mm/ksm.c: don't WARN if page is still mapped in remove_stable_node()
[muen/linux.git] / mm / ksm.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Memory merging support.
4  *
5  * This code enables dynamic sharing of identical pages found in different
6  * memory areas, even if they are not shared by fork()
7  *
8  * Copyright (C) 2008-2009 Red Hat, Inc.
9  * Authors:
10  *      Izik Eidus
11  *      Andrea Arcangeli
12  *      Chris Wright
13  *      Hugh Dickins
14  */
15
16 #include <linux/errno.h>
17 #include <linux/mm.h>
18 #include <linux/fs.h>
19 #include <linux/mman.h>
20 #include <linux/sched.h>
21 #include <linux/sched/mm.h>
22 #include <linux/sched/coredump.h>
23 #include <linux/rwsem.h>
24 #include <linux/pagemap.h>
25 #include <linux/rmap.h>
26 #include <linux/spinlock.h>
27 #include <linux/xxhash.h>
28 #include <linux/delay.h>
29 #include <linux/kthread.h>
30 #include <linux/wait.h>
31 #include <linux/slab.h>
32 #include <linux/rbtree.h>
33 #include <linux/memory.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/swap.h>
36 #include <linux/ksm.h>
37 #include <linux/hashtable.h>
38 #include <linux/freezer.h>
39 #include <linux/oom.h>
40 #include <linux/numa.h>
41
42 #include <asm/tlbflush.h>
43 #include "internal.h"
44
45 #ifdef CONFIG_NUMA
46 #define NUMA(x)         (x)
47 #define DO_NUMA(x)      do { (x); } while (0)
48 #else
49 #define NUMA(x)         (0)
50 #define DO_NUMA(x)      do { } while (0)
51 #endif
52
53 /**
54  * DOC: Overview
55  *
56  * A few notes about the KSM scanning process,
57  * to make it easier to understand the data structures below:
58  *
59  * In order to reduce excessive scanning, KSM sorts the memory pages by their
60  * contents into a data structure that holds pointers to the pages' locations.
61  *
62  * Since the contents of the pages may change at any moment, KSM cannot just
63  * insert the pages into a normal sorted tree and expect it to find anything.
64  * Therefore KSM uses two data structures - the stable and the unstable tree.
65  *
66  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
67  * by their contents.  Because each such page is write-protected, searching on
68  * this tree is fully assured to be working (except when pages are unmapped),
69  * and therefore this tree is called the stable tree.
70  *
71  * The stable tree node includes information required for reverse
72  * mapping from a KSM page to virtual addresses that map this page.
73  *
74  * In order to avoid large latencies of the rmap walks on KSM pages,
75  * KSM maintains two types of nodes in the stable tree:
76  *
77  * * the regular nodes that keep the reverse mapping structures in a
78  *   linked list
79  * * the "chains" that link nodes ("dups") that represent the same
80  *   write protected memory content, but each "dup" corresponds to a
81  *   different KSM page copy of that content
82  *
83  * Internally, the regular nodes, "dups" and "chains" are represented
84  * using the same :c:type:`struct stable_node` structure.
85  *
86  * In addition to the stable tree, KSM uses a second data structure called the
87  * unstable tree: this tree holds pointers to pages which have been found to
88  * be "unchanged for a period of time".  The unstable tree sorts these pages
89  * by their contents, but since they are not write-protected, KSM cannot rely
90  * upon the unstable tree to work correctly - the unstable tree is liable to
91  * be corrupted as its contents are modified, and so it is called unstable.
92  *
93  * KSM solves this problem by several techniques:
94  *
95  * 1) The unstable tree is flushed every time KSM completes scanning all
96  *    memory areas, and then the tree is rebuilt again from the beginning.
97  * 2) KSM will only insert into the unstable tree, pages whose hash value
98  *    has not changed since the previous scan of all memory areas.
99  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
100  *    colors of the nodes and not on their contents, assuring that even when
101  *    the tree gets "corrupted" it won't get out of balance, so scanning time
102  *    remains the same (also, searching and inserting nodes in an rbtree uses
103  *    the same algorithm, so we have no overhead when we flush and rebuild).
104  * 4) KSM never flushes the stable tree, which means that even if it were to
105  *    take 10 attempts to find a page in the unstable tree, once it is found,
106  *    it is secured in the stable tree.  (When we scan a new page, we first
107  *    compare it against the stable tree, and then against the unstable tree.)
108  *
109  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
110  * stable trees and multiple unstable trees: one of each for each NUMA node.
111  */
112
113 /**
114  * struct mm_slot - ksm information per mm that is being scanned
115  * @link: link to the mm_slots hash list
116  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
117  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
118  * @mm: the mm that this information is valid for
119  */
120 struct mm_slot {
121         struct hlist_node link;
122         struct list_head mm_list;
123         struct rmap_item *rmap_list;
124         struct mm_struct *mm;
125 };
126
127 /**
128  * struct ksm_scan - cursor for scanning
129  * @mm_slot: the current mm_slot we are scanning
130  * @address: the next address inside that to be scanned
131  * @rmap_list: link to the next rmap to be scanned in the rmap_list
132  * @seqnr: count of completed full scans (needed when removing unstable node)
133  *
134  * There is only the one ksm_scan instance of this cursor structure.
135  */
136 struct ksm_scan {
137         struct mm_slot *mm_slot;
138         unsigned long address;
139         struct rmap_item **rmap_list;
140         unsigned long seqnr;
141 };
142
143 /**
144  * struct stable_node - node of the stable rbtree
145  * @node: rb node of this ksm page in the stable tree
146  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
147  * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
148  * @list: linked into migrate_nodes, pending placement in the proper node tree
149  * @hlist: hlist head of rmap_items using this ksm page
150  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
151  * @chain_prune_time: time of the last full garbage collection
152  * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
153  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
154  */
155 struct stable_node {
156         union {
157                 struct rb_node node;    /* when node of stable tree */
158                 struct {                /* when listed for migration */
159                         struct list_head *head;
160                         struct {
161                                 struct hlist_node hlist_dup;
162                                 struct list_head list;
163                         };
164                 };
165         };
166         struct hlist_head hlist;
167         union {
168                 unsigned long kpfn;
169                 unsigned long chain_prune_time;
170         };
171         /*
172          * STABLE_NODE_CHAIN can be any negative number in
173          * rmap_hlist_len negative range, but better not -1 to be able
174          * to reliably detect underflows.
175          */
176 #define STABLE_NODE_CHAIN -1024
177         int rmap_hlist_len;
178 #ifdef CONFIG_NUMA
179         int nid;
180 #endif
181 };
182
183 /**
184  * struct rmap_item - reverse mapping item for virtual addresses
185  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
186  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
187  * @nid: NUMA node id of unstable tree in which linked (may not match page)
188  * @mm: the memory structure this rmap_item is pointing into
189  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
190  * @oldchecksum: previous checksum of the page at that virtual address
191  * @node: rb node of this rmap_item in the unstable tree
192  * @head: pointer to stable_node heading this list in the stable tree
193  * @hlist: link into hlist of rmap_items hanging off that stable_node
194  */
195 struct rmap_item {
196         struct rmap_item *rmap_list;
197         union {
198                 struct anon_vma *anon_vma;      /* when stable */
199 #ifdef CONFIG_NUMA
200                 int nid;                /* when node of unstable tree */
201 #endif
202         };
203         struct mm_struct *mm;
204         unsigned long address;          /* + low bits used for flags below */
205         unsigned int oldchecksum;       /* when unstable */
206         union {
207                 struct rb_node node;    /* when node of unstable tree */
208                 struct {                /* when listed from stable tree */
209                         struct stable_node *head;
210                         struct hlist_node hlist;
211                 };
212         };
213 };
214
215 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
216 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
217 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
218 #define KSM_FLAG_MASK   (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
219                                 /* to mask all the flags */
220
221 /* The stable and unstable tree heads */
222 static struct rb_root one_stable_tree[1] = { RB_ROOT };
223 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
224 static struct rb_root *root_stable_tree = one_stable_tree;
225 static struct rb_root *root_unstable_tree = one_unstable_tree;
226
227 /* Recently migrated nodes of stable tree, pending proper placement */
228 static LIST_HEAD(migrate_nodes);
229 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
230
231 #define MM_SLOTS_HASH_BITS 10
232 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
233
234 static struct mm_slot ksm_mm_head = {
235         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
236 };
237 static struct ksm_scan ksm_scan = {
238         .mm_slot = &ksm_mm_head,
239 };
240
241 static struct kmem_cache *rmap_item_cache;
242 static struct kmem_cache *stable_node_cache;
243 static struct kmem_cache *mm_slot_cache;
244
245 /* The number of nodes in the stable tree */
246 static unsigned long ksm_pages_shared;
247
248 /* The number of page slots additionally sharing those nodes */
249 static unsigned long ksm_pages_sharing;
250
251 /* The number of nodes in the unstable tree */
252 static unsigned long ksm_pages_unshared;
253
254 /* The number of rmap_items in use: to calculate pages_volatile */
255 static unsigned long ksm_rmap_items;
256
257 /* The number of stable_node chains */
258 static unsigned long ksm_stable_node_chains;
259
260 /* The number of stable_node dups linked to the stable_node chains */
261 static unsigned long ksm_stable_node_dups;
262
263 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
264 static int ksm_stable_node_chains_prune_millisecs = 2000;
265
266 /* Maximum number of page slots sharing a stable node */
267 static int ksm_max_page_sharing = 256;
268
269 /* Number of pages ksmd should scan in one batch */
270 static unsigned int ksm_thread_pages_to_scan = 100;
271
272 /* Milliseconds ksmd should sleep between batches */
273 static unsigned int ksm_thread_sleep_millisecs = 20;
274
275 /* Checksum of an empty (zeroed) page */
276 static unsigned int zero_checksum __read_mostly;
277
278 /* Whether to merge empty (zeroed) pages with actual zero pages */
279 static bool ksm_use_zero_pages __read_mostly;
280
281 #ifdef CONFIG_NUMA
282 /* Zeroed when merging across nodes is not allowed */
283 static unsigned int ksm_merge_across_nodes = 1;
284 static int ksm_nr_node_ids = 1;
285 #else
286 #define ksm_merge_across_nodes  1U
287 #define ksm_nr_node_ids         1
288 #endif
289
290 #define KSM_RUN_STOP    0
291 #define KSM_RUN_MERGE   1
292 #define KSM_RUN_UNMERGE 2
293 #define KSM_RUN_OFFLINE 4
294 static unsigned long ksm_run = KSM_RUN_STOP;
295 static void wait_while_offlining(void);
296
297 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
298 static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_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 = NUMA_NO_NODE; /* 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 enum get_ksm_page_flags {
670         GET_KSM_PAGE_NOLOCK,
671         GET_KSM_PAGE_LOCK,
672         GET_KSM_PAGE_TRYLOCK
673 };
674
675 /*
676  * get_ksm_page: checks if the page indicated by the stable node
677  * is still its ksm page, despite having held no reference to it.
678  * In which case we can trust the content of the page, and it
679  * returns the gotten page; but if the page has now been zapped,
680  * remove the stale node from the stable tree and return NULL.
681  * But beware, the stable node's page might be being migrated.
682  *
683  * You would expect the stable_node to hold a reference to the ksm page.
684  * But if it increments the page's count, swapping out has to wait for
685  * ksmd to come around again before it can free the page, which may take
686  * seconds or even minutes: much too unresponsive.  So instead we use a
687  * "keyhole reference": access to the ksm page from the stable node peeps
688  * out through its keyhole to see if that page still holds the right key,
689  * pointing back to this stable node.  This relies on freeing a PageAnon
690  * page to reset its page->mapping to NULL, and relies on no other use of
691  * a page to put something that might look like our key in page->mapping.
692  * is on its way to being freed; but it is an anomaly to bear in mind.
693  */
694 static struct page *get_ksm_page(struct stable_node *stable_node,
695                                  enum get_ksm_page_flags flags)
696 {
697         struct page *page;
698         void *expected_mapping;
699         unsigned long kpfn;
700
701         expected_mapping = (void *)((unsigned long)stable_node |
702                                         PAGE_MAPPING_KSM);
703 again:
704         kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
705         page = pfn_to_page(kpfn);
706         if (READ_ONCE(page->mapping) != expected_mapping)
707                 goto stale;
708
709         /*
710          * We cannot do anything with the page while its refcount is 0.
711          * Usually 0 means free, or tail of a higher-order page: in which
712          * case this node is no longer referenced, and should be freed;
713          * however, it might mean that the page is under page_ref_freeze().
714          * The __remove_mapping() case is easy, again the node is now stale;
715          * the same is in reuse_ksm_page() case; but if page is swapcache
716          * in migrate_page_move_mapping(), it might still be our page,
717          * in which case it's essential to keep the node.
718          */
719         while (!get_page_unless_zero(page)) {
720                 /*
721                  * Another check for page->mapping != expected_mapping would
722                  * work here too.  We have chosen the !PageSwapCache test to
723                  * optimize the common case, when the page is or is about to
724                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
725                  * in the ref_freeze section of __remove_mapping(); but Anon
726                  * page->mapping reset to NULL later, in free_pages_prepare().
727                  */
728                 if (!PageSwapCache(page))
729                         goto stale;
730                 cpu_relax();
731         }
732
733         if (READ_ONCE(page->mapping) != expected_mapping) {
734                 put_page(page);
735                 goto stale;
736         }
737
738         if (flags == GET_KSM_PAGE_TRYLOCK) {
739                 if (!trylock_page(page)) {
740                         put_page(page);
741                         return ERR_PTR(-EBUSY);
742                 }
743         } else if (flags == GET_KSM_PAGE_LOCK)
744                 lock_page(page);
745
746         if (flags != GET_KSM_PAGE_NOLOCK) {
747                 if (READ_ONCE(page->mapping) != expected_mapping) {
748                         unlock_page(page);
749                         put_page(page);
750                         goto stale;
751                 }
752         }
753         return page;
754
755 stale:
756         /*
757          * We come here from above when page->mapping or !PageSwapCache
758          * suggests that the node is stale; but it might be under migration.
759          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
760          * before checking whether node->kpfn has been changed.
761          */
762         smp_rmb();
763         if (READ_ONCE(stable_node->kpfn) != kpfn)
764                 goto again;
765         remove_node_from_stable_tree(stable_node);
766         return NULL;
767 }
768
769 /*
770  * Removing rmap_item from stable or unstable tree.
771  * This function will clean the information from the stable/unstable tree.
772  */
773 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
774 {
775         if (rmap_item->address & STABLE_FLAG) {
776                 struct stable_node *stable_node;
777                 struct page *page;
778
779                 stable_node = rmap_item->head;
780                 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
781                 if (!page)
782                         goto out;
783
784                 hlist_del(&rmap_item->hlist);
785                 unlock_page(page);
786                 put_page(page);
787
788                 if (!hlist_empty(&stable_node->hlist))
789                         ksm_pages_sharing--;
790                 else
791                         ksm_pages_shared--;
792                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
793                 stable_node->rmap_hlist_len--;
794
795                 put_anon_vma(rmap_item->anon_vma);
796                 rmap_item->address &= PAGE_MASK;
797
798         } else if (rmap_item->address & UNSTABLE_FLAG) {
799                 unsigned char age;
800                 /*
801                  * Usually ksmd can and must skip the rb_erase, because
802                  * root_unstable_tree was already reset to RB_ROOT.
803                  * But be careful when an mm is exiting: do the rb_erase
804                  * if this rmap_item was inserted by this scan, rather
805                  * than left over from before.
806                  */
807                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
808                 BUG_ON(age > 1);
809                 if (!age)
810                         rb_erase(&rmap_item->node,
811                                  root_unstable_tree + NUMA(rmap_item->nid));
812                 ksm_pages_unshared--;
813                 rmap_item->address &= PAGE_MASK;
814         }
815 out:
816         cond_resched();         /* we're called from many long loops */
817 }
818
819 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
820                                        struct rmap_item **rmap_list)
821 {
822         while (*rmap_list) {
823                 struct rmap_item *rmap_item = *rmap_list;
824                 *rmap_list = rmap_item->rmap_list;
825                 remove_rmap_item_from_tree(rmap_item);
826                 free_rmap_item(rmap_item);
827         }
828 }
829
830 /*
831  * Though it's very tempting to unmerge rmap_items from stable tree rather
832  * than check every pte of a given vma, the locking doesn't quite work for
833  * that - an rmap_item is assigned to the stable tree after inserting ksm
834  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
835  * rmap_items from parent to child at fork time (so as not to waste time
836  * if exit comes before the next scan reaches it).
837  *
838  * Similarly, although we'd like to remove rmap_items (so updating counts
839  * and freeing memory) when unmerging an area, it's easier to leave that
840  * to the next pass of ksmd - consider, for example, how ksmd might be
841  * in cmp_and_merge_page on one of the rmap_items we would be removing.
842  */
843 static int unmerge_ksm_pages(struct vm_area_struct *vma,
844                              unsigned long start, unsigned long end)
845 {
846         unsigned long addr;
847         int err = 0;
848
849         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
850                 if (ksm_test_exit(vma->vm_mm))
851                         break;
852                 if (signal_pending(current))
853                         err = -ERESTARTSYS;
854                 else
855                         err = break_ksm(vma, addr);
856         }
857         return err;
858 }
859
860 static inline struct stable_node *page_stable_node(struct page *page)
861 {
862         return PageKsm(page) ? page_rmapping(page) : NULL;
863 }
864
865 static inline void set_page_stable_node(struct page *page,
866                                         struct stable_node *stable_node)
867 {
868         page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
869 }
870
871 #ifdef CONFIG_SYSFS
872 /*
873  * Only called through the sysfs control interface:
874  */
875 static int remove_stable_node(struct stable_node *stable_node)
876 {
877         struct page *page;
878         int err;
879
880         page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
881         if (!page) {
882                 /*
883                  * get_ksm_page did remove_node_from_stable_tree itself.
884                  */
885                 return 0;
886         }
887
888         /*
889          * Page could be still mapped if this races with __mmput() running in
890          * between ksm_exit() and exit_mmap(). Just refuse to let
891          * merge_across_nodes/max_page_sharing be switched.
892          */
893         err = -EBUSY;
894         if (!page_mapped(page)) {
895                 /*
896                  * The stable node did not yet appear stale to get_ksm_page(),
897                  * since that allows for an unmapped ksm page to be recognized
898                  * right up until it is freed; but the node is safe to remove.
899                  * This page might be in a pagevec waiting to be freed,
900                  * or it might be PageSwapCache (perhaps under writeback),
901                  * or it might have been removed from swapcache a moment ago.
902                  */
903                 set_page_stable_node(page, NULL);
904                 remove_node_from_stable_tree(stable_node);
905                 err = 0;
906         }
907
908         unlock_page(page);
909         put_page(page);
910         return err;
911 }
912
913 static int remove_stable_node_chain(struct stable_node *stable_node,
914                                     struct rb_root *root)
915 {
916         struct stable_node *dup;
917         struct hlist_node *hlist_safe;
918
919         if (!is_stable_node_chain(stable_node)) {
920                 VM_BUG_ON(is_stable_node_dup(stable_node));
921                 if (remove_stable_node(stable_node))
922                         return true;
923                 else
924                         return false;
925         }
926
927         hlist_for_each_entry_safe(dup, hlist_safe,
928                                   &stable_node->hlist, hlist_dup) {
929                 VM_BUG_ON(!is_stable_node_dup(dup));
930                 if (remove_stable_node(dup))
931                         return true;
932         }
933         BUG_ON(!hlist_empty(&stable_node->hlist));
934         free_stable_node_chain(stable_node, root);
935         return false;
936 }
937
938 static int remove_all_stable_nodes(void)
939 {
940         struct stable_node *stable_node, *next;
941         int nid;
942         int err = 0;
943
944         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
945                 while (root_stable_tree[nid].rb_node) {
946                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
947                                                 struct stable_node, node);
948                         if (remove_stable_node_chain(stable_node,
949                                                      root_stable_tree + nid)) {
950                                 err = -EBUSY;
951                                 break;  /* proceed to next nid */
952                         }
953                         cond_resched();
954                 }
955         }
956         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
957                 if (remove_stable_node(stable_node))
958                         err = -EBUSY;
959                 cond_resched();
960         }
961         return err;
962 }
963
964 static int unmerge_and_remove_all_rmap_items(void)
965 {
966         struct mm_slot *mm_slot;
967         struct mm_struct *mm;
968         struct vm_area_struct *vma;
969         int err = 0;
970
971         spin_lock(&ksm_mmlist_lock);
972         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
973                                                 struct mm_slot, mm_list);
974         spin_unlock(&ksm_mmlist_lock);
975
976         for (mm_slot = ksm_scan.mm_slot;
977                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
978                 mm = mm_slot->mm;
979                 down_read(&mm->mmap_sem);
980                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
981                         if (ksm_test_exit(mm))
982                                 break;
983                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
984                                 continue;
985                         err = unmerge_ksm_pages(vma,
986                                                 vma->vm_start, vma->vm_end);
987                         if (err)
988                                 goto error;
989                 }
990
991                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
992                 up_read(&mm->mmap_sem);
993
994                 spin_lock(&ksm_mmlist_lock);
995                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
996                                                 struct mm_slot, mm_list);
997                 if (ksm_test_exit(mm)) {
998                         hash_del(&mm_slot->link);
999                         list_del(&mm_slot->mm_list);
1000                         spin_unlock(&ksm_mmlist_lock);
1001
1002                         free_mm_slot(mm_slot);
1003                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1004                         mmdrop(mm);
1005                 } else
1006                         spin_unlock(&ksm_mmlist_lock);
1007         }
1008
1009         /* Clean up stable nodes, but don't worry if some are still busy */
1010         remove_all_stable_nodes();
1011         ksm_scan.seqnr = 0;
1012         return 0;
1013
1014 error:
1015         up_read(&mm->mmap_sem);
1016         spin_lock(&ksm_mmlist_lock);
1017         ksm_scan.mm_slot = &ksm_mm_head;
1018         spin_unlock(&ksm_mmlist_lock);
1019         return err;
1020 }
1021 #endif /* CONFIG_SYSFS */
1022
1023 static u32 calc_checksum(struct page *page)
1024 {
1025         u32 checksum;
1026         void *addr = kmap_atomic(page);
1027         checksum = xxhash(addr, PAGE_SIZE, 0);
1028         kunmap_atomic(addr);
1029         return checksum;
1030 }
1031
1032 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1033                               pte_t *orig_pte)
1034 {
1035         struct mm_struct *mm = vma->vm_mm;
1036         struct page_vma_mapped_walk pvmw = {
1037                 .page = page,
1038                 .vma = vma,
1039         };
1040         int swapped;
1041         int err = -EFAULT;
1042         struct mmu_notifier_range range;
1043
1044         pvmw.address = page_address_in_vma(page, vma);
1045         if (pvmw.address == -EFAULT)
1046                 goto out;
1047
1048         BUG_ON(PageTransCompound(page));
1049
1050         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1051                                 pvmw.address,
1052                                 pvmw.address + PAGE_SIZE);
1053         mmu_notifier_invalidate_range_start(&range);
1054
1055         if (!page_vma_mapped_walk(&pvmw))
1056                 goto out_mn;
1057         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1058                 goto out_unlock;
1059
1060         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1061             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1062                                                 mm_tlb_flush_pending(mm)) {
1063                 pte_t entry;
1064
1065                 swapped = PageSwapCache(page);
1066                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1067                 /*
1068                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1069                  * take any lock, therefore the check that we are going to make
1070                  * with the pagecount against the mapcount is racey and
1071                  * O_DIRECT can happen right after the check.
1072                  * So we clear the pte and flush the tlb before the check
1073                  * this assure us that no O_DIRECT can happen after the check
1074                  * or in the middle of the check.
1075                  *
1076                  * No need to notify as we are downgrading page table to read
1077                  * only not changing it to point to a new page.
1078                  *
1079                  * See Documentation/vm/mmu_notifier.rst
1080                  */
1081                 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1082                 /*
1083                  * Check that no O_DIRECT or similar I/O is in progress on the
1084                  * page
1085                  */
1086                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1087                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1088                         goto out_unlock;
1089                 }
1090                 if (pte_dirty(entry))
1091                         set_page_dirty(page);
1092
1093                 if (pte_protnone(entry))
1094                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1095                 else
1096                         entry = pte_mkclean(pte_wrprotect(entry));
1097                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1098         }
1099         *orig_pte = *pvmw.pte;
1100         err = 0;
1101
1102 out_unlock:
1103         page_vma_mapped_walk_done(&pvmw);
1104 out_mn:
1105         mmu_notifier_invalidate_range_end(&range);
1106 out:
1107         return err;
1108 }
1109
1110 /**
1111  * replace_page - replace page in vma by new ksm page
1112  * @vma:      vma that holds the pte pointing to page
1113  * @page:     the page we are replacing by kpage
1114  * @kpage:    the ksm page we replace page by
1115  * @orig_pte: the original value of the pte
1116  *
1117  * Returns 0 on success, -EFAULT on failure.
1118  */
1119 static int replace_page(struct vm_area_struct *vma, struct page *page,
1120                         struct page *kpage, pte_t orig_pte)
1121 {
1122         struct mm_struct *mm = vma->vm_mm;
1123         pmd_t *pmd;
1124         pte_t *ptep;
1125         pte_t newpte;
1126         spinlock_t *ptl;
1127         unsigned long addr;
1128         int err = -EFAULT;
1129         struct mmu_notifier_range range;
1130
1131         addr = page_address_in_vma(page, vma);
1132         if (addr == -EFAULT)
1133                 goto out;
1134
1135         pmd = mm_find_pmd(mm, addr);
1136         if (!pmd)
1137                 goto out;
1138
1139         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1140                                 addr + PAGE_SIZE);
1141         mmu_notifier_invalidate_range_start(&range);
1142
1143         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1144         if (!pte_same(*ptep, orig_pte)) {
1145                 pte_unmap_unlock(ptep, ptl);
1146                 goto out_mn;
1147         }
1148
1149         /*
1150          * No need to check ksm_use_zero_pages here: we can only have a
1151          * zero_page here if ksm_use_zero_pages was enabled alreaady.
1152          */
1153         if (!is_zero_pfn(page_to_pfn(kpage))) {
1154                 get_page(kpage);
1155                 page_add_anon_rmap(kpage, vma, addr, false);
1156                 newpte = mk_pte(kpage, vma->vm_page_prot);
1157         } else {
1158                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1159                                                vma->vm_page_prot));
1160                 /*
1161                  * We're replacing an anonymous page with a zero page, which is
1162                  * not anonymous. We need to do proper accounting otherwise we
1163                  * will get wrong values in /proc, and a BUG message in dmesg
1164                  * when tearing down the mm.
1165                  */
1166                 dec_mm_counter(mm, MM_ANONPAGES);
1167         }
1168
1169         flush_cache_page(vma, addr, pte_pfn(*ptep));
1170         /*
1171          * No need to notify as we are replacing a read only page with another
1172          * read only page with the same content.
1173          *
1174          * See Documentation/vm/mmu_notifier.rst
1175          */
1176         ptep_clear_flush(vma, addr, ptep);
1177         set_pte_at_notify(mm, addr, ptep, newpte);
1178
1179         page_remove_rmap(page, false);
1180         if (!page_mapped(page))
1181                 try_to_free_swap(page);
1182         put_page(page);
1183
1184         pte_unmap_unlock(ptep, ptl);
1185         err = 0;
1186 out_mn:
1187         mmu_notifier_invalidate_range_end(&range);
1188 out:
1189         return err;
1190 }
1191
1192 /*
1193  * try_to_merge_one_page - take two pages and merge them into one
1194  * @vma: the vma that holds the pte pointing to page
1195  * @page: the PageAnon page that we want to replace with kpage
1196  * @kpage: the PageKsm page that we want to map instead of page,
1197  *         or NULL the first time when we want to use page as kpage.
1198  *
1199  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1200  */
1201 static int try_to_merge_one_page(struct vm_area_struct *vma,
1202                                  struct page *page, struct page *kpage)
1203 {
1204         pte_t orig_pte = __pte(0);
1205         int err = -EFAULT;
1206
1207         if (page == kpage)                      /* ksm page forked */
1208                 return 0;
1209
1210         if (!PageAnon(page))
1211                 goto out;
1212
1213         /*
1214          * We need the page lock to read a stable PageSwapCache in
1215          * write_protect_page().  We use trylock_page() instead of
1216          * lock_page() because we don't want to wait here - we
1217          * prefer to continue scanning and merging different pages,
1218          * then come back to this page when it is unlocked.
1219          */
1220         if (!trylock_page(page))
1221                 goto out;
1222
1223         if (PageTransCompound(page)) {
1224                 if (split_huge_page(page))
1225                         goto out_unlock;
1226         }
1227
1228         /*
1229          * If this anonymous page is mapped only here, its pte may need
1230          * to be write-protected.  If it's mapped elsewhere, all of its
1231          * ptes are necessarily already write-protected.  But in either
1232          * case, we need to lock and check page_count is not raised.
1233          */
1234         if (write_protect_page(vma, page, &orig_pte) == 0) {
1235                 if (!kpage) {
1236                         /*
1237                          * While we hold page lock, upgrade page from
1238                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1239                          * stable_tree_insert() will update stable_node.
1240                          */
1241                         set_page_stable_node(page, NULL);
1242                         mark_page_accessed(page);
1243                         /*
1244                          * Page reclaim just frees a clean page with no dirty
1245                          * ptes: make sure that the ksm page would be swapped.
1246                          */
1247                         if (!PageDirty(page))
1248                                 SetPageDirty(page);
1249                         err = 0;
1250                 } else if (pages_identical(page, kpage))
1251                         err = replace_page(vma, page, kpage, orig_pte);
1252         }
1253
1254         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1255                 munlock_vma_page(page);
1256                 if (!PageMlocked(kpage)) {
1257                         unlock_page(page);
1258                         lock_page(kpage);
1259                         mlock_vma_page(kpage);
1260                         page = kpage;           /* for final unlock */
1261                 }
1262         }
1263
1264 out_unlock:
1265         unlock_page(page);
1266 out:
1267         return err;
1268 }
1269
1270 /*
1271  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1272  * but no new kernel page is allocated: kpage must already be a ksm page.
1273  *
1274  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1275  */
1276 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1277                                       struct page *page, struct page *kpage)
1278 {
1279         struct mm_struct *mm = rmap_item->mm;
1280         struct vm_area_struct *vma;
1281         int err = -EFAULT;
1282
1283         down_read(&mm->mmap_sem);
1284         vma = find_mergeable_vma(mm, rmap_item->address);
1285         if (!vma)
1286                 goto out;
1287
1288         err = try_to_merge_one_page(vma, page, kpage);
1289         if (err)
1290                 goto out;
1291
1292         /* Unstable nid is in union with stable anon_vma: remove first */
1293         remove_rmap_item_from_tree(rmap_item);
1294
1295         /* Must get reference to anon_vma while still holding mmap_sem */
1296         rmap_item->anon_vma = vma->anon_vma;
1297         get_anon_vma(vma->anon_vma);
1298 out:
1299         up_read(&mm->mmap_sem);
1300         return err;
1301 }
1302
1303 /*
1304  * try_to_merge_two_pages - take two identical pages and prepare them
1305  * to be merged into one page.
1306  *
1307  * This function returns the kpage if we successfully merged two identical
1308  * pages into one ksm page, NULL otherwise.
1309  *
1310  * Note that this function upgrades page to ksm page: if one of the pages
1311  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1312  */
1313 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1314                                            struct page *page,
1315                                            struct rmap_item *tree_rmap_item,
1316                                            struct page *tree_page)
1317 {
1318         int err;
1319
1320         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1321         if (!err) {
1322                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1323                                                         tree_page, page);
1324                 /*
1325                  * If that fails, we have a ksm page with only one pte
1326                  * pointing to it: so break it.
1327                  */
1328                 if (err)
1329                         break_cow(rmap_item);
1330         }
1331         return err ? NULL : page;
1332 }
1333
1334 static __always_inline
1335 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1336 {
1337         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1338         /*
1339          * Check that at least one mapping still exists, otherwise
1340          * there's no much point to merge and share with this
1341          * stable_node, as the underlying tree_page of the other
1342          * sharer is going to be freed soon.
1343          */
1344         return stable_node->rmap_hlist_len &&
1345                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1346 }
1347
1348 static __always_inline
1349 bool is_page_sharing_candidate(struct stable_node *stable_node)
1350 {
1351         return __is_page_sharing_candidate(stable_node, 0);
1352 }
1353
1354 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1355                                     struct stable_node **_stable_node,
1356                                     struct rb_root *root,
1357                                     bool prune_stale_stable_nodes)
1358 {
1359         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1360         struct hlist_node *hlist_safe;
1361         struct page *_tree_page, *tree_page = NULL;
1362         int nr = 0;
1363         int found_rmap_hlist_len;
1364
1365         if (!prune_stale_stable_nodes ||
1366             time_before(jiffies, stable_node->chain_prune_time +
1367                         msecs_to_jiffies(
1368                                 ksm_stable_node_chains_prune_millisecs)))
1369                 prune_stale_stable_nodes = false;
1370         else
1371                 stable_node->chain_prune_time = jiffies;
1372
1373         hlist_for_each_entry_safe(dup, hlist_safe,
1374                                   &stable_node->hlist, hlist_dup) {
1375                 cond_resched();
1376                 /*
1377                  * We must walk all stable_node_dup to prune the stale
1378                  * stable nodes during lookup.
1379                  *
1380                  * get_ksm_page can drop the nodes from the
1381                  * stable_node->hlist if they point to freed pages
1382                  * (that's why we do a _safe walk). The "dup"
1383                  * stable_node parameter itself will be freed from
1384                  * under us if it returns NULL.
1385                  */
1386                 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1387                 if (!_tree_page)
1388                         continue;
1389                 nr += 1;
1390                 if (is_page_sharing_candidate(dup)) {
1391                         if (!found ||
1392                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1393                                 if (found)
1394                                         put_page(tree_page);
1395                                 found = dup;
1396                                 found_rmap_hlist_len = found->rmap_hlist_len;
1397                                 tree_page = _tree_page;
1398
1399                                 /* skip put_page for found dup */
1400                                 if (!prune_stale_stable_nodes)
1401                                         break;
1402                                 continue;
1403                         }
1404                 }
1405                 put_page(_tree_page);
1406         }
1407
1408         if (found) {
1409                 /*
1410                  * nr is counting all dups in the chain only if
1411                  * prune_stale_stable_nodes is true, otherwise we may
1412                  * break the loop at nr == 1 even if there are
1413                  * multiple entries.
1414                  */
1415                 if (prune_stale_stable_nodes && nr == 1) {
1416                         /*
1417                          * If there's not just one entry it would
1418                          * corrupt memory, better BUG_ON. In KSM
1419                          * context with no lock held it's not even
1420                          * fatal.
1421                          */
1422                         BUG_ON(stable_node->hlist.first->next);
1423
1424                         /*
1425                          * There's just one entry and it is below the
1426                          * deduplication limit so drop the chain.
1427                          */
1428                         rb_replace_node(&stable_node->node, &found->node,
1429                                         root);
1430                         free_stable_node(stable_node);
1431                         ksm_stable_node_chains--;
1432                         ksm_stable_node_dups--;
1433                         /*
1434                          * NOTE: the caller depends on the stable_node
1435                          * to be equal to stable_node_dup if the chain
1436                          * was collapsed.
1437                          */
1438                         *_stable_node = found;
1439                         /*
1440                          * Just for robustneess as stable_node is
1441                          * otherwise left as a stable pointer, the
1442                          * compiler shall optimize it away at build
1443                          * time.
1444                          */
1445                         stable_node = NULL;
1446                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1447                            __is_page_sharing_candidate(found, 1)) {
1448                         /*
1449                          * If the found stable_node dup can accept one
1450                          * more future merge (in addition to the one
1451                          * that is underway) and is not at the head of
1452                          * the chain, put it there so next search will
1453                          * be quicker in the !prune_stale_stable_nodes
1454                          * case.
1455                          *
1456                          * NOTE: it would be inaccurate to use nr > 1
1457                          * instead of checking the hlist.first pointer
1458                          * directly, because in the
1459                          * prune_stale_stable_nodes case "nr" isn't
1460                          * the position of the found dup in the chain,
1461                          * but the total number of dups in the chain.
1462                          */
1463                         hlist_del(&found->hlist_dup);
1464                         hlist_add_head(&found->hlist_dup,
1465                                        &stable_node->hlist);
1466                 }
1467         }
1468
1469         *_stable_node_dup = found;
1470         return tree_page;
1471 }
1472
1473 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1474                                                struct rb_root *root)
1475 {
1476         if (!is_stable_node_chain(stable_node))
1477                 return stable_node;
1478         if (hlist_empty(&stable_node->hlist)) {
1479                 free_stable_node_chain(stable_node, root);
1480                 return NULL;
1481         }
1482         return hlist_entry(stable_node->hlist.first,
1483                            typeof(*stable_node), hlist_dup);
1484 }
1485
1486 /*
1487  * Like for get_ksm_page, this function can free the *_stable_node and
1488  * *_stable_node_dup if the returned tree_page is NULL.
1489  *
1490  * It can also free and overwrite *_stable_node with the found
1491  * stable_node_dup if the chain is collapsed (in which case
1492  * *_stable_node will be equal to *_stable_node_dup like if the chain
1493  * never existed). It's up to the caller to verify tree_page is not
1494  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1495  *
1496  * *_stable_node_dup is really a second output parameter of this
1497  * function and will be overwritten in all cases, the caller doesn't
1498  * need to initialize it.
1499  */
1500 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1501                                         struct stable_node **_stable_node,
1502                                         struct rb_root *root,
1503                                         bool prune_stale_stable_nodes)
1504 {
1505         struct stable_node *stable_node = *_stable_node;
1506         if (!is_stable_node_chain(stable_node)) {
1507                 if (is_page_sharing_candidate(stable_node)) {
1508                         *_stable_node_dup = stable_node;
1509                         return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1510                 }
1511                 /*
1512                  * _stable_node_dup set to NULL means the stable_node
1513                  * reached the ksm_max_page_sharing limit.
1514                  */
1515                 *_stable_node_dup = NULL;
1516                 return NULL;
1517         }
1518         return stable_node_dup(_stable_node_dup, _stable_node, root,
1519                                prune_stale_stable_nodes);
1520 }
1521
1522 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1523                                                 struct stable_node **s_n,
1524                                                 struct rb_root *root)
1525 {
1526         return __stable_node_chain(s_n_d, s_n, root, true);
1527 }
1528
1529 static __always_inline struct page *chain(struct stable_node **s_n_d,
1530                                           struct stable_node *s_n,
1531                                           struct rb_root *root)
1532 {
1533         struct stable_node *old_stable_node = s_n;
1534         struct page *tree_page;
1535
1536         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1537         /* not pruning dups so s_n cannot have changed */
1538         VM_BUG_ON(s_n != old_stable_node);
1539         return tree_page;
1540 }
1541
1542 /*
1543  * stable_tree_search - search for page inside the stable tree
1544  *
1545  * This function checks if there is a page inside the stable tree
1546  * with identical content to the page that we are scanning right now.
1547  *
1548  * This function returns the stable tree node of identical content if found,
1549  * NULL otherwise.
1550  */
1551 static struct page *stable_tree_search(struct page *page)
1552 {
1553         int nid;
1554         struct rb_root *root;
1555         struct rb_node **new;
1556         struct rb_node *parent;
1557         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1558         struct stable_node *page_node;
1559
1560         page_node = page_stable_node(page);
1561         if (page_node && page_node->head != &migrate_nodes) {
1562                 /* ksm page forked */
1563                 get_page(page);
1564                 return page;
1565         }
1566
1567         nid = get_kpfn_nid(page_to_pfn(page));
1568         root = root_stable_tree + nid;
1569 again:
1570         new = &root->rb_node;
1571         parent = NULL;
1572
1573         while (*new) {
1574                 struct page *tree_page;
1575                 int ret;
1576
1577                 cond_resched();
1578                 stable_node = rb_entry(*new, struct stable_node, node);
1579                 stable_node_any = NULL;
1580                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1581                 /*
1582                  * NOTE: stable_node may have been freed by
1583                  * chain_prune() if the returned stable_node_dup is
1584                  * not NULL. stable_node_dup may have been inserted in
1585                  * the rbtree instead as a regular stable_node (in
1586                  * order to collapse the stable_node chain if a single
1587                  * stable_node dup was found in it). In such case the
1588                  * stable_node is overwritten by the calleee to point
1589                  * to the stable_node_dup that was collapsed in the
1590                  * stable rbtree and stable_node will be equal to
1591                  * stable_node_dup like if the chain never existed.
1592                  */
1593                 if (!stable_node_dup) {
1594                         /*
1595                          * Either all stable_node dups were full in
1596                          * this stable_node chain, or this chain was
1597                          * empty and should be rb_erased.
1598                          */
1599                         stable_node_any = stable_node_dup_any(stable_node,
1600                                                               root);
1601                         if (!stable_node_any) {
1602                                 /* rb_erase just run */
1603                                 goto again;
1604                         }
1605                         /*
1606                          * Take any of the stable_node dups page of
1607                          * this stable_node chain to let the tree walk
1608                          * continue. All KSM pages belonging to the
1609                          * stable_node dups in a stable_node chain
1610                          * have the same content and they're
1611                          * wrprotected at all times. Any will work
1612                          * fine to continue the walk.
1613                          */
1614                         tree_page = get_ksm_page(stable_node_any,
1615                                                  GET_KSM_PAGE_NOLOCK);
1616                 }
1617                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1618                 if (!tree_page) {
1619                         /*
1620                          * If we walked over a stale stable_node,
1621                          * get_ksm_page() will call rb_erase() and it
1622                          * may rebalance the tree from under us. So
1623                          * restart the search from scratch. Returning
1624                          * NULL would be safe too, but we'd generate
1625                          * false negative insertions just because some
1626                          * stable_node was stale.
1627                          */
1628                         goto again;
1629                 }
1630
1631                 ret = memcmp_pages(page, tree_page);
1632                 put_page(tree_page);
1633
1634                 parent = *new;
1635                 if (ret < 0)
1636                         new = &parent->rb_left;
1637                 else if (ret > 0)
1638                         new = &parent->rb_right;
1639                 else {
1640                         if (page_node) {
1641                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1642                                 /*
1643                                  * Test if the migrated page should be merged
1644                                  * into a stable node dup. If the mapcount is
1645                                  * 1 we can migrate it with another KSM page
1646                                  * without adding it to the chain.
1647                                  */
1648                                 if (page_mapcount(page) > 1)
1649                                         goto chain_append;
1650                         }
1651
1652                         if (!stable_node_dup) {
1653                                 /*
1654                                  * If the stable_node is a chain and
1655                                  * we got a payload match in memcmp
1656                                  * but we cannot merge the scanned
1657                                  * page in any of the existing
1658                                  * stable_node dups because they're
1659                                  * all full, we need to wait the
1660                                  * scanned page to find itself a match
1661                                  * in the unstable tree to create a
1662                                  * brand new KSM page to add later to
1663                                  * the dups of this stable_node.
1664                                  */
1665                                 return NULL;
1666                         }
1667
1668                         /*
1669                          * Lock and unlock the stable_node's page (which
1670                          * might already have been migrated) so that page
1671                          * migration is sure to notice its raised count.
1672                          * It would be more elegant to return stable_node
1673                          * than kpage, but that involves more changes.
1674                          */
1675                         tree_page = get_ksm_page(stable_node_dup,
1676                                                  GET_KSM_PAGE_TRYLOCK);
1677
1678                         if (PTR_ERR(tree_page) == -EBUSY)
1679                                 return ERR_PTR(-EBUSY);
1680
1681                         if (unlikely(!tree_page))
1682                                 /*
1683                                  * The tree may have been rebalanced,
1684                                  * so re-evaluate parent and new.
1685                                  */
1686                                 goto again;
1687                         unlock_page(tree_page);
1688
1689                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1690                             NUMA(stable_node_dup->nid)) {
1691                                 put_page(tree_page);
1692                                 goto replace;
1693                         }
1694                         return tree_page;
1695                 }
1696         }
1697
1698         if (!page_node)
1699                 return NULL;
1700
1701         list_del(&page_node->list);
1702         DO_NUMA(page_node->nid = nid);
1703         rb_link_node(&page_node->node, parent, new);
1704         rb_insert_color(&page_node->node, root);
1705 out:
1706         if (is_page_sharing_candidate(page_node)) {
1707                 get_page(page);
1708                 return page;
1709         } else
1710                 return NULL;
1711
1712 replace:
1713         /*
1714          * If stable_node was a chain and chain_prune collapsed it,
1715          * stable_node has been updated to be the new regular
1716          * stable_node. A collapse of the chain is indistinguishable
1717          * from the case there was no chain in the stable
1718          * rbtree. Otherwise stable_node is the chain and
1719          * stable_node_dup is the dup to replace.
1720          */
1721         if (stable_node_dup == stable_node) {
1722                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1723                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1724                 /* there is no chain */
1725                 if (page_node) {
1726                         VM_BUG_ON(page_node->head != &migrate_nodes);
1727                         list_del(&page_node->list);
1728                         DO_NUMA(page_node->nid = nid);
1729                         rb_replace_node(&stable_node_dup->node,
1730                                         &page_node->node,
1731                                         root);
1732                         if (is_page_sharing_candidate(page_node))
1733                                 get_page(page);
1734                         else
1735                                 page = NULL;
1736                 } else {
1737                         rb_erase(&stable_node_dup->node, root);
1738                         page = NULL;
1739                 }
1740         } else {
1741                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1742                 __stable_node_dup_del(stable_node_dup);
1743                 if (page_node) {
1744                         VM_BUG_ON(page_node->head != &migrate_nodes);
1745                         list_del(&page_node->list);
1746                         DO_NUMA(page_node->nid = nid);
1747                         stable_node_chain_add_dup(page_node, stable_node);
1748                         if (is_page_sharing_candidate(page_node))
1749                                 get_page(page);
1750                         else
1751                                 page = NULL;
1752                 } else {
1753                         page = NULL;
1754                 }
1755         }
1756         stable_node_dup->head = &migrate_nodes;
1757         list_add(&stable_node_dup->list, stable_node_dup->head);
1758         return page;
1759
1760 chain_append:
1761         /* stable_node_dup could be null if it reached the limit */
1762         if (!stable_node_dup)
1763                 stable_node_dup = stable_node_any;
1764         /*
1765          * If stable_node was a chain and chain_prune collapsed it,
1766          * stable_node has been updated to be the new regular
1767          * stable_node. A collapse of the chain is indistinguishable
1768          * from the case there was no chain in the stable
1769          * rbtree. Otherwise stable_node is the chain and
1770          * stable_node_dup is the dup to replace.
1771          */
1772         if (stable_node_dup == stable_node) {
1773                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1774                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1775                 /* chain is missing so create it */
1776                 stable_node = alloc_stable_node_chain(stable_node_dup,
1777                                                       root);
1778                 if (!stable_node)
1779                         return NULL;
1780         }
1781         /*
1782          * Add this stable_node dup that was
1783          * migrated to the stable_node chain
1784          * of the current nid for this page
1785          * content.
1786          */
1787         VM_BUG_ON(!is_stable_node_chain(stable_node));
1788         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1789         VM_BUG_ON(page_node->head != &migrate_nodes);
1790         list_del(&page_node->list);
1791         DO_NUMA(page_node->nid = nid);
1792         stable_node_chain_add_dup(page_node, stable_node);
1793         goto out;
1794 }
1795
1796 /*
1797  * stable_tree_insert - insert stable tree node pointing to new ksm page
1798  * into the stable tree.
1799  *
1800  * This function returns the stable tree node just allocated on success,
1801  * NULL otherwise.
1802  */
1803 static struct stable_node *stable_tree_insert(struct page *kpage)
1804 {
1805         int nid;
1806         unsigned long kpfn;
1807         struct rb_root *root;
1808         struct rb_node **new;
1809         struct rb_node *parent;
1810         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1811         bool need_chain = false;
1812
1813         kpfn = page_to_pfn(kpage);
1814         nid = get_kpfn_nid(kpfn);
1815         root = root_stable_tree + nid;
1816 again:
1817         parent = NULL;
1818         new = &root->rb_node;
1819
1820         while (*new) {
1821                 struct page *tree_page;
1822                 int ret;
1823
1824                 cond_resched();
1825                 stable_node = rb_entry(*new, struct stable_node, node);
1826                 stable_node_any = NULL;
1827                 tree_page = chain(&stable_node_dup, stable_node, root);
1828                 if (!stable_node_dup) {
1829                         /*
1830                          * Either all stable_node dups were full in
1831                          * this stable_node chain, or this chain was
1832                          * empty and should be rb_erased.
1833                          */
1834                         stable_node_any = stable_node_dup_any(stable_node,
1835                                                               root);
1836                         if (!stable_node_any) {
1837                                 /* rb_erase just run */
1838                                 goto again;
1839                         }
1840                         /*
1841                          * Take any of the stable_node dups page of
1842                          * this stable_node chain to let the tree walk
1843                          * continue. All KSM pages belonging to the
1844                          * stable_node dups in a stable_node chain
1845                          * have the same content and they're
1846                          * wrprotected at all times. Any will work
1847                          * fine to continue the walk.
1848                          */
1849                         tree_page = get_ksm_page(stable_node_any,
1850                                                  GET_KSM_PAGE_NOLOCK);
1851                 }
1852                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1853                 if (!tree_page) {
1854                         /*
1855                          * If we walked over a stale stable_node,
1856                          * get_ksm_page() will call rb_erase() and it
1857                          * may rebalance the tree from under us. So
1858                          * restart the search from scratch. Returning
1859                          * NULL would be safe too, but we'd generate
1860                          * false negative insertions just because some
1861                          * stable_node was stale.
1862                          */
1863                         goto again;
1864                 }
1865
1866                 ret = memcmp_pages(kpage, tree_page);
1867                 put_page(tree_page);
1868
1869                 parent = *new;
1870                 if (ret < 0)
1871                         new = &parent->rb_left;
1872                 else if (ret > 0)
1873                         new = &parent->rb_right;
1874                 else {
1875                         need_chain = true;
1876                         break;
1877                 }
1878         }
1879
1880         stable_node_dup = alloc_stable_node();
1881         if (!stable_node_dup)
1882                 return NULL;
1883
1884         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1885         stable_node_dup->kpfn = kpfn;
1886         set_page_stable_node(kpage, stable_node_dup);
1887         stable_node_dup->rmap_hlist_len = 0;
1888         DO_NUMA(stable_node_dup->nid = nid);
1889         if (!need_chain) {
1890                 rb_link_node(&stable_node_dup->node, parent, new);
1891                 rb_insert_color(&stable_node_dup->node, root);
1892         } else {
1893                 if (!is_stable_node_chain(stable_node)) {
1894                         struct stable_node *orig = stable_node;
1895                         /* chain is missing so create it */
1896                         stable_node = alloc_stable_node_chain(orig, root);
1897                         if (!stable_node) {
1898                                 free_stable_node(stable_node_dup);
1899                                 return NULL;
1900                         }
1901                 }
1902                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1903         }
1904
1905         return stable_node_dup;
1906 }
1907
1908 /*
1909  * unstable_tree_search_insert - search for identical page,
1910  * else insert rmap_item into the unstable tree.
1911  *
1912  * This function searches for a page in the unstable tree identical to the
1913  * page currently being scanned; and if no identical page is found in the
1914  * tree, we insert rmap_item as a new object into the unstable tree.
1915  *
1916  * This function returns pointer to rmap_item found to be identical
1917  * to the currently scanned page, NULL otherwise.
1918  *
1919  * This function does both searching and inserting, because they share
1920  * the same walking algorithm in an rbtree.
1921  */
1922 static
1923 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1924                                               struct page *page,
1925                                               struct page **tree_pagep)
1926 {
1927         struct rb_node **new;
1928         struct rb_root *root;
1929         struct rb_node *parent = NULL;
1930         int nid;
1931
1932         nid = get_kpfn_nid(page_to_pfn(page));
1933         root = root_unstable_tree + nid;
1934         new = &root->rb_node;
1935
1936         while (*new) {
1937                 struct rmap_item *tree_rmap_item;
1938                 struct page *tree_page;
1939                 int ret;
1940
1941                 cond_resched();
1942                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1943                 tree_page = get_mergeable_page(tree_rmap_item);
1944                 if (!tree_page)
1945                         return NULL;
1946
1947                 /*
1948                  * Don't substitute a ksm page for a forked page.
1949                  */
1950                 if (page == tree_page) {
1951                         put_page(tree_page);
1952                         return NULL;
1953                 }
1954
1955                 ret = memcmp_pages(page, tree_page);
1956
1957                 parent = *new;
1958                 if (ret < 0) {
1959                         put_page(tree_page);
1960                         new = &parent->rb_left;
1961                 } else if (ret > 0) {
1962                         put_page(tree_page);
1963                         new = &parent->rb_right;
1964                 } else if (!ksm_merge_across_nodes &&
1965                            page_to_nid(tree_page) != nid) {
1966                         /*
1967                          * If tree_page has been migrated to another NUMA node,
1968                          * it will be flushed out and put in the right unstable
1969                          * tree next time: only merge with it when across_nodes.
1970                          */
1971                         put_page(tree_page);
1972                         return NULL;
1973                 } else {
1974                         *tree_pagep = tree_page;
1975                         return tree_rmap_item;
1976                 }
1977         }
1978
1979         rmap_item->address |= UNSTABLE_FLAG;
1980         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1981         DO_NUMA(rmap_item->nid = nid);
1982         rb_link_node(&rmap_item->node, parent, new);
1983         rb_insert_color(&rmap_item->node, root);
1984
1985         ksm_pages_unshared++;
1986         return NULL;
1987 }
1988
1989 /*
1990  * stable_tree_append - add another rmap_item to the linked list of
1991  * rmap_items hanging off a given node of the stable tree, all sharing
1992  * the same ksm page.
1993  */
1994 static void stable_tree_append(struct rmap_item *rmap_item,
1995                                struct stable_node *stable_node,
1996                                bool max_page_sharing_bypass)
1997 {
1998         /*
1999          * rmap won't find this mapping if we don't insert the
2000          * rmap_item in the right stable_node
2001          * duplicate. page_migration could break later if rmap breaks,
2002          * so we can as well crash here. We really need to check for
2003          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2004          * for other negative values as an undeflow if detected here
2005          * for the first time (and not when decreasing rmap_hlist_len)
2006          * would be sign of memory corruption in the stable_node.
2007          */
2008         BUG_ON(stable_node->rmap_hlist_len < 0);
2009
2010         stable_node->rmap_hlist_len++;
2011         if (!max_page_sharing_bypass)
2012                 /* possibly non fatal but unexpected overflow, only warn */
2013                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2014                              ksm_max_page_sharing);
2015
2016         rmap_item->head = stable_node;
2017         rmap_item->address |= STABLE_FLAG;
2018         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2019
2020         if (rmap_item->hlist.next)
2021                 ksm_pages_sharing++;
2022         else
2023                 ksm_pages_shared++;
2024 }
2025
2026 /*
2027  * cmp_and_merge_page - first see if page can be merged into the stable tree;
2028  * if not, compare checksum to previous and if it's the same, see if page can
2029  * be inserted into the unstable tree, or merged with a page already there and
2030  * both transferred to the stable tree.
2031  *
2032  * @page: the page that we are searching identical page to.
2033  * @rmap_item: the reverse mapping into the virtual address of this page
2034  */
2035 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2036 {
2037         struct mm_struct *mm = rmap_item->mm;
2038         struct rmap_item *tree_rmap_item;
2039         struct page *tree_page = NULL;
2040         struct stable_node *stable_node;
2041         struct page *kpage;
2042         unsigned int checksum;
2043         int err;
2044         bool max_page_sharing_bypass = false;
2045
2046         stable_node = page_stable_node(page);
2047         if (stable_node) {
2048                 if (stable_node->head != &migrate_nodes &&
2049                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2050                     NUMA(stable_node->nid)) {
2051                         stable_node_dup_del(stable_node);
2052                         stable_node->head = &migrate_nodes;
2053                         list_add(&stable_node->list, stable_node->head);
2054                 }
2055                 if (stable_node->head != &migrate_nodes &&
2056                     rmap_item->head == stable_node)
2057                         return;
2058                 /*
2059                  * If it's a KSM fork, allow it to go over the sharing limit
2060                  * without warnings.
2061                  */
2062                 if (!is_page_sharing_candidate(stable_node))
2063                         max_page_sharing_bypass = true;
2064         }
2065
2066         /* We first start with searching the page inside the stable tree */
2067         kpage = stable_tree_search(page);
2068         if (kpage == page && rmap_item->head == stable_node) {
2069                 put_page(kpage);
2070                 return;
2071         }
2072
2073         remove_rmap_item_from_tree(rmap_item);
2074
2075         if (kpage) {
2076                 if (PTR_ERR(kpage) == -EBUSY)
2077                         return;
2078
2079                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2080                 if (!err) {
2081                         /*
2082                          * The page was successfully merged:
2083                          * add its rmap_item to the stable tree.
2084                          */
2085                         lock_page(kpage);
2086                         stable_tree_append(rmap_item, page_stable_node(kpage),
2087                                            max_page_sharing_bypass);
2088                         unlock_page(kpage);
2089                 }
2090                 put_page(kpage);
2091                 return;
2092         }
2093
2094         /*
2095          * If the hash value of the page has changed from the last time
2096          * we calculated it, this page is changing frequently: therefore we
2097          * don't want to insert it in the unstable tree, and we don't want
2098          * to waste our time searching for something identical to it there.
2099          */
2100         checksum = calc_checksum(page);
2101         if (rmap_item->oldchecksum != checksum) {
2102                 rmap_item->oldchecksum = checksum;
2103                 return;
2104         }
2105
2106         /*
2107          * Same checksum as an empty page. We attempt to merge it with the
2108          * appropriate zero page if the user enabled this via sysfs.
2109          */
2110         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2111                 struct vm_area_struct *vma;
2112
2113                 down_read(&mm->mmap_sem);
2114                 vma = find_mergeable_vma(mm, rmap_item->address);
2115                 err = try_to_merge_one_page(vma, page,
2116                                             ZERO_PAGE(rmap_item->address));
2117                 up_read(&mm->mmap_sem);
2118                 /*
2119                  * In case of failure, the page was not really empty, so we
2120                  * need to continue. Otherwise we're done.
2121                  */
2122                 if (!err)
2123                         return;
2124         }
2125         tree_rmap_item =
2126                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2127         if (tree_rmap_item) {
2128                 bool split;
2129
2130                 kpage = try_to_merge_two_pages(rmap_item, page,
2131                                                 tree_rmap_item, tree_page);
2132                 /*
2133                  * If both pages we tried to merge belong to the same compound
2134                  * page, then we actually ended up increasing the reference
2135                  * count of the same compound page twice, and split_huge_page
2136                  * failed.
2137                  * Here we set a flag if that happened, and we use it later to
2138                  * try split_huge_page again. Since we call put_page right
2139                  * afterwards, the reference count will be correct and
2140                  * split_huge_page should succeed.
2141                  */
2142                 split = PageTransCompound(page)
2143                         && compound_head(page) == compound_head(tree_page);
2144                 put_page(tree_page);
2145                 if (kpage) {
2146                         /*
2147                          * The pages were successfully merged: insert new
2148                          * node in the stable tree and add both rmap_items.
2149                          */
2150                         lock_page(kpage);
2151                         stable_node = stable_tree_insert(kpage);
2152                         if (stable_node) {
2153                                 stable_tree_append(tree_rmap_item, stable_node,
2154                                                    false);
2155                                 stable_tree_append(rmap_item, stable_node,
2156                                                    false);
2157                         }
2158                         unlock_page(kpage);
2159
2160                         /*
2161                          * If we fail to insert the page into the stable tree,
2162                          * we will have 2 virtual addresses that are pointing
2163                          * to a ksm page left outside the stable tree,
2164                          * in which case we need to break_cow on both.
2165                          */
2166                         if (!stable_node) {
2167                                 break_cow(tree_rmap_item);
2168                                 break_cow(rmap_item);
2169                         }
2170                 } else if (split) {
2171                         /*
2172                          * We are here if we tried to merge two pages and
2173                          * failed because they both belonged to the same
2174                          * compound page. We will split the page now, but no
2175                          * merging will take place.
2176                          * We do not want to add the cost of a full lock; if
2177                          * the page is locked, it is better to skip it and
2178                          * perhaps try again later.
2179                          */
2180                         if (!trylock_page(page))
2181                                 return;
2182                         split_huge_page(page);
2183                         unlock_page(page);
2184                 }
2185         }
2186 }
2187
2188 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2189                                             struct rmap_item **rmap_list,
2190                                             unsigned long addr)
2191 {
2192         struct rmap_item *rmap_item;
2193
2194         while (*rmap_list) {
2195                 rmap_item = *rmap_list;
2196                 if ((rmap_item->address & PAGE_MASK) == addr)
2197                         return rmap_item;
2198                 if (rmap_item->address > addr)
2199                         break;
2200                 *rmap_list = rmap_item->rmap_list;
2201                 remove_rmap_item_from_tree(rmap_item);
2202                 free_rmap_item(rmap_item);
2203         }
2204
2205         rmap_item = alloc_rmap_item();
2206         if (rmap_item) {
2207                 /* It has already been zeroed */
2208                 rmap_item->mm = mm_slot->mm;
2209                 rmap_item->address = addr;
2210                 rmap_item->rmap_list = *rmap_list;
2211                 *rmap_list = rmap_item;
2212         }
2213         return rmap_item;
2214 }
2215
2216 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2217 {
2218         struct mm_struct *mm;
2219         struct mm_slot *slot;
2220         struct vm_area_struct *vma;
2221         struct rmap_item *rmap_item;
2222         int nid;
2223
2224         if (list_empty(&ksm_mm_head.mm_list))
2225                 return NULL;
2226
2227         slot = ksm_scan.mm_slot;
2228         if (slot == &ksm_mm_head) {
2229                 /*
2230                  * A number of pages can hang around indefinitely on per-cpu
2231                  * pagevecs, raised page count preventing write_protect_page
2232                  * from merging them.  Though it doesn't really matter much,
2233                  * it is puzzling to see some stuck in pages_volatile until
2234                  * other activity jostles them out, and they also prevented
2235                  * LTP's KSM test from succeeding deterministically; so drain
2236                  * them here (here rather than on entry to ksm_do_scan(),
2237                  * so we don't IPI too often when pages_to_scan is set low).
2238                  */
2239                 lru_add_drain_all();
2240
2241                 /*
2242                  * Whereas stale stable_nodes on the stable_tree itself
2243                  * get pruned in the regular course of stable_tree_search(),
2244                  * those moved out to the migrate_nodes list can accumulate:
2245                  * so prune them once before each full scan.
2246                  */
2247                 if (!ksm_merge_across_nodes) {
2248                         struct stable_node *stable_node, *next;
2249                         struct page *page;
2250
2251                         list_for_each_entry_safe(stable_node, next,
2252                                                  &migrate_nodes, list) {
2253                                 page = get_ksm_page(stable_node,
2254                                                     GET_KSM_PAGE_NOLOCK);
2255                                 if (page)
2256                                         put_page(page);
2257                                 cond_resched();
2258                         }
2259                 }
2260
2261                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2262                         root_unstable_tree[nid] = RB_ROOT;
2263
2264                 spin_lock(&ksm_mmlist_lock);
2265                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2266                 ksm_scan.mm_slot = slot;
2267                 spin_unlock(&ksm_mmlist_lock);
2268                 /*
2269                  * Although we tested list_empty() above, a racing __ksm_exit
2270                  * of the last mm on the list may have removed it since then.
2271                  */
2272                 if (slot == &ksm_mm_head)
2273                         return NULL;
2274 next_mm:
2275                 ksm_scan.address = 0;
2276                 ksm_scan.rmap_list = &slot->rmap_list;
2277         }
2278
2279         mm = slot->mm;
2280         down_read(&mm->mmap_sem);
2281         if (ksm_test_exit(mm))
2282                 vma = NULL;
2283         else
2284                 vma = find_vma(mm, ksm_scan.address);
2285
2286         for (; vma; vma = vma->vm_next) {
2287                 if (!(vma->vm_flags & VM_MERGEABLE))
2288                         continue;
2289                 if (ksm_scan.address < vma->vm_start)
2290                         ksm_scan.address = vma->vm_start;
2291                 if (!vma->anon_vma)
2292                         ksm_scan.address = vma->vm_end;
2293
2294                 while (ksm_scan.address < vma->vm_end) {
2295                         if (ksm_test_exit(mm))
2296                                 break;
2297                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2298                         if (IS_ERR_OR_NULL(*page)) {
2299                                 ksm_scan.address += PAGE_SIZE;
2300                                 cond_resched();
2301                                 continue;
2302                         }
2303                         if (PageAnon(*page)) {
2304                                 flush_anon_page(vma, *page, ksm_scan.address);
2305                                 flush_dcache_page(*page);
2306                                 rmap_item = get_next_rmap_item(slot,
2307                                         ksm_scan.rmap_list, ksm_scan.address);
2308                                 if (rmap_item) {
2309                                         ksm_scan.rmap_list =
2310                                                         &rmap_item->rmap_list;
2311                                         ksm_scan.address += PAGE_SIZE;
2312                                 } else
2313                                         put_page(*page);
2314                                 up_read(&mm->mmap_sem);
2315                                 return rmap_item;
2316                         }
2317                         put_page(*page);
2318                         ksm_scan.address += PAGE_SIZE;
2319                         cond_resched();
2320                 }
2321         }
2322
2323         if (ksm_test_exit(mm)) {
2324                 ksm_scan.address = 0;
2325                 ksm_scan.rmap_list = &slot->rmap_list;
2326         }
2327         /*
2328          * Nuke all the rmap_items that are above this current rmap:
2329          * because there were no VM_MERGEABLE vmas with such addresses.
2330          */
2331         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2332
2333         spin_lock(&ksm_mmlist_lock);
2334         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2335                                                 struct mm_slot, mm_list);
2336         if (ksm_scan.address == 0) {
2337                 /*
2338                  * We've completed a full scan of all vmas, holding mmap_sem
2339                  * throughout, and found no VM_MERGEABLE: so do the same as
2340                  * __ksm_exit does to remove this mm from all our lists now.
2341                  * This applies either when cleaning up after __ksm_exit
2342                  * (but beware: we can reach here even before __ksm_exit),
2343                  * or when all VM_MERGEABLE areas have been unmapped (and
2344                  * mmap_sem then protects against race with MADV_MERGEABLE).
2345                  */
2346                 hash_del(&slot->link);
2347                 list_del(&slot->mm_list);
2348                 spin_unlock(&ksm_mmlist_lock);
2349
2350                 free_mm_slot(slot);
2351                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2352                 up_read(&mm->mmap_sem);
2353                 mmdrop(mm);
2354         } else {
2355                 up_read(&mm->mmap_sem);
2356                 /*
2357                  * up_read(&mm->mmap_sem) first because after
2358                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2359                  * already have been freed under us by __ksm_exit()
2360                  * because the "mm_slot" is still hashed and
2361                  * ksm_scan.mm_slot doesn't point to it anymore.
2362                  */
2363                 spin_unlock(&ksm_mmlist_lock);
2364         }
2365
2366         /* Repeat until we've completed scanning the whole list */
2367         slot = ksm_scan.mm_slot;
2368         if (slot != &ksm_mm_head)
2369                 goto next_mm;
2370
2371         ksm_scan.seqnr++;
2372         return NULL;
2373 }
2374
2375 /**
2376  * ksm_do_scan  - the ksm scanner main worker function.
2377  * @scan_npages:  number of pages we want to scan before we return.
2378  */
2379 static void ksm_do_scan(unsigned int scan_npages)
2380 {
2381         struct rmap_item *rmap_item;
2382         struct page *uninitialized_var(page);
2383
2384         while (scan_npages-- && likely(!freezing(current))) {
2385                 cond_resched();
2386                 rmap_item = scan_get_next_rmap_item(&page);
2387                 if (!rmap_item)
2388                         return;
2389                 cmp_and_merge_page(page, rmap_item);
2390                 put_page(page);
2391         }
2392 }
2393
2394 static int ksmd_should_run(void)
2395 {
2396         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2397 }
2398
2399 static int ksm_scan_thread(void *nothing)
2400 {
2401         unsigned int sleep_ms;
2402
2403         set_freezable();
2404         set_user_nice(current, 5);
2405
2406         while (!kthread_should_stop()) {
2407                 mutex_lock(&ksm_thread_mutex);
2408                 wait_while_offlining();
2409                 if (ksmd_should_run())
2410                         ksm_do_scan(ksm_thread_pages_to_scan);
2411                 mutex_unlock(&ksm_thread_mutex);
2412
2413                 try_to_freeze();
2414
2415                 if (ksmd_should_run()) {
2416                         sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2417                         wait_event_interruptible_timeout(ksm_iter_wait,
2418                                 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2419                                 msecs_to_jiffies(sleep_ms));
2420                 } else {
2421                         wait_event_freezable(ksm_thread_wait,
2422                                 ksmd_should_run() || kthread_should_stop());
2423                 }
2424         }
2425         return 0;
2426 }
2427
2428 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2429                 unsigned long end, int advice, unsigned long *vm_flags)
2430 {
2431         struct mm_struct *mm = vma->vm_mm;
2432         int err;
2433
2434         switch (advice) {
2435         case MADV_MERGEABLE:
2436                 /*
2437                  * Be somewhat over-protective for now!
2438                  */
2439                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2440                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2441                                  VM_HUGETLB | VM_MIXEDMAP))
2442                         return 0;               /* just ignore the advice */
2443
2444                 if (vma_is_dax(vma))
2445                         return 0;
2446
2447 #ifdef VM_SAO
2448                 if (*vm_flags & VM_SAO)
2449                         return 0;
2450 #endif
2451 #ifdef VM_SPARC_ADI
2452                 if (*vm_flags & VM_SPARC_ADI)
2453                         return 0;
2454 #endif
2455
2456                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2457                         err = __ksm_enter(mm);
2458                         if (err)
2459                                 return err;
2460                 }
2461
2462                 *vm_flags |= VM_MERGEABLE;
2463                 break;
2464
2465         case MADV_UNMERGEABLE:
2466                 if (!(*vm_flags & VM_MERGEABLE))
2467                         return 0;               /* just ignore the advice */
2468
2469                 if (vma->anon_vma) {
2470                         err = unmerge_ksm_pages(vma, start, end);
2471                         if (err)
2472                                 return err;
2473                 }
2474
2475                 *vm_flags &= ~VM_MERGEABLE;
2476                 break;
2477         }
2478
2479         return 0;
2480 }
2481
2482 int __ksm_enter(struct mm_struct *mm)
2483 {
2484         struct mm_slot *mm_slot;
2485         int needs_wakeup;
2486
2487         mm_slot = alloc_mm_slot();
2488         if (!mm_slot)
2489                 return -ENOMEM;
2490
2491         /* Check ksm_run too?  Would need tighter locking */
2492         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2493
2494         spin_lock(&ksm_mmlist_lock);
2495         insert_to_mm_slots_hash(mm, mm_slot);
2496         /*
2497          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2498          * insert just behind the scanning cursor, to let the area settle
2499          * down a little; when fork is followed by immediate exec, we don't
2500          * want ksmd to waste time setting up and tearing down an rmap_list.
2501          *
2502          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2503          * scanning cursor, otherwise KSM pages in newly forked mms will be
2504          * missed: then we might as well insert at the end of the list.
2505          */
2506         if (ksm_run & KSM_RUN_UNMERGE)
2507                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2508         else
2509                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2510         spin_unlock(&ksm_mmlist_lock);
2511
2512         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2513         mmgrab(mm);
2514
2515         if (needs_wakeup)
2516                 wake_up_interruptible(&ksm_thread_wait);
2517
2518         return 0;
2519 }
2520
2521 void __ksm_exit(struct mm_struct *mm)
2522 {
2523         struct mm_slot *mm_slot;
2524         int easy_to_free = 0;
2525
2526         /*
2527          * This process is exiting: if it's straightforward (as is the
2528          * case when ksmd was never running), free mm_slot immediately.
2529          * But if it's at the cursor or has rmap_items linked to it, use
2530          * mmap_sem to synchronize with any break_cows before pagetables
2531          * are freed, and leave the mm_slot on the list for ksmd to free.
2532          * Beware: ksm may already have noticed it exiting and freed the slot.
2533          */
2534
2535         spin_lock(&ksm_mmlist_lock);
2536         mm_slot = get_mm_slot(mm);
2537         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2538                 if (!mm_slot->rmap_list) {
2539                         hash_del(&mm_slot->link);
2540                         list_del(&mm_slot->mm_list);
2541                         easy_to_free = 1;
2542                 } else {
2543                         list_move(&mm_slot->mm_list,
2544                                   &ksm_scan.mm_slot->mm_list);
2545                 }
2546         }
2547         spin_unlock(&ksm_mmlist_lock);
2548
2549         if (easy_to_free) {
2550                 free_mm_slot(mm_slot);
2551                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2552                 mmdrop(mm);
2553         } else if (mm_slot) {
2554                 down_write(&mm->mmap_sem);
2555                 up_write(&mm->mmap_sem);
2556         }
2557 }
2558
2559 struct page *ksm_might_need_to_copy(struct page *page,
2560                         struct vm_area_struct *vma, unsigned long address)
2561 {
2562         struct anon_vma *anon_vma = page_anon_vma(page);
2563         struct page *new_page;
2564
2565         if (PageKsm(page)) {
2566                 if (page_stable_node(page) &&
2567                     !(ksm_run & KSM_RUN_UNMERGE))
2568                         return page;    /* no need to copy it */
2569         } else if (!anon_vma) {
2570                 return page;            /* no need to copy it */
2571         } else if (anon_vma->root == vma->anon_vma->root &&
2572                  page->index == linear_page_index(vma, address)) {
2573                 return page;            /* still no need to copy it */
2574         }
2575         if (!PageUptodate(page))
2576                 return page;            /* let do_swap_page report the error */
2577
2578         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2579         if (new_page) {
2580                 copy_user_highpage(new_page, page, address, vma);
2581
2582                 SetPageDirty(new_page);
2583                 __SetPageUptodate(new_page);
2584                 __SetPageLocked(new_page);
2585         }
2586
2587         return new_page;
2588 }
2589
2590 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2591 {
2592         struct stable_node *stable_node;
2593         struct rmap_item *rmap_item;
2594         int search_new_forks = 0;
2595
2596         VM_BUG_ON_PAGE(!PageKsm(page), page);
2597
2598         /*
2599          * Rely on the page lock to protect against concurrent modifications
2600          * to that page's node of the stable tree.
2601          */
2602         VM_BUG_ON_PAGE(!PageLocked(page), page);
2603
2604         stable_node = page_stable_node(page);
2605         if (!stable_node)
2606                 return;
2607 again:
2608         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2609                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2610                 struct anon_vma_chain *vmac;
2611                 struct vm_area_struct *vma;
2612
2613                 cond_resched();
2614                 anon_vma_lock_read(anon_vma);
2615                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2616                                                0, ULONG_MAX) {
2617                         unsigned long addr;
2618
2619                         cond_resched();
2620                         vma = vmac->vma;
2621
2622                         /* Ignore the stable/unstable/sqnr flags */
2623                         addr = rmap_item->address & ~KSM_FLAG_MASK;
2624
2625                         if (addr < vma->vm_start || addr >= vma->vm_end)
2626                                 continue;
2627                         /*
2628                          * Initially we examine only the vma which covers this
2629                          * rmap_item; but later, if there is still work to do,
2630                          * we examine covering vmas in other mms: in case they
2631                          * were forked from the original since ksmd passed.
2632                          */
2633                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2634                                 continue;
2635
2636                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2637                                 continue;
2638
2639                         if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2640                                 anon_vma_unlock_read(anon_vma);
2641                                 return;
2642                         }
2643                         if (rwc->done && rwc->done(page)) {
2644                                 anon_vma_unlock_read(anon_vma);
2645                                 return;
2646                         }
2647                 }
2648                 anon_vma_unlock_read(anon_vma);
2649         }
2650         if (!search_new_forks++)
2651                 goto again;
2652 }
2653
2654 bool reuse_ksm_page(struct page *page,
2655                     struct vm_area_struct *vma,
2656                     unsigned long address)
2657 {
2658 #ifdef CONFIG_DEBUG_VM
2659         if (WARN_ON(is_zero_pfn(page_to_pfn(page))) ||
2660                         WARN_ON(!page_mapped(page)) ||
2661                         WARN_ON(!PageLocked(page))) {
2662                 dump_page(page, "reuse_ksm_page");
2663                 return false;
2664         }
2665 #endif
2666
2667         if (PageSwapCache(page) || !page_stable_node(page))
2668                 return false;
2669         /* Prohibit parallel get_ksm_page() */
2670         if (!page_ref_freeze(page, 1))
2671                 return false;
2672
2673         page_move_anon_rmap(page, vma);
2674         page->index = linear_page_index(vma, address);
2675         page_ref_unfreeze(page, 1);
2676
2677         return true;
2678 }
2679 #ifdef CONFIG_MIGRATION
2680 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2681 {
2682         struct stable_node *stable_node;
2683
2684         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2685         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2686         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2687
2688         stable_node = page_stable_node(newpage);
2689         if (stable_node) {
2690                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2691                 stable_node->kpfn = page_to_pfn(newpage);
2692                 /*
2693                  * newpage->mapping was set in advance; now we need smp_wmb()
2694                  * to make sure that the new stable_node->kpfn is visible
2695                  * to get_ksm_page() before it can see that oldpage->mapping
2696                  * has gone stale (or that PageSwapCache has been cleared).
2697                  */
2698                 smp_wmb();
2699                 set_page_stable_node(oldpage, NULL);
2700         }
2701 }
2702 #endif /* CONFIG_MIGRATION */
2703
2704 #ifdef CONFIG_MEMORY_HOTREMOVE
2705 static void wait_while_offlining(void)
2706 {
2707         while (ksm_run & KSM_RUN_OFFLINE) {
2708                 mutex_unlock(&ksm_thread_mutex);
2709                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2710                             TASK_UNINTERRUPTIBLE);
2711                 mutex_lock(&ksm_thread_mutex);
2712         }
2713 }
2714
2715 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2716                                          unsigned long start_pfn,
2717                                          unsigned long end_pfn)
2718 {
2719         if (stable_node->kpfn >= start_pfn &&
2720             stable_node->kpfn < end_pfn) {
2721                 /*
2722                  * Don't get_ksm_page, page has already gone:
2723                  * which is why we keep kpfn instead of page*
2724                  */
2725                 remove_node_from_stable_tree(stable_node);
2726                 return true;
2727         }
2728         return false;
2729 }
2730
2731 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2732                                            unsigned long start_pfn,
2733                                            unsigned long end_pfn,
2734                                            struct rb_root *root)
2735 {
2736         struct stable_node *dup;
2737         struct hlist_node *hlist_safe;
2738
2739         if (!is_stable_node_chain(stable_node)) {
2740                 VM_BUG_ON(is_stable_node_dup(stable_node));
2741                 return stable_node_dup_remove_range(stable_node, start_pfn,
2742                                                     end_pfn);
2743         }
2744
2745         hlist_for_each_entry_safe(dup, hlist_safe,
2746                                   &stable_node->hlist, hlist_dup) {
2747                 VM_BUG_ON(!is_stable_node_dup(dup));
2748                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2749         }
2750         if (hlist_empty(&stable_node->hlist)) {
2751                 free_stable_node_chain(stable_node, root);
2752                 return true; /* notify caller that tree was rebalanced */
2753         } else
2754                 return false;
2755 }
2756
2757 static void ksm_check_stable_tree(unsigned long start_pfn,
2758                                   unsigned long end_pfn)
2759 {
2760         struct stable_node *stable_node, *next;
2761         struct rb_node *node;
2762         int nid;
2763
2764         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2765                 node = rb_first(root_stable_tree + nid);
2766                 while (node) {
2767                         stable_node = rb_entry(node, struct stable_node, node);
2768                         if (stable_node_chain_remove_range(stable_node,
2769                                                            start_pfn, end_pfn,
2770                                                            root_stable_tree +
2771                                                            nid))
2772                                 node = rb_first(root_stable_tree + nid);
2773                         else
2774                                 node = rb_next(node);
2775                         cond_resched();
2776                 }
2777         }
2778         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2779                 if (stable_node->kpfn >= start_pfn &&
2780                     stable_node->kpfn < end_pfn)
2781                         remove_node_from_stable_tree(stable_node);
2782                 cond_resched();
2783         }
2784 }
2785
2786 static int ksm_memory_callback(struct notifier_block *self,
2787                                unsigned long action, void *arg)
2788 {
2789         struct memory_notify *mn = arg;
2790
2791         switch (action) {
2792         case MEM_GOING_OFFLINE:
2793                 /*
2794                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2795                  * and remove_all_stable_nodes() while memory is going offline:
2796                  * it is unsafe for them to touch the stable tree at this time.
2797                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2798                  * which do not need the ksm_thread_mutex are all safe.
2799                  */
2800                 mutex_lock(&ksm_thread_mutex);
2801                 ksm_run |= KSM_RUN_OFFLINE;
2802                 mutex_unlock(&ksm_thread_mutex);
2803                 break;
2804
2805         case MEM_OFFLINE:
2806                 /*
2807                  * Most of the work is done by page migration; but there might
2808                  * be a few stable_nodes left over, still pointing to struct
2809                  * pages which have been offlined: prune those from the tree,
2810                  * otherwise get_ksm_page() might later try to access a
2811                  * non-existent struct page.
2812                  */
2813                 ksm_check_stable_tree(mn->start_pfn,
2814                                       mn->start_pfn + mn->nr_pages);
2815                 /* fallthrough */
2816
2817         case MEM_CANCEL_OFFLINE:
2818                 mutex_lock(&ksm_thread_mutex);
2819                 ksm_run &= ~KSM_RUN_OFFLINE;
2820                 mutex_unlock(&ksm_thread_mutex);
2821
2822                 smp_mb();       /* wake_up_bit advises this */
2823                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2824                 break;
2825         }
2826         return NOTIFY_OK;
2827 }
2828 #else
2829 static void wait_while_offlining(void)
2830 {
2831 }
2832 #endif /* CONFIG_MEMORY_HOTREMOVE */
2833
2834 #ifdef CONFIG_SYSFS
2835 /*
2836  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2837  */
2838
2839 #define KSM_ATTR_RO(_name) \
2840         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2841 #define KSM_ATTR(_name) \
2842         static struct kobj_attribute _name##_attr = \
2843                 __ATTR(_name, 0644, _name##_show, _name##_store)
2844
2845 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2846                                     struct kobj_attribute *attr, char *buf)
2847 {
2848         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2849 }
2850
2851 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2852                                      struct kobj_attribute *attr,
2853                                      const char *buf, size_t count)
2854 {
2855         unsigned long msecs;
2856         int err;
2857
2858         err = kstrtoul(buf, 10, &msecs);
2859         if (err || msecs > UINT_MAX)
2860                 return -EINVAL;
2861
2862         ksm_thread_sleep_millisecs = msecs;
2863         wake_up_interruptible(&ksm_iter_wait);
2864
2865         return count;
2866 }
2867 KSM_ATTR(sleep_millisecs);
2868
2869 static ssize_t pages_to_scan_show(struct kobject *kobj,
2870                                   struct kobj_attribute *attr, char *buf)
2871 {
2872         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2873 }
2874
2875 static ssize_t pages_to_scan_store(struct kobject *kobj,
2876                                    struct kobj_attribute *attr,
2877                                    const char *buf, size_t count)
2878 {
2879         int err;
2880         unsigned long nr_pages;
2881
2882         err = kstrtoul(buf, 10, &nr_pages);
2883         if (err || nr_pages > UINT_MAX)
2884                 return -EINVAL;
2885
2886         ksm_thread_pages_to_scan = nr_pages;
2887
2888         return count;
2889 }
2890 KSM_ATTR(pages_to_scan);
2891
2892 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2893                         char *buf)
2894 {
2895         return sprintf(buf, "%lu\n", ksm_run);
2896 }
2897
2898 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2899                          const char *buf, size_t count)
2900 {
2901         int err;
2902         unsigned long flags;
2903
2904         err = kstrtoul(buf, 10, &flags);
2905         if (err || flags > UINT_MAX)
2906                 return -EINVAL;
2907         if (flags > KSM_RUN_UNMERGE)
2908                 return -EINVAL;
2909
2910         /*
2911          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2912          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2913          * breaking COW to free the pages_shared (but leaves mm_slots
2914          * on the list for when ksmd may be set running again).
2915          */
2916
2917         mutex_lock(&ksm_thread_mutex);
2918         wait_while_offlining();
2919         if (ksm_run != flags) {
2920                 ksm_run = flags;
2921                 if (flags & KSM_RUN_UNMERGE) {
2922                         set_current_oom_origin();
2923                         err = unmerge_and_remove_all_rmap_items();
2924                         clear_current_oom_origin();
2925                         if (err) {
2926                                 ksm_run = KSM_RUN_STOP;
2927                                 count = err;
2928                         }
2929                 }
2930         }
2931         mutex_unlock(&ksm_thread_mutex);
2932
2933         if (flags & KSM_RUN_MERGE)
2934                 wake_up_interruptible(&ksm_thread_wait);
2935
2936         return count;
2937 }
2938 KSM_ATTR(run);
2939
2940 #ifdef CONFIG_NUMA
2941 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2942                                 struct kobj_attribute *attr, char *buf)
2943 {
2944         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2945 }
2946
2947 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2948                                    struct kobj_attribute *attr,
2949                                    const char *buf, size_t count)
2950 {
2951         int err;
2952         unsigned long knob;
2953
2954         err = kstrtoul(buf, 10, &knob);
2955         if (err)
2956                 return err;
2957         if (knob > 1)
2958                 return -EINVAL;
2959
2960         mutex_lock(&ksm_thread_mutex);
2961         wait_while_offlining();
2962         if (ksm_merge_across_nodes != knob) {
2963                 if (ksm_pages_shared || remove_all_stable_nodes())
2964                         err = -EBUSY;
2965                 else if (root_stable_tree == one_stable_tree) {
2966                         struct rb_root *buf;
2967                         /*
2968                          * This is the first time that we switch away from the
2969                          * default of merging across nodes: must now allocate
2970                          * a buffer to hold as many roots as may be needed.
2971                          * Allocate stable and unstable together:
2972                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2973                          */
2974                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2975                                       GFP_KERNEL);
2976                         /* Let us assume that RB_ROOT is NULL is zero */
2977                         if (!buf)
2978                                 err = -ENOMEM;
2979                         else {
2980                                 root_stable_tree = buf;
2981                                 root_unstable_tree = buf + nr_node_ids;
2982                                 /* Stable tree is empty but not the unstable */
2983                                 root_unstable_tree[0] = one_unstable_tree[0];
2984                         }
2985                 }
2986                 if (!err) {
2987                         ksm_merge_across_nodes = knob;
2988                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2989                 }
2990         }
2991         mutex_unlock(&ksm_thread_mutex);
2992
2993         return err ? err : count;
2994 }
2995 KSM_ATTR(merge_across_nodes);
2996 #endif
2997
2998 static ssize_t use_zero_pages_show(struct kobject *kobj,
2999                                 struct kobj_attribute *attr, char *buf)
3000 {
3001         return sprintf(buf, "%u\n", ksm_use_zero_pages);
3002 }
3003 static ssize_t use_zero_pages_store(struct kobject *kobj,
3004                                    struct kobj_attribute *attr,
3005                                    const char *buf, size_t count)
3006 {
3007         int err;
3008         bool value;
3009
3010         err = kstrtobool(buf, &value);
3011         if (err)
3012                 return -EINVAL;
3013
3014         ksm_use_zero_pages = value;
3015
3016         return count;
3017 }
3018 KSM_ATTR(use_zero_pages);
3019
3020 static ssize_t max_page_sharing_show(struct kobject *kobj,
3021                                      struct kobj_attribute *attr, char *buf)
3022 {
3023         return sprintf(buf, "%u\n", ksm_max_page_sharing);
3024 }
3025
3026 static ssize_t max_page_sharing_store(struct kobject *kobj,
3027                                       struct kobj_attribute *attr,
3028                                       const char *buf, size_t count)
3029 {
3030         int err;
3031         int knob;
3032
3033         err = kstrtoint(buf, 10, &knob);
3034         if (err)
3035                 return err;
3036         /*
3037          * When a KSM page is created it is shared by 2 mappings. This
3038          * being a signed comparison, it implicitly verifies it's not
3039          * negative.
3040          */
3041         if (knob < 2)
3042                 return -EINVAL;
3043
3044         if (READ_ONCE(ksm_max_page_sharing) == knob)
3045                 return count;
3046
3047         mutex_lock(&ksm_thread_mutex);
3048         wait_while_offlining();
3049         if (ksm_max_page_sharing != knob) {
3050                 if (ksm_pages_shared || remove_all_stable_nodes())
3051                         err = -EBUSY;
3052                 else
3053                         ksm_max_page_sharing = knob;
3054         }
3055         mutex_unlock(&ksm_thread_mutex);
3056
3057         return err ? err : count;
3058 }
3059 KSM_ATTR(max_page_sharing);
3060
3061 static ssize_t pages_shared_show(struct kobject *kobj,
3062                                  struct kobj_attribute *attr, char *buf)
3063 {
3064         return sprintf(buf, "%lu\n", ksm_pages_shared);
3065 }
3066 KSM_ATTR_RO(pages_shared);
3067
3068 static ssize_t pages_sharing_show(struct kobject *kobj,
3069                                   struct kobj_attribute *attr, char *buf)
3070 {
3071         return sprintf(buf, "%lu\n", ksm_pages_sharing);
3072 }
3073 KSM_ATTR_RO(pages_sharing);
3074
3075 static ssize_t pages_unshared_show(struct kobject *kobj,
3076                                    struct kobj_attribute *attr, char *buf)
3077 {
3078         return sprintf(buf, "%lu\n", ksm_pages_unshared);
3079 }
3080 KSM_ATTR_RO(pages_unshared);
3081
3082 static ssize_t pages_volatile_show(struct kobject *kobj,
3083                                    struct kobj_attribute *attr, char *buf)
3084 {
3085         long ksm_pages_volatile;
3086
3087         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3088                                 - ksm_pages_sharing - ksm_pages_unshared;
3089         /*
3090          * It was not worth any locking to calculate that statistic,
3091          * but it might therefore sometimes be negative: conceal that.
3092          */
3093         if (ksm_pages_volatile < 0)
3094                 ksm_pages_volatile = 0;
3095         return sprintf(buf, "%ld\n", ksm_pages_volatile);
3096 }
3097 KSM_ATTR_RO(pages_volatile);
3098
3099 static ssize_t stable_node_dups_show(struct kobject *kobj,
3100                                      struct kobj_attribute *attr, char *buf)
3101 {
3102         return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3103 }
3104 KSM_ATTR_RO(stable_node_dups);
3105
3106 static ssize_t stable_node_chains_show(struct kobject *kobj,
3107                                        struct kobj_attribute *attr, char *buf)
3108 {
3109         return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3110 }
3111 KSM_ATTR_RO(stable_node_chains);
3112
3113 static ssize_t
3114 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3115                                         struct kobj_attribute *attr,
3116                                         char *buf)
3117 {
3118         return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3119 }
3120
3121 static ssize_t
3122 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3123                                          struct kobj_attribute *attr,
3124                                          const char *buf, size_t count)
3125 {
3126         unsigned long msecs;
3127         int err;
3128
3129         err = kstrtoul(buf, 10, &msecs);
3130         if (err || msecs > UINT_MAX)
3131                 return -EINVAL;
3132
3133         ksm_stable_node_chains_prune_millisecs = msecs;
3134
3135         return count;
3136 }
3137 KSM_ATTR(stable_node_chains_prune_millisecs);
3138
3139 static ssize_t full_scans_show(struct kobject *kobj,
3140                                struct kobj_attribute *attr, char *buf)
3141 {
3142         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3143 }
3144 KSM_ATTR_RO(full_scans);
3145
3146 static struct attribute *ksm_attrs[] = {
3147         &sleep_millisecs_attr.attr,
3148         &pages_to_scan_attr.attr,
3149         &run_attr.attr,
3150         &pages_shared_attr.attr,
3151         &pages_sharing_attr.attr,
3152         &pages_unshared_attr.attr,
3153         &pages_volatile_attr.attr,
3154         &full_scans_attr.attr,
3155 #ifdef CONFIG_NUMA
3156         &merge_across_nodes_attr.attr,
3157 #endif
3158         &max_page_sharing_attr.attr,
3159         &stable_node_chains_attr.attr,
3160         &stable_node_dups_attr.attr,
3161         &stable_node_chains_prune_millisecs_attr.attr,
3162         &use_zero_pages_attr.attr,
3163         NULL,
3164 };
3165
3166 static const struct attribute_group ksm_attr_group = {
3167         .attrs = ksm_attrs,
3168         .name = "ksm",
3169 };
3170 #endif /* CONFIG_SYSFS */
3171
3172 static int __init ksm_init(void)
3173 {
3174         struct task_struct *ksm_thread;
3175         int err;
3176
3177         /* The correct value depends on page size and endianness */
3178         zero_checksum = calc_checksum(ZERO_PAGE(0));
3179         /* Default to false for backwards compatibility */
3180         ksm_use_zero_pages = false;
3181
3182         err = ksm_slab_init();
3183         if (err)
3184                 goto out;
3185
3186         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3187         if (IS_ERR(ksm_thread)) {
3188                 pr_err("ksm: creating kthread failed\n");
3189                 err = PTR_ERR(ksm_thread);
3190                 goto out_free;
3191         }
3192
3193 #ifdef CONFIG_SYSFS
3194         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3195         if (err) {
3196                 pr_err("ksm: register sysfs failed\n");
3197                 kthread_stop(ksm_thread);
3198                 goto out_free;
3199         }
3200 #else
3201         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3202
3203 #endif /* CONFIG_SYSFS */
3204
3205 #ifdef CONFIG_MEMORY_HOTREMOVE
3206         /* There is no significance to this priority 100 */
3207         hotplug_memory_notifier(ksm_memory_callback, 100);
3208 #endif
3209         return 0;
3210
3211 out_free:
3212         ksm_slab_free();
3213 out:
3214         return err;
3215 }
3216 subsys_initcall(ksm_init);