_thread_mutex);
2403 
2404 		try_to_freeze();
2405 
2406 		if (ksmd_should_run()) {
2407 			schedule_timeout_interruptible(
2408 				msecs_to_jiffies(ksm_thread_sleep_millisecs));
2409 		} else {
2410 			wait_event_freezable(ksm_thread_wait,
2411 				ksmd_should_run() || kthread_should_stop());
2412 		}
2413 	}
2414 	return 0;
2415 }
2416 
2417 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2418 		unsigned long end, int advice, unsigned long *vm_flags)
2419 {
2420 	struct mm_struct *mm = vma->vm_mm;
2421 	int err;
2422 
2423 	switch (advice) {
2424 	case MADV_MERGEABLE:
2425 		/*
2426 		 * Be somewhat over-protective for now!
2427 		 */
2428 		if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2429 				 VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2430 				 VM_HUGETLB | VM_MIXEDMAP))
2431 			return 0;		/* just ignore the advice */
2432 
2433 		if (vma_is_dax(vma))
2434 			return 0;
2435 
2436 #ifdef VM_SAO
2437 		if (*vm_flags & VM_SAO)
2438 			return 0;
2439 #endif
2440 #ifdef VM_SPARC_ADI
2441 		if (*vm_flags & VM_SPARC_ADI)
2442 			return 0;
2443 #endif
2444 
2445 		if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2446 			err = __ksm_enter(mm);
2447 			if (err)
2448 				return err;
2449 		}
2450 
2451 		*vm_flags |= VM_MERGEABLE;
2452 		break;
2453 
2454 	case MADV_UNMERGEABLE:
2455 		if (!(*vm_flags & VM_MERGEABLE))
2456 			return 0;		/* just ignore the advice */
2457 
2458 		if (vma->anon_vma) {
2459 			err = unmerge_ksm_pages(vma, start, end);
2460 			if (err)
2461 				return err;
2462 		}
2463 
2464 		*vm_flags &= ~VM_MERGEABLE;
2465 		break;
2466 	}
2467 
2468 	return 0;
2469 }
2470 
2471 int __ksm_enter(struct mm_struct *mm)
2472 {
2473 	struct mm_slot *mm_slot;
2474 	int needs_wakeup;
2475 
2476 	mm_slot = alloc_mm_slot();
2477 	if (!mm_slot)
2478 		return -ENOMEM;
2479 
2480 	/* Check ksm_run too?  Would need tighter locking */
2481 	needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2482 
2483 	spin_lock(&ksm_mmlist_lock);
2484 	insert_to_mm_slots_hash(mm, mm_slot);
2485 	/*
2486 	 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2487 	 * insert just behind the scanning cursor, to let the area settle
2488 	 * down a little; when fork is followed by immediate exec, we don't
2489 	 * want ksmd to waste time setting up and tearing down an rmap_list.
2490 	 *
2491 	 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2492 	 * scanning cursor, otherwise KSM pages in newly forked mms will be
2493 	 * missed: then we might as well insert at the end of the list.
2494 	 */
2495 	if (ksm_run & KSM_RUN_UNMERGE)
2496 		list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2497 	else
2498 		list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2499 	spin_unlock(&ksm_mmlist_lock);
2500 
2501 	set_bit(MMF_VM_MERGEABLE, &mm->flags);
2502 	mmgrab(mm);
2503 
2504 	if (needs_wakeup)
2505 		wake_up_interruptible(&ksm_thread_wait);
2506 
2507 	return 0;
2508 }
2509 
2510 void __ksm_exit(struct mm_struct *mm)
2511 {
2512 	struct mm_slot *mm_slot;
2513 	int easy_to_free = 0;
2514 
2515 	/*
2516 	 * This process is exiting: if it's straightforward (as is the
2517 	 * case when ksmd was never running), free mm_slot immediately.
2518 	 * But if it's at the cursor or has rmap_items linked to it, use
2519 	 * mmap_sem to synchronize with any break_cows before pagetables
2520 	 * are freed, and leave the mm_slot on the list for ksmd to free.
2521 	 * Beware: ksm may already have noticed it exiting and freed the slot.
2522 	 */
2523 
2524 	spin_lock(&ksm_mmlist_lock);
2525 	mm_slot = get_mm_slot(mm);
2526 	if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2527 		if (!mm_slot->rmap_list) {
2528 			hash_del(&mm_slot->link);
2529 			list_del(&mm_slot->mm_list);
2530 			easy_to_free = 1;
2531 		} else {
2532 			list_move(&mm_slot->mm_list,
2533 				  &ksm_scan.mm_slot->mm_list);
2534 		}
2535 	}
2536 	spin_unlock(&ksm_mmlist_lock);
2537 
2538 	if (easy_to_free) {
2539 		free_mm_slot(mm_slot);
2540 		clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2541 		mmdrop(mm);
2542 	} else if (mm_slot) {
2543 		down_write(&mm->mmap_sem);
2544 		up_write(&mm->mmap_sem);
2545 	}
2546 }
2547 
2548 struct page *ksm_might_need_to_copy(struct page *page,
2549 			struct vm_area_struct *vma, unsigned long address)
2550 {
2551 	struct anon_vma *anon_vma = page_anon_vma(page);
2552 	struct page *new_page;
2553 
2554 	if (PageKsm(page)) {
2555 		if (page_stable_node(page) &&
2556 		    !(ksm_run & KSM_RUN_UNMERGE))
2557 			return page;	/* no need to copy it */
2558 	} else if (!anon_vma) {
2559 		return page;		/* no need to copy it */
2560 	} else if (anon_vma->root == vma->anon_vma->root &&
2561 		 page->index == linear_page_index(vma, address)) {
2562 		return page;		/* still no need to copy it */
2563 	}
2564 	if (!PageUptodate(page))
2565 		return page;		/* let do_swap_page report the error */
2566 
2567 	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2568 	if (new_page) {
2569 		copy_user_highpage(new_page, page, address, vma);
2570 
2571 		SetPageDirty(new_page);
2572 		__SetPageUptodate(new_page);
2573 		__SetPageLocked(new_page);
2574 	}
2575 
2576 	return new_page;
2577 }
2578 
2579 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2580 {
2581 	struct stable_node *stable_node;
2582 	struct rmap_item *rmap_item;
2583 	int search_new_forks = 0;
2584 
2585 	VM_BUG_ON_PAGE(!PageKsm(page), page);
2586 
2587 	/*
2588 	 * Rely on the page lock to protect against concurrent modifications
2589 	 * to that page's node of the stable tree.
2590 	 */
2591 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2592 
2593 	stable_node = page_stable_node(page);
2594 	if (!stable_node)
2595 		return;
2596 again:
2597 	hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2598 		struct anon_vma *anon_vma = rmap_item->anon_vma;
2599 		struct anon_vma_chain *vmac;
2600 		struct vm_area_struct *vma;
2601 
2602 		cond_resched();
2603 		anon_vma_lock_read(anon_vma);
2604 		anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2605 					       0, ULONG_MAX) {
2606 			unsigned long addr;
2607 
2608 			cond_resched();
2609 			vma = vmac->vma;
2610 
2611 			/* Ignore the stable/unstable/sqnr flags */
2612 			addr = rmap_item->address & ~KSM_FLAG_MASK;
2613 
2614 			if (addr < vma->vm_start || addr >= vma->vm_end)
2615 				continue;
2616 			/*
2617 			 * Initially we examine only the vma which covers this
2618 			 * rmap_item; but later, if there is still work to do,
2619 			 * we examine covering vmas in other mms: in case they
2620 			 * were forked from the original since ksmd passed.
2621 			 */
2622 			if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2623 				continue;
2624 
2625 			if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2626 				continue;
2627 
2628 			if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2629 				anon_vma_unlock_read(anon_vma);
2630 				return;
2631 			}
2632 			if (rwc->done && rwc->done(page)) {
2633 				anon_vma_unlock_read(anon_vma);
2634 				return;
2635 			}
2636 		}
2637 		anon_vma_unlock_read(anon_vma);
2638 	}
2639 	if (!search_new_forks++)
2640 		goto again;
2641 }
2642 
2643 #ifdef CONFIG_MIGRATION
2644 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2645 {
2646 	struct stable_node *stable_node;
2647 
2648 	VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2649 	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2650 	VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2651 
2652 	stable_node = page_stable_node(newpage);
2653 	if (stable_node) {
2654 		VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2655 		stable_node->kpfn = page_to_pfn(newpage);
2656 		/*
2657 		 * newpage->mapping was set in advance; now we need smp_wmb()
2658 		 * to make sure that the new stable_node->kpfn is visible
2659 		 * to get_ksm_page() before it can see that oldpage->mapping
2660 		 * has gone stale (or that PageSwapCache has been cleared).
2661 		 */
2662 		smp_wmb();
2663 		set_page_stable_node(oldpage, NULL);
2664 	}
2665 }
2666 #endif /* CONFIG_MIGRATION */
2667 
2668 #ifdef CONFIG_MEMORY_HOTREMOVE
2669 static void wait_while_offlining(void)
2670 {
2671 	while (ksm_run & KSM_RUN_OFFLINE) {
2672 		mutex_unlock(&ksm_thread_mutex);
2673 		wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2674 			    TASK_UNINTERRUPTIBLE);
2675 		mutex_lock(&ksm_thread_mutex);
2676 	}
2677 }
2678 
2679 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2680 					 unsigned long start_pfn,
2681 					 unsigned long end_pfn)
2682 {
2683 	if (stable_node->kpfn >= start_pfn &&
2684 	    stable_node->kpfn < end_pfn) {
2685 		/*
2686 		 * Don't get_ksm_page, page has already gone:
2687 		 * which is why we keep kpfn instead of page*
2688 		 */
2689 		remove_node_from_stable_tree(stable_node);
2690 		return true;
2691 	}
2692 	return false;
2693 }
2694 
2695 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2696 					   unsigned long start_pfn,
2697 					   unsigned long end_pfn,
2698 					   struct rb_root *root)
2699 {
2700 	struct stable_node *dup;
2701 	struct hlist_node *hlist_safe;
2702 
2703 	if (!is_stable_node_chain(stable_node)) {
2704 		VM_BUG_ON(is_stable_node_dup(stable_node));
2705 		return stable_node_dup_remove_range(stable_node, start_pfn,
2706 						    end_pfn);
2707 	}
2708 
2709 	hlist_for_each_entry_safe(dup, hlist_safe,
2710 				  &stable_node->hlist, hlist_dup) {
2711 		VM_BUG_ON(!is_stable_node_dup(dup));
2712 		stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2713 	}
2714 	if (hlist_empty(&stable_node->hlist)) {
2715 		free_stable_node_chain(stable_node, root);
2716 		return true; /* notify caller that tree was rebalanced */
2717 	} else
2718 		return false;
2719 }
2720 
2721 static void ksm_check_stable_tree(unsigned long start_pfn,
2722 				  unsigned long end_pfn)
2723 {
2724 	struct stable_node *stable_node, *next;
2725 	struct rb_node *node;
2726 	int nid;
2727 
2728 	for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2729 		node = rb_first(root_stable_tree + nid);
2730 		while (node) {
2731 			stable_node = rb_entry(node, struct stable_node, node);
2732 			if (stable_node_chain_remove_range(stable_node,
2733 							   start_pfn, end_pfn,
2734 							   root_stable_tree +
2735 							   nid))
2736 				node = rb_first(root_stable_tree + nid);
2737 			else
2738 				node = rb_next(node);
2739 			cond_resched();
2740 		}
2741 	}
2742 	list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2743 		if (stable_node->kpfn >= start_pfn &&
2744 		    stable_node->kpfn < end_pfn)
2745 			remove_node_from_stable_tree(stable_node);
2746 		cond_resched();
2747 	}
2748 }
2749 
2750 static int ksm_memory_callback(struct notifier_block *self,
2751 			       unsigned long action, void *arg)
2752 {
2753 	struct memory_notify *mn = arg;
2754 
2755 	switch (action) {
2756 	case MEM_GOING_OFFLINE:
2757 		/*
2758 		 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2759 		 * and remove_all_stable_nodes() while memory is going offline:
2760 		 * it is unsafe for them to touch the stable tree at this time.
2761 		 * But unmerge_ksm_pages(), rmap lookups and other entry points
2762 		 * which do not need the ksm_thread_mutex are all safe.
2763 		 */
2764 		mutex_lock(&ksm_thread_mutex);
2765 		ksm_run |= KSM_RUN_OFFLINE;
2766 		mutex_unlock(&ksm_thread_mutex);
2767 		break;
2768 
2769 	case MEM_OFFLINE:
2770 		/*
2771 		 * Most of the work is done by page migration; but there might
2772 		 * be a few stable_nodes left over, still pointing to struct
2773 		 * pages which have been offlined: prune those from the tree,
2774 		 * otherwise get_ksm_page() might later try to access a
2775 		 * non-existent struct page.
2776 		 */
2777 		ksm_check_stable_tree(mn->start_pfn,
2778 				      mn->start_pfn + mn->nr_pages);
2779 		/* fallthrough */
2780 
2781 	case MEM_CANCEL_OFFLINE:
2782 		mutex_lock(&ksm_thread_mutex);
2783 		ksm_run &= ~KSM_RUN_OFFLINE;
2784 		mutex_unlock(&ksm_thread_mutex);
2785 
2786 		smp_mb();	/* wake_up_bit advises this */
2787 		wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2788 		break;
2789 	}
2790 	return NOTIFY_OK;
2791 }
2792 #else
2793 static void wait_while_offlining(void)
2794 {
2795 }
2796 #endif /* CONFIG_MEMORY_HOTREMOVE */
2797 
2798 #ifdef CONFIG_SYSFS
2799 /*
2800  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2801  */
2802 
2803 #define KSM_ATTR_RO(_name) \
2804 	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2805 #define KSM_ATTR(_name) \
2806 	static struct kobj_attribute _name##_attr = \
2807 		__ATTR(_name, 0644, _name##_show, _name##_store)
2808 
2809 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2810 				    struct kobj_attribute *attr, char *buf)
2811 {
2812 	return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2813 }
2814 
2815 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2816 				     struct kobj_attribute *attr,
2817 				     const char *buf, size_t count)
2818 {
2819 	unsigned long msecs;
2820 	int err;
2821 
2822 	err = kstrtoul(buf, 10, &msecs);
2823 	if (err || msecs > UINT_MAX)
2824 		return -EINVAL;
2825 
2826 	ksm_thread_sleep_millisecs = msecs;
2827 
2828 	return count;
2829 }
2830 KSM_ATTR(sleep_millisecs);
2831 
2832 static ssize_t pages_to_scan_show(struct kobject *kobj,
2833 				  struct kobj_attribute *attr, char *buf)
2834 {
2835 	return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2836 }
2837 
2838 static ssize_t pages_to_scan_store(struct kobject *kobj,
2839 				   struct kobj_attribute *attr,
2840 				   const char *buf, size_t count)
2841 {
2842 	int err;
2843 	unsigned long nr_pages;
2844 
2845 	err = kstrtoul(buf, 10, &nr_pages);
2846 	if (err || nr_pages > UINT_MAX)
2847 		return -EINVAL;
2848 
2849 	ksm_thread_pages_to_scan = nr_pages;
2850 
2851 	return count;
2852 }
2853 KSM_ATTR(pages_to_scan);
2854 
2855 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2856 			char *buf)
2857 {
2858 	return sprintf(buf, "%lu\n", ksm_run);
2859 }
2860 
2861 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2862 			 const char *buf, size_t count)
2863 {
2864 	int err;
2865 	unsigned long flags;
2866 
2867 	err = kstrtoul(buf, 10, &flags);
2868 	if (err || flags > UINT_MAX)
2869 		return -EINVAL;
2870 	if (flags > KSM_RUN_UNMERGE)
2871 		return -EINVAL;
2872 
2873 	/*
2874 	 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2875 	 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2876 	 * breaking COW to free the pages_shared (but leaves mm_slots
2877 	 * on the list for when ksmd may be set running again).
2878 	 */
2879 
2880 	mutex_lock(&ksm_thread_mutex);
2881 	wait_while_offlining();
2882 	if (ksm_run != flags) {
2883 		ksm_run = flags;
2884 		if (flags & KSM_RUN_UNMERGE) {
2885 			set_current_oom_origin();
2886 			err = unmerge_and_remove_all_rmap_items();
2887 			clear_current_oom_origin();
2888 			if (err) {
2889 				ksm_run = KSM_RUN_STOP;
2890 				count = err;
2891 			}
2892 		}
2893 	}
2894 	mutex_unlock(&ksm_thread_mutex);
2895 
2896 	if (flags & KSM_RUN_MERGE)
2897 		wake_up_interruptible(&ksm_thread_wait);
2898 
2899 	return count;
2900 }
2901 KSM_ATTR(run);
2902 
2903 #ifdef CONFIG_NUMA
2904 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2905 				struct kobj_attribute *attr, char *buf)
2906 {
2907 	return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2908 }
2909 
2910 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2911 				   struct kobj_attribute *attr,
2912 				   const char *buf, size_t count)
2913 {
2914 	int err;
2915 	unsigned long knob;
2916 
2917 	err = kstrtoul(buf, 10, &knob);
2918 	if (err)
2919 		return err;
2920 	if (knob > 1)
2921 		return -EINVAL;
2922 
2923 	mutex_lock(&ksm_thread_mutex);
2924 	wait_while_offlining();
2925 	if (ksm_merge_across_nodes != knob) {
2926 		if (ksm_pages_shared || remove_all_stable_nodes())
2927 			err = -EBUSY;
2928 		else if (root_stable_tree == one_stable_tree) {
2929 			struct rb_root *buf;
2930 			/*
2931 			 * This is the first time that we switch away from the
2932 			 * default of merging across nodes: must now allocate
2933 			 * a buffer to hold as many roots as may be needed.
2934 			 * Allocate stable and unstable together:
2935 			 * MAXSMP NODES_SHIFT 10 will use 16kB.
2936 			 */
2937 			buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2938 				      GFP_KERNEL);
2939 			/* Let us assume that RB_ROOT is NULL is zero */
2940 			if (!buf)
2941 				err = -ENOMEM;
2942 			else {
2943 				root_stable_tree = buf;
2944 				root_unstable_tree = buf + nr_node_ids;
2945 				/* Stable tree is empty but not the unstable */
2946 				root_unstable_tree[0] = one_unstable_tree[0];
2947 			}
2948 		}
2949 		if (!err) {
2950 			ksm_merge_across_nodes = knob;
2951 			ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2952 		}
2953 	}
2954 	mutex_unlock(&ksm_thread_mutex);
2955 
2956 	return err ? err : count;
2957 }
2958 KSM_ATTR(merge_across_nodes);
2959 #endif
2960 
2961 static ssize_t use_zero_pages_show(struct kobject *kobj,
2962 				struct kobj_attribute *attr, char *buf)
2963 {
2964 	return sprintf(buf, "%u\n", ksm_use_zero_pages);
2965 }
2966 static ssize_t use_zero_pages_store(struct kobject *kobj,
2967 				   struct kobj_attribute *attr,
2968 				   const char *buf, size_t count)
2969 {
2970 	int err;
2971 	bool value;
2972 
2973 	err = kstrtobool(buf, &value);
2974 	if (err)
2975 		return -EINVAL;
2976 
2977 	ksm_use_zero_pages = value;
2978 
2979 	return count;
2980 }
2981 KSM_ATTR(use_zero_pages);
2982 
2983 static ssize_t max_page_sharing_show(struct kobject *kobj,
2984 				     struct kobj_attribute *attr, char *buf)
2985 {
2986 	return sprintf(buf, "%u\n", ksm_max_page_sharing);
2987 }
2988 
2989 static ssize_t max_page_sharing_store(struct kobject *kobj,
2990 				      struct kobj_attribute *attr,
2991 				      const char *buf, size_t count)
2992 {
2993 	int err;
2994 	int knob;
2995 
2996 	err = kstrtoint(buf, 10, &knob);
2997 	if (err)
2998 		return err;
2999 	/*
3000 	 * When a KSM page is created it is shared by 2 mappings. This
3001 	 * being a signed comparison, it implicitly verifies it's not
3002 	 * negative.
3003 	 */
3004 	if (knob < 2)
3005 		return -EINVAL;
3006 
3007 	if (READ_ONCE(ksm_max_page_sharing) == knob)
3008 		return count;
3009 
3010 	mutex_lock(&ksm_thread_mutex);
3011 	wait_while_offlining();
3012 	if (ksm_max_page_sharing != knob) {
3013 		if (ksm_pages_shared || remove_all_stable_nodes())
3014 			err = -EBUSY;
3015 		else
3016 			ksm_max_page_sharing = knob;
3017 	}
3018 	mutex_unlock(&ksm_thread_mutex);
3019 
3020 	return err ? err : count;
3021 }
3022 KSM_ATTR(max_page_sharing);
3023 
3024 static ssize_t pages_shared_show(struct kobject *kobj,
3025 				 struct kobj_attribute *attr, char *buf)
3026 {
3027 	return sprintf(buf, "%lu\n", ksm_pages_shared);
3028 }
3029 KSM_ATTR_RO(pages_shared);
3030 
3031 static ssize_t pages_sharing_show(struct kobject *kobj,
3032 				  struct kobj_attribute *attr, char *buf)
3033 {
3034 	return sprintf(buf, "%lu\n", ksm_pages_sharing);
3035 }
3036 KSM_ATTR_RO(pages_sharing);
3037 
3038 static ssize_t pages_unshared_show(struct kobject *kobj,
3039 				   struct kobj_attribute *attr, char *buf)
3040 {
3041 	return sprintf(buf, "%lu\n", ksm_pages_unshared);
3042 }
3043 KSM_ATTR_RO(pages_unshared);
3044 
3045 static ssize_t pages_volatile_show(struct kobject *kobj,
3046 				   struct kobj_attribute *attr, char *buf)
3047 {
3048 	long ksm_pages_volatile;
3049 
3050 	ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3051 				- ksm_pages_sharing - ksm_pages_unshared;
3052 	/*
3053 	 * It was not worth any locking to calculate that statistic,
3054 	 * but it might therefore sometimes be negative: conceal that.
3055 	 */
3056 	if (ksm_pages_volatile < 0)
3057 		ksm_pages_volatile = 0;
3058 	return sprintf(buf, "%ld\n", ksm_pages_volatile);
3059 }
3060 KSM_ATTR_RO(pages_volatile);
3061 
3062 static ssize_t stable_node_dups_show(struct kobject *kobj,
3063 				     struct kobj_attribute *attr, char *buf)
3064 {
3065 	return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3066 }
3067 KSM_ATTR_RO(stable_node_dups);
3068 
3069 static ssize_t stable_node_chains_show(struct kobject *kobj,
3070 				       struct kobj_attribute *attr, char *buf)
3071 {
3072 	return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3073 }
3074 KSM_ATTR_RO(stable_node_chains);
3075 
3076 static ssize_t
3077 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3078 					struct kobj_attribute *attr,
3079 					char *buf)
3080 {
3081 	return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3082 }
3083 
3084 static ssize_t
3085 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3086 					 struct kobj_attribute *attr,
3087 					 const char *buf, size_t count)
3088 {
3089 	unsigned long msecs;
3090 	int err;
3091 
3092 	err = kstrtoul(buf, 10, &msecs);
3093 	if (err || msecs > UINT_MAX)
3094 		return -EINVAL;
3095 
3096 	ksm_stable_node_chains_prune_millisecs = msecs;
3097 
3098 	return count;
3099 }
3100 KSM_ATTR(stable_node_chains_prune_millisecs);
3101 
3102 static ssize_t full_scans_show(struct kobject *kobj,
3103 			       struct kobj_attribute *attr, char *buf)
3104 {
3105 	return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3106 }
3107 KSM_ATTR_RO(full_scans);
3108 
3109 static struct attribute *ksm_attrs[] = {
3110 	&sleep_millisecs_attr.attr,
3111 	&pages_to_scan_attr.attr,
3112 	&run_attr.attr,
3113 	&pages_shared_attr.attr,
3114 	&pages_sharing_attr.attr,
3115 	&pages_unshared_attr.attr,
3116 	&pages_volatile_attr.attr,
3117 	&full_scans_attr.attr,
3118 #ifdef CONFIG_NUMA
3119 	&merge_across_nodes_attr.attr,
3120 #endif
3121 	&max_page_sharing_attr.attr,
3122 	&stable_node_chains_attr.attr,
3123 	&stable_node_dups_attr.attr,
3124 	&stable_node_chains_prune_millisecs_attr.attr,
3125 	&use_zero_pages_attr.attr,
3126 	NULL,
3127 };
3128 
3129 static const struct attribute_group ksm_attr_group = {
3130 	.attrs = ksm_attrs,
3131 	.name = "ksm",
3132 };
3133 #endif /* CONFIG_SYSFS */
3134 
3135 static int __init ksm_init(void)
3136 {
3137 	struct task_struct *ksm_thread;
3138 	int err;
3139 
3140 	/* The correct value depends on page size and endianness */
3141 	zero_checksum = calc_checksum(ZERO_PAGE(0));
3142 	/* Default to false for backwards compatibility */
3143 	ksm_use_zero_pages = false;
3144 
3145 	err = ksm_slab_init();
3146 	if (err)
3147 		goto out;
3148 
3149 	ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3150 	if (IS_ERR(ksm_thread)) {
3151 		pr_err("ksm: creating kthread failed\n");
3152 		err = PTR_ERR(ksm_thread);
3153 		goto out_free;
3154 	}
3155 
3156 #ifdef CONFIG_SYSFS
3157 	err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3158 	if (err) {
3159 		pr_err("ksm: register sysfs failed\n");
3160 		kthread_stop(ksm_thread);
3161 		goto out_free;
3162 	}
3163 #else
3164 	ksm_run = KSM_RUN_MERGE;	/* no way for user to start it */
3165 
3166 #endif /* CONFIG_SYSFS */
3167 
3168 #ifdef CONFIG_MEMORY_HOTREMOVE
3169 	/* There is no significance to this priority 100 */
3170 	hotplug_memory_notifier(ksm_memory_callback, 100);
3171 #endif
3172 	return 0;
3173 
3174 out_free:
3175 	ksm_slab_free();
3176 out:
3177 	return err;
3178 }
3179 subsys_initcall(ksm_init);
3180