mm: migration: add migrate_entry_wait_huge()
[linux-3.10.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/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hashtable.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
39 #include <linux/numa.h>
40
41 #include <asm/tlbflush.h>
42 #include "internal.h"
43
44 #ifdef CONFIG_NUMA
45 #define NUMA(x)         (x)
46 #define DO_NUMA(x)      do { (x); } while (0)
47 #else
48 #define NUMA(x)         (0)
49 #define DO_NUMA(x)      do { } while (0)
50 #endif
51
52 /*
53  * A few notes about the KSM scanning process,
54  * to make it easier to understand the data structures below:
55  *
56  * In order to reduce excessive scanning, KSM sorts the memory pages by their
57  * contents into a data structure that holds pointers to the pages' locations.
58  *
59  * Since the contents of the pages may change at any moment, KSM cannot just
60  * insert the pages into a normal sorted tree and expect it to find anything.
61  * Therefore KSM uses two data structures - the stable and the unstable tree.
62  *
63  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64  * by their contents.  Because each such page is write-protected, searching on
65  * this tree is fully assured to be working (except when pages are unmapped),
66  * and therefore this tree is called the stable tree.
67  *
68  * In addition to the stable tree, KSM uses a second data structure called the
69  * unstable tree: this tree holds pointers to pages which have been found to
70  * be "unchanged for a period of time".  The unstable tree sorts these pages
71  * by their contents, but since they are not write-protected, KSM cannot rely
72  * upon the unstable tree to work correctly - the unstable tree is liable to
73  * be corrupted as its contents are modified, and so it is called unstable.
74  *
75  * KSM solves this problem by several techniques:
76  *
77  * 1) The unstable tree is flushed every time KSM completes scanning all
78  *    memory areas, and then the tree is rebuilt again from the beginning.
79  * 2) KSM will only insert into the unstable tree, pages whose hash value
80  *    has not changed since the previous scan of all memory areas.
81  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82  *    colors of the nodes and not on their contents, assuring that even when
83  *    the tree gets "corrupted" it won't get out of balance, so scanning time
84  *    remains the same (also, searching and inserting nodes in an rbtree uses
85  *    the same algorithm, so we have no overhead when we flush and rebuild).
86  * 4) KSM never flushes the stable tree, which means that even if it were to
87  *    take 10 attempts to find a page in the unstable tree, once it is found,
88  *    it is secured in the stable tree.  (When we scan a new page, we first
89  *    compare it against the stable tree, and then against the unstable tree.)
90  *
91  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92  * stable trees and multiple unstable trees: one of each for each NUMA node.
93  */
94
95 /**
96  * struct mm_slot - ksm information per mm that is being scanned
97  * @link: link to the mm_slots hash list
98  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
99  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
100  * @mm: the mm that this information is valid for
101  */
102 struct mm_slot {
103         struct hlist_node link;
104         struct list_head mm_list;
105         struct rmap_item *rmap_list;
106         struct mm_struct *mm;
107 };
108
109 /**
110  * struct ksm_scan - cursor for scanning
111  * @mm_slot: the current mm_slot we are scanning
112  * @address: the next address inside that to be scanned
113  * @rmap_list: link to the next rmap to be scanned in the rmap_list
114  * @seqnr: count of completed full scans (needed when removing unstable node)
115  *
116  * There is only the one ksm_scan instance of this cursor structure.
117  */
118 struct ksm_scan {
119         struct mm_slot *mm_slot;
120         unsigned long address;
121         struct rmap_item **rmap_list;
122         unsigned long seqnr;
123 };
124
125 /**
126  * struct stable_node - node of the stable rbtree
127  * @node: rb node of this ksm page in the stable tree
128  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129  * @list: linked into migrate_nodes, pending placement in the proper node tree
130  * @hlist: hlist head of rmap_items using this ksm page
131  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
133  */
134 struct stable_node {
135         union {
136                 struct rb_node node;    /* when node of stable tree */
137                 struct {                /* when listed for migration */
138                         struct list_head *head;
139                         struct list_head list;
140                 };
141         };
142         struct hlist_head hlist;
143         unsigned long kpfn;
144 #ifdef CONFIG_NUMA
145         int nid;
146 #endif
147 };
148
149 /**
150  * struct rmap_item - reverse mapping item for virtual addresses
151  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
152  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
153  * @nid: NUMA node id of unstable tree in which linked (may not match page)
154  * @mm: the memory structure this rmap_item is pointing into
155  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156  * @oldchecksum: previous checksum of the page at that virtual address
157  * @node: rb node of this rmap_item in the unstable tree
158  * @head: pointer to stable_node heading this list in the stable tree
159  * @hlist: link into hlist of rmap_items hanging off that stable_node
160  */
161 struct rmap_item {
162         struct rmap_item *rmap_list;
163         union {
164                 struct anon_vma *anon_vma;      /* when stable */
165 #ifdef CONFIG_NUMA
166                 int nid;                /* when node of unstable tree */
167 #endif
168         };
169         struct mm_struct *mm;
170         unsigned long address;          /* + low bits used for flags below */
171         unsigned int oldchecksum;       /* when unstable */
172         union {
173                 struct rb_node node;    /* when node of unstable tree */
174                 struct {                /* when listed from stable tree */
175                         struct stable_node *head;
176                         struct hlist_node hlist;
177                 };
178         };
179 };
180
181 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
182 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
183 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
184
185 /* The stable and unstable tree heads */
186 static struct rb_root one_stable_tree[1] = { RB_ROOT };
187 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
188 static struct rb_root *root_stable_tree = one_stable_tree;
189 static struct rb_root *root_unstable_tree = one_unstable_tree;
190
191 /* Recently migrated nodes of stable tree, pending proper placement */
192 static LIST_HEAD(migrate_nodes);
193
194 #define MM_SLOTS_HASH_BITS 10
195 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
196
197 static struct mm_slot ksm_mm_head = {
198         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
199 };
200 static struct ksm_scan ksm_scan = {
201         .mm_slot = &ksm_mm_head,
202 };
203
204 static struct kmem_cache *rmap_item_cache;
205 static struct kmem_cache *stable_node_cache;
206 static struct kmem_cache *mm_slot_cache;
207
208 /* The number of nodes in the stable tree */
209 static unsigned long ksm_pages_shared;
210
211 /* The number of page slots additionally sharing those nodes */
212 static unsigned long ksm_pages_sharing;
213
214 /* The number of nodes in the unstable tree */
215 static unsigned long ksm_pages_unshared;
216
217 /* The number of rmap_items in use: to calculate pages_volatile */
218 static unsigned long ksm_rmap_items;
219
220 /* Number of pages ksmd should scan in one batch */
221 static unsigned int ksm_thread_pages_to_scan = 100;
222
223 /* Milliseconds ksmd should sleep between batches */
224 static unsigned int ksm_thread_sleep_millisecs = 20;
225
226 #ifdef CONFIG_NUMA
227 /* Zeroed when merging across nodes is not allowed */
228 static unsigned int ksm_merge_across_nodes = 1;
229 static int ksm_nr_node_ids = 1;
230 #else
231 #define ksm_merge_across_nodes  1U
232 #define ksm_nr_node_ids         1
233 #endif
234
235 #define KSM_RUN_STOP    0
236 #define KSM_RUN_MERGE   1
237 #define KSM_RUN_UNMERGE 2
238 #define KSM_RUN_OFFLINE 4
239 static unsigned long ksm_run = KSM_RUN_STOP;
240 static void wait_while_offlining(void);
241
242 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
243 static DEFINE_MUTEX(ksm_thread_mutex);
244 static DEFINE_SPINLOCK(ksm_mmlist_lock);
245
246 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
247                 sizeof(struct __struct), __alignof__(struct __struct),\
248                 (__flags), NULL)
249
250 static int __init ksm_slab_init(void)
251 {
252         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
253         if (!rmap_item_cache)
254                 goto out;
255
256         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
257         if (!stable_node_cache)
258                 goto out_free1;
259
260         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
261         if (!mm_slot_cache)
262                 goto out_free2;
263
264         return 0;
265
266 out_free2:
267         kmem_cache_destroy(stable_node_cache);
268 out_free1:
269         kmem_cache_destroy(rmap_item_cache);
270 out:
271         return -ENOMEM;
272 }
273
274 static void __init ksm_slab_free(void)
275 {
276         kmem_cache_destroy(mm_slot_cache);
277         kmem_cache_destroy(stable_node_cache);
278         kmem_cache_destroy(rmap_item_cache);
279         mm_slot_cache = NULL;
280 }
281
282 static inline struct rmap_item *alloc_rmap_item(void)
283 {
284         struct rmap_item *rmap_item;
285
286         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
287         if (rmap_item)
288                 ksm_rmap_items++;
289         return rmap_item;
290 }
291
292 static inline void free_rmap_item(struct rmap_item *rmap_item)
293 {
294         ksm_rmap_items--;
295         rmap_item->mm = NULL;   /* debug safety */
296         kmem_cache_free(rmap_item_cache, rmap_item);
297 }
298
299 static inline struct stable_node *alloc_stable_node(void)
300 {
301         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
302 }
303
304 static inline void free_stable_node(struct stable_node *stable_node)
305 {
306         kmem_cache_free(stable_node_cache, stable_node);
307 }
308
309 static inline struct mm_slot *alloc_mm_slot(void)
310 {
311         if (!mm_slot_cache)     /* initialization failed */
312                 return NULL;
313         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
314 }
315
316 static inline void free_mm_slot(struct mm_slot *mm_slot)
317 {
318         kmem_cache_free(mm_slot_cache, mm_slot);
319 }
320
321 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
322 {
323         struct mm_slot *slot;
324
325         hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
326                 if (slot->mm == mm)
327                         return slot;
328
329         return NULL;
330 }
331
332 static void insert_to_mm_slots_hash(struct mm_struct *mm,
333                                     struct mm_slot *mm_slot)
334 {
335         mm_slot->mm = mm;
336         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
337 }
338
339 /*
340  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
341  * page tables after it has passed through ksm_exit() - which, if necessary,
342  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
343  * a special flag: they can just back out as soon as mm_users goes to zero.
344  * ksm_test_exit() is used throughout to make this test for exit: in some
345  * places for correctness, in some places just to avoid unnecessary work.
346  */
347 static inline bool ksm_test_exit(struct mm_struct *mm)
348 {
349         return atomic_read(&mm->mm_users) == 0;
350 }
351
352 /*
353  * We use break_ksm to break COW on a ksm page: it's a stripped down
354  *
355  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
356  *              put_page(page);
357  *
358  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
359  * in case the application has unmapped and remapped mm,addr meanwhile.
360  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
361  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
362  */
363 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
364 {
365         struct page *page;
366         int ret = 0;
367
368         do {
369                 cond_resched();
370                 page = follow_page(vma, addr, FOLL_GET | FOLL_MIGRATION);
371                 if (IS_ERR_OR_NULL(page))
372                         break;
373                 if (PageKsm(page))
374                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
375                                                         FAULT_FLAG_WRITE);
376                 else
377                         ret = VM_FAULT_WRITE;
378                 put_page(page);
379         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
380         /*
381          * We must loop because handle_mm_fault() may back out if there's
382          * any difficulty e.g. if pte accessed bit gets updated concurrently.
383          *
384          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
385          * COW has been broken, even if the vma does not permit VM_WRITE;
386          * but note that a concurrent fault might break PageKsm for us.
387          *
388          * VM_FAULT_SIGBUS could occur if we race with truncation of the
389          * backing file, which also invalidates anonymous pages: that's
390          * okay, that truncation will have unmapped the PageKsm for us.
391          *
392          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
393          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
394          * current task has TIF_MEMDIE set, and will be OOM killed on return
395          * to user; and ksmd, having no mm, would never be chosen for that.
396          *
397          * But if the mm is in a limited mem_cgroup, then the fault may fail
398          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
399          * even ksmd can fail in this way - though it's usually breaking ksm
400          * just to undo a merge it made a moment before, so unlikely to oom.
401          *
402          * That's a pity: we might therefore have more kernel pages allocated
403          * than we're counting as nodes in the stable tree; but ksm_do_scan
404          * will retry to break_cow on each pass, so should recover the page
405          * in due course.  The important thing is to not let VM_MERGEABLE
406          * be cleared while any such pages might remain in the area.
407          */
408         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
409 }
410
411 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
412                 unsigned long addr)
413 {
414         struct vm_area_struct *vma;
415         if (ksm_test_exit(mm))
416                 return NULL;
417         vma = find_vma(mm, addr);
418         if (!vma || vma->vm_start > addr)
419                 return NULL;
420         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
421                 return NULL;
422         return vma;
423 }
424
425 static void break_cow(struct rmap_item *rmap_item)
426 {
427         struct mm_struct *mm = rmap_item->mm;
428         unsigned long addr = rmap_item->address;
429         struct vm_area_struct *vma;
430
431         /*
432          * It is not an accident that whenever we want to break COW
433          * to undo, we also need to drop a reference to the anon_vma.
434          */
435         put_anon_vma(rmap_item->anon_vma);
436
437         down_read(&mm->mmap_sem);
438         vma = find_mergeable_vma(mm, addr);
439         if (vma)
440                 break_ksm(vma, addr);
441         up_read(&mm->mmap_sem);
442 }
443
444 static struct page *page_trans_compound_anon(struct page *page)
445 {
446         if (PageTransCompound(page)) {
447                 struct page *head = compound_trans_head(page);
448                 /*
449                  * head may actually be splitted and freed from under
450                  * us but it's ok here.
451                  */
452                 if (PageAnon(head))
453                         return head;
454         }
455         return NULL;
456 }
457
458 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
459 {
460         struct mm_struct *mm = rmap_item->mm;
461         unsigned long addr = rmap_item->address;
462         struct vm_area_struct *vma;
463         struct page *page;
464
465         down_read(&mm->mmap_sem);
466         vma = find_mergeable_vma(mm, addr);
467         if (!vma)
468                 goto out;
469
470         page = follow_page(vma, addr, FOLL_GET);
471         if (IS_ERR_OR_NULL(page))
472                 goto out;
473         if (PageAnon(page) || page_trans_compound_anon(page)) {
474                 flush_anon_page(vma, page, addr);
475                 flush_dcache_page(page);
476         } else {
477                 put_page(page);
478 out:            page = NULL;
479         }
480         up_read(&mm->mmap_sem);
481         return page;
482 }
483
484 /*
485  * This helper is used for getting right index into array of tree roots.
486  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
487  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
488  * every node has its own stable and unstable tree.
489  */
490 static inline int get_kpfn_nid(unsigned long kpfn)
491 {
492         return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
493 }
494
495 static void remove_node_from_stable_tree(struct stable_node *stable_node)
496 {
497         struct rmap_item *rmap_item;
498
499         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
500                 if (rmap_item->hlist.next)
501                         ksm_pages_sharing--;
502                 else
503                         ksm_pages_shared--;
504                 put_anon_vma(rmap_item->anon_vma);
505                 rmap_item->address &= PAGE_MASK;
506                 cond_resched();
507         }
508
509         if (stable_node->head == &migrate_nodes)
510                 list_del(&stable_node->list);
511         else
512                 rb_erase(&stable_node->node,
513                          root_stable_tree + NUMA(stable_node->nid));
514         free_stable_node(stable_node);
515 }
516
517 /*
518  * get_ksm_page: checks if the page indicated by the stable node
519  * is still its ksm page, despite having held no reference to it.
520  * In which case we can trust the content of the page, and it
521  * returns the gotten page; but if the page has now been zapped,
522  * remove the stale node from the stable tree and return NULL.
523  * But beware, the stable node's page might be being migrated.
524  *
525  * You would expect the stable_node to hold a reference to the ksm page.
526  * But if it increments the page's count, swapping out has to wait for
527  * ksmd to come around again before it can free the page, which may take
528  * seconds or even minutes: much too unresponsive.  So instead we use a
529  * "keyhole reference": access to the ksm page from the stable node peeps
530  * out through its keyhole to see if that page still holds the right key,
531  * pointing back to this stable node.  This relies on freeing a PageAnon
532  * page to reset its page->mapping to NULL, and relies on no other use of
533  * a page to put something that might look like our key in page->mapping.
534  * is on its way to being freed; but it is an anomaly to bear in mind.
535  */
536 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
537 {
538         struct page *page;
539         void *expected_mapping;
540         unsigned long kpfn;
541
542         expected_mapping = (void *)stable_node +
543                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
544 again:
545         kpfn = ACCESS_ONCE(stable_node->kpfn);
546         page = pfn_to_page(kpfn);
547
548         /*
549          * page is computed from kpfn, so on most architectures reading
550          * page->mapping is naturally ordered after reading node->kpfn,
551          * but on Alpha we need to be more careful.
552          */
553         smp_read_barrier_depends();
554         if (ACCESS_ONCE(page->mapping) != expected_mapping)
555                 goto stale;
556
557         /*
558          * We cannot do anything with the page while its refcount is 0.
559          * Usually 0 means free, or tail of a higher-order page: in which
560          * case this node is no longer referenced, and should be freed;
561          * however, it might mean that the page is under page_freeze_refs().
562          * The __remove_mapping() case is easy, again the node is now stale;
563          * but if page is swapcache in migrate_page_move_mapping(), it might
564          * still be our page, in which case it's essential to keep the node.
565          */
566         while (!get_page_unless_zero(page)) {
567                 /*
568                  * Another check for page->mapping != expected_mapping would
569                  * work here too.  We have chosen the !PageSwapCache test to
570                  * optimize the common case, when the page is or is about to
571                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
572                  * in the freeze_refs section of __remove_mapping(); but Anon
573                  * page->mapping reset to NULL later, in free_pages_prepare().
574                  */
575                 if (!PageSwapCache(page))
576                         goto stale;
577                 cpu_relax();
578         }
579
580         if (ACCESS_ONCE(page->mapping) != expected_mapping) {
581                 put_page(page);
582                 goto stale;
583         }
584
585         if (lock_it) {
586                 lock_page(page);
587                 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
588                         unlock_page(page);
589                         put_page(page);
590                         goto stale;
591                 }
592         }
593         return page;
594
595 stale:
596         /*
597          * We come here from above when page->mapping or !PageSwapCache
598          * suggests that the node is stale; but it might be under migration.
599          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
600          * before checking whether node->kpfn has been changed.
601          */
602         smp_rmb();
603         if (ACCESS_ONCE(stable_node->kpfn) != kpfn)
604                 goto again;
605         remove_node_from_stable_tree(stable_node);
606         return NULL;
607 }
608
609 /*
610  * Removing rmap_item from stable or unstable tree.
611  * This function will clean the information from the stable/unstable tree.
612  */
613 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
614 {
615         if (rmap_item->address & STABLE_FLAG) {
616                 struct stable_node *stable_node;
617                 struct page *page;
618
619                 stable_node = rmap_item->head;
620                 page = get_ksm_page(stable_node, true);
621                 if (!page)
622                         goto out;
623
624                 hlist_del(&rmap_item->hlist);
625                 unlock_page(page);
626                 put_page(page);
627
628                 if (stable_node->hlist.first)
629                         ksm_pages_sharing--;
630                 else
631                         ksm_pages_shared--;
632
633                 put_anon_vma(rmap_item->anon_vma);
634                 rmap_item->address &= PAGE_MASK;
635
636         } else if (rmap_item->address & UNSTABLE_FLAG) {
637                 unsigned char age;
638                 /*
639                  * Usually ksmd can and must skip the rb_erase, because
640                  * root_unstable_tree was already reset to RB_ROOT.
641                  * But be careful when an mm is exiting: do the rb_erase
642                  * if this rmap_item was inserted by this scan, rather
643                  * than left over from before.
644                  */
645                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
646                 BUG_ON(age > 1);
647                 if (!age)
648                         rb_erase(&rmap_item->node,
649                                  root_unstable_tree + NUMA(rmap_item->nid));
650                 ksm_pages_unshared--;
651                 rmap_item->address &= PAGE_MASK;
652         }
653 out:
654         cond_resched();         /* we're called from many long loops */
655 }
656
657 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
658                                        struct rmap_item **rmap_list)
659 {
660         while (*rmap_list) {
661                 struct rmap_item *rmap_item = *rmap_list;
662                 *rmap_list = rmap_item->rmap_list;
663                 remove_rmap_item_from_tree(rmap_item);
664                 free_rmap_item(rmap_item);
665         }
666 }
667
668 /*
669  * Though it's very tempting to unmerge rmap_items from stable tree rather
670  * than check every pte of a given vma, the locking doesn't quite work for
671  * that - an rmap_item is assigned to the stable tree after inserting ksm
672  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
673  * rmap_items from parent to child at fork time (so as not to waste time
674  * if exit comes before the next scan reaches it).
675  *
676  * Similarly, although we'd like to remove rmap_items (so updating counts
677  * and freeing memory) when unmerging an area, it's easier to leave that
678  * to the next pass of ksmd - consider, for example, how ksmd might be
679  * in cmp_and_merge_page on one of the rmap_items we would be removing.
680  */
681 static int unmerge_ksm_pages(struct vm_area_struct *vma,
682                              unsigned long start, unsigned long end)
683 {
684         unsigned long addr;
685         int err = 0;
686
687         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
688                 if (ksm_test_exit(vma->vm_mm))
689                         break;
690                 if (signal_pending(current))
691                         err = -ERESTARTSYS;
692                 else
693                         err = break_ksm(vma, addr);
694         }
695         return err;
696 }
697
698 #ifdef CONFIG_SYSFS
699 /*
700  * Only called through the sysfs control interface:
701  */
702 static int remove_stable_node(struct stable_node *stable_node)
703 {
704         struct page *page;
705         int err;
706
707         page = get_ksm_page(stable_node, true);
708         if (!page) {
709                 /*
710                  * get_ksm_page did remove_node_from_stable_tree itself.
711                  */
712                 return 0;
713         }
714
715         if (WARN_ON_ONCE(page_mapped(page))) {
716                 /*
717                  * This should not happen: but if it does, just refuse to let
718                  * merge_across_nodes be switched - there is no need to panic.
719                  */
720                 err = -EBUSY;
721         } else {
722                 /*
723                  * The stable node did not yet appear stale to get_ksm_page(),
724                  * since that allows for an unmapped ksm page to be recognized
725                  * right up until it is freed; but the node is safe to remove.
726                  * This page might be in a pagevec waiting to be freed,
727                  * or it might be PageSwapCache (perhaps under writeback),
728                  * or it might have been removed from swapcache a moment ago.
729                  */
730                 set_page_stable_node(page, NULL);
731                 remove_node_from_stable_tree(stable_node);
732                 err = 0;
733         }
734
735         unlock_page(page);
736         put_page(page);
737         return err;
738 }
739
740 static int remove_all_stable_nodes(void)
741 {
742         struct stable_node *stable_node;
743         struct list_head *this, *next;
744         int nid;
745         int err = 0;
746
747         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
748                 while (root_stable_tree[nid].rb_node) {
749                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
750                                                 struct stable_node, node);
751                         if (remove_stable_node(stable_node)) {
752                                 err = -EBUSY;
753                                 break;  /* proceed to next nid */
754                         }
755                         cond_resched();
756                 }
757         }
758         list_for_each_safe(this, next, &migrate_nodes) {
759                 stable_node = list_entry(this, struct stable_node, list);
760                 if (remove_stable_node(stable_node))
761                         err = -EBUSY;
762                 cond_resched();
763         }
764         return err;
765 }
766
767 static int unmerge_and_remove_all_rmap_items(void)
768 {
769         struct mm_slot *mm_slot;
770         struct mm_struct *mm;
771         struct vm_area_struct *vma;
772         int err = 0;
773
774         spin_lock(&ksm_mmlist_lock);
775         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
776                                                 struct mm_slot, mm_list);
777         spin_unlock(&ksm_mmlist_lock);
778
779         for (mm_slot = ksm_scan.mm_slot;
780                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
781                 mm = mm_slot->mm;
782                 down_read(&mm->mmap_sem);
783                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
784                         if (ksm_test_exit(mm))
785                                 break;
786                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
787                                 continue;
788                         err = unmerge_ksm_pages(vma,
789                                                 vma->vm_start, vma->vm_end);
790                         if (err)
791                                 goto error;
792                 }
793
794                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
795
796                 spin_lock(&ksm_mmlist_lock);
797                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
798                                                 struct mm_slot, mm_list);
799                 if (ksm_test_exit(mm)) {
800                         hash_del(&mm_slot->link);
801                         list_del(&mm_slot->mm_list);
802                         spin_unlock(&ksm_mmlist_lock);
803
804                         free_mm_slot(mm_slot);
805                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
806                         up_read(&mm->mmap_sem);
807                         mmdrop(mm);
808                 } else {
809                         spin_unlock(&ksm_mmlist_lock);
810                         up_read(&mm->mmap_sem);
811                 }
812         }
813
814         /* Clean up stable nodes, but don't worry if some are still busy */
815         remove_all_stable_nodes();
816         ksm_scan.seqnr = 0;
817         return 0;
818
819 error:
820         up_read(&mm->mmap_sem);
821         spin_lock(&ksm_mmlist_lock);
822         ksm_scan.mm_slot = &ksm_mm_head;
823         spin_unlock(&ksm_mmlist_lock);
824         return err;
825 }
826 #endif /* CONFIG_SYSFS */
827
828 static u32 calc_checksum(struct page *page)
829 {
830         u32 checksum;
831         void *addr = kmap_atomic(page);
832         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
833         kunmap_atomic(addr);
834         return checksum;
835 }
836
837 static int memcmp_pages(struct page *page1, struct page *page2)
838 {
839         char *addr1, *addr2;
840         int ret;
841
842         addr1 = kmap_atomic(page1);
843         addr2 = kmap_atomic(page2);
844         ret = memcmp(addr1, addr2, PAGE_SIZE);
845         kunmap_atomic(addr2);
846         kunmap_atomic(addr1);
847         return ret;
848 }
849
850 static inline int pages_identical(struct page *page1, struct page *page2)
851 {
852         return !memcmp_pages(page1, page2);
853 }
854
855 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
856                               pte_t *orig_pte)
857 {
858         struct mm_struct *mm = vma->vm_mm;
859         unsigned long addr;
860         pte_t *ptep;
861         spinlock_t *ptl;
862         int swapped;
863         int err = -EFAULT;
864         unsigned long mmun_start;       /* For mmu_notifiers */
865         unsigned long mmun_end;         /* For mmu_notifiers */
866
867         addr = page_address_in_vma(page, vma);
868         if (addr == -EFAULT)
869                 goto out;
870
871         BUG_ON(PageTransCompound(page));
872
873         mmun_start = addr;
874         mmun_end   = addr + PAGE_SIZE;
875         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
876
877         ptep = page_check_address(page, mm, addr, &ptl, 0);
878         if (!ptep)
879                 goto out_mn;
880
881         if (pte_write(*ptep) || pte_dirty(*ptep)) {
882                 pte_t entry;
883
884                 swapped = PageSwapCache(page);
885                 flush_cache_page(vma, addr, page_to_pfn(page));
886                 /*
887                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
888                  * take any lock, therefore the check that we are going to make
889                  * with the pagecount against the mapcount is racey and
890                  * O_DIRECT can happen right after the check.
891                  * So we clear the pte and flush the tlb before the check
892                  * this assure us that no O_DIRECT can happen after the check
893                  * or in the middle of the check.
894                  */
895                 entry = ptep_clear_flush(vma, addr, ptep);
896                 /*
897                  * Check that no O_DIRECT or similar I/O is in progress on the
898                  * page
899                  */
900                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
901                         set_pte_at(mm, addr, ptep, entry);
902                         goto out_unlock;
903                 }
904                 if (pte_dirty(entry))
905                         set_page_dirty(page);
906                 entry = pte_mkclean(pte_wrprotect(entry));
907                 set_pte_at_notify(mm, addr, ptep, entry);
908         }
909         *orig_pte = *ptep;
910         err = 0;
911
912 out_unlock:
913         pte_unmap_unlock(ptep, ptl);
914 out_mn:
915         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
916 out:
917         return err;
918 }
919
920 /**
921  * replace_page - replace page in vma by new ksm page
922  * @vma:      vma that holds the pte pointing to page
923  * @page:     the page we are replacing by kpage
924  * @kpage:    the ksm page we replace page by
925  * @orig_pte: the original value of the pte
926  *
927  * Returns 0 on success, -EFAULT on failure.
928  */
929 static int replace_page(struct vm_area_struct *vma, struct page *page,
930                         struct page *kpage, pte_t orig_pte)
931 {
932         struct mm_struct *mm = vma->vm_mm;
933         pmd_t *pmd;
934         pte_t *ptep;
935         spinlock_t *ptl;
936         unsigned long addr;
937         int err = -EFAULT;
938         unsigned long mmun_start;       /* For mmu_notifiers */
939         unsigned long mmun_end;         /* For mmu_notifiers */
940
941         addr = page_address_in_vma(page, vma);
942         if (addr == -EFAULT)
943                 goto out;
944
945         pmd = mm_find_pmd(mm, addr);
946         if (!pmd)
947                 goto out;
948         BUG_ON(pmd_trans_huge(*pmd));
949
950         mmun_start = addr;
951         mmun_end   = addr + PAGE_SIZE;
952         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
953
954         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
955         if (!pte_same(*ptep, orig_pte)) {
956                 pte_unmap_unlock(ptep, ptl);
957                 goto out_mn;
958         }
959
960         get_page(kpage);
961         page_add_anon_rmap(kpage, vma, addr);
962
963         flush_cache_page(vma, addr, pte_pfn(*ptep));
964         ptep_clear_flush(vma, addr, ptep);
965         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
966
967         page_remove_rmap(page);
968         if (!page_mapped(page))
969                 try_to_free_swap(page);
970         put_page(page);
971
972         pte_unmap_unlock(ptep, ptl);
973         err = 0;
974 out_mn:
975         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
976 out:
977         return err;
978 }
979
980 static int page_trans_compound_anon_split(struct page *page)
981 {
982         int ret = 0;
983         struct page *transhuge_head = page_trans_compound_anon(page);
984         if (transhuge_head) {
985                 /* Get the reference on the head to split it. */
986                 if (get_page_unless_zero(transhuge_head)) {
987                         /*
988                          * Recheck we got the reference while the head
989                          * was still anonymous.
990                          */
991                         if (PageAnon(transhuge_head))
992                                 ret = split_huge_page(transhuge_head);
993                         else
994                                 /*
995                                  * Retry later if split_huge_page run
996                                  * from under us.
997                                  */
998                                 ret = 1;
999                         put_page(transhuge_head);
1000                 } else
1001                         /* Retry later if split_huge_page run from under us. */
1002                         ret = 1;
1003         }
1004         return ret;
1005 }
1006
1007 /*
1008  * try_to_merge_one_page - take two pages and merge them into one
1009  * @vma: the vma that holds the pte pointing to page
1010  * @page: the PageAnon page that we want to replace with kpage
1011  * @kpage: the PageKsm page that we want to map instead of page,
1012  *         or NULL the first time when we want to use page as kpage.
1013  *
1014  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1015  */
1016 static int try_to_merge_one_page(struct vm_area_struct *vma,
1017                                  struct page *page, struct page *kpage)
1018 {
1019         pte_t orig_pte = __pte(0);
1020         int err = -EFAULT;
1021
1022         if (page == kpage)                      /* ksm page forked */
1023                 return 0;
1024
1025         if (!(vma->vm_flags & VM_MERGEABLE))
1026                 goto out;
1027         if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1028                 goto out;
1029         BUG_ON(PageTransCompound(page));
1030         if (!PageAnon(page))
1031                 goto out;
1032
1033         /*
1034          * We need the page lock to read a stable PageSwapCache in
1035          * write_protect_page().  We use trylock_page() instead of
1036          * lock_page() because we don't want to wait here - we
1037          * prefer to continue scanning and merging different pages,
1038          * then come back to this page when it is unlocked.
1039          */
1040         if (!trylock_page(page))
1041                 goto out;
1042         /*
1043          * If this anonymous page is mapped only here, its pte may need
1044          * to be write-protected.  If it's mapped elsewhere, all of its
1045          * ptes are necessarily already write-protected.  But in either
1046          * case, we need to lock and check page_count is not raised.
1047          */
1048         if (write_protect_page(vma, page, &orig_pte) == 0) {
1049                 if (!kpage) {
1050                         /*
1051                          * While we hold page lock, upgrade page from
1052                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1053                          * stable_tree_insert() will update stable_node.
1054                          */
1055                         set_page_stable_node(page, NULL);
1056                         mark_page_accessed(page);
1057                         err = 0;
1058                 } else if (pages_identical(page, kpage))
1059                         err = replace_page(vma, page, kpage, orig_pte);
1060         }
1061
1062         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1063                 munlock_vma_page(page);
1064                 if (!PageMlocked(kpage)) {
1065                         unlock_page(page);
1066                         lock_page(kpage);
1067                         mlock_vma_page(kpage);
1068                         page = kpage;           /* for final unlock */
1069                 }
1070         }
1071
1072         unlock_page(page);
1073 out:
1074         return err;
1075 }
1076
1077 /*
1078  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1079  * but no new kernel page is allocated: kpage must already be a ksm page.
1080  *
1081  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1082  */
1083 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1084                                       struct page *page, struct page *kpage)
1085 {
1086         struct mm_struct *mm = rmap_item->mm;
1087         struct vm_area_struct *vma;
1088         int err = -EFAULT;
1089
1090         down_read(&mm->mmap_sem);
1091         if (ksm_test_exit(mm))
1092                 goto out;
1093         vma = find_vma(mm, rmap_item->address);
1094         if (!vma || vma->vm_start > rmap_item->address)
1095                 goto out;
1096
1097         err = try_to_merge_one_page(vma, page, kpage);
1098         if (err)
1099                 goto out;
1100
1101         /* Unstable nid is in union with stable anon_vma: remove first */
1102         remove_rmap_item_from_tree(rmap_item);
1103
1104         /* Must get reference to anon_vma while still holding mmap_sem */
1105         rmap_item->anon_vma = vma->anon_vma;
1106         get_anon_vma(vma->anon_vma);
1107 out:
1108         up_read(&mm->mmap_sem);
1109         return err;
1110 }
1111
1112 /*
1113  * try_to_merge_two_pages - take two identical pages and prepare them
1114  * to be merged into one page.
1115  *
1116  * This function returns the kpage if we successfully merged two identical
1117  * pages into one ksm page, NULL otherwise.
1118  *
1119  * Note that this function upgrades page to ksm page: if one of the pages
1120  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1121  */
1122 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1123                                            struct page *page,
1124                                            struct rmap_item *tree_rmap_item,
1125                                            struct page *tree_page)
1126 {
1127         int err;
1128
1129         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1130         if (!err) {
1131                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1132                                                         tree_page, page);
1133                 /*
1134                  * If that fails, we have a ksm page with only one pte
1135                  * pointing to it: so break it.
1136                  */
1137                 if (err)
1138                         break_cow(rmap_item);
1139         }
1140         return err ? NULL : page;
1141 }
1142
1143 /*
1144  * stable_tree_search - search for page inside the stable tree
1145  *
1146  * This function checks if there is a page inside the stable tree
1147  * with identical content to the page that we are scanning right now.
1148  *
1149  * This function returns the stable tree node of identical content if found,
1150  * NULL otherwise.
1151  */
1152 static struct page *stable_tree_search(struct page *page)
1153 {
1154         int nid;
1155         struct rb_root *root;
1156         struct rb_node **new;
1157         struct rb_node *parent;
1158         struct stable_node *stable_node;
1159         struct stable_node *page_node;
1160
1161         page_node = page_stable_node(page);
1162         if (page_node && page_node->head != &migrate_nodes) {
1163                 /* ksm page forked */
1164                 get_page(page);
1165                 return page;
1166         }
1167
1168         nid = get_kpfn_nid(page_to_pfn(page));
1169         root = root_stable_tree + nid;
1170 again:
1171         new = &root->rb_node;
1172         parent = NULL;
1173
1174         while (*new) {
1175                 struct page *tree_page;
1176                 int ret;
1177
1178                 cond_resched();
1179                 stable_node = rb_entry(*new, struct stable_node, node);
1180                 tree_page = get_ksm_page(stable_node, false);
1181                 if (!tree_page)
1182                         return NULL;
1183
1184                 ret = memcmp_pages(page, tree_page);
1185                 put_page(tree_page);
1186
1187                 parent = *new;
1188                 if (ret < 0)
1189                         new = &parent->rb_left;
1190                 else if (ret > 0)
1191                         new = &parent->rb_right;
1192                 else {
1193                         /*
1194                          * Lock and unlock the stable_node's page (which
1195                          * might already have been migrated) so that page
1196                          * migration is sure to notice its raised count.
1197                          * It would be more elegant to return stable_node
1198                          * than kpage, but that involves more changes.
1199                          */
1200                         tree_page = get_ksm_page(stable_node, true);
1201                         if (tree_page) {
1202                                 unlock_page(tree_page);
1203                                 if (get_kpfn_nid(stable_node->kpfn) !=
1204                                                 NUMA(stable_node->nid)) {
1205                                         put_page(tree_page);
1206                                         goto replace;
1207                                 }
1208                                 return tree_page;
1209                         }
1210                         /*
1211                          * There is now a place for page_node, but the tree may
1212                          * have been rebalanced, so re-evaluate parent and new.
1213                          */
1214                         if (page_node)
1215                                 goto again;
1216                         return NULL;
1217                 }
1218         }
1219
1220         if (!page_node)
1221                 return NULL;
1222
1223         list_del(&page_node->list);
1224         DO_NUMA(page_node->nid = nid);
1225         rb_link_node(&page_node->node, parent, new);
1226         rb_insert_color(&page_node->node, root);
1227         get_page(page);
1228         return page;
1229
1230 replace:
1231         if (page_node) {
1232                 list_del(&page_node->list);
1233                 DO_NUMA(page_node->nid = nid);
1234                 rb_replace_node(&stable_node->node, &page_node->node, root);
1235                 get_page(page);
1236         } else {
1237                 rb_erase(&stable_node->node, root);
1238                 page = NULL;
1239         }
1240         stable_node->head = &migrate_nodes;
1241         list_add(&stable_node->list, stable_node->head);
1242         return page;
1243 }
1244
1245 /*
1246  * stable_tree_insert - insert stable tree node pointing to new ksm page
1247  * into the stable tree.
1248  *
1249  * This function returns the stable tree node just allocated on success,
1250  * NULL otherwise.
1251  */
1252 static struct stable_node *stable_tree_insert(struct page *kpage)
1253 {
1254         int nid;
1255         unsigned long kpfn;
1256         struct rb_root *root;
1257         struct rb_node **new;
1258         struct rb_node *parent = NULL;
1259         struct stable_node *stable_node;
1260
1261         kpfn = page_to_pfn(kpage);
1262         nid = get_kpfn_nid(kpfn);
1263         root = root_stable_tree + nid;
1264         new = &root->rb_node;
1265
1266         while (*new) {
1267                 struct page *tree_page;
1268                 int ret;
1269
1270                 cond_resched();
1271                 stable_node = rb_entry(*new, struct stable_node, node);
1272                 tree_page = get_ksm_page(stable_node, false);
1273                 if (!tree_page)
1274                         return NULL;
1275
1276                 ret = memcmp_pages(kpage, tree_page);
1277                 put_page(tree_page);
1278
1279                 parent = *new;
1280                 if (ret < 0)
1281                         new = &parent->rb_left;
1282                 else if (ret > 0)
1283                         new = &parent->rb_right;
1284                 else {
1285                         /*
1286                          * It is not a bug that stable_tree_search() didn't
1287                          * find this node: because at that time our page was
1288                          * not yet write-protected, so may have changed since.
1289                          */
1290                         return NULL;
1291                 }
1292         }
1293
1294         stable_node = alloc_stable_node();
1295         if (!stable_node)
1296                 return NULL;
1297
1298         INIT_HLIST_HEAD(&stable_node->hlist);
1299         stable_node->kpfn = kpfn;
1300         set_page_stable_node(kpage, stable_node);
1301         DO_NUMA(stable_node->nid = nid);
1302         rb_link_node(&stable_node->node, parent, new);
1303         rb_insert_color(&stable_node->node, root);
1304
1305         return stable_node;
1306 }
1307
1308 /*
1309  * unstable_tree_search_insert - search for identical page,
1310  * else insert rmap_item into the unstable tree.
1311  *
1312  * This function searches for a page in the unstable tree identical to the
1313  * page currently being scanned; and if no identical page is found in the
1314  * tree, we insert rmap_item as a new object into the unstable tree.
1315  *
1316  * This function returns pointer to rmap_item found to be identical
1317  * to the currently scanned page, NULL otherwise.
1318  *
1319  * This function does both searching and inserting, because they share
1320  * the same walking algorithm in an rbtree.
1321  */
1322 static
1323 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1324                                               struct page *page,
1325                                               struct page **tree_pagep)
1326 {
1327         struct rb_node **new;
1328         struct rb_root *root;
1329         struct rb_node *parent = NULL;
1330         int nid;
1331
1332         nid = get_kpfn_nid(page_to_pfn(page));
1333         root = root_unstable_tree + nid;
1334         new = &root->rb_node;
1335
1336         while (*new) {
1337                 struct rmap_item *tree_rmap_item;
1338                 struct page *tree_page;
1339                 int ret;
1340
1341                 cond_resched();
1342                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1343                 tree_page = get_mergeable_page(tree_rmap_item);
1344                 if (IS_ERR_OR_NULL(tree_page))
1345                         return NULL;
1346
1347                 /*
1348                  * Don't substitute a ksm page for a forked page.
1349                  */
1350                 if (page == tree_page) {
1351                         put_page(tree_page);
1352                         return NULL;
1353                 }
1354
1355                 ret = memcmp_pages(page, tree_page);
1356
1357                 parent = *new;
1358                 if (ret < 0) {
1359                         put_page(tree_page);
1360                         new = &parent->rb_left;
1361                 } else if (ret > 0) {
1362                         put_page(tree_page);
1363                         new = &parent->rb_right;
1364                 } else if (!ksm_merge_across_nodes &&
1365                            page_to_nid(tree_page) != nid) {
1366                         /*
1367                          * If tree_page has been migrated to another NUMA node,
1368                          * it will be flushed out and put in the right unstable
1369                          * tree next time: only merge with it when across_nodes.
1370                          */
1371                         put_page(tree_page);
1372                         return NULL;
1373                 } else {
1374                         *tree_pagep = tree_page;
1375                         return tree_rmap_item;
1376                 }
1377         }
1378
1379         rmap_item->address |= UNSTABLE_FLAG;
1380         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1381         DO_NUMA(rmap_item->nid = nid);
1382         rb_link_node(&rmap_item->node, parent, new);
1383         rb_insert_color(&rmap_item->node, root);
1384
1385         ksm_pages_unshared++;
1386         return NULL;
1387 }
1388
1389 /*
1390  * stable_tree_append - add another rmap_item to the linked list of
1391  * rmap_items hanging off a given node of the stable tree, all sharing
1392  * the same ksm page.
1393  */
1394 static void stable_tree_append(struct rmap_item *rmap_item,
1395                                struct stable_node *stable_node)
1396 {
1397         rmap_item->head = stable_node;
1398         rmap_item->address |= STABLE_FLAG;
1399         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1400
1401         if (rmap_item->hlist.next)
1402                 ksm_pages_sharing++;
1403         else
1404                 ksm_pages_shared++;
1405 }
1406
1407 /*
1408  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1409  * if not, compare checksum to previous and if it's the same, see if page can
1410  * be inserted into the unstable tree, or merged with a page already there and
1411  * both transferred to the stable tree.
1412  *
1413  * @page: the page that we are searching identical page to.
1414  * @rmap_item: the reverse mapping into the virtual address of this page
1415  */
1416 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1417 {
1418         struct rmap_item *tree_rmap_item;
1419         struct page *tree_page = NULL;
1420         struct stable_node *stable_node;
1421         struct page *kpage;
1422         unsigned int checksum;
1423         int err;
1424
1425         stable_node = page_stable_node(page);
1426         if (stable_node) {
1427                 if (stable_node->head != &migrate_nodes &&
1428                     get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1429                         rb_erase(&stable_node->node,
1430                                  root_stable_tree + NUMA(stable_node->nid));
1431                         stable_node->head = &migrate_nodes;
1432                         list_add(&stable_node->list, stable_node->head);
1433                 }
1434                 if (stable_node->head != &migrate_nodes &&
1435                     rmap_item->head == stable_node)
1436                         return;
1437         }
1438
1439         /* We first start with searching the page inside the stable tree */
1440         kpage = stable_tree_search(page);
1441         if (kpage == page && rmap_item->head == stable_node) {
1442                 put_page(kpage);
1443                 return;
1444         }
1445
1446         remove_rmap_item_from_tree(rmap_item);
1447
1448         if (kpage) {
1449                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1450                 if (!err) {
1451                         /*
1452                          * The page was successfully merged:
1453                          * add its rmap_item to the stable tree.
1454                          */
1455                         lock_page(kpage);
1456                         stable_tree_append(rmap_item, page_stable_node(kpage));
1457                         unlock_page(kpage);
1458                 }
1459                 put_page(kpage);
1460                 return;
1461         }
1462
1463         /*
1464          * If the hash value of the page has changed from the last time
1465          * we calculated it, this page is changing frequently: therefore we
1466          * don't want to insert it in the unstable tree, and we don't want
1467          * to waste our time searching for something identical to it there.
1468          */
1469         checksum = calc_checksum(page);
1470         if (rmap_item->oldchecksum != checksum) {
1471                 rmap_item->oldchecksum = checksum;
1472                 return;
1473         }
1474
1475         tree_rmap_item =
1476                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1477         if (tree_rmap_item) {
1478                 kpage = try_to_merge_two_pages(rmap_item, page,
1479                                                 tree_rmap_item, tree_page);
1480                 put_page(tree_page);
1481                 if (kpage) {
1482                         /*
1483                          * The pages were successfully merged: insert new
1484                          * node in the stable tree and add both rmap_items.
1485                          */
1486                         lock_page(kpage);
1487                         stable_node = stable_tree_insert(kpage);
1488                         if (stable_node) {
1489                                 stable_tree_append(tree_rmap_item, stable_node);
1490                                 stable_tree_append(rmap_item, stable_node);
1491                         }
1492                         unlock_page(kpage);
1493
1494                         /*
1495                          * If we fail to insert the page into the stable tree,
1496                          * we will have 2 virtual addresses that are pointing
1497                          * to a ksm page left outside the stable tree,
1498                          * in which case we need to break_cow on both.
1499                          */
1500                         if (!stable_node) {
1501                                 break_cow(tree_rmap_item);
1502                                 break_cow(rmap_item);
1503                         }
1504                 }
1505         }
1506 }
1507
1508 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1509                                             struct rmap_item **rmap_list,
1510                                             unsigned long addr)
1511 {
1512         struct rmap_item *rmap_item;
1513
1514         while (*rmap_list) {
1515                 rmap_item = *rmap_list;
1516                 if ((rmap_item->address & PAGE_MASK) == addr)
1517                         return rmap_item;
1518                 if (rmap_item->address > addr)
1519                         break;
1520                 *rmap_list = rmap_item->rmap_list;
1521                 remove_rmap_item_from_tree(rmap_item);
1522                 free_rmap_item(rmap_item);
1523         }
1524
1525         rmap_item = alloc_rmap_item();
1526         if (rmap_item) {
1527                 /* It has already been zeroed */
1528                 rmap_item->mm = mm_slot->mm;
1529                 rmap_item->address = addr;
1530                 rmap_item->rmap_list = *rmap_list;
1531                 *rmap_list = rmap_item;
1532         }
1533         return rmap_item;
1534 }
1535
1536 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1537 {
1538         struct mm_struct *mm;
1539         struct mm_slot *slot;
1540         struct vm_area_struct *vma;
1541         struct rmap_item *rmap_item;
1542         int nid;
1543
1544         if (list_empty(&ksm_mm_head.mm_list))
1545                 return NULL;
1546
1547         slot = ksm_scan.mm_slot;
1548         if (slot == &ksm_mm_head) {
1549                 /*
1550                  * A number of pages can hang around indefinitely on per-cpu
1551                  * pagevecs, raised page count preventing write_protect_page
1552                  * from merging them.  Though it doesn't really matter much,
1553                  * it is puzzling to see some stuck in pages_volatile until
1554                  * other activity jostles them out, and they also prevented
1555                  * LTP's KSM test from succeeding deterministically; so drain
1556                  * them here (here rather than on entry to ksm_do_scan(),
1557                  * so we don't IPI too often when pages_to_scan is set low).
1558                  */
1559                 lru_add_drain_all();
1560
1561                 /*
1562                  * Whereas stale stable_nodes on the stable_tree itself
1563                  * get pruned in the regular course of stable_tree_search(),
1564                  * those moved out to the migrate_nodes list can accumulate:
1565                  * so prune them once before each full scan.
1566                  */
1567                 if (!ksm_merge_across_nodes) {
1568                         struct stable_node *stable_node;
1569                         struct list_head *this, *next;
1570                         struct page *page;
1571
1572                         list_for_each_safe(this, next, &migrate_nodes) {
1573                                 stable_node = list_entry(this,
1574                                                 struct stable_node, list);
1575                                 page = get_ksm_page(stable_node, false);
1576                                 if (page)
1577                                         put_page(page);
1578                                 cond_resched();
1579                         }
1580                 }
1581
1582                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1583                         root_unstable_tree[nid] = RB_ROOT;
1584
1585                 spin_lock(&ksm_mmlist_lock);
1586                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1587                 ksm_scan.mm_slot = slot;
1588                 spin_unlock(&ksm_mmlist_lock);
1589                 /*
1590                  * Although we tested list_empty() above, a racing __ksm_exit
1591                  * of the last mm on the list may have removed it since then.
1592                  */
1593                 if (slot == &ksm_mm_head)
1594                         return NULL;
1595 next_mm:
1596                 ksm_scan.address = 0;
1597                 ksm_scan.rmap_list = &slot->rmap_list;
1598         }
1599
1600         mm = slot->mm;
1601         down_read(&mm->mmap_sem);
1602         if (ksm_test_exit(mm))
1603                 vma = NULL;
1604         else
1605                 vma = find_vma(mm, ksm_scan.address);
1606
1607         for (; vma; vma = vma->vm_next) {
1608                 if (!(vma->vm_flags & VM_MERGEABLE))
1609                         continue;
1610                 if (ksm_scan.address < vma->vm_start)
1611                         ksm_scan.address = vma->vm_start;
1612                 if (!vma->anon_vma)
1613                         ksm_scan.address = vma->vm_end;
1614
1615                 while (ksm_scan.address < vma->vm_end) {
1616                         if (ksm_test_exit(mm))
1617                                 break;
1618                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1619                         if (IS_ERR_OR_NULL(*page)) {
1620                                 ksm_scan.address += PAGE_SIZE;
1621                                 cond_resched();
1622                                 continue;
1623                         }
1624                         if (PageAnon(*page) ||
1625                             page_trans_compound_anon(*page)) {
1626                                 flush_anon_page(vma, *page, ksm_scan.address);
1627                                 flush_dcache_page(*page);
1628                                 rmap_item = get_next_rmap_item(slot,
1629                                         ksm_scan.rmap_list, ksm_scan.address);
1630                                 if (rmap_item) {
1631                                         ksm_scan.rmap_list =
1632                                                         &rmap_item->rmap_list;
1633                                         ksm_scan.address += PAGE_SIZE;
1634                                 } else
1635                                         put_page(*page);
1636                                 up_read(&mm->mmap_sem);
1637                                 return rmap_item;
1638                         }
1639                         put_page(*page);
1640                         ksm_scan.address += PAGE_SIZE;
1641                         cond_resched();
1642                 }
1643         }
1644
1645         if (ksm_test_exit(mm)) {
1646                 ksm_scan.address = 0;
1647                 ksm_scan.rmap_list = &slot->rmap_list;
1648         }
1649         /*
1650          * Nuke all the rmap_items that are above this current rmap:
1651          * because there were no VM_MERGEABLE vmas with such addresses.
1652          */
1653         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1654
1655         spin_lock(&ksm_mmlist_lock);
1656         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1657                                                 struct mm_slot, mm_list);
1658         if (ksm_scan.address == 0) {
1659                 /*
1660                  * We've completed a full scan of all vmas, holding mmap_sem
1661                  * throughout, and found no VM_MERGEABLE: so do the same as
1662                  * __ksm_exit does to remove this mm from all our lists now.
1663                  * This applies either when cleaning up after __ksm_exit
1664                  * (but beware: we can reach here even before __ksm_exit),
1665                  * or when all VM_MERGEABLE areas have been unmapped (and
1666                  * mmap_sem then protects against race with MADV_MERGEABLE).
1667                  */
1668                 hash_del(&slot->link);
1669                 list_del(&slot->mm_list);
1670                 spin_unlock(&ksm_mmlist_lock);
1671
1672                 free_mm_slot(slot);
1673                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1674                 up_read(&mm->mmap_sem);
1675                 mmdrop(mm);
1676         } else {
1677                 spin_unlock(&ksm_mmlist_lock);
1678                 up_read(&mm->mmap_sem);
1679         }
1680
1681         /* Repeat until we've completed scanning the whole list */
1682         slot = ksm_scan.mm_slot;
1683         if (slot != &ksm_mm_head)
1684                 goto next_mm;
1685
1686         ksm_scan.seqnr++;
1687         return NULL;
1688 }
1689
1690 /**
1691  * ksm_do_scan  - the ksm scanner main worker function.
1692  * @scan_npages - number of pages we want to scan before we return.
1693  */
1694 static void ksm_do_scan(unsigned int scan_npages)
1695 {
1696         struct rmap_item *rmap_item;
1697         struct page *uninitialized_var(page);
1698
1699         while (scan_npages-- && likely(!freezing(current))) {
1700                 cond_resched();
1701                 rmap_item = scan_get_next_rmap_item(&page);
1702                 if (!rmap_item)
1703                         return;
1704                 cmp_and_merge_page(page, rmap_item);
1705                 put_page(page);
1706         }
1707 }
1708
1709 static int ksmd_should_run(void)
1710 {
1711         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1712 }
1713
1714 static int ksm_scan_thread(void *nothing)
1715 {
1716         set_freezable();
1717         set_user_nice(current, 5);
1718
1719         while (!kthread_should_stop()) {
1720                 mutex_lock(&ksm_thread_mutex);
1721                 wait_while_offlining();
1722                 if (ksmd_should_run())
1723                         ksm_do_scan(ksm_thread_pages_to_scan);
1724                 mutex_unlock(&ksm_thread_mutex);
1725
1726                 try_to_freeze();
1727
1728                 if (ksmd_should_run()) {
1729                         schedule_timeout_interruptible(
1730                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1731                 } else {
1732                         wait_event_freezable(ksm_thread_wait,
1733                                 ksmd_should_run() || kthread_should_stop());
1734                 }
1735         }
1736         return 0;
1737 }
1738
1739 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1740                 unsigned long end, int advice, unsigned long *vm_flags)
1741 {
1742         struct mm_struct *mm = vma->vm_mm;
1743         int err;
1744
1745         switch (advice) {
1746         case MADV_MERGEABLE:
1747                 /*
1748                  * Be somewhat over-protective for now!
1749                  */
1750                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1751                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1752                                  VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1753                         return 0;               /* just ignore the advice */
1754
1755 #ifdef VM_SAO
1756                 if (*vm_flags & VM_SAO)
1757                         return 0;
1758 #endif
1759
1760                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1761                         err = __ksm_enter(mm);
1762                         if (err)
1763                                 return err;
1764                 }
1765
1766                 *vm_flags |= VM_MERGEABLE;
1767                 break;
1768
1769         case MADV_UNMERGEABLE:
1770                 if (!(*vm_flags & VM_MERGEABLE))
1771                         return 0;               /* just ignore the advice */
1772
1773                 if (vma->anon_vma) {
1774                         err = unmerge_ksm_pages(vma, start, end);
1775                         if (err)
1776                                 return err;
1777                 }
1778
1779                 *vm_flags &= ~VM_MERGEABLE;
1780                 break;
1781         }
1782
1783         return 0;
1784 }
1785
1786 int __ksm_enter(struct mm_struct *mm)
1787 {
1788         struct mm_slot *mm_slot;
1789         int needs_wakeup;
1790
1791         mm_slot = alloc_mm_slot();
1792         if (!mm_slot)
1793                 return -ENOMEM;
1794
1795         /* Check ksm_run too?  Would need tighter locking */
1796         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1797
1798         spin_lock(&ksm_mmlist_lock);
1799         insert_to_mm_slots_hash(mm, mm_slot);
1800         /*
1801          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1802          * insert just behind the scanning cursor, to let the area settle
1803          * down a little; when fork is followed by immediate exec, we don't
1804          * want ksmd to waste time setting up and tearing down an rmap_list.
1805          *
1806          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1807          * scanning cursor, otherwise KSM pages in newly forked mms will be
1808          * missed: then we might as well insert at the end of the list.
1809          */
1810         if (ksm_run & KSM_RUN_UNMERGE)
1811                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1812         else
1813                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1814         spin_unlock(&ksm_mmlist_lock);
1815
1816         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1817         atomic_inc(&mm->mm_count);
1818
1819         if (needs_wakeup)
1820                 wake_up_interruptible(&ksm_thread_wait);
1821
1822         return 0;
1823 }
1824
1825 void __ksm_exit(struct mm_struct *mm)
1826 {
1827         struct mm_slot *mm_slot;
1828         int easy_to_free = 0;
1829
1830         /*
1831          * This process is exiting: if it's straightforward (as is the
1832          * case when ksmd was never running), free mm_slot immediately.
1833          * But if it's at the cursor or has rmap_items linked to it, use
1834          * mmap_sem to synchronize with any break_cows before pagetables
1835          * are freed, and leave the mm_slot on the list for ksmd to free.
1836          * Beware: ksm may already have noticed it exiting and freed the slot.
1837          */
1838
1839         spin_lock(&ksm_mmlist_lock);
1840         mm_slot = get_mm_slot(mm);
1841         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1842                 if (!mm_slot->rmap_list) {
1843                         hash_del(&mm_slot->link);
1844                         list_del(&mm_slot->mm_list);
1845                         easy_to_free = 1;
1846                 } else {
1847                         list_move(&mm_slot->mm_list,
1848                                   &ksm_scan.mm_slot->mm_list);
1849                 }
1850         }
1851         spin_unlock(&ksm_mmlist_lock);
1852
1853         if (easy_to_free) {
1854                 free_mm_slot(mm_slot);
1855                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1856                 mmdrop(mm);
1857         } else if (mm_slot) {
1858                 down_write(&mm->mmap_sem);
1859                 up_write(&mm->mmap_sem);
1860         }
1861 }
1862
1863 struct page *ksm_might_need_to_copy(struct page *page,
1864                         struct vm_area_struct *vma, unsigned long address)
1865 {
1866         struct anon_vma *anon_vma = page_anon_vma(page);
1867         struct page *new_page;
1868
1869         if (PageKsm(page)) {
1870                 if (page_stable_node(page) &&
1871                     !(ksm_run & KSM_RUN_UNMERGE))
1872                         return page;    /* no need to copy it */
1873         } else if (!anon_vma) {
1874                 return page;            /* no need to copy it */
1875         } else if (anon_vma->root == vma->anon_vma->root &&
1876                  page->index == linear_page_index(vma, address)) {
1877                 return page;            /* still no need to copy it */
1878         }
1879         if (!PageUptodate(page))
1880                 return page;            /* let do_swap_page report the error */
1881
1882         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1883         if (new_page) {
1884                 copy_user_highpage(new_page, page, address, vma);
1885
1886                 SetPageDirty(new_page);
1887                 __SetPageUptodate(new_page);
1888                 __set_page_locked(new_page);
1889         }
1890
1891         return new_page;
1892 }
1893
1894 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1895                         unsigned long *vm_flags)
1896 {
1897         struct stable_node *stable_node;
1898         struct rmap_item *rmap_item;
1899         unsigned int mapcount = page_mapcount(page);
1900         int referenced = 0;
1901         int search_new_forks = 0;
1902
1903         VM_BUG_ON(!PageKsm(page));
1904         VM_BUG_ON(!PageLocked(page));
1905
1906         stable_node = page_stable_node(page);
1907         if (!stable_node)
1908                 return 0;
1909 again:
1910         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1911                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1912                 struct anon_vma_chain *vmac;
1913                 struct vm_area_struct *vma;
1914
1915                 anon_vma_lock_read(anon_vma);
1916                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1917                                                0, ULONG_MAX) {
1918                         vma = vmac->vma;
1919                         if (rmap_item->address < vma->vm_start ||
1920                             rmap_item->address >= vma->vm_end)
1921                                 continue;
1922                         /*
1923                          * Initially we examine only the vma which covers this
1924                          * rmap_item; but later, if there is still work to do,
1925                          * we examine covering vmas in other mms: in case they
1926                          * were forked from the original since ksmd passed.
1927                          */
1928                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1929                                 continue;
1930
1931                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1932                                 continue;
1933
1934                         referenced += page_referenced_one(page, vma,
1935                                 rmap_item->address, &mapcount, vm_flags);
1936                         if (!search_new_forks || !mapcount)
1937                                 break;
1938                 }
1939                 anon_vma_unlock_read(anon_vma);
1940                 if (!mapcount)
1941                         goto out;
1942         }
1943         if (!search_new_forks++)
1944                 goto again;
1945 out:
1946         return referenced;
1947 }
1948
1949 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1950 {
1951         struct stable_node *stable_node;
1952         struct rmap_item *rmap_item;
1953         int ret = SWAP_AGAIN;
1954         int search_new_forks = 0;
1955
1956         VM_BUG_ON(!PageKsm(page));
1957         VM_BUG_ON(!PageLocked(page));
1958
1959         stable_node = page_stable_node(page);
1960         if (!stable_node)
1961                 return SWAP_FAIL;
1962 again:
1963         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1964                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1965                 struct anon_vma_chain *vmac;
1966                 struct vm_area_struct *vma;
1967
1968                 anon_vma_lock_read(anon_vma);
1969                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1970                                                0, ULONG_MAX) {
1971                         vma = vmac->vma;
1972                         if (rmap_item->address < vma->vm_start ||
1973                             rmap_item->address >= vma->vm_end)
1974                                 continue;
1975                         /*
1976                          * Initially we examine only the vma which covers this
1977                          * rmap_item; but later, if there is still work to do,
1978                          * we examine covering vmas in other mms: in case they
1979                          * were forked from the original since ksmd passed.
1980                          */
1981                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1982                                 continue;
1983
1984                         ret = try_to_unmap_one(page, vma,
1985                                         rmap_item->address, flags);
1986                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1987                                 anon_vma_unlock_read(anon_vma);
1988                                 goto out;
1989                         }
1990                 }
1991                 anon_vma_unlock_read(anon_vma);
1992         }
1993         if (!search_new_forks++)
1994                 goto again;
1995 out:
1996         return ret;
1997 }
1998
1999 #ifdef CONFIG_MIGRATION
2000 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
2001                   struct vm_area_struct *, unsigned long, void *), void *arg)
2002 {
2003         struct stable_node *stable_node;
2004         struct rmap_item *rmap_item;
2005         int ret = SWAP_AGAIN;
2006         int search_new_forks = 0;
2007
2008         VM_BUG_ON(!PageKsm(page));
2009         VM_BUG_ON(!PageLocked(page));
2010
2011         stable_node = page_stable_node(page);
2012         if (!stable_node)
2013                 return ret;
2014 again:
2015         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2016                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2017                 struct anon_vma_chain *vmac;
2018                 struct vm_area_struct *vma;
2019
2020                 anon_vma_lock_read(anon_vma);
2021                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2022                                                0, ULONG_MAX) {
2023                         vma = vmac->vma;
2024                         if (rmap_item->address < vma->vm_start ||
2025                             rmap_item->address >= vma->vm_end)
2026                                 continue;
2027                         /*
2028                          * Initially we examine only the vma which covers this
2029                          * rmap_item; but later, if there is still work to do,
2030                          * we examine covering vmas in other mms: in case they
2031                          * were forked from the original since ksmd passed.
2032                          */
2033                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2034                                 continue;
2035
2036                         ret = rmap_one(page, vma, rmap_item->address, arg);
2037                         if (ret != SWAP_AGAIN) {
2038                                 anon_vma_unlock_read(anon_vma);
2039                                 goto out;
2040                         }
2041                 }
2042                 anon_vma_unlock_read(anon_vma);
2043         }
2044         if (!search_new_forks++)
2045                 goto again;
2046 out:
2047         return ret;
2048 }
2049
2050 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2051 {
2052         struct stable_node *stable_node;
2053
2054         VM_BUG_ON(!PageLocked(oldpage));
2055         VM_BUG_ON(!PageLocked(newpage));
2056         VM_BUG_ON(newpage->mapping != oldpage->mapping);
2057
2058         stable_node = page_stable_node(newpage);
2059         if (stable_node) {
2060                 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
2061                 stable_node->kpfn = page_to_pfn(newpage);
2062                 /*
2063                  * newpage->mapping was set in advance; now we need smp_wmb()
2064                  * to make sure that the new stable_node->kpfn is visible
2065                  * to get_ksm_page() before it can see that oldpage->mapping
2066                  * has gone stale (or that PageSwapCache has been cleared).
2067                  */
2068                 smp_wmb();
2069                 set_page_stable_node(oldpage, NULL);
2070         }
2071 }
2072 #endif /* CONFIG_MIGRATION */
2073
2074 #ifdef CONFIG_MEMORY_HOTREMOVE
2075 static int just_wait(void *word)
2076 {
2077         schedule();
2078         return 0;
2079 }
2080
2081 static void wait_while_offlining(void)
2082 {
2083         while (ksm_run & KSM_RUN_OFFLINE) {
2084                 mutex_unlock(&ksm_thread_mutex);
2085                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2086                                 just_wait, TASK_UNINTERRUPTIBLE);
2087                 mutex_lock(&ksm_thread_mutex);
2088         }
2089 }
2090
2091 static void ksm_check_stable_tree(unsigned long start_pfn,
2092                                   unsigned long end_pfn)
2093 {
2094         struct stable_node *stable_node;
2095         struct list_head *this, *next;
2096         struct rb_node *node;
2097         int nid;
2098
2099         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2100                 node = rb_first(root_stable_tree + nid);
2101                 while (node) {
2102                         stable_node = rb_entry(node, struct stable_node, node);
2103                         if (stable_node->kpfn >= start_pfn &&
2104                             stable_node->kpfn < end_pfn) {
2105                                 /*
2106                                  * Don't get_ksm_page, page has already gone:
2107                                  * which is why we keep kpfn instead of page*
2108                                  */
2109                                 remove_node_from_stable_tree(stable_node);
2110                                 node = rb_first(root_stable_tree + nid);
2111                         } else
2112                                 node = rb_next(node);
2113                         cond_resched();
2114                 }
2115         }
2116         list_for_each_safe(this, next, &migrate_nodes) {
2117                 stable_node = list_entry(this, struct stable_node, list);
2118                 if (stable_node->kpfn >= start_pfn &&
2119                     stable_node->kpfn < end_pfn)
2120                         remove_node_from_stable_tree(stable_node);
2121                 cond_resched();
2122         }
2123 }
2124
2125 static int ksm_memory_callback(struct notifier_block *self,
2126                                unsigned long action, void *arg)
2127 {
2128         struct memory_notify *mn = arg;
2129
2130         switch (action) {
2131         case MEM_GOING_OFFLINE:
2132                 /*
2133                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2134                  * and remove_all_stable_nodes() while memory is going offline:
2135                  * it is unsafe for them to touch the stable tree at this time.
2136                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2137                  * which do not need the ksm_thread_mutex are all safe.
2138                  */
2139                 mutex_lock(&ksm_thread_mutex);
2140                 ksm_run |= KSM_RUN_OFFLINE;
2141                 mutex_unlock(&ksm_thread_mutex);
2142                 break;
2143
2144         case MEM_OFFLINE:
2145                 /*
2146                  * Most of the work is done by page migration; but there might
2147                  * be a few stable_nodes left over, still pointing to struct
2148                  * pages which have been offlined: prune those from the tree,
2149                  * otherwise get_ksm_page() might later try to access a
2150                  * non-existent struct page.
2151                  */
2152                 ksm_check_stable_tree(mn->start_pfn,
2153                                       mn->start_pfn + mn->nr_pages);
2154                 /* fallthrough */
2155
2156         case MEM_CANCEL_OFFLINE:
2157                 mutex_lock(&ksm_thread_mutex);
2158                 ksm_run &= ~KSM_RUN_OFFLINE;
2159                 mutex_unlock(&ksm_thread_mutex);
2160
2161                 smp_mb();       /* wake_up_bit advises this */
2162                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2163                 break;
2164         }
2165         return NOTIFY_OK;
2166 }
2167 #else
2168 static void wait_while_offlining(void)
2169 {
2170 }
2171 #endif /* CONFIG_MEMORY_HOTREMOVE */
2172
2173 #ifdef CONFIG_SYSFS
2174 /*
2175  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2176  */
2177
2178 #define KSM_ATTR_RO(_name) \
2179         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2180 #define KSM_ATTR(_name) \
2181         static struct kobj_attribute _name##_attr = \
2182                 __ATTR(_name, 0644, _name##_show, _name##_store)
2183
2184 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2185                                     struct kobj_attribute *attr, char *buf)
2186 {
2187         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2188 }
2189
2190 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2191                                      struct kobj_attribute *attr,
2192                                      const char *buf, size_t count)
2193 {
2194         unsigned long msecs;
2195         int err;
2196
2197         err = strict_strtoul(buf, 10, &msecs);
2198         if (err || msecs > UINT_MAX)
2199                 return -EINVAL;
2200
2201         ksm_thread_sleep_millisecs = msecs;
2202
2203         return count;
2204 }
2205 KSM_ATTR(sleep_millisecs);
2206
2207 static ssize_t pages_to_scan_show(struct kobject *kobj,
2208                                   struct kobj_attribute *attr, char *buf)
2209 {
2210         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2211 }
2212
2213 static ssize_t pages_to_scan_store(struct kobject *kobj,
2214                                    struct kobj_attribute *attr,
2215                                    const char *buf, size_t count)
2216 {
2217         int err;
2218         unsigned long nr_pages;
2219
2220         err = strict_strtoul(buf, 10, &nr_pages);
2221         if (err || nr_pages > UINT_MAX)
2222                 return -EINVAL;
2223
2224         ksm_thread_pages_to_scan = nr_pages;
2225
2226         return count;
2227 }
2228 KSM_ATTR(pages_to_scan);
2229
2230 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2231                         char *buf)
2232 {
2233         return sprintf(buf, "%lu\n", ksm_run);
2234 }
2235
2236 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2237                          const char *buf, size_t count)
2238 {
2239         int err;
2240         unsigned long flags;
2241
2242         err = strict_strtoul(buf, 10, &flags);
2243         if (err || flags > UINT_MAX)
2244                 return -EINVAL;
2245         if (flags > KSM_RUN_UNMERGE)
2246                 return -EINVAL;
2247
2248         /*
2249          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2250          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2251          * breaking COW to free the pages_shared (but leaves mm_slots
2252          * on the list for when ksmd may be set running again).
2253          */
2254
2255         mutex_lock(&ksm_thread_mutex);
2256         wait_while_offlining();
2257         if (ksm_run != flags) {
2258                 ksm_run = flags;
2259                 if (flags & KSM_RUN_UNMERGE) {
2260                         set_current_oom_origin();
2261                         err = unmerge_and_remove_all_rmap_items();
2262                         clear_current_oom_origin();
2263                         if (err) {
2264                                 ksm_run = KSM_RUN_STOP;
2265                                 count = err;
2266                         }
2267                 }
2268         }
2269         mutex_unlock(&ksm_thread_mutex);
2270
2271         if (flags & KSM_RUN_MERGE)
2272                 wake_up_interruptible(&ksm_thread_wait);
2273
2274         return count;
2275 }
2276 KSM_ATTR(run);
2277
2278 #ifdef CONFIG_NUMA
2279 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2280                                 struct kobj_attribute *attr, char *buf)
2281 {
2282         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2283 }
2284
2285 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2286                                    struct kobj_attribute *attr,
2287                                    const char *buf, size_t count)
2288 {
2289         int err;
2290         unsigned long knob;
2291
2292         err = kstrtoul(buf, 10, &knob);
2293         if (err)
2294                 return err;
2295         if (knob > 1)
2296                 return -EINVAL;
2297
2298         mutex_lock(&ksm_thread_mutex);
2299         wait_while_offlining();
2300         if (ksm_merge_across_nodes != knob) {
2301                 if (ksm_pages_shared || remove_all_stable_nodes())
2302                         err = -EBUSY;
2303                 else if (root_stable_tree == one_stable_tree) {
2304                         struct rb_root *buf;
2305                         /*
2306                          * This is the first time that we switch away from the
2307                          * default of merging across nodes: must now allocate
2308                          * a buffer to hold as many roots as may be needed.
2309                          * Allocate stable and unstable together:
2310                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2311                          */
2312                         buf = kcalloc(nr_node_ids + nr_node_ids,
2313                                 sizeof(*buf), GFP_KERNEL | __GFP_ZERO);
2314                         /* Let us assume that RB_ROOT is NULL is zero */
2315                         if (!buf)
2316                                 err = -ENOMEM;
2317                         else {
2318                                 root_stable_tree = buf;
2319                                 root_unstable_tree = buf + nr_node_ids;
2320                                 /* Stable tree is empty but not the unstable */
2321                                 root_unstable_tree[0] = one_unstable_tree[0];
2322                         }
2323                 }
2324                 if (!err) {
2325                         ksm_merge_across_nodes = knob;
2326                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2327                 }
2328         }
2329         mutex_unlock(&ksm_thread_mutex);
2330
2331         return err ? err : count;
2332 }
2333 KSM_ATTR(merge_across_nodes);
2334 #endif
2335
2336 static ssize_t pages_shared_show(struct kobject *kobj,
2337                                  struct kobj_attribute *attr, char *buf)
2338 {
2339         return sprintf(buf, "%lu\n", ksm_pages_shared);
2340 }
2341 KSM_ATTR_RO(pages_shared);
2342
2343 static ssize_t pages_sharing_show(struct kobject *kobj,
2344                                   struct kobj_attribute *attr, char *buf)
2345 {
2346         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2347 }
2348 KSM_ATTR_RO(pages_sharing);
2349
2350 static ssize_t pages_unshared_show(struct kobject *kobj,
2351                                    struct kobj_attribute *attr, char *buf)
2352 {
2353         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2354 }
2355 KSM_ATTR_RO(pages_unshared);
2356
2357 static ssize_t pages_volatile_show(struct kobject *kobj,
2358                                    struct kobj_attribute *attr, char *buf)
2359 {
2360         long ksm_pages_volatile;
2361
2362         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2363                                 - ksm_pages_sharing - ksm_pages_unshared;
2364         /*
2365          * It was not worth any locking to calculate that statistic,
2366          * but it might therefore sometimes be negative: conceal that.
2367          */
2368         if (ksm_pages_volatile < 0)
2369                 ksm_pages_volatile = 0;
2370         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2371 }
2372 KSM_ATTR_RO(pages_volatile);
2373
2374 static ssize_t full_scans_show(struct kobject *kobj,
2375                                struct kobj_attribute *attr, char *buf)
2376 {
2377         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2378 }
2379 KSM_ATTR_RO(full_scans);
2380
2381 static struct attribute *ksm_attrs[] = {
2382         &sleep_millisecs_attr.attr,
2383         &pages_to_scan_attr.attr,
2384         &run_attr.attr,
2385         &pages_shared_attr.attr,
2386         &pages_sharing_attr.attr,
2387         &pages_unshared_attr.attr,
2388         &pages_volatile_attr.attr,
2389         &full_scans_attr.attr,
2390 #ifdef CONFIG_NUMA
2391         &merge_across_nodes_attr.attr,
2392 #endif
2393         NULL,
2394 };
2395
2396 static struct attribute_group ksm_attr_group = {
2397         .attrs = ksm_attrs,
2398         .name = "ksm",
2399 };
2400 #endif /* CONFIG_SYSFS */
2401
2402 static int __init ksm_init(void)
2403 {
2404         struct task_struct *ksm_thread;
2405         int err;
2406
2407         err = ksm_slab_init();
2408         if (err)
2409                 goto out;
2410
2411         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2412         if (IS_ERR(ksm_thread)) {
2413                 printk(KERN_ERR "ksm: creating kthread failed\n");
2414                 err = PTR_ERR(ksm_thread);
2415                 goto out_free;
2416         }
2417
2418 #ifdef CONFIG_SYSFS
2419         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2420         if (err) {
2421                 printk(KERN_ERR "ksm: register sysfs failed\n");
2422                 kthread_stop(ksm_thread);
2423                 goto out_free;
2424         }
2425 #else
2426         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
2427
2428 #endif /* CONFIG_SYSFS */
2429
2430 #ifdef CONFIG_MEMORY_HOTREMOVE
2431         /* There is no significance to this priority 100 */
2432         hotplug_memory_notifier(ksm_memory_callback, 100);
2433 #endif
2434         return 0;
2435
2436 out_free:
2437         ksm_slab_free();
2438 out:
2439         return err;
2440 }
2441 module_init(ksm_init)