Added license information
[kvmfornfv.git] / kernel / 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_SIGSEGV | 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_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 = READ_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 (READ_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 (READ_ONCE(page->mapping) != expected_mapping) {
581                 put_page(page);
582                 goto stale;
583         }
584
585         if (lock_it) {
586                 lock_page(page);
587                 if (READ_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 (READ_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_notify(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
949         mmun_start = addr;
950         mmun_end   = addr + PAGE_SIZE;
951         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
952
953         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
954         if (!pte_same(*ptep, orig_pte)) {
955                 pte_unmap_unlock(ptep, ptl);
956                 goto out_mn;
957         }
958
959         get_page(kpage);
960         page_add_anon_rmap(kpage, vma, addr);
961
962         flush_cache_page(vma, addr, pte_pfn(*ptep));
963         ptep_clear_flush_notify(vma, addr, ptep);
964         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
965
966         page_remove_rmap(page);
967         if (!page_mapped(page))
968                 try_to_free_swap(page);
969         put_page(page);
970
971         pte_unmap_unlock(ptep, ptl);
972         err = 0;
973 out_mn:
974         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
975 out:
976         return err;
977 }
978
979 static int page_trans_compound_anon_split(struct page *page)
980 {
981         int ret = 0;
982         struct page *transhuge_head = page_trans_compound_anon(page);
983         if (transhuge_head) {
984                 /* Get the reference on the head to split it. */
985                 if (get_page_unless_zero(transhuge_head)) {
986                         /*
987                          * Recheck we got the reference while the head
988                          * was still anonymous.
989                          */
990                         if (PageAnon(transhuge_head))
991                                 ret = split_huge_page(transhuge_head);
992                         else
993                                 /*
994                                  * Retry later if split_huge_page run
995                                  * from under us.
996                                  */
997                                 ret = 1;
998                         put_page(transhuge_head);
999                 } else
1000                         /* Retry later if split_huge_page run from under us. */
1001                         ret = 1;
1002         }
1003         return ret;
1004 }
1005
1006 /*
1007  * try_to_merge_one_page - take two pages and merge them into one
1008  * @vma: the vma that holds the pte pointing to page
1009  * @page: the PageAnon page that we want to replace with kpage
1010  * @kpage: the PageKsm page that we want to map instead of page,
1011  *         or NULL the first time when we want to use page as kpage.
1012  *
1013  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1014  */
1015 static int try_to_merge_one_page(struct vm_area_struct *vma,
1016                                  struct page *page, struct page *kpage)
1017 {
1018         pte_t orig_pte = __pte(0);
1019         int err = -EFAULT;
1020
1021         if (page == kpage)                      /* ksm page forked */
1022                 return 0;
1023
1024         if (!(vma->vm_flags & VM_MERGEABLE))
1025                 goto out;
1026         if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1027                 goto out;
1028         BUG_ON(PageTransCompound(page));
1029         if (!PageAnon(page))
1030                 goto out;
1031
1032         /*
1033          * We need the page lock to read a stable PageSwapCache in
1034          * write_protect_page().  We use trylock_page() instead of
1035          * lock_page() because we don't want to wait here - we
1036          * prefer to continue scanning and merging different pages,
1037          * then come back to this page when it is unlocked.
1038          */
1039         if (!trylock_page(page))
1040                 goto out;
1041         /*
1042          * If this anonymous page is mapped only here, its pte may need
1043          * to be write-protected.  If it's mapped elsewhere, all of its
1044          * ptes are necessarily already write-protected.  But in either
1045          * case, we need to lock and check page_count is not raised.
1046          */
1047         if (write_protect_page(vma, page, &orig_pte) == 0) {
1048                 if (!kpage) {
1049                         /*
1050                          * While we hold page lock, upgrade page from
1051                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1052                          * stable_tree_insert() will update stable_node.
1053                          */
1054                         set_page_stable_node(page, NULL);
1055                         mark_page_accessed(page);
1056                         err = 0;
1057                 } else if (pages_identical(page, kpage))
1058                         err = replace_page(vma, page, kpage, orig_pte);
1059         }
1060
1061         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1062                 munlock_vma_page(page);
1063                 if (!PageMlocked(kpage)) {
1064                         unlock_page(page);
1065                         lock_page(kpage);
1066                         mlock_vma_page(kpage);
1067                         page = kpage;           /* for final unlock */
1068                 }
1069         }
1070
1071         unlock_page(page);
1072 out:
1073         return err;
1074 }
1075
1076 /*
1077  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1078  * but no new kernel page is allocated: kpage must already be a ksm page.
1079  *
1080  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1081  */
1082 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1083                                       struct page *page, struct page *kpage)
1084 {
1085         struct mm_struct *mm = rmap_item->mm;
1086         struct vm_area_struct *vma;
1087         int err = -EFAULT;
1088
1089         down_read(&mm->mmap_sem);
1090         if (ksm_test_exit(mm))
1091                 goto out;
1092         vma = find_vma(mm, rmap_item->address);
1093         if (!vma || vma->vm_start > rmap_item->address)
1094                 goto out;
1095
1096         err = try_to_merge_one_page(vma, page, kpage);
1097         if (err)
1098                 goto out;
1099
1100         /* Unstable nid is in union with stable anon_vma: remove first */
1101         remove_rmap_item_from_tree(rmap_item);
1102
1103         /* Must get reference to anon_vma while still holding mmap_sem */
1104         rmap_item->anon_vma = vma->anon_vma;
1105         get_anon_vma(vma->anon_vma);
1106 out:
1107         up_read(&mm->mmap_sem);
1108         return err;
1109 }
1110
1111 /*
1112  * try_to_merge_two_pages - take two identical pages and prepare them
1113  * to be merged into one page.
1114  *
1115  * This function returns the kpage if we successfully merged two identical
1116  * pages into one ksm page, NULL otherwise.
1117  *
1118  * Note that this function upgrades page to ksm page: if one of the pages
1119  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1120  */
1121 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1122                                            struct page *page,
1123                                            struct rmap_item *tree_rmap_item,
1124                                            struct page *tree_page)
1125 {
1126         int err;
1127
1128         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1129         if (!err) {
1130                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1131                                                         tree_page, page);
1132                 /*
1133                  * If that fails, we have a ksm page with only one pte
1134                  * pointing to it: so break it.
1135                  */
1136                 if (err)
1137                         break_cow(rmap_item);
1138         }
1139         return err ? NULL : page;
1140 }
1141
1142 /*
1143  * stable_tree_search - search for page inside the stable tree
1144  *
1145  * This function checks if there is a page inside the stable tree
1146  * with identical content to the page that we are scanning right now.
1147  *
1148  * This function returns the stable tree node of identical content if found,
1149  * NULL otherwise.
1150  */
1151 static struct page *stable_tree_search(struct page *page)
1152 {
1153         int nid;
1154         struct rb_root *root;
1155         struct rb_node **new;
1156         struct rb_node *parent;
1157         struct stable_node *stable_node;
1158         struct stable_node *page_node;
1159
1160         page_node = page_stable_node(page);
1161         if (page_node && page_node->head != &migrate_nodes) {
1162                 /* ksm page forked */
1163                 get_page(page);
1164                 return page;
1165         }
1166
1167         nid = get_kpfn_nid(page_to_pfn(page));
1168         root = root_stable_tree + nid;
1169 again:
1170         new = &root->rb_node;
1171         parent = NULL;
1172
1173         while (*new) {
1174                 struct page *tree_page;
1175                 int ret;
1176
1177                 cond_resched();
1178                 stable_node = rb_entry(*new, struct stable_node, node);
1179                 tree_page = get_ksm_page(stable_node, false);
1180                 if (!tree_page)
1181                         return NULL;
1182
1183                 ret = memcmp_pages(page, tree_page);
1184                 put_page(tree_page);
1185
1186                 parent = *new;
1187                 if (ret < 0)
1188                         new = &parent->rb_left;
1189                 else if (ret > 0)
1190                         new = &parent->rb_right;
1191                 else {
1192                         /*
1193                          * Lock and unlock the stable_node's page (which
1194                          * might already have been migrated) so that page
1195                          * migration is sure to notice its raised count.
1196                          * It would be more elegant to return stable_node
1197                          * than kpage, but that involves more changes.
1198                          */
1199                         tree_page = get_ksm_page(stable_node, true);
1200                         if (tree_page) {
1201                                 unlock_page(tree_page);
1202                                 if (get_kpfn_nid(stable_node->kpfn) !=
1203                                                 NUMA(stable_node->nid)) {
1204                                         put_page(tree_page);
1205                                         goto replace;
1206                                 }
1207                                 return tree_page;
1208                         }
1209                         /*
1210                          * There is now a place for page_node, but the tree may
1211                          * have been rebalanced, so re-evaluate parent and new.
1212                          */
1213                         if (page_node)
1214                                 goto again;
1215                         return NULL;
1216                 }
1217         }
1218
1219         if (!page_node)
1220                 return NULL;
1221
1222         list_del(&page_node->list);
1223         DO_NUMA(page_node->nid = nid);
1224         rb_link_node(&page_node->node, parent, new);
1225         rb_insert_color(&page_node->node, root);
1226         get_page(page);
1227         return page;
1228
1229 replace:
1230         if (page_node) {
1231                 list_del(&page_node->list);
1232                 DO_NUMA(page_node->nid = nid);
1233                 rb_replace_node(&stable_node->node, &page_node->node, root);
1234                 get_page(page);
1235         } else {
1236                 rb_erase(&stable_node->node, root);
1237                 page = NULL;
1238         }
1239         stable_node->head = &migrate_nodes;
1240         list_add(&stable_node->list, stable_node->head);
1241         return page;
1242 }
1243
1244 /*
1245  * stable_tree_insert - insert stable tree node pointing to new ksm page
1246  * into the stable tree.
1247  *
1248  * This function returns the stable tree node just allocated on success,
1249  * NULL otherwise.
1250  */
1251 static struct stable_node *stable_tree_insert(struct page *kpage)
1252 {
1253         int nid;
1254         unsigned long kpfn;
1255         struct rb_root *root;
1256         struct rb_node **new;
1257         struct rb_node *parent = NULL;
1258         struct stable_node *stable_node;
1259
1260         kpfn = page_to_pfn(kpage);
1261         nid = get_kpfn_nid(kpfn);
1262         root = root_stable_tree + nid;
1263         new = &root->rb_node;
1264
1265         while (*new) {
1266                 struct page *tree_page;
1267                 int ret;
1268
1269                 cond_resched();
1270                 stable_node = rb_entry(*new, struct stable_node, node);
1271                 tree_page = get_ksm_page(stable_node, false);
1272                 if (!tree_page)
1273                         return NULL;
1274
1275                 ret = memcmp_pages(kpage, tree_page);
1276                 put_page(tree_page);
1277
1278                 parent = *new;
1279                 if (ret < 0)
1280                         new = &parent->rb_left;
1281                 else if (ret > 0)
1282                         new = &parent->rb_right;
1283                 else {
1284                         /*
1285                          * It is not a bug that stable_tree_search() didn't
1286                          * find this node: because at that time our page was
1287                          * not yet write-protected, so may have changed since.
1288                          */
1289                         return NULL;
1290                 }
1291         }
1292
1293         stable_node = alloc_stable_node();
1294         if (!stable_node)
1295                 return NULL;
1296
1297         INIT_HLIST_HEAD(&stable_node->hlist);
1298         stable_node->kpfn = kpfn;
1299         set_page_stable_node(kpage, stable_node);
1300         DO_NUMA(stable_node->nid = nid);
1301         rb_link_node(&stable_node->node, parent, new);
1302         rb_insert_color(&stable_node->node, root);
1303
1304         return stable_node;
1305 }
1306
1307 /*
1308  * unstable_tree_search_insert - search for identical page,
1309  * else insert rmap_item into the unstable tree.
1310  *
1311  * This function searches for a page in the unstable tree identical to the
1312  * page currently being scanned; and if no identical page is found in the
1313  * tree, we insert rmap_item as a new object into the unstable tree.
1314  *
1315  * This function returns pointer to rmap_item found to be identical
1316  * to the currently scanned page, NULL otherwise.
1317  *
1318  * This function does both searching and inserting, because they share
1319  * the same walking algorithm in an rbtree.
1320  */
1321 static
1322 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1323                                               struct page *page,
1324                                               struct page **tree_pagep)
1325 {
1326         struct rb_node **new;
1327         struct rb_root *root;
1328         struct rb_node *parent = NULL;
1329         int nid;
1330
1331         nid = get_kpfn_nid(page_to_pfn(page));
1332         root = root_unstable_tree + nid;
1333         new = &root->rb_node;
1334
1335         while (*new) {
1336                 struct rmap_item *tree_rmap_item;
1337                 struct page *tree_page;
1338                 int ret;
1339
1340                 cond_resched();
1341                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1342                 tree_page = get_mergeable_page(tree_rmap_item);
1343                 if (IS_ERR_OR_NULL(tree_page))
1344                         return NULL;
1345
1346                 /*
1347                  * Don't substitute a ksm page for a forked page.
1348                  */
1349                 if (page == tree_page) {
1350                         put_page(tree_page);
1351                         return NULL;
1352                 }
1353
1354                 ret = memcmp_pages(page, tree_page);
1355
1356                 parent = *new;
1357                 if (ret < 0) {
1358                         put_page(tree_page);
1359                         new = &parent->rb_left;
1360                 } else if (ret > 0) {
1361                         put_page(tree_page);
1362                         new = &parent->rb_right;
1363                 } else if (!ksm_merge_across_nodes &&
1364                            page_to_nid(tree_page) != nid) {
1365                         /*
1366                          * If tree_page has been migrated to another NUMA node,
1367                          * it will be flushed out and put in the right unstable
1368                          * tree next time: only merge with it when across_nodes.
1369                          */
1370                         put_page(tree_page);
1371                         return NULL;
1372                 } else {
1373                         *tree_pagep = tree_page;
1374                         return tree_rmap_item;
1375                 }
1376         }
1377
1378         rmap_item->address |= UNSTABLE_FLAG;
1379         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1380         DO_NUMA(rmap_item->nid = nid);
1381         rb_link_node(&rmap_item->node, parent, new);
1382         rb_insert_color(&rmap_item->node, root);
1383
1384         ksm_pages_unshared++;
1385         return NULL;
1386 }
1387
1388 /*
1389  * stable_tree_append - add another rmap_item to the linked list of
1390  * rmap_items hanging off a given node of the stable tree, all sharing
1391  * the same ksm page.
1392  */
1393 static void stable_tree_append(struct rmap_item *rmap_item,
1394                                struct stable_node *stable_node)
1395 {
1396         rmap_item->head = stable_node;
1397         rmap_item->address |= STABLE_FLAG;
1398         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1399
1400         if (rmap_item->hlist.next)
1401                 ksm_pages_sharing++;
1402         else
1403                 ksm_pages_shared++;
1404 }
1405
1406 /*
1407  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1408  * if not, compare checksum to previous and if it's the same, see if page can
1409  * be inserted into the unstable tree, or merged with a page already there and
1410  * both transferred to the stable tree.
1411  *
1412  * @page: the page that we are searching identical page to.
1413  * @rmap_item: the reverse mapping into the virtual address of this page
1414  */
1415 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1416 {
1417         struct rmap_item *tree_rmap_item;
1418         struct page *tree_page = NULL;
1419         struct stable_node *stable_node;
1420         struct page *kpage;
1421         unsigned int checksum;
1422         int err;
1423
1424         stable_node = page_stable_node(page);
1425         if (stable_node) {
1426                 if (stable_node->head != &migrate_nodes &&
1427                     get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1428                         rb_erase(&stable_node->node,
1429                                  root_stable_tree + NUMA(stable_node->nid));
1430                         stable_node->head = &migrate_nodes;
1431                         list_add(&stable_node->list, stable_node->head);
1432                 }
1433                 if (stable_node->head != &migrate_nodes &&
1434                     rmap_item->head == stable_node)
1435                         return;
1436         }
1437
1438         /* We first start with searching the page inside the stable tree */
1439         kpage = stable_tree_search(page);
1440         if (kpage == page && rmap_item->head == stable_node) {
1441                 put_page(kpage);
1442                 return;
1443         }
1444
1445         remove_rmap_item_from_tree(rmap_item);
1446
1447         if (kpage) {
1448                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1449                 if (!err) {
1450                         /*
1451                          * The page was successfully merged:
1452                          * add its rmap_item to the stable tree.
1453                          */
1454                         lock_page(kpage);
1455                         stable_tree_append(rmap_item, page_stable_node(kpage));
1456                         unlock_page(kpage);
1457                 }
1458                 put_page(kpage);
1459                 return;
1460         }
1461
1462         /*
1463          * If the hash value of the page has changed from the last time
1464          * we calculated it, this page is changing frequently: therefore we
1465          * don't want to insert it in the unstable tree, and we don't want
1466          * to waste our time searching for something identical to it there.
1467          */
1468         checksum = calc_checksum(page);
1469         if (rmap_item->oldchecksum != checksum) {
1470                 rmap_item->oldchecksum = checksum;
1471                 return;
1472         }
1473
1474         tree_rmap_item =
1475                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1476         if (tree_rmap_item) {
1477                 kpage = try_to_merge_two_pages(rmap_item, page,
1478                                                 tree_rmap_item, tree_page);
1479                 put_page(tree_page);
1480                 if (kpage) {
1481                         /*
1482                          * The pages were successfully merged: insert new
1483                          * node in the stable tree and add both rmap_items.
1484                          */
1485                         lock_page(kpage);
1486                         stable_node = stable_tree_insert(kpage);
1487                         if (stable_node) {
1488                                 stable_tree_append(tree_rmap_item, stable_node);
1489                                 stable_tree_append(rmap_item, stable_node);
1490                         }
1491                         unlock_page(kpage);
1492
1493                         /*
1494                          * If we fail to insert the page into the stable tree,
1495                          * we will have 2 virtual addresses that are pointing
1496                          * to a ksm page left outside the stable tree,
1497                          * in which case we need to break_cow on both.
1498                          */
1499                         if (!stable_node) {
1500                                 break_cow(tree_rmap_item);
1501                                 break_cow(rmap_item);
1502                         }
1503                 }
1504         }
1505 }
1506
1507 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1508                                             struct rmap_item **rmap_list,
1509                                             unsigned long addr)
1510 {
1511         struct rmap_item *rmap_item;
1512
1513         while (*rmap_list) {
1514                 rmap_item = *rmap_list;
1515                 if ((rmap_item->address & PAGE_MASK) == addr)
1516                         return rmap_item;
1517                 if (rmap_item->address > addr)
1518                         break;
1519                 *rmap_list = rmap_item->rmap_list;
1520                 remove_rmap_item_from_tree(rmap_item);
1521                 free_rmap_item(rmap_item);
1522         }
1523
1524         rmap_item = alloc_rmap_item();
1525         if (rmap_item) {
1526                 /* It has already been zeroed */
1527                 rmap_item->mm = mm_slot->mm;
1528                 rmap_item->address = addr;
1529                 rmap_item->rmap_list = *rmap_list;
1530                 *rmap_list = rmap_item;
1531         }
1532         return rmap_item;
1533 }
1534
1535 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1536 {
1537         struct mm_struct *mm;
1538         struct mm_slot *slot;
1539         struct vm_area_struct *vma;
1540         struct rmap_item *rmap_item;
1541         int nid;
1542
1543         if (list_empty(&ksm_mm_head.mm_list))
1544                 return NULL;
1545
1546         slot = ksm_scan.mm_slot;
1547         if (slot == &ksm_mm_head) {
1548                 /*
1549                  * A number of pages can hang around indefinitely on per-cpu
1550                  * pagevecs, raised page count preventing write_protect_page
1551                  * from merging them.  Though it doesn't really matter much,
1552                  * it is puzzling to see some stuck in pages_volatile until
1553                  * other activity jostles them out, and they also prevented
1554                  * LTP's KSM test from succeeding deterministically; so drain
1555                  * them here (here rather than on entry to ksm_do_scan(),
1556                  * so we don't IPI too often when pages_to_scan is set low).
1557                  */
1558                 lru_add_drain_all();
1559
1560                 /*
1561                  * Whereas stale stable_nodes on the stable_tree itself
1562                  * get pruned in the regular course of stable_tree_search(),
1563                  * those moved out to the migrate_nodes list can accumulate:
1564                  * so prune them once before each full scan.
1565                  */
1566                 if (!ksm_merge_across_nodes) {
1567                         struct stable_node *stable_node;
1568                         struct list_head *this, *next;
1569                         struct page *page;
1570
1571                         list_for_each_safe(this, next, &migrate_nodes) {
1572                                 stable_node = list_entry(this,
1573                                                 struct stable_node, list);
1574                                 page = get_ksm_page(stable_node, false);
1575                                 if (page)
1576                                         put_page(page);
1577                                 cond_resched();
1578                         }
1579                 }
1580
1581                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1582                         root_unstable_tree[nid] = RB_ROOT;
1583
1584                 spin_lock(&ksm_mmlist_lock);
1585                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1586                 ksm_scan.mm_slot = slot;
1587                 spin_unlock(&ksm_mmlist_lock);
1588                 /*
1589                  * Although we tested list_empty() above, a racing __ksm_exit
1590                  * of the last mm on the list may have removed it since then.
1591                  */
1592                 if (slot == &ksm_mm_head)
1593                         return NULL;
1594 next_mm:
1595                 ksm_scan.address = 0;
1596                 ksm_scan.rmap_list = &slot->rmap_list;
1597         }
1598
1599         mm = slot->mm;
1600         down_read(&mm->mmap_sem);
1601         if (ksm_test_exit(mm))
1602                 vma = NULL;
1603         else
1604                 vma = find_vma(mm, ksm_scan.address);
1605
1606         for (; vma; vma = vma->vm_next) {
1607                 if (!(vma->vm_flags & VM_MERGEABLE))
1608                         continue;
1609                 if (ksm_scan.address < vma->vm_start)
1610                         ksm_scan.address = vma->vm_start;
1611                 if (!vma->anon_vma)
1612                         ksm_scan.address = vma->vm_end;
1613
1614                 while (ksm_scan.address < vma->vm_end) {
1615                         if (ksm_test_exit(mm))
1616                                 break;
1617                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1618                         if (IS_ERR_OR_NULL(*page)) {
1619                                 ksm_scan.address += PAGE_SIZE;
1620                                 cond_resched();
1621                                 continue;
1622                         }
1623                         if (PageAnon(*page) ||
1624                             page_trans_compound_anon(*page)) {
1625                                 flush_anon_page(vma, *page, ksm_scan.address);
1626                                 flush_dcache_page(*page);
1627                                 rmap_item = get_next_rmap_item(slot,
1628                                         ksm_scan.rmap_list, ksm_scan.address);
1629                                 if (rmap_item) {
1630                                         ksm_scan.rmap_list =
1631                                                         &rmap_item->rmap_list;
1632                                         ksm_scan.address += PAGE_SIZE;
1633                                 } else
1634                                         put_page(*page);
1635                                 up_read(&mm->mmap_sem);
1636                                 return rmap_item;
1637                         }
1638                         put_page(*page);
1639                         ksm_scan.address += PAGE_SIZE;
1640                         cond_resched();
1641                 }
1642         }
1643
1644         if (ksm_test_exit(mm)) {
1645                 ksm_scan.address = 0;
1646                 ksm_scan.rmap_list = &slot->rmap_list;
1647         }
1648         /*
1649          * Nuke all the rmap_items that are above this current rmap:
1650          * because there were no VM_MERGEABLE vmas with such addresses.
1651          */
1652         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1653
1654         spin_lock(&ksm_mmlist_lock);
1655         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1656                                                 struct mm_slot, mm_list);
1657         if (ksm_scan.address == 0) {
1658                 /*
1659                  * We've completed a full scan of all vmas, holding mmap_sem
1660                  * throughout, and found no VM_MERGEABLE: so do the same as
1661                  * __ksm_exit does to remove this mm from all our lists now.
1662                  * This applies either when cleaning up after __ksm_exit
1663                  * (but beware: we can reach here even before __ksm_exit),
1664                  * or when all VM_MERGEABLE areas have been unmapped (and
1665                  * mmap_sem then protects against race with MADV_MERGEABLE).
1666                  */
1667                 hash_del(&slot->link);
1668                 list_del(&slot->mm_list);
1669                 spin_unlock(&ksm_mmlist_lock);
1670
1671                 free_mm_slot(slot);
1672                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1673                 up_read(&mm->mmap_sem);
1674                 mmdrop(mm);
1675         } else {
1676                 spin_unlock(&ksm_mmlist_lock);
1677                 up_read(&mm->mmap_sem);
1678         }
1679
1680         /* Repeat until we've completed scanning the whole list */
1681         slot = ksm_scan.mm_slot;
1682         if (slot != &ksm_mm_head)
1683                 goto next_mm;
1684
1685         ksm_scan.seqnr++;
1686         return NULL;
1687 }
1688
1689 /**
1690  * ksm_do_scan  - the ksm scanner main worker function.
1691  * @scan_npages - number of pages we want to scan before we return.
1692  */
1693 static void ksm_do_scan(unsigned int scan_npages)
1694 {
1695         struct rmap_item *rmap_item;
1696         struct page *uninitialized_var(page);
1697
1698         while (scan_npages-- && likely(!freezing(current))) {
1699                 cond_resched();
1700                 rmap_item = scan_get_next_rmap_item(&page);
1701                 if (!rmap_item)
1702                         return;
1703                 cmp_and_merge_page(page, rmap_item);
1704                 put_page(page);
1705         }
1706 }
1707
1708 static int ksmd_should_run(void)
1709 {
1710         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1711 }
1712
1713 static int ksm_scan_thread(void *nothing)
1714 {
1715         set_freezable();
1716         set_user_nice(current, 5);
1717
1718         while (!kthread_should_stop()) {
1719                 mutex_lock(&ksm_thread_mutex);
1720                 wait_while_offlining();
1721                 if (ksmd_should_run())
1722                         ksm_do_scan(ksm_thread_pages_to_scan);
1723                 mutex_unlock(&ksm_thread_mutex);
1724
1725                 try_to_freeze();
1726
1727                 if (ksmd_should_run()) {
1728                         schedule_timeout_interruptible(
1729                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1730                 } else {
1731                         wait_event_freezable(ksm_thread_wait,
1732                                 ksmd_should_run() || kthread_should_stop());
1733                 }
1734         }
1735         return 0;
1736 }
1737
1738 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1739                 unsigned long end, int advice, unsigned long *vm_flags)
1740 {
1741         struct mm_struct *mm = vma->vm_mm;
1742         int err;
1743
1744         switch (advice) {
1745         case MADV_MERGEABLE:
1746                 /*
1747                  * Be somewhat over-protective for now!
1748                  */
1749                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1750                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1751                                  VM_HUGETLB | VM_MIXEDMAP))
1752                         return 0;               /* just ignore the advice */
1753
1754 #ifdef VM_SAO
1755                 if (*vm_flags & VM_SAO)
1756                         return 0;
1757 #endif
1758
1759                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1760                         err = __ksm_enter(mm);
1761                         if (err)
1762                                 return err;
1763                 }
1764
1765                 *vm_flags |= VM_MERGEABLE;
1766                 break;
1767
1768         case MADV_UNMERGEABLE:
1769                 if (!(*vm_flags & VM_MERGEABLE))
1770                         return 0;               /* just ignore the advice */
1771
1772                 if (vma->anon_vma) {
1773                         err = unmerge_ksm_pages(vma, start, end);
1774                         if (err)
1775                                 return err;
1776                 }
1777
1778                 *vm_flags &= ~VM_MERGEABLE;
1779                 break;
1780         }
1781
1782         return 0;
1783 }
1784
1785 int __ksm_enter(struct mm_struct *mm)
1786 {
1787         struct mm_slot *mm_slot;
1788         int needs_wakeup;
1789
1790         mm_slot = alloc_mm_slot();
1791         if (!mm_slot)
1792                 return -ENOMEM;
1793
1794         /* Check ksm_run too?  Would need tighter locking */
1795         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1796
1797         spin_lock(&ksm_mmlist_lock);
1798         insert_to_mm_slots_hash(mm, mm_slot);
1799         /*
1800          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1801          * insert just behind the scanning cursor, to let the area settle
1802          * down a little; when fork is followed by immediate exec, we don't
1803          * want ksmd to waste time setting up and tearing down an rmap_list.
1804          *
1805          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1806          * scanning cursor, otherwise KSM pages in newly forked mms will be
1807          * missed: then we might as well insert at the end of the list.
1808          */
1809         if (ksm_run & KSM_RUN_UNMERGE)
1810                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1811         else
1812                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1813         spin_unlock(&ksm_mmlist_lock);
1814
1815         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1816         atomic_inc(&mm->mm_count);
1817
1818         if (needs_wakeup)
1819                 wake_up_interruptible(&ksm_thread_wait);
1820
1821         return 0;
1822 }
1823
1824 void __ksm_exit(struct mm_struct *mm)
1825 {
1826         struct mm_slot *mm_slot;
1827         int easy_to_free = 0;
1828
1829         /*
1830          * This process is exiting: if it's straightforward (as is the
1831          * case when ksmd was never running), free mm_slot immediately.
1832          * But if it's at the cursor or has rmap_items linked to it, use
1833          * mmap_sem to synchronize with any break_cows before pagetables
1834          * are freed, and leave the mm_slot on the list for ksmd to free.
1835          * Beware: ksm may already have noticed it exiting and freed the slot.
1836          */
1837
1838         spin_lock(&ksm_mmlist_lock);
1839         mm_slot = get_mm_slot(mm);
1840         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1841                 if (!mm_slot->rmap_list) {
1842                         hash_del(&mm_slot->link);
1843                         list_del(&mm_slot->mm_list);
1844                         easy_to_free = 1;
1845                 } else {
1846                         list_move(&mm_slot->mm_list,
1847                                   &ksm_scan.mm_slot->mm_list);
1848                 }
1849         }
1850         spin_unlock(&ksm_mmlist_lock);
1851
1852         if (easy_to_free) {
1853                 free_mm_slot(mm_slot);
1854                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1855                 mmdrop(mm);
1856         } else if (mm_slot) {
1857                 down_write(&mm->mmap_sem);
1858                 up_write(&mm->mmap_sem);
1859         }
1860 }
1861
1862 struct page *ksm_might_need_to_copy(struct page *page,
1863                         struct vm_area_struct *vma, unsigned long address)
1864 {
1865         struct anon_vma *anon_vma = page_anon_vma(page);
1866         struct page *new_page;
1867
1868         if (PageKsm(page)) {
1869                 if (page_stable_node(page) &&
1870                     !(ksm_run & KSM_RUN_UNMERGE))
1871                         return page;    /* no need to copy it */
1872         } else if (!anon_vma) {
1873                 return page;            /* no need to copy it */
1874         } else if (anon_vma->root == vma->anon_vma->root &&
1875                  page->index == linear_page_index(vma, address)) {
1876                 return page;            /* still no need to copy it */
1877         }
1878         if (!PageUptodate(page))
1879                 return page;            /* let do_swap_page report the error */
1880
1881         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1882         if (new_page) {
1883                 copy_user_highpage(new_page, page, address, vma);
1884
1885                 SetPageDirty(new_page);
1886                 __SetPageUptodate(new_page);
1887                 __set_page_locked(new_page);
1888         }
1889
1890         return new_page;
1891 }
1892
1893 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1894 {
1895         struct stable_node *stable_node;
1896         struct rmap_item *rmap_item;
1897         int ret = SWAP_AGAIN;
1898         int search_new_forks = 0;
1899
1900         VM_BUG_ON_PAGE(!PageKsm(page), page);
1901
1902         /*
1903          * Rely on the page lock to protect against concurrent modifications
1904          * to that page's node of the stable tree.
1905          */
1906         VM_BUG_ON_PAGE(!PageLocked(page), page);
1907
1908         stable_node = page_stable_node(page);
1909         if (!stable_node)
1910                 return ret;
1911 again:
1912         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1913                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1914                 struct anon_vma_chain *vmac;
1915                 struct vm_area_struct *vma;
1916
1917                 anon_vma_lock_read(anon_vma);
1918                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1919                                                0, ULONG_MAX) {
1920                         vma = vmac->vma;
1921                         if (rmap_item->address < vma->vm_start ||
1922                             rmap_item->address >= vma->vm_end)
1923                                 continue;
1924                         /*
1925                          * Initially we examine only the vma which covers this
1926                          * rmap_item; but later, if there is still work to do,
1927                          * we examine covering vmas in other mms: in case they
1928                          * were forked from the original since ksmd passed.
1929                          */
1930                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1931                                 continue;
1932
1933                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1934                                 continue;
1935
1936                         ret = rwc->rmap_one(page, vma,
1937                                         rmap_item->address, rwc->arg);
1938                         if (ret != SWAP_AGAIN) {
1939                                 anon_vma_unlock_read(anon_vma);
1940                                 goto out;
1941                         }
1942                         if (rwc->done && rwc->done(page)) {
1943                                 anon_vma_unlock_read(anon_vma);
1944                                 goto out;
1945                         }
1946                 }
1947                 anon_vma_unlock_read(anon_vma);
1948         }
1949         if (!search_new_forks++)
1950                 goto again;
1951 out:
1952         return ret;
1953 }
1954
1955 #ifdef CONFIG_MIGRATION
1956 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1957 {
1958         struct stable_node *stable_node;
1959
1960         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1961         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1962         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
1963
1964         stable_node = page_stable_node(newpage);
1965         if (stable_node) {
1966                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
1967                 stable_node->kpfn = page_to_pfn(newpage);
1968                 /*
1969                  * newpage->mapping was set in advance; now we need smp_wmb()
1970                  * to make sure that the new stable_node->kpfn is visible
1971                  * to get_ksm_page() before it can see that oldpage->mapping
1972                  * has gone stale (or that PageSwapCache has been cleared).
1973                  */
1974                 smp_wmb();
1975                 set_page_stable_node(oldpage, NULL);
1976         }
1977 }
1978 #endif /* CONFIG_MIGRATION */
1979
1980 #ifdef CONFIG_MEMORY_HOTREMOVE
1981 static void wait_while_offlining(void)
1982 {
1983         while (ksm_run & KSM_RUN_OFFLINE) {
1984                 mutex_unlock(&ksm_thread_mutex);
1985                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
1986                             TASK_UNINTERRUPTIBLE);
1987                 mutex_lock(&ksm_thread_mutex);
1988         }
1989 }
1990
1991 static void ksm_check_stable_tree(unsigned long start_pfn,
1992                                   unsigned long end_pfn)
1993 {
1994         struct stable_node *stable_node;
1995         struct list_head *this, *next;
1996         struct rb_node *node;
1997         int nid;
1998
1999         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2000                 node = rb_first(root_stable_tree + nid);
2001                 while (node) {
2002                         stable_node = rb_entry(node, struct stable_node, node);
2003                         if (stable_node->kpfn >= start_pfn &&
2004                             stable_node->kpfn < end_pfn) {
2005                                 /*
2006                                  * Don't get_ksm_page, page has already gone:
2007                                  * which is why we keep kpfn instead of page*
2008                                  */
2009                                 remove_node_from_stable_tree(stable_node);
2010                                 node = rb_first(root_stable_tree + nid);
2011                         } else
2012                                 node = rb_next(node);
2013                         cond_resched();
2014                 }
2015         }
2016         list_for_each_safe(this, next, &migrate_nodes) {
2017                 stable_node = list_entry(this, struct stable_node, list);
2018                 if (stable_node->kpfn >= start_pfn &&
2019                     stable_node->kpfn < end_pfn)
2020                         remove_node_from_stable_tree(stable_node);
2021                 cond_resched();
2022         }
2023 }
2024
2025 static int ksm_memory_callback(struct notifier_block *self,
2026                                unsigned long action, void *arg)
2027 {
2028         struct memory_notify *mn = arg;
2029
2030         switch (action) {
2031         case MEM_GOING_OFFLINE:
2032                 /*
2033                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2034                  * and remove_all_stable_nodes() while memory is going offline:
2035                  * it is unsafe for them to touch the stable tree at this time.
2036                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2037                  * which do not need the ksm_thread_mutex are all safe.
2038                  */
2039                 mutex_lock(&ksm_thread_mutex);
2040                 ksm_run |= KSM_RUN_OFFLINE;
2041                 mutex_unlock(&ksm_thread_mutex);
2042                 break;
2043
2044         case MEM_OFFLINE:
2045                 /*
2046                  * Most of the work is done by page migration; but there might
2047                  * be a few stable_nodes left over, still pointing to struct
2048                  * pages which have been offlined: prune those from the tree,
2049                  * otherwise get_ksm_page() might later try to access a
2050                  * non-existent struct page.
2051                  */
2052                 ksm_check_stable_tree(mn->start_pfn,
2053                                       mn->start_pfn + mn->nr_pages);
2054                 /* fallthrough */
2055
2056         case MEM_CANCEL_OFFLINE:
2057                 mutex_lock(&ksm_thread_mutex);
2058                 ksm_run &= ~KSM_RUN_OFFLINE;
2059                 mutex_unlock(&ksm_thread_mutex);
2060
2061                 smp_mb();       /* wake_up_bit advises this */
2062                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2063                 break;
2064         }
2065         return NOTIFY_OK;
2066 }
2067 #else
2068 static void wait_while_offlining(void)
2069 {
2070 }
2071 #endif /* CONFIG_MEMORY_HOTREMOVE */
2072
2073 #ifdef CONFIG_SYSFS
2074 /*
2075  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2076  */
2077
2078 #define KSM_ATTR_RO(_name) \
2079         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2080 #define KSM_ATTR(_name) \
2081         static struct kobj_attribute _name##_attr = \
2082                 __ATTR(_name, 0644, _name##_show, _name##_store)
2083
2084 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2085                                     struct kobj_attribute *attr, char *buf)
2086 {
2087         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2088 }
2089
2090 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2091                                      struct kobj_attribute *attr,
2092                                      const char *buf, size_t count)
2093 {
2094         unsigned long msecs;
2095         int err;
2096
2097         err = kstrtoul(buf, 10, &msecs);
2098         if (err || msecs > UINT_MAX)
2099                 return -EINVAL;
2100
2101         ksm_thread_sleep_millisecs = msecs;
2102
2103         return count;
2104 }
2105 KSM_ATTR(sleep_millisecs);
2106
2107 static ssize_t pages_to_scan_show(struct kobject *kobj,
2108                                   struct kobj_attribute *attr, char *buf)
2109 {
2110         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2111 }
2112
2113 static ssize_t pages_to_scan_store(struct kobject *kobj,
2114                                    struct kobj_attribute *attr,
2115                                    const char *buf, size_t count)
2116 {
2117         int err;
2118         unsigned long nr_pages;
2119
2120         err = kstrtoul(buf, 10, &nr_pages);
2121         if (err || nr_pages > UINT_MAX)
2122                 return -EINVAL;
2123
2124         ksm_thread_pages_to_scan = nr_pages;
2125
2126         return count;
2127 }
2128 KSM_ATTR(pages_to_scan);
2129
2130 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2131                         char *buf)
2132 {
2133         return sprintf(buf, "%lu\n", ksm_run);
2134 }
2135
2136 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2137                          const char *buf, size_t count)
2138 {
2139         int err;
2140         unsigned long flags;
2141
2142         err = kstrtoul(buf, 10, &flags);
2143         if (err || flags > UINT_MAX)
2144                 return -EINVAL;
2145         if (flags > KSM_RUN_UNMERGE)
2146                 return -EINVAL;
2147
2148         /*
2149          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2150          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2151          * breaking COW to free the pages_shared (but leaves mm_slots
2152          * on the list for when ksmd may be set running again).
2153          */
2154
2155         mutex_lock(&ksm_thread_mutex);
2156         wait_while_offlining();
2157         if (ksm_run != flags) {
2158                 ksm_run = flags;
2159                 if (flags & KSM_RUN_UNMERGE) {
2160                         set_current_oom_origin();
2161                         err = unmerge_and_remove_all_rmap_items();
2162                         clear_current_oom_origin();
2163                         if (err) {
2164                                 ksm_run = KSM_RUN_STOP;
2165                                 count = err;
2166                         }
2167                 }
2168         }
2169         mutex_unlock(&ksm_thread_mutex);
2170
2171         if (flags & KSM_RUN_MERGE)
2172                 wake_up_interruptible(&ksm_thread_wait);
2173
2174         return count;
2175 }
2176 KSM_ATTR(run);
2177
2178 #ifdef CONFIG_NUMA
2179 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2180                                 struct kobj_attribute *attr, char *buf)
2181 {
2182         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2183 }
2184
2185 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2186                                    struct kobj_attribute *attr,
2187                                    const char *buf, size_t count)
2188 {
2189         int err;
2190         unsigned long knob;
2191
2192         err = kstrtoul(buf, 10, &knob);
2193         if (err)
2194                 return err;
2195         if (knob > 1)
2196                 return -EINVAL;
2197
2198         mutex_lock(&ksm_thread_mutex);
2199         wait_while_offlining();
2200         if (ksm_merge_across_nodes != knob) {
2201                 if (ksm_pages_shared || remove_all_stable_nodes())
2202                         err = -EBUSY;
2203                 else if (root_stable_tree == one_stable_tree) {
2204                         struct rb_root *buf;
2205                         /*
2206                          * This is the first time that we switch away from the
2207                          * default of merging across nodes: must now allocate
2208                          * a buffer to hold as many roots as may be needed.
2209                          * Allocate stable and unstable together:
2210                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2211                          */
2212                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2213                                       GFP_KERNEL);
2214                         /* Let us assume that RB_ROOT is NULL is zero */
2215                         if (!buf)
2216                                 err = -ENOMEM;
2217                         else {
2218                                 root_stable_tree = buf;
2219                                 root_unstable_tree = buf + nr_node_ids;
2220                                 /* Stable tree is empty but not the unstable */
2221                                 root_unstable_tree[0] = one_unstable_tree[0];
2222                         }
2223                 }
2224                 if (!err) {
2225                         ksm_merge_across_nodes = knob;
2226                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2227                 }
2228         }
2229         mutex_unlock(&ksm_thread_mutex);
2230
2231         return err ? err : count;
2232 }
2233 KSM_ATTR(merge_across_nodes);
2234 #endif
2235
2236 static ssize_t pages_shared_show(struct kobject *kobj,
2237                                  struct kobj_attribute *attr, char *buf)
2238 {
2239         return sprintf(buf, "%lu\n", ksm_pages_shared);
2240 }
2241 KSM_ATTR_RO(pages_shared);
2242
2243 static ssize_t pages_sharing_show(struct kobject *kobj,
2244                                   struct kobj_attribute *attr, char *buf)
2245 {
2246         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2247 }
2248 KSM_ATTR_RO(pages_sharing);
2249
2250 static ssize_t pages_unshared_show(struct kobject *kobj,
2251                                    struct kobj_attribute *attr, char *buf)
2252 {
2253         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2254 }
2255 KSM_ATTR_RO(pages_unshared);
2256
2257 static ssize_t pages_volatile_show(struct kobject *kobj,
2258                                    struct kobj_attribute *attr, char *buf)
2259 {
2260         long ksm_pages_volatile;
2261
2262         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2263                                 - ksm_pages_sharing - ksm_pages_unshared;
2264         /*
2265          * It was not worth any locking to calculate that statistic,
2266          * but it might therefore sometimes be negative: conceal that.
2267          */
2268         if (ksm_pages_volatile < 0)
2269                 ksm_pages_volatile = 0;
2270         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2271 }
2272 KSM_ATTR_RO(pages_volatile);
2273
2274 static ssize_t full_scans_show(struct kobject *kobj,
2275                                struct kobj_attribute *attr, char *buf)
2276 {
2277         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2278 }
2279 KSM_ATTR_RO(full_scans);
2280
2281 static struct attribute *ksm_attrs[] = {
2282         &sleep_millisecs_attr.attr,
2283         &pages_to_scan_attr.attr,
2284         &run_attr.attr,
2285         &pages_shared_attr.attr,
2286         &pages_sharing_attr.attr,
2287         &pages_unshared_attr.attr,
2288         &pages_volatile_attr.attr,
2289         &full_scans_attr.attr,
2290 #ifdef CONFIG_NUMA
2291         &merge_across_nodes_attr.attr,
2292 #endif
2293         NULL,
2294 };
2295
2296 static struct attribute_group ksm_attr_group = {
2297         .attrs = ksm_attrs,
2298         .name = "ksm",
2299 };
2300 #endif /* CONFIG_SYSFS */
2301
2302 static int __init ksm_init(void)
2303 {
2304         struct task_struct *ksm_thread;
2305         int err;
2306
2307         err = ksm_slab_init();
2308         if (err)
2309                 goto out;
2310
2311         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2312         if (IS_ERR(ksm_thread)) {
2313                 pr_err("ksm: creating kthread failed\n");
2314                 err = PTR_ERR(ksm_thread);
2315                 goto out_free;
2316         }
2317
2318 #ifdef CONFIG_SYSFS
2319         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2320         if (err) {
2321                 pr_err("ksm: register sysfs failed\n");
2322                 kthread_stop(ksm_thread);
2323                 goto out_free;
2324         }
2325 #else
2326         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
2327
2328 #endif /* CONFIG_SYSFS */
2329
2330 #ifdef CONFIG_MEMORY_HOTREMOVE
2331         /* There is no significance to this priority 100 */
2332         hotplug_memory_notifier(ksm_memory_callback, 100);
2333 #endif
2334         return 0;
2335
2336 out_free:
2337         ksm_slab_free();
2338 out:
2339         return err;
2340 }
2341 subsys_initcall(ksm_init);