Merge tag 'x86-urgent-2022-08-13' of git://git.kernel.org/pub/scm/linux/kernel/git...
[platform/kernel/linux-starfive.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/mm_inline.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/xxhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
42
43 #include <asm/tlbflush.h>
44 #include "internal.h"
45
46 #ifdef CONFIG_NUMA
47 #define NUMA(x)         (x)
48 #define DO_NUMA(x)      do { (x); } while (0)
49 #else
50 #define NUMA(x)         (0)
51 #define DO_NUMA(x)      do { } while (0)
52 #endif
53
54 /**
55  * DOC: Overview
56  *
57  * A few notes about the KSM scanning process,
58  * to make it easier to understand the data structures below:
59  *
60  * In order to reduce excessive scanning, KSM sorts the memory pages by their
61  * contents into a data structure that holds pointers to the pages' locations.
62  *
63  * Since the contents of the pages may change at any moment, KSM cannot just
64  * insert the pages into a normal sorted tree and expect it to find anything.
65  * Therefore KSM uses two data structures - the stable and the unstable tree.
66  *
67  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68  * by their contents.  Because each such page is write-protected, searching on
69  * this tree is fully assured to be working (except when pages are unmapped),
70  * and therefore this tree is called the stable tree.
71  *
72  * The stable tree node includes information required for reverse
73  * mapping from a KSM page to virtual addresses that map this page.
74  *
75  * In order to avoid large latencies of the rmap walks on KSM pages,
76  * KSM maintains two types of nodes in the stable tree:
77  *
78  * * the regular nodes that keep the reverse mapping structures in a
79  *   linked list
80  * * the "chains" that link nodes ("dups") that represent the same
81  *   write protected memory content, but each "dup" corresponds to a
82  *   different KSM page copy of that content
83  *
84  * Internally, the regular nodes, "dups" and "chains" are represented
85  * using the same struct stable_node structure.
86  *
87  * In addition to the stable tree, KSM uses a second data structure called the
88  * unstable tree: this tree holds pointers to pages which have been found to
89  * be "unchanged for a period of time".  The unstable tree sorts these pages
90  * by their contents, but since they are not write-protected, KSM cannot rely
91  * upon the unstable tree to work correctly - the unstable tree is liable to
92  * be corrupted as its contents are modified, and so it is called unstable.
93  *
94  * KSM solves this problem by several techniques:
95  *
96  * 1) The unstable tree is flushed every time KSM completes scanning all
97  *    memory areas, and then the tree is rebuilt again from the beginning.
98  * 2) KSM will only insert into the unstable tree, pages whose hash value
99  *    has not changed since the previous scan of all memory areas.
100  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
101  *    colors of the nodes and not on their contents, assuring that even when
102  *    the tree gets "corrupted" it won't get out of balance, so scanning time
103  *    remains the same (also, searching and inserting nodes in an rbtree uses
104  *    the same algorithm, so we have no overhead when we flush and rebuild).
105  * 4) KSM never flushes the stable tree, which means that even if it were to
106  *    take 10 attempts to find a page in the unstable tree, once it is found,
107  *    it is secured in the stable tree.  (When we scan a new page, we first
108  *    compare it against the stable tree, and then against the unstable tree.)
109  *
110  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
111  * stable trees and multiple unstable trees: one of each for each NUMA node.
112  */
113
114 /**
115  * struct mm_slot - ksm information per mm that is being scanned
116  * @link: link to the mm_slots hash list
117  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
118  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
119  * @mm: the mm that this information is valid for
120  */
121 struct mm_slot {
122         struct hlist_node link;
123         struct list_head mm_list;
124         struct rmap_item *rmap_list;
125         struct mm_struct *mm;
126 };
127
128 /**
129  * struct ksm_scan - cursor for scanning
130  * @mm_slot: the current mm_slot we are scanning
131  * @address: the next address inside that to be scanned
132  * @rmap_list: link to the next rmap to be scanned in the rmap_list
133  * @seqnr: count of completed full scans (needed when removing unstable node)
134  *
135  * There is only the one ksm_scan instance of this cursor structure.
136  */
137 struct ksm_scan {
138         struct mm_slot *mm_slot;
139         unsigned long address;
140         struct rmap_item **rmap_list;
141         unsigned long seqnr;
142 };
143
144 /**
145  * struct stable_node - node of the stable rbtree
146  * @node: rb node of this ksm page in the stable tree
147  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
148  * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
149  * @list: linked into migrate_nodes, pending placement in the proper node tree
150  * @hlist: hlist head of rmap_items using this ksm page
151  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
152  * @chain_prune_time: time of the last full garbage collection
153  * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
154  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
155  */
156 struct stable_node {
157         union {
158                 struct rb_node node;    /* when node of stable tree */
159                 struct {                /* when listed for migration */
160                         struct list_head *head;
161                         struct {
162                                 struct hlist_node hlist_dup;
163                                 struct list_head list;
164                         };
165                 };
166         };
167         struct hlist_head hlist;
168         union {
169                 unsigned long kpfn;
170                 unsigned long chain_prune_time;
171         };
172         /*
173          * STABLE_NODE_CHAIN can be any negative number in
174          * rmap_hlist_len negative range, but better not -1 to be able
175          * to reliably detect underflows.
176          */
177 #define STABLE_NODE_CHAIN -1024
178         int rmap_hlist_len;
179 #ifdef CONFIG_NUMA
180         int nid;
181 #endif
182 };
183
184 /**
185  * struct rmap_item - reverse mapping item for virtual addresses
186  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
187  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
188  * @nid: NUMA node id of unstable tree in which linked (may not match page)
189  * @mm: the memory structure this rmap_item is pointing into
190  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
191  * @oldchecksum: previous checksum of the page at that virtual address
192  * @node: rb node of this rmap_item in the unstable tree
193  * @head: pointer to stable_node heading this list in the stable tree
194  * @hlist: link into hlist of rmap_items hanging off that stable_node
195  */
196 struct rmap_item {
197         struct rmap_item *rmap_list;
198         union {
199                 struct anon_vma *anon_vma;      /* when stable */
200 #ifdef CONFIG_NUMA
201                 int nid;                /* when node of unstable tree */
202 #endif
203         };
204         struct mm_struct *mm;
205         unsigned long address;          /* + low bits used for flags below */
206         unsigned int oldchecksum;       /* when unstable */
207         union {
208                 struct rb_node node;    /* when node of unstable tree */
209                 struct {                /* when listed from stable tree */
210                         struct stable_node *head;
211                         struct hlist_node hlist;
212                 };
213         };
214 };
215
216 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
217 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
218 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
219
220 /* The stable and unstable tree heads */
221 static struct rb_root one_stable_tree[1] = { RB_ROOT };
222 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
223 static struct rb_root *root_stable_tree = one_stable_tree;
224 static struct rb_root *root_unstable_tree = one_unstable_tree;
225
226 /* Recently migrated nodes of stable tree, pending proper placement */
227 static LIST_HEAD(migrate_nodes);
228 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
229
230 #define MM_SLOTS_HASH_BITS 10
231 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
232
233 static struct mm_slot ksm_mm_head = {
234         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
235 };
236 static struct ksm_scan ksm_scan = {
237         .mm_slot = &ksm_mm_head,
238 };
239
240 static struct kmem_cache *rmap_item_cache;
241 static struct kmem_cache *stable_node_cache;
242 static struct kmem_cache *mm_slot_cache;
243
244 /* The number of nodes in the stable tree */
245 static unsigned long ksm_pages_shared;
246
247 /* The number of page slots additionally sharing those nodes */
248 static unsigned long ksm_pages_sharing;
249
250 /* The number of nodes in the unstable tree */
251 static unsigned long ksm_pages_unshared;
252
253 /* The number of rmap_items in use: to calculate pages_volatile */
254 static unsigned long ksm_rmap_items;
255
256 /* The number of stable_node chains */
257 static unsigned long ksm_stable_node_chains;
258
259 /* The number of stable_node dups linked to the stable_node chains */
260 static unsigned long ksm_stable_node_dups;
261
262 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
263 static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
264
265 /* Maximum number of page slots sharing a stable node */
266 static int ksm_max_page_sharing = 256;
267
268 /* Number of pages ksmd should scan in one batch */
269 static unsigned int ksm_thread_pages_to_scan = 100;
270
271 /* Milliseconds ksmd should sleep between batches */
272 static unsigned int ksm_thread_sleep_millisecs = 20;
273
274 /* Checksum of an empty (zeroed) page */
275 static unsigned int zero_checksum __read_mostly;
276
277 /* Whether to merge empty (zeroed) pages with actual zero pages */
278 static bool ksm_use_zero_pages __read_mostly;
279
280 #ifdef CONFIG_NUMA
281 /* Zeroed when merging across nodes is not allowed */
282 static unsigned int ksm_merge_across_nodes = 1;
283 static int ksm_nr_node_ids = 1;
284 #else
285 #define ksm_merge_across_nodes  1U
286 #define ksm_nr_node_ids         1
287 #endif
288
289 #define KSM_RUN_STOP    0
290 #define KSM_RUN_MERGE   1
291 #define KSM_RUN_UNMERGE 2
292 #define KSM_RUN_OFFLINE 4
293 static unsigned long ksm_run = KSM_RUN_STOP;
294 static void wait_while_offlining(void);
295
296 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
297 static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
298 static DEFINE_MUTEX(ksm_thread_mutex);
299 static DEFINE_SPINLOCK(ksm_mmlist_lock);
300
301 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
302                 sizeof(struct __struct), __alignof__(struct __struct),\
303                 (__flags), NULL)
304
305 static int __init ksm_slab_init(void)
306 {
307         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
308         if (!rmap_item_cache)
309                 goto out;
310
311         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
312         if (!stable_node_cache)
313                 goto out_free1;
314
315         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
316         if (!mm_slot_cache)
317                 goto out_free2;
318
319         return 0;
320
321 out_free2:
322         kmem_cache_destroy(stable_node_cache);
323 out_free1:
324         kmem_cache_destroy(rmap_item_cache);
325 out:
326         return -ENOMEM;
327 }
328
329 static void __init ksm_slab_free(void)
330 {
331         kmem_cache_destroy(mm_slot_cache);
332         kmem_cache_destroy(stable_node_cache);
333         kmem_cache_destroy(rmap_item_cache);
334         mm_slot_cache = NULL;
335 }
336
337 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
338 {
339         return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
340 }
341
342 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
343 {
344         return dup->head == STABLE_NODE_DUP_HEAD;
345 }
346
347 static inline void stable_node_chain_add_dup(struct stable_node *dup,
348                                              struct stable_node *chain)
349 {
350         VM_BUG_ON(is_stable_node_dup(dup));
351         dup->head = STABLE_NODE_DUP_HEAD;
352         VM_BUG_ON(!is_stable_node_chain(chain));
353         hlist_add_head(&dup->hlist_dup, &chain->hlist);
354         ksm_stable_node_dups++;
355 }
356
357 static inline void __stable_node_dup_del(struct stable_node *dup)
358 {
359         VM_BUG_ON(!is_stable_node_dup(dup));
360         hlist_del(&dup->hlist_dup);
361         ksm_stable_node_dups--;
362 }
363
364 static inline void stable_node_dup_del(struct stable_node *dup)
365 {
366         VM_BUG_ON(is_stable_node_chain(dup));
367         if (is_stable_node_dup(dup))
368                 __stable_node_dup_del(dup);
369         else
370                 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
371 #ifdef CONFIG_DEBUG_VM
372         dup->head = NULL;
373 #endif
374 }
375
376 static inline struct rmap_item *alloc_rmap_item(void)
377 {
378         struct rmap_item *rmap_item;
379
380         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
381                                                 __GFP_NORETRY | __GFP_NOWARN);
382         if (rmap_item)
383                 ksm_rmap_items++;
384         return rmap_item;
385 }
386
387 static inline void free_rmap_item(struct rmap_item *rmap_item)
388 {
389         ksm_rmap_items--;
390         rmap_item->mm = NULL;   /* debug safety */
391         kmem_cache_free(rmap_item_cache, rmap_item);
392 }
393
394 static inline struct stable_node *alloc_stable_node(void)
395 {
396         /*
397          * The allocation can take too long with GFP_KERNEL when memory is under
398          * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
399          * grants access to memory reserves, helping to avoid this problem.
400          */
401         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
402 }
403
404 static inline void free_stable_node(struct stable_node *stable_node)
405 {
406         VM_BUG_ON(stable_node->rmap_hlist_len &&
407                   !is_stable_node_chain(stable_node));
408         kmem_cache_free(stable_node_cache, stable_node);
409 }
410
411 static inline struct mm_slot *alloc_mm_slot(void)
412 {
413         if (!mm_slot_cache)     /* initialization failed */
414                 return NULL;
415         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
416 }
417
418 static inline void free_mm_slot(struct mm_slot *mm_slot)
419 {
420         kmem_cache_free(mm_slot_cache, mm_slot);
421 }
422
423 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
424 {
425         struct mm_slot *slot;
426
427         hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
428                 if (slot->mm == mm)
429                         return slot;
430
431         return NULL;
432 }
433
434 static void insert_to_mm_slots_hash(struct mm_struct *mm,
435                                     struct mm_slot *mm_slot)
436 {
437         mm_slot->mm = mm;
438         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
439 }
440
441 /*
442  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
443  * page tables after it has passed through ksm_exit() - which, if necessary,
444  * takes mmap_lock briefly to serialize against them.  ksm_exit() does not set
445  * a special flag: they can just back out as soon as mm_users goes to zero.
446  * ksm_test_exit() is used throughout to make this test for exit: in some
447  * places for correctness, in some places just to avoid unnecessary work.
448  */
449 static inline bool ksm_test_exit(struct mm_struct *mm)
450 {
451         return atomic_read(&mm->mm_users) == 0;
452 }
453
454 /*
455  * We use break_ksm to break COW on a ksm page: it's a stripped down
456  *
457  *      if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1)
458  *              put_page(page);
459  *
460  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
461  * in case the application has unmapped and remapped mm,addr meanwhile.
462  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
463  * mmap of /dev/mem, where we would not want to touch it.
464  *
465  * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
466  * of the process that owns 'vma'.  We also do not want to enforce
467  * protection keys here anyway.
468  */
469 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
470 {
471         struct page *page;
472         vm_fault_t ret = 0;
473
474         do {
475                 cond_resched();
476                 page = follow_page(vma, addr,
477                                 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
478                 if (IS_ERR_OR_NULL(page) || is_zone_device_page(page))
479                         break;
480                 if (PageKsm(page))
481                         ret = handle_mm_fault(vma, addr,
482                                               FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE,
483                                               NULL);
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 = vma_lookup(mm, addr);
526         if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
527                 return NULL;
528         return vma;
529 }
530
531 static void break_cow(struct rmap_item *rmap_item)
532 {
533         struct mm_struct *mm = rmap_item->mm;
534         unsigned long addr = rmap_item->address;
535         struct vm_area_struct *vma;
536
537         /*
538          * It is not an accident that whenever we want to break COW
539          * to undo, we also need to drop a reference to the anon_vma.
540          */
541         put_anon_vma(rmap_item->anon_vma);
542
543         mmap_read_lock(mm);
544         vma = find_mergeable_vma(mm, addr);
545         if (vma)
546                 break_ksm(vma, addr);
547         mmap_read_unlock(mm);
548 }
549
550 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
551 {
552         struct mm_struct *mm = rmap_item->mm;
553         unsigned long addr = rmap_item->address;
554         struct vm_area_struct *vma;
555         struct page *page;
556
557         mmap_read_lock(mm);
558         vma = find_mergeable_vma(mm, addr);
559         if (!vma)
560                 goto out;
561
562         page = follow_page(vma, addr, FOLL_GET);
563         if (IS_ERR_OR_NULL(page) || is_zone_device_page(page))
564                 goto out;
565         if (PageAnon(page)) {
566                 flush_anon_page(vma, page, addr);
567                 flush_dcache_page(page);
568         } else {
569                 put_page(page);
570 out:
571                 page = NULL;
572         }
573         mmap_read_unlock(mm);
574         return page;
575 }
576
577 /*
578  * This helper is used for getting right index into array of tree roots.
579  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
580  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
581  * every node has its own stable and unstable tree.
582  */
583 static inline int get_kpfn_nid(unsigned long kpfn)
584 {
585         return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
586 }
587
588 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
589                                                    struct rb_root *root)
590 {
591         struct stable_node *chain = alloc_stable_node();
592         VM_BUG_ON(is_stable_node_chain(dup));
593         if (likely(chain)) {
594                 INIT_HLIST_HEAD(&chain->hlist);
595                 chain->chain_prune_time = jiffies;
596                 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
597 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
598                 chain->nid = NUMA_NO_NODE; /* debug */
599 #endif
600                 ksm_stable_node_chains++;
601
602                 /*
603                  * Put the stable node chain in the first dimension of
604                  * the stable tree and at the same time remove the old
605                  * stable node.
606                  */
607                 rb_replace_node(&dup->node, &chain->node, root);
608
609                 /*
610                  * Move the old stable node to the second dimension
611                  * queued in the hlist_dup. The invariant is that all
612                  * dup stable_nodes in the chain->hlist point to pages
613                  * that are write protected and have the exact same
614                  * content.
615                  */
616                 stable_node_chain_add_dup(dup, chain);
617         }
618         return chain;
619 }
620
621 static inline void free_stable_node_chain(struct stable_node *chain,
622                                           struct rb_root *root)
623 {
624         rb_erase(&chain->node, root);
625         free_stable_node(chain);
626         ksm_stable_node_chains--;
627 }
628
629 static void remove_node_from_stable_tree(struct stable_node *stable_node)
630 {
631         struct rmap_item *rmap_item;
632
633         /* check it's not STABLE_NODE_CHAIN or negative */
634         BUG_ON(stable_node->rmap_hlist_len < 0);
635
636         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
637                 if (rmap_item->hlist.next)
638                         ksm_pages_sharing--;
639                 else
640                         ksm_pages_shared--;
641
642                 rmap_item->mm->ksm_merging_pages--;
643
644                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
645                 stable_node->rmap_hlist_len--;
646                 put_anon_vma(rmap_item->anon_vma);
647                 rmap_item->address &= PAGE_MASK;
648                 cond_resched();
649         }
650
651         /*
652          * We need the second aligned pointer of the migrate_nodes
653          * list_head to stay clear from the rb_parent_color union
654          * (aligned and different than any node) and also different
655          * from &migrate_nodes. This will verify that future list.h changes
656          * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
657          */
658         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
659         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
660
661         if (stable_node->head == &migrate_nodes)
662                 list_del(&stable_node->list);
663         else
664                 stable_node_dup_del(stable_node);
665         free_stable_node(stable_node);
666 }
667
668 enum get_ksm_page_flags {
669         GET_KSM_PAGE_NOLOCK,
670         GET_KSM_PAGE_LOCK,
671         GET_KSM_PAGE_TRYLOCK
672 };
673
674 /*
675  * get_ksm_page: checks if the page indicated by the stable node
676  * is still its ksm page, despite having held no reference to it.
677  * In which case we can trust the content of the page, and it
678  * returns the gotten page; but if the page has now been zapped,
679  * remove the stale node from the stable tree and return NULL.
680  * But beware, the stable node's page might be being migrated.
681  *
682  * You would expect the stable_node to hold a reference to the ksm page.
683  * But if it increments the page's count, swapping out has to wait for
684  * ksmd to come around again before it can free the page, which may take
685  * seconds or even minutes: much too unresponsive.  So instead we use a
686  * "keyhole reference": access to the ksm page from the stable node peeps
687  * out through its keyhole to see if that page still holds the right key,
688  * pointing back to this stable node.  This relies on freeing a PageAnon
689  * page to reset its page->mapping to NULL, and relies on no other use of
690  * a page to put something that might look like our key in page->mapping.
691  * is on its way to being freed; but it is an anomaly to bear in mind.
692  */
693 static struct page *get_ksm_page(struct stable_node *stable_node,
694                                  enum get_ksm_page_flags flags)
695 {
696         struct page *page;
697         void *expected_mapping;
698         unsigned long kpfn;
699
700         expected_mapping = (void *)((unsigned long)stable_node |
701                                         PAGE_MAPPING_KSM);
702 again:
703         kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
704         page = pfn_to_page(kpfn);
705         if (READ_ONCE(page->mapping) != expected_mapping)
706                 goto stale;
707
708         /*
709          * We cannot do anything with the page while its refcount is 0.
710          * Usually 0 means free, or tail of a higher-order page: in which
711          * case this node is no longer referenced, and should be freed;
712          * however, it might mean that the page is under page_ref_freeze().
713          * The __remove_mapping() case is easy, again the node is now stale;
714          * the same is in reuse_ksm_page() case; but if page is swapcache
715          * in folio_migrate_mapping(), it might still be our page,
716          * in which case it's essential to keep the node.
717          */
718         while (!get_page_unless_zero(page)) {
719                 /*
720                  * Another check for page->mapping != expected_mapping would
721                  * work here too.  We have chosen the !PageSwapCache test to
722                  * optimize the common case, when the page is or is about to
723                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
724                  * in the ref_freeze section of __remove_mapping(); but Anon
725                  * page->mapping reset to NULL later, in free_pages_prepare().
726                  */
727                 if (!PageSwapCache(page))
728                         goto stale;
729                 cpu_relax();
730         }
731
732         if (READ_ONCE(page->mapping) != expected_mapping) {
733                 put_page(page);
734                 goto stale;
735         }
736
737         if (flags == GET_KSM_PAGE_TRYLOCK) {
738                 if (!trylock_page(page)) {
739                         put_page(page);
740                         return ERR_PTR(-EBUSY);
741                 }
742         } else if (flags == GET_KSM_PAGE_LOCK)
743                 lock_page(page);
744
745         if (flags != GET_KSM_PAGE_NOLOCK) {
746                 if (READ_ONCE(page->mapping) != expected_mapping) {
747                         unlock_page(page);
748                         put_page(page);
749                         goto stale;
750                 }
751         }
752         return page;
753
754 stale:
755         /*
756          * We come here from above when page->mapping or !PageSwapCache
757          * suggests that the node is stale; but it might be under migration.
758          * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(),
759          * before checking whether node->kpfn has been changed.
760          */
761         smp_rmb();
762         if (READ_ONCE(stable_node->kpfn) != kpfn)
763                 goto again;
764         remove_node_from_stable_tree(stable_node);
765         return NULL;
766 }
767
768 /*
769  * Removing rmap_item from stable or unstable tree.
770  * This function will clean the information from the stable/unstable tree.
771  */
772 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
773 {
774         if (rmap_item->address & STABLE_FLAG) {
775                 struct stable_node *stable_node;
776                 struct page *page;
777
778                 stable_node = rmap_item->head;
779                 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
780                 if (!page)
781                         goto out;
782
783                 hlist_del(&rmap_item->hlist);
784                 unlock_page(page);
785                 put_page(page);
786
787                 if (!hlist_empty(&stable_node->hlist))
788                         ksm_pages_sharing--;
789                 else
790                         ksm_pages_shared--;
791
792                 rmap_item->mm->ksm_merging_pages--;
793
794                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
795                 stable_node->rmap_hlist_len--;
796
797                 put_anon_vma(rmap_item->anon_vma);
798                 rmap_item->head = NULL;
799                 rmap_item->address &= PAGE_MASK;
800
801         } else if (rmap_item->address & UNSTABLE_FLAG) {
802                 unsigned char age;
803                 /*
804                  * Usually ksmd can and must skip the rb_erase, because
805                  * root_unstable_tree was already reset to RB_ROOT.
806                  * But be careful when an mm is exiting: do the rb_erase
807                  * if this rmap_item was inserted by this scan, rather
808                  * than left over from before.
809                  */
810                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
811                 BUG_ON(age > 1);
812                 if (!age)
813                         rb_erase(&rmap_item->node,
814                                  root_unstable_tree + NUMA(rmap_item->nid));
815                 ksm_pages_unshared--;
816                 rmap_item->address &= PAGE_MASK;
817         }
818 out:
819         cond_resched();         /* we're called from many long loops */
820 }
821
822 static void remove_trailing_rmap_items(struct rmap_item **rmap_list)
823 {
824         while (*rmap_list) {
825                 struct rmap_item *rmap_item = *rmap_list;
826                 *rmap_list = rmap_item->rmap_list;
827                 remove_rmap_item_from_tree(rmap_item);
828                 free_rmap_item(rmap_item);
829         }
830 }
831
832 /*
833  * Though it's very tempting to unmerge rmap_items from stable tree rather
834  * than check every pte of a given vma, the locking doesn't quite work for
835  * that - an rmap_item is assigned to the stable tree after inserting ksm
836  * page and upping mmap_lock.  Nor does it fit with the way we skip dup'ing
837  * rmap_items from parent to child at fork time (so as not to waste time
838  * if exit comes before the next scan reaches it).
839  *
840  * Similarly, although we'd like to remove rmap_items (so updating counts
841  * and freeing memory) when unmerging an area, it's easier to leave that
842  * to the next pass of ksmd - consider, for example, how ksmd might be
843  * in cmp_and_merge_page on one of the rmap_items we would be removing.
844  */
845 static int unmerge_ksm_pages(struct vm_area_struct *vma,
846                              unsigned long start, unsigned long end)
847 {
848         unsigned long addr;
849         int err = 0;
850
851         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
852                 if (ksm_test_exit(vma->vm_mm))
853                         break;
854                 if (signal_pending(current))
855                         err = -ERESTARTSYS;
856                 else
857                         err = break_ksm(vma, addr);
858         }
859         return err;
860 }
861
862 static inline struct stable_node *folio_stable_node(struct folio *folio)
863 {
864         return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL;
865 }
866
867 static inline struct stable_node *page_stable_node(struct page *page)
868 {
869         return folio_stable_node(page_folio(page));
870 }
871
872 static inline void set_page_stable_node(struct page *page,
873                                         struct stable_node *stable_node)
874 {
875         VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page);
876         page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
877 }
878
879 #ifdef CONFIG_SYSFS
880 /*
881  * Only called through the sysfs control interface:
882  */
883 static int remove_stable_node(struct stable_node *stable_node)
884 {
885         struct page *page;
886         int err;
887
888         page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
889         if (!page) {
890                 /*
891                  * get_ksm_page did remove_node_from_stable_tree itself.
892                  */
893                 return 0;
894         }
895
896         /*
897          * Page could be still mapped if this races with __mmput() running in
898          * between ksm_exit() and exit_mmap(). Just refuse to let
899          * merge_across_nodes/max_page_sharing be switched.
900          */
901         err = -EBUSY;
902         if (!page_mapped(page)) {
903                 /*
904                  * The stable node did not yet appear stale to get_ksm_page(),
905                  * since that allows for an unmapped ksm page to be recognized
906                  * right up until it is freed; but the node is safe to remove.
907                  * This page might be in a pagevec waiting to be freed,
908                  * or it might be PageSwapCache (perhaps under writeback),
909                  * or it might have been removed from swapcache a moment ago.
910                  */
911                 set_page_stable_node(page, NULL);
912                 remove_node_from_stable_tree(stable_node);
913                 err = 0;
914         }
915
916         unlock_page(page);
917         put_page(page);
918         return err;
919 }
920
921 static int remove_stable_node_chain(struct stable_node *stable_node,
922                                     struct rb_root *root)
923 {
924         struct stable_node *dup;
925         struct hlist_node *hlist_safe;
926
927         if (!is_stable_node_chain(stable_node)) {
928                 VM_BUG_ON(is_stable_node_dup(stable_node));
929                 if (remove_stable_node(stable_node))
930                         return true;
931                 else
932                         return false;
933         }
934
935         hlist_for_each_entry_safe(dup, hlist_safe,
936                                   &stable_node->hlist, hlist_dup) {
937                 VM_BUG_ON(!is_stable_node_dup(dup));
938                 if (remove_stable_node(dup))
939                         return true;
940         }
941         BUG_ON(!hlist_empty(&stable_node->hlist));
942         free_stable_node_chain(stable_node, root);
943         return false;
944 }
945
946 static int remove_all_stable_nodes(void)
947 {
948         struct stable_node *stable_node, *next;
949         int nid;
950         int err = 0;
951
952         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
953                 while (root_stable_tree[nid].rb_node) {
954                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
955                                                 struct stable_node, node);
956                         if (remove_stable_node_chain(stable_node,
957                                                      root_stable_tree + nid)) {
958                                 err = -EBUSY;
959                                 break;  /* proceed to next nid */
960                         }
961                         cond_resched();
962                 }
963         }
964         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
965                 if (remove_stable_node(stable_node))
966                         err = -EBUSY;
967                 cond_resched();
968         }
969         return err;
970 }
971
972 static int unmerge_and_remove_all_rmap_items(void)
973 {
974         struct mm_slot *mm_slot;
975         struct mm_struct *mm;
976         struct vm_area_struct *vma;
977         int err = 0;
978
979         spin_lock(&ksm_mmlist_lock);
980         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
981                                                 struct mm_slot, mm_list);
982         spin_unlock(&ksm_mmlist_lock);
983
984         for (mm_slot = ksm_scan.mm_slot;
985                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
986                 mm = mm_slot->mm;
987                 mmap_read_lock(mm);
988                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
989                         if (ksm_test_exit(mm))
990                                 break;
991                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
992                                 continue;
993                         err = unmerge_ksm_pages(vma,
994                                                 vma->vm_start, vma->vm_end);
995                         if (err)
996                                 goto error;
997                 }
998
999                 remove_trailing_rmap_items(&mm_slot->rmap_list);
1000                 mmap_read_unlock(mm);
1001
1002                 spin_lock(&ksm_mmlist_lock);
1003                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
1004                                                 struct mm_slot, mm_list);
1005                 if (ksm_test_exit(mm)) {
1006                         hash_del(&mm_slot->link);
1007                         list_del(&mm_slot->mm_list);
1008                         spin_unlock(&ksm_mmlist_lock);
1009
1010                         free_mm_slot(mm_slot);
1011                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1012                         mmdrop(mm);
1013                 } else
1014                         spin_unlock(&ksm_mmlist_lock);
1015         }
1016
1017         /* Clean up stable nodes, but don't worry if some are still busy */
1018         remove_all_stable_nodes();
1019         ksm_scan.seqnr = 0;
1020         return 0;
1021
1022 error:
1023         mmap_read_unlock(mm);
1024         spin_lock(&ksm_mmlist_lock);
1025         ksm_scan.mm_slot = &ksm_mm_head;
1026         spin_unlock(&ksm_mmlist_lock);
1027         return err;
1028 }
1029 #endif /* CONFIG_SYSFS */
1030
1031 static u32 calc_checksum(struct page *page)
1032 {
1033         u32 checksum;
1034         void *addr = kmap_atomic(page);
1035         checksum = xxhash(addr, PAGE_SIZE, 0);
1036         kunmap_atomic(addr);
1037         return checksum;
1038 }
1039
1040 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1041                               pte_t *orig_pte)
1042 {
1043         struct mm_struct *mm = vma->vm_mm;
1044         DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0);
1045         int swapped;
1046         int err = -EFAULT;
1047         struct mmu_notifier_range range;
1048         bool anon_exclusive;
1049
1050         pvmw.address = page_address_in_vma(page, vma);
1051         if (pvmw.address == -EFAULT)
1052                 goto out;
1053
1054         BUG_ON(PageTransCompound(page));
1055
1056         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1057                                 pvmw.address,
1058                                 pvmw.address + PAGE_SIZE);
1059         mmu_notifier_invalidate_range_start(&range);
1060
1061         if (!page_vma_mapped_walk(&pvmw))
1062                 goto out_mn;
1063         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1064                 goto out_unlock;
1065
1066         anon_exclusive = PageAnonExclusive(page);
1067         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1068             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1069             anon_exclusive || mm_tlb_flush_pending(mm)) {
1070                 pte_t entry;
1071
1072                 swapped = PageSwapCache(page);
1073                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1074                 /*
1075                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1076                  * take any lock, therefore the check that we are going to make
1077                  * with the pagecount against the mapcount is racy and
1078                  * O_DIRECT can happen right after the check.
1079                  * So we clear the pte and flush the tlb before the check
1080                  * this assure us that no O_DIRECT can happen after the check
1081                  * or in the middle of the check.
1082                  *
1083                  * No need to notify as we are downgrading page table to read
1084                  * only not changing it to point to a new page.
1085                  *
1086                  * See Documentation/mm/mmu_notifier.rst
1087                  */
1088                 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1089                 /*
1090                  * Check that no O_DIRECT or similar I/O is in progress on the
1091                  * page
1092                  */
1093                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1094                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1095                         goto out_unlock;
1096                 }
1097
1098                 if (anon_exclusive && page_try_share_anon_rmap(page)) {
1099                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1100                         goto out_unlock;
1101                 }
1102
1103                 if (pte_dirty(entry))
1104                         set_page_dirty(page);
1105
1106                 if (pte_protnone(entry))
1107                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1108                 else
1109                         entry = pte_mkclean(pte_wrprotect(entry));
1110                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1111         }
1112         *orig_pte = *pvmw.pte;
1113         err = 0;
1114
1115 out_unlock:
1116         page_vma_mapped_walk_done(&pvmw);
1117 out_mn:
1118         mmu_notifier_invalidate_range_end(&range);
1119 out:
1120         return err;
1121 }
1122
1123 /**
1124  * replace_page - replace page in vma by new ksm page
1125  * @vma:      vma that holds the pte pointing to page
1126  * @page:     the page we are replacing by kpage
1127  * @kpage:    the ksm page we replace page by
1128  * @orig_pte: the original value of the pte
1129  *
1130  * Returns 0 on success, -EFAULT on failure.
1131  */
1132 static int replace_page(struct vm_area_struct *vma, struct page *page,
1133                         struct page *kpage, pte_t orig_pte)
1134 {
1135         struct mm_struct *mm = vma->vm_mm;
1136         pmd_t *pmd;
1137         pte_t *ptep;
1138         pte_t newpte;
1139         spinlock_t *ptl;
1140         unsigned long addr;
1141         int err = -EFAULT;
1142         struct mmu_notifier_range range;
1143
1144         addr = page_address_in_vma(page, vma);
1145         if (addr == -EFAULT)
1146                 goto out;
1147
1148         pmd = mm_find_pmd(mm, addr);
1149         if (!pmd)
1150                 goto out;
1151
1152         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1153                                 addr + PAGE_SIZE);
1154         mmu_notifier_invalidate_range_start(&range);
1155
1156         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1157         if (!pte_same(*ptep, orig_pte)) {
1158                 pte_unmap_unlock(ptep, ptl);
1159                 goto out_mn;
1160         }
1161         VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
1162         VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage);
1163
1164         /*
1165          * No need to check ksm_use_zero_pages here: we can only have a
1166          * zero_page here if ksm_use_zero_pages was enabled already.
1167          */
1168         if (!is_zero_pfn(page_to_pfn(kpage))) {
1169                 get_page(kpage);
1170                 page_add_anon_rmap(kpage, vma, addr, RMAP_NONE);
1171                 newpte = mk_pte(kpage, vma->vm_page_prot);
1172         } else {
1173                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1174                                                vma->vm_page_prot));
1175                 /*
1176                  * We're replacing an anonymous page with a zero page, which is
1177                  * not anonymous. We need to do proper accounting otherwise we
1178                  * will get wrong values in /proc, and a BUG message in dmesg
1179                  * when tearing down the mm.
1180                  */
1181                 dec_mm_counter(mm, MM_ANONPAGES);
1182         }
1183
1184         flush_cache_page(vma, addr, pte_pfn(*ptep));
1185         /*
1186          * No need to notify as we are replacing a read only page with another
1187          * read only page with the same content.
1188          *
1189          * See Documentation/mm/mmu_notifier.rst
1190          */
1191         ptep_clear_flush(vma, addr, ptep);
1192         set_pte_at_notify(mm, addr, ptep, newpte);
1193
1194         page_remove_rmap(page, vma, false);
1195         if (!page_mapped(page))
1196                 try_to_free_swap(page);
1197         put_page(page);
1198
1199         pte_unmap_unlock(ptep, ptl);
1200         err = 0;
1201 out_mn:
1202         mmu_notifier_invalidate_range_end(&range);
1203 out:
1204         return err;
1205 }
1206
1207 /*
1208  * try_to_merge_one_page - take two pages and merge them into one
1209  * @vma: the vma that holds the pte pointing to page
1210  * @page: the PageAnon page that we want to replace with kpage
1211  * @kpage: the PageKsm page that we want to map instead of page,
1212  *         or NULL the first time when we want to use page as kpage.
1213  *
1214  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1215  */
1216 static int try_to_merge_one_page(struct vm_area_struct *vma,
1217                                  struct page *page, struct page *kpage)
1218 {
1219         pte_t orig_pte = __pte(0);
1220         int err = -EFAULT;
1221
1222         if (page == kpage)                      /* ksm page forked */
1223                 return 0;
1224
1225         if (!PageAnon(page))
1226                 goto out;
1227
1228         /*
1229          * We need the page lock to read a stable PageSwapCache in
1230          * write_protect_page().  We use trylock_page() instead of
1231          * lock_page() because we don't want to wait here - we
1232          * prefer to continue scanning and merging different pages,
1233          * then come back to this page when it is unlocked.
1234          */
1235         if (!trylock_page(page))
1236                 goto out;
1237
1238         if (PageTransCompound(page)) {
1239                 if (split_huge_page(page))
1240                         goto out_unlock;
1241         }
1242
1243         /*
1244          * If this anonymous page is mapped only here, its pte may need
1245          * to be write-protected.  If it's mapped elsewhere, all of its
1246          * ptes are necessarily already write-protected.  But in either
1247          * case, we need to lock and check page_count is not raised.
1248          */
1249         if (write_protect_page(vma, page, &orig_pte) == 0) {
1250                 if (!kpage) {
1251                         /*
1252                          * While we hold page lock, upgrade page from
1253                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1254                          * stable_tree_insert() will update stable_node.
1255                          */
1256                         set_page_stable_node(page, NULL);
1257                         mark_page_accessed(page);
1258                         /*
1259                          * Page reclaim just frees a clean page with no dirty
1260                          * ptes: make sure that the ksm page would be swapped.
1261                          */
1262                         if (!PageDirty(page))
1263                                 SetPageDirty(page);
1264                         err = 0;
1265                 } else if (pages_identical(page, kpage))
1266                         err = replace_page(vma, page, kpage, orig_pte);
1267         }
1268
1269 out_unlock:
1270         unlock_page(page);
1271 out:
1272         return err;
1273 }
1274
1275 /*
1276  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1277  * but no new kernel page is allocated: kpage must already be a ksm page.
1278  *
1279  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1280  */
1281 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1282                                       struct page *page, struct page *kpage)
1283 {
1284         struct mm_struct *mm = rmap_item->mm;
1285         struct vm_area_struct *vma;
1286         int err = -EFAULT;
1287
1288         mmap_read_lock(mm);
1289         vma = find_mergeable_vma(mm, rmap_item->address);
1290         if (!vma)
1291                 goto out;
1292
1293         err = try_to_merge_one_page(vma, page, kpage);
1294         if (err)
1295                 goto out;
1296
1297         /* Unstable nid is in union with stable anon_vma: remove first */
1298         remove_rmap_item_from_tree(rmap_item);
1299
1300         /* Must get reference to anon_vma while still holding mmap_lock */
1301         rmap_item->anon_vma = vma->anon_vma;
1302         get_anon_vma(vma->anon_vma);
1303 out:
1304         mmap_read_unlock(mm);
1305         return err;
1306 }
1307
1308 /*
1309  * try_to_merge_two_pages - take two identical pages and prepare them
1310  * to be merged into one page.
1311  *
1312  * This function returns the kpage if we successfully merged two identical
1313  * pages into one ksm page, NULL otherwise.
1314  *
1315  * Note that this function upgrades page to ksm page: if one of the pages
1316  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1317  */
1318 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1319                                            struct page *page,
1320                                            struct rmap_item *tree_rmap_item,
1321                                            struct page *tree_page)
1322 {
1323         int err;
1324
1325         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1326         if (!err) {
1327                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1328                                                         tree_page, page);
1329                 /*
1330                  * If that fails, we have a ksm page with only one pte
1331                  * pointing to it: so break it.
1332                  */
1333                 if (err)
1334                         break_cow(rmap_item);
1335         }
1336         return err ? NULL : page;
1337 }
1338
1339 static __always_inline
1340 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1341 {
1342         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1343         /*
1344          * Check that at least one mapping still exists, otherwise
1345          * there's no much point to merge and share with this
1346          * stable_node, as the underlying tree_page of the other
1347          * sharer is going to be freed soon.
1348          */
1349         return stable_node->rmap_hlist_len &&
1350                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1351 }
1352
1353 static __always_inline
1354 bool is_page_sharing_candidate(struct stable_node *stable_node)
1355 {
1356         return __is_page_sharing_candidate(stable_node, 0);
1357 }
1358
1359 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1360                                     struct stable_node **_stable_node,
1361                                     struct rb_root *root,
1362                                     bool prune_stale_stable_nodes)
1363 {
1364         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1365         struct hlist_node *hlist_safe;
1366         struct page *_tree_page, *tree_page = NULL;
1367         int nr = 0;
1368         int found_rmap_hlist_len;
1369
1370         if (!prune_stale_stable_nodes ||
1371             time_before(jiffies, stable_node->chain_prune_time +
1372                         msecs_to_jiffies(
1373                                 ksm_stable_node_chains_prune_millisecs)))
1374                 prune_stale_stable_nodes = false;
1375         else
1376                 stable_node->chain_prune_time = jiffies;
1377
1378         hlist_for_each_entry_safe(dup, hlist_safe,
1379                                   &stable_node->hlist, hlist_dup) {
1380                 cond_resched();
1381                 /*
1382                  * We must walk all stable_node_dup to prune the stale
1383                  * stable nodes during lookup.
1384                  *
1385                  * get_ksm_page can drop the nodes from the
1386                  * stable_node->hlist if they point to freed pages
1387                  * (that's why we do a _safe walk). The "dup"
1388                  * stable_node parameter itself will be freed from
1389                  * under us if it returns NULL.
1390                  */
1391                 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1392                 if (!_tree_page)
1393                         continue;
1394                 nr += 1;
1395                 if (is_page_sharing_candidate(dup)) {
1396                         if (!found ||
1397                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1398                                 if (found)
1399                                         put_page(tree_page);
1400                                 found = dup;
1401                                 found_rmap_hlist_len = found->rmap_hlist_len;
1402                                 tree_page = _tree_page;
1403
1404                                 /* skip put_page for found dup */
1405                                 if (!prune_stale_stable_nodes)
1406                                         break;
1407                                 continue;
1408                         }
1409                 }
1410                 put_page(_tree_page);
1411         }
1412
1413         if (found) {
1414                 /*
1415                  * nr is counting all dups in the chain only if
1416                  * prune_stale_stable_nodes is true, otherwise we may
1417                  * break the loop at nr == 1 even if there are
1418                  * multiple entries.
1419                  */
1420                 if (prune_stale_stable_nodes && nr == 1) {
1421                         /*
1422                          * If there's not just one entry it would
1423                          * corrupt memory, better BUG_ON. In KSM
1424                          * context with no lock held it's not even
1425                          * fatal.
1426                          */
1427                         BUG_ON(stable_node->hlist.first->next);
1428
1429                         /*
1430                          * There's just one entry and it is below the
1431                          * deduplication limit so drop the chain.
1432                          */
1433                         rb_replace_node(&stable_node->node, &found->node,
1434                                         root);
1435                         free_stable_node(stable_node);
1436                         ksm_stable_node_chains--;
1437                         ksm_stable_node_dups--;
1438                         /*
1439                          * NOTE: the caller depends on the stable_node
1440                          * to be equal to stable_node_dup if the chain
1441                          * was collapsed.
1442                          */
1443                         *_stable_node = found;
1444                         /*
1445                          * Just for robustness, as stable_node is
1446                          * otherwise left as a stable pointer, the
1447                          * compiler shall optimize it away at build
1448                          * time.
1449                          */
1450                         stable_node = NULL;
1451                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1452                            __is_page_sharing_candidate(found, 1)) {
1453                         /*
1454                          * If the found stable_node dup can accept one
1455                          * more future merge (in addition to the one
1456                          * that is underway) and is not at the head of
1457                          * the chain, put it there so next search will
1458                          * be quicker in the !prune_stale_stable_nodes
1459                          * case.
1460                          *
1461                          * NOTE: it would be inaccurate to use nr > 1
1462                          * instead of checking the hlist.first pointer
1463                          * directly, because in the
1464                          * prune_stale_stable_nodes case "nr" isn't
1465                          * the position of the found dup in the chain,
1466                          * but the total number of dups in the chain.
1467                          */
1468                         hlist_del(&found->hlist_dup);
1469                         hlist_add_head(&found->hlist_dup,
1470                                        &stable_node->hlist);
1471                 }
1472         }
1473
1474         *_stable_node_dup = found;
1475         return tree_page;
1476 }
1477
1478 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1479                                                struct rb_root *root)
1480 {
1481         if (!is_stable_node_chain(stable_node))
1482                 return stable_node;
1483         if (hlist_empty(&stable_node->hlist)) {
1484                 free_stable_node_chain(stable_node, root);
1485                 return NULL;
1486         }
1487         return hlist_entry(stable_node->hlist.first,
1488                            typeof(*stable_node), hlist_dup);
1489 }
1490
1491 /*
1492  * Like for get_ksm_page, this function can free the *_stable_node and
1493  * *_stable_node_dup if the returned tree_page is NULL.
1494  *
1495  * It can also free and overwrite *_stable_node with the found
1496  * stable_node_dup if the chain is collapsed (in which case
1497  * *_stable_node will be equal to *_stable_node_dup like if the chain
1498  * never existed). It's up to the caller to verify tree_page is not
1499  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1500  *
1501  * *_stable_node_dup is really a second output parameter of this
1502  * function and will be overwritten in all cases, the caller doesn't
1503  * need to initialize it.
1504  */
1505 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1506                                         struct stable_node **_stable_node,
1507                                         struct rb_root *root,
1508                                         bool prune_stale_stable_nodes)
1509 {
1510         struct stable_node *stable_node = *_stable_node;
1511         if (!is_stable_node_chain(stable_node)) {
1512                 if (is_page_sharing_candidate(stable_node)) {
1513                         *_stable_node_dup = stable_node;
1514                         return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1515                 }
1516                 /*
1517                  * _stable_node_dup set to NULL means the stable_node
1518                  * reached the ksm_max_page_sharing limit.
1519                  */
1520                 *_stable_node_dup = NULL;
1521                 return NULL;
1522         }
1523         return stable_node_dup(_stable_node_dup, _stable_node, root,
1524                                prune_stale_stable_nodes);
1525 }
1526
1527 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1528                                                 struct stable_node **s_n,
1529                                                 struct rb_root *root)
1530 {
1531         return __stable_node_chain(s_n_d, s_n, root, true);
1532 }
1533
1534 static __always_inline struct page *chain(struct stable_node **s_n_d,
1535                                           struct stable_node *s_n,
1536                                           struct rb_root *root)
1537 {
1538         struct stable_node *old_stable_node = s_n;
1539         struct page *tree_page;
1540
1541         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1542         /* not pruning dups so s_n cannot have changed */
1543         VM_BUG_ON(s_n != old_stable_node);
1544         return tree_page;
1545 }
1546
1547 /*
1548  * stable_tree_search - search for page inside the stable tree
1549  *
1550  * This function checks if there is a page inside the stable tree
1551  * with identical content to the page that we are scanning right now.
1552  *
1553  * This function returns the stable tree node of identical content if found,
1554  * NULL otherwise.
1555  */
1556 static struct page *stable_tree_search(struct page *page)
1557 {
1558         int nid;
1559         struct rb_root *root;
1560         struct rb_node **new;
1561         struct rb_node *parent;
1562         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1563         struct stable_node *page_node;
1564
1565         page_node = page_stable_node(page);
1566         if (page_node && page_node->head != &migrate_nodes) {
1567                 /* ksm page forked */
1568                 get_page(page);
1569                 return page;
1570         }
1571
1572         nid = get_kpfn_nid(page_to_pfn(page));
1573         root = root_stable_tree + nid;
1574 again:
1575         new = &root->rb_node;
1576         parent = NULL;
1577
1578         while (*new) {
1579                 struct page *tree_page;
1580                 int ret;
1581
1582                 cond_resched();
1583                 stable_node = rb_entry(*new, struct stable_node, node);
1584                 stable_node_any = NULL;
1585                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1586                 /*
1587                  * NOTE: stable_node may have been freed by
1588                  * chain_prune() if the returned stable_node_dup is
1589                  * not NULL. stable_node_dup may have been inserted in
1590                  * the rbtree instead as a regular stable_node (in
1591                  * order to collapse the stable_node chain if a single
1592                  * stable_node dup was found in it). In such case the
1593                  * stable_node is overwritten by the callee to point
1594                  * to the stable_node_dup that was collapsed in the
1595                  * stable rbtree and stable_node will be equal to
1596                  * stable_node_dup like if the chain never existed.
1597                  */
1598                 if (!stable_node_dup) {
1599                         /*
1600                          * Either all stable_node dups were full in
1601                          * this stable_node chain, or this chain was
1602                          * empty and should be rb_erased.
1603                          */
1604                         stable_node_any = stable_node_dup_any(stable_node,
1605                                                               root);
1606                         if (!stable_node_any) {
1607                                 /* rb_erase just run */
1608                                 goto again;
1609                         }
1610                         /*
1611                          * Take any of the stable_node dups page of
1612                          * this stable_node chain to let the tree walk
1613                          * continue. All KSM pages belonging to the
1614                          * stable_node dups in a stable_node chain
1615                          * have the same content and they're
1616                          * write protected at all times. Any will work
1617                          * fine to continue the walk.
1618                          */
1619                         tree_page = get_ksm_page(stable_node_any,
1620                                                  GET_KSM_PAGE_NOLOCK);
1621                 }
1622                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1623                 if (!tree_page) {
1624                         /*
1625                          * If we walked over a stale stable_node,
1626                          * get_ksm_page() will call rb_erase() and it
1627                          * may rebalance the tree from under us. So
1628                          * restart the search from scratch. Returning
1629                          * NULL would be safe too, but we'd generate
1630                          * false negative insertions just because some
1631                          * stable_node was stale.
1632                          */
1633                         goto again;
1634                 }
1635
1636                 ret = memcmp_pages(page, tree_page);
1637                 put_page(tree_page);
1638
1639                 parent = *new;
1640                 if (ret < 0)
1641                         new = &parent->rb_left;
1642                 else if (ret > 0)
1643                         new = &parent->rb_right;
1644                 else {
1645                         if (page_node) {
1646                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1647                                 /*
1648                                  * Test if the migrated page should be merged
1649                                  * into a stable node dup. If the mapcount is
1650                                  * 1 we can migrate it with another KSM page
1651                                  * without adding it to the chain.
1652                                  */
1653                                 if (page_mapcount(page) > 1)
1654                                         goto chain_append;
1655                         }
1656
1657                         if (!stable_node_dup) {
1658                                 /*
1659                                  * If the stable_node is a chain and
1660                                  * we got a payload match in memcmp
1661                                  * but we cannot merge the scanned
1662                                  * page in any of the existing
1663                                  * stable_node dups because they're
1664                                  * all full, we need to wait the
1665                                  * scanned page to find itself a match
1666                                  * in the unstable tree to create a
1667                                  * brand new KSM page to add later to
1668                                  * the dups of this stable_node.
1669                                  */
1670                                 return NULL;
1671                         }
1672
1673                         /*
1674                          * Lock and unlock the stable_node's page (which
1675                          * might already have been migrated) so that page
1676                          * migration is sure to notice its raised count.
1677                          * It would be more elegant to return stable_node
1678                          * than kpage, but that involves more changes.
1679                          */
1680                         tree_page = get_ksm_page(stable_node_dup,
1681                                                  GET_KSM_PAGE_TRYLOCK);
1682
1683                         if (PTR_ERR(tree_page) == -EBUSY)
1684                                 return ERR_PTR(-EBUSY);
1685
1686                         if (unlikely(!tree_page))
1687                                 /*
1688                                  * The tree may have been rebalanced,
1689                                  * so re-evaluate parent and new.
1690                                  */
1691                                 goto again;
1692                         unlock_page(tree_page);
1693
1694                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1695                             NUMA(stable_node_dup->nid)) {
1696                                 put_page(tree_page);
1697                                 goto replace;
1698                         }
1699                         return tree_page;
1700                 }
1701         }
1702
1703         if (!page_node)
1704                 return NULL;
1705
1706         list_del(&page_node->list);
1707         DO_NUMA(page_node->nid = nid);
1708         rb_link_node(&page_node->node, parent, new);
1709         rb_insert_color(&page_node->node, root);
1710 out:
1711         if (is_page_sharing_candidate(page_node)) {
1712                 get_page(page);
1713                 return page;
1714         } else
1715                 return NULL;
1716
1717 replace:
1718         /*
1719          * If stable_node was a chain and chain_prune collapsed it,
1720          * stable_node has been updated to be the new regular
1721          * stable_node. A collapse of the chain is indistinguishable
1722          * from the case there was no chain in the stable
1723          * rbtree. Otherwise stable_node is the chain and
1724          * stable_node_dup is the dup to replace.
1725          */
1726         if (stable_node_dup == stable_node) {
1727                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1728                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1729                 /* there is no chain */
1730                 if (page_node) {
1731                         VM_BUG_ON(page_node->head != &migrate_nodes);
1732                         list_del(&page_node->list);
1733                         DO_NUMA(page_node->nid = nid);
1734                         rb_replace_node(&stable_node_dup->node,
1735                                         &page_node->node,
1736                                         root);
1737                         if (is_page_sharing_candidate(page_node))
1738                                 get_page(page);
1739                         else
1740                                 page = NULL;
1741                 } else {
1742                         rb_erase(&stable_node_dup->node, root);
1743                         page = NULL;
1744                 }
1745         } else {
1746                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1747                 __stable_node_dup_del(stable_node_dup);
1748                 if (page_node) {
1749                         VM_BUG_ON(page_node->head != &migrate_nodes);
1750                         list_del(&page_node->list);
1751                         DO_NUMA(page_node->nid = nid);
1752                         stable_node_chain_add_dup(page_node, stable_node);
1753                         if (is_page_sharing_candidate(page_node))
1754                                 get_page(page);
1755                         else
1756                                 page = NULL;
1757                 } else {
1758                         page = NULL;
1759                 }
1760         }
1761         stable_node_dup->head = &migrate_nodes;
1762         list_add(&stable_node_dup->list, stable_node_dup->head);
1763         return page;
1764
1765 chain_append:
1766         /* stable_node_dup could be null if it reached the limit */
1767         if (!stable_node_dup)
1768                 stable_node_dup = stable_node_any;
1769         /*
1770          * If stable_node was a chain and chain_prune collapsed it,
1771          * stable_node has been updated to be the new regular
1772          * stable_node. A collapse of the chain is indistinguishable
1773          * from the case there was no chain in the stable
1774          * rbtree. Otherwise stable_node is the chain and
1775          * stable_node_dup is the dup to replace.
1776          */
1777         if (stable_node_dup == stable_node) {
1778                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1779                 /* chain is missing so create it */
1780                 stable_node = alloc_stable_node_chain(stable_node_dup,
1781                                                       root);
1782                 if (!stable_node)
1783                         return NULL;
1784         }
1785         /*
1786          * Add this stable_node dup that was
1787          * migrated to the stable_node chain
1788          * of the current nid for this page
1789          * content.
1790          */
1791         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1792         VM_BUG_ON(page_node->head != &migrate_nodes);
1793         list_del(&page_node->list);
1794         DO_NUMA(page_node->nid = nid);
1795         stable_node_chain_add_dup(page_node, stable_node);
1796         goto out;
1797 }
1798
1799 /*
1800  * stable_tree_insert - insert stable tree node pointing to new ksm page
1801  * into the stable tree.
1802  *
1803  * This function returns the stable tree node just allocated on success,
1804  * NULL otherwise.
1805  */
1806 static struct stable_node *stable_tree_insert(struct page *kpage)
1807 {
1808         int nid;
1809         unsigned long kpfn;
1810         struct rb_root *root;
1811         struct rb_node **new;
1812         struct rb_node *parent;
1813         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1814         bool need_chain = false;
1815
1816         kpfn = page_to_pfn(kpage);
1817         nid = get_kpfn_nid(kpfn);
1818         root = root_stable_tree + nid;
1819 again:
1820         parent = NULL;
1821         new = &root->rb_node;
1822
1823         while (*new) {
1824                 struct page *tree_page;
1825                 int ret;
1826
1827                 cond_resched();
1828                 stable_node = rb_entry(*new, struct stable_node, node);
1829                 stable_node_any = NULL;
1830                 tree_page = chain(&stable_node_dup, stable_node, root);
1831                 if (!stable_node_dup) {
1832                         /*
1833                          * Either all stable_node dups were full in
1834                          * this stable_node chain, or this chain was
1835                          * empty and should be rb_erased.
1836                          */
1837                         stable_node_any = stable_node_dup_any(stable_node,
1838                                                               root);
1839                         if (!stable_node_any) {
1840                                 /* rb_erase just run */
1841                                 goto again;
1842                         }
1843                         /*
1844                          * Take any of the stable_node dups page of
1845                          * this stable_node chain to let the tree walk
1846                          * continue. All KSM pages belonging to the
1847                          * stable_node dups in a stable_node chain
1848                          * have the same content and they're
1849                          * write protected at all times. Any will work
1850                          * fine to continue the walk.
1851                          */
1852                         tree_page = get_ksm_page(stable_node_any,
1853                                                  GET_KSM_PAGE_NOLOCK);
1854                 }
1855                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1856                 if (!tree_page) {
1857                         /*
1858                          * If we walked over a stale stable_node,
1859                          * get_ksm_page() will call rb_erase() and it
1860                          * may rebalance the tree from under us. So
1861                          * restart the search from scratch. Returning
1862                          * NULL would be safe too, but we'd generate
1863                          * false negative insertions just because some
1864                          * stable_node was stale.
1865                          */
1866                         goto again;
1867                 }
1868
1869                 ret = memcmp_pages(kpage, tree_page);
1870                 put_page(tree_page);
1871
1872                 parent = *new;
1873                 if (ret < 0)
1874                         new = &parent->rb_left;
1875                 else if (ret > 0)
1876                         new = &parent->rb_right;
1877                 else {
1878                         need_chain = true;
1879                         break;
1880                 }
1881         }
1882
1883         stable_node_dup = alloc_stable_node();
1884         if (!stable_node_dup)
1885                 return NULL;
1886
1887         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1888         stable_node_dup->kpfn = kpfn;
1889         set_page_stable_node(kpage, stable_node_dup);
1890         stable_node_dup->rmap_hlist_len = 0;
1891         DO_NUMA(stable_node_dup->nid = nid);
1892         if (!need_chain) {
1893                 rb_link_node(&stable_node_dup->node, parent, new);
1894                 rb_insert_color(&stable_node_dup->node, root);
1895         } else {
1896                 if (!is_stable_node_chain(stable_node)) {
1897                         struct stable_node *orig = stable_node;
1898                         /* chain is missing so create it */
1899                         stable_node = alloc_stable_node_chain(orig, root);
1900                         if (!stable_node) {
1901                                 free_stable_node(stable_node_dup);
1902                                 return NULL;
1903                         }
1904                 }
1905                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1906         }
1907
1908         return stable_node_dup;
1909 }
1910
1911 /*
1912  * unstable_tree_search_insert - search for identical page,
1913  * else insert rmap_item into the unstable tree.
1914  *
1915  * This function searches for a page in the unstable tree identical to the
1916  * page currently being scanned; and if no identical page is found in the
1917  * tree, we insert rmap_item as a new object into the unstable tree.
1918  *
1919  * This function returns pointer to rmap_item found to be identical
1920  * to the currently scanned page, NULL otherwise.
1921  *
1922  * This function does both searching and inserting, because they share
1923  * the same walking algorithm in an rbtree.
1924  */
1925 static
1926 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1927                                               struct page *page,
1928                                               struct page **tree_pagep)
1929 {
1930         struct rb_node **new;
1931         struct rb_root *root;
1932         struct rb_node *parent = NULL;
1933         int nid;
1934
1935         nid = get_kpfn_nid(page_to_pfn(page));
1936         root = root_unstable_tree + nid;
1937         new = &root->rb_node;
1938
1939         while (*new) {
1940                 struct rmap_item *tree_rmap_item;
1941                 struct page *tree_page;
1942                 int ret;
1943
1944                 cond_resched();
1945                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1946                 tree_page = get_mergeable_page(tree_rmap_item);
1947                 if (!tree_page)
1948                         return NULL;
1949
1950                 /*
1951                  * Don't substitute a ksm page for a forked page.
1952                  */
1953                 if (page == tree_page) {
1954                         put_page(tree_page);
1955                         return NULL;
1956                 }
1957
1958                 ret = memcmp_pages(page, tree_page);
1959
1960                 parent = *new;
1961                 if (ret < 0) {
1962                         put_page(tree_page);
1963                         new = &parent->rb_left;
1964                 } else if (ret > 0) {
1965                         put_page(tree_page);
1966                         new = &parent->rb_right;
1967                 } else if (!ksm_merge_across_nodes &&
1968                            page_to_nid(tree_page) != nid) {
1969                         /*
1970                          * If tree_page has been migrated to another NUMA node,
1971                          * it will be flushed out and put in the right unstable
1972                          * tree next time: only merge with it when across_nodes.
1973                          */
1974                         put_page(tree_page);
1975                         return NULL;
1976                 } else {
1977                         *tree_pagep = tree_page;
1978                         return tree_rmap_item;
1979                 }
1980         }
1981
1982         rmap_item->address |= UNSTABLE_FLAG;
1983         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1984         DO_NUMA(rmap_item->nid = nid);
1985         rb_link_node(&rmap_item->node, parent, new);
1986         rb_insert_color(&rmap_item->node, root);
1987
1988         ksm_pages_unshared++;
1989         return NULL;
1990 }
1991
1992 /*
1993  * stable_tree_append - add another rmap_item to the linked list of
1994  * rmap_items hanging off a given node of the stable tree, all sharing
1995  * the same ksm page.
1996  */
1997 static void stable_tree_append(struct rmap_item *rmap_item,
1998                                struct stable_node *stable_node,
1999                                bool max_page_sharing_bypass)
2000 {
2001         /*
2002          * rmap won't find this mapping if we don't insert the
2003          * rmap_item in the right stable_node
2004          * duplicate. page_migration could break later if rmap breaks,
2005          * so we can as well crash here. We really need to check for
2006          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2007          * for other negative values as an underflow if detected here
2008          * for the first time (and not when decreasing rmap_hlist_len)
2009          * would be sign of memory corruption in the stable_node.
2010          */
2011         BUG_ON(stable_node->rmap_hlist_len < 0);
2012
2013         stable_node->rmap_hlist_len++;
2014         if (!max_page_sharing_bypass)
2015                 /* possibly non fatal but unexpected overflow, only warn */
2016                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2017                              ksm_max_page_sharing);
2018
2019         rmap_item->head = stable_node;
2020         rmap_item->address |= STABLE_FLAG;
2021         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2022
2023         if (rmap_item->hlist.next)
2024                 ksm_pages_sharing++;
2025         else
2026                 ksm_pages_shared++;
2027
2028         rmap_item->mm->ksm_merging_pages++;
2029 }
2030
2031 /*
2032  * cmp_and_merge_page - first see if page can be merged into the stable tree;
2033  * if not, compare checksum to previous and if it's the same, see if page can
2034  * be inserted into the unstable tree, or merged with a page already there and
2035  * both transferred to the stable tree.
2036  *
2037  * @page: the page that we are searching identical page to.
2038  * @rmap_item: the reverse mapping into the virtual address of this page
2039  */
2040 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2041 {
2042         struct mm_struct *mm = rmap_item->mm;
2043         struct rmap_item *tree_rmap_item;
2044         struct page *tree_page = NULL;
2045         struct stable_node *stable_node;
2046         struct page *kpage;
2047         unsigned int checksum;
2048         int err;
2049         bool max_page_sharing_bypass = false;
2050
2051         stable_node = page_stable_node(page);
2052         if (stable_node) {
2053                 if (stable_node->head != &migrate_nodes &&
2054                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2055                     NUMA(stable_node->nid)) {
2056                         stable_node_dup_del(stable_node);
2057                         stable_node->head = &migrate_nodes;
2058                         list_add(&stable_node->list, stable_node->head);
2059                 }
2060                 if (stable_node->head != &migrate_nodes &&
2061                     rmap_item->head == stable_node)
2062                         return;
2063                 /*
2064                  * If it's a KSM fork, allow it to go over the sharing limit
2065                  * without warnings.
2066                  */
2067                 if (!is_page_sharing_candidate(stable_node))
2068                         max_page_sharing_bypass = true;
2069         }
2070
2071         /* We first start with searching the page inside the stable tree */
2072         kpage = stable_tree_search(page);
2073         if (kpage == page && rmap_item->head == stable_node) {
2074                 put_page(kpage);
2075                 return;
2076         }
2077
2078         remove_rmap_item_from_tree(rmap_item);
2079
2080         if (kpage) {
2081                 if (PTR_ERR(kpage) == -EBUSY)
2082                         return;
2083
2084                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2085                 if (!err) {
2086                         /*
2087                          * The page was successfully merged:
2088                          * add its rmap_item to the stable tree.
2089                          */
2090                         lock_page(kpage);
2091                         stable_tree_append(rmap_item, page_stable_node(kpage),
2092                                            max_page_sharing_bypass);
2093                         unlock_page(kpage);
2094                 }
2095                 put_page(kpage);
2096                 return;
2097         }
2098
2099         /*
2100          * If the hash value of the page has changed from the last time
2101          * we calculated it, this page is changing frequently: therefore we
2102          * don't want to insert it in the unstable tree, and we don't want
2103          * to waste our time searching for something identical to it there.
2104          */
2105         checksum = calc_checksum(page);
2106         if (rmap_item->oldchecksum != checksum) {
2107                 rmap_item->oldchecksum = checksum;
2108                 return;
2109         }
2110
2111         /*
2112          * Same checksum as an empty page. We attempt to merge it with the
2113          * appropriate zero page if the user enabled this via sysfs.
2114          */
2115         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2116                 struct vm_area_struct *vma;
2117
2118                 mmap_read_lock(mm);
2119                 vma = find_mergeable_vma(mm, rmap_item->address);
2120                 if (vma) {
2121                         err = try_to_merge_one_page(vma, page,
2122                                         ZERO_PAGE(rmap_item->address));
2123                 } else {
2124                         /*
2125                          * If the vma is out of date, we do not need to
2126                          * continue.
2127                          */
2128                         err = 0;
2129                 }
2130                 mmap_read_unlock(mm);
2131                 /*
2132                  * In case of failure, the page was not really empty, so we
2133                  * need to continue. Otherwise we're done.
2134                  */
2135                 if (!err)
2136                         return;
2137         }
2138         tree_rmap_item =
2139                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2140         if (tree_rmap_item) {
2141                 bool split;
2142
2143                 kpage = try_to_merge_two_pages(rmap_item, page,
2144                                                 tree_rmap_item, tree_page);
2145                 /*
2146                  * If both pages we tried to merge belong to the same compound
2147                  * page, then we actually ended up increasing the reference
2148                  * count of the same compound page twice, and split_huge_page
2149                  * failed.
2150                  * Here we set a flag if that happened, and we use it later to
2151                  * try split_huge_page again. Since we call put_page right
2152                  * afterwards, the reference count will be correct and
2153                  * split_huge_page should succeed.
2154                  */
2155                 split = PageTransCompound(page)
2156                         && compound_head(page) == compound_head(tree_page);
2157                 put_page(tree_page);
2158                 if (kpage) {
2159                         /*
2160                          * The pages were successfully merged: insert new
2161                          * node in the stable tree and add both rmap_items.
2162                          */
2163                         lock_page(kpage);
2164                         stable_node = stable_tree_insert(kpage);
2165                         if (stable_node) {
2166                                 stable_tree_append(tree_rmap_item, stable_node,
2167                                                    false);
2168                                 stable_tree_append(rmap_item, stable_node,
2169                                                    false);
2170                         }
2171                         unlock_page(kpage);
2172
2173                         /*
2174                          * If we fail to insert the page into the stable tree,
2175                          * we will have 2 virtual addresses that are pointing
2176                          * to a ksm page left outside the stable tree,
2177                          * in which case we need to break_cow on both.
2178                          */
2179                         if (!stable_node) {
2180                                 break_cow(tree_rmap_item);
2181                                 break_cow(rmap_item);
2182                         }
2183                 } else if (split) {
2184                         /*
2185                          * We are here if we tried to merge two pages and
2186                          * failed because they both belonged to the same
2187                          * compound page. We will split the page now, but no
2188                          * merging will take place.
2189                          * We do not want to add the cost of a full lock; if
2190                          * the page is locked, it is better to skip it and
2191                          * perhaps try again later.
2192                          */
2193                         if (!trylock_page(page))
2194                                 return;
2195                         split_huge_page(page);
2196                         unlock_page(page);
2197                 }
2198         }
2199 }
2200
2201 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2202                                             struct rmap_item **rmap_list,
2203                                             unsigned long addr)
2204 {
2205         struct rmap_item *rmap_item;
2206
2207         while (*rmap_list) {
2208                 rmap_item = *rmap_list;
2209                 if ((rmap_item->address & PAGE_MASK) == addr)
2210                         return rmap_item;
2211                 if (rmap_item->address > addr)
2212                         break;
2213                 *rmap_list = rmap_item->rmap_list;
2214                 remove_rmap_item_from_tree(rmap_item);
2215                 free_rmap_item(rmap_item);
2216         }
2217
2218         rmap_item = alloc_rmap_item();
2219         if (rmap_item) {
2220                 /* It has already been zeroed */
2221                 rmap_item->mm = mm_slot->mm;
2222                 rmap_item->address = addr;
2223                 rmap_item->rmap_list = *rmap_list;
2224                 *rmap_list = rmap_item;
2225         }
2226         return rmap_item;
2227 }
2228
2229 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2230 {
2231         struct mm_struct *mm;
2232         struct mm_slot *slot;
2233         struct vm_area_struct *vma;
2234         struct rmap_item *rmap_item;
2235         int nid;
2236
2237         if (list_empty(&ksm_mm_head.mm_list))
2238                 return NULL;
2239
2240         slot = ksm_scan.mm_slot;
2241         if (slot == &ksm_mm_head) {
2242                 /*
2243                  * A number of pages can hang around indefinitely on per-cpu
2244                  * pagevecs, raised page count preventing write_protect_page
2245                  * from merging them.  Though it doesn't really matter much,
2246                  * it is puzzling to see some stuck in pages_volatile until
2247                  * other activity jostles them out, and they also prevented
2248                  * LTP's KSM test from succeeding deterministically; so drain
2249                  * them here (here rather than on entry to ksm_do_scan(),
2250                  * so we don't IPI too often when pages_to_scan is set low).
2251                  */
2252                 lru_add_drain_all();
2253
2254                 /*
2255                  * Whereas stale stable_nodes on the stable_tree itself
2256                  * get pruned in the regular course of stable_tree_search(),
2257                  * those moved out to the migrate_nodes list can accumulate:
2258                  * so prune them once before each full scan.
2259                  */
2260                 if (!ksm_merge_across_nodes) {
2261                         struct stable_node *stable_node, *next;
2262                         struct page *page;
2263
2264                         list_for_each_entry_safe(stable_node, next,
2265                                                  &migrate_nodes, list) {
2266                                 page = get_ksm_page(stable_node,
2267                                                     GET_KSM_PAGE_NOLOCK);
2268                                 if (page)
2269                                         put_page(page);
2270                                 cond_resched();
2271                         }
2272                 }
2273
2274                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2275                         root_unstable_tree[nid] = RB_ROOT;
2276
2277                 spin_lock(&ksm_mmlist_lock);
2278                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2279                 ksm_scan.mm_slot = slot;
2280                 spin_unlock(&ksm_mmlist_lock);
2281                 /*
2282                  * Although we tested list_empty() above, a racing __ksm_exit
2283                  * of the last mm on the list may have removed it since then.
2284                  */
2285                 if (slot == &ksm_mm_head)
2286                         return NULL;
2287 next_mm:
2288                 ksm_scan.address = 0;
2289                 ksm_scan.rmap_list = &slot->rmap_list;
2290         }
2291
2292         mm = slot->mm;
2293         mmap_read_lock(mm);
2294         if (ksm_test_exit(mm))
2295                 vma = NULL;
2296         else
2297                 vma = find_vma(mm, ksm_scan.address);
2298
2299         for (; vma; vma = vma->vm_next) {
2300                 if (!(vma->vm_flags & VM_MERGEABLE))
2301                         continue;
2302                 if (ksm_scan.address < vma->vm_start)
2303                         ksm_scan.address = vma->vm_start;
2304                 if (!vma->anon_vma)
2305                         ksm_scan.address = vma->vm_end;
2306
2307                 while (ksm_scan.address < vma->vm_end) {
2308                         if (ksm_test_exit(mm))
2309                                 break;
2310                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2311                         if (IS_ERR_OR_NULL(*page) || is_zone_device_page(*page)) {
2312                                 ksm_scan.address += PAGE_SIZE;
2313                                 cond_resched();
2314                                 continue;
2315                         }
2316                         if (PageAnon(*page)) {
2317                                 flush_anon_page(vma, *page, ksm_scan.address);
2318                                 flush_dcache_page(*page);
2319                                 rmap_item = get_next_rmap_item(slot,
2320                                         ksm_scan.rmap_list, ksm_scan.address);
2321                                 if (rmap_item) {
2322                                         ksm_scan.rmap_list =
2323                                                         &rmap_item->rmap_list;
2324                                         ksm_scan.address += PAGE_SIZE;
2325                                 } else
2326                                         put_page(*page);
2327                                 mmap_read_unlock(mm);
2328                                 return rmap_item;
2329                         }
2330                         put_page(*page);
2331                         ksm_scan.address += PAGE_SIZE;
2332                         cond_resched();
2333                 }
2334         }
2335
2336         if (ksm_test_exit(mm)) {
2337                 ksm_scan.address = 0;
2338                 ksm_scan.rmap_list = &slot->rmap_list;
2339         }
2340         /*
2341          * Nuke all the rmap_items that are above this current rmap:
2342          * because there were no VM_MERGEABLE vmas with such addresses.
2343          */
2344         remove_trailing_rmap_items(ksm_scan.rmap_list);
2345
2346         spin_lock(&ksm_mmlist_lock);
2347         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2348                                                 struct mm_slot, mm_list);
2349         if (ksm_scan.address == 0) {
2350                 /*
2351                  * We've completed a full scan of all vmas, holding mmap_lock
2352                  * throughout, and found no VM_MERGEABLE: so do the same as
2353                  * __ksm_exit does to remove this mm from all our lists now.
2354                  * This applies either when cleaning up after __ksm_exit
2355                  * (but beware: we can reach here even before __ksm_exit),
2356                  * or when all VM_MERGEABLE areas have been unmapped (and
2357                  * mmap_lock then protects against race with MADV_MERGEABLE).
2358                  */
2359                 hash_del(&slot->link);
2360                 list_del(&slot->mm_list);
2361                 spin_unlock(&ksm_mmlist_lock);
2362
2363                 free_mm_slot(slot);
2364                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2365                 mmap_read_unlock(mm);
2366                 mmdrop(mm);
2367         } else {
2368                 mmap_read_unlock(mm);
2369                 /*
2370                  * mmap_read_unlock(mm) first because after
2371                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2372                  * already have been freed under us by __ksm_exit()
2373                  * because the "mm_slot" is still hashed and
2374                  * ksm_scan.mm_slot doesn't point to it anymore.
2375                  */
2376                 spin_unlock(&ksm_mmlist_lock);
2377         }
2378
2379         /* Repeat until we've completed scanning the whole list */
2380         slot = ksm_scan.mm_slot;
2381         if (slot != &ksm_mm_head)
2382                 goto next_mm;
2383
2384         ksm_scan.seqnr++;
2385         return NULL;
2386 }
2387
2388 /**
2389  * ksm_do_scan  - the ksm scanner main worker function.
2390  * @scan_npages:  number of pages we want to scan before we return.
2391  */
2392 static void ksm_do_scan(unsigned int scan_npages)
2393 {
2394         struct rmap_item *rmap_item;
2395         struct page *page;
2396
2397         while (scan_npages-- && likely(!freezing(current))) {
2398                 cond_resched();
2399                 rmap_item = scan_get_next_rmap_item(&page);
2400                 if (!rmap_item)
2401                         return;
2402                 cmp_and_merge_page(page, rmap_item);
2403                 put_page(page);
2404         }
2405 }
2406
2407 static int ksmd_should_run(void)
2408 {
2409         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2410 }
2411
2412 static int ksm_scan_thread(void *nothing)
2413 {
2414         unsigned int sleep_ms;
2415
2416         set_freezable();
2417         set_user_nice(current, 5);
2418
2419         while (!kthread_should_stop()) {
2420                 mutex_lock(&ksm_thread_mutex);
2421                 wait_while_offlining();
2422                 if (ksmd_should_run())
2423                         ksm_do_scan(ksm_thread_pages_to_scan);
2424                 mutex_unlock(&ksm_thread_mutex);
2425
2426                 try_to_freeze();
2427
2428                 if (ksmd_should_run()) {
2429                         sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2430                         wait_event_interruptible_timeout(ksm_iter_wait,
2431                                 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2432                                 msecs_to_jiffies(sleep_ms));
2433                 } else {
2434                         wait_event_freezable(ksm_thread_wait,
2435                                 ksmd_should_run() || kthread_should_stop());
2436                 }
2437         }
2438         return 0;
2439 }
2440
2441 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2442                 unsigned long end, int advice, unsigned long *vm_flags)
2443 {
2444         struct mm_struct *mm = vma->vm_mm;
2445         int err;
2446
2447         switch (advice) {
2448         case MADV_MERGEABLE:
2449                 /*
2450                  * Be somewhat over-protective for now!
2451                  */
2452                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2453                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2454                                  VM_HUGETLB | VM_MIXEDMAP))
2455                         return 0;               /* just ignore the advice */
2456
2457                 if (vma_is_dax(vma))
2458                         return 0;
2459
2460 #ifdef VM_SAO
2461                 if (*vm_flags & VM_SAO)
2462                         return 0;
2463 #endif
2464 #ifdef VM_SPARC_ADI
2465                 if (*vm_flags & VM_SPARC_ADI)
2466                         return 0;
2467 #endif
2468
2469                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2470                         err = __ksm_enter(mm);
2471                         if (err)
2472                                 return err;
2473                 }
2474
2475                 *vm_flags |= VM_MERGEABLE;
2476                 break;
2477
2478         case MADV_UNMERGEABLE:
2479                 if (!(*vm_flags & VM_MERGEABLE))
2480                         return 0;               /* just ignore the advice */
2481
2482                 if (vma->anon_vma) {
2483                         err = unmerge_ksm_pages(vma, start, end);
2484                         if (err)
2485                                 return err;
2486                 }
2487
2488                 *vm_flags &= ~VM_MERGEABLE;
2489                 break;
2490         }
2491
2492         return 0;
2493 }
2494 EXPORT_SYMBOL_GPL(ksm_madvise);
2495
2496 int __ksm_enter(struct mm_struct *mm)
2497 {
2498         struct mm_slot *mm_slot;
2499         int needs_wakeup;
2500
2501         mm_slot = alloc_mm_slot();
2502         if (!mm_slot)
2503                 return -ENOMEM;
2504
2505         /* Check ksm_run too?  Would need tighter locking */
2506         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2507
2508         spin_lock(&ksm_mmlist_lock);
2509         insert_to_mm_slots_hash(mm, mm_slot);
2510         /*
2511          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2512          * insert just behind the scanning cursor, to let the area settle
2513          * down a little; when fork is followed by immediate exec, we don't
2514          * want ksmd to waste time setting up and tearing down an rmap_list.
2515          *
2516          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2517          * scanning cursor, otherwise KSM pages in newly forked mms will be
2518          * missed: then we might as well insert at the end of the list.
2519          */
2520         if (ksm_run & KSM_RUN_UNMERGE)
2521                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2522         else
2523                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2524         spin_unlock(&ksm_mmlist_lock);
2525
2526         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2527         mmgrab(mm);
2528
2529         if (needs_wakeup)
2530                 wake_up_interruptible(&ksm_thread_wait);
2531
2532         return 0;
2533 }
2534
2535 void __ksm_exit(struct mm_struct *mm)
2536 {
2537         struct mm_slot *mm_slot;
2538         int easy_to_free = 0;
2539
2540         /*
2541          * This process is exiting: if it's straightforward (as is the
2542          * case when ksmd was never running), free mm_slot immediately.
2543          * But if it's at the cursor or has rmap_items linked to it, use
2544          * mmap_lock to synchronize with any break_cows before pagetables
2545          * are freed, and leave the mm_slot on the list for ksmd to free.
2546          * Beware: ksm may already have noticed it exiting and freed the slot.
2547          */
2548
2549         spin_lock(&ksm_mmlist_lock);
2550         mm_slot = get_mm_slot(mm);
2551         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2552                 if (!mm_slot->rmap_list) {
2553                         hash_del(&mm_slot->link);
2554                         list_del(&mm_slot->mm_list);
2555                         easy_to_free = 1;
2556                 } else {
2557                         list_move(&mm_slot->mm_list,
2558                                   &ksm_scan.mm_slot->mm_list);
2559                 }
2560         }
2561         spin_unlock(&ksm_mmlist_lock);
2562
2563         if (easy_to_free) {
2564                 free_mm_slot(mm_slot);
2565                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2566                 mmdrop(mm);
2567         } else if (mm_slot) {
2568                 mmap_write_lock(mm);
2569                 mmap_write_unlock(mm);
2570         }
2571 }
2572
2573 struct page *ksm_might_need_to_copy(struct page *page,
2574                         struct vm_area_struct *vma, unsigned long address)
2575 {
2576         struct folio *folio = page_folio(page);
2577         struct anon_vma *anon_vma = folio_anon_vma(folio);
2578         struct page *new_page;
2579
2580         if (PageKsm(page)) {
2581                 if (page_stable_node(page) &&
2582                     !(ksm_run & KSM_RUN_UNMERGE))
2583                         return page;    /* no need to copy it */
2584         } else if (!anon_vma) {
2585                 return page;            /* no need to copy it */
2586         } else if (page->index == linear_page_index(vma, address) &&
2587                         anon_vma->root == vma->anon_vma->root) {
2588                 return page;            /* still no need to copy it */
2589         }
2590         if (!PageUptodate(page))
2591                 return page;            /* let do_swap_page report the error */
2592
2593         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2594         if (new_page &&
2595             mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) {
2596                 put_page(new_page);
2597                 new_page = NULL;
2598         }
2599         if (new_page) {
2600                 copy_user_highpage(new_page, page, address, vma);
2601
2602                 SetPageDirty(new_page);
2603                 __SetPageUptodate(new_page);
2604                 __SetPageLocked(new_page);
2605 #ifdef CONFIG_SWAP
2606                 count_vm_event(KSM_SWPIN_COPY);
2607 #endif
2608         }
2609
2610         return new_page;
2611 }
2612
2613 void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc)
2614 {
2615         struct stable_node *stable_node;
2616         struct rmap_item *rmap_item;
2617         int search_new_forks = 0;
2618
2619         VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio);
2620
2621         /*
2622          * Rely on the page lock to protect against concurrent modifications
2623          * to that page's node of the stable tree.
2624          */
2625         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2626
2627         stable_node = folio_stable_node(folio);
2628         if (!stable_node)
2629                 return;
2630 again:
2631         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2632                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2633                 struct anon_vma_chain *vmac;
2634                 struct vm_area_struct *vma;
2635
2636                 cond_resched();
2637                 if (!anon_vma_trylock_read(anon_vma)) {
2638                         if (rwc->try_lock) {
2639                                 rwc->contended = true;
2640                                 return;
2641                         }
2642                         anon_vma_lock_read(anon_vma);
2643                 }
2644                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2645                                                0, ULONG_MAX) {
2646                         unsigned long addr;
2647
2648                         cond_resched();
2649                         vma = vmac->vma;
2650
2651                         /* Ignore the stable/unstable/sqnr flags */
2652                         addr = rmap_item->address & PAGE_MASK;
2653
2654                         if (addr < vma->vm_start || addr >= vma->vm_end)
2655                                 continue;
2656                         /*
2657                          * Initially we examine only the vma which covers this
2658                          * rmap_item; but later, if there is still work to do,
2659                          * we examine covering vmas in other mms: in case they
2660                          * were forked from the original since ksmd passed.
2661                          */
2662                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2663                                 continue;
2664
2665                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2666                                 continue;
2667
2668                         if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) {
2669                                 anon_vma_unlock_read(anon_vma);
2670                                 return;
2671                         }
2672                         if (rwc->done && rwc->done(folio)) {
2673                                 anon_vma_unlock_read(anon_vma);
2674                                 return;
2675                         }
2676                 }
2677                 anon_vma_unlock_read(anon_vma);
2678         }
2679         if (!search_new_forks++)
2680                 goto again;
2681 }
2682
2683 #ifdef CONFIG_MIGRATION
2684 void folio_migrate_ksm(struct folio *newfolio, struct folio *folio)
2685 {
2686         struct stable_node *stable_node;
2687
2688         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2689         VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio);
2690         VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio);
2691
2692         stable_node = folio_stable_node(folio);
2693         if (stable_node) {
2694                 VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio);
2695                 stable_node->kpfn = folio_pfn(newfolio);
2696                 /*
2697                  * newfolio->mapping was set in advance; now we need smp_wmb()
2698                  * to make sure that the new stable_node->kpfn is visible
2699                  * to get_ksm_page() before it can see that folio->mapping
2700                  * has gone stale (or that folio_test_swapcache has been cleared).
2701                  */
2702                 smp_wmb();
2703                 set_page_stable_node(&folio->page, NULL);
2704         }
2705 }
2706 #endif /* CONFIG_MIGRATION */
2707
2708 #ifdef CONFIG_MEMORY_HOTREMOVE
2709 static void wait_while_offlining(void)
2710 {
2711         while (ksm_run & KSM_RUN_OFFLINE) {
2712                 mutex_unlock(&ksm_thread_mutex);
2713                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2714                             TASK_UNINTERRUPTIBLE);
2715                 mutex_lock(&ksm_thread_mutex);
2716         }
2717 }
2718
2719 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2720                                          unsigned long start_pfn,
2721                                          unsigned long end_pfn)
2722 {
2723         if (stable_node->kpfn >= start_pfn &&
2724             stable_node->kpfn < end_pfn) {
2725                 /*
2726                  * Don't get_ksm_page, page has already gone:
2727                  * which is why we keep kpfn instead of page*
2728                  */
2729                 remove_node_from_stable_tree(stable_node);
2730                 return true;
2731         }
2732         return false;
2733 }
2734
2735 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2736                                            unsigned long start_pfn,
2737                                            unsigned long end_pfn,
2738                                            struct rb_root *root)
2739 {
2740         struct stable_node *dup;
2741         struct hlist_node *hlist_safe;
2742
2743         if (!is_stable_node_chain(stable_node)) {
2744                 VM_BUG_ON(is_stable_node_dup(stable_node));
2745                 return stable_node_dup_remove_range(stable_node, start_pfn,
2746                                                     end_pfn);
2747         }
2748
2749         hlist_for_each_entry_safe(dup, hlist_safe,
2750                                   &stable_node->hlist, hlist_dup) {
2751                 VM_BUG_ON(!is_stable_node_dup(dup));
2752                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2753         }
2754         if (hlist_empty(&stable_node->hlist)) {
2755                 free_stable_node_chain(stable_node, root);
2756                 return true; /* notify caller that tree was rebalanced */
2757         } else
2758                 return false;
2759 }
2760
2761 static void ksm_check_stable_tree(unsigned long start_pfn,
2762                                   unsigned long end_pfn)
2763 {
2764         struct stable_node *stable_node, *next;
2765         struct rb_node *node;
2766         int nid;
2767
2768         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2769                 node = rb_first(root_stable_tree + nid);
2770                 while (node) {
2771                         stable_node = rb_entry(node, struct stable_node, node);
2772                         if (stable_node_chain_remove_range(stable_node,
2773                                                            start_pfn, end_pfn,
2774                                                            root_stable_tree +
2775                                                            nid))
2776                                 node = rb_first(root_stable_tree + nid);
2777                         else
2778                                 node = rb_next(node);
2779                         cond_resched();
2780                 }
2781         }
2782         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2783                 if (stable_node->kpfn >= start_pfn &&
2784                     stable_node->kpfn < end_pfn)
2785                         remove_node_from_stable_tree(stable_node);
2786                 cond_resched();
2787         }
2788 }
2789
2790 static int ksm_memory_callback(struct notifier_block *self,
2791                                unsigned long action, void *arg)
2792 {
2793         struct memory_notify *mn = arg;
2794
2795         switch (action) {
2796         case MEM_GOING_OFFLINE:
2797                 /*
2798                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2799                  * and remove_all_stable_nodes() while memory is going offline:
2800                  * it is unsafe for them to touch the stable tree at this time.
2801                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2802                  * which do not need the ksm_thread_mutex are all safe.
2803                  */
2804                 mutex_lock(&ksm_thread_mutex);
2805                 ksm_run |= KSM_RUN_OFFLINE;
2806                 mutex_unlock(&ksm_thread_mutex);
2807                 break;
2808
2809         case MEM_OFFLINE:
2810                 /*
2811                  * Most of the work is done by page migration; but there might
2812                  * be a few stable_nodes left over, still pointing to struct
2813                  * pages which have been offlined: prune those from the tree,
2814                  * otherwise get_ksm_page() might later try to access a
2815                  * non-existent struct page.
2816                  */
2817                 ksm_check_stable_tree(mn->start_pfn,
2818                                       mn->start_pfn + mn->nr_pages);
2819                 fallthrough;
2820         case MEM_CANCEL_OFFLINE:
2821                 mutex_lock(&ksm_thread_mutex);
2822                 ksm_run &= ~KSM_RUN_OFFLINE;
2823                 mutex_unlock(&ksm_thread_mutex);
2824
2825                 smp_mb();       /* wake_up_bit advises this */
2826                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2827                 break;
2828         }
2829         return NOTIFY_OK;
2830 }
2831 #else
2832 static void wait_while_offlining(void)
2833 {
2834 }
2835 #endif /* CONFIG_MEMORY_HOTREMOVE */
2836
2837 #ifdef CONFIG_SYSFS
2838 /*
2839  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2840  */
2841
2842 #define KSM_ATTR_RO(_name) \
2843         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2844 #define KSM_ATTR(_name) \
2845         static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
2846
2847 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2848                                     struct kobj_attribute *attr, char *buf)
2849 {
2850         return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
2851 }
2852
2853 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2854                                      struct kobj_attribute *attr,
2855                                      const char *buf, size_t count)
2856 {
2857         unsigned int msecs;
2858         int err;
2859
2860         err = kstrtouint(buf, 10, &msecs);
2861         if (err)
2862                 return -EINVAL;
2863
2864         ksm_thread_sleep_millisecs = msecs;
2865         wake_up_interruptible(&ksm_iter_wait);
2866
2867         return count;
2868 }
2869 KSM_ATTR(sleep_millisecs);
2870
2871 static ssize_t pages_to_scan_show(struct kobject *kobj,
2872                                   struct kobj_attribute *attr, char *buf)
2873 {
2874         return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
2875 }
2876
2877 static ssize_t pages_to_scan_store(struct kobject *kobj,
2878                                    struct kobj_attribute *attr,
2879                                    const char *buf, size_t count)
2880 {
2881         unsigned int nr_pages;
2882         int err;
2883
2884         err = kstrtouint(buf, 10, &nr_pages);
2885         if (err)
2886                 return -EINVAL;
2887
2888         ksm_thread_pages_to_scan = nr_pages;
2889
2890         return count;
2891 }
2892 KSM_ATTR(pages_to_scan);
2893
2894 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2895                         char *buf)
2896 {
2897         return sysfs_emit(buf, "%lu\n", ksm_run);
2898 }
2899
2900 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2901                          const char *buf, size_t count)
2902 {
2903         unsigned int flags;
2904         int err;
2905
2906         err = kstrtouint(buf, 10, &flags);
2907         if (err)
2908                 return -EINVAL;
2909         if (flags > KSM_RUN_UNMERGE)
2910                 return -EINVAL;
2911
2912         /*
2913          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2914          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2915          * breaking COW to free the pages_shared (but leaves mm_slots
2916          * on the list for when ksmd may be set running again).
2917          */
2918
2919         mutex_lock(&ksm_thread_mutex);
2920         wait_while_offlining();
2921         if (ksm_run != flags) {
2922                 ksm_run = flags;
2923                 if (flags & KSM_RUN_UNMERGE) {
2924                         set_current_oom_origin();
2925                         err = unmerge_and_remove_all_rmap_items();
2926                         clear_current_oom_origin();
2927                         if (err) {
2928                                 ksm_run = KSM_RUN_STOP;
2929                                 count = err;
2930                         }
2931                 }
2932         }
2933         mutex_unlock(&ksm_thread_mutex);
2934
2935         if (flags & KSM_RUN_MERGE)
2936                 wake_up_interruptible(&ksm_thread_wait);
2937
2938         return count;
2939 }
2940 KSM_ATTR(run);
2941
2942 #ifdef CONFIG_NUMA
2943 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2944                                        struct kobj_attribute *attr, char *buf)
2945 {
2946         return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
2947 }
2948
2949 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2950                                    struct kobj_attribute *attr,
2951                                    const char *buf, size_t count)
2952 {
2953         int err;
2954         unsigned long knob;
2955
2956         err = kstrtoul(buf, 10, &knob);
2957         if (err)
2958                 return err;
2959         if (knob > 1)
2960                 return -EINVAL;
2961
2962         mutex_lock(&ksm_thread_mutex);
2963         wait_while_offlining();
2964         if (ksm_merge_across_nodes != knob) {
2965                 if (ksm_pages_shared || remove_all_stable_nodes())
2966                         err = -EBUSY;
2967                 else if (root_stable_tree == one_stable_tree) {
2968                         struct rb_root *buf;
2969                         /*
2970                          * This is the first time that we switch away from the
2971                          * default of merging across nodes: must now allocate
2972                          * a buffer to hold as many roots as may be needed.
2973                          * Allocate stable and unstable together:
2974                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2975                          */
2976                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2977                                       GFP_KERNEL);
2978                         /* Let us assume that RB_ROOT is NULL is zero */
2979                         if (!buf)
2980                                 err = -ENOMEM;
2981                         else {
2982                                 root_stable_tree = buf;
2983                                 root_unstable_tree = buf + nr_node_ids;
2984                                 /* Stable tree is empty but not the unstable */
2985                                 root_unstable_tree[0] = one_unstable_tree[0];
2986                         }
2987                 }
2988                 if (!err) {
2989                         ksm_merge_across_nodes = knob;
2990                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2991                 }
2992         }
2993         mutex_unlock(&ksm_thread_mutex);
2994
2995         return err ? err : count;
2996 }
2997 KSM_ATTR(merge_across_nodes);
2998 #endif
2999
3000 static ssize_t use_zero_pages_show(struct kobject *kobj,
3001                                    struct kobj_attribute *attr, char *buf)
3002 {
3003         return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
3004 }
3005 static ssize_t use_zero_pages_store(struct kobject *kobj,
3006                                    struct kobj_attribute *attr,
3007                                    const char *buf, size_t count)
3008 {
3009         int err;
3010         bool value;
3011
3012         err = kstrtobool(buf, &value);
3013         if (err)
3014                 return -EINVAL;
3015
3016         ksm_use_zero_pages = value;
3017
3018         return count;
3019 }
3020 KSM_ATTR(use_zero_pages);
3021
3022 static ssize_t max_page_sharing_show(struct kobject *kobj,
3023                                      struct kobj_attribute *attr, char *buf)
3024 {
3025         return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3026 }
3027
3028 static ssize_t max_page_sharing_store(struct kobject *kobj,
3029                                       struct kobj_attribute *attr,
3030                                       const char *buf, size_t count)
3031 {
3032         int err;
3033         int knob;
3034
3035         err = kstrtoint(buf, 10, &knob);
3036         if (err)
3037                 return err;
3038         /*
3039          * When a KSM page is created it is shared by 2 mappings. This
3040          * being a signed comparison, it implicitly verifies it's not
3041          * negative.
3042          */
3043         if (knob < 2)
3044                 return -EINVAL;
3045
3046         if (READ_ONCE(ksm_max_page_sharing) == knob)
3047                 return count;
3048
3049         mutex_lock(&ksm_thread_mutex);
3050         wait_while_offlining();
3051         if (ksm_max_page_sharing != knob) {
3052                 if (ksm_pages_shared || remove_all_stable_nodes())
3053                         err = -EBUSY;
3054                 else
3055                         ksm_max_page_sharing = knob;
3056         }
3057         mutex_unlock(&ksm_thread_mutex);
3058
3059         return err ? err : count;
3060 }
3061 KSM_ATTR(max_page_sharing);
3062
3063 static ssize_t pages_shared_show(struct kobject *kobj,
3064                                  struct kobj_attribute *attr, char *buf)
3065 {
3066         return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3067 }
3068 KSM_ATTR_RO(pages_shared);
3069
3070 static ssize_t pages_sharing_show(struct kobject *kobj,
3071                                   struct kobj_attribute *attr, char *buf)
3072 {
3073         return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3074 }
3075 KSM_ATTR_RO(pages_sharing);
3076
3077 static ssize_t pages_unshared_show(struct kobject *kobj,
3078                                    struct kobj_attribute *attr, char *buf)
3079 {
3080         return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3081 }
3082 KSM_ATTR_RO(pages_unshared);
3083
3084 static ssize_t pages_volatile_show(struct kobject *kobj,
3085                                    struct kobj_attribute *attr, char *buf)
3086 {
3087         long ksm_pages_volatile;
3088
3089         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3090                                 - ksm_pages_sharing - ksm_pages_unshared;
3091         /*
3092          * It was not worth any locking to calculate that statistic,
3093          * but it might therefore sometimes be negative: conceal that.
3094          */
3095         if (ksm_pages_volatile < 0)
3096                 ksm_pages_volatile = 0;
3097         return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3098 }
3099 KSM_ATTR_RO(pages_volatile);
3100
3101 static ssize_t stable_node_dups_show(struct kobject *kobj,
3102                                      struct kobj_attribute *attr, char *buf)
3103 {
3104         return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3105 }
3106 KSM_ATTR_RO(stable_node_dups);
3107
3108 static ssize_t stable_node_chains_show(struct kobject *kobj,
3109                                        struct kobj_attribute *attr, char *buf)
3110 {
3111         return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3112 }
3113 KSM_ATTR_RO(stable_node_chains);
3114
3115 static ssize_t
3116 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3117                                         struct kobj_attribute *attr,
3118                                         char *buf)
3119 {
3120         return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3121 }
3122
3123 static ssize_t
3124 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3125                                          struct kobj_attribute *attr,
3126                                          const char *buf, size_t count)
3127 {
3128         unsigned int msecs;
3129         int err;
3130
3131         err = kstrtouint(buf, 10, &msecs);
3132         if (err)
3133                 return -EINVAL;
3134
3135         ksm_stable_node_chains_prune_millisecs = msecs;
3136
3137         return count;
3138 }
3139 KSM_ATTR(stable_node_chains_prune_millisecs);
3140
3141 static ssize_t full_scans_show(struct kobject *kobj,
3142                                struct kobj_attribute *attr, char *buf)
3143 {
3144         return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3145 }
3146 KSM_ATTR_RO(full_scans);
3147
3148 static struct attribute *ksm_attrs[] = {
3149         &sleep_millisecs_attr.attr,
3150         &pages_to_scan_attr.attr,
3151         &run_attr.attr,
3152         &pages_shared_attr.attr,
3153         &pages_sharing_attr.attr,
3154         &pages_unshared_attr.attr,
3155         &pages_volatile_attr.attr,
3156         &full_scans_attr.attr,
3157 #ifdef CONFIG_NUMA
3158         &merge_across_nodes_attr.attr,
3159 #endif
3160         &max_page_sharing_attr.attr,
3161         &stable_node_chains_attr.attr,
3162         &stable_node_dups_attr.attr,
3163         &stable_node_chains_prune_millisecs_attr.attr,
3164         &use_zero_pages_attr.attr,
3165         NULL,
3166 };
3167
3168 static const struct attribute_group ksm_attr_group = {
3169         .attrs = ksm_attrs,
3170         .name = "ksm",
3171 };
3172 #endif /* CONFIG_SYSFS */
3173
3174 static int __init ksm_init(void)
3175 {
3176         struct task_struct *ksm_thread;
3177         int err;
3178
3179         /* The correct value depends on page size and endianness */
3180         zero_checksum = calc_checksum(ZERO_PAGE(0));
3181         /* Default to false for backwards compatibility */
3182         ksm_use_zero_pages = false;
3183
3184         err = ksm_slab_init();
3185         if (err)
3186                 goto out;
3187
3188         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3189         if (IS_ERR(ksm_thread)) {
3190                 pr_err("ksm: creating kthread failed\n");
3191                 err = PTR_ERR(ksm_thread);
3192                 goto out_free;
3193         }
3194
3195 #ifdef CONFIG_SYSFS
3196         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3197         if (err) {
3198                 pr_err("ksm: register sysfs failed\n");
3199                 kthread_stop(ksm_thread);
3200                 goto out_free;
3201         }
3202 #else
3203         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3204
3205 #endif /* CONFIG_SYSFS */
3206
3207 #ifdef CONFIG_MEMORY_HOTREMOVE
3208         /* There is no significance to this priority 100 */
3209         hotplug_memory_notifier(ksm_memory_callback, 100);
3210 #endif
3211         return 0;
3212
3213 out_free:
3214         ksm_slab_free();
3215 out:
3216         return err;
3217 }
3218 subsys_initcall(ksm_init);