]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - mm/ksm.c
Merge tag 'fbdev-v4.13-rc5' of git://github.com/bzolnier/linux
[karo-tx-linux.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/jhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
42
43 #include <asm/tlbflush.h>
44 #include "internal.h"
45
46 #ifdef CONFIG_NUMA
47 #define NUMA(x)         (x)
48 #define DO_NUMA(x)      do { (x); } while (0)
49 #else
50 #define NUMA(x)         (0)
51 #define DO_NUMA(x)      do { } while (0)
52 #endif
53
54 /*
55  * A few notes about the KSM scanning process,
56  * to make it easier to understand the data structures below:
57  *
58  * In order to reduce excessive scanning, KSM sorts the memory pages by their
59  * contents into a data structure that holds pointers to the pages' locations.
60  *
61  * Since the contents of the pages may change at any moment, KSM cannot just
62  * insert the pages into a normal sorted tree and expect it to find anything.
63  * Therefore KSM uses two data structures - the stable and the unstable tree.
64  *
65  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66  * by their contents.  Because each such page is write-protected, searching on
67  * this tree is fully assured to be working (except when pages are unmapped),
68  * and therefore this tree is called the stable tree.
69  *
70  * In addition to the stable tree, KSM uses a second data structure called the
71  * unstable tree: this tree holds pointers to pages which have been found to
72  * be "unchanged for a period of time".  The unstable tree sorts these pages
73  * by their contents, but since they are not write-protected, KSM cannot rely
74  * upon the unstable tree to work correctly - the unstable tree is liable to
75  * be corrupted as its contents are modified, and so it is called unstable.
76  *
77  * KSM solves this problem by several techniques:
78  *
79  * 1) The unstable tree is flushed every time KSM completes scanning all
80  *    memory areas, and then the tree is rebuilt again from the beginning.
81  * 2) KSM will only insert into the unstable tree, pages whose hash value
82  *    has not changed since the previous scan of all memory areas.
83  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84  *    colors of the nodes and not on their contents, assuring that even when
85  *    the tree gets "corrupted" it won't get out of balance, so scanning time
86  *    remains the same (also, searching and inserting nodes in an rbtree uses
87  *    the same algorithm, so we have no overhead when we flush and rebuild).
88  * 4) KSM never flushes the stable tree, which means that even if it were to
89  *    take 10 attempts to find a page in the unstable tree, once it is found,
90  *    it is secured in the stable tree.  (When we scan a new page, we first
91  *    compare it against the stable tree, and then against the unstable tree.)
92  *
93  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94  * stable trees and multiple unstable trees: one of each for each NUMA node.
95  */
96
97 /**
98  * struct mm_slot - ksm information per mm that is being scanned
99  * @link: link to the mm_slots hash list
100  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
101  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
102  * @mm: the mm that this information is valid for
103  */
104 struct mm_slot {
105         struct hlist_node link;
106         struct list_head mm_list;
107         struct rmap_item *rmap_list;
108         struct mm_struct *mm;
109 };
110
111 /**
112  * struct ksm_scan - cursor for scanning
113  * @mm_slot: the current mm_slot we are scanning
114  * @address: the next address inside that to be scanned
115  * @rmap_list: link to the next rmap to be scanned in the rmap_list
116  * @seqnr: count of completed full scans (needed when removing unstable node)
117  *
118  * There is only the one ksm_scan instance of this cursor structure.
119  */
120 struct ksm_scan {
121         struct mm_slot *mm_slot;
122         unsigned long address;
123         struct rmap_item **rmap_list;
124         unsigned long seqnr;
125 };
126
127 /**
128  * struct stable_node - node of the stable rbtree
129  * @node: rb node of this ksm page in the stable tree
130  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131  * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
132  * @list: linked into migrate_nodes, pending placement in the proper node tree
133  * @hlist: hlist head of rmap_items using this ksm page
134  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
135  * @chain_prune_time: time of the last full garbage collection
136  * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
137  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
138  */
139 struct stable_node {
140         union {
141                 struct rb_node node;    /* when node of stable tree */
142                 struct {                /* when listed for migration */
143                         struct list_head *head;
144                         struct {
145                                 struct hlist_node hlist_dup;
146                                 struct list_head list;
147                         };
148                 };
149         };
150         struct hlist_head hlist;
151         union {
152                 unsigned long kpfn;
153                 unsigned long chain_prune_time;
154         };
155         /*
156          * STABLE_NODE_CHAIN can be any negative number in
157          * rmap_hlist_len negative range, but better not -1 to be able
158          * to reliably detect underflows.
159          */
160 #define STABLE_NODE_CHAIN -1024
161         int rmap_hlist_len;
162 #ifdef CONFIG_NUMA
163         int nid;
164 #endif
165 };
166
167 /**
168  * struct rmap_item - reverse mapping item for virtual addresses
169  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
170  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
171  * @nid: NUMA node id of unstable tree in which linked (may not match page)
172  * @mm: the memory structure this rmap_item is pointing into
173  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
174  * @oldchecksum: previous checksum of the page at that virtual address
175  * @node: rb node of this rmap_item in the unstable tree
176  * @head: pointer to stable_node heading this list in the stable tree
177  * @hlist: link into hlist of rmap_items hanging off that stable_node
178  */
179 struct rmap_item {
180         struct rmap_item *rmap_list;
181         union {
182                 struct anon_vma *anon_vma;      /* when stable */
183 #ifdef CONFIG_NUMA
184                 int nid;                /* when node of unstable tree */
185 #endif
186         };
187         struct mm_struct *mm;
188         unsigned long address;          /* + low bits used for flags below */
189         unsigned int oldchecksum;       /* when unstable */
190         union {
191                 struct rb_node node;    /* when node of unstable tree */
192                 struct {                /* when listed from stable tree */
193                         struct stable_node *head;
194                         struct hlist_node hlist;
195                 };
196         };
197 };
198
199 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
200 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
201 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
202
203 /* The stable and unstable tree heads */
204 static struct rb_root one_stable_tree[1] = { RB_ROOT };
205 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
206 static struct rb_root *root_stable_tree = one_stable_tree;
207 static struct rb_root *root_unstable_tree = one_unstable_tree;
208
209 /* Recently migrated nodes of stable tree, pending proper placement */
210 static LIST_HEAD(migrate_nodes);
211 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
212
213 #define MM_SLOTS_HASH_BITS 10
214 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
215
216 static struct mm_slot ksm_mm_head = {
217         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
218 };
219 static struct ksm_scan ksm_scan = {
220         .mm_slot = &ksm_mm_head,
221 };
222
223 static struct kmem_cache *rmap_item_cache;
224 static struct kmem_cache *stable_node_cache;
225 static struct kmem_cache *mm_slot_cache;
226
227 /* The number of nodes in the stable tree */
228 static unsigned long ksm_pages_shared;
229
230 /* The number of page slots additionally sharing those nodes */
231 static unsigned long ksm_pages_sharing;
232
233 /* The number of nodes in the unstable tree */
234 static unsigned long ksm_pages_unshared;
235
236 /* The number of rmap_items in use: to calculate pages_volatile */
237 static unsigned long ksm_rmap_items;
238
239 /* The number of stable_node chains */
240 static unsigned long ksm_stable_node_chains;
241
242 /* The number of stable_node dups linked to the stable_node chains */
243 static unsigned long ksm_stable_node_dups;
244
245 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
246 static int ksm_stable_node_chains_prune_millisecs = 2000;
247
248 /* Maximum number of page slots sharing a stable node */
249 static int ksm_max_page_sharing = 256;
250
251 /* Number of pages ksmd should scan in one batch */
252 static unsigned int ksm_thread_pages_to_scan = 100;
253
254 /* Milliseconds ksmd should sleep between batches */
255 static unsigned int ksm_thread_sleep_millisecs = 20;
256
257 /* Checksum of an empty (zeroed) page */
258 static unsigned int zero_checksum __read_mostly;
259
260 /* Whether to merge empty (zeroed) pages with actual zero pages */
261 static bool ksm_use_zero_pages __read_mostly;
262
263 #ifdef CONFIG_NUMA
264 /* Zeroed when merging across nodes is not allowed */
265 static unsigned int ksm_merge_across_nodes = 1;
266 static int ksm_nr_node_ids = 1;
267 #else
268 #define ksm_merge_across_nodes  1U
269 #define ksm_nr_node_ids         1
270 #endif
271
272 #define KSM_RUN_STOP    0
273 #define KSM_RUN_MERGE   1
274 #define KSM_RUN_UNMERGE 2
275 #define KSM_RUN_OFFLINE 4
276 static unsigned long ksm_run = KSM_RUN_STOP;
277 static void wait_while_offlining(void);
278
279 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
280 static DEFINE_MUTEX(ksm_thread_mutex);
281 static DEFINE_SPINLOCK(ksm_mmlist_lock);
282
283 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
284                 sizeof(struct __struct), __alignof__(struct __struct),\
285                 (__flags), NULL)
286
287 static int __init ksm_slab_init(void)
288 {
289         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
290         if (!rmap_item_cache)
291                 goto out;
292
293         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
294         if (!stable_node_cache)
295                 goto out_free1;
296
297         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
298         if (!mm_slot_cache)
299                 goto out_free2;
300
301         return 0;
302
303 out_free2:
304         kmem_cache_destroy(stable_node_cache);
305 out_free1:
306         kmem_cache_destroy(rmap_item_cache);
307 out:
308         return -ENOMEM;
309 }
310
311 static void __init ksm_slab_free(void)
312 {
313         kmem_cache_destroy(mm_slot_cache);
314         kmem_cache_destroy(stable_node_cache);
315         kmem_cache_destroy(rmap_item_cache);
316         mm_slot_cache = NULL;
317 }
318
319 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
320 {
321         return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
322 }
323
324 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
325 {
326         return dup->head == STABLE_NODE_DUP_HEAD;
327 }
328
329 static inline void stable_node_chain_add_dup(struct stable_node *dup,
330                                              struct stable_node *chain)
331 {
332         VM_BUG_ON(is_stable_node_dup(dup));
333         dup->head = STABLE_NODE_DUP_HEAD;
334         VM_BUG_ON(!is_stable_node_chain(chain));
335         hlist_add_head(&dup->hlist_dup, &chain->hlist);
336         ksm_stable_node_dups++;
337 }
338
339 static inline void __stable_node_dup_del(struct stable_node *dup)
340 {
341         VM_BUG_ON(!is_stable_node_dup(dup));
342         hlist_del(&dup->hlist_dup);
343         ksm_stable_node_dups--;
344 }
345
346 static inline void stable_node_dup_del(struct stable_node *dup)
347 {
348         VM_BUG_ON(is_stable_node_chain(dup));
349         if (is_stable_node_dup(dup))
350                 __stable_node_dup_del(dup);
351         else
352                 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
353 #ifdef CONFIG_DEBUG_VM
354         dup->head = NULL;
355 #endif
356 }
357
358 static inline struct rmap_item *alloc_rmap_item(void)
359 {
360         struct rmap_item *rmap_item;
361
362         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
363                                                 __GFP_NORETRY | __GFP_NOWARN);
364         if (rmap_item)
365                 ksm_rmap_items++;
366         return rmap_item;
367 }
368
369 static inline void free_rmap_item(struct rmap_item *rmap_item)
370 {
371         ksm_rmap_items--;
372         rmap_item->mm = NULL;   /* debug safety */
373         kmem_cache_free(rmap_item_cache, rmap_item);
374 }
375
376 static inline struct stable_node *alloc_stable_node(void)
377 {
378         /*
379          * The allocation can take too long with GFP_KERNEL when memory is under
380          * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
381          * grants access to memory reserves, helping to avoid this problem.
382          */
383         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
384 }
385
386 static inline void free_stable_node(struct stable_node *stable_node)
387 {
388         VM_BUG_ON(stable_node->rmap_hlist_len &&
389                   !is_stable_node_chain(stable_node));
390         kmem_cache_free(stable_node_cache, stable_node);
391 }
392
393 static inline struct mm_slot *alloc_mm_slot(void)
394 {
395         if (!mm_slot_cache)     /* initialization failed */
396                 return NULL;
397         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
398 }
399
400 static inline void free_mm_slot(struct mm_slot *mm_slot)
401 {
402         kmem_cache_free(mm_slot_cache, mm_slot);
403 }
404
405 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
406 {
407         struct mm_slot *slot;
408
409         hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
410                 if (slot->mm == mm)
411                         return slot;
412
413         return NULL;
414 }
415
416 static void insert_to_mm_slots_hash(struct mm_struct *mm,
417                                     struct mm_slot *mm_slot)
418 {
419         mm_slot->mm = mm;
420         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
421 }
422
423 /*
424  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
425  * page tables after it has passed through ksm_exit() - which, if necessary,
426  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
427  * a special flag: they can just back out as soon as mm_users goes to zero.
428  * ksm_test_exit() is used throughout to make this test for exit: in some
429  * places for correctness, in some places just to avoid unnecessary work.
430  */
431 static inline bool ksm_test_exit(struct mm_struct *mm)
432 {
433         return atomic_read(&mm->mm_users) == 0;
434 }
435
436 /*
437  * We use break_ksm to break COW on a ksm page: it's a stripped down
438  *
439  *      if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
440  *              put_page(page);
441  *
442  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
443  * in case the application has unmapped and remapped mm,addr meanwhile.
444  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
445  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
446  *
447  * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
448  * of the process that owns 'vma'.  We also do not want to enforce
449  * protection keys here anyway.
450  */
451 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
452 {
453         struct page *page;
454         int ret = 0;
455
456         do {
457                 cond_resched();
458                 page = follow_page(vma, addr,
459                                 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
460                 if (IS_ERR_OR_NULL(page))
461                         break;
462                 if (PageKsm(page))
463                         ret = handle_mm_fault(vma, addr,
464                                         FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
465                 else
466                         ret = VM_FAULT_WRITE;
467                 put_page(page);
468         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
469         /*
470          * We must loop because handle_mm_fault() may back out if there's
471          * any difficulty e.g. if pte accessed bit gets updated concurrently.
472          *
473          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
474          * COW has been broken, even if the vma does not permit VM_WRITE;
475          * but note that a concurrent fault might break PageKsm for us.
476          *
477          * VM_FAULT_SIGBUS could occur if we race with truncation of the
478          * backing file, which also invalidates anonymous pages: that's
479          * okay, that truncation will have unmapped the PageKsm for us.
480          *
481          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
482          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
483          * current task has TIF_MEMDIE set, and will be OOM killed on return
484          * to user; and ksmd, having no mm, would never be chosen for that.
485          *
486          * But if the mm is in a limited mem_cgroup, then the fault may fail
487          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
488          * even ksmd can fail in this way - though it's usually breaking ksm
489          * just to undo a merge it made a moment before, so unlikely to oom.
490          *
491          * That's a pity: we might therefore have more kernel pages allocated
492          * than we're counting as nodes in the stable tree; but ksm_do_scan
493          * will retry to break_cow on each pass, so should recover the page
494          * in due course.  The important thing is to not let VM_MERGEABLE
495          * be cleared while any such pages might remain in the area.
496          */
497         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
498 }
499
500 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
501                 unsigned long addr)
502 {
503         struct vm_area_struct *vma;
504         if (ksm_test_exit(mm))
505                 return NULL;
506         vma = find_vma(mm, addr);
507         if (!vma || vma->vm_start > addr)
508                 return NULL;
509         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
510                 return NULL;
511         return vma;
512 }
513
514 static void break_cow(struct rmap_item *rmap_item)
515 {
516         struct mm_struct *mm = rmap_item->mm;
517         unsigned long addr = rmap_item->address;
518         struct vm_area_struct *vma;
519
520         /*
521          * It is not an accident that whenever we want to break COW
522          * to undo, we also need to drop a reference to the anon_vma.
523          */
524         put_anon_vma(rmap_item->anon_vma);
525
526         down_read(&mm->mmap_sem);
527         vma = find_mergeable_vma(mm, addr);
528         if (vma)
529                 break_ksm(vma, addr);
530         up_read(&mm->mmap_sem);
531 }
532
533 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
534 {
535         struct mm_struct *mm = rmap_item->mm;
536         unsigned long addr = rmap_item->address;
537         struct vm_area_struct *vma;
538         struct page *page;
539
540         down_read(&mm->mmap_sem);
541         vma = find_mergeable_vma(mm, addr);
542         if (!vma)
543                 goto out;
544
545         page = follow_page(vma, addr, FOLL_GET);
546         if (IS_ERR_OR_NULL(page))
547                 goto out;
548         if (PageAnon(page)) {
549                 flush_anon_page(vma, page, addr);
550                 flush_dcache_page(page);
551         } else {
552                 put_page(page);
553 out:
554                 page = NULL;
555         }
556         up_read(&mm->mmap_sem);
557         return page;
558 }
559
560 /*
561  * This helper is used for getting right index into array of tree roots.
562  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
563  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
564  * every node has its own stable and unstable tree.
565  */
566 static inline int get_kpfn_nid(unsigned long kpfn)
567 {
568         return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
569 }
570
571 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
572                                                    struct rb_root *root)
573 {
574         struct stable_node *chain = alloc_stable_node();
575         VM_BUG_ON(is_stable_node_chain(dup));
576         if (likely(chain)) {
577                 INIT_HLIST_HEAD(&chain->hlist);
578                 chain->chain_prune_time = jiffies;
579                 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
580 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
581                 chain->nid = -1; /* debug */
582 #endif
583                 ksm_stable_node_chains++;
584
585                 /*
586                  * Put the stable node chain in the first dimension of
587                  * the stable tree and at the same time remove the old
588                  * stable node.
589                  */
590                 rb_replace_node(&dup->node, &chain->node, root);
591
592                 /*
593                  * Move the old stable node to the second dimension
594                  * queued in the hlist_dup. The invariant is that all
595                  * dup stable_nodes in the chain->hlist point to pages
596                  * that are wrprotected and have the exact same
597                  * content.
598                  */
599                 stable_node_chain_add_dup(dup, chain);
600         }
601         return chain;
602 }
603
604 static inline void free_stable_node_chain(struct stable_node *chain,
605                                           struct rb_root *root)
606 {
607         rb_erase(&chain->node, root);
608         free_stable_node(chain);
609         ksm_stable_node_chains--;
610 }
611
612 static void remove_node_from_stable_tree(struct stable_node *stable_node)
613 {
614         struct rmap_item *rmap_item;
615
616         /* check it's not STABLE_NODE_CHAIN or negative */
617         BUG_ON(stable_node->rmap_hlist_len < 0);
618
619         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
620                 if (rmap_item->hlist.next)
621                         ksm_pages_sharing--;
622                 else
623                         ksm_pages_shared--;
624                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
625                 stable_node->rmap_hlist_len--;
626                 put_anon_vma(rmap_item->anon_vma);
627                 rmap_item->address &= PAGE_MASK;
628                 cond_resched();
629         }
630
631         /*
632          * We need the second aligned pointer of the migrate_nodes
633          * list_head to stay clear from the rb_parent_color union
634          * (aligned and different than any node) and also different
635          * from &migrate_nodes. This will verify that future list.h changes
636          * don't break STABLE_NODE_DUP_HEAD.
637          */
638 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
639         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
640         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
641 #endif
642
643         if (stable_node->head == &migrate_nodes)
644                 list_del(&stable_node->list);
645         else
646                 stable_node_dup_del(stable_node);
647         free_stable_node(stable_node);
648 }
649
650 /*
651  * get_ksm_page: checks if the page indicated by the stable node
652  * is still its ksm page, despite having held no reference to it.
653  * In which case we can trust the content of the page, and it
654  * returns the gotten page; but if the page has now been zapped,
655  * remove the stale node from the stable tree and return NULL.
656  * But beware, the stable node's page might be being migrated.
657  *
658  * You would expect the stable_node to hold a reference to the ksm page.
659  * But if it increments the page's count, swapping out has to wait for
660  * ksmd to come around again before it can free the page, which may take
661  * seconds or even minutes: much too unresponsive.  So instead we use a
662  * "keyhole reference": access to the ksm page from the stable node peeps
663  * out through its keyhole to see if that page still holds the right key,
664  * pointing back to this stable node.  This relies on freeing a PageAnon
665  * page to reset its page->mapping to NULL, and relies on no other use of
666  * a page to put something that might look like our key in page->mapping.
667  * is on its way to being freed; but it is an anomaly to bear in mind.
668  */
669 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
670 {
671         struct page *page;
672         void *expected_mapping;
673         unsigned long kpfn;
674
675         expected_mapping = (void *)((unsigned long)stable_node |
676                                         PAGE_MAPPING_KSM);
677 again:
678         kpfn = READ_ONCE(stable_node->kpfn);
679         page = pfn_to_page(kpfn);
680
681         /*
682          * page is computed from kpfn, so on most architectures reading
683          * page->mapping is naturally ordered after reading node->kpfn,
684          * but on Alpha we need to be more careful.
685          */
686         smp_read_barrier_depends();
687         if (READ_ONCE(page->mapping) != expected_mapping)
688                 goto stale;
689
690         /*
691          * We cannot do anything with the page while its refcount is 0.
692          * Usually 0 means free, or tail of a higher-order page: in which
693          * case this node is no longer referenced, and should be freed;
694          * however, it might mean that the page is under page_freeze_refs().
695          * The __remove_mapping() case is easy, again the node is now stale;
696          * but if page is swapcache in migrate_page_move_mapping(), it might
697          * still be our page, in which case it's essential to keep the node.
698          */
699         while (!get_page_unless_zero(page)) {
700                 /*
701                  * Another check for page->mapping != expected_mapping would
702                  * work here too.  We have chosen the !PageSwapCache test to
703                  * optimize the common case, when the page is or is about to
704                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
705                  * in the freeze_refs section of __remove_mapping(); but Anon
706                  * page->mapping reset to NULL later, in free_pages_prepare().
707                  */
708                 if (!PageSwapCache(page))
709                         goto stale;
710                 cpu_relax();
711         }
712
713         if (READ_ONCE(page->mapping) != expected_mapping) {
714                 put_page(page);
715                 goto stale;
716         }
717
718         if (lock_it) {
719                 lock_page(page);
720                 if (READ_ONCE(page->mapping) != expected_mapping) {
721                         unlock_page(page);
722                         put_page(page);
723                         goto stale;
724                 }
725         }
726         return page;
727
728 stale:
729         /*
730          * We come here from above when page->mapping or !PageSwapCache
731          * suggests that the node is stale; but it might be under migration.
732          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
733          * before checking whether node->kpfn has been changed.
734          */
735         smp_rmb();
736         if (READ_ONCE(stable_node->kpfn) != kpfn)
737                 goto again;
738         remove_node_from_stable_tree(stable_node);
739         return NULL;
740 }
741
742 /*
743  * Removing rmap_item from stable or unstable tree.
744  * This function will clean the information from the stable/unstable tree.
745  */
746 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
747 {
748         if (rmap_item->address & STABLE_FLAG) {
749                 struct stable_node *stable_node;
750                 struct page *page;
751
752                 stable_node = rmap_item->head;
753                 page = get_ksm_page(stable_node, true);
754                 if (!page)
755                         goto out;
756
757                 hlist_del(&rmap_item->hlist);
758                 unlock_page(page);
759                 put_page(page);
760
761                 if (!hlist_empty(&stable_node->hlist))
762                         ksm_pages_sharing--;
763                 else
764                         ksm_pages_shared--;
765                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
766                 stable_node->rmap_hlist_len--;
767
768                 put_anon_vma(rmap_item->anon_vma);
769                 rmap_item->address &= PAGE_MASK;
770
771         } else if (rmap_item->address & UNSTABLE_FLAG) {
772                 unsigned char age;
773                 /*
774                  * Usually ksmd can and must skip the rb_erase, because
775                  * root_unstable_tree was already reset to RB_ROOT.
776                  * But be careful when an mm is exiting: do the rb_erase
777                  * if this rmap_item was inserted by this scan, rather
778                  * than left over from before.
779                  */
780                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
781                 BUG_ON(age > 1);
782                 if (!age)
783                         rb_erase(&rmap_item->node,
784                                  root_unstable_tree + NUMA(rmap_item->nid));
785                 ksm_pages_unshared--;
786                 rmap_item->address &= PAGE_MASK;
787         }
788 out:
789         cond_resched();         /* we're called from many long loops */
790 }
791
792 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
793                                        struct rmap_item **rmap_list)
794 {
795         while (*rmap_list) {
796                 struct rmap_item *rmap_item = *rmap_list;
797                 *rmap_list = rmap_item->rmap_list;
798                 remove_rmap_item_from_tree(rmap_item);
799                 free_rmap_item(rmap_item);
800         }
801 }
802
803 /*
804  * Though it's very tempting to unmerge rmap_items from stable tree rather
805  * than check every pte of a given vma, the locking doesn't quite work for
806  * that - an rmap_item is assigned to the stable tree after inserting ksm
807  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
808  * rmap_items from parent to child at fork time (so as not to waste time
809  * if exit comes before the next scan reaches it).
810  *
811  * Similarly, although we'd like to remove rmap_items (so updating counts
812  * and freeing memory) when unmerging an area, it's easier to leave that
813  * to the next pass of ksmd - consider, for example, how ksmd might be
814  * in cmp_and_merge_page on one of the rmap_items we would be removing.
815  */
816 static int unmerge_ksm_pages(struct vm_area_struct *vma,
817                              unsigned long start, unsigned long end)
818 {
819         unsigned long addr;
820         int err = 0;
821
822         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
823                 if (ksm_test_exit(vma->vm_mm))
824                         break;
825                 if (signal_pending(current))
826                         err = -ERESTARTSYS;
827                 else
828                         err = break_ksm(vma, addr);
829         }
830         return err;
831 }
832
833 #ifdef CONFIG_SYSFS
834 /*
835  * Only called through the sysfs control interface:
836  */
837 static int remove_stable_node(struct stable_node *stable_node)
838 {
839         struct page *page;
840         int err;
841
842         page = get_ksm_page(stable_node, true);
843         if (!page) {
844                 /*
845                  * get_ksm_page did remove_node_from_stable_tree itself.
846                  */
847                 return 0;
848         }
849
850         if (WARN_ON_ONCE(page_mapped(page))) {
851                 /*
852                  * This should not happen: but if it does, just refuse to let
853                  * merge_across_nodes be switched - there is no need to panic.
854                  */
855                 err = -EBUSY;
856         } else {
857                 /*
858                  * The stable node did not yet appear stale to get_ksm_page(),
859                  * since that allows for an unmapped ksm page to be recognized
860                  * right up until it is freed; but the node is safe to remove.
861                  * This page might be in a pagevec waiting to be freed,
862                  * or it might be PageSwapCache (perhaps under writeback),
863                  * or it might have been removed from swapcache a moment ago.
864                  */
865                 set_page_stable_node(page, NULL);
866                 remove_node_from_stable_tree(stable_node);
867                 err = 0;
868         }
869
870         unlock_page(page);
871         put_page(page);
872         return err;
873 }
874
875 static int remove_stable_node_chain(struct stable_node *stable_node,
876                                     struct rb_root *root)
877 {
878         struct stable_node *dup;
879         struct hlist_node *hlist_safe;
880
881         if (!is_stable_node_chain(stable_node)) {
882                 VM_BUG_ON(is_stable_node_dup(stable_node));
883                 if (remove_stable_node(stable_node))
884                         return true;
885                 else
886                         return false;
887         }
888
889         hlist_for_each_entry_safe(dup, hlist_safe,
890                                   &stable_node->hlist, hlist_dup) {
891                 VM_BUG_ON(!is_stable_node_dup(dup));
892                 if (remove_stable_node(dup))
893                         return true;
894         }
895         BUG_ON(!hlist_empty(&stable_node->hlist));
896         free_stable_node_chain(stable_node, root);
897         return false;
898 }
899
900 static int remove_all_stable_nodes(void)
901 {
902         struct stable_node *stable_node, *next;
903         int nid;
904         int err = 0;
905
906         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
907                 while (root_stable_tree[nid].rb_node) {
908                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
909                                                 struct stable_node, node);
910                         if (remove_stable_node_chain(stable_node,
911                                                      root_stable_tree + nid)) {
912                                 err = -EBUSY;
913                                 break;  /* proceed to next nid */
914                         }
915                         cond_resched();
916                 }
917         }
918         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
919                 if (remove_stable_node(stable_node))
920                         err = -EBUSY;
921                 cond_resched();
922         }
923         return err;
924 }
925
926 static int unmerge_and_remove_all_rmap_items(void)
927 {
928         struct mm_slot *mm_slot;
929         struct mm_struct *mm;
930         struct vm_area_struct *vma;
931         int err = 0;
932
933         spin_lock(&ksm_mmlist_lock);
934         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
935                                                 struct mm_slot, mm_list);
936         spin_unlock(&ksm_mmlist_lock);
937
938         for (mm_slot = ksm_scan.mm_slot;
939                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
940                 mm = mm_slot->mm;
941                 down_read(&mm->mmap_sem);
942                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
943                         if (ksm_test_exit(mm))
944                                 break;
945                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
946                                 continue;
947                         err = unmerge_ksm_pages(vma,
948                                                 vma->vm_start, vma->vm_end);
949                         if (err)
950                                 goto error;
951                 }
952
953                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
954                 up_read(&mm->mmap_sem);
955
956                 spin_lock(&ksm_mmlist_lock);
957                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
958                                                 struct mm_slot, mm_list);
959                 if (ksm_test_exit(mm)) {
960                         hash_del(&mm_slot->link);
961                         list_del(&mm_slot->mm_list);
962                         spin_unlock(&ksm_mmlist_lock);
963
964                         free_mm_slot(mm_slot);
965                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
966                         mmdrop(mm);
967                 } else
968                         spin_unlock(&ksm_mmlist_lock);
969         }
970
971         /* Clean up stable nodes, but don't worry if some are still busy */
972         remove_all_stable_nodes();
973         ksm_scan.seqnr = 0;
974         return 0;
975
976 error:
977         up_read(&mm->mmap_sem);
978         spin_lock(&ksm_mmlist_lock);
979         ksm_scan.mm_slot = &ksm_mm_head;
980         spin_unlock(&ksm_mmlist_lock);
981         return err;
982 }
983 #endif /* CONFIG_SYSFS */
984
985 static u32 calc_checksum(struct page *page)
986 {
987         u32 checksum;
988         void *addr = kmap_atomic(page);
989         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
990         kunmap_atomic(addr);
991         return checksum;
992 }
993
994 static int memcmp_pages(struct page *page1, struct page *page2)
995 {
996         char *addr1, *addr2;
997         int ret;
998
999         addr1 = kmap_atomic(page1);
1000         addr2 = kmap_atomic(page2);
1001         ret = memcmp(addr1, addr2, PAGE_SIZE);
1002         kunmap_atomic(addr2);
1003         kunmap_atomic(addr1);
1004         return ret;
1005 }
1006
1007 static inline int pages_identical(struct page *page1, struct page *page2)
1008 {
1009         return !memcmp_pages(page1, page2);
1010 }
1011
1012 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1013                               pte_t *orig_pte)
1014 {
1015         struct mm_struct *mm = vma->vm_mm;
1016         struct page_vma_mapped_walk pvmw = {
1017                 .page = page,
1018                 .vma = vma,
1019         };
1020         int swapped;
1021         int err = -EFAULT;
1022         unsigned long mmun_start;       /* For mmu_notifiers */
1023         unsigned long mmun_end;         /* For mmu_notifiers */
1024
1025         pvmw.address = page_address_in_vma(page, vma);
1026         if (pvmw.address == -EFAULT)
1027                 goto out;
1028
1029         BUG_ON(PageTransCompound(page));
1030
1031         mmun_start = pvmw.address;
1032         mmun_end   = pvmw.address + PAGE_SIZE;
1033         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1034
1035         if (!page_vma_mapped_walk(&pvmw))
1036                 goto out_mn;
1037         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1038                 goto out_unlock;
1039
1040         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1041             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1042                                                 mm_tlb_flush_pending(mm)) {
1043                 pte_t entry;
1044
1045                 swapped = PageSwapCache(page);
1046                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1047                 /*
1048                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1049                  * take any lock, therefore the check that we are going to make
1050                  * with the pagecount against the mapcount is racey and
1051                  * O_DIRECT can happen right after the check.
1052                  * So we clear the pte and flush the tlb before the check
1053                  * this assure us that no O_DIRECT can happen after the check
1054                  * or in the middle of the check.
1055                  */
1056                 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
1057                 /*
1058                  * Check that no O_DIRECT or similar I/O is in progress on the
1059                  * page
1060                  */
1061                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1062                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1063                         goto out_unlock;
1064                 }
1065                 if (pte_dirty(entry))
1066                         set_page_dirty(page);
1067
1068                 if (pte_protnone(entry))
1069                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1070                 else
1071                         entry = pte_mkclean(pte_wrprotect(entry));
1072                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1073         }
1074         *orig_pte = *pvmw.pte;
1075         err = 0;
1076
1077 out_unlock:
1078         page_vma_mapped_walk_done(&pvmw);
1079 out_mn:
1080         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1081 out:
1082         return err;
1083 }
1084
1085 /**
1086  * replace_page - replace page in vma by new ksm page
1087  * @vma:      vma that holds the pte pointing to page
1088  * @page:     the page we are replacing by kpage
1089  * @kpage:    the ksm page we replace page by
1090  * @orig_pte: the original value of the pte
1091  *
1092  * Returns 0 on success, -EFAULT on failure.
1093  */
1094 static int replace_page(struct vm_area_struct *vma, struct page *page,
1095                         struct page *kpage, pte_t orig_pte)
1096 {
1097         struct mm_struct *mm = vma->vm_mm;
1098         pmd_t *pmd;
1099         pte_t *ptep;
1100         pte_t newpte;
1101         spinlock_t *ptl;
1102         unsigned long addr;
1103         int err = -EFAULT;
1104         unsigned long mmun_start;       /* For mmu_notifiers */
1105         unsigned long mmun_end;         /* For mmu_notifiers */
1106
1107         addr = page_address_in_vma(page, vma);
1108         if (addr == -EFAULT)
1109                 goto out;
1110
1111         pmd = mm_find_pmd(mm, addr);
1112         if (!pmd)
1113                 goto out;
1114
1115         mmun_start = addr;
1116         mmun_end   = addr + PAGE_SIZE;
1117         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1118
1119         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1120         if (!pte_same(*ptep, orig_pte)) {
1121                 pte_unmap_unlock(ptep, ptl);
1122                 goto out_mn;
1123         }
1124
1125         /*
1126          * No need to check ksm_use_zero_pages here: we can only have a
1127          * zero_page here if ksm_use_zero_pages was enabled alreaady.
1128          */
1129         if (!is_zero_pfn(page_to_pfn(kpage))) {
1130                 get_page(kpage);
1131                 page_add_anon_rmap(kpage, vma, addr, false);
1132                 newpte = mk_pte(kpage, vma->vm_page_prot);
1133         } else {
1134                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1135                                                vma->vm_page_prot));
1136         }
1137
1138         flush_cache_page(vma, addr, pte_pfn(*ptep));
1139         ptep_clear_flush_notify(vma, addr, ptep);
1140         set_pte_at_notify(mm, addr, ptep, newpte);
1141
1142         page_remove_rmap(page, false);
1143         if (!page_mapped(page))
1144                 try_to_free_swap(page);
1145         put_page(page);
1146
1147         pte_unmap_unlock(ptep, ptl);
1148         err = 0;
1149 out_mn:
1150         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1151 out:
1152         return err;
1153 }
1154
1155 /*
1156  * try_to_merge_one_page - take two pages and merge them into one
1157  * @vma: the vma that holds the pte pointing to page
1158  * @page: the PageAnon page that we want to replace with kpage
1159  * @kpage: the PageKsm page that we want to map instead of page,
1160  *         or NULL the first time when we want to use page as kpage.
1161  *
1162  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1163  */
1164 static int try_to_merge_one_page(struct vm_area_struct *vma,
1165                                  struct page *page, struct page *kpage)
1166 {
1167         pte_t orig_pte = __pte(0);
1168         int err = -EFAULT;
1169
1170         if (page == kpage)                      /* ksm page forked */
1171                 return 0;
1172
1173         if (!PageAnon(page))
1174                 goto out;
1175
1176         /*
1177          * We need the page lock to read a stable PageSwapCache in
1178          * write_protect_page().  We use trylock_page() instead of
1179          * lock_page() because we don't want to wait here - we
1180          * prefer to continue scanning and merging different pages,
1181          * then come back to this page when it is unlocked.
1182          */
1183         if (!trylock_page(page))
1184                 goto out;
1185
1186         if (PageTransCompound(page)) {
1187                 if (split_huge_page(page))
1188                         goto out_unlock;
1189         }
1190
1191         /*
1192          * If this anonymous page is mapped only here, its pte may need
1193          * to be write-protected.  If it's mapped elsewhere, all of its
1194          * ptes are necessarily already write-protected.  But in either
1195          * case, we need to lock and check page_count is not raised.
1196          */
1197         if (write_protect_page(vma, page, &orig_pte) == 0) {
1198                 if (!kpage) {
1199                         /*
1200                          * While we hold page lock, upgrade page from
1201                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1202                          * stable_tree_insert() will update stable_node.
1203                          */
1204                         set_page_stable_node(page, NULL);
1205                         mark_page_accessed(page);
1206                         /*
1207                          * Page reclaim just frees a clean page with no dirty
1208                          * ptes: make sure that the ksm page would be swapped.
1209                          */
1210                         if (!PageDirty(page))
1211                                 SetPageDirty(page);
1212                         err = 0;
1213                 } else if (pages_identical(page, kpage))
1214                         err = replace_page(vma, page, kpage, orig_pte);
1215         }
1216
1217         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1218                 munlock_vma_page(page);
1219                 if (!PageMlocked(kpage)) {
1220                         unlock_page(page);
1221                         lock_page(kpage);
1222                         mlock_vma_page(kpage);
1223                         page = kpage;           /* for final unlock */
1224                 }
1225         }
1226
1227 out_unlock:
1228         unlock_page(page);
1229 out:
1230         return err;
1231 }
1232
1233 /*
1234  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1235  * but no new kernel page is allocated: kpage must already be a ksm page.
1236  *
1237  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1238  */
1239 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1240                                       struct page *page, struct page *kpage)
1241 {
1242         struct mm_struct *mm = rmap_item->mm;
1243         struct vm_area_struct *vma;
1244         int err = -EFAULT;
1245
1246         down_read(&mm->mmap_sem);
1247         vma = find_mergeable_vma(mm, rmap_item->address);
1248         if (!vma)
1249                 goto out;
1250
1251         err = try_to_merge_one_page(vma, page, kpage);
1252         if (err)
1253                 goto out;
1254
1255         /* Unstable nid is in union with stable anon_vma: remove first */
1256         remove_rmap_item_from_tree(rmap_item);
1257
1258         /* Must get reference to anon_vma while still holding mmap_sem */
1259         rmap_item->anon_vma = vma->anon_vma;
1260         get_anon_vma(vma->anon_vma);
1261 out:
1262         up_read(&mm->mmap_sem);
1263         return err;
1264 }
1265
1266 /*
1267  * try_to_merge_two_pages - take two identical pages and prepare them
1268  * to be merged into one page.
1269  *
1270  * This function returns the kpage if we successfully merged two identical
1271  * pages into one ksm page, NULL otherwise.
1272  *
1273  * Note that this function upgrades page to ksm page: if one of the pages
1274  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1275  */
1276 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1277                                            struct page *page,
1278                                            struct rmap_item *tree_rmap_item,
1279                                            struct page *tree_page)
1280 {
1281         int err;
1282
1283         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1284         if (!err) {
1285                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1286                                                         tree_page, page);
1287                 /*
1288                  * If that fails, we have a ksm page with only one pte
1289                  * pointing to it: so break it.
1290                  */
1291                 if (err)
1292                         break_cow(rmap_item);
1293         }
1294         return err ? NULL : page;
1295 }
1296
1297 static __always_inline
1298 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1299 {
1300         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1301         /*
1302          * Check that at least one mapping still exists, otherwise
1303          * there's no much point to merge and share with this
1304          * stable_node, as the underlying tree_page of the other
1305          * sharer is going to be freed soon.
1306          */
1307         return stable_node->rmap_hlist_len &&
1308                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1309 }
1310
1311 static __always_inline
1312 bool is_page_sharing_candidate(struct stable_node *stable_node)
1313 {
1314         return __is_page_sharing_candidate(stable_node, 0);
1315 }
1316
1317 struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1318                              struct stable_node **_stable_node,
1319                              struct rb_root *root,
1320                              bool prune_stale_stable_nodes)
1321 {
1322         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1323         struct hlist_node *hlist_safe;
1324         struct page *_tree_page, *tree_page = NULL;
1325         int nr = 0;
1326         int found_rmap_hlist_len;
1327
1328         if (!prune_stale_stable_nodes ||
1329             time_before(jiffies, stable_node->chain_prune_time +
1330                         msecs_to_jiffies(
1331                                 ksm_stable_node_chains_prune_millisecs)))
1332                 prune_stale_stable_nodes = false;
1333         else
1334                 stable_node->chain_prune_time = jiffies;
1335
1336         hlist_for_each_entry_safe(dup, hlist_safe,
1337                                   &stable_node->hlist, hlist_dup) {
1338                 cond_resched();
1339                 /*
1340                  * We must walk all stable_node_dup to prune the stale
1341                  * stable nodes during lookup.
1342                  *
1343                  * get_ksm_page can drop the nodes from the
1344                  * stable_node->hlist if they point to freed pages
1345                  * (that's why we do a _safe walk). The "dup"
1346                  * stable_node parameter itself will be freed from
1347                  * under us if it returns NULL.
1348                  */
1349                 _tree_page = get_ksm_page(dup, false);
1350                 if (!_tree_page)
1351                         continue;
1352                 nr += 1;
1353                 if (is_page_sharing_candidate(dup)) {
1354                         if (!found ||
1355                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1356                                 if (found)
1357                                         put_page(tree_page);
1358                                 found = dup;
1359                                 found_rmap_hlist_len = found->rmap_hlist_len;
1360                                 tree_page = _tree_page;
1361
1362                                 /* skip put_page for found dup */
1363                                 if (!prune_stale_stable_nodes)
1364                                         break;
1365                                 continue;
1366                         }
1367                 }
1368                 put_page(_tree_page);
1369         }
1370
1371         if (found) {
1372                 /*
1373                  * nr is counting all dups in the chain only if
1374                  * prune_stale_stable_nodes is true, otherwise we may
1375                  * break the loop at nr == 1 even if there are
1376                  * multiple entries.
1377                  */
1378                 if (prune_stale_stable_nodes && nr == 1) {
1379                         /*
1380                          * If there's not just one entry it would
1381                          * corrupt memory, better BUG_ON. In KSM
1382                          * context with no lock held it's not even
1383                          * fatal.
1384                          */
1385                         BUG_ON(stable_node->hlist.first->next);
1386
1387                         /*
1388                          * There's just one entry and it is below the
1389                          * deduplication limit so drop the chain.
1390                          */
1391                         rb_replace_node(&stable_node->node, &found->node,
1392                                         root);
1393                         free_stable_node(stable_node);
1394                         ksm_stable_node_chains--;
1395                         ksm_stable_node_dups--;
1396                         /*
1397                          * NOTE: the caller depends on the stable_node
1398                          * to be equal to stable_node_dup if the chain
1399                          * was collapsed.
1400                          */
1401                         *_stable_node = found;
1402                         /*
1403                          * Just for robustneess as stable_node is
1404                          * otherwise left as a stable pointer, the
1405                          * compiler shall optimize it away at build
1406                          * time.
1407                          */
1408                         stable_node = NULL;
1409                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1410                            __is_page_sharing_candidate(found, 1)) {
1411                         /*
1412                          * If the found stable_node dup can accept one
1413                          * more future merge (in addition to the one
1414                          * that is underway) and is not at the head of
1415                          * the chain, put it there so next search will
1416                          * be quicker in the !prune_stale_stable_nodes
1417                          * case.
1418                          *
1419                          * NOTE: it would be inaccurate to use nr > 1
1420                          * instead of checking the hlist.first pointer
1421                          * directly, because in the
1422                          * prune_stale_stable_nodes case "nr" isn't
1423                          * the position of the found dup in the chain,
1424                          * but the total number of dups in the chain.
1425                          */
1426                         hlist_del(&found->hlist_dup);
1427                         hlist_add_head(&found->hlist_dup,
1428                                        &stable_node->hlist);
1429                 }
1430         }
1431
1432         *_stable_node_dup = found;
1433         return tree_page;
1434 }
1435
1436 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1437                                                struct rb_root *root)
1438 {
1439         if (!is_stable_node_chain(stable_node))
1440                 return stable_node;
1441         if (hlist_empty(&stable_node->hlist)) {
1442                 free_stable_node_chain(stable_node, root);
1443                 return NULL;
1444         }
1445         return hlist_entry(stable_node->hlist.first,
1446                            typeof(*stable_node), hlist_dup);
1447 }
1448
1449 /*
1450  * Like for get_ksm_page, this function can free the *_stable_node and
1451  * *_stable_node_dup if the returned tree_page is NULL.
1452  *
1453  * It can also free and overwrite *_stable_node with the found
1454  * stable_node_dup if the chain is collapsed (in which case
1455  * *_stable_node will be equal to *_stable_node_dup like if the chain
1456  * never existed). It's up to the caller to verify tree_page is not
1457  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1458  *
1459  * *_stable_node_dup is really a second output parameter of this
1460  * function and will be overwritten in all cases, the caller doesn't
1461  * need to initialize it.
1462  */
1463 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1464                                         struct stable_node **_stable_node,
1465                                         struct rb_root *root,
1466                                         bool prune_stale_stable_nodes)
1467 {
1468         struct stable_node *stable_node = *_stable_node;
1469         if (!is_stable_node_chain(stable_node)) {
1470                 if (is_page_sharing_candidate(stable_node)) {
1471                         *_stable_node_dup = stable_node;
1472                         return get_ksm_page(stable_node, false);
1473                 }
1474                 /*
1475                  * _stable_node_dup set to NULL means the stable_node
1476                  * reached the ksm_max_page_sharing limit.
1477                  */
1478                 *_stable_node_dup = NULL;
1479                 return NULL;
1480         }
1481         return stable_node_dup(_stable_node_dup, _stable_node, root,
1482                                prune_stale_stable_nodes);
1483 }
1484
1485 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1486                                                 struct stable_node **s_n,
1487                                                 struct rb_root *root)
1488 {
1489         return __stable_node_chain(s_n_d, s_n, root, true);
1490 }
1491
1492 static __always_inline struct page *chain(struct stable_node **s_n_d,
1493                                           struct stable_node *s_n,
1494                                           struct rb_root *root)
1495 {
1496         struct stable_node *old_stable_node = s_n;
1497         struct page *tree_page;
1498
1499         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1500         /* not pruning dups so s_n cannot have changed */
1501         VM_BUG_ON(s_n != old_stable_node);
1502         return tree_page;
1503 }
1504
1505 /*
1506  * stable_tree_search - search for page inside the stable tree
1507  *
1508  * This function checks if there is a page inside the stable tree
1509  * with identical content to the page that we are scanning right now.
1510  *
1511  * This function returns the stable tree node of identical content if found,
1512  * NULL otherwise.
1513  */
1514 static struct page *stable_tree_search(struct page *page)
1515 {
1516         int nid;
1517         struct rb_root *root;
1518         struct rb_node **new;
1519         struct rb_node *parent;
1520         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1521         struct stable_node *page_node;
1522
1523         page_node = page_stable_node(page);
1524         if (page_node && page_node->head != &migrate_nodes) {
1525                 /* ksm page forked */
1526                 get_page(page);
1527                 return page;
1528         }
1529
1530         nid = get_kpfn_nid(page_to_pfn(page));
1531         root = root_stable_tree + nid;
1532 again:
1533         new = &root->rb_node;
1534         parent = NULL;
1535
1536         while (*new) {
1537                 struct page *tree_page;
1538                 int ret;
1539
1540                 cond_resched();
1541                 stable_node = rb_entry(*new, struct stable_node, node);
1542                 stable_node_any = NULL;
1543                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1544                 /*
1545                  * NOTE: stable_node may have been freed by
1546                  * chain_prune() if the returned stable_node_dup is
1547                  * not NULL. stable_node_dup may have been inserted in
1548                  * the rbtree instead as a regular stable_node (in
1549                  * order to collapse the stable_node chain if a single
1550                  * stable_node dup was found in it). In such case the
1551                  * stable_node is overwritten by the calleee to point
1552                  * to the stable_node_dup that was collapsed in the
1553                  * stable rbtree and stable_node will be equal to
1554                  * stable_node_dup like if the chain never existed.
1555                  */
1556                 if (!stable_node_dup) {
1557                         /*
1558                          * Either all stable_node dups were full in
1559                          * this stable_node chain, or this chain was
1560                          * empty and should be rb_erased.
1561                          */
1562                         stable_node_any = stable_node_dup_any(stable_node,
1563                                                               root);
1564                         if (!stable_node_any) {
1565                                 /* rb_erase just run */
1566                                 goto again;
1567                         }
1568                         /*
1569                          * Take any of the stable_node dups page of
1570                          * this stable_node chain to let the tree walk
1571                          * continue. All KSM pages belonging to the
1572                          * stable_node dups in a stable_node chain
1573                          * have the same content and they're
1574                          * wrprotected at all times. Any will work
1575                          * fine to continue the walk.
1576                          */
1577                         tree_page = get_ksm_page(stable_node_any, false);
1578                 }
1579                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1580                 if (!tree_page) {
1581                         /*
1582                          * If we walked over a stale stable_node,
1583                          * get_ksm_page() will call rb_erase() and it
1584                          * may rebalance the tree from under us. So
1585                          * restart the search from scratch. Returning
1586                          * NULL would be safe too, but we'd generate
1587                          * false negative insertions just because some
1588                          * stable_node was stale.
1589                          */
1590                         goto again;
1591                 }
1592
1593                 ret = memcmp_pages(page, tree_page);
1594                 put_page(tree_page);
1595
1596                 parent = *new;
1597                 if (ret < 0)
1598                         new = &parent->rb_left;
1599                 else if (ret > 0)
1600                         new = &parent->rb_right;
1601                 else {
1602                         if (page_node) {
1603                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1604                                 /*
1605                                  * Test if the migrated page should be merged
1606                                  * into a stable node dup. If the mapcount is
1607                                  * 1 we can migrate it with another KSM page
1608                                  * without adding it to the chain.
1609                                  */
1610                                 if (page_mapcount(page) > 1)
1611                                         goto chain_append;
1612                         }
1613
1614                         if (!stable_node_dup) {
1615                                 /*
1616                                  * If the stable_node is a chain and
1617                                  * we got a payload match in memcmp
1618                                  * but we cannot merge the scanned
1619                                  * page in any of the existing
1620                                  * stable_node dups because they're
1621                                  * all full, we need to wait the
1622                                  * scanned page to find itself a match
1623                                  * in the unstable tree to create a
1624                                  * brand new KSM page to add later to
1625                                  * the dups of this stable_node.
1626                                  */
1627                                 return NULL;
1628                         }
1629
1630                         /*
1631                          * Lock and unlock the stable_node's page (which
1632                          * might already have been migrated) so that page
1633                          * migration is sure to notice its raised count.
1634                          * It would be more elegant to return stable_node
1635                          * than kpage, but that involves more changes.
1636                          */
1637                         tree_page = get_ksm_page(stable_node_dup, true);
1638                         if (unlikely(!tree_page))
1639                                 /*
1640                                  * The tree may have been rebalanced,
1641                                  * so re-evaluate parent and new.
1642                                  */
1643                                 goto again;
1644                         unlock_page(tree_page);
1645
1646                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1647                             NUMA(stable_node_dup->nid)) {
1648                                 put_page(tree_page);
1649                                 goto replace;
1650                         }
1651                         return tree_page;
1652                 }
1653         }
1654
1655         if (!page_node)
1656                 return NULL;
1657
1658         list_del(&page_node->list);
1659         DO_NUMA(page_node->nid = nid);
1660         rb_link_node(&page_node->node, parent, new);
1661         rb_insert_color(&page_node->node, root);
1662 out:
1663         if (is_page_sharing_candidate(page_node)) {
1664                 get_page(page);
1665                 return page;
1666         } else
1667                 return NULL;
1668
1669 replace:
1670         /*
1671          * If stable_node was a chain and chain_prune collapsed it,
1672          * stable_node has been updated to be the new regular
1673          * stable_node. A collapse of the chain is indistinguishable
1674          * from the case there was no chain in the stable
1675          * rbtree. Otherwise stable_node is the chain and
1676          * stable_node_dup is the dup to replace.
1677          */
1678         if (stable_node_dup == stable_node) {
1679                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1680                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1681                 /* there is no chain */
1682                 if (page_node) {
1683                         VM_BUG_ON(page_node->head != &migrate_nodes);
1684                         list_del(&page_node->list);
1685                         DO_NUMA(page_node->nid = nid);
1686                         rb_replace_node(&stable_node_dup->node,
1687                                         &page_node->node,
1688                                         root);
1689                         if (is_page_sharing_candidate(page_node))
1690                                 get_page(page);
1691                         else
1692                                 page = NULL;
1693                 } else {
1694                         rb_erase(&stable_node_dup->node, root);
1695                         page = NULL;
1696                 }
1697         } else {
1698                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1699                 __stable_node_dup_del(stable_node_dup);
1700                 if (page_node) {
1701                         VM_BUG_ON(page_node->head != &migrate_nodes);
1702                         list_del(&page_node->list);
1703                         DO_NUMA(page_node->nid = nid);
1704                         stable_node_chain_add_dup(page_node, stable_node);
1705                         if (is_page_sharing_candidate(page_node))
1706                                 get_page(page);
1707                         else
1708                                 page = NULL;
1709                 } else {
1710                         page = NULL;
1711                 }
1712         }
1713         stable_node_dup->head = &migrate_nodes;
1714         list_add(&stable_node_dup->list, stable_node_dup->head);
1715         return page;
1716
1717 chain_append:
1718         /* stable_node_dup could be null if it reached the limit */
1719         if (!stable_node_dup)
1720                 stable_node_dup = stable_node_any;
1721         /*
1722          * If stable_node was a chain and chain_prune collapsed it,
1723          * stable_node has been updated to be the new regular
1724          * stable_node. A collapse of the chain is indistinguishable
1725          * from the case there was no chain in the stable
1726          * rbtree. Otherwise stable_node is the chain and
1727          * stable_node_dup is the dup to replace.
1728          */
1729         if (stable_node_dup == stable_node) {
1730                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1731                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1732                 /* chain is missing so create it */
1733                 stable_node = alloc_stable_node_chain(stable_node_dup,
1734                                                       root);
1735                 if (!stable_node)
1736                         return NULL;
1737         }
1738         /*
1739          * Add this stable_node dup that was
1740          * migrated to the stable_node chain
1741          * of the current nid for this page
1742          * content.
1743          */
1744         VM_BUG_ON(!is_stable_node_chain(stable_node));
1745         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1746         VM_BUG_ON(page_node->head != &migrate_nodes);
1747         list_del(&page_node->list);
1748         DO_NUMA(page_node->nid = nid);
1749         stable_node_chain_add_dup(page_node, stable_node);
1750         goto out;
1751 }
1752
1753 /*
1754  * stable_tree_insert - insert stable tree node pointing to new ksm page
1755  * into the stable tree.
1756  *
1757  * This function returns the stable tree node just allocated on success,
1758  * NULL otherwise.
1759  */
1760 static struct stable_node *stable_tree_insert(struct page *kpage)
1761 {
1762         int nid;
1763         unsigned long kpfn;
1764         struct rb_root *root;
1765         struct rb_node **new;
1766         struct rb_node *parent;
1767         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1768         bool need_chain = false;
1769
1770         kpfn = page_to_pfn(kpage);
1771         nid = get_kpfn_nid(kpfn);
1772         root = root_stable_tree + nid;
1773 again:
1774         parent = NULL;
1775         new = &root->rb_node;
1776
1777         while (*new) {
1778                 struct page *tree_page;
1779                 int ret;
1780
1781                 cond_resched();
1782                 stable_node = rb_entry(*new, struct stable_node, node);
1783                 stable_node_any = NULL;
1784                 tree_page = chain(&stable_node_dup, stable_node, root);
1785                 if (!stable_node_dup) {
1786                         /*
1787                          * Either all stable_node dups were full in
1788                          * this stable_node chain, or this chain was
1789                          * empty and should be rb_erased.
1790                          */
1791                         stable_node_any = stable_node_dup_any(stable_node,
1792                                                               root);
1793                         if (!stable_node_any) {
1794                                 /* rb_erase just run */
1795                                 goto again;
1796                         }
1797                         /*
1798                          * Take any of the stable_node dups page of
1799                          * this stable_node chain to let the tree walk
1800                          * continue. All KSM pages belonging to the
1801                          * stable_node dups in a stable_node chain
1802                          * have the same content and they're
1803                          * wrprotected at all times. Any will work
1804                          * fine to continue the walk.
1805                          */
1806                         tree_page = get_ksm_page(stable_node_any, false);
1807                 }
1808                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1809                 if (!tree_page) {
1810                         /*
1811                          * If we walked over a stale stable_node,
1812                          * get_ksm_page() will call rb_erase() and it
1813                          * may rebalance the tree from under us. So
1814                          * restart the search from scratch. Returning
1815                          * NULL would be safe too, but we'd generate
1816                          * false negative insertions just because some
1817                          * stable_node was stale.
1818                          */
1819                         goto again;
1820                 }
1821
1822                 ret = memcmp_pages(kpage, tree_page);
1823                 put_page(tree_page);
1824
1825                 parent = *new;
1826                 if (ret < 0)
1827                         new = &parent->rb_left;
1828                 else if (ret > 0)
1829                         new = &parent->rb_right;
1830                 else {
1831                         need_chain = true;
1832                         break;
1833                 }
1834         }
1835
1836         stable_node_dup = alloc_stable_node();
1837         if (!stable_node_dup)
1838                 return NULL;
1839
1840         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1841         stable_node_dup->kpfn = kpfn;
1842         set_page_stable_node(kpage, stable_node_dup);
1843         stable_node_dup->rmap_hlist_len = 0;
1844         DO_NUMA(stable_node_dup->nid = nid);
1845         if (!need_chain) {
1846                 rb_link_node(&stable_node_dup->node, parent, new);
1847                 rb_insert_color(&stable_node_dup->node, root);
1848         } else {
1849                 if (!is_stable_node_chain(stable_node)) {
1850                         struct stable_node *orig = stable_node;
1851                         /* chain is missing so create it */
1852                         stable_node = alloc_stable_node_chain(orig, root);
1853                         if (!stable_node) {
1854                                 free_stable_node(stable_node_dup);
1855                                 return NULL;
1856                         }
1857                 }
1858                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1859         }
1860
1861         return stable_node_dup;
1862 }
1863
1864 /*
1865  * unstable_tree_search_insert - search for identical page,
1866  * else insert rmap_item into the unstable tree.
1867  *
1868  * This function searches for a page in the unstable tree identical to the
1869  * page currently being scanned; and if no identical page is found in the
1870  * tree, we insert rmap_item as a new object into the unstable tree.
1871  *
1872  * This function returns pointer to rmap_item found to be identical
1873  * to the currently scanned page, NULL otherwise.
1874  *
1875  * This function does both searching and inserting, because they share
1876  * the same walking algorithm in an rbtree.
1877  */
1878 static
1879 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1880                                               struct page *page,
1881                                               struct page **tree_pagep)
1882 {
1883         struct rb_node **new;
1884         struct rb_root *root;
1885         struct rb_node *parent = NULL;
1886         int nid;
1887
1888         nid = get_kpfn_nid(page_to_pfn(page));
1889         root = root_unstable_tree + nid;
1890         new = &root->rb_node;
1891
1892         while (*new) {
1893                 struct rmap_item *tree_rmap_item;
1894                 struct page *tree_page;
1895                 int ret;
1896
1897                 cond_resched();
1898                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1899                 tree_page = get_mergeable_page(tree_rmap_item);
1900                 if (!tree_page)
1901                         return NULL;
1902
1903                 /*
1904                  * Don't substitute a ksm page for a forked page.
1905                  */
1906                 if (page == tree_page) {
1907                         put_page(tree_page);
1908                         return NULL;
1909                 }
1910
1911                 ret = memcmp_pages(page, tree_page);
1912
1913                 parent = *new;
1914                 if (ret < 0) {
1915                         put_page(tree_page);
1916                         new = &parent->rb_left;
1917                 } else if (ret > 0) {
1918                         put_page(tree_page);
1919                         new = &parent->rb_right;
1920                 } else if (!ksm_merge_across_nodes &&
1921                            page_to_nid(tree_page) != nid) {
1922                         /*
1923                          * If tree_page has been migrated to another NUMA node,
1924                          * it will be flushed out and put in the right unstable
1925                          * tree next time: only merge with it when across_nodes.
1926                          */
1927                         put_page(tree_page);
1928                         return NULL;
1929                 } else {
1930                         *tree_pagep = tree_page;
1931                         return tree_rmap_item;
1932                 }
1933         }
1934
1935         rmap_item->address |= UNSTABLE_FLAG;
1936         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1937         DO_NUMA(rmap_item->nid = nid);
1938         rb_link_node(&rmap_item->node, parent, new);
1939         rb_insert_color(&rmap_item->node, root);
1940
1941         ksm_pages_unshared++;
1942         return NULL;
1943 }
1944
1945 /*
1946  * stable_tree_append - add another rmap_item to the linked list of
1947  * rmap_items hanging off a given node of the stable tree, all sharing
1948  * the same ksm page.
1949  */
1950 static void stable_tree_append(struct rmap_item *rmap_item,
1951                                struct stable_node *stable_node,
1952                                bool max_page_sharing_bypass)
1953 {
1954         /*
1955          * rmap won't find this mapping if we don't insert the
1956          * rmap_item in the right stable_node
1957          * duplicate. page_migration could break later if rmap breaks,
1958          * so we can as well crash here. We really need to check for
1959          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1960          * for other negative values as an undeflow if detected here
1961          * for the first time (and not when decreasing rmap_hlist_len)
1962          * would be sign of memory corruption in the stable_node.
1963          */
1964         BUG_ON(stable_node->rmap_hlist_len < 0);
1965
1966         stable_node->rmap_hlist_len++;
1967         if (!max_page_sharing_bypass)
1968                 /* possibly non fatal but unexpected overflow, only warn */
1969                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1970                              ksm_max_page_sharing);
1971
1972         rmap_item->head = stable_node;
1973         rmap_item->address |= STABLE_FLAG;
1974         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1975
1976         if (rmap_item->hlist.next)
1977                 ksm_pages_sharing++;
1978         else
1979                 ksm_pages_shared++;
1980 }
1981
1982 /*
1983  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1984  * if not, compare checksum to previous and if it's the same, see if page can
1985  * be inserted into the unstable tree, or merged with a page already there and
1986  * both transferred to the stable tree.
1987  *
1988  * @page: the page that we are searching identical page to.
1989  * @rmap_item: the reverse mapping into the virtual address of this page
1990  */
1991 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1992 {
1993         struct rmap_item *tree_rmap_item;
1994         struct page *tree_page = NULL;
1995         struct stable_node *stable_node;
1996         struct page *kpage;
1997         unsigned int checksum;
1998         int err;
1999         bool max_page_sharing_bypass = false;
2000
2001         stable_node = page_stable_node(page);
2002         if (stable_node) {
2003                 if (stable_node->head != &migrate_nodes &&
2004                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2005                     NUMA(stable_node->nid)) {
2006                         stable_node_dup_del(stable_node);
2007                         stable_node->head = &migrate_nodes;
2008                         list_add(&stable_node->list, stable_node->head);
2009                 }
2010                 if (stable_node->head != &migrate_nodes &&
2011                     rmap_item->head == stable_node)
2012                         return;
2013                 /*
2014                  * If it's a KSM fork, allow it to go over the sharing limit
2015                  * without warnings.
2016                  */
2017                 if (!is_page_sharing_candidate(stable_node))
2018                         max_page_sharing_bypass = true;
2019         }
2020
2021         /* We first start with searching the page inside the stable tree */
2022         kpage = stable_tree_search(page);
2023         if (kpage == page && rmap_item->head == stable_node) {
2024                 put_page(kpage);
2025                 return;
2026         }
2027
2028         remove_rmap_item_from_tree(rmap_item);
2029
2030         if (kpage) {
2031                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2032                 if (!err) {
2033                         /*
2034                          * The page was successfully merged:
2035                          * add its rmap_item to the stable tree.
2036                          */
2037                         lock_page(kpage);
2038                         stable_tree_append(rmap_item, page_stable_node(kpage),
2039                                            max_page_sharing_bypass);
2040                         unlock_page(kpage);
2041                 }
2042                 put_page(kpage);
2043                 return;
2044         }
2045
2046         /*
2047          * If the hash value of the page has changed from the last time
2048          * we calculated it, this page is changing frequently: therefore we
2049          * don't want to insert it in the unstable tree, and we don't want
2050          * to waste our time searching for something identical to it there.
2051          */
2052         checksum = calc_checksum(page);
2053         if (rmap_item->oldchecksum != checksum) {
2054                 rmap_item->oldchecksum = checksum;
2055                 return;
2056         }
2057
2058         /*
2059          * Same checksum as an empty page. We attempt to merge it with the
2060          * appropriate zero page if the user enabled this via sysfs.
2061          */
2062         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2063                 struct vm_area_struct *vma;
2064
2065                 vma = find_mergeable_vma(rmap_item->mm, rmap_item->address);
2066                 err = try_to_merge_one_page(vma, page,
2067                                             ZERO_PAGE(rmap_item->address));
2068                 /*
2069                  * In case of failure, the page was not really empty, so we
2070                  * need to continue. Otherwise we're done.
2071                  */
2072                 if (!err)
2073                         return;
2074         }
2075         tree_rmap_item =
2076                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2077         if (tree_rmap_item) {
2078                 kpage = try_to_merge_two_pages(rmap_item, page,
2079                                                 tree_rmap_item, tree_page);
2080                 put_page(tree_page);
2081                 if (kpage) {
2082                         /*
2083                          * The pages were successfully merged: insert new
2084                          * node in the stable tree and add both rmap_items.
2085                          */
2086                         lock_page(kpage);
2087                         stable_node = stable_tree_insert(kpage);
2088                         if (stable_node) {
2089                                 stable_tree_append(tree_rmap_item, stable_node,
2090                                                    false);
2091                                 stable_tree_append(rmap_item, stable_node,
2092                                                    false);
2093                         }
2094                         unlock_page(kpage);
2095
2096                         /*
2097                          * If we fail to insert the page into the stable tree,
2098                          * we will have 2 virtual addresses that are pointing
2099                          * to a ksm page left outside the stable tree,
2100                          * in which case we need to break_cow on both.
2101                          */
2102                         if (!stable_node) {
2103                                 break_cow(tree_rmap_item);
2104                                 break_cow(rmap_item);
2105                         }
2106                 }
2107         }
2108 }
2109
2110 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2111                                             struct rmap_item **rmap_list,
2112                                             unsigned long addr)
2113 {
2114         struct rmap_item *rmap_item;
2115
2116         while (*rmap_list) {
2117                 rmap_item = *rmap_list;
2118                 if ((rmap_item->address & PAGE_MASK) == addr)
2119                         return rmap_item;
2120                 if (rmap_item->address > addr)
2121                         break;
2122                 *rmap_list = rmap_item->rmap_list;
2123                 remove_rmap_item_from_tree(rmap_item);
2124                 free_rmap_item(rmap_item);
2125         }
2126
2127         rmap_item = alloc_rmap_item();
2128         if (rmap_item) {
2129                 /* It has already been zeroed */
2130                 rmap_item->mm = mm_slot->mm;
2131                 rmap_item->address = addr;
2132                 rmap_item->rmap_list = *rmap_list;
2133                 *rmap_list = rmap_item;
2134         }
2135         return rmap_item;
2136 }
2137
2138 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2139 {
2140         struct mm_struct *mm;
2141         struct mm_slot *slot;
2142         struct vm_area_struct *vma;
2143         struct rmap_item *rmap_item;
2144         int nid;
2145
2146         if (list_empty(&ksm_mm_head.mm_list))
2147                 return NULL;
2148
2149         slot = ksm_scan.mm_slot;
2150         if (slot == &ksm_mm_head) {
2151                 /*
2152                  * A number of pages can hang around indefinitely on per-cpu
2153                  * pagevecs, raised page count preventing write_protect_page
2154                  * from merging them.  Though it doesn't really matter much,
2155                  * it is puzzling to see some stuck in pages_volatile until
2156                  * other activity jostles them out, and they also prevented
2157                  * LTP's KSM test from succeeding deterministically; so drain
2158                  * them here (here rather than on entry to ksm_do_scan(),
2159                  * so we don't IPI too often when pages_to_scan is set low).
2160                  */
2161                 lru_add_drain_all();
2162
2163                 /*
2164                  * Whereas stale stable_nodes on the stable_tree itself
2165                  * get pruned in the regular course of stable_tree_search(),
2166                  * those moved out to the migrate_nodes list can accumulate:
2167                  * so prune them once before each full scan.
2168                  */
2169                 if (!ksm_merge_across_nodes) {
2170                         struct stable_node *stable_node, *next;
2171                         struct page *page;
2172
2173                         list_for_each_entry_safe(stable_node, next,
2174                                                  &migrate_nodes, list) {
2175                                 page = get_ksm_page(stable_node, false);
2176                                 if (page)
2177                                         put_page(page);
2178                                 cond_resched();
2179                         }
2180                 }
2181
2182                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2183                         root_unstable_tree[nid] = RB_ROOT;
2184
2185                 spin_lock(&ksm_mmlist_lock);
2186                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2187                 ksm_scan.mm_slot = slot;
2188                 spin_unlock(&ksm_mmlist_lock);
2189                 /*
2190                  * Although we tested list_empty() above, a racing __ksm_exit
2191                  * of the last mm on the list may have removed it since then.
2192                  */
2193                 if (slot == &ksm_mm_head)
2194                         return NULL;
2195 next_mm:
2196                 ksm_scan.address = 0;
2197                 ksm_scan.rmap_list = &slot->rmap_list;
2198         }
2199
2200         mm = slot->mm;
2201         down_read(&mm->mmap_sem);
2202         if (ksm_test_exit(mm))
2203                 vma = NULL;
2204         else
2205                 vma = find_vma(mm, ksm_scan.address);
2206
2207         for (; vma; vma = vma->vm_next) {
2208                 if (!(vma->vm_flags & VM_MERGEABLE))
2209                         continue;
2210                 if (ksm_scan.address < vma->vm_start)
2211                         ksm_scan.address = vma->vm_start;
2212                 if (!vma->anon_vma)
2213                         ksm_scan.address = vma->vm_end;
2214
2215                 while (ksm_scan.address < vma->vm_end) {
2216                         if (ksm_test_exit(mm))
2217                                 break;
2218                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2219                         if (IS_ERR_OR_NULL(*page)) {
2220                                 ksm_scan.address += PAGE_SIZE;
2221                                 cond_resched();
2222                                 continue;
2223                         }
2224                         if (PageAnon(*page)) {
2225                                 flush_anon_page(vma, *page, ksm_scan.address);
2226                                 flush_dcache_page(*page);
2227                                 rmap_item = get_next_rmap_item(slot,
2228                                         ksm_scan.rmap_list, ksm_scan.address);
2229                                 if (rmap_item) {
2230                                         ksm_scan.rmap_list =
2231                                                         &rmap_item->rmap_list;
2232                                         ksm_scan.address += PAGE_SIZE;
2233                                 } else
2234                                         put_page(*page);
2235                                 up_read(&mm->mmap_sem);
2236                                 return rmap_item;
2237                         }
2238                         put_page(*page);
2239                         ksm_scan.address += PAGE_SIZE;
2240                         cond_resched();
2241                 }
2242         }
2243
2244         if (ksm_test_exit(mm)) {
2245                 ksm_scan.address = 0;
2246                 ksm_scan.rmap_list = &slot->rmap_list;
2247         }
2248         /*
2249          * Nuke all the rmap_items that are above this current rmap:
2250          * because there were no VM_MERGEABLE vmas with such addresses.
2251          */
2252         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2253
2254         spin_lock(&ksm_mmlist_lock);
2255         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2256                                                 struct mm_slot, mm_list);
2257         if (ksm_scan.address == 0) {
2258                 /*
2259                  * We've completed a full scan of all vmas, holding mmap_sem
2260                  * throughout, and found no VM_MERGEABLE: so do the same as
2261                  * __ksm_exit does to remove this mm from all our lists now.
2262                  * This applies either when cleaning up after __ksm_exit
2263                  * (but beware: we can reach here even before __ksm_exit),
2264                  * or when all VM_MERGEABLE areas have been unmapped (and
2265                  * mmap_sem then protects against race with MADV_MERGEABLE).
2266                  */
2267                 hash_del(&slot->link);
2268                 list_del(&slot->mm_list);
2269                 spin_unlock(&ksm_mmlist_lock);
2270
2271                 free_mm_slot(slot);
2272                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2273                 up_read(&mm->mmap_sem);
2274                 mmdrop(mm);
2275         } else {
2276                 up_read(&mm->mmap_sem);
2277                 /*
2278                  * up_read(&mm->mmap_sem) first because after
2279                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2280                  * already have been freed under us by __ksm_exit()
2281                  * because the "mm_slot" is still hashed and
2282                  * ksm_scan.mm_slot doesn't point to it anymore.
2283                  */
2284                 spin_unlock(&ksm_mmlist_lock);
2285         }
2286
2287         /* Repeat until we've completed scanning the whole list */
2288         slot = ksm_scan.mm_slot;
2289         if (slot != &ksm_mm_head)
2290                 goto next_mm;
2291
2292         ksm_scan.seqnr++;
2293         return NULL;
2294 }
2295
2296 /**
2297  * ksm_do_scan  - the ksm scanner main worker function.
2298  * @scan_npages - number of pages we want to scan before we return.
2299  */
2300 static void ksm_do_scan(unsigned int scan_npages)
2301 {
2302         struct rmap_item *rmap_item;
2303         struct page *uninitialized_var(page);
2304
2305         while (scan_npages-- && likely(!freezing(current))) {
2306                 cond_resched();
2307                 rmap_item = scan_get_next_rmap_item(&page);
2308                 if (!rmap_item)
2309                         return;
2310                 cmp_and_merge_page(page, rmap_item);
2311                 put_page(page);
2312         }
2313 }
2314
2315 static int ksmd_should_run(void)
2316 {
2317         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2318 }
2319
2320 static int ksm_scan_thread(void *nothing)
2321 {
2322         set_freezable();
2323         set_user_nice(current, 5);
2324
2325         while (!kthread_should_stop()) {
2326                 mutex_lock(&ksm_thread_mutex);
2327                 wait_while_offlining();
2328                 if (ksmd_should_run())
2329                         ksm_do_scan(ksm_thread_pages_to_scan);
2330                 mutex_unlock(&ksm_thread_mutex);
2331
2332                 try_to_freeze();
2333
2334                 if (ksmd_should_run()) {
2335                         schedule_timeout_interruptible(
2336                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2337                 } else {
2338                         wait_event_freezable(ksm_thread_wait,
2339                                 ksmd_should_run() || kthread_should_stop());
2340                 }
2341         }
2342         return 0;
2343 }
2344
2345 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2346                 unsigned long end, int advice, unsigned long *vm_flags)
2347 {
2348         struct mm_struct *mm = vma->vm_mm;
2349         int err;
2350
2351         switch (advice) {
2352         case MADV_MERGEABLE:
2353                 /*
2354                  * Be somewhat over-protective for now!
2355                  */
2356                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2357                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2358                                  VM_HUGETLB | VM_MIXEDMAP))
2359                         return 0;               /* just ignore the advice */
2360
2361 #ifdef VM_SAO
2362                 if (*vm_flags & VM_SAO)
2363                         return 0;
2364 #endif
2365
2366                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2367                         err = __ksm_enter(mm);
2368                         if (err)
2369                                 return err;
2370                 }
2371
2372                 *vm_flags |= VM_MERGEABLE;
2373                 break;
2374
2375         case MADV_UNMERGEABLE:
2376                 if (!(*vm_flags & VM_MERGEABLE))
2377                         return 0;               /* just ignore the advice */
2378
2379                 if (vma->anon_vma) {
2380                         err = unmerge_ksm_pages(vma, start, end);
2381                         if (err)
2382                                 return err;
2383                 }
2384
2385                 *vm_flags &= ~VM_MERGEABLE;
2386                 break;
2387         }
2388
2389         return 0;
2390 }
2391
2392 int __ksm_enter(struct mm_struct *mm)
2393 {
2394         struct mm_slot *mm_slot;
2395         int needs_wakeup;
2396
2397         mm_slot = alloc_mm_slot();
2398         if (!mm_slot)
2399                 return -ENOMEM;
2400
2401         /* Check ksm_run too?  Would need tighter locking */
2402         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2403
2404         spin_lock(&ksm_mmlist_lock);
2405         insert_to_mm_slots_hash(mm, mm_slot);
2406         /*
2407          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2408          * insert just behind the scanning cursor, to let the area settle
2409          * down a little; when fork is followed by immediate exec, we don't
2410          * want ksmd to waste time setting up and tearing down an rmap_list.
2411          *
2412          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2413          * scanning cursor, otherwise KSM pages in newly forked mms will be
2414          * missed: then we might as well insert at the end of the list.
2415          */
2416         if (ksm_run & KSM_RUN_UNMERGE)
2417                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2418         else
2419                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2420         spin_unlock(&ksm_mmlist_lock);
2421
2422         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2423         mmgrab(mm);
2424
2425         if (needs_wakeup)
2426                 wake_up_interruptible(&ksm_thread_wait);
2427
2428         return 0;
2429 }
2430
2431 void __ksm_exit(struct mm_struct *mm)
2432 {
2433         struct mm_slot *mm_slot;
2434         int easy_to_free = 0;
2435
2436         /*
2437          * This process is exiting: if it's straightforward (as is the
2438          * case when ksmd was never running), free mm_slot immediately.
2439          * But if it's at the cursor or has rmap_items linked to it, use
2440          * mmap_sem to synchronize with any break_cows before pagetables
2441          * are freed, and leave the mm_slot on the list for ksmd to free.
2442          * Beware: ksm may already have noticed it exiting and freed the slot.
2443          */
2444
2445         spin_lock(&ksm_mmlist_lock);
2446         mm_slot = get_mm_slot(mm);
2447         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2448                 if (!mm_slot->rmap_list) {
2449                         hash_del(&mm_slot->link);
2450                         list_del(&mm_slot->mm_list);
2451                         easy_to_free = 1;
2452                 } else {
2453                         list_move(&mm_slot->mm_list,
2454                                   &ksm_scan.mm_slot->mm_list);
2455                 }
2456         }
2457         spin_unlock(&ksm_mmlist_lock);
2458
2459         if (easy_to_free) {
2460                 free_mm_slot(mm_slot);
2461                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2462                 mmdrop(mm);
2463         } else if (mm_slot) {
2464                 down_write(&mm->mmap_sem);
2465                 up_write(&mm->mmap_sem);
2466         }
2467 }
2468
2469 struct page *ksm_might_need_to_copy(struct page *page,
2470                         struct vm_area_struct *vma, unsigned long address)
2471 {
2472         struct anon_vma *anon_vma = page_anon_vma(page);
2473         struct page *new_page;
2474
2475         if (PageKsm(page)) {
2476                 if (page_stable_node(page) &&
2477                     !(ksm_run & KSM_RUN_UNMERGE))
2478                         return page;    /* no need to copy it */
2479         } else if (!anon_vma) {
2480                 return page;            /* no need to copy it */
2481         } else if (anon_vma->root == vma->anon_vma->root &&
2482                  page->index == linear_page_index(vma, address)) {
2483                 return page;            /* still no need to copy it */
2484         }
2485         if (!PageUptodate(page))
2486                 return page;            /* let do_swap_page report the error */
2487
2488         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2489         if (new_page) {
2490                 copy_user_highpage(new_page, page, address, vma);
2491
2492                 SetPageDirty(new_page);
2493                 __SetPageUptodate(new_page);
2494                 __SetPageLocked(new_page);
2495         }
2496
2497         return new_page;
2498 }
2499
2500 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2501 {
2502         struct stable_node *stable_node;
2503         struct rmap_item *rmap_item;
2504         int search_new_forks = 0;
2505
2506         VM_BUG_ON_PAGE(!PageKsm(page), page);
2507
2508         /*
2509          * Rely on the page lock to protect against concurrent modifications
2510          * to that page's node of the stable tree.
2511          */
2512         VM_BUG_ON_PAGE(!PageLocked(page), page);
2513
2514         stable_node = page_stable_node(page);
2515         if (!stable_node)
2516                 return;
2517 again:
2518         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2519                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2520                 struct anon_vma_chain *vmac;
2521                 struct vm_area_struct *vma;
2522
2523                 cond_resched();
2524                 anon_vma_lock_read(anon_vma);
2525                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2526                                                0, ULONG_MAX) {
2527                         cond_resched();
2528                         vma = vmac->vma;
2529                         if (rmap_item->address < vma->vm_start ||
2530                             rmap_item->address >= vma->vm_end)
2531                                 continue;
2532                         /*
2533                          * Initially we examine only the vma which covers this
2534                          * rmap_item; but later, if there is still work to do,
2535                          * we examine covering vmas in other mms: in case they
2536                          * were forked from the original since ksmd passed.
2537                          */
2538                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2539                                 continue;
2540
2541                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2542                                 continue;
2543
2544                         if (!rwc->rmap_one(page, vma,
2545                                         rmap_item->address, rwc->arg)) {
2546                                 anon_vma_unlock_read(anon_vma);
2547                                 return;
2548                         }
2549                         if (rwc->done && rwc->done(page)) {
2550                                 anon_vma_unlock_read(anon_vma);
2551                                 return;
2552                         }
2553                 }
2554                 anon_vma_unlock_read(anon_vma);
2555         }
2556         if (!search_new_forks++)
2557                 goto again;
2558 }
2559
2560 #ifdef CONFIG_MIGRATION
2561 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2562 {
2563         struct stable_node *stable_node;
2564
2565         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2566         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2567         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2568
2569         stable_node = page_stable_node(newpage);
2570         if (stable_node) {
2571                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2572                 stable_node->kpfn = page_to_pfn(newpage);
2573                 /*
2574                  * newpage->mapping was set in advance; now we need smp_wmb()
2575                  * to make sure that the new stable_node->kpfn is visible
2576                  * to get_ksm_page() before it can see that oldpage->mapping
2577                  * has gone stale (or that PageSwapCache has been cleared).
2578                  */
2579                 smp_wmb();
2580                 set_page_stable_node(oldpage, NULL);
2581         }
2582 }
2583 #endif /* CONFIG_MIGRATION */
2584
2585 #ifdef CONFIG_MEMORY_HOTREMOVE
2586 static void wait_while_offlining(void)
2587 {
2588         while (ksm_run & KSM_RUN_OFFLINE) {
2589                 mutex_unlock(&ksm_thread_mutex);
2590                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2591                             TASK_UNINTERRUPTIBLE);
2592                 mutex_lock(&ksm_thread_mutex);
2593         }
2594 }
2595
2596 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2597                                          unsigned long start_pfn,
2598                                          unsigned long end_pfn)
2599 {
2600         if (stable_node->kpfn >= start_pfn &&
2601             stable_node->kpfn < end_pfn) {
2602                 /*
2603                  * Don't get_ksm_page, page has already gone:
2604                  * which is why we keep kpfn instead of page*
2605                  */
2606                 remove_node_from_stable_tree(stable_node);
2607                 return true;
2608         }
2609         return false;
2610 }
2611
2612 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2613                                            unsigned long start_pfn,
2614                                            unsigned long end_pfn,
2615                                            struct rb_root *root)
2616 {
2617         struct stable_node *dup;
2618         struct hlist_node *hlist_safe;
2619
2620         if (!is_stable_node_chain(stable_node)) {
2621                 VM_BUG_ON(is_stable_node_dup(stable_node));
2622                 return stable_node_dup_remove_range(stable_node, start_pfn,
2623                                                     end_pfn);
2624         }
2625
2626         hlist_for_each_entry_safe(dup, hlist_safe,
2627                                   &stable_node->hlist, hlist_dup) {
2628                 VM_BUG_ON(!is_stable_node_dup(dup));
2629                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2630         }
2631         if (hlist_empty(&stable_node->hlist)) {
2632                 free_stable_node_chain(stable_node, root);
2633                 return true; /* notify caller that tree was rebalanced */
2634         } else
2635                 return false;
2636 }
2637
2638 static void ksm_check_stable_tree(unsigned long start_pfn,
2639                                   unsigned long end_pfn)
2640 {
2641         struct stable_node *stable_node, *next;
2642         struct rb_node *node;
2643         int nid;
2644
2645         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2646                 node = rb_first(root_stable_tree + nid);
2647                 while (node) {
2648                         stable_node = rb_entry(node, struct stable_node, node);
2649                         if (stable_node_chain_remove_range(stable_node,
2650                                                            start_pfn, end_pfn,
2651                                                            root_stable_tree +
2652                                                            nid))
2653                                 node = rb_first(root_stable_tree + nid);
2654                         else
2655                                 node = rb_next(node);
2656                         cond_resched();
2657                 }
2658         }
2659         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2660                 if (stable_node->kpfn >= start_pfn &&
2661                     stable_node->kpfn < end_pfn)
2662                         remove_node_from_stable_tree(stable_node);
2663                 cond_resched();
2664         }
2665 }
2666
2667 static int ksm_memory_callback(struct notifier_block *self,
2668                                unsigned long action, void *arg)
2669 {
2670         struct memory_notify *mn = arg;
2671
2672         switch (action) {
2673         case MEM_GOING_OFFLINE:
2674                 /*
2675                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2676                  * and remove_all_stable_nodes() while memory is going offline:
2677                  * it is unsafe for them to touch the stable tree at this time.
2678                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2679                  * which do not need the ksm_thread_mutex are all safe.
2680                  */
2681                 mutex_lock(&ksm_thread_mutex);
2682                 ksm_run |= KSM_RUN_OFFLINE;
2683                 mutex_unlock(&ksm_thread_mutex);
2684                 break;
2685
2686         case MEM_OFFLINE:
2687                 /*
2688                  * Most of the work is done by page migration; but there might
2689                  * be a few stable_nodes left over, still pointing to struct
2690                  * pages which have been offlined: prune those from the tree,
2691                  * otherwise get_ksm_page() might later try to access a
2692                  * non-existent struct page.
2693                  */
2694                 ksm_check_stable_tree(mn->start_pfn,
2695                                       mn->start_pfn + mn->nr_pages);
2696                 /* fallthrough */
2697
2698         case MEM_CANCEL_OFFLINE:
2699                 mutex_lock(&ksm_thread_mutex);
2700                 ksm_run &= ~KSM_RUN_OFFLINE;
2701                 mutex_unlock(&ksm_thread_mutex);
2702
2703                 smp_mb();       /* wake_up_bit advises this */
2704                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2705                 break;
2706         }
2707         return NOTIFY_OK;
2708 }
2709 #else
2710 static void wait_while_offlining(void)
2711 {
2712 }
2713 #endif /* CONFIG_MEMORY_HOTREMOVE */
2714
2715 #ifdef CONFIG_SYSFS
2716 /*
2717  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2718  */
2719
2720 #define KSM_ATTR_RO(_name) \
2721         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2722 #define KSM_ATTR(_name) \
2723         static struct kobj_attribute _name##_attr = \
2724                 __ATTR(_name, 0644, _name##_show, _name##_store)
2725
2726 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2727                                     struct kobj_attribute *attr, char *buf)
2728 {
2729         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2730 }
2731
2732 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2733                                      struct kobj_attribute *attr,
2734                                      const char *buf, size_t count)
2735 {
2736         unsigned long msecs;
2737         int err;
2738
2739         err = kstrtoul(buf, 10, &msecs);
2740         if (err || msecs > UINT_MAX)
2741                 return -EINVAL;
2742
2743         ksm_thread_sleep_millisecs = msecs;
2744
2745         return count;
2746 }
2747 KSM_ATTR(sleep_millisecs);
2748
2749 static ssize_t pages_to_scan_show(struct kobject *kobj,
2750                                   struct kobj_attribute *attr, char *buf)
2751 {
2752         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2753 }
2754
2755 static ssize_t pages_to_scan_store(struct kobject *kobj,
2756                                    struct kobj_attribute *attr,
2757                                    const char *buf, size_t count)
2758 {
2759         int err;
2760         unsigned long nr_pages;
2761
2762         err = kstrtoul(buf, 10, &nr_pages);
2763         if (err || nr_pages > UINT_MAX)
2764                 return -EINVAL;
2765
2766         ksm_thread_pages_to_scan = nr_pages;
2767
2768         return count;
2769 }
2770 KSM_ATTR(pages_to_scan);
2771
2772 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2773                         char *buf)
2774 {
2775         return sprintf(buf, "%lu\n", ksm_run);
2776 }
2777
2778 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2779                          const char *buf, size_t count)
2780 {
2781         int err;
2782         unsigned long flags;
2783
2784         err = kstrtoul(buf, 10, &flags);
2785         if (err || flags > UINT_MAX)
2786                 return -EINVAL;
2787         if (flags > KSM_RUN_UNMERGE)
2788                 return -EINVAL;
2789
2790         /*
2791          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2792          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2793          * breaking COW to free the pages_shared (but leaves mm_slots
2794          * on the list for when ksmd may be set running again).
2795          */
2796
2797         mutex_lock(&ksm_thread_mutex);
2798         wait_while_offlining();
2799         if (ksm_run != flags) {
2800                 ksm_run = flags;
2801                 if (flags & KSM_RUN_UNMERGE) {
2802                         set_current_oom_origin();
2803                         err = unmerge_and_remove_all_rmap_items();
2804                         clear_current_oom_origin();
2805                         if (err) {
2806                                 ksm_run = KSM_RUN_STOP;
2807                                 count = err;
2808                         }
2809                 }
2810         }
2811         mutex_unlock(&ksm_thread_mutex);
2812
2813         if (flags & KSM_RUN_MERGE)
2814                 wake_up_interruptible(&ksm_thread_wait);
2815
2816         return count;
2817 }
2818 KSM_ATTR(run);
2819
2820 #ifdef CONFIG_NUMA
2821 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2822                                 struct kobj_attribute *attr, char *buf)
2823 {
2824         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2825 }
2826
2827 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2828                                    struct kobj_attribute *attr,
2829                                    const char *buf, size_t count)
2830 {
2831         int err;
2832         unsigned long knob;
2833
2834         err = kstrtoul(buf, 10, &knob);
2835         if (err)
2836                 return err;
2837         if (knob > 1)
2838                 return -EINVAL;
2839
2840         mutex_lock(&ksm_thread_mutex);
2841         wait_while_offlining();
2842         if (ksm_merge_across_nodes != knob) {
2843                 if (ksm_pages_shared || remove_all_stable_nodes())
2844                         err = -EBUSY;
2845                 else if (root_stable_tree == one_stable_tree) {
2846                         struct rb_root *buf;
2847                         /*
2848                          * This is the first time that we switch away from the
2849                          * default of merging across nodes: must now allocate
2850                          * a buffer to hold as many roots as may be needed.
2851                          * Allocate stable and unstable together:
2852                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2853                          */
2854                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2855                                       GFP_KERNEL);
2856                         /* Let us assume that RB_ROOT is NULL is zero */
2857                         if (!buf)
2858                                 err = -ENOMEM;
2859                         else {
2860                                 root_stable_tree = buf;
2861                                 root_unstable_tree = buf + nr_node_ids;
2862                                 /* Stable tree is empty but not the unstable */
2863                                 root_unstable_tree[0] = one_unstable_tree[0];
2864                         }
2865                 }
2866                 if (!err) {
2867                         ksm_merge_across_nodes = knob;
2868                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2869                 }
2870         }
2871         mutex_unlock(&ksm_thread_mutex);
2872
2873         return err ? err : count;
2874 }
2875 KSM_ATTR(merge_across_nodes);
2876 #endif
2877
2878 static ssize_t use_zero_pages_show(struct kobject *kobj,
2879                                 struct kobj_attribute *attr, char *buf)
2880 {
2881         return sprintf(buf, "%u\n", ksm_use_zero_pages);
2882 }
2883 static ssize_t use_zero_pages_store(struct kobject *kobj,
2884                                    struct kobj_attribute *attr,
2885                                    const char *buf, size_t count)
2886 {
2887         int err;
2888         bool value;
2889
2890         err = kstrtobool(buf, &value);
2891         if (err)
2892                 return -EINVAL;
2893
2894         ksm_use_zero_pages = value;
2895
2896         return count;
2897 }
2898 KSM_ATTR(use_zero_pages);
2899
2900 static ssize_t max_page_sharing_show(struct kobject *kobj,
2901                                      struct kobj_attribute *attr, char *buf)
2902 {
2903         return sprintf(buf, "%u\n", ksm_max_page_sharing);
2904 }
2905
2906 static ssize_t max_page_sharing_store(struct kobject *kobj,
2907                                       struct kobj_attribute *attr,
2908                                       const char *buf, size_t count)
2909 {
2910         int err;
2911         int knob;
2912
2913         err = kstrtoint(buf, 10, &knob);
2914         if (err)
2915                 return err;
2916         /*
2917          * When a KSM page is created it is shared by 2 mappings. This
2918          * being a signed comparison, it implicitly verifies it's not
2919          * negative.
2920          */
2921         if (knob < 2)
2922                 return -EINVAL;
2923
2924         if (READ_ONCE(ksm_max_page_sharing) == knob)
2925                 return count;
2926
2927         mutex_lock(&ksm_thread_mutex);
2928         wait_while_offlining();
2929         if (ksm_max_page_sharing != knob) {
2930                 if (ksm_pages_shared || remove_all_stable_nodes())
2931                         err = -EBUSY;
2932                 else
2933                         ksm_max_page_sharing = knob;
2934         }
2935         mutex_unlock(&ksm_thread_mutex);
2936
2937         return err ? err : count;
2938 }
2939 KSM_ATTR(max_page_sharing);
2940
2941 static ssize_t pages_shared_show(struct kobject *kobj,
2942                                  struct kobj_attribute *attr, char *buf)
2943 {
2944         return sprintf(buf, "%lu\n", ksm_pages_shared);
2945 }
2946 KSM_ATTR_RO(pages_shared);
2947
2948 static ssize_t pages_sharing_show(struct kobject *kobj,
2949                                   struct kobj_attribute *attr, char *buf)
2950 {
2951         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2952 }
2953 KSM_ATTR_RO(pages_sharing);
2954
2955 static ssize_t pages_unshared_show(struct kobject *kobj,
2956                                    struct kobj_attribute *attr, char *buf)
2957 {
2958         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2959 }
2960 KSM_ATTR_RO(pages_unshared);
2961
2962 static ssize_t pages_volatile_show(struct kobject *kobj,
2963                                    struct kobj_attribute *attr, char *buf)
2964 {
2965         long ksm_pages_volatile;
2966
2967         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2968                                 - ksm_pages_sharing - ksm_pages_unshared;
2969         /*
2970          * It was not worth any locking to calculate that statistic,
2971          * but it might therefore sometimes be negative: conceal that.
2972          */
2973         if (ksm_pages_volatile < 0)
2974                 ksm_pages_volatile = 0;
2975         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2976 }
2977 KSM_ATTR_RO(pages_volatile);
2978
2979 static ssize_t stable_node_dups_show(struct kobject *kobj,
2980                                      struct kobj_attribute *attr, char *buf)
2981 {
2982         return sprintf(buf, "%lu\n", ksm_stable_node_dups);
2983 }
2984 KSM_ATTR_RO(stable_node_dups);
2985
2986 static ssize_t stable_node_chains_show(struct kobject *kobj,
2987                                        struct kobj_attribute *attr, char *buf)
2988 {
2989         return sprintf(buf, "%lu\n", ksm_stable_node_chains);
2990 }
2991 KSM_ATTR_RO(stable_node_chains);
2992
2993 static ssize_t
2994 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
2995                                         struct kobj_attribute *attr,
2996                                         char *buf)
2997 {
2998         return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
2999 }
3000
3001 static ssize_t
3002 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3003                                          struct kobj_attribute *attr,
3004                                          const char *buf, size_t count)
3005 {
3006         unsigned long msecs;
3007         int err;
3008
3009         err = kstrtoul(buf, 10, &msecs);
3010         if (err || msecs > UINT_MAX)
3011                 return -EINVAL;
3012
3013         ksm_stable_node_chains_prune_millisecs = msecs;
3014
3015         return count;
3016 }
3017 KSM_ATTR(stable_node_chains_prune_millisecs);
3018
3019 static ssize_t full_scans_show(struct kobject *kobj,
3020                                struct kobj_attribute *attr, char *buf)
3021 {
3022         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3023 }
3024 KSM_ATTR_RO(full_scans);
3025
3026 static struct attribute *ksm_attrs[] = {
3027         &sleep_millisecs_attr.attr,
3028         &pages_to_scan_attr.attr,
3029         &run_attr.attr,
3030         &pages_shared_attr.attr,
3031         &pages_sharing_attr.attr,
3032         &pages_unshared_attr.attr,
3033         &pages_volatile_attr.attr,
3034         &full_scans_attr.attr,
3035 #ifdef CONFIG_NUMA
3036         &merge_across_nodes_attr.attr,
3037 #endif
3038         &max_page_sharing_attr.attr,
3039         &stable_node_chains_attr.attr,
3040         &stable_node_dups_attr.attr,
3041         &stable_node_chains_prune_millisecs_attr.attr,
3042         &use_zero_pages_attr.attr,
3043         NULL,
3044 };
3045
3046 static struct attribute_group ksm_attr_group = {
3047         .attrs = ksm_attrs,
3048         .name = "ksm",
3049 };
3050 #endif /* CONFIG_SYSFS */
3051
3052 static int __init ksm_init(void)
3053 {
3054         struct task_struct *ksm_thread;
3055         int err;
3056
3057         /* The correct value depends on page size and endianness */
3058         zero_checksum = calc_checksum(ZERO_PAGE(0));
3059         /* Default to false for backwards compatibility */
3060         ksm_use_zero_pages = false;
3061
3062         err = ksm_slab_init();
3063         if (err)
3064                 goto out;
3065
3066         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3067         if (IS_ERR(ksm_thread)) {
3068                 pr_err("ksm: creating kthread failed\n");
3069                 err = PTR_ERR(ksm_thread);
3070                 goto out_free;
3071         }
3072
3073 #ifdef CONFIG_SYSFS
3074         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3075         if (err) {
3076                 pr_err("ksm: register sysfs failed\n");
3077                 kthread_stop(ksm_thread);
3078                 goto out_free;
3079         }
3080 #else
3081         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3082
3083 #endif /* CONFIG_SYSFS */
3084
3085 #ifdef CONFIG_MEMORY_HOTREMOVE
3086         /* There is no significance to this priority 100 */
3087         hotplug_memory_notifier(ksm_memory_callback, 100);
3088 #endif
3089         return 0;
3090
3091 out_free:
3092         ksm_slab_free();
3093 out:
3094         return err;
3095 }
3096 subsys_initcall(ksm_init);