Merge tag 'arm64-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux
[platform/kernel/linux-starfive.git] / mm / ksm.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Memory merging support.
4  *
5  * This code enables dynamic sharing of identical pages found in different
6  * memory areas, even if they are not shared by fork()
7  *
8  * Copyright (C) 2008-2009 Red Hat, Inc.
9  * Authors:
10  *      Izik Eidus
11  *      Andrea Arcangeli
12  *      Chris Wright
13  *      Hugh Dickins
14  */
15
16 #include <linux/errno.h>
17 #include <linux/mm.h>
18 #include <linux/fs.h>
19 #include <linux/mman.h>
20 #include <linux/sched.h>
21 #include <linux/sched/mm.h>
22 #include <linux/sched/coredump.h>
23 #include <linux/rwsem.h>
24 #include <linux/pagemap.h>
25 #include <linux/rmap.h>
26 #include <linux/spinlock.h>
27 #include <linux/xxhash.h>
28 #include <linux/delay.h>
29 #include <linux/kthread.h>
30 #include <linux/wait.h>
31 #include <linux/slab.h>
32 #include <linux/rbtree.h>
33 #include <linux/memory.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/swap.h>
36 #include <linux/ksm.h>
37 #include <linux/hashtable.h>
38 #include <linux/freezer.h>
39 #include <linux/oom.h>
40 #include <linux/numa.h>
41
42 #include <asm/tlbflush.h>
43 #include "internal.h"
44
45 #ifdef CONFIG_NUMA
46 #define NUMA(x)         (x)
47 #define DO_NUMA(x)      do { (x); } while (0)
48 #else
49 #define NUMA(x)         (0)
50 #define DO_NUMA(x)      do { } while (0)
51 #endif
52
53 /**
54  * DOC: Overview
55  *
56  * A few notes about the KSM scanning process,
57  * to make it easier to understand the data structures below:
58  *
59  * In order to reduce excessive scanning, KSM sorts the memory pages by their
60  * contents into a data structure that holds pointers to the pages' locations.
61  *
62  * Since the contents of the pages may change at any moment, KSM cannot just
63  * insert the pages into a normal sorted tree and expect it to find anything.
64  * Therefore KSM uses two data structures - the stable and the unstable tree.
65  *
66  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
67  * by their contents.  Because each such page is write-protected, searching on
68  * this tree is fully assured to be working (except when pages are unmapped),
69  * and therefore this tree is called the stable tree.
70  *
71  * The stable tree node includes information required for reverse
72  * mapping from a KSM page to virtual addresses that map this page.
73  *
74  * In order to avoid large latencies of the rmap walks on KSM pages,
75  * KSM maintains two types of nodes in the stable tree:
76  *
77  * * the regular nodes that keep the reverse mapping structures in a
78  *   linked list
79  * * the "chains" that link nodes ("dups") that represent the same
80  *   write protected memory content, but each "dup" corresponds to a
81  *   different KSM page copy of that content
82  *
83  * Internally, the regular nodes, "dups" and "chains" are represented
84  * using the same struct stable_node structure.
85  *
86  * In addition to the stable tree, KSM uses a second data structure called the
87  * unstable tree: this tree holds pointers to pages which have been found to
88  * be "unchanged for a period of time".  The unstable tree sorts these pages
89  * by their contents, but since they are not write-protected, KSM cannot rely
90  * upon the unstable tree to work correctly - the unstable tree is liable to
91  * be corrupted as its contents are modified, and so it is called unstable.
92  *
93  * KSM solves this problem by several techniques:
94  *
95  * 1) The unstable tree is flushed every time KSM completes scanning all
96  *    memory areas, and then the tree is rebuilt again from the beginning.
97  * 2) KSM will only insert into the unstable tree, pages whose hash value
98  *    has not changed since the previous scan of all memory areas.
99  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
100  *    colors of the nodes and not on their contents, assuring that even when
101  *    the tree gets "corrupted" it won't get out of balance, so scanning time
102  *    remains the same (also, searching and inserting nodes in an rbtree uses
103  *    the same algorithm, so we have no overhead when we flush and rebuild).
104  * 4) KSM never flushes the stable tree, which means that even if it were to
105  *    take 10 attempts to find a page in the unstable tree, once it is found,
106  *    it is secured in the stable tree.  (When we scan a new page, we first
107  *    compare it against the stable tree, and then against the unstable tree.)
108  *
109  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
110  * stable trees and multiple unstable trees: one of each for each NUMA node.
111  */
112
113 /**
114  * struct mm_slot - ksm information per mm that is being scanned
115  * @link: link to the mm_slots hash list
116  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
117  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
118  * @mm: the mm that this information is valid for
119  */
120 struct mm_slot {
121         struct hlist_node link;
122         struct list_head mm_list;
123         struct rmap_item *rmap_list;
124         struct mm_struct *mm;
125 };
126
127 /**
128  * struct ksm_scan - cursor for scanning
129  * @mm_slot: the current mm_slot we are scanning
130  * @address: the next address inside that to be scanned
131  * @rmap_list: link to the next rmap to be scanned in the rmap_list
132  * @seqnr: count of completed full scans (needed when removing unstable node)
133  *
134  * There is only the one ksm_scan instance of this cursor structure.
135  */
136 struct ksm_scan {
137         struct mm_slot *mm_slot;
138         unsigned long address;
139         struct rmap_item **rmap_list;
140         unsigned long seqnr;
141 };
142
143 /**
144  * struct stable_node - node of the stable rbtree
145  * @node: rb node of this ksm page in the stable tree
146  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
147  * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
148  * @list: linked into migrate_nodes, pending placement in the proper node tree
149  * @hlist: hlist head of rmap_items using this ksm page
150  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
151  * @chain_prune_time: time of the last full garbage collection
152  * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
153  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
154  */
155 struct stable_node {
156         union {
157                 struct rb_node node;    /* when node of stable tree */
158                 struct {                /* when listed for migration */
159                         struct list_head *head;
160                         struct {
161                                 struct hlist_node hlist_dup;
162                                 struct list_head list;
163                         };
164                 };
165         };
166         struct hlist_head hlist;
167         union {
168                 unsigned long kpfn;
169                 unsigned long chain_prune_time;
170         };
171         /*
172          * STABLE_NODE_CHAIN can be any negative number in
173          * rmap_hlist_len negative range, but better not -1 to be able
174          * to reliably detect underflows.
175          */
176 #define STABLE_NODE_CHAIN -1024
177         int rmap_hlist_len;
178 #ifdef CONFIG_NUMA
179         int nid;
180 #endif
181 };
182
183 /**
184  * struct rmap_item - reverse mapping item for virtual addresses
185  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
186  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
187  * @nid: NUMA node id of unstable tree in which linked (may not match page)
188  * @mm: the memory structure this rmap_item is pointing into
189  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
190  * @oldchecksum: previous checksum of the page at that virtual address
191  * @node: rb node of this rmap_item in the unstable tree
192  * @head: pointer to stable_node heading this list in the stable tree
193  * @hlist: link into hlist of rmap_items hanging off that stable_node
194  */
195 struct rmap_item {
196         struct rmap_item *rmap_list;
197         union {
198                 struct anon_vma *anon_vma;      /* when stable */
199 #ifdef CONFIG_NUMA
200                 int nid;                /* when node of unstable tree */
201 #endif
202         };
203         struct mm_struct *mm;
204         unsigned long address;          /* + low bits used for flags below */
205         unsigned int oldchecksum;       /* when unstable */
206         union {
207                 struct rb_node node;    /* when node of unstable tree */
208                 struct {                /* when listed from stable tree */
209                         struct stable_node *head;
210                         struct hlist_node hlist;
211                 };
212         };
213 };
214
215 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
216 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
217 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
218
219 /* The stable and unstable tree heads */
220 static struct rb_root one_stable_tree[1] = { RB_ROOT };
221 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
222 static struct rb_root *root_stable_tree = one_stable_tree;
223 static struct rb_root *root_unstable_tree = one_unstable_tree;
224
225 /* Recently migrated nodes of stable tree, pending proper placement */
226 static LIST_HEAD(migrate_nodes);
227 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
228
229 #define MM_SLOTS_HASH_BITS 10
230 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
231
232 static struct mm_slot ksm_mm_head = {
233         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
234 };
235 static struct ksm_scan ksm_scan = {
236         .mm_slot = &ksm_mm_head,
237 };
238
239 static struct kmem_cache *rmap_item_cache;
240 static struct kmem_cache *stable_node_cache;
241 static struct kmem_cache *mm_slot_cache;
242
243 /* The number of nodes in the stable tree */
244 static unsigned long ksm_pages_shared;
245
246 /* The number of page slots additionally sharing those nodes */
247 static unsigned long ksm_pages_sharing;
248
249 /* The number of nodes in the unstable tree */
250 static unsigned long ksm_pages_unshared;
251
252 /* The number of rmap_items in use: to calculate pages_volatile */
253 static unsigned long ksm_rmap_items;
254
255 /* The number of stable_node chains */
256 static unsigned long ksm_stable_node_chains;
257
258 /* The number of stable_node dups linked to the stable_node chains */
259 static unsigned long ksm_stable_node_dups;
260
261 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
262 static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
263
264 /* Maximum number of page slots sharing a stable node */
265 static int ksm_max_page_sharing = 256;
266
267 /* Number of pages ksmd should scan in one batch */
268 static unsigned int ksm_thread_pages_to_scan = 100;
269
270 /* Milliseconds ksmd should sleep between batches */
271 static unsigned int ksm_thread_sleep_millisecs = 20;
272
273 /* Checksum of an empty (zeroed) page */
274 static unsigned int zero_checksum __read_mostly;
275
276 /* Whether to merge empty (zeroed) pages with actual zero pages */
277 static bool ksm_use_zero_pages __read_mostly;
278
279 #ifdef CONFIG_NUMA
280 /* Zeroed when merging across nodes is not allowed */
281 static unsigned int ksm_merge_across_nodes = 1;
282 static int ksm_nr_node_ids = 1;
283 #else
284 #define ksm_merge_across_nodes  1U
285 #define ksm_nr_node_ids         1
286 #endif
287
288 #define KSM_RUN_STOP    0
289 #define KSM_RUN_MERGE   1
290 #define KSM_RUN_UNMERGE 2
291 #define KSM_RUN_OFFLINE 4
292 static unsigned long ksm_run = KSM_RUN_STOP;
293 static void wait_while_offlining(void);
294
295 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
296 static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
297 static DEFINE_MUTEX(ksm_thread_mutex);
298 static DEFINE_SPINLOCK(ksm_mmlist_lock);
299
300 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
301                 sizeof(struct __struct), __alignof__(struct __struct),\
302                 (__flags), NULL)
303
304 static int __init ksm_slab_init(void)
305 {
306         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
307         if (!rmap_item_cache)
308                 goto out;
309
310         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
311         if (!stable_node_cache)
312                 goto out_free1;
313
314         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
315         if (!mm_slot_cache)
316                 goto out_free2;
317
318         return 0;
319
320 out_free2:
321         kmem_cache_destroy(stable_node_cache);
322 out_free1:
323         kmem_cache_destroy(rmap_item_cache);
324 out:
325         return -ENOMEM;
326 }
327
328 static void __init ksm_slab_free(void)
329 {
330         kmem_cache_destroy(mm_slot_cache);
331         kmem_cache_destroy(stable_node_cache);
332         kmem_cache_destroy(rmap_item_cache);
333         mm_slot_cache = NULL;
334 }
335
336 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
337 {
338         return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
339 }
340
341 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
342 {
343         return dup->head == STABLE_NODE_DUP_HEAD;
344 }
345
346 static inline void stable_node_chain_add_dup(struct stable_node *dup,
347                                              struct stable_node *chain)
348 {
349         VM_BUG_ON(is_stable_node_dup(dup));
350         dup->head = STABLE_NODE_DUP_HEAD;
351         VM_BUG_ON(!is_stable_node_chain(chain));
352         hlist_add_head(&dup->hlist_dup, &chain->hlist);
353         ksm_stable_node_dups++;
354 }
355
356 static inline void __stable_node_dup_del(struct stable_node *dup)
357 {
358         VM_BUG_ON(!is_stable_node_dup(dup));
359         hlist_del(&dup->hlist_dup);
360         ksm_stable_node_dups--;
361 }
362
363 static inline void stable_node_dup_del(struct stable_node *dup)
364 {
365         VM_BUG_ON(is_stable_node_chain(dup));
366         if (is_stable_node_dup(dup))
367                 __stable_node_dup_del(dup);
368         else
369                 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
370 #ifdef CONFIG_DEBUG_VM
371         dup->head = NULL;
372 #endif
373 }
374
375 static inline struct rmap_item *alloc_rmap_item(void)
376 {
377         struct rmap_item *rmap_item;
378
379         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
380                                                 __GFP_NORETRY | __GFP_NOWARN);
381         if (rmap_item)
382                 ksm_rmap_items++;
383         return rmap_item;
384 }
385
386 static inline void free_rmap_item(struct rmap_item *rmap_item)
387 {
388         ksm_rmap_items--;
389         rmap_item->mm = NULL;   /* debug safety */
390         kmem_cache_free(rmap_item_cache, rmap_item);
391 }
392
393 static inline struct stable_node *alloc_stable_node(void)
394 {
395         /*
396          * The allocation can take too long with GFP_KERNEL when memory is under
397          * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
398          * grants access to memory reserves, helping to avoid this problem.
399          */
400         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
401 }
402
403 static inline void free_stable_node(struct stable_node *stable_node)
404 {
405         VM_BUG_ON(stable_node->rmap_hlist_len &&
406                   !is_stable_node_chain(stable_node));
407         kmem_cache_free(stable_node_cache, stable_node);
408 }
409
410 static inline struct mm_slot *alloc_mm_slot(void)
411 {
412         if (!mm_slot_cache)     /* initialization failed */
413                 return NULL;
414         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
415 }
416
417 static inline void free_mm_slot(struct mm_slot *mm_slot)
418 {
419         kmem_cache_free(mm_slot_cache, mm_slot);
420 }
421
422 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
423 {
424         struct mm_slot *slot;
425
426         hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
427                 if (slot->mm == mm)
428                         return slot;
429
430         return NULL;
431 }
432
433 static void insert_to_mm_slots_hash(struct mm_struct *mm,
434                                     struct mm_slot *mm_slot)
435 {
436         mm_slot->mm = mm;
437         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
438 }
439
440 /*
441  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
442  * page tables after it has passed through ksm_exit() - which, if necessary,
443  * takes mmap_lock briefly to serialize against them.  ksm_exit() does not set
444  * a special flag: they can just back out as soon as mm_users goes to zero.
445  * ksm_test_exit() is used throughout to make this test for exit: in some
446  * places for correctness, in some places just to avoid unnecessary work.
447  */
448 static inline bool ksm_test_exit(struct mm_struct *mm)
449 {
450         return atomic_read(&mm->mm_users) == 0;
451 }
452
453 /*
454  * We use break_ksm to break COW on a ksm page: it's a stripped down
455  *
456  *      if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1)
457  *              put_page(page);
458  *
459  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
460  * in case the application has unmapped and remapped mm,addr meanwhile.
461  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
462  * mmap of /dev/mem, where we would not want to touch it.
463  *
464  * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
465  * of the process that owns 'vma'.  We also do not want to enforce
466  * protection keys here anyway.
467  */
468 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
469 {
470         struct page *page;
471         vm_fault_t ret = 0;
472
473         do {
474                 cond_resched();
475                 page = follow_page(vma, addr,
476                                 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
477                 if (IS_ERR_OR_NULL(page))
478                         break;
479                 if (PageKsm(page))
480                         ret = handle_mm_fault(vma, addr,
481                                               FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE,
482                                               NULL);
483                 else
484                         ret = VM_FAULT_WRITE;
485                 put_page(page);
486         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
487         /*
488          * We must loop because handle_mm_fault() may back out if there's
489          * any difficulty e.g. if pte accessed bit gets updated concurrently.
490          *
491          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
492          * COW has been broken, even if the vma does not permit VM_WRITE;
493          * but note that a concurrent fault might break PageKsm for us.
494          *
495          * VM_FAULT_SIGBUS could occur if we race with truncation of the
496          * backing file, which also invalidates anonymous pages: that's
497          * okay, that truncation will have unmapped the PageKsm for us.
498          *
499          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
500          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
501          * current task has TIF_MEMDIE set, and will be OOM killed on return
502          * to user; and ksmd, having no mm, would never be chosen for that.
503          *
504          * But if the mm is in a limited mem_cgroup, then the fault may fail
505          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
506          * even ksmd can fail in this way - though it's usually breaking ksm
507          * just to undo a merge it made a moment before, so unlikely to oom.
508          *
509          * That's a pity: we might therefore have more kernel pages allocated
510          * than we're counting as nodes in the stable tree; but ksm_do_scan
511          * will retry to break_cow on each pass, so should recover the page
512          * in due course.  The important thing is to not let VM_MERGEABLE
513          * be cleared while any such pages might remain in the area.
514          */
515         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
516 }
517
518 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
519                 unsigned long addr)
520 {
521         struct vm_area_struct *vma;
522         if (ksm_test_exit(mm))
523                 return NULL;
524         vma = vma_lookup(mm, addr);
525         if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
526                 return NULL;
527         return vma;
528 }
529
530 static void break_cow(struct rmap_item *rmap_item)
531 {
532         struct mm_struct *mm = rmap_item->mm;
533         unsigned long addr = rmap_item->address;
534         struct vm_area_struct *vma;
535
536         /*
537          * It is not an accident that whenever we want to break COW
538          * to undo, we also need to drop a reference to the anon_vma.
539          */
540         put_anon_vma(rmap_item->anon_vma);
541
542         mmap_read_lock(mm);
543         vma = find_mergeable_vma(mm, addr);
544         if (vma)
545                 break_ksm(vma, addr);
546         mmap_read_unlock(mm);
547 }
548
549 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
550 {
551         struct mm_struct *mm = rmap_item->mm;
552         unsigned long addr = rmap_item->address;
553         struct vm_area_struct *vma;
554         struct page *page;
555
556         mmap_read_lock(mm);
557         vma = find_mergeable_vma(mm, addr);
558         if (!vma)
559                 goto out;
560
561         page = follow_page(vma, addr, FOLL_GET);
562         if (IS_ERR_OR_NULL(page))
563                 goto out;
564         if (PageAnon(page)) {
565                 flush_anon_page(vma, page, addr);
566                 flush_dcache_page(page);
567         } else {
568                 put_page(page);
569 out:
570                 page = NULL;
571         }
572         mmap_read_unlock(mm);
573         return page;
574 }
575
576 /*
577  * This helper is used for getting right index into array of tree roots.
578  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
579  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
580  * every node has its own stable and unstable tree.
581  */
582 static inline int get_kpfn_nid(unsigned long kpfn)
583 {
584         return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
585 }
586
587 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
588                                                    struct rb_root *root)
589 {
590         struct stable_node *chain = alloc_stable_node();
591         VM_BUG_ON(is_stable_node_chain(dup));
592         if (likely(chain)) {
593                 INIT_HLIST_HEAD(&chain->hlist);
594                 chain->chain_prune_time = jiffies;
595                 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
596 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
597                 chain->nid = NUMA_NO_NODE; /* debug */
598 #endif
599                 ksm_stable_node_chains++;
600
601                 /*
602                  * Put the stable node chain in the first dimension of
603                  * the stable tree and at the same time remove the old
604                  * stable node.
605                  */
606                 rb_replace_node(&dup->node, &chain->node, root);
607
608                 /*
609                  * Move the old stable node to the second dimension
610                  * queued in the hlist_dup. The invariant is that all
611                  * dup stable_nodes in the chain->hlist point to pages
612                  * that are write protected and have the exact same
613                  * content.
614                  */
615                 stable_node_chain_add_dup(dup, chain);
616         }
617         return chain;
618 }
619
620 static inline void free_stable_node_chain(struct stable_node *chain,
621                                           struct rb_root *root)
622 {
623         rb_erase(&chain->node, root);
624         free_stable_node(chain);
625         ksm_stable_node_chains--;
626 }
627
628 static void remove_node_from_stable_tree(struct stable_node *stable_node)
629 {
630         struct rmap_item *rmap_item;
631
632         /* check it's not STABLE_NODE_CHAIN or negative */
633         BUG_ON(stable_node->rmap_hlist_len < 0);
634
635         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
636                 if (rmap_item->hlist.next)
637                         ksm_pages_sharing--;
638                 else
639                         ksm_pages_shared--;
640                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
641                 stable_node->rmap_hlist_len--;
642                 put_anon_vma(rmap_item->anon_vma);
643                 rmap_item->address &= PAGE_MASK;
644                 cond_resched();
645         }
646
647         /*
648          * We need the second aligned pointer of the migrate_nodes
649          * list_head to stay clear from the rb_parent_color union
650          * (aligned and different than any node) and also different
651          * from &migrate_nodes. This will verify that future list.h changes
652          * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
653          */
654         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
655         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
656
657         if (stable_node->head == &migrate_nodes)
658                 list_del(&stable_node->list);
659         else
660                 stable_node_dup_del(stable_node);
661         free_stable_node(stable_node);
662 }
663
664 enum get_ksm_page_flags {
665         GET_KSM_PAGE_NOLOCK,
666         GET_KSM_PAGE_LOCK,
667         GET_KSM_PAGE_TRYLOCK
668 };
669
670 /*
671  * get_ksm_page: checks if the page indicated by the stable node
672  * is still its ksm page, despite having held no reference to it.
673  * In which case we can trust the content of the page, and it
674  * returns the gotten page; but if the page has now been zapped,
675  * remove the stale node from the stable tree and return NULL.
676  * But beware, the stable node's page might be being migrated.
677  *
678  * You would expect the stable_node to hold a reference to the ksm page.
679  * But if it increments the page's count, swapping out has to wait for
680  * ksmd to come around again before it can free the page, which may take
681  * seconds or even minutes: much too unresponsive.  So instead we use a
682  * "keyhole reference": access to the ksm page from the stable node peeps
683  * out through its keyhole to see if that page still holds the right key,
684  * pointing back to this stable node.  This relies on freeing a PageAnon
685  * page to reset its page->mapping to NULL, and relies on no other use of
686  * a page to put something that might look like our key in page->mapping.
687  * is on its way to being freed; but it is an anomaly to bear in mind.
688  */
689 static struct page *get_ksm_page(struct stable_node *stable_node,
690                                  enum get_ksm_page_flags flags)
691 {
692         struct page *page;
693         void *expected_mapping;
694         unsigned long kpfn;
695
696         expected_mapping = (void *)((unsigned long)stable_node |
697                                         PAGE_MAPPING_KSM);
698 again:
699         kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
700         page = pfn_to_page(kpfn);
701         if (READ_ONCE(page->mapping) != expected_mapping)
702                 goto stale;
703
704         /*
705          * We cannot do anything with the page while its refcount is 0.
706          * Usually 0 means free, or tail of a higher-order page: in which
707          * case this node is no longer referenced, and should be freed;
708          * however, it might mean that the page is under page_ref_freeze().
709          * The __remove_mapping() case is easy, again the node is now stale;
710          * the same is in reuse_ksm_page() case; but if page is swapcache
711          * in migrate_page_move_mapping(), it might still be our page,
712          * in which case it's essential to keep the node.
713          */
714         while (!get_page_unless_zero(page)) {
715                 /*
716                  * Another check for page->mapping != expected_mapping would
717                  * work here too.  We have chosen the !PageSwapCache test to
718                  * optimize the common case, when the page is or is about to
719                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
720                  * in the ref_freeze section of __remove_mapping(); but Anon
721                  * page->mapping reset to NULL later, in free_pages_prepare().
722                  */
723                 if (!PageSwapCache(page))
724                         goto stale;
725                 cpu_relax();
726         }
727
728         if (READ_ONCE(page->mapping) != expected_mapping) {
729                 put_page(page);
730                 goto stale;
731         }
732
733         if (flags == GET_KSM_PAGE_TRYLOCK) {
734                 if (!trylock_page(page)) {
735                         put_page(page);
736                         return ERR_PTR(-EBUSY);
737                 }
738         } else if (flags == GET_KSM_PAGE_LOCK)
739                 lock_page(page);
740
741         if (flags != GET_KSM_PAGE_NOLOCK) {
742                 if (READ_ONCE(page->mapping) != expected_mapping) {
743                         unlock_page(page);
744                         put_page(page);
745                         goto stale;
746                 }
747         }
748         return page;
749
750 stale:
751         /*
752          * We come here from above when page->mapping or !PageSwapCache
753          * suggests that the node is stale; but it might be under migration.
754          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
755          * before checking whether node->kpfn has been changed.
756          */
757         smp_rmb();
758         if (READ_ONCE(stable_node->kpfn) != kpfn)
759                 goto again;
760         remove_node_from_stable_tree(stable_node);
761         return NULL;
762 }
763
764 /*
765  * Removing rmap_item from stable or unstable tree.
766  * This function will clean the information from the stable/unstable tree.
767  */
768 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
769 {
770         if (rmap_item->address & STABLE_FLAG) {
771                 struct stable_node *stable_node;
772                 struct page *page;
773
774                 stable_node = rmap_item->head;
775                 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
776                 if (!page)
777                         goto out;
778
779                 hlist_del(&rmap_item->hlist);
780                 unlock_page(page);
781                 put_page(page);
782
783                 if (!hlist_empty(&stable_node->hlist))
784                         ksm_pages_sharing--;
785                 else
786                         ksm_pages_shared--;
787                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
788                 stable_node->rmap_hlist_len--;
789
790                 put_anon_vma(rmap_item->anon_vma);
791                 rmap_item->head = NULL;
792                 rmap_item->address &= PAGE_MASK;
793
794         } else if (rmap_item->address & UNSTABLE_FLAG) {
795                 unsigned char age;
796                 /*
797                  * Usually ksmd can and must skip the rb_erase, because
798                  * root_unstable_tree was already reset to RB_ROOT.
799                  * But be careful when an mm is exiting: do the rb_erase
800                  * if this rmap_item was inserted by this scan, rather
801                  * than left over from before.
802                  */
803                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
804                 BUG_ON(age > 1);
805                 if (!age)
806                         rb_erase(&rmap_item->node,
807                                  root_unstable_tree + NUMA(rmap_item->nid));
808                 ksm_pages_unshared--;
809                 rmap_item->address &= PAGE_MASK;
810         }
811 out:
812         cond_resched();         /* we're called from many long loops */
813 }
814
815 static void remove_trailing_rmap_items(struct rmap_item **rmap_list)
816 {
817         while (*rmap_list) {
818                 struct rmap_item *rmap_item = *rmap_list;
819                 *rmap_list = rmap_item->rmap_list;
820                 remove_rmap_item_from_tree(rmap_item);
821                 free_rmap_item(rmap_item);
822         }
823 }
824
825 /*
826  * Though it's very tempting to unmerge rmap_items from stable tree rather
827  * than check every pte of a given vma, the locking doesn't quite work for
828  * that - an rmap_item is assigned to the stable tree after inserting ksm
829  * page and upping mmap_lock.  Nor does it fit with the way we skip dup'ing
830  * rmap_items from parent to child at fork time (so as not to waste time
831  * if exit comes before the next scan reaches it).
832  *
833  * Similarly, although we'd like to remove rmap_items (so updating counts
834  * and freeing memory) when unmerging an area, it's easier to leave that
835  * to the next pass of ksmd - consider, for example, how ksmd might be
836  * in cmp_and_merge_page on one of the rmap_items we would be removing.
837  */
838 static int unmerge_ksm_pages(struct vm_area_struct *vma,
839                              unsigned long start, unsigned long end)
840 {
841         unsigned long addr;
842         int err = 0;
843
844         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
845                 if (ksm_test_exit(vma->vm_mm))
846                         break;
847                 if (signal_pending(current))
848                         err = -ERESTARTSYS;
849                 else
850                         err = break_ksm(vma, addr);
851         }
852         return err;
853 }
854
855 static inline struct stable_node *page_stable_node(struct page *page)
856 {
857         return PageKsm(page) ? page_rmapping(page) : NULL;
858 }
859
860 static inline void set_page_stable_node(struct page *page,
861                                         struct stable_node *stable_node)
862 {
863         page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
864 }
865
866 #ifdef CONFIG_SYSFS
867 /*
868  * Only called through the sysfs control interface:
869  */
870 static int remove_stable_node(struct stable_node *stable_node)
871 {
872         struct page *page;
873         int err;
874
875         page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
876         if (!page) {
877                 /*
878                  * get_ksm_page did remove_node_from_stable_tree itself.
879                  */
880                 return 0;
881         }
882
883         /*
884          * Page could be still mapped if this races with __mmput() running in
885          * between ksm_exit() and exit_mmap(). Just refuse to let
886          * merge_across_nodes/max_page_sharing be switched.
887          */
888         err = -EBUSY;
889         if (!page_mapped(page)) {
890                 /*
891                  * The stable node did not yet appear stale to get_ksm_page(),
892                  * since that allows for an unmapped ksm page to be recognized
893                  * right up until it is freed; but the node is safe to remove.
894                  * This page might be in a pagevec waiting to be freed,
895                  * or it might be PageSwapCache (perhaps under writeback),
896                  * or it might have been removed from swapcache a moment ago.
897                  */
898                 set_page_stable_node(page, NULL);
899                 remove_node_from_stable_tree(stable_node);
900                 err = 0;
901         }
902
903         unlock_page(page);
904         put_page(page);
905         return err;
906 }
907
908 static int remove_stable_node_chain(struct stable_node *stable_node,
909                                     struct rb_root *root)
910 {
911         struct stable_node *dup;
912         struct hlist_node *hlist_safe;
913
914         if (!is_stable_node_chain(stable_node)) {
915                 VM_BUG_ON(is_stable_node_dup(stable_node));
916                 if (remove_stable_node(stable_node))
917                         return true;
918                 else
919                         return false;
920         }
921
922         hlist_for_each_entry_safe(dup, hlist_safe,
923                                   &stable_node->hlist, hlist_dup) {
924                 VM_BUG_ON(!is_stable_node_dup(dup));
925                 if (remove_stable_node(dup))
926                         return true;
927         }
928         BUG_ON(!hlist_empty(&stable_node->hlist));
929         free_stable_node_chain(stable_node, root);
930         return false;
931 }
932
933 static int remove_all_stable_nodes(void)
934 {
935         struct stable_node *stable_node, *next;
936         int nid;
937         int err = 0;
938
939         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
940                 while (root_stable_tree[nid].rb_node) {
941                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
942                                                 struct stable_node, node);
943                         if (remove_stable_node_chain(stable_node,
944                                                      root_stable_tree + nid)) {
945                                 err = -EBUSY;
946                                 break;  /* proceed to next nid */
947                         }
948                         cond_resched();
949                 }
950         }
951         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
952                 if (remove_stable_node(stable_node))
953                         err = -EBUSY;
954                 cond_resched();
955         }
956         return err;
957 }
958
959 static int unmerge_and_remove_all_rmap_items(void)
960 {
961         struct mm_slot *mm_slot;
962         struct mm_struct *mm;
963         struct vm_area_struct *vma;
964         int err = 0;
965
966         spin_lock(&ksm_mmlist_lock);
967         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
968                                                 struct mm_slot, mm_list);
969         spin_unlock(&ksm_mmlist_lock);
970
971         for (mm_slot = ksm_scan.mm_slot;
972                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
973                 mm = mm_slot->mm;
974                 mmap_read_lock(mm);
975                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
976                         if (ksm_test_exit(mm))
977                                 break;
978                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
979                                 continue;
980                         err = unmerge_ksm_pages(vma,
981                                                 vma->vm_start, vma->vm_end);
982                         if (err)
983                                 goto error;
984                 }
985
986                 remove_trailing_rmap_items(&mm_slot->rmap_list);
987                 mmap_read_unlock(mm);
988
989                 spin_lock(&ksm_mmlist_lock);
990                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
991                                                 struct mm_slot, mm_list);
992                 if (ksm_test_exit(mm)) {
993                         hash_del(&mm_slot->link);
994                         list_del(&mm_slot->mm_list);
995                         spin_unlock(&ksm_mmlist_lock);
996
997                         free_mm_slot(mm_slot);
998                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
999                         mmdrop(mm);
1000                 } else
1001                         spin_unlock(&ksm_mmlist_lock);
1002         }
1003
1004         /* Clean up stable nodes, but don't worry if some are still busy */
1005         remove_all_stable_nodes();
1006         ksm_scan.seqnr = 0;
1007         return 0;
1008
1009 error:
1010         mmap_read_unlock(mm);
1011         spin_lock(&ksm_mmlist_lock);
1012         ksm_scan.mm_slot = &ksm_mm_head;
1013         spin_unlock(&ksm_mmlist_lock);
1014         return err;
1015 }
1016 #endif /* CONFIG_SYSFS */
1017
1018 static u32 calc_checksum(struct page *page)
1019 {
1020         u32 checksum;
1021         void *addr = kmap_atomic(page);
1022         checksum = xxhash(addr, PAGE_SIZE, 0);
1023         kunmap_atomic(addr);
1024         return checksum;
1025 }
1026
1027 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1028                               pte_t *orig_pte)
1029 {
1030         struct mm_struct *mm = vma->vm_mm;
1031         struct page_vma_mapped_walk pvmw = {
1032                 .page = page,
1033                 .vma = vma,
1034         };
1035         int swapped;
1036         int err = -EFAULT;
1037         struct mmu_notifier_range range;
1038
1039         pvmw.address = page_address_in_vma(page, vma);
1040         if (pvmw.address == -EFAULT)
1041                 goto out;
1042
1043         BUG_ON(PageTransCompound(page));
1044
1045         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1046                                 pvmw.address,
1047                                 pvmw.address + PAGE_SIZE);
1048         mmu_notifier_invalidate_range_start(&range);
1049
1050         if (!page_vma_mapped_walk(&pvmw))
1051                 goto out_mn;
1052         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1053                 goto out_unlock;
1054
1055         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1056             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1057                                                 mm_tlb_flush_pending(mm)) {
1058                 pte_t entry;
1059
1060                 swapped = PageSwapCache(page);
1061                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1062                 /*
1063                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1064                  * take any lock, therefore the check that we are going to make
1065                  * with the pagecount against the mapcount is racy and
1066                  * O_DIRECT can happen right after the check.
1067                  * So we clear the pte and flush the tlb before the check
1068                  * this assure us that no O_DIRECT can happen after the check
1069                  * or in the middle of the check.
1070                  *
1071                  * No need to notify as we are downgrading page table to read
1072                  * only not changing it to point to a new page.
1073                  *
1074                  * See Documentation/vm/mmu_notifier.rst
1075                  */
1076                 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1077                 /*
1078                  * Check that no O_DIRECT or similar I/O is in progress on the
1079                  * page
1080                  */
1081                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1082                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1083                         goto out_unlock;
1084                 }
1085                 if (pte_dirty(entry))
1086                         set_page_dirty(page);
1087
1088                 if (pte_protnone(entry))
1089                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1090                 else
1091                         entry = pte_mkclean(pte_wrprotect(entry));
1092                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1093         }
1094         *orig_pte = *pvmw.pte;
1095         err = 0;
1096
1097 out_unlock:
1098         page_vma_mapped_walk_done(&pvmw);
1099 out_mn:
1100         mmu_notifier_invalidate_range_end(&range);
1101 out:
1102         return err;
1103 }
1104
1105 /**
1106  * replace_page - replace page in vma by new ksm page
1107  * @vma:      vma that holds the pte pointing to page
1108  * @page:     the page we are replacing by kpage
1109  * @kpage:    the ksm page we replace page by
1110  * @orig_pte: the original value of the pte
1111  *
1112  * Returns 0 on success, -EFAULT on failure.
1113  */
1114 static int replace_page(struct vm_area_struct *vma, struct page *page,
1115                         struct page *kpage, pte_t orig_pte)
1116 {
1117         struct mm_struct *mm = vma->vm_mm;
1118         pmd_t *pmd;
1119         pte_t *ptep;
1120         pte_t newpte;
1121         spinlock_t *ptl;
1122         unsigned long addr;
1123         int err = -EFAULT;
1124         struct mmu_notifier_range range;
1125
1126         addr = page_address_in_vma(page, vma);
1127         if (addr == -EFAULT)
1128                 goto out;
1129
1130         pmd = mm_find_pmd(mm, addr);
1131         if (!pmd)
1132                 goto out;
1133
1134         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1135                                 addr + PAGE_SIZE);
1136         mmu_notifier_invalidate_range_start(&range);
1137
1138         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1139         if (!pte_same(*ptep, orig_pte)) {
1140                 pte_unmap_unlock(ptep, ptl);
1141                 goto out_mn;
1142         }
1143
1144         /*
1145          * No need to check ksm_use_zero_pages here: we can only have a
1146          * zero_page here if ksm_use_zero_pages was enabled already.
1147          */
1148         if (!is_zero_pfn(page_to_pfn(kpage))) {
1149                 get_page(kpage);
1150                 page_add_anon_rmap(kpage, vma, addr, false);
1151                 newpte = mk_pte(kpage, vma->vm_page_prot);
1152         } else {
1153                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1154                                                vma->vm_page_prot));
1155                 /*
1156                  * We're replacing an anonymous page with a zero page, which is
1157                  * not anonymous. We need to do proper accounting otherwise we
1158                  * will get wrong values in /proc, and a BUG message in dmesg
1159                  * when tearing down the mm.
1160                  */
1161                 dec_mm_counter(mm, MM_ANONPAGES);
1162         }
1163
1164         flush_cache_page(vma, addr, pte_pfn(*ptep));
1165         /*
1166          * No need to notify as we are replacing a read only page with another
1167          * read only page with the same content.
1168          *
1169          * See Documentation/vm/mmu_notifier.rst
1170          */
1171         ptep_clear_flush(vma, addr, ptep);
1172         set_pte_at_notify(mm, addr, ptep, newpte);
1173
1174         page_remove_rmap(page, false);
1175         if (!page_mapped(page))
1176                 try_to_free_swap(page);
1177         put_page(page);
1178
1179         pte_unmap_unlock(ptep, ptl);
1180         err = 0;
1181 out_mn:
1182         mmu_notifier_invalidate_range_end(&range);
1183 out:
1184         return err;
1185 }
1186
1187 /*
1188  * try_to_merge_one_page - take two pages and merge them into one
1189  * @vma: the vma that holds the pte pointing to page
1190  * @page: the PageAnon page that we want to replace with kpage
1191  * @kpage: the PageKsm page that we want to map instead of page,
1192  *         or NULL the first time when we want to use page as kpage.
1193  *
1194  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1195  */
1196 static int try_to_merge_one_page(struct vm_area_struct *vma,
1197                                  struct page *page, struct page *kpage)
1198 {
1199         pte_t orig_pte = __pte(0);
1200         int err = -EFAULT;
1201
1202         if (page == kpage)                      /* ksm page forked */
1203                 return 0;
1204
1205         if (!PageAnon(page))
1206                 goto out;
1207
1208         /*
1209          * We need the page lock to read a stable PageSwapCache in
1210          * write_protect_page().  We use trylock_page() instead of
1211          * lock_page() because we don't want to wait here - we
1212          * prefer to continue scanning and merging different pages,
1213          * then come back to this page when it is unlocked.
1214          */
1215         if (!trylock_page(page))
1216                 goto out;
1217
1218         if (PageTransCompound(page)) {
1219                 if (split_huge_page(page))
1220                         goto out_unlock;
1221         }
1222
1223         /*
1224          * If this anonymous page is mapped only here, its pte may need
1225          * to be write-protected.  If it's mapped elsewhere, all of its
1226          * ptes are necessarily already write-protected.  But in either
1227          * case, we need to lock and check page_count is not raised.
1228          */
1229         if (write_protect_page(vma, page, &orig_pte) == 0) {
1230                 if (!kpage) {
1231                         /*
1232                          * While we hold page lock, upgrade page from
1233                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1234                          * stable_tree_insert() will update stable_node.
1235                          */
1236                         set_page_stable_node(page, NULL);
1237                         mark_page_accessed(page);
1238                         /*
1239                          * Page reclaim just frees a clean page with no dirty
1240                          * ptes: make sure that the ksm page would be swapped.
1241                          */
1242                         if (!PageDirty(page))
1243                                 SetPageDirty(page);
1244                         err = 0;
1245                 } else if (pages_identical(page, kpage))
1246                         err = replace_page(vma, page, kpage, orig_pte);
1247         }
1248
1249         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1250                 munlock_vma_page(page);
1251                 if (!PageMlocked(kpage)) {
1252                         unlock_page(page);
1253                         lock_page(kpage);
1254                         mlock_vma_page(kpage);
1255                         page = kpage;           /* for final unlock */
1256                 }
1257         }
1258
1259 out_unlock:
1260         unlock_page(page);
1261 out:
1262         return err;
1263 }
1264
1265 /*
1266  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1267  * but no new kernel page is allocated: kpage must already be a ksm page.
1268  *
1269  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1270  */
1271 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1272                                       struct page *page, struct page *kpage)
1273 {
1274         struct mm_struct *mm = rmap_item->mm;
1275         struct vm_area_struct *vma;
1276         int err = -EFAULT;
1277
1278         mmap_read_lock(mm);
1279         vma = find_mergeable_vma(mm, rmap_item->address);
1280         if (!vma)
1281                 goto out;
1282
1283         err = try_to_merge_one_page(vma, page, kpage);
1284         if (err)
1285                 goto out;
1286
1287         /* Unstable nid is in union with stable anon_vma: remove first */
1288         remove_rmap_item_from_tree(rmap_item);
1289
1290         /* Must get reference to anon_vma while still holding mmap_lock */
1291         rmap_item->anon_vma = vma->anon_vma;
1292         get_anon_vma(vma->anon_vma);
1293 out:
1294         mmap_read_unlock(mm);
1295         return err;
1296 }
1297
1298 /*
1299  * try_to_merge_two_pages - take two identical pages and prepare them
1300  * to be merged into one page.
1301  *
1302  * This function returns the kpage if we successfully merged two identical
1303  * pages into one ksm page, NULL otherwise.
1304  *
1305  * Note that this function upgrades page to ksm page: if one of the pages
1306  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1307  */
1308 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1309                                            struct page *page,
1310                                            struct rmap_item *tree_rmap_item,
1311                                            struct page *tree_page)
1312 {
1313         int err;
1314
1315         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1316         if (!err) {
1317                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1318                                                         tree_page, page);
1319                 /*
1320                  * If that fails, we have a ksm page with only one pte
1321                  * pointing to it: so break it.
1322                  */
1323                 if (err)
1324                         break_cow(rmap_item);
1325         }
1326         return err ? NULL : page;
1327 }
1328
1329 static __always_inline
1330 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1331 {
1332         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1333         /*
1334          * Check that at least one mapping still exists, otherwise
1335          * there's no much point to merge and share with this
1336          * stable_node, as the underlying tree_page of the other
1337          * sharer is going to be freed soon.
1338          */
1339         return stable_node->rmap_hlist_len &&
1340                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1341 }
1342
1343 static __always_inline
1344 bool is_page_sharing_candidate(struct stable_node *stable_node)
1345 {
1346         return __is_page_sharing_candidate(stable_node, 0);
1347 }
1348
1349 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1350                                     struct stable_node **_stable_node,
1351                                     struct rb_root *root,
1352                                     bool prune_stale_stable_nodes)
1353 {
1354         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1355         struct hlist_node *hlist_safe;
1356         struct page *_tree_page, *tree_page = NULL;
1357         int nr = 0;
1358         int found_rmap_hlist_len;
1359
1360         if (!prune_stale_stable_nodes ||
1361             time_before(jiffies, stable_node->chain_prune_time +
1362                         msecs_to_jiffies(
1363                                 ksm_stable_node_chains_prune_millisecs)))
1364                 prune_stale_stable_nodes = false;
1365         else
1366                 stable_node->chain_prune_time = jiffies;
1367
1368         hlist_for_each_entry_safe(dup, hlist_safe,
1369                                   &stable_node->hlist, hlist_dup) {
1370                 cond_resched();
1371                 /*
1372                  * We must walk all stable_node_dup to prune the stale
1373                  * stable nodes during lookup.
1374                  *
1375                  * get_ksm_page can drop the nodes from the
1376                  * stable_node->hlist if they point to freed pages
1377                  * (that's why we do a _safe walk). The "dup"
1378                  * stable_node parameter itself will be freed from
1379                  * under us if it returns NULL.
1380                  */
1381                 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1382                 if (!_tree_page)
1383                         continue;
1384                 nr += 1;
1385                 if (is_page_sharing_candidate(dup)) {
1386                         if (!found ||
1387                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1388                                 if (found)
1389                                         put_page(tree_page);
1390                                 found = dup;
1391                                 found_rmap_hlist_len = found->rmap_hlist_len;
1392                                 tree_page = _tree_page;
1393
1394                                 /* skip put_page for found dup */
1395                                 if (!prune_stale_stable_nodes)
1396                                         break;
1397                                 continue;
1398                         }
1399                 }
1400                 put_page(_tree_page);
1401         }
1402
1403         if (found) {
1404                 /*
1405                  * nr is counting all dups in the chain only if
1406                  * prune_stale_stable_nodes is true, otherwise we may
1407                  * break the loop at nr == 1 even if there are
1408                  * multiple entries.
1409                  */
1410                 if (prune_stale_stable_nodes && nr == 1) {
1411                         /*
1412                          * If there's not just one entry it would
1413                          * corrupt memory, better BUG_ON. In KSM
1414                          * context with no lock held it's not even
1415                          * fatal.
1416                          */
1417                         BUG_ON(stable_node->hlist.first->next);
1418
1419                         /*
1420                          * There's just one entry and it is below the
1421                          * deduplication limit so drop the chain.
1422                          */
1423                         rb_replace_node(&stable_node->node, &found->node,
1424                                         root);
1425                         free_stable_node(stable_node);
1426                         ksm_stable_node_chains--;
1427                         ksm_stable_node_dups--;
1428                         /*
1429                          * NOTE: the caller depends on the stable_node
1430                          * to be equal to stable_node_dup if the chain
1431                          * was collapsed.
1432                          */
1433                         *_stable_node = found;
1434                         /*
1435                          * Just for robustness, as stable_node is
1436                          * otherwise left as a stable pointer, the
1437                          * compiler shall optimize it away at build
1438                          * time.
1439                          */
1440                         stable_node = NULL;
1441                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1442                            __is_page_sharing_candidate(found, 1)) {
1443                         /*
1444                          * If the found stable_node dup can accept one
1445                          * more future merge (in addition to the one
1446                          * that is underway) and is not at the head of
1447                          * the chain, put it there so next search will
1448                          * be quicker in the !prune_stale_stable_nodes
1449                          * case.
1450                          *
1451                          * NOTE: it would be inaccurate to use nr > 1
1452                          * instead of checking the hlist.first pointer
1453                          * directly, because in the
1454                          * prune_stale_stable_nodes case "nr" isn't
1455                          * the position of the found dup in the chain,
1456                          * but the total number of dups in the chain.
1457                          */
1458                         hlist_del(&found->hlist_dup);
1459                         hlist_add_head(&found->hlist_dup,
1460                                        &stable_node->hlist);
1461                 }
1462         }
1463
1464         *_stable_node_dup = found;
1465         return tree_page;
1466 }
1467
1468 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1469                                                struct rb_root *root)
1470 {
1471         if (!is_stable_node_chain(stable_node))
1472                 return stable_node;
1473         if (hlist_empty(&stable_node->hlist)) {
1474                 free_stable_node_chain(stable_node, root);
1475                 return NULL;
1476         }
1477         return hlist_entry(stable_node->hlist.first,
1478                            typeof(*stable_node), hlist_dup);
1479 }
1480
1481 /*
1482  * Like for get_ksm_page, this function can free the *_stable_node and
1483  * *_stable_node_dup if the returned tree_page is NULL.
1484  *
1485  * It can also free and overwrite *_stable_node with the found
1486  * stable_node_dup if the chain is collapsed (in which case
1487  * *_stable_node will be equal to *_stable_node_dup like if the chain
1488  * never existed). It's up to the caller to verify tree_page is not
1489  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1490  *
1491  * *_stable_node_dup is really a second output parameter of this
1492  * function and will be overwritten in all cases, the caller doesn't
1493  * need to initialize it.
1494  */
1495 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1496                                         struct stable_node **_stable_node,
1497                                         struct rb_root *root,
1498                                         bool prune_stale_stable_nodes)
1499 {
1500         struct stable_node *stable_node = *_stable_node;
1501         if (!is_stable_node_chain(stable_node)) {
1502                 if (is_page_sharing_candidate(stable_node)) {
1503                         *_stable_node_dup = stable_node;
1504                         return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1505                 }
1506                 /*
1507                  * _stable_node_dup set to NULL means the stable_node
1508                  * reached the ksm_max_page_sharing limit.
1509                  */
1510                 *_stable_node_dup = NULL;
1511                 return NULL;
1512         }
1513         return stable_node_dup(_stable_node_dup, _stable_node, root,
1514                                prune_stale_stable_nodes);
1515 }
1516
1517 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1518                                                 struct stable_node **s_n,
1519                                                 struct rb_root *root)
1520 {
1521         return __stable_node_chain(s_n_d, s_n, root, true);
1522 }
1523
1524 static __always_inline struct page *chain(struct stable_node **s_n_d,
1525                                           struct stable_node *s_n,
1526                                           struct rb_root *root)
1527 {
1528         struct stable_node *old_stable_node = s_n;
1529         struct page *tree_page;
1530
1531         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1532         /* not pruning dups so s_n cannot have changed */
1533         VM_BUG_ON(s_n != old_stable_node);
1534         return tree_page;
1535 }
1536
1537 /*
1538  * stable_tree_search - search for page inside the stable tree
1539  *
1540  * This function checks if there is a page inside the stable tree
1541  * with identical content to the page that we are scanning right now.
1542  *
1543  * This function returns the stable tree node of identical content if found,
1544  * NULL otherwise.
1545  */
1546 static struct page *stable_tree_search(struct page *page)
1547 {
1548         int nid;
1549         struct rb_root *root;
1550         struct rb_node **new;
1551         struct rb_node *parent;
1552         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1553         struct stable_node *page_node;
1554
1555         page_node = page_stable_node(page);
1556         if (page_node && page_node->head != &migrate_nodes) {
1557                 /* ksm page forked */
1558                 get_page(page);
1559                 return page;
1560         }
1561
1562         nid = get_kpfn_nid(page_to_pfn(page));
1563         root = root_stable_tree + nid;
1564 again:
1565         new = &root->rb_node;
1566         parent = NULL;
1567
1568         while (*new) {
1569                 struct page *tree_page;
1570                 int ret;
1571
1572                 cond_resched();
1573                 stable_node = rb_entry(*new, struct stable_node, node);
1574                 stable_node_any = NULL;
1575                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1576                 /*
1577                  * NOTE: stable_node may have been freed by
1578                  * chain_prune() if the returned stable_node_dup is
1579                  * not NULL. stable_node_dup may have been inserted in
1580                  * the rbtree instead as a regular stable_node (in
1581                  * order to collapse the stable_node chain if a single
1582                  * stable_node dup was found in it). In such case the
1583                  * stable_node is overwritten by the calleee to point
1584                  * to the stable_node_dup that was collapsed in the
1585                  * stable rbtree and stable_node will be equal to
1586                  * stable_node_dup like if the chain never existed.
1587                  */
1588                 if (!stable_node_dup) {
1589                         /*
1590                          * Either all stable_node dups were full in
1591                          * this stable_node chain, or this chain was
1592                          * empty and should be rb_erased.
1593                          */
1594                         stable_node_any = stable_node_dup_any(stable_node,
1595                                                               root);
1596                         if (!stable_node_any) {
1597                                 /* rb_erase just run */
1598                                 goto again;
1599                         }
1600                         /*
1601                          * Take any of the stable_node dups page of
1602                          * this stable_node chain to let the tree walk
1603                          * continue. All KSM pages belonging to the
1604                          * stable_node dups in a stable_node chain
1605                          * have the same content and they're
1606                          * write protected at all times. Any will work
1607                          * fine to continue the walk.
1608                          */
1609                         tree_page = get_ksm_page(stable_node_any,
1610                                                  GET_KSM_PAGE_NOLOCK);
1611                 }
1612                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1613                 if (!tree_page) {
1614                         /*
1615                          * If we walked over a stale stable_node,
1616                          * get_ksm_page() will call rb_erase() and it
1617                          * may rebalance the tree from under us. So
1618                          * restart the search from scratch. Returning
1619                          * NULL would be safe too, but we'd generate
1620                          * false negative insertions just because some
1621                          * stable_node was stale.
1622                          */
1623                         goto again;
1624                 }
1625
1626                 ret = memcmp_pages(page, tree_page);
1627                 put_page(tree_page);
1628
1629                 parent = *new;
1630                 if (ret < 0)
1631                         new = &parent->rb_left;
1632                 else if (ret > 0)
1633                         new = &parent->rb_right;
1634                 else {
1635                         if (page_node) {
1636                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1637                                 /*
1638                                  * Test if the migrated page should be merged
1639                                  * into a stable node dup. If the mapcount is
1640                                  * 1 we can migrate it with another KSM page
1641                                  * without adding it to the chain.
1642                                  */
1643                                 if (page_mapcount(page) > 1)
1644                                         goto chain_append;
1645                         }
1646
1647                         if (!stable_node_dup) {
1648                                 /*
1649                                  * If the stable_node is a chain and
1650                                  * we got a payload match in memcmp
1651                                  * but we cannot merge the scanned
1652                                  * page in any of the existing
1653                                  * stable_node dups because they're
1654                                  * all full, we need to wait the
1655                                  * scanned page to find itself a match
1656                                  * in the unstable tree to create a
1657                                  * brand new KSM page to add later to
1658                                  * the dups of this stable_node.
1659                                  */
1660                                 return NULL;
1661                         }
1662
1663                         /*
1664                          * Lock and unlock the stable_node's page (which
1665                          * might already have been migrated) so that page
1666                          * migration is sure to notice its raised count.
1667                          * It would be more elegant to return stable_node
1668                          * than kpage, but that involves more changes.
1669                          */
1670                         tree_page = get_ksm_page(stable_node_dup,
1671                                                  GET_KSM_PAGE_TRYLOCK);
1672
1673                         if (PTR_ERR(tree_page) == -EBUSY)
1674                                 return ERR_PTR(-EBUSY);
1675
1676                         if (unlikely(!tree_page))
1677                                 /*
1678                                  * The tree may have been rebalanced,
1679                                  * so re-evaluate parent and new.
1680                                  */
1681                                 goto again;
1682                         unlock_page(tree_page);
1683
1684                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1685                             NUMA(stable_node_dup->nid)) {
1686                                 put_page(tree_page);
1687                                 goto replace;
1688                         }
1689                         return tree_page;
1690                 }
1691         }
1692
1693         if (!page_node)
1694                 return NULL;
1695
1696         list_del(&page_node->list);
1697         DO_NUMA(page_node->nid = nid);
1698         rb_link_node(&page_node->node, parent, new);
1699         rb_insert_color(&page_node->node, root);
1700 out:
1701         if (is_page_sharing_candidate(page_node)) {
1702                 get_page(page);
1703                 return page;
1704         } else
1705                 return NULL;
1706
1707 replace:
1708         /*
1709          * If stable_node was a chain and chain_prune collapsed it,
1710          * stable_node has been updated to be the new regular
1711          * stable_node. A collapse of the chain is indistinguishable
1712          * from the case there was no chain in the stable
1713          * rbtree. Otherwise stable_node is the chain and
1714          * stable_node_dup is the dup to replace.
1715          */
1716         if (stable_node_dup == stable_node) {
1717                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1718                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1719                 /* there is no chain */
1720                 if (page_node) {
1721                         VM_BUG_ON(page_node->head != &migrate_nodes);
1722                         list_del(&page_node->list);
1723                         DO_NUMA(page_node->nid = nid);
1724                         rb_replace_node(&stable_node_dup->node,
1725                                         &page_node->node,
1726                                         root);
1727                         if (is_page_sharing_candidate(page_node))
1728                                 get_page(page);
1729                         else
1730                                 page = NULL;
1731                 } else {
1732                         rb_erase(&stable_node_dup->node, root);
1733                         page = NULL;
1734                 }
1735         } else {
1736                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1737                 __stable_node_dup_del(stable_node_dup);
1738                 if (page_node) {
1739                         VM_BUG_ON(page_node->head != &migrate_nodes);
1740                         list_del(&page_node->list);
1741                         DO_NUMA(page_node->nid = nid);
1742                         stable_node_chain_add_dup(page_node, stable_node);
1743                         if (is_page_sharing_candidate(page_node))
1744                                 get_page(page);
1745                         else
1746                                 page = NULL;
1747                 } else {
1748                         page = NULL;
1749                 }
1750         }
1751         stable_node_dup->head = &migrate_nodes;
1752         list_add(&stable_node_dup->list, stable_node_dup->head);
1753         return page;
1754
1755 chain_append:
1756         /* stable_node_dup could be null if it reached the limit */
1757         if (!stable_node_dup)
1758                 stable_node_dup = stable_node_any;
1759         /*
1760          * If stable_node was a chain and chain_prune collapsed it,
1761          * stable_node has been updated to be the new regular
1762          * stable_node. A collapse of the chain is indistinguishable
1763          * from the case there was no chain in the stable
1764          * rbtree. Otherwise stable_node is the chain and
1765          * stable_node_dup is the dup to replace.
1766          */
1767         if (stable_node_dup == stable_node) {
1768                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1769                 /* chain is missing so create it */
1770                 stable_node = alloc_stable_node_chain(stable_node_dup,
1771                                                       root);
1772                 if (!stable_node)
1773                         return NULL;
1774         }
1775         /*
1776          * Add this stable_node dup that was
1777          * migrated to the stable_node chain
1778          * of the current nid for this page
1779          * content.
1780          */
1781         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1782         VM_BUG_ON(page_node->head != &migrate_nodes);
1783         list_del(&page_node->list);
1784         DO_NUMA(page_node->nid = nid);
1785         stable_node_chain_add_dup(page_node, stable_node);
1786         goto out;
1787 }
1788
1789 /*
1790  * stable_tree_insert - insert stable tree node pointing to new ksm page
1791  * into the stable tree.
1792  *
1793  * This function returns the stable tree node just allocated on success,
1794  * NULL otherwise.
1795  */
1796 static struct stable_node *stable_tree_insert(struct page *kpage)
1797 {
1798         int nid;
1799         unsigned long kpfn;
1800         struct rb_root *root;
1801         struct rb_node **new;
1802         struct rb_node *parent;
1803         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1804         bool need_chain = false;
1805
1806         kpfn = page_to_pfn(kpage);
1807         nid = get_kpfn_nid(kpfn);
1808         root = root_stable_tree + nid;
1809 again:
1810         parent = NULL;
1811         new = &root->rb_node;
1812
1813         while (*new) {
1814                 struct page *tree_page;
1815                 int ret;
1816
1817                 cond_resched();
1818                 stable_node = rb_entry(*new, struct stable_node, node);
1819                 stable_node_any = NULL;
1820                 tree_page = chain(&stable_node_dup, stable_node, root);
1821                 if (!stable_node_dup) {
1822                         /*
1823                          * Either all stable_node dups were full in
1824                          * this stable_node chain, or this chain was
1825                          * empty and should be rb_erased.
1826                          */
1827                         stable_node_any = stable_node_dup_any(stable_node,
1828                                                               root);
1829                         if (!stable_node_any) {
1830                                 /* rb_erase just run */
1831                                 goto again;
1832                         }
1833                         /*
1834                          * Take any of the stable_node dups page of
1835                          * this stable_node chain to let the tree walk
1836                          * continue. All KSM pages belonging to the
1837                          * stable_node dups in a stable_node chain
1838                          * have the same content and they're
1839                          * write protected at all times. Any will work
1840                          * fine to continue the walk.
1841                          */
1842                         tree_page = get_ksm_page(stable_node_any,
1843                                                  GET_KSM_PAGE_NOLOCK);
1844                 }
1845                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1846                 if (!tree_page) {
1847                         /*
1848                          * If we walked over a stale stable_node,
1849                          * get_ksm_page() will call rb_erase() and it
1850                          * may rebalance the tree from under us. So
1851                          * restart the search from scratch. Returning
1852                          * NULL would be safe too, but we'd generate
1853                          * false negative insertions just because some
1854                          * stable_node was stale.
1855                          */
1856                         goto again;
1857                 }
1858
1859                 ret = memcmp_pages(kpage, tree_page);
1860                 put_page(tree_page);
1861
1862                 parent = *new;
1863                 if (ret < 0)
1864                         new = &parent->rb_left;
1865                 else if (ret > 0)
1866                         new = &parent->rb_right;
1867                 else {
1868                         need_chain = true;
1869                         break;
1870                 }
1871         }
1872
1873         stable_node_dup = alloc_stable_node();
1874         if (!stable_node_dup)
1875                 return NULL;
1876
1877         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1878         stable_node_dup->kpfn = kpfn;
1879         set_page_stable_node(kpage, stable_node_dup);
1880         stable_node_dup->rmap_hlist_len = 0;
1881         DO_NUMA(stable_node_dup->nid = nid);
1882         if (!need_chain) {
1883                 rb_link_node(&stable_node_dup->node, parent, new);
1884                 rb_insert_color(&stable_node_dup->node, root);
1885         } else {
1886                 if (!is_stable_node_chain(stable_node)) {
1887                         struct stable_node *orig = stable_node;
1888                         /* chain is missing so create it */
1889                         stable_node = alloc_stable_node_chain(orig, root);
1890                         if (!stable_node) {
1891                                 free_stable_node(stable_node_dup);
1892                                 return NULL;
1893                         }
1894                 }
1895                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1896         }
1897
1898         return stable_node_dup;
1899 }
1900
1901 /*
1902  * unstable_tree_search_insert - search for identical page,
1903  * else insert rmap_item into the unstable tree.
1904  *
1905  * This function searches for a page in the unstable tree identical to the
1906  * page currently being scanned; and if no identical page is found in the
1907  * tree, we insert rmap_item as a new object into the unstable tree.
1908  *
1909  * This function returns pointer to rmap_item found to be identical
1910  * to the currently scanned page, NULL otherwise.
1911  *
1912  * This function does both searching and inserting, because they share
1913  * the same walking algorithm in an rbtree.
1914  */
1915 static
1916 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1917                                               struct page *page,
1918                                               struct page **tree_pagep)
1919 {
1920         struct rb_node **new;
1921         struct rb_root *root;
1922         struct rb_node *parent = NULL;
1923         int nid;
1924
1925         nid = get_kpfn_nid(page_to_pfn(page));
1926         root = root_unstable_tree + nid;
1927         new = &root->rb_node;
1928
1929         while (*new) {
1930                 struct rmap_item *tree_rmap_item;
1931                 struct page *tree_page;
1932                 int ret;
1933
1934                 cond_resched();
1935                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1936                 tree_page = get_mergeable_page(tree_rmap_item);
1937                 if (!tree_page)
1938                         return NULL;
1939
1940                 /*
1941                  * Don't substitute a ksm page for a forked page.
1942                  */
1943                 if (page == tree_page) {
1944                         put_page(tree_page);
1945                         return NULL;
1946                 }
1947
1948                 ret = memcmp_pages(page, tree_page);
1949
1950                 parent = *new;
1951                 if (ret < 0) {
1952                         put_page(tree_page);
1953                         new = &parent->rb_left;
1954                 } else if (ret > 0) {
1955                         put_page(tree_page);
1956                         new = &parent->rb_right;
1957                 } else if (!ksm_merge_across_nodes &&
1958                            page_to_nid(tree_page) != nid) {
1959                         /*
1960                          * If tree_page has been migrated to another NUMA node,
1961                          * it will be flushed out and put in the right unstable
1962                          * tree next time: only merge with it when across_nodes.
1963                          */
1964                         put_page(tree_page);
1965                         return NULL;
1966                 } else {
1967                         *tree_pagep = tree_page;
1968                         return tree_rmap_item;
1969                 }
1970         }
1971
1972         rmap_item->address |= UNSTABLE_FLAG;
1973         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1974         DO_NUMA(rmap_item->nid = nid);
1975         rb_link_node(&rmap_item->node, parent, new);
1976         rb_insert_color(&rmap_item->node, root);
1977
1978         ksm_pages_unshared++;
1979         return NULL;
1980 }
1981
1982 /*
1983  * stable_tree_append - add another rmap_item to the linked list of
1984  * rmap_items hanging off a given node of the stable tree, all sharing
1985  * the same ksm page.
1986  */
1987 static void stable_tree_append(struct rmap_item *rmap_item,
1988                                struct stable_node *stable_node,
1989                                bool max_page_sharing_bypass)
1990 {
1991         /*
1992          * rmap won't find this mapping if we don't insert the
1993          * rmap_item in the right stable_node
1994          * duplicate. page_migration could break later if rmap breaks,
1995          * so we can as well crash here. We really need to check for
1996          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1997          * for other negative values as an underflow if detected here
1998          * for the first time (and not when decreasing rmap_hlist_len)
1999          * would be sign of memory corruption in the stable_node.
2000          */
2001         BUG_ON(stable_node->rmap_hlist_len < 0);
2002
2003         stable_node->rmap_hlist_len++;
2004         if (!max_page_sharing_bypass)
2005                 /* possibly non fatal but unexpected overflow, only warn */
2006                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2007                              ksm_max_page_sharing);
2008
2009         rmap_item->head = stable_node;
2010         rmap_item->address |= STABLE_FLAG;
2011         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2012
2013         if (rmap_item->hlist.next)
2014                 ksm_pages_sharing++;
2015         else
2016                 ksm_pages_shared++;
2017 }
2018
2019 /*
2020  * cmp_and_merge_page - first see if page can be merged into the stable tree;
2021  * if not, compare checksum to previous and if it's the same, see if page can
2022  * be inserted into the unstable tree, or merged with a page already there and
2023  * both transferred to the stable tree.
2024  *
2025  * @page: the page that we are searching identical page to.
2026  * @rmap_item: the reverse mapping into the virtual address of this page
2027  */
2028 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2029 {
2030         struct mm_struct *mm = rmap_item->mm;
2031         struct rmap_item *tree_rmap_item;
2032         struct page *tree_page = NULL;
2033         struct stable_node *stable_node;
2034         struct page *kpage;
2035         unsigned int checksum;
2036         int err;
2037         bool max_page_sharing_bypass = false;
2038
2039         stable_node = page_stable_node(page);
2040         if (stable_node) {
2041                 if (stable_node->head != &migrate_nodes &&
2042                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2043                     NUMA(stable_node->nid)) {
2044                         stable_node_dup_del(stable_node);
2045                         stable_node->head = &migrate_nodes;
2046                         list_add(&stable_node->list, stable_node->head);
2047                 }
2048                 if (stable_node->head != &migrate_nodes &&
2049                     rmap_item->head == stable_node)
2050                         return;
2051                 /*
2052                  * If it's a KSM fork, allow it to go over the sharing limit
2053                  * without warnings.
2054                  */
2055                 if (!is_page_sharing_candidate(stable_node))
2056                         max_page_sharing_bypass = true;
2057         }
2058
2059         /* We first start with searching the page inside the stable tree */
2060         kpage = stable_tree_search(page);
2061         if (kpage == page && rmap_item->head == stable_node) {
2062                 put_page(kpage);
2063                 return;
2064         }
2065
2066         remove_rmap_item_from_tree(rmap_item);
2067
2068         if (kpage) {
2069                 if (PTR_ERR(kpage) == -EBUSY)
2070                         return;
2071
2072                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2073                 if (!err) {
2074                         /*
2075                          * The page was successfully merged:
2076                          * add its rmap_item to the stable tree.
2077                          */
2078                         lock_page(kpage);
2079                         stable_tree_append(rmap_item, page_stable_node(kpage),
2080                                            max_page_sharing_bypass);
2081                         unlock_page(kpage);
2082                 }
2083                 put_page(kpage);
2084                 return;
2085         }
2086
2087         /*
2088          * If the hash value of the page has changed from the last time
2089          * we calculated it, this page is changing frequently: therefore we
2090          * don't want to insert it in the unstable tree, and we don't want
2091          * to waste our time searching for something identical to it there.
2092          */
2093         checksum = calc_checksum(page);
2094         if (rmap_item->oldchecksum != checksum) {
2095                 rmap_item->oldchecksum = checksum;
2096                 return;
2097         }
2098
2099         /*
2100          * Same checksum as an empty page. We attempt to merge it with the
2101          * appropriate zero page if the user enabled this via sysfs.
2102          */
2103         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2104                 struct vm_area_struct *vma;
2105
2106                 mmap_read_lock(mm);
2107                 vma = find_mergeable_vma(mm, rmap_item->address);
2108                 if (vma) {
2109                         err = try_to_merge_one_page(vma, page,
2110                                         ZERO_PAGE(rmap_item->address));
2111                 } else {
2112                         /*
2113                          * If the vma is out of date, we do not need to
2114                          * continue.
2115                          */
2116                         err = 0;
2117                 }
2118                 mmap_read_unlock(mm);
2119                 /*
2120                  * In case of failure, the page was not really empty, so we
2121                  * need to continue. Otherwise we're done.
2122                  */
2123                 if (!err)
2124                         return;
2125         }
2126         tree_rmap_item =
2127                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2128         if (tree_rmap_item) {
2129                 bool split;
2130
2131                 kpage = try_to_merge_two_pages(rmap_item, page,
2132                                                 tree_rmap_item, tree_page);
2133                 /*
2134                  * If both pages we tried to merge belong to the same compound
2135                  * page, then we actually ended up increasing the reference
2136                  * count of the same compound page twice, and split_huge_page
2137                  * failed.
2138                  * Here we set a flag if that happened, and we use it later to
2139                  * try split_huge_page again. Since we call put_page right
2140                  * afterwards, the reference count will be correct and
2141                  * split_huge_page should succeed.
2142                  */
2143                 split = PageTransCompound(page)
2144                         && compound_head(page) == compound_head(tree_page);
2145                 put_page(tree_page);
2146                 if (kpage) {
2147                         /*
2148                          * The pages were successfully merged: insert new
2149                          * node in the stable tree and add both rmap_items.
2150                          */
2151                         lock_page(kpage);
2152                         stable_node = stable_tree_insert(kpage);
2153                         if (stable_node) {
2154                                 stable_tree_append(tree_rmap_item, stable_node,
2155                                                    false);
2156                                 stable_tree_append(rmap_item, stable_node,
2157                                                    false);
2158                         }
2159                         unlock_page(kpage);
2160
2161                         /*
2162                          * If we fail to insert the page into the stable tree,
2163                          * we will have 2 virtual addresses that are pointing
2164                          * to a ksm page left outside the stable tree,
2165                          * in which case we need to break_cow on both.
2166                          */
2167                         if (!stable_node) {
2168                                 break_cow(tree_rmap_item);
2169                                 break_cow(rmap_item);
2170                         }
2171                 } else if (split) {
2172                         /*
2173                          * We are here if we tried to merge two pages and
2174                          * failed because they both belonged to the same
2175                          * compound page. We will split the page now, but no
2176                          * merging will take place.
2177                          * We do not want to add the cost of a full lock; if
2178                          * the page is locked, it is better to skip it and
2179                          * perhaps try again later.
2180                          */
2181                         if (!trylock_page(page))
2182                                 return;
2183                         split_huge_page(page);
2184                         unlock_page(page);
2185                 }
2186         }
2187 }
2188
2189 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2190                                             struct rmap_item **rmap_list,
2191                                             unsigned long addr)
2192 {
2193         struct rmap_item *rmap_item;
2194
2195         while (*rmap_list) {
2196                 rmap_item = *rmap_list;
2197                 if ((rmap_item->address & PAGE_MASK) == addr)
2198                         return rmap_item;
2199                 if (rmap_item->address > addr)
2200                         break;
2201                 *rmap_list = rmap_item->rmap_list;
2202                 remove_rmap_item_from_tree(rmap_item);
2203                 free_rmap_item(rmap_item);
2204         }
2205
2206         rmap_item = alloc_rmap_item();
2207         if (rmap_item) {
2208                 /* It has already been zeroed */
2209                 rmap_item->mm = mm_slot->mm;
2210                 rmap_item->address = addr;
2211                 rmap_item->rmap_list = *rmap_list;
2212                 *rmap_list = rmap_item;
2213         }
2214         return rmap_item;
2215 }
2216
2217 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2218 {
2219         struct mm_struct *mm;
2220         struct mm_slot *slot;
2221         struct vm_area_struct *vma;
2222         struct rmap_item *rmap_item;
2223         int nid;
2224
2225         if (list_empty(&ksm_mm_head.mm_list))
2226                 return NULL;
2227
2228         slot = ksm_scan.mm_slot;
2229         if (slot == &ksm_mm_head) {
2230                 /*
2231                  * A number of pages can hang around indefinitely on per-cpu
2232                  * pagevecs, raised page count preventing write_protect_page
2233                  * from merging them.  Though it doesn't really matter much,
2234                  * it is puzzling to see some stuck in pages_volatile until
2235                  * other activity jostles them out, and they also prevented
2236                  * LTP's KSM test from succeeding deterministically; so drain
2237                  * them here (here rather than on entry to ksm_do_scan(),
2238                  * so we don't IPI too often when pages_to_scan is set low).
2239                  */
2240                 lru_add_drain_all();
2241
2242                 /*
2243                  * Whereas stale stable_nodes on the stable_tree itself
2244                  * get pruned in the regular course of stable_tree_search(),
2245                  * those moved out to the migrate_nodes list can accumulate:
2246                  * so prune them once before each full scan.
2247                  */
2248                 if (!ksm_merge_across_nodes) {
2249                         struct stable_node *stable_node, *next;
2250                         struct page *page;
2251
2252                         list_for_each_entry_safe(stable_node, next,
2253                                                  &migrate_nodes, list) {
2254                                 page = get_ksm_page(stable_node,
2255                                                     GET_KSM_PAGE_NOLOCK);
2256                                 if (page)
2257                                         put_page(page);
2258                                 cond_resched();
2259                         }
2260                 }
2261
2262                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2263                         root_unstable_tree[nid] = RB_ROOT;
2264
2265                 spin_lock(&ksm_mmlist_lock);
2266                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2267                 ksm_scan.mm_slot = slot;
2268                 spin_unlock(&ksm_mmlist_lock);
2269                 /*
2270                  * Although we tested list_empty() above, a racing __ksm_exit
2271                  * of the last mm on the list may have removed it since then.
2272                  */
2273                 if (slot == &ksm_mm_head)
2274                         return NULL;
2275 next_mm:
2276                 ksm_scan.address = 0;
2277                 ksm_scan.rmap_list = &slot->rmap_list;
2278         }
2279
2280         mm = slot->mm;
2281         mmap_read_lock(mm);
2282         if (ksm_test_exit(mm))
2283                 vma = NULL;
2284         else
2285                 vma = find_vma(mm, ksm_scan.address);
2286
2287         for (; vma; vma = vma->vm_next) {
2288                 if (!(vma->vm_flags & VM_MERGEABLE))
2289                         continue;
2290                 if (ksm_scan.address < vma->vm_start)
2291                         ksm_scan.address = vma->vm_start;
2292                 if (!vma->anon_vma)
2293                         ksm_scan.address = vma->vm_end;
2294
2295                 while (ksm_scan.address < vma->vm_end) {
2296                         if (ksm_test_exit(mm))
2297                                 break;
2298                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2299                         if (IS_ERR_OR_NULL(*page)) {
2300                                 ksm_scan.address += PAGE_SIZE;
2301                                 cond_resched();
2302                                 continue;
2303                         }
2304                         if (PageAnon(*page)) {
2305                                 flush_anon_page(vma, *page, ksm_scan.address);
2306                                 flush_dcache_page(*page);
2307                                 rmap_item = get_next_rmap_item(slot,
2308                                         ksm_scan.rmap_list, ksm_scan.address);
2309                                 if (rmap_item) {
2310                                         ksm_scan.rmap_list =
2311                                                         &rmap_item->rmap_list;
2312                                         ksm_scan.address += PAGE_SIZE;
2313                                 } else
2314                                         put_page(*page);
2315                                 mmap_read_unlock(mm);
2316                                 return rmap_item;
2317                         }
2318                         put_page(*page);
2319                         ksm_scan.address += PAGE_SIZE;
2320                         cond_resched();
2321                 }
2322         }
2323
2324         if (ksm_test_exit(mm)) {
2325                 ksm_scan.address = 0;
2326                 ksm_scan.rmap_list = &slot->rmap_list;
2327         }
2328         /*
2329          * Nuke all the rmap_items that are above this current rmap:
2330          * because there were no VM_MERGEABLE vmas with such addresses.
2331          */
2332         remove_trailing_rmap_items(ksm_scan.rmap_list);
2333
2334         spin_lock(&ksm_mmlist_lock);
2335         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2336                                                 struct mm_slot, mm_list);
2337         if (ksm_scan.address == 0) {
2338                 /*
2339                  * We've completed a full scan of all vmas, holding mmap_lock
2340                  * throughout, and found no VM_MERGEABLE: so do the same as
2341                  * __ksm_exit does to remove this mm from all our lists now.
2342                  * This applies either when cleaning up after __ksm_exit
2343                  * (but beware: we can reach here even before __ksm_exit),
2344                  * or when all VM_MERGEABLE areas have been unmapped (and
2345                  * mmap_lock then protects against race with MADV_MERGEABLE).
2346                  */
2347                 hash_del(&slot->link);
2348                 list_del(&slot->mm_list);
2349                 spin_unlock(&ksm_mmlist_lock);
2350
2351                 free_mm_slot(slot);
2352                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2353                 mmap_read_unlock(mm);
2354                 mmdrop(mm);
2355         } else {
2356                 mmap_read_unlock(mm);
2357                 /*
2358                  * mmap_read_unlock(mm) first because after
2359                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2360                  * already have been freed under us by __ksm_exit()
2361                  * because the "mm_slot" is still hashed and
2362                  * ksm_scan.mm_slot doesn't point to it anymore.
2363                  */
2364                 spin_unlock(&ksm_mmlist_lock);
2365         }
2366
2367         /* Repeat until we've completed scanning the whole list */
2368         slot = ksm_scan.mm_slot;
2369         if (slot != &ksm_mm_head)
2370                 goto next_mm;
2371
2372         ksm_scan.seqnr++;
2373         return NULL;
2374 }
2375
2376 /**
2377  * ksm_do_scan  - the ksm scanner main worker function.
2378  * @scan_npages:  number of pages we want to scan before we return.
2379  */
2380 static void ksm_do_scan(unsigned int scan_npages)
2381 {
2382         struct rmap_item *rmap_item;
2383         struct page *page;
2384
2385         while (scan_npages-- && likely(!freezing(current))) {
2386                 cond_resched();
2387                 rmap_item = scan_get_next_rmap_item(&page);
2388                 if (!rmap_item)
2389                         return;
2390                 cmp_and_merge_page(page, rmap_item);
2391                 put_page(page);
2392         }
2393 }
2394
2395 static int ksmd_should_run(void)
2396 {
2397         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2398 }
2399
2400 static int ksm_scan_thread(void *nothing)
2401 {
2402         unsigned int sleep_ms;
2403
2404         set_freezable();
2405         set_user_nice(current, 5);
2406
2407         while (!kthread_should_stop()) {
2408                 mutex_lock(&ksm_thread_mutex);
2409                 wait_while_offlining();
2410                 if (ksmd_should_run())
2411                         ksm_do_scan(ksm_thread_pages_to_scan);
2412                 mutex_unlock(&ksm_thread_mutex);
2413
2414                 try_to_freeze();
2415
2416                 if (ksmd_should_run()) {
2417                         sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2418                         wait_event_interruptible_timeout(ksm_iter_wait,
2419                                 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2420                                 msecs_to_jiffies(sleep_ms));
2421                 } else {
2422                         wait_event_freezable(ksm_thread_wait,
2423                                 ksmd_should_run() || kthread_should_stop());
2424                 }
2425         }
2426         return 0;
2427 }
2428
2429 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2430                 unsigned long end, int advice, unsigned long *vm_flags)
2431 {
2432         struct mm_struct *mm = vma->vm_mm;
2433         int err;
2434
2435         switch (advice) {
2436         case MADV_MERGEABLE:
2437                 /*
2438                  * Be somewhat over-protective for now!
2439                  */
2440                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2441                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2442                                  VM_HUGETLB | VM_MIXEDMAP))
2443                         return 0;               /* just ignore the advice */
2444
2445                 if (vma_is_dax(vma))
2446                         return 0;
2447
2448 #ifdef VM_SAO
2449                 if (*vm_flags & VM_SAO)
2450                         return 0;
2451 #endif
2452 #ifdef VM_SPARC_ADI
2453                 if (*vm_flags & VM_SPARC_ADI)
2454                         return 0;
2455 #endif
2456
2457                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2458                         err = __ksm_enter(mm);
2459                         if (err)
2460                                 return err;
2461                 }
2462
2463                 *vm_flags |= VM_MERGEABLE;
2464                 break;
2465
2466         case MADV_UNMERGEABLE:
2467                 if (!(*vm_flags & VM_MERGEABLE))
2468                         return 0;               /* just ignore the advice */
2469
2470                 if (vma->anon_vma) {
2471                         err = unmerge_ksm_pages(vma, start, end);
2472                         if (err)
2473                                 return err;
2474                 }
2475
2476                 *vm_flags &= ~VM_MERGEABLE;
2477                 break;
2478         }
2479
2480         return 0;
2481 }
2482 EXPORT_SYMBOL_GPL(ksm_madvise);
2483
2484 int __ksm_enter(struct mm_struct *mm)
2485 {
2486         struct mm_slot *mm_slot;
2487         int needs_wakeup;
2488
2489         mm_slot = alloc_mm_slot();
2490         if (!mm_slot)
2491                 return -ENOMEM;
2492
2493         /* Check ksm_run too?  Would need tighter locking */
2494         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2495
2496         spin_lock(&ksm_mmlist_lock);
2497         insert_to_mm_slots_hash(mm, mm_slot);
2498         /*
2499          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2500          * insert just behind the scanning cursor, to let the area settle
2501          * down a little; when fork is followed by immediate exec, we don't
2502          * want ksmd to waste time setting up and tearing down an rmap_list.
2503          *
2504          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2505          * scanning cursor, otherwise KSM pages in newly forked mms will be
2506          * missed: then we might as well insert at the end of the list.
2507          */
2508         if (ksm_run & KSM_RUN_UNMERGE)
2509                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2510         else
2511                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2512         spin_unlock(&ksm_mmlist_lock);
2513
2514         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2515         mmgrab(mm);
2516
2517         if (needs_wakeup)
2518                 wake_up_interruptible(&ksm_thread_wait);
2519
2520         return 0;
2521 }
2522
2523 void __ksm_exit(struct mm_struct *mm)
2524 {
2525         struct mm_slot *mm_slot;
2526         int easy_to_free = 0;
2527
2528         /*
2529          * This process is exiting: if it's straightforward (as is the
2530          * case when ksmd was never running), free mm_slot immediately.
2531          * But if it's at the cursor or has rmap_items linked to it, use
2532          * mmap_lock to synchronize with any break_cows before pagetables
2533          * are freed, and leave the mm_slot on the list for ksmd to free.
2534          * Beware: ksm may already have noticed it exiting and freed the slot.
2535          */
2536
2537         spin_lock(&ksm_mmlist_lock);
2538         mm_slot = get_mm_slot(mm);
2539         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2540                 if (!mm_slot->rmap_list) {
2541                         hash_del(&mm_slot->link);
2542                         list_del(&mm_slot->mm_list);
2543                         easy_to_free = 1;
2544                 } else {
2545                         list_move(&mm_slot->mm_list,
2546                                   &ksm_scan.mm_slot->mm_list);
2547                 }
2548         }
2549         spin_unlock(&ksm_mmlist_lock);
2550
2551         if (easy_to_free) {
2552                 free_mm_slot(mm_slot);
2553                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2554                 mmdrop(mm);
2555         } else if (mm_slot) {
2556                 mmap_write_lock(mm);
2557                 mmap_write_unlock(mm);
2558         }
2559 }
2560
2561 struct page *ksm_might_need_to_copy(struct page *page,
2562                         struct vm_area_struct *vma, unsigned long address)
2563 {
2564         struct anon_vma *anon_vma = page_anon_vma(page);
2565         struct page *new_page;
2566
2567         if (PageKsm(page)) {
2568                 if (page_stable_node(page) &&
2569                     !(ksm_run & KSM_RUN_UNMERGE))
2570                         return page;    /* no need to copy it */
2571         } else if (!anon_vma) {
2572                 return page;            /* no need to copy it */
2573         } else if (anon_vma->root == vma->anon_vma->root &&
2574                  page->index == linear_page_index(vma, address)) {
2575                 return page;            /* still no need to copy it */
2576         }
2577         if (!PageUptodate(page))
2578                 return page;            /* let do_swap_page report the error */
2579
2580         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2581         if (new_page && mem_cgroup_charge(new_page, vma->vm_mm, GFP_KERNEL)) {
2582                 put_page(new_page);
2583                 new_page = NULL;
2584         }
2585         if (new_page) {
2586                 copy_user_highpage(new_page, page, address, vma);
2587
2588                 SetPageDirty(new_page);
2589                 __SetPageUptodate(new_page);
2590                 __SetPageLocked(new_page);
2591         }
2592
2593         return new_page;
2594 }
2595
2596 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2597 {
2598         struct stable_node *stable_node;
2599         struct rmap_item *rmap_item;
2600         int search_new_forks = 0;
2601
2602         VM_BUG_ON_PAGE(!PageKsm(page), page);
2603
2604         /*
2605          * Rely on the page lock to protect against concurrent modifications
2606          * to that page's node of the stable tree.
2607          */
2608         VM_BUG_ON_PAGE(!PageLocked(page), page);
2609
2610         stable_node = page_stable_node(page);
2611         if (!stable_node)
2612                 return;
2613 again:
2614         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2615                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2616                 struct anon_vma_chain *vmac;
2617                 struct vm_area_struct *vma;
2618
2619                 cond_resched();
2620                 anon_vma_lock_read(anon_vma);
2621                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2622                                                0, ULONG_MAX) {
2623                         unsigned long addr;
2624
2625                         cond_resched();
2626                         vma = vmac->vma;
2627
2628                         /* Ignore the stable/unstable/sqnr flags */
2629                         addr = rmap_item->address & PAGE_MASK;
2630
2631                         if (addr < vma->vm_start || addr >= vma->vm_end)
2632                                 continue;
2633                         /*
2634                          * Initially we examine only the vma which covers this
2635                          * rmap_item; but later, if there is still work to do,
2636                          * we examine covering vmas in other mms: in case they
2637                          * were forked from the original since ksmd passed.
2638                          */
2639                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2640                                 continue;
2641
2642                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2643                                 continue;
2644
2645                         if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2646                                 anon_vma_unlock_read(anon_vma);
2647                                 return;
2648                         }
2649                         if (rwc->done && rwc->done(page)) {
2650                                 anon_vma_unlock_read(anon_vma);
2651                                 return;
2652                         }
2653                 }
2654                 anon_vma_unlock_read(anon_vma);
2655         }
2656         if (!search_new_forks++)
2657                 goto again;
2658 }
2659
2660 #ifdef CONFIG_MIGRATION
2661 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2662 {
2663         struct stable_node *stable_node;
2664
2665         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2666         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2667         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2668
2669         stable_node = page_stable_node(newpage);
2670         if (stable_node) {
2671                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2672                 stable_node->kpfn = page_to_pfn(newpage);
2673                 /*
2674                  * newpage->mapping was set in advance; now we need smp_wmb()
2675                  * to make sure that the new stable_node->kpfn is visible
2676                  * to get_ksm_page() before it can see that oldpage->mapping
2677                  * has gone stale (or that PageSwapCache has been cleared).
2678                  */
2679                 smp_wmb();
2680                 set_page_stable_node(oldpage, NULL);
2681         }
2682 }
2683 #endif /* CONFIG_MIGRATION */
2684
2685 #ifdef CONFIG_MEMORY_HOTREMOVE
2686 static void wait_while_offlining(void)
2687 {
2688         while (ksm_run & KSM_RUN_OFFLINE) {
2689                 mutex_unlock(&ksm_thread_mutex);
2690                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2691                             TASK_UNINTERRUPTIBLE);
2692                 mutex_lock(&ksm_thread_mutex);
2693         }
2694 }
2695
2696 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2697                                          unsigned long start_pfn,
2698                                          unsigned long end_pfn)
2699 {
2700         if (stable_node->kpfn >= start_pfn &&
2701             stable_node->kpfn < end_pfn) {
2702                 /*
2703                  * Don't get_ksm_page, page has already gone:
2704                  * which is why we keep kpfn instead of page*
2705                  */
2706                 remove_node_from_stable_tree(stable_node);
2707                 return true;
2708         }
2709         return false;
2710 }
2711
2712 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2713                                            unsigned long start_pfn,
2714                                            unsigned long end_pfn,
2715                                            struct rb_root *root)
2716 {
2717         struct stable_node *dup;
2718         struct hlist_node *hlist_safe;
2719
2720         if (!is_stable_node_chain(stable_node)) {
2721                 VM_BUG_ON(is_stable_node_dup(stable_node));
2722                 return stable_node_dup_remove_range(stable_node, start_pfn,
2723                                                     end_pfn);
2724         }
2725
2726         hlist_for_each_entry_safe(dup, hlist_safe,
2727                                   &stable_node->hlist, hlist_dup) {
2728                 VM_BUG_ON(!is_stable_node_dup(dup));
2729                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2730         }
2731         if (hlist_empty(&stable_node->hlist)) {
2732                 free_stable_node_chain(stable_node, root);
2733                 return true; /* notify caller that tree was rebalanced */
2734         } else
2735                 return false;
2736 }
2737
2738 static void ksm_check_stable_tree(unsigned long start_pfn,
2739                                   unsigned long end_pfn)
2740 {
2741         struct stable_node *stable_node, *next;
2742         struct rb_node *node;
2743         int nid;
2744
2745         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2746                 node = rb_first(root_stable_tree + nid);
2747                 while (node) {
2748                         stable_node = rb_entry(node, struct stable_node, node);
2749                         if (stable_node_chain_remove_range(stable_node,
2750                                                            start_pfn, end_pfn,
2751                                                            root_stable_tree +
2752                                                            nid))
2753                                 node = rb_first(root_stable_tree + nid);
2754                         else
2755                                 node = rb_next(node);
2756                         cond_resched();
2757                 }
2758         }
2759         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2760                 if (stable_node->kpfn >= start_pfn &&
2761                     stable_node->kpfn < end_pfn)
2762                         remove_node_from_stable_tree(stable_node);
2763                 cond_resched();
2764         }
2765 }
2766
2767 static int ksm_memory_callback(struct notifier_block *self,
2768                                unsigned long action, void *arg)
2769 {
2770         struct memory_notify *mn = arg;
2771
2772         switch (action) {
2773         case MEM_GOING_OFFLINE:
2774                 /*
2775                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2776                  * and remove_all_stable_nodes() while memory is going offline:
2777                  * it is unsafe for them to touch the stable tree at this time.
2778                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2779                  * which do not need the ksm_thread_mutex are all safe.
2780                  */
2781                 mutex_lock(&ksm_thread_mutex);
2782                 ksm_run |= KSM_RUN_OFFLINE;
2783                 mutex_unlock(&ksm_thread_mutex);
2784                 break;
2785
2786         case MEM_OFFLINE:
2787                 /*
2788                  * Most of the work is done by page migration; but there might
2789                  * be a few stable_nodes left over, still pointing to struct
2790                  * pages which have been offlined: prune those from the tree,
2791                  * otherwise get_ksm_page() might later try to access a
2792                  * non-existent struct page.
2793                  */
2794                 ksm_check_stable_tree(mn->start_pfn,
2795                                       mn->start_pfn + mn->nr_pages);
2796                 fallthrough;
2797         case MEM_CANCEL_OFFLINE:
2798                 mutex_lock(&ksm_thread_mutex);
2799                 ksm_run &= ~KSM_RUN_OFFLINE;
2800                 mutex_unlock(&ksm_thread_mutex);
2801
2802                 smp_mb();       /* wake_up_bit advises this */
2803                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2804                 break;
2805         }
2806         return NOTIFY_OK;
2807 }
2808 #else
2809 static void wait_while_offlining(void)
2810 {
2811 }
2812 #endif /* CONFIG_MEMORY_HOTREMOVE */
2813
2814 #ifdef CONFIG_SYSFS
2815 /*
2816  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2817  */
2818
2819 #define KSM_ATTR_RO(_name) \
2820         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2821 #define KSM_ATTR(_name) \
2822         static struct kobj_attribute _name##_attr = \
2823                 __ATTR(_name, 0644, _name##_show, _name##_store)
2824
2825 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2826                                     struct kobj_attribute *attr, char *buf)
2827 {
2828         return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
2829 }
2830
2831 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2832                                      struct kobj_attribute *attr,
2833                                      const char *buf, size_t count)
2834 {
2835         unsigned int msecs;
2836         int err;
2837
2838         err = kstrtouint(buf, 10, &msecs);
2839         if (err)
2840                 return -EINVAL;
2841
2842         ksm_thread_sleep_millisecs = msecs;
2843         wake_up_interruptible(&ksm_iter_wait);
2844
2845         return count;
2846 }
2847 KSM_ATTR(sleep_millisecs);
2848
2849 static ssize_t pages_to_scan_show(struct kobject *kobj,
2850                                   struct kobj_attribute *attr, char *buf)
2851 {
2852         return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
2853 }
2854
2855 static ssize_t pages_to_scan_store(struct kobject *kobj,
2856                                    struct kobj_attribute *attr,
2857                                    const char *buf, size_t count)
2858 {
2859         unsigned int nr_pages;
2860         int err;
2861
2862         err = kstrtouint(buf, 10, &nr_pages);
2863         if (err)
2864                 return -EINVAL;
2865
2866         ksm_thread_pages_to_scan = nr_pages;
2867
2868         return count;
2869 }
2870 KSM_ATTR(pages_to_scan);
2871
2872 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2873                         char *buf)
2874 {
2875         return sysfs_emit(buf, "%lu\n", ksm_run);
2876 }
2877
2878 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2879                          const char *buf, size_t count)
2880 {
2881         unsigned int flags;
2882         int err;
2883
2884         err = kstrtouint(buf, 10, &flags);
2885         if (err)
2886                 return -EINVAL;
2887         if (flags > KSM_RUN_UNMERGE)
2888                 return -EINVAL;
2889
2890         /*
2891          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2892          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2893          * breaking COW to free the pages_shared (but leaves mm_slots
2894          * on the list for when ksmd may be set running again).
2895          */
2896
2897         mutex_lock(&ksm_thread_mutex);
2898         wait_while_offlining();
2899         if (ksm_run != flags) {
2900                 ksm_run = flags;
2901                 if (flags & KSM_RUN_UNMERGE) {
2902                         set_current_oom_origin();
2903                         err = unmerge_and_remove_all_rmap_items();
2904                         clear_current_oom_origin();
2905                         if (err) {
2906                                 ksm_run = KSM_RUN_STOP;
2907                                 count = err;
2908                         }
2909                 }
2910         }
2911         mutex_unlock(&ksm_thread_mutex);
2912
2913         if (flags & KSM_RUN_MERGE)
2914                 wake_up_interruptible(&ksm_thread_wait);
2915
2916         return count;
2917 }
2918 KSM_ATTR(run);
2919
2920 #ifdef CONFIG_NUMA
2921 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2922                                        struct kobj_attribute *attr, char *buf)
2923 {
2924         return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
2925 }
2926
2927 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2928                                    struct kobj_attribute *attr,
2929                                    const char *buf, size_t count)
2930 {
2931         int err;
2932         unsigned long knob;
2933
2934         err = kstrtoul(buf, 10, &knob);
2935         if (err)
2936                 return err;
2937         if (knob > 1)
2938                 return -EINVAL;
2939
2940         mutex_lock(&ksm_thread_mutex);
2941         wait_while_offlining();
2942         if (ksm_merge_across_nodes != knob) {
2943                 if (ksm_pages_shared || remove_all_stable_nodes())
2944                         err = -EBUSY;
2945                 else if (root_stable_tree == one_stable_tree) {
2946                         struct rb_root *buf;
2947                         /*
2948                          * This is the first time that we switch away from the
2949                          * default of merging across nodes: must now allocate
2950                          * a buffer to hold as many roots as may be needed.
2951                          * Allocate stable and unstable together:
2952                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2953                          */
2954                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2955                                       GFP_KERNEL);
2956                         /* Let us assume that RB_ROOT is NULL is zero */
2957                         if (!buf)
2958                                 err = -ENOMEM;
2959                         else {
2960                                 root_stable_tree = buf;
2961                                 root_unstable_tree = buf + nr_node_ids;
2962                                 /* Stable tree is empty but not the unstable */
2963                                 root_unstable_tree[0] = one_unstable_tree[0];
2964                         }
2965                 }
2966                 if (!err) {
2967                         ksm_merge_across_nodes = knob;
2968                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2969                 }
2970         }
2971         mutex_unlock(&ksm_thread_mutex);
2972
2973         return err ? err : count;
2974 }
2975 KSM_ATTR(merge_across_nodes);
2976 #endif
2977
2978 static ssize_t use_zero_pages_show(struct kobject *kobj,
2979                                    struct kobj_attribute *attr, char *buf)
2980 {
2981         return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
2982 }
2983 static ssize_t use_zero_pages_store(struct kobject *kobj,
2984                                    struct kobj_attribute *attr,
2985                                    const char *buf, size_t count)
2986 {
2987         int err;
2988         bool value;
2989
2990         err = kstrtobool(buf, &value);
2991         if (err)
2992                 return -EINVAL;
2993
2994         ksm_use_zero_pages = value;
2995
2996         return count;
2997 }
2998 KSM_ATTR(use_zero_pages);
2999
3000 static ssize_t max_page_sharing_show(struct kobject *kobj,
3001                                      struct kobj_attribute *attr, char *buf)
3002 {
3003         return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3004 }
3005
3006 static ssize_t max_page_sharing_store(struct kobject *kobj,
3007                                       struct kobj_attribute *attr,
3008                                       const char *buf, size_t count)
3009 {
3010         int err;
3011         int knob;
3012
3013         err = kstrtoint(buf, 10, &knob);
3014         if (err)
3015                 return err;
3016         /*
3017          * When a KSM page is created it is shared by 2 mappings. This
3018          * being a signed comparison, it implicitly verifies it's not
3019          * negative.
3020          */
3021         if (knob < 2)
3022                 return -EINVAL;
3023
3024         if (READ_ONCE(ksm_max_page_sharing) == knob)
3025                 return count;
3026
3027         mutex_lock(&ksm_thread_mutex);
3028         wait_while_offlining();
3029         if (ksm_max_page_sharing != knob) {
3030                 if (ksm_pages_shared || remove_all_stable_nodes())
3031                         err = -EBUSY;
3032                 else
3033                         ksm_max_page_sharing = knob;
3034         }
3035         mutex_unlock(&ksm_thread_mutex);
3036
3037         return err ? err : count;
3038 }
3039 KSM_ATTR(max_page_sharing);
3040
3041 static ssize_t pages_shared_show(struct kobject *kobj,
3042                                  struct kobj_attribute *attr, char *buf)
3043 {
3044         return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3045 }
3046 KSM_ATTR_RO(pages_shared);
3047
3048 static ssize_t pages_sharing_show(struct kobject *kobj,
3049                                   struct kobj_attribute *attr, char *buf)
3050 {
3051         return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3052 }
3053 KSM_ATTR_RO(pages_sharing);
3054
3055 static ssize_t pages_unshared_show(struct kobject *kobj,
3056                                    struct kobj_attribute *attr, char *buf)
3057 {
3058         return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3059 }
3060 KSM_ATTR_RO(pages_unshared);
3061
3062 static ssize_t pages_volatile_show(struct kobject *kobj,
3063                                    struct kobj_attribute *attr, char *buf)
3064 {
3065         long ksm_pages_volatile;
3066
3067         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3068                                 - ksm_pages_sharing - ksm_pages_unshared;
3069         /*
3070          * It was not worth any locking to calculate that statistic,
3071          * but it might therefore sometimes be negative: conceal that.
3072          */
3073         if (ksm_pages_volatile < 0)
3074                 ksm_pages_volatile = 0;
3075         return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3076 }
3077 KSM_ATTR_RO(pages_volatile);
3078
3079 static ssize_t stable_node_dups_show(struct kobject *kobj,
3080                                      struct kobj_attribute *attr, char *buf)
3081 {
3082         return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3083 }
3084 KSM_ATTR_RO(stable_node_dups);
3085
3086 static ssize_t stable_node_chains_show(struct kobject *kobj,
3087                                        struct kobj_attribute *attr, char *buf)
3088 {
3089         return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3090 }
3091 KSM_ATTR_RO(stable_node_chains);
3092
3093 static ssize_t
3094 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3095                                         struct kobj_attribute *attr,
3096                                         char *buf)
3097 {
3098         return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3099 }
3100
3101 static ssize_t
3102 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3103                                          struct kobj_attribute *attr,
3104                                          const char *buf, size_t count)
3105 {
3106         unsigned int msecs;
3107         int err;
3108
3109         err = kstrtouint(buf, 10, &msecs);
3110         if (err)
3111                 return -EINVAL;
3112
3113         ksm_stable_node_chains_prune_millisecs = msecs;
3114
3115         return count;
3116 }
3117 KSM_ATTR(stable_node_chains_prune_millisecs);
3118
3119 static ssize_t full_scans_show(struct kobject *kobj,
3120                                struct kobj_attribute *attr, char *buf)
3121 {
3122         return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3123 }
3124 KSM_ATTR_RO(full_scans);
3125
3126 static struct attribute *ksm_attrs[] = {
3127         &sleep_millisecs_attr.attr,
3128         &pages_to_scan_attr.attr,
3129         &run_attr.attr,
3130         &pages_shared_attr.attr,
3131         &pages_sharing_attr.attr,
3132         &pages_unshared_attr.attr,
3133         &pages_volatile_attr.attr,
3134         &full_scans_attr.attr,
3135 #ifdef CONFIG_NUMA
3136         &merge_across_nodes_attr.attr,
3137 #endif
3138         &max_page_sharing_attr.attr,
3139         &stable_node_chains_attr.attr,
3140         &stable_node_dups_attr.attr,
3141         &stable_node_chains_prune_millisecs_attr.attr,
3142         &use_zero_pages_attr.attr,
3143         NULL,
3144 };
3145
3146 static const struct attribute_group ksm_attr_group = {
3147         .attrs = ksm_attrs,
3148         .name = "ksm",
3149 };
3150 #endif /* CONFIG_SYSFS */
3151
3152 static int __init ksm_init(void)
3153 {
3154         struct task_struct *ksm_thread;
3155         int err;
3156
3157         /* The correct value depends on page size and endianness */
3158         zero_checksum = calc_checksum(ZERO_PAGE(0));
3159         /* Default to false for backwards compatibility */
3160         ksm_use_zero_pages = false;
3161
3162         err = ksm_slab_init();
3163         if (err)
3164                 goto out;
3165
3166         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3167         if (IS_ERR(ksm_thread)) {
3168                 pr_err("ksm: creating kthread failed\n");
3169                 err = PTR_ERR(ksm_thread);
3170                 goto out_free;
3171         }
3172
3173 #ifdef CONFIG_SYSFS
3174         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3175         if (err) {
3176                 pr_err("ksm: register sysfs failed\n");
3177                 kthread_stop(ksm_thread);
3178                 goto out_free;
3179         }
3180 #else
3181         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3182
3183 #endif /* CONFIG_SYSFS */
3184
3185 #ifdef CONFIG_MEMORY_HOTREMOVE
3186         /* There is no significance to this priority 100 */
3187         hotplug_memory_notifier(ksm_memory_callback, 100);
3188 #endif
3189         return 0;
3190
3191 out_free:
3192         ksm_slab_free();
3193 out:
3194         return err;
3195 }
3196 subsys_initcall(ksm_init);