Merge tag 'regmap-v5.18' of git://git.kernel.org/pub/scm/linux/kernel/git/broonie...
[linux-2.6-microblaze.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))
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))
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                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
642                 stable_node->rmap_hlist_len--;
643                 put_anon_vma(rmap_item->anon_vma);
644                 rmap_item->address &= PAGE_MASK;
645                 cond_resched();
646         }
647
648         /*
649          * We need the second aligned pointer of the migrate_nodes
650          * list_head to stay clear from the rb_parent_color union
651          * (aligned and different than any node) and also different
652          * from &migrate_nodes. This will verify that future list.h changes
653          * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
654          */
655         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
656         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
657
658         if (stable_node->head == &migrate_nodes)
659                 list_del(&stable_node->list);
660         else
661                 stable_node_dup_del(stable_node);
662         free_stable_node(stable_node);
663 }
664
665 enum get_ksm_page_flags {
666         GET_KSM_PAGE_NOLOCK,
667         GET_KSM_PAGE_LOCK,
668         GET_KSM_PAGE_TRYLOCK
669 };
670
671 /*
672  * get_ksm_page: checks if the page indicated by the stable node
673  * is still its ksm page, despite having held no reference to it.
674  * In which case we can trust the content of the page, and it
675  * returns the gotten page; but if the page has now been zapped,
676  * remove the stale node from the stable tree and return NULL.
677  * But beware, the stable node's page might be being migrated.
678  *
679  * You would expect the stable_node to hold a reference to the ksm page.
680  * But if it increments the page's count, swapping out has to wait for
681  * ksmd to come around again before it can free the page, which may take
682  * seconds or even minutes: much too unresponsive.  So instead we use a
683  * "keyhole reference": access to the ksm page from the stable node peeps
684  * out through its keyhole to see if that page still holds the right key,
685  * pointing back to this stable node.  This relies on freeing a PageAnon
686  * page to reset its page->mapping to NULL, and relies on no other use of
687  * a page to put something that might look like our key in page->mapping.
688  * is on its way to being freed; but it is an anomaly to bear in mind.
689  */
690 static struct page *get_ksm_page(struct stable_node *stable_node,
691                                  enum get_ksm_page_flags flags)
692 {
693         struct page *page;
694         void *expected_mapping;
695         unsigned long kpfn;
696
697         expected_mapping = (void *)((unsigned long)stable_node |
698                                         PAGE_MAPPING_KSM);
699 again:
700         kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
701         page = pfn_to_page(kpfn);
702         if (READ_ONCE(page->mapping) != expected_mapping)
703                 goto stale;
704
705         /*
706          * We cannot do anything with the page while its refcount is 0.
707          * Usually 0 means free, or tail of a higher-order page: in which
708          * case this node is no longer referenced, and should be freed;
709          * however, it might mean that the page is under page_ref_freeze().
710          * The __remove_mapping() case is easy, again the node is now stale;
711          * the same is in reuse_ksm_page() case; but if page is swapcache
712          * in migrate_page_move_mapping(), it might still be our page,
713          * in which case it's essential to keep the node.
714          */
715         while (!get_page_unless_zero(page)) {
716                 /*
717                  * Another check for page->mapping != expected_mapping would
718                  * work here too.  We have chosen the !PageSwapCache test to
719                  * optimize the common case, when the page is or is about to
720                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
721                  * in the ref_freeze section of __remove_mapping(); but Anon
722                  * page->mapping reset to NULL later, in free_pages_prepare().
723                  */
724                 if (!PageSwapCache(page))
725                         goto stale;
726                 cpu_relax();
727         }
728
729         if (READ_ONCE(page->mapping) != expected_mapping) {
730                 put_page(page);
731                 goto stale;
732         }
733
734         if (flags == GET_KSM_PAGE_TRYLOCK) {
735                 if (!trylock_page(page)) {
736                         put_page(page);
737                         return ERR_PTR(-EBUSY);
738                 }
739         } else if (flags == GET_KSM_PAGE_LOCK)
740                 lock_page(page);
741
742         if (flags != GET_KSM_PAGE_NOLOCK) {
743                 if (READ_ONCE(page->mapping) != expected_mapping) {
744                         unlock_page(page);
745                         put_page(page);
746                         goto stale;
747                 }
748         }
749         return page;
750
751 stale:
752         /*
753          * We come here from above when page->mapping or !PageSwapCache
754          * suggests that the node is stale; but it might be under migration.
755          * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(),
756          * before checking whether node->kpfn has been changed.
757          */
758         smp_rmb();
759         if (READ_ONCE(stable_node->kpfn) != kpfn)
760                 goto again;
761         remove_node_from_stable_tree(stable_node);
762         return NULL;
763 }
764
765 /*
766  * Removing rmap_item from stable or unstable tree.
767  * This function will clean the information from the stable/unstable tree.
768  */
769 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
770 {
771         if (rmap_item->address & STABLE_FLAG) {
772                 struct stable_node *stable_node;
773                 struct page *page;
774
775                 stable_node = rmap_item->head;
776                 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
777                 if (!page)
778                         goto out;
779
780                 hlist_del(&rmap_item->hlist);
781                 unlock_page(page);
782                 put_page(page);
783
784                 if (!hlist_empty(&stable_node->hlist))
785                         ksm_pages_sharing--;
786                 else
787                         ksm_pages_shared--;
788                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
789                 stable_node->rmap_hlist_len--;
790
791                 put_anon_vma(rmap_item->anon_vma);
792                 rmap_item->head = NULL;
793                 rmap_item->address &= PAGE_MASK;
794
795         } else if (rmap_item->address & UNSTABLE_FLAG) {
796                 unsigned char age;
797                 /*
798                  * Usually ksmd can and must skip the rb_erase, because
799                  * root_unstable_tree was already reset to RB_ROOT.
800                  * But be careful when an mm is exiting: do the rb_erase
801                  * if this rmap_item was inserted by this scan, rather
802                  * than left over from before.
803                  */
804                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
805                 BUG_ON(age > 1);
806                 if (!age)
807                         rb_erase(&rmap_item->node,
808                                  root_unstable_tree + NUMA(rmap_item->nid));
809                 ksm_pages_unshared--;
810                 rmap_item->address &= PAGE_MASK;
811         }
812 out:
813         cond_resched();         /* we're called from many long loops */
814 }
815
816 static void remove_trailing_rmap_items(struct rmap_item **rmap_list)
817 {
818         while (*rmap_list) {
819                 struct rmap_item *rmap_item = *rmap_list;
820                 *rmap_list = rmap_item->rmap_list;
821                 remove_rmap_item_from_tree(rmap_item);
822                 free_rmap_item(rmap_item);
823         }
824 }
825
826 /*
827  * Though it's very tempting to unmerge rmap_items from stable tree rather
828  * than check every pte of a given vma, the locking doesn't quite work for
829  * that - an rmap_item is assigned to the stable tree after inserting ksm
830  * page and upping mmap_lock.  Nor does it fit with the way we skip dup'ing
831  * rmap_items from parent to child at fork time (so as not to waste time
832  * if exit comes before the next scan reaches it).
833  *
834  * Similarly, although we'd like to remove rmap_items (so updating counts
835  * and freeing memory) when unmerging an area, it's easier to leave that
836  * to the next pass of ksmd - consider, for example, how ksmd might be
837  * in cmp_and_merge_page on one of the rmap_items we would be removing.
838  */
839 static int unmerge_ksm_pages(struct vm_area_struct *vma,
840                              unsigned long start, unsigned long end)
841 {
842         unsigned long addr;
843         int err = 0;
844
845         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
846                 if (ksm_test_exit(vma->vm_mm))
847                         break;
848                 if (signal_pending(current))
849                         err = -ERESTARTSYS;
850                 else
851                         err = break_ksm(vma, addr);
852         }
853         return err;
854 }
855
856 static inline struct stable_node *folio_stable_node(struct folio *folio)
857 {
858         return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL;
859 }
860
861 static inline struct stable_node *page_stable_node(struct page *page)
862 {
863         return folio_stable_node(page_folio(page));
864 }
865
866 static inline void set_page_stable_node(struct page *page,
867                                         struct stable_node *stable_node)
868 {
869         page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
870 }
871
872 #ifdef CONFIG_SYSFS
873 /*
874  * Only called through the sysfs control interface:
875  */
876 static int remove_stable_node(struct stable_node *stable_node)
877 {
878         struct page *page;
879         int err;
880
881         page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
882         if (!page) {
883                 /*
884                  * get_ksm_page did remove_node_from_stable_tree itself.
885                  */
886                 return 0;
887         }
888
889         /*
890          * Page could be still mapped if this races with __mmput() running in
891          * between ksm_exit() and exit_mmap(). Just refuse to let
892          * merge_across_nodes/max_page_sharing be switched.
893          */
894         err = -EBUSY;
895         if (!page_mapped(page)) {
896                 /*
897                  * The stable node did not yet appear stale to get_ksm_page(),
898                  * since that allows for an unmapped ksm page to be recognized
899                  * right up until it is freed; but the node is safe to remove.
900                  * This page might be in a pagevec waiting to be freed,
901                  * or it might be PageSwapCache (perhaps under writeback),
902                  * or it might have been removed from swapcache a moment ago.
903                  */
904                 set_page_stable_node(page, NULL);
905                 remove_node_from_stable_tree(stable_node);
906                 err = 0;
907         }
908
909         unlock_page(page);
910         put_page(page);
911         return err;
912 }
913
914 static int remove_stable_node_chain(struct stable_node *stable_node,
915                                     struct rb_root *root)
916 {
917         struct stable_node *dup;
918         struct hlist_node *hlist_safe;
919
920         if (!is_stable_node_chain(stable_node)) {
921                 VM_BUG_ON(is_stable_node_dup(stable_node));
922                 if (remove_stable_node(stable_node))
923                         return true;
924                 else
925                         return false;
926         }
927
928         hlist_for_each_entry_safe(dup, hlist_safe,
929                                   &stable_node->hlist, hlist_dup) {
930                 VM_BUG_ON(!is_stable_node_dup(dup));
931                 if (remove_stable_node(dup))
932                         return true;
933         }
934         BUG_ON(!hlist_empty(&stable_node->hlist));
935         free_stable_node_chain(stable_node, root);
936         return false;
937 }
938
939 static int remove_all_stable_nodes(void)
940 {
941         struct stable_node *stable_node, *next;
942         int nid;
943         int err = 0;
944
945         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
946                 while (root_stable_tree[nid].rb_node) {
947                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
948                                                 struct stable_node, node);
949                         if (remove_stable_node_chain(stable_node,
950                                                      root_stable_tree + nid)) {
951                                 err = -EBUSY;
952                                 break;  /* proceed to next nid */
953                         }
954                         cond_resched();
955                 }
956         }
957         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
958                 if (remove_stable_node(stable_node))
959                         err = -EBUSY;
960                 cond_resched();
961         }
962         return err;
963 }
964
965 static int unmerge_and_remove_all_rmap_items(void)
966 {
967         struct mm_slot *mm_slot;
968         struct mm_struct *mm;
969         struct vm_area_struct *vma;
970         int err = 0;
971
972         spin_lock(&ksm_mmlist_lock);
973         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
974                                                 struct mm_slot, mm_list);
975         spin_unlock(&ksm_mmlist_lock);
976
977         for (mm_slot = ksm_scan.mm_slot;
978                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
979                 mm = mm_slot->mm;
980                 mmap_read_lock(mm);
981                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
982                         if (ksm_test_exit(mm))
983                                 break;
984                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
985                                 continue;
986                         err = unmerge_ksm_pages(vma,
987                                                 vma->vm_start, vma->vm_end);
988                         if (err)
989                                 goto error;
990                 }
991
992                 remove_trailing_rmap_items(&mm_slot->rmap_list);
993                 mmap_read_unlock(mm);
994
995                 spin_lock(&ksm_mmlist_lock);
996                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
997                                                 struct mm_slot, mm_list);
998                 if (ksm_test_exit(mm)) {
999                         hash_del(&mm_slot->link);
1000                         list_del(&mm_slot->mm_list);
1001                         spin_unlock(&ksm_mmlist_lock);
1002
1003                         free_mm_slot(mm_slot);
1004                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1005                         mmdrop(mm);
1006                 } else
1007                         spin_unlock(&ksm_mmlist_lock);
1008         }
1009
1010         /* Clean up stable nodes, but don't worry if some are still busy */
1011         remove_all_stable_nodes();
1012         ksm_scan.seqnr = 0;
1013         return 0;
1014
1015 error:
1016         mmap_read_unlock(mm);
1017         spin_lock(&ksm_mmlist_lock);
1018         ksm_scan.mm_slot = &ksm_mm_head;
1019         spin_unlock(&ksm_mmlist_lock);
1020         return err;
1021 }
1022 #endif /* CONFIG_SYSFS */
1023
1024 static u32 calc_checksum(struct page *page)
1025 {
1026         u32 checksum;
1027         void *addr = kmap_atomic(page);
1028         checksum = xxhash(addr, PAGE_SIZE, 0);
1029         kunmap_atomic(addr);
1030         return checksum;
1031 }
1032
1033 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1034                               pte_t *orig_pte)
1035 {
1036         struct mm_struct *mm = vma->vm_mm;
1037         struct page_vma_mapped_walk pvmw = {
1038                 .page = page,
1039                 .vma = vma,
1040         };
1041         int swapped;
1042         int err = -EFAULT;
1043         struct mmu_notifier_range range;
1044
1045         pvmw.address = page_address_in_vma(page, vma);
1046         if (pvmw.address == -EFAULT)
1047                 goto out;
1048
1049         BUG_ON(PageTransCompound(page));
1050
1051         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1052                                 pvmw.address,
1053                                 pvmw.address + PAGE_SIZE);
1054         mmu_notifier_invalidate_range_start(&range);
1055
1056         if (!page_vma_mapped_walk(&pvmw))
1057                 goto out_mn;
1058         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1059                 goto out_unlock;
1060
1061         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1062             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1063                                                 mm_tlb_flush_pending(mm)) {
1064                 pte_t entry;
1065
1066                 swapped = PageSwapCache(page);
1067                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1068                 /*
1069                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1070                  * take any lock, therefore the check that we are going to make
1071                  * with the pagecount against the mapcount is racy and
1072                  * O_DIRECT can happen right after the check.
1073                  * So we clear the pte and flush the tlb before the check
1074                  * this assure us that no O_DIRECT can happen after the check
1075                  * or in the middle of the check.
1076                  *
1077                  * No need to notify as we are downgrading page table to read
1078                  * only not changing it to point to a new page.
1079                  *
1080                  * See Documentation/vm/mmu_notifier.rst
1081                  */
1082                 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1083                 /*
1084                  * Check that no O_DIRECT or similar I/O is in progress on the
1085                  * page
1086                  */
1087                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1088                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1089                         goto out_unlock;
1090                 }
1091                 if (pte_dirty(entry))
1092                         set_page_dirty(page);
1093
1094                 if (pte_protnone(entry))
1095                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1096                 else
1097                         entry = pte_mkclean(pte_wrprotect(entry));
1098                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1099         }
1100         *orig_pte = *pvmw.pte;
1101         err = 0;
1102
1103 out_unlock:
1104         page_vma_mapped_walk_done(&pvmw);
1105 out_mn:
1106         mmu_notifier_invalidate_range_end(&range);
1107 out:
1108         return err;
1109 }
1110
1111 /**
1112  * replace_page - replace page in vma by new ksm page
1113  * @vma:      vma that holds the pte pointing to page
1114  * @page:     the page we are replacing by kpage
1115  * @kpage:    the ksm page we replace page by
1116  * @orig_pte: the original value of the pte
1117  *
1118  * Returns 0 on success, -EFAULT on failure.
1119  */
1120 static int replace_page(struct vm_area_struct *vma, struct page *page,
1121                         struct page *kpage, pte_t orig_pte)
1122 {
1123         struct mm_struct *mm = vma->vm_mm;
1124         pmd_t *pmd;
1125         pte_t *ptep;
1126         pte_t newpte;
1127         spinlock_t *ptl;
1128         unsigned long addr;
1129         int err = -EFAULT;
1130         struct mmu_notifier_range range;
1131
1132         addr = page_address_in_vma(page, vma);
1133         if (addr == -EFAULT)
1134                 goto out;
1135
1136         pmd = mm_find_pmd(mm, addr);
1137         if (!pmd)
1138                 goto out;
1139
1140         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1141                                 addr + PAGE_SIZE);
1142         mmu_notifier_invalidate_range_start(&range);
1143
1144         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1145         if (!pte_same(*ptep, orig_pte)) {
1146                 pte_unmap_unlock(ptep, ptl);
1147                 goto out_mn;
1148         }
1149
1150         /*
1151          * No need to check ksm_use_zero_pages here: we can only have a
1152          * zero_page here if ksm_use_zero_pages was enabled already.
1153          */
1154         if (!is_zero_pfn(page_to_pfn(kpage))) {
1155                 get_page(kpage);
1156                 page_add_anon_rmap(kpage, vma, addr, false);
1157                 newpte = mk_pte(kpage, vma->vm_page_prot);
1158         } else {
1159                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1160                                                vma->vm_page_prot));
1161                 /*
1162                  * We're replacing an anonymous page with a zero page, which is
1163                  * not anonymous. We need to do proper accounting otherwise we
1164                  * will get wrong values in /proc, and a BUG message in dmesg
1165                  * when tearing down the mm.
1166                  */
1167                 dec_mm_counter(mm, MM_ANONPAGES);
1168         }
1169
1170         flush_cache_page(vma, addr, pte_pfn(*ptep));
1171         /*
1172          * No need to notify as we are replacing a read only page with another
1173          * read only page with the same content.
1174          *
1175          * See Documentation/vm/mmu_notifier.rst
1176          */
1177         ptep_clear_flush(vma, addr, ptep);
1178         set_pte_at_notify(mm, addr, ptep, newpte);
1179
1180         page_remove_rmap(page, false);
1181         if (!page_mapped(page))
1182                 try_to_free_swap(page);
1183         put_page(page);
1184
1185         pte_unmap_unlock(ptep, ptl);
1186         err = 0;
1187 out_mn:
1188         mmu_notifier_invalidate_range_end(&range);
1189 out:
1190         return err;
1191 }
1192
1193 /*
1194  * try_to_merge_one_page - take two pages and merge them into one
1195  * @vma: the vma that holds the pte pointing to page
1196  * @page: the PageAnon page that we want to replace with kpage
1197  * @kpage: the PageKsm page that we want to map instead of page,
1198  *         or NULL the first time when we want to use page as kpage.
1199  *
1200  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1201  */
1202 static int try_to_merge_one_page(struct vm_area_struct *vma,
1203                                  struct page *page, struct page *kpage)
1204 {
1205         pte_t orig_pte = __pte(0);
1206         int err = -EFAULT;
1207
1208         if (page == kpage)                      /* ksm page forked */
1209                 return 0;
1210
1211         if (!PageAnon(page))
1212                 goto out;
1213
1214         /*
1215          * We need the page lock to read a stable PageSwapCache in
1216          * write_protect_page().  We use trylock_page() instead of
1217          * lock_page() because we don't want to wait here - we
1218          * prefer to continue scanning and merging different pages,
1219          * then come back to this page when it is unlocked.
1220          */
1221         if (!trylock_page(page))
1222                 goto out;
1223
1224         if (PageTransCompound(page)) {
1225                 if (split_huge_page(page))
1226                         goto out_unlock;
1227         }
1228
1229         /*
1230          * If this anonymous page is mapped only here, its pte may need
1231          * to be write-protected.  If it's mapped elsewhere, all of its
1232          * ptes are necessarily already write-protected.  But in either
1233          * case, we need to lock and check page_count is not raised.
1234          */
1235         if (write_protect_page(vma, page, &orig_pte) == 0) {
1236                 if (!kpage) {
1237                         /*
1238                          * While we hold page lock, upgrade page from
1239                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1240                          * stable_tree_insert() will update stable_node.
1241                          */
1242                         set_page_stable_node(page, NULL);
1243                         mark_page_accessed(page);
1244                         /*
1245                          * Page reclaim just frees a clean page with no dirty
1246                          * ptes: make sure that the ksm page would be swapped.
1247                          */
1248                         if (!PageDirty(page))
1249                                 SetPageDirty(page);
1250                         err = 0;
1251                 } else if (pages_identical(page, kpage))
1252                         err = replace_page(vma, page, kpage, orig_pte);
1253         }
1254
1255         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1256                 munlock_vma_page(page);
1257                 if (!PageMlocked(kpage)) {
1258                         unlock_page(page);
1259                         lock_page(kpage);
1260                         mlock_vma_page(kpage);
1261                         page = kpage;           /* for final unlock */
1262                 }
1263         }
1264
1265 out_unlock:
1266         unlock_page(page);
1267 out:
1268         return err;
1269 }
1270
1271 /*
1272  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1273  * but no new kernel page is allocated: kpage must already be a ksm page.
1274  *
1275  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1276  */
1277 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1278                                       struct page *page, struct page *kpage)
1279 {
1280         struct mm_struct *mm = rmap_item->mm;
1281         struct vm_area_struct *vma;
1282         int err = -EFAULT;
1283
1284         mmap_read_lock(mm);
1285         vma = find_mergeable_vma(mm, rmap_item->address);
1286         if (!vma)
1287                 goto out;
1288
1289         err = try_to_merge_one_page(vma, page, kpage);
1290         if (err)
1291                 goto out;
1292
1293         /* Unstable nid is in union with stable anon_vma: remove first */
1294         remove_rmap_item_from_tree(rmap_item);
1295
1296         /* Must get reference to anon_vma while still holding mmap_lock */
1297         rmap_item->anon_vma = vma->anon_vma;
1298         get_anon_vma(vma->anon_vma);
1299 out:
1300         mmap_read_unlock(mm);
1301         return err;
1302 }
1303
1304 /*
1305  * try_to_merge_two_pages - take two identical pages and prepare them
1306  * to be merged into one page.
1307  *
1308  * This function returns the kpage if we successfully merged two identical
1309  * pages into one ksm page, NULL otherwise.
1310  *
1311  * Note that this function upgrades page to ksm page: if one of the pages
1312  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1313  */
1314 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1315                                            struct page *page,
1316                                            struct rmap_item *tree_rmap_item,
1317                                            struct page *tree_page)
1318 {
1319         int err;
1320
1321         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1322         if (!err) {
1323                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1324                                                         tree_page, page);
1325                 /*
1326                  * If that fails, we have a ksm page with only one pte
1327                  * pointing to it: so break it.
1328                  */
1329                 if (err)
1330                         break_cow(rmap_item);
1331         }
1332         return err ? NULL : page;
1333 }
1334
1335 static __always_inline
1336 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1337 {
1338         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1339         /*
1340          * Check that at least one mapping still exists, otherwise
1341          * there's no much point to merge and share with this
1342          * stable_node, as the underlying tree_page of the other
1343          * sharer is going to be freed soon.
1344          */
1345         return stable_node->rmap_hlist_len &&
1346                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1347 }
1348
1349 static __always_inline
1350 bool is_page_sharing_candidate(struct stable_node *stable_node)
1351 {
1352         return __is_page_sharing_candidate(stable_node, 0);
1353 }
1354
1355 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1356                                     struct stable_node **_stable_node,
1357                                     struct rb_root *root,
1358                                     bool prune_stale_stable_nodes)
1359 {
1360         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1361         struct hlist_node *hlist_safe;
1362         struct page *_tree_page, *tree_page = NULL;
1363         int nr = 0;
1364         int found_rmap_hlist_len;
1365
1366         if (!prune_stale_stable_nodes ||
1367             time_before(jiffies, stable_node->chain_prune_time +
1368                         msecs_to_jiffies(
1369                                 ksm_stable_node_chains_prune_millisecs)))
1370                 prune_stale_stable_nodes = false;
1371         else
1372                 stable_node->chain_prune_time = jiffies;
1373
1374         hlist_for_each_entry_safe(dup, hlist_safe,
1375                                   &stable_node->hlist, hlist_dup) {
1376                 cond_resched();
1377                 /*
1378                  * We must walk all stable_node_dup to prune the stale
1379                  * stable nodes during lookup.
1380                  *
1381                  * get_ksm_page can drop the nodes from the
1382                  * stable_node->hlist if they point to freed pages
1383                  * (that's why we do a _safe walk). The "dup"
1384                  * stable_node parameter itself will be freed from
1385                  * under us if it returns NULL.
1386                  */
1387                 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1388                 if (!_tree_page)
1389                         continue;
1390                 nr += 1;
1391                 if (is_page_sharing_candidate(dup)) {
1392                         if (!found ||
1393                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1394                                 if (found)
1395                                         put_page(tree_page);
1396                                 found = dup;
1397                                 found_rmap_hlist_len = found->rmap_hlist_len;
1398                                 tree_page = _tree_page;
1399
1400                                 /* skip put_page for found dup */
1401                                 if (!prune_stale_stable_nodes)
1402                                         break;
1403                                 continue;
1404                         }
1405                 }
1406                 put_page(_tree_page);
1407         }
1408
1409         if (found) {
1410                 /*
1411                  * nr is counting all dups in the chain only if
1412                  * prune_stale_stable_nodes is true, otherwise we may
1413                  * break the loop at nr == 1 even if there are
1414                  * multiple entries.
1415                  */
1416                 if (prune_stale_stable_nodes && nr == 1) {
1417                         /*
1418                          * If there's not just one entry it would
1419                          * corrupt memory, better BUG_ON. In KSM
1420                          * context with no lock held it's not even
1421                          * fatal.
1422                          */
1423                         BUG_ON(stable_node->hlist.first->next);
1424
1425                         /*
1426                          * There's just one entry and it is below the
1427                          * deduplication limit so drop the chain.
1428                          */
1429                         rb_replace_node(&stable_node->node, &found->node,
1430                                         root);
1431                         free_stable_node(stable_node);
1432                         ksm_stable_node_chains--;
1433                         ksm_stable_node_dups--;
1434                         /*
1435                          * NOTE: the caller depends on the stable_node
1436                          * to be equal to stable_node_dup if the chain
1437                          * was collapsed.
1438                          */
1439                         *_stable_node = found;
1440                         /*
1441                          * Just for robustness, as stable_node is
1442                          * otherwise left as a stable pointer, the
1443                          * compiler shall optimize it away at build
1444                          * time.
1445                          */
1446                         stable_node = NULL;
1447                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1448                            __is_page_sharing_candidate(found, 1)) {
1449                         /*
1450                          * If the found stable_node dup can accept one
1451                          * more future merge (in addition to the one
1452                          * that is underway) and is not at the head of
1453                          * the chain, put it there so next search will
1454                          * be quicker in the !prune_stale_stable_nodes
1455                          * case.
1456                          *
1457                          * NOTE: it would be inaccurate to use nr > 1
1458                          * instead of checking the hlist.first pointer
1459                          * directly, because in the
1460                          * prune_stale_stable_nodes case "nr" isn't
1461                          * the position of the found dup in the chain,
1462                          * but the total number of dups in the chain.
1463                          */
1464                         hlist_del(&found->hlist_dup);
1465                         hlist_add_head(&found->hlist_dup,
1466                                        &stable_node->hlist);
1467                 }
1468         }
1469
1470         *_stable_node_dup = found;
1471         return tree_page;
1472 }
1473
1474 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1475                                                struct rb_root *root)
1476 {
1477         if (!is_stable_node_chain(stable_node))
1478                 return stable_node;
1479         if (hlist_empty(&stable_node->hlist)) {
1480                 free_stable_node_chain(stable_node, root);
1481                 return NULL;
1482         }
1483         return hlist_entry(stable_node->hlist.first,
1484                            typeof(*stable_node), hlist_dup);
1485 }
1486
1487 /*
1488  * Like for get_ksm_page, this function can free the *_stable_node and
1489  * *_stable_node_dup if the returned tree_page is NULL.
1490  *
1491  * It can also free and overwrite *_stable_node with the found
1492  * stable_node_dup if the chain is collapsed (in which case
1493  * *_stable_node will be equal to *_stable_node_dup like if the chain
1494  * never existed). It's up to the caller to verify tree_page is not
1495  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1496  *
1497  * *_stable_node_dup is really a second output parameter of this
1498  * function and will be overwritten in all cases, the caller doesn't
1499  * need to initialize it.
1500  */
1501 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1502                                         struct stable_node **_stable_node,
1503                                         struct rb_root *root,
1504                                         bool prune_stale_stable_nodes)
1505 {
1506         struct stable_node *stable_node = *_stable_node;
1507         if (!is_stable_node_chain(stable_node)) {
1508                 if (is_page_sharing_candidate(stable_node)) {
1509                         *_stable_node_dup = stable_node;
1510                         return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1511                 }
1512                 /*
1513                  * _stable_node_dup set to NULL means the stable_node
1514                  * reached the ksm_max_page_sharing limit.
1515                  */
1516                 *_stable_node_dup = NULL;
1517                 return NULL;
1518         }
1519         return stable_node_dup(_stable_node_dup, _stable_node, root,
1520                                prune_stale_stable_nodes);
1521 }
1522
1523 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1524                                                 struct stable_node **s_n,
1525                                                 struct rb_root *root)
1526 {
1527         return __stable_node_chain(s_n_d, s_n, root, true);
1528 }
1529
1530 static __always_inline struct page *chain(struct stable_node **s_n_d,
1531                                           struct stable_node *s_n,
1532                                           struct rb_root *root)
1533 {
1534         struct stable_node *old_stable_node = s_n;
1535         struct page *tree_page;
1536
1537         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1538         /* not pruning dups so s_n cannot have changed */
1539         VM_BUG_ON(s_n != old_stable_node);
1540         return tree_page;
1541 }
1542
1543 /*
1544  * stable_tree_search - search for page inside the stable tree
1545  *
1546  * This function checks if there is a page inside the stable tree
1547  * with identical content to the page that we are scanning right now.
1548  *
1549  * This function returns the stable tree node of identical content if found,
1550  * NULL otherwise.
1551  */
1552 static struct page *stable_tree_search(struct page *page)
1553 {
1554         int nid;
1555         struct rb_root *root;
1556         struct rb_node **new;
1557         struct rb_node *parent;
1558         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1559         struct stable_node *page_node;
1560
1561         page_node = page_stable_node(page);
1562         if (page_node && page_node->head != &migrate_nodes) {
1563                 /* ksm page forked */
1564                 get_page(page);
1565                 return page;
1566         }
1567
1568         nid = get_kpfn_nid(page_to_pfn(page));
1569         root = root_stable_tree + nid;
1570 again:
1571         new = &root->rb_node;
1572         parent = NULL;
1573
1574         while (*new) {
1575                 struct page *tree_page;
1576                 int ret;
1577
1578                 cond_resched();
1579                 stable_node = rb_entry(*new, struct stable_node, node);
1580                 stable_node_any = NULL;
1581                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1582                 /*
1583                  * NOTE: stable_node may have been freed by
1584                  * chain_prune() if the returned stable_node_dup is
1585                  * not NULL. stable_node_dup may have been inserted in
1586                  * the rbtree instead as a regular stable_node (in
1587                  * order to collapse the stable_node chain if a single
1588                  * stable_node dup was found in it). In such case the
1589                  * stable_node is overwritten by the calleee to point
1590                  * to the stable_node_dup that was collapsed in the
1591                  * stable rbtree and stable_node will be equal to
1592                  * stable_node_dup like if the chain never existed.
1593                  */
1594                 if (!stable_node_dup) {
1595                         /*
1596                          * Either all stable_node dups were full in
1597                          * this stable_node chain, or this chain was
1598                          * empty and should be rb_erased.
1599                          */
1600                         stable_node_any = stable_node_dup_any(stable_node,
1601                                                               root);
1602                         if (!stable_node_any) {
1603                                 /* rb_erase just run */
1604                                 goto again;
1605                         }
1606                         /*
1607                          * Take any of the stable_node dups page of
1608                          * this stable_node chain to let the tree walk
1609                          * continue. All KSM pages belonging to the
1610                          * stable_node dups in a stable_node chain
1611                          * have the same content and they're
1612                          * write protected at all times. Any will work
1613                          * fine to continue the walk.
1614                          */
1615                         tree_page = get_ksm_page(stable_node_any,
1616                                                  GET_KSM_PAGE_NOLOCK);
1617                 }
1618                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1619                 if (!tree_page) {
1620                         /*
1621                          * If we walked over a stale stable_node,
1622                          * get_ksm_page() will call rb_erase() and it
1623                          * may rebalance the tree from under us. So
1624                          * restart the search from scratch. Returning
1625                          * NULL would be safe too, but we'd generate
1626                          * false negative insertions just because some
1627                          * stable_node was stale.
1628                          */
1629                         goto again;
1630                 }
1631
1632                 ret = memcmp_pages(page, tree_page);
1633                 put_page(tree_page);
1634
1635                 parent = *new;
1636                 if (ret < 0)
1637                         new = &parent->rb_left;
1638                 else if (ret > 0)
1639                         new = &parent->rb_right;
1640                 else {
1641                         if (page_node) {
1642                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1643                                 /*
1644                                  * Test if the migrated page should be merged
1645                                  * into a stable node dup. If the mapcount is
1646                                  * 1 we can migrate it with another KSM page
1647                                  * without adding it to the chain.
1648                                  */
1649                                 if (page_mapcount(page) > 1)
1650                                         goto chain_append;
1651                         }
1652
1653                         if (!stable_node_dup) {
1654                                 /*
1655                                  * If the stable_node is a chain and
1656                                  * we got a payload match in memcmp
1657                                  * but we cannot merge the scanned
1658                                  * page in any of the existing
1659                                  * stable_node dups because they're
1660                                  * all full, we need to wait the
1661                                  * scanned page to find itself a match
1662                                  * in the unstable tree to create a
1663                                  * brand new KSM page to add later to
1664                                  * the dups of this stable_node.
1665                                  */
1666                                 return NULL;
1667                         }
1668
1669                         /*
1670                          * Lock and unlock the stable_node's page (which
1671                          * might already have been migrated) so that page
1672                          * migration is sure to notice its raised count.
1673                          * It would be more elegant to return stable_node
1674                          * than kpage, but that involves more changes.
1675                          */
1676                         tree_page = get_ksm_page(stable_node_dup,
1677                                                  GET_KSM_PAGE_TRYLOCK);
1678
1679                         if (PTR_ERR(tree_page) == -EBUSY)
1680                                 return ERR_PTR(-EBUSY);
1681
1682                         if (unlikely(!tree_page))
1683                                 /*
1684                                  * The tree may have been rebalanced,
1685                                  * so re-evaluate parent and new.
1686                                  */
1687                                 goto again;
1688                         unlock_page(tree_page);
1689
1690                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1691                             NUMA(stable_node_dup->nid)) {
1692                                 put_page(tree_page);
1693                                 goto replace;
1694                         }
1695                         return tree_page;
1696                 }
1697         }
1698
1699         if (!page_node)
1700                 return NULL;
1701
1702         list_del(&page_node->list);
1703         DO_NUMA(page_node->nid = nid);
1704         rb_link_node(&page_node->node, parent, new);
1705         rb_insert_color(&page_node->node, root);
1706 out:
1707         if (is_page_sharing_candidate(page_node)) {
1708                 get_page(page);
1709                 return page;
1710         } else
1711                 return NULL;
1712
1713 replace:
1714         /*
1715          * If stable_node was a chain and chain_prune collapsed it,
1716          * stable_node has been updated to be the new regular
1717          * stable_node. A collapse of the chain is indistinguishable
1718          * from the case there was no chain in the stable
1719          * rbtree. Otherwise stable_node is the chain and
1720          * stable_node_dup is the dup to replace.
1721          */
1722         if (stable_node_dup == stable_node) {
1723                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1724                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1725                 /* there is no chain */
1726                 if (page_node) {
1727                         VM_BUG_ON(page_node->head != &migrate_nodes);
1728                         list_del(&page_node->list);
1729                         DO_NUMA(page_node->nid = nid);
1730                         rb_replace_node(&stable_node_dup->node,
1731                                         &page_node->node,
1732                                         root);
1733                         if (is_page_sharing_candidate(page_node))
1734                                 get_page(page);
1735                         else
1736                                 page = NULL;
1737                 } else {
1738                         rb_erase(&stable_node_dup->node, root);
1739                         page = NULL;
1740                 }
1741         } else {
1742                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1743                 __stable_node_dup_del(stable_node_dup);
1744                 if (page_node) {
1745                         VM_BUG_ON(page_node->head != &migrate_nodes);
1746                         list_del(&page_node->list);
1747                         DO_NUMA(page_node->nid = nid);
1748                         stable_node_chain_add_dup(page_node, stable_node);
1749                         if (is_page_sharing_candidate(page_node))
1750                                 get_page(page);
1751                         else
1752                                 page = NULL;
1753                 } else {
1754                         page = NULL;
1755                 }
1756         }
1757         stable_node_dup->head = &migrate_nodes;
1758         list_add(&stable_node_dup->list, stable_node_dup->head);
1759         return page;
1760
1761 chain_append:
1762         /* stable_node_dup could be null if it reached the limit */
1763         if (!stable_node_dup)
1764                 stable_node_dup = stable_node_any;
1765         /*
1766          * If stable_node was a chain and chain_prune collapsed it,
1767          * stable_node has been updated to be the new regular
1768          * stable_node. A collapse of the chain is indistinguishable
1769          * from the case there was no chain in the stable
1770          * rbtree. Otherwise stable_node is the chain and
1771          * stable_node_dup is the dup to replace.
1772          */
1773         if (stable_node_dup == stable_node) {
1774                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1775                 /* chain is missing so create it */
1776                 stable_node = alloc_stable_node_chain(stable_node_dup,
1777                                                       root);
1778                 if (!stable_node)
1779                         return NULL;
1780         }
1781         /*
1782          * Add this stable_node dup that was
1783          * migrated to the stable_node chain
1784          * of the current nid for this page
1785          * content.
1786          */
1787         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1788         VM_BUG_ON(page_node->head != &migrate_nodes);
1789         list_del(&page_node->list);
1790         DO_NUMA(page_node->nid = nid);
1791         stable_node_chain_add_dup(page_node, stable_node);
1792         goto out;
1793 }
1794
1795 /*
1796  * stable_tree_insert - insert stable tree node pointing to new ksm page
1797  * into the stable tree.
1798  *
1799  * This function returns the stable tree node just allocated on success,
1800  * NULL otherwise.
1801  */
1802 static struct stable_node *stable_tree_insert(struct page *kpage)
1803 {
1804         int nid;
1805         unsigned long kpfn;
1806         struct rb_root *root;
1807         struct rb_node **new;
1808         struct rb_node *parent;
1809         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1810         bool need_chain = false;
1811
1812         kpfn = page_to_pfn(kpage);
1813         nid = get_kpfn_nid(kpfn);
1814         root = root_stable_tree + nid;
1815 again:
1816         parent = NULL;
1817         new = &root->rb_node;
1818
1819         while (*new) {
1820                 struct page *tree_page;
1821                 int ret;
1822
1823                 cond_resched();
1824                 stable_node = rb_entry(*new, struct stable_node, node);
1825                 stable_node_any = NULL;
1826                 tree_page = chain(&stable_node_dup, stable_node, root);
1827                 if (!stable_node_dup) {
1828                         /*
1829                          * Either all stable_node dups were full in
1830                          * this stable_node chain, or this chain was
1831                          * empty and should be rb_erased.
1832                          */
1833                         stable_node_any = stable_node_dup_any(stable_node,
1834                                                               root);
1835                         if (!stable_node_any) {
1836                                 /* rb_erase just run */
1837                                 goto again;
1838                         }
1839                         /*
1840                          * Take any of the stable_node dups page of
1841                          * this stable_node chain to let the tree walk
1842                          * continue. All KSM pages belonging to the
1843                          * stable_node dups in a stable_node chain
1844                          * have the same content and they're
1845                          * write protected at all times. Any will work
1846                          * fine to continue the walk.
1847                          */
1848                         tree_page = get_ksm_page(stable_node_any,
1849                                                  GET_KSM_PAGE_NOLOCK);
1850                 }
1851                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1852                 if (!tree_page) {
1853                         /*
1854                          * If we walked over a stale stable_node,
1855                          * get_ksm_page() will call rb_erase() and it
1856                          * may rebalance the tree from under us. So
1857                          * restart the search from scratch. Returning
1858                          * NULL would be safe too, but we'd generate
1859                          * false negative insertions just because some
1860                          * stable_node was stale.
1861                          */
1862                         goto again;
1863                 }
1864
1865                 ret = memcmp_pages(kpage, tree_page);
1866                 put_page(tree_page);
1867
1868                 parent = *new;
1869                 if (ret < 0)
1870                         new = &parent->rb_left;
1871                 else if (ret > 0)
1872                         new = &parent->rb_right;
1873                 else {
1874                         need_chain = true;
1875                         break;
1876                 }
1877         }
1878
1879         stable_node_dup = alloc_stable_node();
1880         if (!stable_node_dup)
1881                 return NULL;
1882
1883         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1884         stable_node_dup->kpfn = kpfn;
1885         set_page_stable_node(kpage, stable_node_dup);
1886         stable_node_dup->rmap_hlist_len = 0;
1887         DO_NUMA(stable_node_dup->nid = nid);
1888         if (!need_chain) {
1889                 rb_link_node(&stable_node_dup->node, parent, new);
1890                 rb_insert_color(&stable_node_dup->node, root);
1891         } else {
1892                 if (!is_stable_node_chain(stable_node)) {
1893                         struct stable_node *orig = stable_node;
1894                         /* chain is missing so create it */
1895                         stable_node = alloc_stable_node_chain(orig, root);
1896                         if (!stable_node) {
1897                                 free_stable_node(stable_node_dup);
1898                                 return NULL;
1899                         }
1900                 }
1901                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1902         }
1903
1904         return stable_node_dup;
1905 }
1906
1907 /*
1908  * unstable_tree_search_insert - search for identical page,
1909  * else insert rmap_item into the unstable tree.
1910  *
1911  * This function searches for a page in the unstable tree identical to the
1912  * page currently being scanned; and if no identical page is found in the
1913  * tree, we insert rmap_item as a new object into the unstable tree.
1914  *
1915  * This function returns pointer to rmap_item found to be identical
1916  * to the currently scanned page, NULL otherwise.
1917  *
1918  * This function does both searching and inserting, because they share
1919  * the same walking algorithm in an rbtree.
1920  */
1921 static
1922 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1923                                               struct page *page,
1924                                               struct page **tree_pagep)
1925 {
1926         struct rb_node **new;
1927         struct rb_root *root;
1928         struct rb_node *parent = NULL;
1929         int nid;
1930
1931         nid = get_kpfn_nid(page_to_pfn(page));
1932         root = root_unstable_tree + nid;
1933         new = &root->rb_node;
1934
1935         while (*new) {
1936                 struct rmap_item *tree_rmap_item;
1937                 struct page *tree_page;
1938                 int ret;
1939
1940                 cond_resched();
1941                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1942                 tree_page = get_mergeable_page(tree_rmap_item);
1943                 if (!tree_page)
1944                         return NULL;
1945
1946                 /*
1947                  * Don't substitute a ksm page for a forked page.
1948                  */
1949                 if (page == tree_page) {
1950                         put_page(tree_page);
1951                         return NULL;
1952                 }
1953
1954                 ret = memcmp_pages(page, tree_page);
1955
1956                 parent = *new;
1957                 if (ret < 0) {
1958                         put_page(tree_page);
1959                         new = &parent->rb_left;
1960                 } else if (ret > 0) {
1961                         put_page(tree_page);
1962                         new = &parent->rb_right;
1963                 } else if (!ksm_merge_across_nodes &&
1964                            page_to_nid(tree_page) != nid) {
1965                         /*
1966                          * If tree_page has been migrated to another NUMA node,
1967                          * it will be flushed out and put in the right unstable
1968                          * tree next time: only merge with it when across_nodes.
1969                          */
1970                         put_page(tree_page);
1971                         return NULL;
1972                 } else {
1973                         *tree_pagep = tree_page;
1974                         return tree_rmap_item;
1975                 }
1976         }
1977
1978         rmap_item->address |= UNSTABLE_FLAG;
1979         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1980         DO_NUMA(rmap_item->nid = nid);
1981         rb_link_node(&rmap_item->node, parent, new);
1982         rb_insert_color(&rmap_item->node, root);
1983
1984         ksm_pages_unshared++;
1985         return NULL;
1986 }
1987
1988 /*
1989  * stable_tree_append - add another rmap_item to the linked list of
1990  * rmap_items hanging off a given node of the stable tree, all sharing
1991  * the same ksm page.
1992  */
1993 static void stable_tree_append(struct rmap_item *rmap_item,
1994                                struct stable_node *stable_node,
1995                                bool max_page_sharing_bypass)
1996 {
1997         /*
1998          * rmap won't find this mapping if we don't insert the
1999          * rmap_item in the right stable_node
2000          * duplicate. page_migration could break later if rmap breaks,
2001          * so we can as well crash here. We really need to check for
2002          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2003          * for other negative values as an underflow if detected here
2004          * for the first time (and not when decreasing rmap_hlist_len)
2005          * would be sign of memory corruption in the stable_node.
2006          */
2007         BUG_ON(stable_node->rmap_hlist_len < 0);
2008
2009         stable_node->rmap_hlist_len++;
2010         if (!max_page_sharing_bypass)
2011                 /* possibly non fatal but unexpected overflow, only warn */
2012                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2013                              ksm_max_page_sharing);
2014
2015         rmap_item->head = stable_node;
2016         rmap_item->address |= STABLE_FLAG;
2017         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2018
2019         if (rmap_item->hlist.next)
2020                 ksm_pages_sharing++;
2021         else
2022                 ksm_pages_shared++;
2023 }
2024
2025 /*
2026  * cmp_and_merge_page - first see if page can be merged into the stable tree;
2027  * if not, compare checksum to previous and if it's the same, see if page can
2028  * be inserted into the unstable tree, or merged with a page already there and
2029  * both transferred to the stable tree.
2030  *
2031  * @page: the page that we are searching identical page to.
2032  * @rmap_item: the reverse mapping into the virtual address of this page
2033  */
2034 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2035 {
2036         struct mm_struct *mm = rmap_item->mm;
2037         struct rmap_item *tree_rmap_item;
2038         struct page *tree_page = NULL;
2039         struct stable_node *stable_node;
2040         struct page *kpage;
2041         unsigned int checksum;
2042         int err;
2043         bool max_page_sharing_bypass = false;
2044
2045         stable_node = page_stable_node(page);
2046         if (stable_node) {
2047                 if (stable_node->head != &migrate_nodes &&
2048                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2049                     NUMA(stable_node->nid)) {
2050                         stable_node_dup_del(stable_node);
2051                         stable_node->head = &migrate_nodes;
2052                         list_add(&stable_node->list, stable_node->head);
2053                 }
2054                 if (stable_node->head != &migrate_nodes &&
2055                     rmap_item->head == stable_node)
2056                         return;
2057                 /*
2058                  * If it's a KSM fork, allow it to go over the sharing limit
2059                  * without warnings.
2060                  */
2061                 if (!is_page_sharing_candidate(stable_node))
2062                         max_page_sharing_bypass = true;
2063         }
2064
2065         /* We first start with searching the page inside the stable tree */
2066         kpage = stable_tree_search(page);
2067         if (kpage == page && rmap_item->head == stable_node) {
2068                 put_page(kpage);
2069                 return;
2070         }
2071
2072         remove_rmap_item_from_tree(rmap_item);
2073
2074         if (kpage) {
2075                 if (PTR_ERR(kpage) == -EBUSY)
2076                         return;
2077
2078                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2079                 if (!err) {
2080                         /*
2081                          * The page was successfully merged:
2082                          * add its rmap_item to the stable tree.
2083                          */
2084                         lock_page(kpage);
2085                         stable_tree_append(rmap_item, page_stable_node(kpage),
2086                                            max_page_sharing_bypass);
2087                         unlock_page(kpage);
2088                 }
2089                 put_page(kpage);
2090                 return;
2091         }
2092
2093         /*
2094          * If the hash value of the page has changed from the last time
2095          * we calculated it, this page is changing frequently: therefore we
2096          * don't want to insert it in the unstable tree, and we don't want
2097          * to waste our time searching for something identical to it there.
2098          */
2099         checksum = calc_checksum(page);
2100         if (rmap_item->oldchecksum != checksum) {
2101                 rmap_item->oldchecksum = checksum;
2102                 return;
2103         }
2104
2105         /*
2106          * Same checksum as an empty page. We attempt to merge it with the
2107          * appropriate zero page if the user enabled this via sysfs.
2108          */
2109         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2110                 struct vm_area_struct *vma;
2111
2112                 mmap_read_lock(mm);
2113                 vma = find_mergeable_vma(mm, rmap_item->address);
2114                 if (vma) {
2115                         err = try_to_merge_one_page(vma, page,
2116                                         ZERO_PAGE(rmap_item->address));
2117                 } else {
2118                         /*
2119                          * If the vma is out of date, we do not need to
2120                          * continue.
2121                          */
2122                         err = 0;
2123                 }
2124                 mmap_read_unlock(mm);
2125                 /*
2126                  * In case of failure, the page was not really empty, so we
2127                  * need to continue. Otherwise we're done.
2128                  */
2129                 if (!err)
2130                         return;
2131         }
2132         tree_rmap_item =
2133                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2134         if (tree_rmap_item) {
2135                 bool split;
2136
2137                 kpage = try_to_merge_two_pages(rmap_item, page,
2138                                                 tree_rmap_item, tree_page);
2139                 /*
2140                  * If both pages we tried to merge belong to the same compound
2141                  * page, then we actually ended up increasing the reference
2142                  * count of the same compound page twice, and split_huge_page
2143                  * failed.
2144                  * Here we set a flag if that happened, and we use it later to
2145                  * try split_huge_page again. Since we call put_page right
2146                  * afterwards, the reference count will be correct and
2147                  * split_huge_page should succeed.
2148                  */
2149                 split = PageTransCompound(page)
2150                         && compound_head(page) == compound_head(tree_page);
2151                 put_page(tree_page);
2152                 if (kpage) {
2153                         /*
2154                          * The pages were successfully merged: insert new
2155                          * node in the stable tree and add both rmap_items.
2156                          */
2157                         lock_page(kpage);
2158                         stable_node = stable_tree_insert(kpage);
2159                         if (stable_node) {
2160                                 stable_tree_append(tree_rmap_item, stable_node,
2161                                                    false);
2162                                 stable_tree_append(rmap_item, stable_node,
2163                                                    false);
2164                         }
2165                         unlock_page(kpage);
2166
2167                         /*
2168                          * If we fail to insert the page into the stable tree,
2169                          * we will have 2 virtual addresses that are pointing
2170                          * to a ksm page left outside the stable tree,
2171                          * in which case we need to break_cow on both.
2172                          */
2173                         if (!stable_node) {
2174                                 break_cow(tree_rmap_item);
2175                                 break_cow(rmap_item);
2176                         }
2177                 } else if (split) {
2178                         /*
2179                          * We are here if we tried to merge two pages and
2180                          * failed because they both belonged to the same
2181                          * compound page. We will split the page now, but no
2182                          * merging will take place.
2183                          * We do not want to add the cost of a full lock; if
2184                          * the page is locked, it is better to skip it and
2185                          * perhaps try again later.
2186                          */
2187                         if (!trylock_page(page))
2188                                 return;
2189                         split_huge_page(page);
2190                         unlock_page(page);
2191                 }
2192         }
2193 }
2194
2195 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2196                                             struct rmap_item **rmap_list,
2197                                             unsigned long addr)
2198 {
2199         struct rmap_item *rmap_item;
2200
2201         while (*rmap_list) {
2202                 rmap_item = *rmap_list;
2203                 if ((rmap_item->address & PAGE_MASK) == addr)
2204                         return rmap_item;
2205                 if (rmap_item->address > addr)
2206                         break;
2207                 *rmap_list = rmap_item->rmap_list;
2208                 remove_rmap_item_from_tree(rmap_item);
2209                 free_rmap_item(rmap_item);
2210         }
2211
2212         rmap_item = alloc_rmap_item();
2213         if (rmap_item) {
2214                 /* It has already been zeroed */
2215                 rmap_item->mm = mm_slot->mm;
2216                 rmap_item->address = addr;
2217                 rmap_item->rmap_list = *rmap_list;
2218                 *rmap_list = rmap_item;
2219         }
2220         return rmap_item;
2221 }
2222
2223 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2224 {
2225         struct mm_struct *mm;
2226         struct mm_slot *slot;
2227         struct vm_area_struct *vma;
2228         struct rmap_item *rmap_item;
2229         int nid;
2230
2231         if (list_empty(&ksm_mm_head.mm_list))
2232                 return NULL;
2233
2234         slot = ksm_scan.mm_slot;
2235         if (slot == &ksm_mm_head) {
2236                 /*
2237                  * A number of pages can hang around indefinitely on per-cpu
2238                  * pagevecs, raised page count preventing write_protect_page
2239                  * from merging them.  Though it doesn't really matter much,
2240                  * it is puzzling to see some stuck in pages_volatile until
2241                  * other activity jostles them out, and they also prevented
2242                  * LTP's KSM test from succeeding deterministically; so drain
2243                  * them here (here rather than on entry to ksm_do_scan(),
2244                  * so we don't IPI too often when pages_to_scan is set low).
2245                  */
2246                 lru_add_drain_all();
2247
2248                 /*
2249                  * Whereas stale stable_nodes on the stable_tree itself
2250                  * get pruned in the regular course of stable_tree_search(),
2251                  * those moved out to the migrate_nodes list can accumulate:
2252                  * so prune them once before each full scan.
2253                  */
2254                 if (!ksm_merge_across_nodes) {
2255                         struct stable_node *stable_node, *next;
2256                         struct page *page;
2257
2258                         list_for_each_entry_safe(stable_node, next,
2259                                                  &migrate_nodes, list) {
2260                                 page = get_ksm_page(stable_node,
2261                                                     GET_KSM_PAGE_NOLOCK);
2262                                 if (page)
2263                                         put_page(page);
2264                                 cond_resched();
2265                         }
2266                 }
2267
2268                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2269                         root_unstable_tree[nid] = RB_ROOT;
2270
2271                 spin_lock(&ksm_mmlist_lock);
2272                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2273                 ksm_scan.mm_slot = slot;
2274                 spin_unlock(&ksm_mmlist_lock);
2275                 /*
2276                  * Although we tested list_empty() above, a racing __ksm_exit
2277                  * of the last mm on the list may have removed it since then.
2278                  */
2279                 if (slot == &ksm_mm_head)
2280                         return NULL;
2281 next_mm:
2282                 ksm_scan.address = 0;
2283                 ksm_scan.rmap_list = &slot->rmap_list;
2284         }
2285
2286         mm = slot->mm;
2287         mmap_read_lock(mm);
2288         if (ksm_test_exit(mm))
2289                 vma = NULL;
2290         else
2291                 vma = find_vma(mm, ksm_scan.address);
2292
2293         for (; vma; vma = vma->vm_next) {
2294                 if (!(vma->vm_flags & VM_MERGEABLE))
2295                         continue;
2296                 if (ksm_scan.address < vma->vm_start)
2297                         ksm_scan.address = vma->vm_start;
2298                 if (!vma->anon_vma)
2299                         ksm_scan.address = vma->vm_end;
2300
2301                 while (ksm_scan.address < vma->vm_end) {
2302                         if (ksm_test_exit(mm))
2303                                 break;
2304                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2305                         if (IS_ERR_OR_NULL(*page)) {
2306                                 ksm_scan.address += PAGE_SIZE;
2307                                 cond_resched();
2308                                 continue;
2309                         }
2310                         if (PageAnon(*page)) {
2311                                 flush_anon_page(vma, *page, ksm_scan.address);
2312                                 flush_dcache_page(*page);
2313                                 rmap_item = get_next_rmap_item(slot,
2314                                         ksm_scan.rmap_list, ksm_scan.address);
2315                                 if (rmap_item) {
2316                                         ksm_scan.rmap_list =
2317                                                         &rmap_item->rmap_list;
2318                                         ksm_scan.address += PAGE_SIZE;
2319                                 } else
2320                                         put_page(*page);
2321                                 mmap_read_unlock(mm);
2322                                 return rmap_item;
2323                         }
2324                         put_page(*page);
2325                         ksm_scan.address += PAGE_SIZE;
2326                         cond_resched();
2327                 }
2328         }
2329
2330         if (ksm_test_exit(mm)) {
2331                 ksm_scan.address = 0;
2332                 ksm_scan.rmap_list = &slot->rmap_list;
2333         }
2334         /*
2335          * Nuke all the rmap_items that are above this current rmap:
2336          * because there were no VM_MERGEABLE vmas with such addresses.
2337          */
2338         remove_trailing_rmap_items(ksm_scan.rmap_list);
2339
2340         spin_lock(&ksm_mmlist_lock);
2341         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2342                                                 struct mm_slot, mm_list);
2343         if (ksm_scan.address == 0) {
2344                 /*
2345                  * We've completed a full scan of all vmas, holding mmap_lock
2346                  * throughout, and found no VM_MERGEABLE: so do the same as
2347                  * __ksm_exit does to remove this mm from all our lists now.
2348                  * This applies either when cleaning up after __ksm_exit
2349                  * (but beware: we can reach here even before __ksm_exit),
2350                  * or when all VM_MERGEABLE areas have been unmapped (and
2351                  * mmap_lock then protects against race with MADV_MERGEABLE).
2352                  */
2353                 hash_del(&slot->link);
2354                 list_del(&slot->mm_list);
2355                 spin_unlock(&ksm_mmlist_lock);
2356
2357                 free_mm_slot(slot);
2358                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2359                 mmap_read_unlock(mm);
2360                 mmdrop(mm);
2361         } else {
2362                 mmap_read_unlock(mm);
2363                 /*
2364                  * mmap_read_unlock(mm) first because after
2365                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2366                  * already have been freed under us by __ksm_exit()
2367                  * because the "mm_slot" is still hashed and
2368                  * ksm_scan.mm_slot doesn't point to it anymore.
2369                  */
2370                 spin_unlock(&ksm_mmlist_lock);
2371         }
2372
2373         /* Repeat until we've completed scanning the whole list */
2374         slot = ksm_scan.mm_slot;
2375         if (slot != &ksm_mm_head)
2376                 goto next_mm;
2377
2378         ksm_scan.seqnr++;
2379         return NULL;
2380 }
2381
2382 /**
2383  * ksm_do_scan  - the ksm scanner main worker function.
2384  * @scan_npages:  number of pages we want to scan before we return.
2385  */
2386 static void ksm_do_scan(unsigned int scan_npages)
2387 {
2388         struct rmap_item *rmap_item;
2389         struct page *page;
2390
2391         while (scan_npages-- && likely(!freezing(current))) {
2392                 cond_resched();
2393                 rmap_item = scan_get_next_rmap_item(&page);
2394                 if (!rmap_item)
2395                         return;
2396                 cmp_and_merge_page(page, rmap_item);
2397                 put_page(page);
2398         }
2399 }
2400
2401 static int ksmd_should_run(void)
2402 {
2403         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2404 }
2405
2406 static int ksm_scan_thread(void *nothing)
2407 {
2408         unsigned int sleep_ms;
2409
2410         set_freezable();
2411         set_user_nice(current, 5);
2412
2413         while (!kthread_should_stop()) {
2414                 mutex_lock(&ksm_thread_mutex);
2415                 wait_while_offlining();
2416                 if (ksmd_should_run())
2417                         ksm_do_scan(ksm_thread_pages_to_scan);
2418                 mutex_unlock(&ksm_thread_mutex);
2419
2420                 try_to_freeze();
2421
2422                 if (ksmd_should_run()) {
2423                         sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2424                         wait_event_interruptible_timeout(ksm_iter_wait,
2425                                 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2426                                 msecs_to_jiffies(sleep_ms));
2427                 } else {
2428                         wait_event_freezable(ksm_thread_wait,
2429                                 ksmd_should_run() || kthread_should_stop());
2430                 }
2431         }
2432         return 0;
2433 }
2434
2435 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2436                 unsigned long end, int advice, unsigned long *vm_flags)
2437 {
2438         struct mm_struct *mm = vma->vm_mm;
2439         int err;
2440
2441         switch (advice) {
2442         case MADV_MERGEABLE:
2443                 /*
2444                  * Be somewhat over-protective for now!
2445                  */
2446                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2447                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2448                                  VM_HUGETLB | VM_MIXEDMAP))
2449                         return 0;               /* just ignore the advice */
2450
2451                 if (vma_is_dax(vma))
2452                         return 0;
2453
2454 #ifdef VM_SAO
2455                 if (*vm_flags & VM_SAO)
2456                         return 0;
2457 #endif
2458 #ifdef VM_SPARC_ADI
2459                 if (*vm_flags & VM_SPARC_ADI)
2460                         return 0;
2461 #endif
2462
2463                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2464                         err = __ksm_enter(mm);
2465                         if (err)
2466                                 return err;
2467                 }
2468
2469                 *vm_flags |= VM_MERGEABLE;
2470                 break;
2471
2472         case MADV_UNMERGEABLE:
2473                 if (!(*vm_flags & VM_MERGEABLE))
2474                         return 0;               /* just ignore the advice */
2475
2476                 if (vma->anon_vma) {
2477                         err = unmerge_ksm_pages(vma, start, end);
2478                         if (err)
2479                                 return err;
2480                 }
2481
2482                 *vm_flags &= ~VM_MERGEABLE;
2483                 break;
2484         }
2485
2486         return 0;
2487 }
2488 EXPORT_SYMBOL_GPL(ksm_madvise);
2489
2490 int __ksm_enter(struct mm_struct *mm)
2491 {
2492         struct mm_slot *mm_slot;
2493         int needs_wakeup;
2494
2495         mm_slot = alloc_mm_slot();
2496         if (!mm_slot)
2497                 return -ENOMEM;
2498
2499         /* Check ksm_run too?  Would need tighter locking */
2500         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2501
2502         spin_lock(&ksm_mmlist_lock);
2503         insert_to_mm_slots_hash(mm, mm_slot);
2504         /*
2505          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2506          * insert just behind the scanning cursor, to let the area settle
2507          * down a little; when fork is followed by immediate exec, we don't
2508          * want ksmd to waste time setting up and tearing down an rmap_list.
2509          *
2510          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2511          * scanning cursor, otherwise KSM pages in newly forked mms will be
2512          * missed: then we might as well insert at the end of the list.
2513          */
2514         if (ksm_run & KSM_RUN_UNMERGE)
2515                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2516         else
2517                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2518         spin_unlock(&ksm_mmlist_lock);
2519
2520         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2521         mmgrab(mm);
2522
2523         if (needs_wakeup)
2524                 wake_up_interruptible(&ksm_thread_wait);
2525
2526         return 0;
2527 }
2528
2529 void __ksm_exit(struct mm_struct *mm)
2530 {
2531         struct mm_slot *mm_slot;
2532         int easy_to_free = 0;
2533
2534         /*
2535          * This process is exiting: if it's straightforward (as is the
2536          * case when ksmd was never running), free mm_slot immediately.
2537          * But if it's at the cursor or has rmap_items linked to it, use
2538          * mmap_lock to synchronize with any break_cows before pagetables
2539          * are freed, and leave the mm_slot on the list for ksmd to free.
2540          * Beware: ksm may already have noticed it exiting and freed the slot.
2541          */
2542
2543         spin_lock(&ksm_mmlist_lock);
2544         mm_slot = get_mm_slot(mm);
2545         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2546                 if (!mm_slot->rmap_list) {
2547                         hash_del(&mm_slot->link);
2548                         list_del(&mm_slot->mm_list);
2549                         easy_to_free = 1;
2550                 } else {
2551                         list_move(&mm_slot->mm_list,
2552                                   &ksm_scan.mm_slot->mm_list);
2553                 }
2554         }
2555         spin_unlock(&ksm_mmlist_lock);
2556
2557         if (easy_to_free) {
2558                 free_mm_slot(mm_slot);
2559                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2560                 mmdrop(mm);
2561         } else if (mm_slot) {
2562                 mmap_write_lock(mm);
2563                 mmap_write_unlock(mm);
2564         }
2565 }
2566
2567 struct page *ksm_might_need_to_copy(struct page *page,
2568                         struct vm_area_struct *vma, unsigned long address)
2569 {
2570         struct anon_vma *anon_vma = page_anon_vma(page);
2571         struct page *new_page;
2572
2573         if (PageKsm(page)) {
2574                 if (page_stable_node(page) &&
2575                     !(ksm_run & KSM_RUN_UNMERGE))
2576                         return page;    /* no need to copy it */
2577         } else if (!anon_vma) {
2578                 return page;            /* no need to copy it */
2579         } else if (page->index == linear_page_index(vma, address) &&
2580                         anon_vma->root == vma->anon_vma->root) {
2581                 return page;            /* still no need to copy it */
2582         }
2583         if (!PageUptodate(page))
2584                 return page;            /* let do_swap_page report the error */
2585
2586         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2587         if (new_page &&
2588             mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) {
2589                 put_page(new_page);
2590                 new_page = NULL;
2591         }
2592         if (new_page) {
2593                 copy_user_highpage(new_page, page, address, vma);
2594
2595                 SetPageDirty(new_page);
2596                 __SetPageUptodate(new_page);
2597                 __SetPageLocked(new_page);
2598         }
2599
2600         return new_page;
2601 }
2602
2603 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2604 {
2605         struct stable_node *stable_node;
2606         struct rmap_item *rmap_item;
2607         int search_new_forks = 0;
2608
2609         VM_BUG_ON_PAGE(!PageKsm(page), page);
2610
2611         /*
2612          * Rely on the page lock to protect against concurrent modifications
2613          * to that page's node of the stable tree.
2614          */
2615         VM_BUG_ON_PAGE(!PageLocked(page), page);
2616
2617         stable_node = page_stable_node(page);
2618         if (!stable_node)
2619                 return;
2620 again:
2621         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2622                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2623                 struct anon_vma_chain *vmac;
2624                 struct vm_area_struct *vma;
2625
2626                 cond_resched();
2627                 anon_vma_lock_read(anon_vma);
2628                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2629                                                0, ULONG_MAX) {
2630                         unsigned long addr;
2631
2632                         cond_resched();
2633                         vma = vmac->vma;
2634
2635                         /* Ignore the stable/unstable/sqnr flags */
2636                         addr = rmap_item->address & PAGE_MASK;
2637
2638                         if (addr < vma->vm_start || addr >= vma->vm_end)
2639                                 continue;
2640                         /*
2641                          * Initially we examine only the vma which covers this
2642                          * rmap_item; but later, if there is still work to do,
2643                          * we examine covering vmas in other mms: in case they
2644                          * were forked from the original since ksmd passed.
2645                          */
2646                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2647                                 continue;
2648
2649                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2650                                 continue;
2651
2652                         if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2653                                 anon_vma_unlock_read(anon_vma);
2654                                 return;
2655                         }
2656                         if (rwc->done && rwc->done(page)) {
2657                                 anon_vma_unlock_read(anon_vma);
2658                                 return;
2659                         }
2660                 }
2661                 anon_vma_unlock_read(anon_vma);
2662         }
2663         if (!search_new_forks++)
2664                 goto again;
2665 }
2666
2667 #ifdef CONFIG_MIGRATION
2668 void folio_migrate_ksm(struct folio *newfolio, struct folio *folio)
2669 {
2670         struct stable_node *stable_node;
2671
2672         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2673         VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio);
2674         VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio);
2675
2676         stable_node = folio_stable_node(folio);
2677         if (stable_node) {
2678                 VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio);
2679                 stable_node->kpfn = folio_pfn(newfolio);
2680                 /*
2681                  * newfolio->mapping was set in advance; now we need smp_wmb()
2682                  * to make sure that the new stable_node->kpfn is visible
2683                  * to get_ksm_page() before it can see that folio->mapping
2684                  * has gone stale (or that folio_test_swapcache has been cleared).
2685                  */
2686                 smp_wmb();
2687                 set_page_stable_node(&folio->page, NULL);
2688         }
2689 }
2690 #endif /* CONFIG_MIGRATION */
2691
2692 #ifdef CONFIG_MEMORY_HOTREMOVE
2693 static void wait_while_offlining(void)
2694 {
2695         while (ksm_run & KSM_RUN_OFFLINE) {
2696                 mutex_unlock(&ksm_thread_mutex);
2697                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2698                             TASK_UNINTERRUPTIBLE);
2699                 mutex_lock(&ksm_thread_mutex);
2700         }
2701 }
2702
2703 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2704                                          unsigned long start_pfn,
2705                                          unsigned long end_pfn)
2706 {
2707         if (stable_node->kpfn >= start_pfn &&
2708             stable_node->kpfn < end_pfn) {
2709                 /*
2710                  * Don't get_ksm_page, page has already gone:
2711                  * which is why we keep kpfn instead of page*
2712                  */
2713                 remove_node_from_stable_tree(stable_node);
2714                 return true;
2715         }
2716         return false;
2717 }
2718
2719 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2720                                            unsigned long start_pfn,
2721                                            unsigned long end_pfn,
2722                                            struct rb_root *root)
2723 {
2724         struct stable_node *dup;
2725         struct hlist_node *hlist_safe;
2726
2727         if (!is_stable_node_chain(stable_node)) {
2728                 VM_BUG_ON(is_stable_node_dup(stable_node));
2729                 return stable_node_dup_remove_range(stable_node, start_pfn,
2730                                                     end_pfn);
2731         }
2732
2733         hlist_for_each_entry_safe(dup, hlist_safe,
2734                                   &stable_node->hlist, hlist_dup) {
2735                 VM_BUG_ON(!is_stable_node_dup(dup));
2736                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2737         }
2738         if (hlist_empty(&stable_node->hlist)) {
2739                 free_stable_node_chain(stable_node, root);
2740                 return true; /* notify caller that tree was rebalanced */
2741         } else
2742                 return false;
2743 }
2744
2745 static void ksm_check_stable_tree(unsigned long start_pfn,
2746                                   unsigned long end_pfn)
2747 {
2748         struct stable_node *stable_node, *next;
2749         struct rb_node *node;
2750         int nid;
2751
2752         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2753                 node = rb_first(root_stable_tree + nid);
2754                 while (node) {
2755                         stable_node = rb_entry(node, struct stable_node, node);
2756                         if (stable_node_chain_remove_range(stable_node,
2757                                                            start_pfn, end_pfn,
2758                                                            root_stable_tree +
2759                                                            nid))
2760                                 node = rb_first(root_stable_tree + nid);
2761                         else
2762                                 node = rb_next(node);
2763                         cond_resched();
2764                 }
2765         }
2766         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2767                 if (stable_node->kpfn >= start_pfn &&
2768                     stable_node->kpfn < end_pfn)
2769                         remove_node_from_stable_tree(stable_node);
2770                 cond_resched();
2771         }
2772 }
2773
2774 static int ksm_memory_callback(struct notifier_block *self,
2775                                unsigned long action, void *arg)
2776 {
2777         struct memory_notify *mn = arg;
2778
2779         switch (action) {
2780         case MEM_GOING_OFFLINE:
2781                 /*
2782                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2783                  * and remove_all_stable_nodes() while memory is going offline:
2784                  * it is unsafe for them to touch the stable tree at this time.
2785                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2786                  * which do not need the ksm_thread_mutex are all safe.
2787                  */
2788                 mutex_lock(&ksm_thread_mutex);
2789                 ksm_run |= KSM_RUN_OFFLINE;
2790                 mutex_unlock(&ksm_thread_mutex);
2791                 break;
2792
2793         case MEM_OFFLINE:
2794                 /*
2795                  * Most of the work is done by page migration; but there might
2796                  * be a few stable_nodes left over, still pointing to struct
2797                  * pages which have been offlined: prune those from the tree,
2798                  * otherwise get_ksm_page() might later try to access a
2799                  * non-existent struct page.
2800                  */
2801                 ksm_check_stable_tree(mn->start_pfn,
2802                                       mn->start_pfn + mn->nr_pages);
2803                 fallthrough;
2804         case MEM_CANCEL_OFFLINE:
2805                 mutex_lock(&ksm_thread_mutex);
2806                 ksm_run &= ~KSM_RUN_OFFLINE;
2807                 mutex_unlock(&ksm_thread_mutex);
2808
2809                 smp_mb();       /* wake_up_bit advises this */
2810                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2811                 break;
2812         }
2813         return NOTIFY_OK;
2814 }
2815 #else
2816 static void wait_while_offlining(void)
2817 {
2818 }
2819 #endif /* CONFIG_MEMORY_HOTREMOVE */
2820
2821 #ifdef CONFIG_SYSFS
2822 /*
2823  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2824  */
2825
2826 #define KSM_ATTR_RO(_name) \
2827         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2828 #define KSM_ATTR(_name) \
2829         static struct kobj_attribute _name##_attr = \
2830                 __ATTR(_name, 0644, _name##_show, _name##_store)
2831
2832 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2833                                     struct kobj_attribute *attr, char *buf)
2834 {
2835         return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
2836 }
2837
2838 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2839                                      struct kobj_attribute *attr,
2840                                      const char *buf, size_t count)
2841 {
2842         unsigned int msecs;
2843         int err;
2844
2845         err = kstrtouint(buf, 10, &msecs);
2846         if (err)
2847                 return -EINVAL;
2848
2849         ksm_thread_sleep_millisecs = msecs;
2850         wake_up_interruptible(&ksm_iter_wait);
2851
2852         return count;
2853 }
2854 KSM_ATTR(sleep_millisecs);
2855
2856 static ssize_t pages_to_scan_show(struct kobject *kobj,
2857                                   struct kobj_attribute *attr, char *buf)
2858 {
2859         return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
2860 }
2861
2862 static ssize_t pages_to_scan_store(struct kobject *kobj,
2863                                    struct kobj_attribute *attr,
2864                                    const char *buf, size_t count)
2865 {
2866         unsigned int nr_pages;
2867         int err;
2868
2869         err = kstrtouint(buf, 10, &nr_pages);
2870         if (err)
2871                 return -EINVAL;
2872
2873         ksm_thread_pages_to_scan = nr_pages;
2874
2875         return count;
2876 }
2877 KSM_ATTR(pages_to_scan);
2878
2879 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2880                         char *buf)
2881 {
2882         return sysfs_emit(buf, "%lu\n", ksm_run);
2883 }
2884
2885 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2886                          const char *buf, size_t count)
2887 {
2888         unsigned int flags;
2889         int err;
2890
2891         err = kstrtouint(buf, 10, &flags);
2892         if (err)
2893                 return -EINVAL;
2894         if (flags > KSM_RUN_UNMERGE)
2895                 return -EINVAL;
2896
2897         /*
2898          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2899          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2900          * breaking COW to free the pages_shared (but leaves mm_slots
2901          * on the list for when ksmd may be set running again).
2902          */
2903
2904         mutex_lock(&ksm_thread_mutex);
2905         wait_while_offlining();
2906         if (ksm_run != flags) {
2907                 ksm_run = flags;
2908                 if (flags & KSM_RUN_UNMERGE) {
2909                         set_current_oom_origin();
2910                         err = unmerge_and_remove_all_rmap_items();
2911                         clear_current_oom_origin();
2912                         if (err) {
2913                                 ksm_run = KSM_RUN_STOP;
2914                                 count = err;
2915                         }
2916                 }
2917         }
2918         mutex_unlock(&ksm_thread_mutex);
2919
2920         if (flags & KSM_RUN_MERGE)
2921                 wake_up_interruptible(&ksm_thread_wait);
2922
2923         return count;
2924 }
2925 KSM_ATTR(run);
2926
2927 #ifdef CONFIG_NUMA
2928 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2929                                        struct kobj_attribute *attr, char *buf)
2930 {
2931         return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
2932 }
2933
2934 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2935                                    struct kobj_attribute *attr,
2936                                    const char *buf, size_t count)
2937 {
2938         int err;
2939         unsigned long knob;
2940
2941         err = kstrtoul(buf, 10, &knob);
2942         if (err)
2943                 return err;
2944         if (knob > 1)
2945                 return -EINVAL;
2946
2947         mutex_lock(&ksm_thread_mutex);
2948         wait_while_offlining();
2949         if (ksm_merge_across_nodes != knob) {
2950                 if (ksm_pages_shared || remove_all_stable_nodes())
2951                         err = -EBUSY;
2952                 else if (root_stable_tree == one_stable_tree) {
2953                         struct rb_root *buf;
2954                         /*
2955                          * This is the first time that we switch away from the
2956                          * default of merging across nodes: must now allocate
2957                          * a buffer to hold as many roots as may be needed.
2958                          * Allocate stable and unstable together:
2959                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2960                          */
2961                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2962                                       GFP_KERNEL);
2963                         /* Let us assume that RB_ROOT is NULL is zero */
2964                         if (!buf)
2965                                 err = -ENOMEM;
2966                         else {
2967                                 root_stable_tree = buf;
2968                                 root_unstable_tree = buf + nr_node_ids;
2969                                 /* Stable tree is empty but not the unstable */
2970                                 root_unstable_tree[0] = one_unstable_tree[0];
2971                         }
2972                 }
2973                 if (!err) {
2974                         ksm_merge_across_nodes = knob;
2975                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2976                 }
2977         }
2978         mutex_unlock(&ksm_thread_mutex);
2979
2980         return err ? err : count;
2981 }
2982 KSM_ATTR(merge_across_nodes);
2983 #endif
2984
2985 static ssize_t use_zero_pages_show(struct kobject *kobj,
2986                                    struct kobj_attribute *attr, char *buf)
2987 {
2988         return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
2989 }
2990 static ssize_t use_zero_pages_store(struct kobject *kobj,
2991                                    struct kobj_attribute *attr,
2992                                    const char *buf, size_t count)
2993 {
2994         int err;
2995         bool value;
2996
2997         err = kstrtobool(buf, &value);
2998         if (err)
2999                 return -EINVAL;
3000
3001         ksm_use_zero_pages = value;
3002
3003         return count;
3004 }
3005 KSM_ATTR(use_zero_pages);
3006
3007 static ssize_t max_page_sharing_show(struct kobject *kobj,
3008                                      struct kobj_attribute *attr, char *buf)
3009 {
3010         return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3011 }
3012
3013 static ssize_t max_page_sharing_store(struct kobject *kobj,
3014                                       struct kobj_attribute *attr,
3015                                       const char *buf, size_t count)
3016 {
3017         int err;
3018         int knob;
3019
3020         err = kstrtoint(buf, 10, &knob);
3021         if (err)
3022                 return err;
3023         /*
3024          * When a KSM page is created it is shared by 2 mappings. This
3025          * being a signed comparison, it implicitly verifies it's not
3026          * negative.
3027          */
3028         if (knob < 2)
3029                 return -EINVAL;
3030
3031         if (READ_ONCE(ksm_max_page_sharing) == knob)
3032                 return count;
3033
3034         mutex_lock(&ksm_thread_mutex);
3035         wait_while_offlining();
3036         if (ksm_max_page_sharing != knob) {
3037                 if (ksm_pages_shared || remove_all_stable_nodes())
3038                         err = -EBUSY;
3039                 else
3040                         ksm_max_page_sharing = knob;
3041         }
3042         mutex_unlock(&ksm_thread_mutex);
3043
3044         return err ? err : count;
3045 }
3046 KSM_ATTR(max_page_sharing);
3047
3048 static ssize_t pages_shared_show(struct kobject *kobj,
3049                                  struct kobj_attribute *attr, char *buf)
3050 {
3051         return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3052 }
3053 KSM_ATTR_RO(pages_shared);
3054
3055 static ssize_t pages_sharing_show(struct kobject *kobj,
3056                                   struct kobj_attribute *attr, char *buf)
3057 {
3058         return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3059 }
3060 KSM_ATTR_RO(pages_sharing);
3061
3062 static ssize_t pages_unshared_show(struct kobject *kobj,
3063                                    struct kobj_attribute *attr, char *buf)
3064 {
3065         return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3066 }
3067 KSM_ATTR_RO(pages_unshared);
3068
3069 static ssize_t pages_volatile_show(struct kobject *kobj,
3070                                    struct kobj_attribute *attr, char *buf)
3071 {
3072         long ksm_pages_volatile;
3073
3074         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3075                                 - ksm_pages_sharing - ksm_pages_unshared;
3076         /*
3077          * It was not worth any locking to calculate that statistic,
3078          * but it might therefore sometimes be negative: conceal that.
3079          */
3080         if (ksm_pages_volatile < 0)
3081                 ksm_pages_volatile = 0;
3082         return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3083 }
3084 KSM_ATTR_RO(pages_volatile);
3085
3086 static ssize_t stable_node_dups_show(struct kobject *kobj,
3087                                      struct kobj_attribute *attr, char *buf)
3088 {
3089         return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3090 }
3091 KSM_ATTR_RO(stable_node_dups);
3092
3093 static ssize_t stable_node_chains_show(struct kobject *kobj,
3094                                        struct kobj_attribute *attr, char *buf)
3095 {
3096         return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3097 }
3098 KSM_ATTR_RO(stable_node_chains);
3099
3100 static ssize_t
3101 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3102                                         struct kobj_attribute *attr,
3103                                         char *buf)
3104 {
3105         return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3106 }
3107
3108 static ssize_t
3109 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3110                                          struct kobj_attribute *attr,
3111                                          const char *buf, size_t count)
3112 {
3113         unsigned int msecs;
3114         int err;
3115
3116         err = kstrtouint(buf, 10, &msecs);
3117         if (err)
3118                 return -EINVAL;
3119
3120         ksm_stable_node_chains_prune_millisecs = msecs;
3121
3122         return count;
3123 }
3124 KSM_ATTR(stable_node_chains_prune_millisecs);
3125
3126 static ssize_t full_scans_show(struct kobject *kobj,
3127                                struct kobj_attribute *attr, char *buf)
3128 {
3129         return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3130 }
3131 KSM_ATTR_RO(full_scans);
3132
3133 static struct attribute *ksm_attrs[] = {
3134         &sleep_millisecs_attr.attr,
3135         &pages_to_scan_attr.attr,
3136         &run_attr.attr,
3137         &pages_shared_attr.attr,
3138         &pages_sharing_attr.attr,
3139         &pages_unshared_attr.attr,
3140         &pages_volatile_attr.attr,
3141         &full_scans_attr.attr,
3142 #ifdef CONFIG_NUMA
3143         &merge_across_nodes_attr.attr,
3144 #endif
3145         &max_page_sharing_attr.attr,
3146         &stable_node_chains_attr.attr,
3147         &stable_node_dups_attr.attr,
3148         &stable_node_chains_prune_millisecs_attr.attr,
3149         &use_zero_pages_attr.attr,
3150         NULL,
3151 };
3152
3153 static const struct attribute_group ksm_attr_group = {
3154         .attrs = ksm_attrs,
3155         .name = "ksm",
3156 };
3157 #endif /* CONFIG_SYSFS */
3158
3159 static int __init ksm_init(void)
3160 {
3161         struct task_struct *ksm_thread;
3162         int err;
3163
3164         /* The correct value depends on page size and endianness */
3165         zero_checksum = calc_checksum(ZERO_PAGE(0));
3166         /* Default to false for backwards compatibility */
3167         ksm_use_zero_pages = false;
3168
3169         err = ksm_slab_init();
3170         if (err)
3171                 goto out;
3172
3173         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3174         if (IS_ERR(ksm_thread)) {
3175                 pr_err("ksm: creating kthread failed\n");
3176                 err = PTR_ERR(ksm_thread);
3177                 goto out_free;
3178         }
3179
3180 #ifdef CONFIG_SYSFS
3181         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3182         if (err) {
3183                 pr_err("ksm: register sysfs failed\n");
3184                 kthread_stop(ksm_thread);
3185                 goto out_free;
3186         }
3187 #else
3188         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3189
3190 #endif /* CONFIG_SYSFS */
3191
3192 #ifdef CONFIG_MEMORY_HOTREMOVE
3193         /* There is no significance to this priority 100 */
3194         hotplug_memory_notifier(ksm_memory_callback, 100);
3195 #endif
3196         return 0;
3197
3198 out_free:
3199         ksm_slab_free();
3200 out:
3201         return err;
3202 }
3203 subsys_initcall(ksm_init);