mm/fadvise.c: drain all pagevecs if POSIX_FADV_DONTNEED fails to discard all pages
[platform/adaptation/renesas_rcar/renesas_kernel.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hashtable.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
39 #include <linux/numa.h>
40
41 #include <asm/tlbflush.h>
42 #include "internal.h"
43
44 #ifdef CONFIG_NUMA
45 #define NUMA(x)         (x)
46 #define DO_NUMA(x)      do { (x); } while (0)
47 #else
48 #define NUMA(x)         (0)
49 #define DO_NUMA(x)      do { } while (0)
50 #endif
51
52 /*
53  * A few notes about the KSM scanning process,
54  * to make it easier to understand the data structures below:
55  *
56  * In order to reduce excessive scanning, KSM sorts the memory pages by their
57  * contents into a data structure that holds pointers to the pages' locations.
58  *
59  * Since the contents of the pages may change at any moment, KSM cannot just
60  * insert the pages into a normal sorted tree and expect it to find anything.
61  * Therefore KSM uses two data structures - the stable and the unstable tree.
62  *
63  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64  * by their contents.  Because each such page is write-protected, searching on
65  * this tree is fully assured to be working (except when pages are unmapped),
66  * and therefore this tree is called the stable tree.
67  *
68  * In addition to the stable tree, KSM uses a second data structure called the
69  * unstable tree: this tree holds pointers to pages which have been found to
70  * be "unchanged for a period of time".  The unstable tree sorts these pages
71  * by their contents, but since they are not write-protected, KSM cannot rely
72  * upon the unstable tree to work correctly - the unstable tree is liable to
73  * be corrupted as its contents are modified, and so it is called unstable.
74  *
75  * KSM solves this problem by several techniques:
76  *
77  * 1) The unstable tree is flushed every time KSM completes scanning all
78  *    memory areas, and then the tree is rebuilt again from the beginning.
79  * 2) KSM will only insert into the unstable tree, pages whose hash value
80  *    has not changed since the previous scan of all memory areas.
81  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82  *    colors of the nodes and not on their contents, assuring that even when
83  *    the tree gets "corrupted" it won't get out of balance, so scanning time
84  *    remains the same (also, searching and inserting nodes in an rbtree uses
85  *    the same algorithm, so we have no overhead when we flush and rebuild).
86  * 4) KSM never flushes the stable tree, which means that even if it were to
87  *    take 10 attempts to find a page in the unstable tree, once it is found,
88  *    it is secured in the stable tree.  (When we scan a new page, we first
89  *    compare it against the stable tree, and then against the unstable tree.)
90  */
91
92 /**
93  * struct mm_slot - ksm information per mm that is being scanned
94  * @link: link to the mm_slots hash list
95  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
96  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
97  * @mm: the mm that this information is valid for
98  */
99 struct mm_slot {
100         struct hlist_node link;
101         struct list_head mm_list;
102         struct rmap_item *rmap_list;
103         struct mm_struct *mm;
104 };
105
106 /**
107  * struct ksm_scan - cursor for scanning
108  * @mm_slot: the current mm_slot we are scanning
109  * @address: the next address inside that to be scanned
110  * @rmap_list: link to the next rmap to be scanned in the rmap_list
111  * @seqnr: count of completed full scans (needed when removing unstable node)
112  *
113  * There is only the one ksm_scan instance of this cursor structure.
114  */
115 struct ksm_scan {
116         struct mm_slot *mm_slot;
117         unsigned long address;
118         struct rmap_item **rmap_list;
119         unsigned long seqnr;
120 };
121
122 /**
123  * struct stable_node - node of the stable rbtree
124  * @node: rb node of this ksm page in the stable tree
125  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
126  * @list: linked into migrate_nodes, pending placement in the proper node tree
127  * @hlist: hlist head of rmap_items using this ksm page
128  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
129  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
130  */
131 struct stable_node {
132         union {
133                 struct rb_node node;    /* when node of stable tree */
134                 struct {                /* when listed for migration */
135                         struct list_head *head;
136                         struct list_head list;
137                 };
138         };
139         struct hlist_head hlist;
140         unsigned long kpfn;
141 #ifdef CONFIG_NUMA
142         int nid;
143 #endif
144 };
145
146 /**
147  * struct rmap_item - reverse mapping item for virtual addresses
148  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
149  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
150  * @mm: the memory structure this rmap_item is pointing into
151  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
152  * @oldchecksum: previous checksum of the page at that virtual address
153  * @nid: NUMA node id of unstable tree in which linked (may not match page)
154  * @node: rb node of this rmap_item in the unstable tree
155  * @head: pointer to stable_node heading this list in the stable tree
156  * @hlist: link into hlist of rmap_items hanging off that stable_node
157  */
158 struct rmap_item {
159         struct rmap_item *rmap_list;
160         struct anon_vma *anon_vma;      /* when stable */
161         struct mm_struct *mm;
162         unsigned long address;          /* + low bits used for flags below */
163         unsigned int oldchecksum;       /* when unstable */
164 #ifdef CONFIG_NUMA
165         int nid;
166 #endif
167         union {
168                 struct rb_node node;    /* when node of unstable tree */
169                 struct {                /* when listed from stable tree */
170                         struct stable_node *head;
171                         struct hlist_node hlist;
172                 };
173         };
174 };
175
176 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
177 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
178 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
179
180 /* The stable and unstable tree heads */
181 static struct rb_root root_unstable_tree[MAX_NUMNODES];
182 static struct rb_root root_stable_tree[MAX_NUMNODES];
183
184 /* Recently migrated nodes of stable tree, pending proper placement */
185 static LIST_HEAD(migrate_nodes);
186
187 #define MM_SLOTS_HASH_BITS 10
188 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
189
190 static struct mm_slot ksm_mm_head = {
191         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
192 };
193 static struct ksm_scan ksm_scan = {
194         .mm_slot = &ksm_mm_head,
195 };
196
197 static struct kmem_cache *rmap_item_cache;
198 static struct kmem_cache *stable_node_cache;
199 static struct kmem_cache *mm_slot_cache;
200
201 /* The number of nodes in the stable tree */
202 static unsigned long ksm_pages_shared;
203
204 /* The number of page slots additionally sharing those nodes */
205 static unsigned long ksm_pages_sharing;
206
207 /* The number of nodes in the unstable tree */
208 static unsigned long ksm_pages_unshared;
209
210 /* The number of rmap_items in use: to calculate pages_volatile */
211 static unsigned long ksm_rmap_items;
212
213 /* Number of pages ksmd should scan in one batch */
214 static unsigned int ksm_thread_pages_to_scan = 100;
215
216 /* Milliseconds ksmd should sleep between batches */
217 static unsigned int ksm_thread_sleep_millisecs = 20;
218
219 #ifdef CONFIG_NUMA
220 /* Zeroed when merging across nodes is not allowed */
221 static unsigned int ksm_merge_across_nodes = 1;
222 #else
223 #define ksm_merge_across_nodes  1U
224 #endif
225
226 #define KSM_RUN_STOP    0
227 #define KSM_RUN_MERGE   1
228 #define KSM_RUN_UNMERGE 2
229 #define KSM_RUN_OFFLINE 4
230 static unsigned long ksm_run = KSM_RUN_STOP;
231 static void wait_while_offlining(void);
232
233 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
234 static DEFINE_MUTEX(ksm_thread_mutex);
235 static DEFINE_SPINLOCK(ksm_mmlist_lock);
236
237 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
238                 sizeof(struct __struct), __alignof__(struct __struct),\
239                 (__flags), NULL)
240
241 static int __init ksm_slab_init(void)
242 {
243         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
244         if (!rmap_item_cache)
245                 goto out;
246
247         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
248         if (!stable_node_cache)
249                 goto out_free1;
250
251         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
252         if (!mm_slot_cache)
253                 goto out_free2;
254
255         return 0;
256
257 out_free2:
258         kmem_cache_destroy(stable_node_cache);
259 out_free1:
260         kmem_cache_destroy(rmap_item_cache);
261 out:
262         return -ENOMEM;
263 }
264
265 static void __init ksm_slab_free(void)
266 {
267         kmem_cache_destroy(mm_slot_cache);
268         kmem_cache_destroy(stable_node_cache);
269         kmem_cache_destroy(rmap_item_cache);
270         mm_slot_cache = NULL;
271 }
272
273 static inline struct rmap_item *alloc_rmap_item(void)
274 {
275         struct rmap_item *rmap_item;
276
277         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
278         if (rmap_item)
279                 ksm_rmap_items++;
280         return rmap_item;
281 }
282
283 static inline void free_rmap_item(struct rmap_item *rmap_item)
284 {
285         ksm_rmap_items--;
286         rmap_item->mm = NULL;   /* debug safety */
287         kmem_cache_free(rmap_item_cache, rmap_item);
288 }
289
290 static inline struct stable_node *alloc_stable_node(void)
291 {
292         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
293 }
294
295 static inline void free_stable_node(struct stable_node *stable_node)
296 {
297         kmem_cache_free(stable_node_cache, stable_node);
298 }
299
300 static inline struct mm_slot *alloc_mm_slot(void)
301 {
302         if (!mm_slot_cache)     /* initialization failed */
303                 return NULL;
304         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
305 }
306
307 static inline void free_mm_slot(struct mm_slot *mm_slot)
308 {
309         kmem_cache_free(mm_slot_cache, mm_slot);
310 }
311
312 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
313 {
314         struct hlist_node *node;
315         struct mm_slot *slot;
316
317         hash_for_each_possible(mm_slots_hash, slot, node, link, (unsigned long)mm)
318                 if (slot->mm == mm)
319                         return slot;
320
321         return NULL;
322 }
323
324 static void insert_to_mm_slots_hash(struct mm_struct *mm,
325                                     struct mm_slot *mm_slot)
326 {
327         mm_slot->mm = mm;
328         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
329 }
330
331 /*
332  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
333  * page tables after it has passed through ksm_exit() - which, if necessary,
334  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
335  * a special flag: they can just back out as soon as mm_users goes to zero.
336  * ksm_test_exit() is used throughout to make this test for exit: in some
337  * places for correctness, in some places just to avoid unnecessary work.
338  */
339 static inline bool ksm_test_exit(struct mm_struct *mm)
340 {
341         return atomic_read(&mm->mm_users) == 0;
342 }
343
344 /*
345  * We use break_ksm to break COW on a ksm page: it's a stripped down
346  *
347  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
348  *              put_page(page);
349  *
350  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
351  * in case the application has unmapped and remapped mm,addr meanwhile.
352  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
353  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
354  */
355 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
356 {
357         struct page *page;
358         int ret = 0;
359
360         do {
361                 cond_resched();
362                 page = follow_page(vma, addr, FOLL_GET);
363                 if (IS_ERR_OR_NULL(page))
364                         break;
365                 if (PageKsm(page))
366                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
367                                                         FAULT_FLAG_WRITE);
368                 else
369                         ret = VM_FAULT_WRITE;
370                 put_page(page);
371         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
372         /*
373          * We must loop because handle_mm_fault() may back out if there's
374          * any difficulty e.g. if pte accessed bit gets updated concurrently.
375          *
376          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
377          * COW has been broken, even if the vma does not permit VM_WRITE;
378          * but note that a concurrent fault might break PageKsm for us.
379          *
380          * VM_FAULT_SIGBUS could occur if we race with truncation of the
381          * backing file, which also invalidates anonymous pages: that's
382          * okay, that truncation will have unmapped the PageKsm for us.
383          *
384          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
385          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
386          * current task has TIF_MEMDIE set, and will be OOM killed on return
387          * to user; and ksmd, having no mm, would never be chosen for that.
388          *
389          * But if the mm is in a limited mem_cgroup, then the fault may fail
390          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
391          * even ksmd can fail in this way - though it's usually breaking ksm
392          * just to undo a merge it made a moment before, so unlikely to oom.
393          *
394          * That's a pity: we might therefore have more kernel pages allocated
395          * than we're counting as nodes in the stable tree; but ksm_do_scan
396          * will retry to break_cow on each pass, so should recover the page
397          * in due course.  The important thing is to not let VM_MERGEABLE
398          * be cleared while any such pages might remain in the area.
399          */
400         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
401 }
402
403 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
404                 unsigned long addr)
405 {
406         struct vm_area_struct *vma;
407         if (ksm_test_exit(mm))
408                 return NULL;
409         vma = find_vma(mm, addr);
410         if (!vma || vma->vm_start > addr)
411                 return NULL;
412         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
413                 return NULL;
414         return vma;
415 }
416
417 static void break_cow(struct rmap_item *rmap_item)
418 {
419         struct mm_struct *mm = rmap_item->mm;
420         unsigned long addr = rmap_item->address;
421         struct vm_area_struct *vma;
422
423         /*
424          * It is not an accident that whenever we want to break COW
425          * to undo, we also need to drop a reference to the anon_vma.
426          */
427         put_anon_vma(rmap_item->anon_vma);
428
429         down_read(&mm->mmap_sem);
430         vma = find_mergeable_vma(mm, addr);
431         if (vma)
432                 break_ksm(vma, addr);
433         up_read(&mm->mmap_sem);
434 }
435
436 static struct page *page_trans_compound_anon(struct page *page)
437 {
438         if (PageTransCompound(page)) {
439                 struct page *head = compound_trans_head(page);
440                 /*
441                  * head may actually be splitted and freed from under
442                  * us but it's ok here.
443                  */
444                 if (PageAnon(head))
445                         return head;
446         }
447         return NULL;
448 }
449
450 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
451 {
452         struct mm_struct *mm = rmap_item->mm;
453         unsigned long addr = rmap_item->address;
454         struct vm_area_struct *vma;
455         struct page *page;
456
457         down_read(&mm->mmap_sem);
458         vma = find_mergeable_vma(mm, addr);
459         if (!vma)
460                 goto out;
461
462         page = follow_page(vma, addr, FOLL_GET);
463         if (IS_ERR_OR_NULL(page))
464                 goto out;
465         if (PageAnon(page) || page_trans_compound_anon(page)) {
466                 flush_anon_page(vma, page, addr);
467                 flush_dcache_page(page);
468         } else {
469                 put_page(page);
470 out:            page = NULL;
471         }
472         up_read(&mm->mmap_sem);
473         return page;
474 }
475
476 /*
477  * This helper is used for getting right index into array of tree roots.
478  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
479  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
480  * every node has its own stable and unstable tree.
481  */
482 static inline int get_kpfn_nid(unsigned long kpfn)
483 {
484         return ksm_merge_across_nodes ? 0 : pfn_to_nid(kpfn);
485 }
486
487 static void remove_node_from_stable_tree(struct stable_node *stable_node)
488 {
489         struct rmap_item *rmap_item;
490         struct hlist_node *hlist;
491
492         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
493                 if (rmap_item->hlist.next)
494                         ksm_pages_sharing--;
495                 else
496                         ksm_pages_shared--;
497                 put_anon_vma(rmap_item->anon_vma);
498                 rmap_item->address &= PAGE_MASK;
499                 cond_resched();
500         }
501
502         if (stable_node->head == &migrate_nodes)
503                 list_del(&stable_node->list);
504         else
505                 rb_erase(&stable_node->node,
506                          &root_stable_tree[NUMA(stable_node->nid)]);
507         free_stable_node(stable_node);
508 }
509
510 /*
511  * get_ksm_page: checks if the page indicated by the stable node
512  * is still its ksm page, despite having held no reference to it.
513  * In which case we can trust the content of the page, and it
514  * returns the gotten page; but if the page has now been zapped,
515  * remove the stale node from the stable tree and return NULL.
516  * But beware, the stable node's page might be being migrated.
517  *
518  * You would expect the stable_node to hold a reference to the ksm page.
519  * But if it increments the page's count, swapping out has to wait for
520  * ksmd to come around again before it can free the page, which may take
521  * seconds or even minutes: much too unresponsive.  So instead we use a
522  * "keyhole reference": access to the ksm page from the stable node peeps
523  * out through its keyhole to see if that page still holds the right key,
524  * pointing back to this stable node.  This relies on freeing a PageAnon
525  * page to reset its page->mapping to NULL, and relies on no other use of
526  * a page to put something that might look like our key in page->mapping.
527  * is on its way to being freed; but it is an anomaly to bear in mind.
528  */
529 static struct page *get_ksm_page(struct stable_node *stable_node, bool locked)
530 {
531         struct page *page;
532         void *expected_mapping;
533         unsigned long kpfn;
534
535         expected_mapping = (void *)stable_node +
536                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
537 again:
538         kpfn = ACCESS_ONCE(stable_node->kpfn);
539         page = pfn_to_page(kpfn);
540
541         /*
542          * page is computed from kpfn, so on most architectures reading
543          * page->mapping is naturally ordered after reading node->kpfn,
544          * but on Alpha we need to be more careful.
545          */
546         smp_read_barrier_depends();
547         if (ACCESS_ONCE(page->mapping) != expected_mapping)
548                 goto stale;
549
550         /*
551          * We cannot do anything with the page while its refcount is 0.
552          * Usually 0 means free, or tail of a higher-order page: in which
553          * case this node is no longer referenced, and should be freed;
554          * however, it might mean that the page is under page_freeze_refs().
555          * The __remove_mapping() case is easy, again the node is now stale;
556          * but if page is swapcache in migrate_page_move_mapping(), it might
557          * still be our page, in which case it's essential to keep the node.
558          */
559         while (!get_page_unless_zero(page)) {
560                 /*
561                  * Another check for page->mapping != expected_mapping would
562                  * work here too.  We have chosen the !PageSwapCache test to
563                  * optimize the common case, when the page is or is about to
564                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
565                  * in the freeze_refs section of __remove_mapping(); but Anon
566                  * page->mapping reset to NULL later, in free_pages_prepare().
567                  */
568                 if (!PageSwapCache(page))
569                         goto stale;
570                 cpu_relax();
571         }
572
573         if (ACCESS_ONCE(page->mapping) != expected_mapping) {
574                 put_page(page);
575                 goto stale;
576         }
577
578         if (locked) {
579                 lock_page(page);
580                 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
581                         unlock_page(page);
582                         put_page(page);
583                         goto stale;
584                 }
585         }
586         return page;
587
588 stale:
589         /*
590          * We come here from above when page->mapping or !PageSwapCache
591          * suggests that the node is stale; but it might be under migration.
592          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
593          * before checking whether node->kpfn has been changed.
594          */
595         smp_rmb();
596         if (ACCESS_ONCE(stable_node->kpfn) != kpfn)
597                 goto again;
598         remove_node_from_stable_tree(stable_node);
599         return NULL;
600 }
601
602 /*
603  * Removing rmap_item from stable or unstable tree.
604  * This function will clean the information from the stable/unstable tree.
605  */
606 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
607 {
608         if (rmap_item->address & STABLE_FLAG) {
609                 struct stable_node *stable_node;
610                 struct page *page;
611
612                 stable_node = rmap_item->head;
613                 page = get_ksm_page(stable_node, true);
614                 if (!page)
615                         goto out;
616
617                 hlist_del(&rmap_item->hlist);
618                 unlock_page(page);
619                 put_page(page);
620
621                 if (stable_node->hlist.first)
622                         ksm_pages_sharing--;
623                 else
624                         ksm_pages_shared--;
625
626                 put_anon_vma(rmap_item->anon_vma);
627                 rmap_item->address &= PAGE_MASK;
628
629         } else if (rmap_item->address & UNSTABLE_FLAG) {
630                 unsigned char age;
631                 /*
632                  * Usually ksmd can and must skip the rb_erase, because
633                  * root_unstable_tree was already reset to RB_ROOT.
634                  * But be careful when an mm is exiting: do the rb_erase
635                  * if this rmap_item was inserted by this scan, rather
636                  * than left over from before.
637                  */
638                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
639                 BUG_ON(age > 1);
640                 if (!age)
641                         rb_erase(&rmap_item->node,
642                                  &root_unstable_tree[NUMA(rmap_item->nid)]);
643                 ksm_pages_unshared--;
644                 rmap_item->address &= PAGE_MASK;
645         }
646 out:
647         cond_resched();         /* we're called from many long loops */
648 }
649
650 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
651                                        struct rmap_item **rmap_list)
652 {
653         while (*rmap_list) {
654                 struct rmap_item *rmap_item = *rmap_list;
655                 *rmap_list = rmap_item->rmap_list;
656                 remove_rmap_item_from_tree(rmap_item);
657                 free_rmap_item(rmap_item);
658         }
659 }
660
661 /*
662  * Though it's very tempting to unmerge rmap_items from stable tree rather
663  * than check every pte of a given vma, the locking doesn't quite work for
664  * that - an rmap_item is assigned to the stable tree after inserting ksm
665  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
666  * rmap_items from parent to child at fork time (so as not to waste time
667  * if exit comes before the next scan reaches it).
668  *
669  * Similarly, although we'd like to remove rmap_items (so updating counts
670  * and freeing memory) when unmerging an area, it's easier to leave that
671  * to the next pass of ksmd - consider, for example, how ksmd might be
672  * in cmp_and_merge_page on one of the rmap_items we would be removing.
673  */
674 static int unmerge_ksm_pages(struct vm_area_struct *vma,
675                              unsigned long start, unsigned long end)
676 {
677         unsigned long addr;
678         int err = 0;
679
680         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
681                 if (ksm_test_exit(vma->vm_mm))
682                         break;
683                 if (signal_pending(current))
684                         err = -ERESTARTSYS;
685                 else
686                         err = break_ksm(vma, addr);
687         }
688         return err;
689 }
690
691 #ifdef CONFIG_SYSFS
692 /*
693  * Only called through the sysfs control interface:
694  */
695 static int remove_stable_node(struct stable_node *stable_node)
696 {
697         struct page *page;
698         int err;
699
700         page = get_ksm_page(stable_node, true);
701         if (!page) {
702                 /*
703                  * get_ksm_page did remove_node_from_stable_tree itself.
704                  */
705                 return 0;
706         }
707
708         if (WARN_ON_ONCE(page_mapped(page)))
709                 err = -EBUSY;
710         else {
711                 /*
712                  * This page might be in a pagevec waiting to be freed,
713                  * or it might be PageSwapCache (perhaps under writeback),
714                  * or it might have been removed from swapcache a moment ago.
715                  */
716                 set_page_stable_node(page, NULL);
717                 remove_node_from_stable_tree(stable_node);
718                 err = 0;
719         }
720
721         unlock_page(page);
722         put_page(page);
723         return err;
724 }
725
726 static int remove_all_stable_nodes(void)
727 {
728         struct stable_node *stable_node;
729         struct list_head *this, *next;
730         int nid;
731         int err = 0;
732
733         for (nid = 0; nid < nr_node_ids; nid++) {
734                 while (root_stable_tree[nid].rb_node) {
735                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
736                                                 struct stable_node, node);
737                         if (remove_stable_node(stable_node)) {
738                                 err = -EBUSY;
739                                 break;  /* proceed to next nid */
740                         }
741                         cond_resched();
742                 }
743         }
744         list_for_each_safe(this, next, &migrate_nodes) {
745                 stable_node = list_entry(this, struct stable_node, list);
746                 if (remove_stable_node(stable_node))
747                         err = -EBUSY;
748                 cond_resched();
749         }
750         return err;
751 }
752
753 static int unmerge_and_remove_all_rmap_items(void)
754 {
755         struct mm_slot *mm_slot;
756         struct mm_struct *mm;
757         struct vm_area_struct *vma;
758         int err = 0;
759
760         spin_lock(&ksm_mmlist_lock);
761         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
762                                                 struct mm_slot, mm_list);
763         spin_unlock(&ksm_mmlist_lock);
764
765         for (mm_slot = ksm_scan.mm_slot;
766                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
767                 mm = mm_slot->mm;
768                 down_read(&mm->mmap_sem);
769                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
770                         if (ksm_test_exit(mm))
771                                 break;
772                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
773                                 continue;
774                         err = unmerge_ksm_pages(vma,
775                                                 vma->vm_start, vma->vm_end);
776                         if (err)
777                                 goto error;
778                 }
779
780                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
781
782                 spin_lock(&ksm_mmlist_lock);
783                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
784                                                 struct mm_slot, mm_list);
785                 if (ksm_test_exit(mm)) {
786                         hash_del(&mm_slot->link);
787                         list_del(&mm_slot->mm_list);
788                         spin_unlock(&ksm_mmlist_lock);
789
790                         free_mm_slot(mm_slot);
791                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
792                         up_read(&mm->mmap_sem);
793                         mmdrop(mm);
794                 } else {
795                         spin_unlock(&ksm_mmlist_lock);
796                         up_read(&mm->mmap_sem);
797                 }
798         }
799
800         /* Clean up stable nodes, but don't worry if some are still busy */
801         remove_all_stable_nodes();
802         ksm_scan.seqnr = 0;
803         return 0;
804
805 error:
806         up_read(&mm->mmap_sem);
807         spin_lock(&ksm_mmlist_lock);
808         ksm_scan.mm_slot = &ksm_mm_head;
809         spin_unlock(&ksm_mmlist_lock);
810         return err;
811 }
812 #endif /* CONFIG_SYSFS */
813
814 static u32 calc_checksum(struct page *page)
815 {
816         u32 checksum;
817         void *addr = kmap_atomic(page);
818         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
819         kunmap_atomic(addr);
820         return checksum;
821 }
822
823 static int memcmp_pages(struct page *page1, struct page *page2)
824 {
825         char *addr1, *addr2;
826         int ret;
827
828         addr1 = kmap_atomic(page1);
829         addr2 = kmap_atomic(page2);
830         ret = memcmp(addr1, addr2, PAGE_SIZE);
831         kunmap_atomic(addr2);
832         kunmap_atomic(addr1);
833         return ret;
834 }
835
836 static inline int pages_identical(struct page *page1, struct page *page2)
837 {
838         return !memcmp_pages(page1, page2);
839 }
840
841 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
842                               pte_t *orig_pte)
843 {
844         struct mm_struct *mm = vma->vm_mm;
845         unsigned long addr;
846         pte_t *ptep;
847         spinlock_t *ptl;
848         int swapped;
849         int err = -EFAULT;
850         unsigned long mmun_start;       /* For mmu_notifiers */
851         unsigned long mmun_end;         /* For mmu_notifiers */
852
853         addr = page_address_in_vma(page, vma);
854         if (addr == -EFAULT)
855                 goto out;
856
857         BUG_ON(PageTransCompound(page));
858
859         mmun_start = addr;
860         mmun_end   = addr + PAGE_SIZE;
861         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
862
863         ptep = page_check_address(page, mm, addr, &ptl, 0);
864         if (!ptep)
865                 goto out_mn;
866
867         if (pte_write(*ptep) || pte_dirty(*ptep)) {
868                 pte_t entry;
869
870                 swapped = PageSwapCache(page);
871                 flush_cache_page(vma, addr, page_to_pfn(page));
872                 /*
873                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
874                  * take any lock, therefore the check that we are going to make
875                  * with the pagecount against the mapcount is racey and
876                  * O_DIRECT can happen right after the check.
877                  * So we clear the pte and flush the tlb before the check
878                  * this assure us that no O_DIRECT can happen after the check
879                  * or in the middle of the check.
880                  */
881                 entry = ptep_clear_flush(vma, addr, ptep);
882                 /*
883                  * Check that no O_DIRECT or similar I/O is in progress on the
884                  * page
885                  */
886                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
887                         set_pte_at(mm, addr, ptep, entry);
888                         goto out_unlock;
889                 }
890                 if (pte_dirty(entry))
891                         set_page_dirty(page);
892                 entry = pte_mkclean(pte_wrprotect(entry));
893                 set_pte_at_notify(mm, addr, ptep, entry);
894         }
895         *orig_pte = *ptep;
896         err = 0;
897
898 out_unlock:
899         pte_unmap_unlock(ptep, ptl);
900 out_mn:
901         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
902 out:
903         return err;
904 }
905
906 /**
907  * replace_page - replace page in vma by new ksm page
908  * @vma:      vma that holds the pte pointing to page
909  * @page:     the page we are replacing by kpage
910  * @kpage:    the ksm page we replace page by
911  * @orig_pte: the original value of the pte
912  *
913  * Returns 0 on success, -EFAULT on failure.
914  */
915 static int replace_page(struct vm_area_struct *vma, struct page *page,
916                         struct page *kpage, pte_t orig_pte)
917 {
918         struct mm_struct *mm = vma->vm_mm;
919         pmd_t *pmd;
920         pte_t *ptep;
921         spinlock_t *ptl;
922         unsigned long addr;
923         int err = -EFAULT;
924         unsigned long mmun_start;       /* For mmu_notifiers */
925         unsigned long mmun_end;         /* For mmu_notifiers */
926
927         addr = page_address_in_vma(page, vma);
928         if (addr == -EFAULT)
929                 goto out;
930
931         pmd = mm_find_pmd(mm, addr);
932         if (!pmd)
933                 goto out;
934         BUG_ON(pmd_trans_huge(*pmd));
935
936         mmun_start = addr;
937         mmun_end   = addr + PAGE_SIZE;
938         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
939
940         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
941         if (!pte_same(*ptep, orig_pte)) {
942                 pte_unmap_unlock(ptep, ptl);
943                 goto out_mn;
944         }
945
946         get_page(kpage);
947         page_add_anon_rmap(kpage, vma, addr);
948
949         flush_cache_page(vma, addr, pte_pfn(*ptep));
950         ptep_clear_flush(vma, addr, ptep);
951         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
952
953         page_remove_rmap(page);
954         if (!page_mapped(page))
955                 try_to_free_swap(page);
956         put_page(page);
957
958         pte_unmap_unlock(ptep, ptl);
959         err = 0;
960 out_mn:
961         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
962 out:
963         return err;
964 }
965
966 static int page_trans_compound_anon_split(struct page *page)
967 {
968         int ret = 0;
969         struct page *transhuge_head = page_trans_compound_anon(page);
970         if (transhuge_head) {
971                 /* Get the reference on the head to split it. */
972                 if (get_page_unless_zero(transhuge_head)) {
973                         /*
974                          * Recheck we got the reference while the head
975                          * was still anonymous.
976                          */
977                         if (PageAnon(transhuge_head))
978                                 ret = split_huge_page(transhuge_head);
979                         else
980                                 /*
981                                  * Retry later if split_huge_page run
982                                  * from under us.
983                                  */
984                                 ret = 1;
985                         put_page(transhuge_head);
986                 } else
987                         /* Retry later if split_huge_page run from under us. */
988                         ret = 1;
989         }
990         return ret;
991 }
992
993 /*
994  * try_to_merge_one_page - take two pages and merge them into one
995  * @vma: the vma that holds the pte pointing to page
996  * @page: the PageAnon page that we want to replace with kpage
997  * @kpage: the PageKsm page that we want to map instead of page,
998  *         or NULL the first time when we want to use page as kpage.
999  *
1000  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1001  */
1002 static int try_to_merge_one_page(struct vm_area_struct *vma,
1003                                  struct page *page, struct page *kpage)
1004 {
1005         pte_t orig_pte = __pte(0);
1006         int err = -EFAULT;
1007
1008         if (page == kpage)                      /* ksm page forked */
1009                 return 0;
1010
1011         if (!(vma->vm_flags & VM_MERGEABLE))
1012                 goto out;
1013         if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1014                 goto out;
1015         BUG_ON(PageTransCompound(page));
1016         if (!PageAnon(page))
1017                 goto out;
1018
1019         /*
1020          * We need the page lock to read a stable PageSwapCache in
1021          * write_protect_page().  We use trylock_page() instead of
1022          * lock_page() because we don't want to wait here - we
1023          * prefer to continue scanning and merging different pages,
1024          * then come back to this page when it is unlocked.
1025          */
1026         if (!trylock_page(page))
1027                 goto out;
1028         /*
1029          * If this anonymous page is mapped only here, its pte may need
1030          * to be write-protected.  If it's mapped elsewhere, all of its
1031          * ptes are necessarily already write-protected.  But in either
1032          * case, we need to lock and check page_count is not raised.
1033          */
1034         if (write_protect_page(vma, page, &orig_pte) == 0) {
1035                 if (!kpage) {
1036                         /*
1037                          * While we hold page lock, upgrade page from
1038                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1039                          * stable_tree_insert() will update stable_node.
1040                          */
1041                         set_page_stable_node(page, NULL);
1042                         mark_page_accessed(page);
1043                         err = 0;
1044                 } else if (pages_identical(page, kpage))
1045                         err = replace_page(vma, page, kpage, orig_pte);
1046         }
1047
1048         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1049                 munlock_vma_page(page);
1050                 if (!PageMlocked(kpage)) {
1051                         unlock_page(page);
1052                         lock_page(kpage);
1053                         mlock_vma_page(kpage);
1054                         page = kpage;           /* for final unlock */
1055                 }
1056         }
1057
1058         unlock_page(page);
1059 out:
1060         return err;
1061 }
1062
1063 /*
1064  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1065  * but no new kernel page is allocated: kpage must already be a ksm page.
1066  *
1067  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1068  */
1069 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1070                                       struct page *page, struct page *kpage)
1071 {
1072         struct mm_struct *mm = rmap_item->mm;
1073         struct vm_area_struct *vma;
1074         int err = -EFAULT;
1075
1076         down_read(&mm->mmap_sem);
1077         if (ksm_test_exit(mm))
1078                 goto out;
1079         vma = find_vma(mm, rmap_item->address);
1080         if (!vma || vma->vm_start > rmap_item->address)
1081                 goto out;
1082
1083         err = try_to_merge_one_page(vma, page, kpage);
1084         if (err)
1085                 goto out;
1086
1087         /* Must get reference to anon_vma while still holding mmap_sem */
1088         rmap_item->anon_vma = vma->anon_vma;
1089         get_anon_vma(vma->anon_vma);
1090 out:
1091         up_read(&mm->mmap_sem);
1092         return err;
1093 }
1094
1095 /*
1096  * try_to_merge_two_pages - take two identical pages and prepare them
1097  * to be merged into one page.
1098  *
1099  * This function returns the kpage if we successfully merged two identical
1100  * pages into one ksm page, NULL otherwise.
1101  *
1102  * Note that this function upgrades page to ksm page: if one of the pages
1103  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1104  */
1105 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1106                                            struct page *page,
1107                                            struct rmap_item *tree_rmap_item,
1108                                            struct page *tree_page)
1109 {
1110         int err;
1111
1112         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1113         if (!err) {
1114                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1115                                                         tree_page, page);
1116                 /*
1117                  * If that fails, we have a ksm page with only one pte
1118                  * pointing to it: so break it.
1119                  */
1120                 if (err)
1121                         break_cow(rmap_item);
1122         }
1123         return err ? NULL : page;
1124 }
1125
1126 /*
1127  * stable_tree_search - search for page inside the stable tree
1128  *
1129  * This function checks if there is a page inside the stable tree
1130  * with identical content to the page that we are scanning right now.
1131  *
1132  * This function returns the stable tree node of identical content if found,
1133  * NULL otherwise.
1134  */
1135 static struct page *stable_tree_search(struct page *page)
1136 {
1137         int nid;
1138         struct rb_node **new;
1139         struct rb_node *parent;
1140         struct stable_node *stable_node;
1141         struct stable_node *page_node;
1142
1143         page_node = page_stable_node(page);
1144         if (page_node && page_node->head != &migrate_nodes) {
1145                 /* ksm page forked */
1146                 get_page(page);
1147                 return page;
1148         }
1149
1150         nid = get_kpfn_nid(page_to_pfn(page));
1151 again:
1152         new = &root_stable_tree[nid].rb_node;
1153         parent = NULL;
1154
1155         while (*new) {
1156                 struct page *tree_page;
1157                 int ret;
1158
1159                 cond_resched();
1160                 stable_node = rb_entry(*new, struct stable_node, node);
1161                 tree_page = get_ksm_page(stable_node, false);
1162                 if (!tree_page)
1163                         return NULL;
1164
1165                 ret = memcmp_pages(page, tree_page);
1166                 put_page(tree_page);
1167
1168                 parent = *new;
1169                 if (ret < 0)
1170                         new = &parent->rb_left;
1171                 else if (ret > 0)
1172                         new = &parent->rb_right;
1173                 else {
1174                         /*
1175                          * Lock and unlock the stable_node's page (which
1176                          * might already have been migrated) so that page
1177                          * migration is sure to notice its raised count.
1178                          * It would be more elegant to return stable_node
1179                          * than kpage, but that involves more changes.
1180                          */
1181                         tree_page = get_ksm_page(stable_node, true);
1182                         if (tree_page) {
1183                                 unlock_page(tree_page);
1184                                 if (get_kpfn_nid(stable_node->kpfn) !=
1185                                                 NUMA(stable_node->nid)) {
1186                                         put_page(tree_page);
1187                                         goto replace;
1188                                 }
1189                                 return tree_page;
1190                         }
1191                         /*
1192                          * There is now a place for page_node, but the tree may
1193                          * have been rebalanced, so re-evaluate parent and new.
1194                          */
1195                         if (page_node)
1196                                 goto again;
1197                         return NULL;
1198                 }
1199         }
1200
1201         if (!page_node)
1202                 return NULL;
1203
1204         list_del(&page_node->list);
1205         DO_NUMA(page_node->nid = nid);
1206         rb_link_node(&page_node->node, parent, new);
1207         rb_insert_color(&page_node->node, &root_stable_tree[nid]);
1208         get_page(page);
1209         return page;
1210
1211 replace:
1212         if (page_node) {
1213                 list_del(&page_node->list);
1214                 DO_NUMA(page_node->nid = nid);
1215                 rb_replace_node(&stable_node->node,
1216                                 &page_node->node, &root_stable_tree[nid]);
1217                 get_page(page);
1218         } else {
1219                 rb_erase(&stable_node->node, &root_stable_tree[nid]);
1220                 page = NULL;
1221         }
1222         stable_node->head = &migrate_nodes;
1223         list_add(&stable_node->list, stable_node->head);
1224         return page;
1225 }
1226
1227 /*
1228  * stable_tree_insert - insert stable tree node pointing to new ksm page
1229  * into the stable tree.
1230  *
1231  * This function returns the stable tree node just allocated on success,
1232  * NULL otherwise.
1233  */
1234 static struct stable_node *stable_tree_insert(struct page *kpage)
1235 {
1236         int nid;
1237         unsigned long kpfn;
1238         struct rb_node **new;
1239         struct rb_node *parent = NULL;
1240         struct stable_node *stable_node;
1241
1242         kpfn = page_to_pfn(kpage);
1243         nid = get_kpfn_nid(kpfn);
1244         new = &root_stable_tree[nid].rb_node;
1245
1246         while (*new) {
1247                 struct page *tree_page;
1248                 int ret;
1249
1250                 cond_resched();
1251                 stable_node = rb_entry(*new, struct stable_node, node);
1252                 tree_page = get_ksm_page(stable_node, false);
1253                 if (!tree_page)
1254                         return NULL;
1255
1256                 ret = memcmp_pages(kpage, tree_page);
1257                 put_page(tree_page);
1258
1259                 parent = *new;
1260                 if (ret < 0)
1261                         new = &parent->rb_left;
1262                 else if (ret > 0)
1263                         new = &parent->rb_right;
1264                 else {
1265                         /*
1266                          * It is not a bug that stable_tree_search() didn't
1267                          * find this node: because at that time our page was
1268                          * not yet write-protected, so may have changed since.
1269                          */
1270                         return NULL;
1271                 }
1272         }
1273
1274         stable_node = alloc_stable_node();
1275         if (!stable_node)
1276                 return NULL;
1277
1278         INIT_HLIST_HEAD(&stable_node->hlist);
1279         stable_node->kpfn = kpfn;
1280         set_page_stable_node(kpage, stable_node);
1281         DO_NUMA(stable_node->nid = nid);
1282         rb_link_node(&stable_node->node, parent, new);
1283         rb_insert_color(&stable_node->node, &root_stable_tree[nid]);
1284
1285         return stable_node;
1286 }
1287
1288 /*
1289  * unstable_tree_search_insert - search for identical page,
1290  * else insert rmap_item into the unstable tree.
1291  *
1292  * This function searches for a page in the unstable tree identical to the
1293  * page currently being scanned; and if no identical page is found in the
1294  * tree, we insert rmap_item as a new object into the unstable tree.
1295  *
1296  * This function returns pointer to rmap_item found to be identical
1297  * to the currently scanned page, NULL otherwise.
1298  *
1299  * This function does both searching and inserting, because they share
1300  * the same walking algorithm in an rbtree.
1301  */
1302 static
1303 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1304                                               struct page *page,
1305                                               struct page **tree_pagep)
1306 {
1307         struct rb_node **new;
1308         struct rb_root *root;
1309         struct rb_node *parent = NULL;
1310         int nid;
1311
1312         nid = get_kpfn_nid(page_to_pfn(page));
1313         root = &root_unstable_tree[nid];
1314         new = &root->rb_node;
1315
1316         while (*new) {
1317                 struct rmap_item *tree_rmap_item;
1318                 struct page *tree_page;
1319                 int ret;
1320
1321                 cond_resched();
1322                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1323                 tree_page = get_mergeable_page(tree_rmap_item);
1324                 if (IS_ERR_OR_NULL(tree_page))
1325                         return NULL;
1326
1327                 /*
1328                  * Don't substitute a ksm page for a forked page.
1329                  */
1330                 if (page == tree_page) {
1331                         put_page(tree_page);
1332                         return NULL;
1333                 }
1334
1335                 /*
1336                  * If tree_page has been migrated to another NUMA node, it
1337                  * will be flushed out and put into the right unstable tree
1338                  * next time: only merge with it if merge_across_nodes.
1339                  */
1340                 if (!ksm_merge_across_nodes && page_to_nid(tree_page) != nid) {
1341                         put_page(tree_page);
1342                         return NULL;
1343                 }
1344
1345                 ret = memcmp_pages(page, tree_page);
1346
1347                 parent = *new;
1348                 if (ret < 0) {
1349                         put_page(tree_page);
1350                         new = &parent->rb_left;
1351                 } else if (ret > 0) {
1352                         put_page(tree_page);
1353                         new = &parent->rb_right;
1354                 } else {
1355                         *tree_pagep = tree_page;
1356                         return tree_rmap_item;
1357                 }
1358         }
1359
1360         rmap_item->address |= UNSTABLE_FLAG;
1361         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1362         DO_NUMA(rmap_item->nid = nid);
1363         rb_link_node(&rmap_item->node, parent, new);
1364         rb_insert_color(&rmap_item->node, root);
1365
1366         ksm_pages_unshared++;
1367         return NULL;
1368 }
1369
1370 /*
1371  * stable_tree_append - add another rmap_item to the linked list of
1372  * rmap_items hanging off a given node of the stable tree, all sharing
1373  * the same ksm page.
1374  */
1375 static void stable_tree_append(struct rmap_item *rmap_item,
1376                                struct stable_node *stable_node)
1377 {
1378         rmap_item->head = stable_node;
1379         rmap_item->address |= STABLE_FLAG;
1380         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1381
1382         if (rmap_item->hlist.next)
1383                 ksm_pages_sharing++;
1384         else
1385                 ksm_pages_shared++;
1386 }
1387
1388 /*
1389  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1390  * if not, compare checksum to previous and if it's the same, see if page can
1391  * be inserted into the unstable tree, or merged with a page already there and
1392  * both transferred to the stable tree.
1393  *
1394  * @page: the page that we are searching identical page to.
1395  * @rmap_item: the reverse mapping into the virtual address of this page
1396  */
1397 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1398 {
1399         struct rmap_item *tree_rmap_item;
1400         struct page *tree_page = NULL;
1401         struct stable_node *stable_node;
1402         struct page *kpage;
1403         unsigned int checksum;
1404         int err;
1405
1406         stable_node = page_stable_node(page);
1407         if (stable_node) {
1408                 if (stable_node->head != &migrate_nodes &&
1409                     get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1410                         rb_erase(&stable_node->node,
1411                                  &root_stable_tree[NUMA(stable_node->nid)]);
1412                         stable_node->head = &migrate_nodes;
1413                         list_add(&stable_node->list, stable_node->head);
1414                 }
1415                 if (stable_node->head != &migrate_nodes &&
1416                     rmap_item->head == stable_node)
1417                         return;
1418         }
1419
1420         /* We first start with searching the page inside the stable tree */
1421         kpage = stable_tree_search(page);
1422         if (kpage == page && rmap_item->head == stable_node) {
1423                 put_page(kpage);
1424                 return;
1425         }
1426
1427         remove_rmap_item_from_tree(rmap_item);
1428
1429         if (kpage) {
1430                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1431                 if (!err) {
1432                         /*
1433                          * The page was successfully merged:
1434                          * add its rmap_item to the stable tree.
1435                          */
1436                         lock_page(kpage);
1437                         stable_tree_append(rmap_item, page_stable_node(kpage));
1438                         unlock_page(kpage);
1439                 }
1440                 put_page(kpage);
1441                 return;
1442         }
1443
1444         /*
1445          * If the hash value of the page has changed from the last time
1446          * we calculated it, this page is changing frequently: therefore we
1447          * don't want to insert it in the unstable tree, and we don't want
1448          * to waste our time searching for something identical to it there.
1449          */
1450         checksum = calc_checksum(page);
1451         if (rmap_item->oldchecksum != checksum) {
1452                 rmap_item->oldchecksum = checksum;
1453                 return;
1454         }
1455
1456         tree_rmap_item =
1457                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1458         if (tree_rmap_item) {
1459                 kpage = try_to_merge_two_pages(rmap_item, page,
1460                                                 tree_rmap_item, tree_page);
1461                 put_page(tree_page);
1462                 /*
1463                  * As soon as we merge this page, we want to remove the
1464                  * rmap_item of the page we have merged with from the unstable
1465                  * tree, and insert it instead as new node in the stable tree.
1466                  */
1467                 if (kpage) {
1468                         remove_rmap_item_from_tree(tree_rmap_item);
1469
1470                         lock_page(kpage);
1471                         stable_node = stable_tree_insert(kpage);
1472                         if (stable_node) {
1473                                 stable_tree_append(tree_rmap_item, stable_node);
1474                                 stable_tree_append(rmap_item, stable_node);
1475                         }
1476                         unlock_page(kpage);
1477
1478                         /*
1479                          * If we fail to insert the page into the stable tree,
1480                          * we will have 2 virtual addresses that are pointing
1481                          * to a ksm page left outside the stable tree,
1482                          * in which case we need to break_cow on both.
1483                          */
1484                         if (!stable_node) {
1485                                 break_cow(tree_rmap_item);
1486                                 break_cow(rmap_item);
1487                         }
1488                 }
1489         }
1490 }
1491
1492 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1493                                             struct rmap_item **rmap_list,
1494                                             unsigned long addr)
1495 {
1496         struct rmap_item *rmap_item;
1497
1498         while (*rmap_list) {
1499                 rmap_item = *rmap_list;
1500                 if ((rmap_item->address & PAGE_MASK) == addr)
1501                         return rmap_item;
1502                 if (rmap_item->address > addr)
1503                         break;
1504                 *rmap_list = rmap_item->rmap_list;
1505                 remove_rmap_item_from_tree(rmap_item);
1506                 free_rmap_item(rmap_item);
1507         }
1508
1509         rmap_item = alloc_rmap_item();
1510         if (rmap_item) {
1511                 /* It has already been zeroed */
1512                 rmap_item->mm = mm_slot->mm;
1513                 rmap_item->address = addr;
1514                 rmap_item->rmap_list = *rmap_list;
1515                 *rmap_list = rmap_item;
1516         }
1517         return rmap_item;
1518 }
1519
1520 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1521 {
1522         struct mm_struct *mm;
1523         struct mm_slot *slot;
1524         struct vm_area_struct *vma;
1525         struct rmap_item *rmap_item;
1526         int nid;
1527
1528         if (list_empty(&ksm_mm_head.mm_list))
1529                 return NULL;
1530
1531         slot = ksm_scan.mm_slot;
1532         if (slot == &ksm_mm_head) {
1533                 /*
1534                  * A number of pages can hang around indefinitely on per-cpu
1535                  * pagevecs, raised page count preventing write_protect_page
1536                  * from merging them.  Though it doesn't really matter much,
1537                  * it is puzzling to see some stuck in pages_volatile until
1538                  * other activity jostles them out, and they also prevented
1539                  * LTP's KSM test from succeeding deterministically; so drain
1540                  * them here (here rather than on entry to ksm_do_scan(),
1541                  * so we don't IPI too often when pages_to_scan is set low).
1542                  */
1543                 lru_add_drain_all();
1544
1545                 /*
1546                  * Whereas stale stable_nodes on the stable_tree itself
1547                  * get pruned in the regular course of stable_tree_search(),
1548                  * those moved out to the migrate_nodes list can accumulate:
1549                  * so prune them once before each full scan.
1550                  */
1551                 if (!ksm_merge_across_nodes) {
1552                         struct stable_node *stable_node;
1553                         struct list_head *this, *next;
1554                         struct page *page;
1555
1556                         list_for_each_safe(this, next, &migrate_nodes) {
1557                                 stable_node = list_entry(this,
1558                                                 struct stable_node, list);
1559                                 page = get_ksm_page(stable_node, false);
1560                                 if (page)
1561                                         put_page(page);
1562                                 cond_resched();
1563                         }
1564                 }
1565
1566                 for (nid = 0; nid < nr_node_ids; nid++)
1567                         root_unstable_tree[nid] = RB_ROOT;
1568
1569                 spin_lock(&ksm_mmlist_lock);
1570                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1571                 ksm_scan.mm_slot = slot;
1572                 spin_unlock(&ksm_mmlist_lock);
1573                 /*
1574                  * Although we tested list_empty() above, a racing __ksm_exit
1575                  * of the last mm on the list may have removed it since then.
1576                  */
1577                 if (slot == &ksm_mm_head)
1578                         return NULL;
1579 next_mm:
1580                 ksm_scan.address = 0;
1581                 ksm_scan.rmap_list = &slot->rmap_list;
1582         }
1583
1584         mm = slot->mm;
1585         down_read(&mm->mmap_sem);
1586         if (ksm_test_exit(mm))
1587                 vma = NULL;
1588         else
1589                 vma = find_vma(mm, ksm_scan.address);
1590
1591         for (; vma; vma = vma->vm_next) {
1592                 if (!(vma->vm_flags & VM_MERGEABLE))
1593                         continue;
1594                 if (ksm_scan.address < vma->vm_start)
1595                         ksm_scan.address = vma->vm_start;
1596                 if (!vma->anon_vma)
1597                         ksm_scan.address = vma->vm_end;
1598
1599                 while (ksm_scan.address < vma->vm_end) {
1600                         if (ksm_test_exit(mm))
1601                                 break;
1602                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1603                         if (IS_ERR_OR_NULL(*page)) {
1604                                 ksm_scan.address += PAGE_SIZE;
1605                                 cond_resched();
1606                                 continue;
1607                         }
1608                         if (PageAnon(*page) ||
1609                             page_trans_compound_anon(*page)) {
1610                                 flush_anon_page(vma, *page, ksm_scan.address);
1611                                 flush_dcache_page(*page);
1612                                 rmap_item = get_next_rmap_item(slot,
1613                                         ksm_scan.rmap_list, ksm_scan.address);
1614                                 if (rmap_item) {
1615                                         ksm_scan.rmap_list =
1616                                                         &rmap_item->rmap_list;
1617                                         ksm_scan.address += PAGE_SIZE;
1618                                 } else
1619                                         put_page(*page);
1620                                 up_read(&mm->mmap_sem);
1621                                 return rmap_item;
1622                         }
1623                         put_page(*page);
1624                         ksm_scan.address += PAGE_SIZE;
1625                         cond_resched();
1626                 }
1627         }
1628
1629         if (ksm_test_exit(mm)) {
1630                 ksm_scan.address = 0;
1631                 ksm_scan.rmap_list = &slot->rmap_list;
1632         }
1633         /*
1634          * Nuke all the rmap_items that are above this current rmap:
1635          * because there were no VM_MERGEABLE vmas with such addresses.
1636          */
1637         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1638
1639         spin_lock(&ksm_mmlist_lock);
1640         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1641                                                 struct mm_slot, mm_list);
1642         if (ksm_scan.address == 0) {
1643                 /*
1644                  * We've completed a full scan of all vmas, holding mmap_sem
1645                  * throughout, and found no VM_MERGEABLE: so do the same as
1646                  * __ksm_exit does to remove this mm from all our lists now.
1647                  * This applies either when cleaning up after __ksm_exit
1648                  * (but beware: we can reach here even before __ksm_exit),
1649                  * or when all VM_MERGEABLE areas have been unmapped (and
1650                  * mmap_sem then protects against race with MADV_MERGEABLE).
1651                  */
1652                 hash_del(&slot->link);
1653                 list_del(&slot->mm_list);
1654                 spin_unlock(&ksm_mmlist_lock);
1655
1656                 free_mm_slot(slot);
1657                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1658                 up_read(&mm->mmap_sem);
1659                 mmdrop(mm);
1660         } else {
1661                 spin_unlock(&ksm_mmlist_lock);
1662                 up_read(&mm->mmap_sem);
1663         }
1664
1665         /* Repeat until we've completed scanning the whole list */
1666         slot = ksm_scan.mm_slot;
1667         if (slot != &ksm_mm_head)
1668                 goto next_mm;
1669
1670         ksm_scan.seqnr++;
1671         return NULL;
1672 }
1673
1674 /**
1675  * ksm_do_scan  - the ksm scanner main worker function.
1676  * @scan_npages - number of pages we want to scan before we return.
1677  */
1678 static void ksm_do_scan(unsigned int scan_npages)
1679 {
1680         struct rmap_item *rmap_item;
1681         struct page *uninitialized_var(page);
1682
1683         while (scan_npages-- && likely(!freezing(current))) {
1684                 cond_resched();
1685                 rmap_item = scan_get_next_rmap_item(&page);
1686                 if (!rmap_item)
1687                         return;
1688                 cmp_and_merge_page(page, rmap_item);
1689                 put_page(page);
1690         }
1691 }
1692
1693 static int ksmd_should_run(void)
1694 {
1695         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1696 }
1697
1698 static int ksm_scan_thread(void *nothing)
1699 {
1700         set_freezable();
1701         set_user_nice(current, 5);
1702
1703         while (!kthread_should_stop()) {
1704                 mutex_lock(&ksm_thread_mutex);
1705                 wait_while_offlining();
1706                 if (ksmd_should_run())
1707                         ksm_do_scan(ksm_thread_pages_to_scan);
1708                 mutex_unlock(&ksm_thread_mutex);
1709
1710                 try_to_freeze();
1711
1712                 if (ksmd_should_run()) {
1713                         schedule_timeout_interruptible(
1714                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1715                 } else {
1716                         wait_event_freezable(ksm_thread_wait,
1717                                 ksmd_should_run() || kthread_should_stop());
1718                 }
1719         }
1720         return 0;
1721 }
1722
1723 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1724                 unsigned long end, int advice, unsigned long *vm_flags)
1725 {
1726         struct mm_struct *mm = vma->vm_mm;
1727         int err;
1728
1729         switch (advice) {
1730         case MADV_MERGEABLE:
1731                 /*
1732                  * Be somewhat over-protective for now!
1733                  */
1734                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1735                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1736                                  VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1737                         return 0;               /* just ignore the advice */
1738
1739 #ifdef VM_SAO
1740                 if (*vm_flags & VM_SAO)
1741                         return 0;
1742 #endif
1743
1744                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1745                         err = __ksm_enter(mm);
1746                         if (err)
1747                                 return err;
1748                 }
1749
1750                 *vm_flags |= VM_MERGEABLE;
1751                 break;
1752
1753         case MADV_UNMERGEABLE:
1754                 if (!(*vm_flags & VM_MERGEABLE))
1755                         return 0;               /* just ignore the advice */
1756
1757                 if (vma->anon_vma) {
1758                         err = unmerge_ksm_pages(vma, start, end);
1759                         if (err)
1760                                 return err;
1761                 }
1762
1763                 *vm_flags &= ~VM_MERGEABLE;
1764                 break;
1765         }
1766
1767         return 0;
1768 }
1769
1770 int __ksm_enter(struct mm_struct *mm)
1771 {
1772         struct mm_slot *mm_slot;
1773         int needs_wakeup;
1774
1775         mm_slot = alloc_mm_slot();
1776         if (!mm_slot)
1777                 return -ENOMEM;
1778
1779         /* Check ksm_run too?  Would need tighter locking */
1780         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1781
1782         spin_lock(&ksm_mmlist_lock);
1783         insert_to_mm_slots_hash(mm, mm_slot);
1784         /*
1785          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1786          * insert just behind the scanning cursor, to let the area settle
1787          * down a little; when fork is followed by immediate exec, we don't
1788          * want ksmd to waste time setting up and tearing down an rmap_list.
1789          *
1790          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1791          * scanning cursor, otherwise KSM pages in newly forked mms will be
1792          * missed: then we might as well insert at the end of the list.
1793          */
1794         if (ksm_run & KSM_RUN_UNMERGE)
1795                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1796         else
1797                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1798         spin_unlock(&ksm_mmlist_lock);
1799
1800         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1801         atomic_inc(&mm->mm_count);
1802
1803         if (needs_wakeup)
1804                 wake_up_interruptible(&ksm_thread_wait);
1805
1806         return 0;
1807 }
1808
1809 void __ksm_exit(struct mm_struct *mm)
1810 {
1811         struct mm_slot *mm_slot;
1812         int easy_to_free = 0;
1813
1814         /*
1815          * This process is exiting: if it's straightforward (as is the
1816          * case when ksmd was never running), free mm_slot immediately.
1817          * But if it's at the cursor or has rmap_items linked to it, use
1818          * mmap_sem to synchronize with any break_cows before pagetables
1819          * are freed, and leave the mm_slot on the list for ksmd to free.
1820          * Beware: ksm may already have noticed it exiting and freed the slot.
1821          */
1822
1823         spin_lock(&ksm_mmlist_lock);
1824         mm_slot = get_mm_slot(mm);
1825         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1826                 if (!mm_slot->rmap_list) {
1827                         hash_del(&mm_slot->link);
1828                         list_del(&mm_slot->mm_list);
1829                         easy_to_free = 1;
1830                 } else {
1831                         list_move(&mm_slot->mm_list,
1832                                   &ksm_scan.mm_slot->mm_list);
1833                 }
1834         }
1835         spin_unlock(&ksm_mmlist_lock);
1836
1837         if (easy_to_free) {
1838                 free_mm_slot(mm_slot);
1839                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1840                 mmdrop(mm);
1841         } else if (mm_slot) {
1842                 down_write(&mm->mmap_sem);
1843                 up_write(&mm->mmap_sem);
1844         }
1845 }
1846
1847 struct page *ksm_might_need_to_copy(struct page *page,
1848                         struct vm_area_struct *vma, unsigned long address)
1849 {
1850         struct anon_vma *anon_vma = page_anon_vma(page);
1851         struct page *new_page;
1852
1853         if (PageKsm(page)) {
1854                 if (page_stable_node(page) &&
1855                     !(ksm_run & KSM_RUN_UNMERGE))
1856                         return page;    /* no need to copy it */
1857         } else if (!anon_vma) {
1858                 return page;            /* no need to copy it */
1859         } else if (anon_vma->root == vma->anon_vma->root &&
1860                  page->index == linear_page_index(vma, address)) {
1861                 return page;            /* still no need to copy it */
1862         }
1863         if (!PageUptodate(page))
1864                 return page;            /* let do_swap_page report the error */
1865
1866         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1867         if (new_page) {
1868                 copy_user_highpage(new_page, page, address, vma);
1869
1870                 SetPageDirty(new_page);
1871                 __SetPageUptodate(new_page);
1872                 __set_page_locked(new_page);
1873         }
1874
1875         return new_page;
1876 }
1877
1878 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1879                         unsigned long *vm_flags)
1880 {
1881         struct stable_node *stable_node;
1882         struct rmap_item *rmap_item;
1883         struct hlist_node *hlist;
1884         unsigned int mapcount = page_mapcount(page);
1885         int referenced = 0;
1886         int search_new_forks = 0;
1887
1888         VM_BUG_ON(!PageKsm(page));
1889         VM_BUG_ON(!PageLocked(page));
1890
1891         stable_node = page_stable_node(page);
1892         if (!stable_node)
1893                 return 0;
1894 again:
1895         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1896                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1897                 struct anon_vma_chain *vmac;
1898                 struct vm_area_struct *vma;
1899
1900                 anon_vma_lock_read(anon_vma);
1901                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1902                                                0, ULONG_MAX) {
1903                         vma = vmac->vma;
1904                         if (rmap_item->address < vma->vm_start ||
1905                             rmap_item->address >= vma->vm_end)
1906                                 continue;
1907                         /*
1908                          * Initially we examine only the vma which covers this
1909                          * rmap_item; but later, if there is still work to do,
1910                          * we examine covering vmas in other mms: in case they
1911                          * were forked from the original since ksmd passed.
1912                          */
1913                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1914                                 continue;
1915
1916                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1917                                 continue;
1918
1919                         referenced += page_referenced_one(page, vma,
1920                                 rmap_item->address, &mapcount, vm_flags);
1921                         if (!search_new_forks || !mapcount)
1922                                 break;
1923                 }
1924                 anon_vma_unlock_read(anon_vma);
1925                 if (!mapcount)
1926                         goto out;
1927         }
1928         if (!search_new_forks++)
1929                 goto again;
1930 out:
1931         return referenced;
1932 }
1933
1934 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1935 {
1936         struct stable_node *stable_node;
1937         struct hlist_node *hlist;
1938         struct rmap_item *rmap_item;
1939         int ret = SWAP_AGAIN;
1940         int search_new_forks = 0;
1941
1942         VM_BUG_ON(!PageKsm(page));
1943         VM_BUG_ON(!PageLocked(page));
1944
1945         stable_node = page_stable_node(page);
1946         if (!stable_node)
1947                 return SWAP_FAIL;
1948 again:
1949         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1950                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1951                 struct anon_vma_chain *vmac;
1952                 struct vm_area_struct *vma;
1953
1954                 anon_vma_lock_read(anon_vma);
1955                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1956                                                0, ULONG_MAX) {
1957                         vma = vmac->vma;
1958                         if (rmap_item->address < vma->vm_start ||
1959                             rmap_item->address >= vma->vm_end)
1960                                 continue;
1961                         /*
1962                          * Initially we examine only the vma which covers this
1963                          * rmap_item; but later, if there is still work to do,
1964                          * we examine covering vmas in other mms: in case they
1965                          * were forked from the original since ksmd passed.
1966                          */
1967                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1968                                 continue;
1969
1970                         ret = try_to_unmap_one(page, vma,
1971                                         rmap_item->address, flags);
1972                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1973                                 anon_vma_unlock_read(anon_vma);
1974                                 goto out;
1975                         }
1976                 }
1977                 anon_vma_unlock_read(anon_vma);
1978         }
1979         if (!search_new_forks++)
1980                 goto again;
1981 out:
1982         return ret;
1983 }
1984
1985 #ifdef CONFIG_MIGRATION
1986 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1987                   struct vm_area_struct *, unsigned long, void *), void *arg)
1988 {
1989         struct stable_node *stable_node;
1990         struct hlist_node *hlist;
1991         struct rmap_item *rmap_item;
1992         int ret = SWAP_AGAIN;
1993         int search_new_forks = 0;
1994
1995         VM_BUG_ON(!PageKsm(page));
1996         VM_BUG_ON(!PageLocked(page));
1997
1998         stable_node = page_stable_node(page);
1999         if (!stable_node)
2000                 return ret;
2001 again:
2002         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
2003                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2004                 struct anon_vma_chain *vmac;
2005                 struct vm_area_struct *vma;
2006
2007                 anon_vma_lock_read(anon_vma);
2008                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2009                                                0, ULONG_MAX) {
2010                         vma = vmac->vma;
2011                         if (rmap_item->address < vma->vm_start ||
2012                             rmap_item->address >= vma->vm_end)
2013                                 continue;
2014                         /*
2015                          * Initially we examine only the vma which covers this
2016                          * rmap_item; but later, if there is still work to do,
2017                          * we examine covering vmas in other mms: in case they
2018                          * were forked from the original since ksmd passed.
2019                          */
2020                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2021                                 continue;
2022
2023                         ret = rmap_one(page, vma, rmap_item->address, arg);
2024                         if (ret != SWAP_AGAIN) {
2025                                 anon_vma_unlock_read(anon_vma);
2026                                 goto out;
2027                         }
2028                 }
2029                 anon_vma_unlock_read(anon_vma);
2030         }
2031         if (!search_new_forks++)
2032                 goto again;
2033 out:
2034         return ret;
2035 }
2036
2037 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2038 {
2039         struct stable_node *stable_node;
2040
2041         VM_BUG_ON(!PageLocked(oldpage));
2042         VM_BUG_ON(!PageLocked(newpage));
2043         VM_BUG_ON(newpage->mapping != oldpage->mapping);
2044
2045         stable_node = page_stable_node(newpage);
2046         if (stable_node) {
2047                 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
2048                 stable_node->kpfn = page_to_pfn(newpage);
2049                 /*
2050                  * newpage->mapping was set in advance; now we need smp_wmb()
2051                  * to make sure that the new stable_node->kpfn is visible
2052                  * to get_ksm_page() before it can see that oldpage->mapping
2053                  * has gone stale (or that PageSwapCache has been cleared).
2054                  */
2055                 smp_wmb();
2056                 set_page_stable_node(oldpage, NULL);
2057         }
2058 }
2059 #endif /* CONFIG_MIGRATION */
2060
2061 #ifdef CONFIG_MEMORY_HOTREMOVE
2062 static int just_wait(void *word)
2063 {
2064         schedule();
2065         return 0;
2066 }
2067
2068 static void wait_while_offlining(void)
2069 {
2070         while (ksm_run & KSM_RUN_OFFLINE) {
2071                 mutex_unlock(&ksm_thread_mutex);
2072                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2073                                 just_wait, TASK_UNINTERRUPTIBLE);
2074                 mutex_lock(&ksm_thread_mutex);
2075         }
2076 }
2077
2078 static void ksm_check_stable_tree(unsigned long start_pfn,
2079                                   unsigned long end_pfn)
2080 {
2081         struct stable_node *stable_node;
2082         struct list_head *this, *next;
2083         struct rb_node *node;
2084         int nid;
2085
2086         for (nid = 0; nid < nr_node_ids; nid++) {
2087                 node = rb_first(&root_stable_tree[nid]);
2088                 while (node) {
2089                         stable_node = rb_entry(node, struct stable_node, node);
2090                         if (stable_node->kpfn >= start_pfn &&
2091                             stable_node->kpfn < end_pfn) {
2092                                 /*
2093                                  * Don't get_ksm_page, page has already gone:
2094                                  * which is why we keep kpfn instead of page*
2095                                  */
2096                                 remove_node_from_stable_tree(stable_node);
2097                                 node = rb_first(&root_stable_tree[nid]);
2098                         } else
2099                                 node = rb_next(node);
2100                         cond_resched();
2101                 }
2102         }
2103         list_for_each_safe(this, next, &migrate_nodes) {
2104                 stable_node = list_entry(this, struct stable_node, list);
2105                 if (stable_node->kpfn >= start_pfn &&
2106                     stable_node->kpfn < end_pfn)
2107                         remove_node_from_stable_tree(stable_node);
2108                 cond_resched();
2109         }
2110 }
2111
2112 static int ksm_memory_callback(struct notifier_block *self,
2113                                unsigned long action, void *arg)
2114 {
2115         struct memory_notify *mn = arg;
2116
2117         switch (action) {
2118         case MEM_GOING_OFFLINE:
2119                 /*
2120                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2121                  * and remove_all_stable_nodes() while memory is going offline:
2122                  * it is unsafe for them to touch the stable tree at this time.
2123                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2124                  * which do not need the ksm_thread_mutex are all safe.
2125                  */
2126                 mutex_lock(&ksm_thread_mutex);
2127                 ksm_run |= KSM_RUN_OFFLINE;
2128                 mutex_unlock(&ksm_thread_mutex);
2129                 break;
2130
2131         case MEM_OFFLINE:
2132                 /*
2133                  * Most of the work is done by page migration; but there might
2134                  * be a few stable_nodes left over, still pointing to struct
2135                  * pages which have been offlined: prune those from the tree,
2136                  * otherwise get_ksm_page() might later try to access a
2137                  * non-existent struct page.
2138                  */
2139                 ksm_check_stable_tree(mn->start_pfn,
2140                                       mn->start_pfn + mn->nr_pages);
2141                 /* fallthrough */
2142
2143         case MEM_CANCEL_OFFLINE:
2144                 mutex_lock(&ksm_thread_mutex);
2145                 ksm_run &= ~KSM_RUN_OFFLINE;
2146                 mutex_unlock(&ksm_thread_mutex);
2147
2148                 smp_mb();       /* wake_up_bit advises this */
2149                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2150                 break;
2151         }
2152         return NOTIFY_OK;
2153 }
2154 #else
2155 static void wait_while_offlining(void)
2156 {
2157 }
2158 #endif /* CONFIG_MEMORY_HOTREMOVE */
2159
2160 #ifdef CONFIG_SYSFS
2161 /*
2162  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2163  */
2164
2165 #define KSM_ATTR_RO(_name) \
2166         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2167 #define KSM_ATTR(_name) \
2168         static struct kobj_attribute _name##_attr = \
2169                 __ATTR(_name, 0644, _name##_show, _name##_store)
2170
2171 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2172                                     struct kobj_attribute *attr, char *buf)
2173 {
2174         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2175 }
2176
2177 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2178                                      struct kobj_attribute *attr,
2179                                      const char *buf, size_t count)
2180 {
2181         unsigned long msecs;
2182         int err;
2183
2184         err = strict_strtoul(buf, 10, &msecs);
2185         if (err || msecs > UINT_MAX)
2186                 return -EINVAL;
2187
2188         ksm_thread_sleep_millisecs = msecs;
2189
2190         return count;
2191 }
2192 KSM_ATTR(sleep_millisecs);
2193
2194 static ssize_t pages_to_scan_show(struct kobject *kobj,
2195                                   struct kobj_attribute *attr, char *buf)
2196 {
2197         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2198 }
2199
2200 static ssize_t pages_to_scan_store(struct kobject *kobj,
2201                                    struct kobj_attribute *attr,
2202                                    const char *buf, size_t count)
2203 {
2204         int err;
2205         unsigned long nr_pages;
2206
2207         err = strict_strtoul(buf, 10, &nr_pages);
2208         if (err || nr_pages > UINT_MAX)
2209                 return -EINVAL;
2210
2211         ksm_thread_pages_to_scan = nr_pages;
2212
2213         return count;
2214 }
2215 KSM_ATTR(pages_to_scan);
2216
2217 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2218                         char *buf)
2219 {
2220         return sprintf(buf, "%lu\n", ksm_run);
2221 }
2222
2223 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2224                          const char *buf, size_t count)
2225 {
2226         int err;
2227         unsigned long flags;
2228
2229         err = strict_strtoul(buf, 10, &flags);
2230         if (err || flags > UINT_MAX)
2231                 return -EINVAL;
2232         if (flags > KSM_RUN_UNMERGE)
2233                 return -EINVAL;
2234
2235         /*
2236          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2237          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2238          * breaking COW to free the pages_shared (but leaves mm_slots
2239          * on the list for when ksmd may be set running again).
2240          */
2241
2242         mutex_lock(&ksm_thread_mutex);
2243         wait_while_offlining();
2244         if (ksm_run != flags) {
2245                 ksm_run = flags;
2246                 if (flags & KSM_RUN_UNMERGE) {
2247                         set_current_oom_origin();
2248                         err = unmerge_and_remove_all_rmap_items();
2249                         clear_current_oom_origin();
2250                         if (err) {
2251                                 ksm_run = KSM_RUN_STOP;
2252                                 count = err;
2253                         }
2254                 }
2255         }
2256         mutex_unlock(&ksm_thread_mutex);
2257
2258         if (flags & KSM_RUN_MERGE)
2259                 wake_up_interruptible(&ksm_thread_wait);
2260
2261         return count;
2262 }
2263 KSM_ATTR(run);
2264
2265 #ifdef CONFIG_NUMA
2266 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2267                                 struct kobj_attribute *attr, char *buf)
2268 {
2269         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2270 }
2271
2272 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2273                                    struct kobj_attribute *attr,
2274                                    const char *buf, size_t count)
2275 {
2276         int err;
2277         unsigned long knob;
2278
2279         err = kstrtoul(buf, 10, &knob);
2280         if (err)
2281                 return err;
2282         if (knob > 1)
2283                 return -EINVAL;
2284
2285         mutex_lock(&ksm_thread_mutex);
2286         wait_while_offlining();
2287         if (ksm_merge_across_nodes != knob) {
2288                 if (ksm_pages_shared || remove_all_stable_nodes())
2289                         err = -EBUSY;
2290                 else
2291                         ksm_merge_across_nodes = knob;
2292         }
2293         mutex_unlock(&ksm_thread_mutex);
2294
2295         return err ? err : count;
2296 }
2297 KSM_ATTR(merge_across_nodes);
2298 #endif
2299
2300 static ssize_t pages_shared_show(struct kobject *kobj,
2301                                  struct kobj_attribute *attr, char *buf)
2302 {
2303         return sprintf(buf, "%lu\n", ksm_pages_shared);
2304 }
2305 KSM_ATTR_RO(pages_shared);
2306
2307 static ssize_t pages_sharing_show(struct kobject *kobj,
2308                                   struct kobj_attribute *attr, char *buf)
2309 {
2310         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2311 }
2312 KSM_ATTR_RO(pages_sharing);
2313
2314 static ssize_t pages_unshared_show(struct kobject *kobj,
2315                                    struct kobj_attribute *attr, char *buf)
2316 {
2317         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2318 }
2319 KSM_ATTR_RO(pages_unshared);
2320
2321 static ssize_t pages_volatile_show(struct kobject *kobj,
2322                                    struct kobj_attribute *attr, char *buf)
2323 {
2324         long ksm_pages_volatile;
2325
2326         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2327                                 - ksm_pages_sharing - ksm_pages_unshared;
2328         /*
2329          * It was not worth any locking to calculate that statistic,
2330          * but it might therefore sometimes be negative: conceal that.
2331          */
2332         if (ksm_pages_volatile < 0)
2333                 ksm_pages_volatile = 0;
2334         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2335 }
2336 KSM_ATTR_RO(pages_volatile);
2337
2338 static ssize_t full_scans_show(struct kobject *kobj,
2339                                struct kobj_attribute *attr, char *buf)
2340 {
2341         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2342 }
2343 KSM_ATTR_RO(full_scans);
2344
2345 static struct attribute *ksm_attrs[] = {
2346         &sleep_millisecs_attr.attr,
2347         &pages_to_scan_attr.attr,
2348         &run_attr.attr,
2349         &pages_shared_attr.attr,
2350         &pages_sharing_attr.attr,
2351         &pages_unshared_attr.attr,
2352         &pages_volatile_attr.attr,
2353         &full_scans_attr.attr,
2354 #ifdef CONFIG_NUMA
2355         &merge_across_nodes_attr.attr,
2356 #endif
2357         NULL,
2358 };
2359
2360 static struct attribute_group ksm_attr_group = {
2361         .attrs = ksm_attrs,
2362         .name = "ksm",
2363 };
2364 #endif /* CONFIG_SYSFS */
2365
2366 static int __init ksm_init(void)
2367 {
2368         struct task_struct *ksm_thread;
2369         int err;
2370         int nid;
2371
2372         err = ksm_slab_init();
2373         if (err)
2374                 goto out;
2375
2376         for (nid = 0; nid < nr_node_ids; nid++)
2377                 root_stable_tree[nid] = RB_ROOT;
2378
2379         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2380         if (IS_ERR(ksm_thread)) {
2381                 printk(KERN_ERR "ksm: creating kthread failed\n");
2382                 err = PTR_ERR(ksm_thread);
2383                 goto out_free;
2384         }
2385
2386 #ifdef CONFIG_SYSFS
2387         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2388         if (err) {
2389                 printk(KERN_ERR "ksm: register sysfs failed\n");
2390                 kthread_stop(ksm_thread);
2391                 goto out_free;
2392         }
2393 #else
2394         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
2395
2396 #endif /* CONFIG_SYSFS */
2397
2398 #ifdef CONFIG_MEMORY_HOTREMOVE
2399         /* There is no significance to this priority 100 */
2400         hotplug_memory_notifier(ksm_memory_callback, 100);
2401 #endif
2402         return 0;
2403
2404 out_free:
2405         ksm_slab_free();
2406 out:
2407         return err;
2408 }
2409 module_init(ksm_init)