mm: extend KSM refcounts to the anon_vma root
[linux-3.10.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36
37 #include <asm/tlbflush.h>
38 #include "internal.h"
39
40 /*
41  * A few notes about the KSM scanning process,
42  * to make it easier to understand the data structures below:
43  *
44  * In order to reduce excessive scanning, KSM sorts the memory pages by their
45  * contents into a data structure that holds pointers to the pages' locations.
46  *
47  * Since the contents of the pages may change at any moment, KSM cannot just
48  * insert the pages into a normal sorted tree and expect it to find anything.
49  * Therefore KSM uses two data structures - the stable and the unstable tree.
50  *
51  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
52  * by their contents.  Because each such page is write-protected, searching on
53  * this tree is fully assured to be working (except when pages are unmapped),
54  * and therefore this tree is called the stable tree.
55  *
56  * In addition to the stable tree, KSM uses a second data structure called the
57  * unstable tree: this tree holds pointers to pages which have been found to
58  * be "unchanged for a period of time".  The unstable tree sorts these pages
59  * by their contents, but since they are not write-protected, KSM cannot rely
60  * upon the unstable tree to work correctly - the unstable tree is liable to
61  * be corrupted as its contents are modified, and so it is called unstable.
62  *
63  * KSM solves this problem by several techniques:
64  *
65  * 1) The unstable tree is flushed every time KSM completes scanning all
66  *    memory areas, and then the tree is rebuilt again from the beginning.
67  * 2) KSM will only insert into the unstable tree, pages whose hash value
68  *    has not changed since the previous scan of all memory areas.
69  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
70  *    colors of the nodes and not on their contents, assuring that even when
71  *    the tree gets "corrupted" it won't get out of balance, so scanning time
72  *    remains the same (also, searching and inserting nodes in an rbtree uses
73  *    the same algorithm, so we have no overhead when we flush and rebuild).
74  * 4) KSM never flushes the stable tree, which means that even if it were to
75  *    take 10 attempts to find a page in the unstable tree, once it is found,
76  *    it is secured in the stable tree.  (When we scan a new page, we first
77  *    compare it against the stable tree, and then against the unstable tree.)
78  */
79
80 /**
81  * struct mm_slot - ksm information per mm that is being scanned
82  * @link: link to the mm_slots hash list
83  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
84  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
85  * @mm: the mm that this information is valid for
86  */
87 struct mm_slot {
88         struct hlist_node link;
89         struct list_head mm_list;
90         struct rmap_item *rmap_list;
91         struct mm_struct *mm;
92 };
93
94 /**
95  * struct ksm_scan - cursor for scanning
96  * @mm_slot: the current mm_slot we are scanning
97  * @address: the next address inside that to be scanned
98  * @rmap_list: link to the next rmap to be scanned in the rmap_list
99  * @seqnr: count of completed full scans (needed when removing unstable node)
100  *
101  * There is only the one ksm_scan instance of this cursor structure.
102  */
103 struct ksm_scan {
104         struct mm_slot *mm_slot;
105         unsigned long address;
106         struct rmap_item **rmap_list;
107         unsigned long seqnr;
108 };
109
110 /**
111  * struct stable_node - node of the stable rbtree
112  * @node: rb node of this ksm page in the stable tree
113  * @hlist: hlist head of rmap_items using this ksm page
114  * @kpfn: page frame number of this ksm page
115  */
116 struct stable_node {
117         struct rb_node node;
118         struct hlist_head hlist;
119         unsigned long kpfn;
120 };
121
122 /**
123  * struct rmap_item - reverse mapping item for virtual addresses
124  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
125  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
126  * @mm: the memory structure this rmap_item is pointing into
127  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
128  * @oldchecksum: previous checksum of the page at that virtual address
129  * @node: rb node of this rmap_item in the unstable tree
130  * @head: pointer to stable_node heading this list in the stable tree
131  * @hlist: link into hlist of rmap_items hanging off that stable_node
132  */
133 struct rmap_item {
134         struct rmap_item *rmap_list;
135         struct anon_vma *anon_vma;      /* when stable */
136         struct mm_struct *mm;
137         unsigned long address;          /* + low bits used for flags below */
138         unsigned int oldchecksum;       /* when unstable */
139         union {
140                 struct rb_node node;    /* when node of unstable tree */
141                 struct {                /* when listed from stable tree */
142                         struct stable_node *head;
143                         struct hlist_node hlist;
144                 };
145         };
146 };
147
148 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
149 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
150 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
151
152 /* The stable and unstable tree heads */
153 static struct rb_root root_stable_tree = RB_ROOT;
154 static struct rb_root root_unstable_tree = RB_ROOT;
155
156 #define MM_SLOTS_HASH_HEADS 1024
157 static struct hlist_head *mm_slots_hash;
158
159 static struct mm_slot ksm_mm_head = {
160         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
161 };
162 static struct ksm_scan ksm_scan = {
163         .mm_slot = &ksm_mm_head,
164 };
165
166 static struct kmem_cache *rmap_item_cache;
167 static struct kmem_cache *stable_node_cache;
168 static struct kmem_cache *mm_slot_cache;
169
170 /* The number of nodes in the stable tree */
171 static unsigned long ksm_pages_shared;
172
173 /* The number of page slots additionally sharing those nodes */
174 static unsigned long ksm_pages_sharing;
175
176 /* The number of nodes in the unstable tree */
177 static unsigned long ksm_pages_unshared;
178
179 /* The number of rmap_items in use: to calculate pages_volatile */
180 static unsigned long ksm_rmap_items;
181
182 /* Number of pages ksmd should scan in one batch */
183 static unsigned int ksm_thread_pages_to_scan = 100;
184
185 /* Milliseconds ksmd should sleep between batches */
186 static unsigned int ksm_thread_sleep_millisecs = 20;
187
188 #define KSM_RUN_STOP    0
189 #define KSM_RUN_MERGE   1
190 #define KSM_RUN_UNMERGE 2
191 static unsigned int ksm_run = KSM_RUN_STOP;
192
193 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
194 static DEFINE_MUTEX(ksm_thread_mutex);
195 static DEFINE_SPINLOCK(ksm_mmlist_lock);
196
197 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
198                 sizeof(struct __struct), __alignof__(struct __struct),\
199                 (__flags), NULL)
200
201 static int __init ksm_slab_init(void)
202 {
203         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
204         if (!rmap_item_cache)
205                 goto out;
206
207         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
208         if (!stable_node_cache)
209                 goto out_free1;
210
211         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
212         if (!mm_slot_cache)
213                 goto out_free2;
214
215         return 0;
216
217 out_free2:
218         kmem_cache_destroy(stable_node_cache);
219 out_free1:
220         kmem_cache_destroy(rmap_item_cache);
221 out:
222         return -ENOMEM;
223 }
224
225 static void __init ksm_slab_free(void)
226 {
227         kmem_cache_destroy(mm_slot_cache);
228         kmem_cache_destroy(stable_node_cache);
229         kmem_cache_destroy(rmap_item_cache);
230         mm_slot_cache = NULL;
231 }
232
233 static inline struct rmap_item *alloc_rmap_item(void)
234 {
235         struct rmap_item *rmap_item;
236
237         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
238         if (rmap_item)
239                 ksm_rmap_items++;
240         return rmap_item;
241 }
242
243 static inline void free_rmap_item(struct rmap_item *rmap_item)
244 {
245         ksm_rmap_items--;
246         rmap_item->mm = NULL;   /* debug safety */
247         kmem_cache_free(rmap_item_cache, rmap_item);
248 }
249
250 static inline struct stable_node *alloc_stable_node(void)
251 {
252         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
253 }
254
255 static inline void free_stable_node(struct stable_node *stable_node)
256 {
257         kmem_cache_free(stable_node_cache, stable_node);
258 }
259
260 static inline struct mm_slot *alloc_mm_slot(void)
261 {
262         if (!mm_slot_cache)     /* initialization failed */
263                 return NULL;
264         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
265 }
266
267 static inline void free_mm_slot(struct mm_slot *mm_slot)
268 {
269         kmem_cache_free(mm_slot_cache, mm_slot);
270 }
271
272 static int __init mm_slots_hash_init(void)
273 {
274         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
275                                 GFP_KERNEL);
276         if (!mm_slots_hash)
277                 return -ENOMEM;
278         return 0;
279 }
280
281 static void __init mm_slots_hash_free(void)
282 {
283         kfree(mm_slots_hash);
284 }
285
286 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
287 {
288         struct mm_slot *mm_slot;
289         struct hlist_head *bucket;
290         struct hlist_node *node;
291
292         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
293                                 % MM_SLOTS_HASH_HEADS];
294         hlist_for_each_entry(mm_slot, node, bucket, link) {
295                 if (mm == mm_slot->mm)
296                         return mm_slot;
297         }
298         return NULL;
299 }
300
301 static void insert_to_mm_slots_hash(struct mm_struct *mm,
302                                     struct mm_slot *mm_slot)
303 {
304         struct hlist_head *bucket;
305
306         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
307                                 % MM_SLOTS_HASH_HEADS];
308         mm_slot->mm = mm;
309         hlist_add_head(&mm_slot->link, bucket);
310 }
311
312 static inline int in_stable_tree(struct rmap_item *rmap_item)
313 {
314         return rmap_item->address & STABLE_FLAG;
315 }
316
317 static void hold_anon_vma(struct rmap_item *rmap_item,
318                           struct anon_vma *anon_vma)
319 {
320         rmap_item->anon_vma = anon_vma;
321         get_anon_vma(anon_vma);
322 }
323
324 static void ksm_drop_anon_vma(struct rmap_item *rmap_item)
325 {
326         struct anon_vma *anon_vma = rmap_item->anon_vma;
327
328         drop_anon_vma(anon_vma);
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 void break_cow(struct rmap_item *rmap_item)
404 {
405         struct mm_struct *mm = rmap_item->mm;
406         unsigned long addr = rmap_item->address;
407         struct vm_area_struct *vma;
408
409         /*
410          * It is not an accident that whenever we want to break COW
411          * to undo, we also need to drop a reference to the anon_vma.
412          */
413         ksm_drop_anon_vma(rmap_item);
414
415         down_read(&mm->mmap_sem);
416         if (ksm_test_exit(mm))
417                 goto out;
418         vma = find_vma(mm, addr);
419         if (!vma || vma->vm_start > addr)
420                 goto out;
421         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
422                 goto out;
423         break_ksm(vma, addr);
424 out:
425         up_read(&mm->mmap_sem);
426 }
427
428 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
429 {
430         struct mm_struct *mm = rmap_item->mm;
431         unsigned long addr = rmap_item->address;
432         struct vm_area_struct *vma;
433         struct page *page;
434
435         down_read(&mm->mmap_sem);
436         if (ksm_test_exit(mm))
437                 goto out;
438         vma = find_vma(mm, addr);
439         if (!vma || vma->vm_start > addr)
440                 goto out;
441         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
442                 goto out;
443
444         page = follow_page(vma, addr, FOLL_GET);
445         if (IS_ERR_OR_NULL(page))
446                 goto out;
447         if (PageAnon(page)) {
448                 flush_anon_page(vma, page, addr);
449                 flush_dcache_page(page);
450         } else {
451                 put_page(page);
452 out:            page = NULL;
453         }
454         up_read(&mm->mmap_sem);
455         return page;
456 }
457
458 static void remove_node_from_stable_tree(struct stable_node *stable_node)
459 {
460         struct rmap_item *rmap_item;
461         struct hlist_node *hlist;
462
463         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
464                 if (rmap_item->hlist.next)
465                         ksm_pages_sharing--;
466                 else
467                         ksm_pages_shared--;
468                 ksm_drop_anon_vma(rmap_item);
469                 rmap_item->address &= PAGE_MASK;
470                 cond_resched();
471         }
472
473         rb_erase(&stable_node->node, &root_stable_tree);
474         free_stable_node(stable_node);
475 }
476
477 /*
478  * get_ksm_page: checks if the page indicated by the stable node
479  * is still its ksm page, despite having held no reference to it.
480  * In which case we can trust the content of the page, and it
481  * returns the gotten page; but if the page has now been zapped,
482  * remove the stale node from the stable tree and return NULL.
483  *
484  * You would expect the stable_node to hold a reference to the ksm page.
485  * But if it increments the page's count, swapping out has to wait for
486  * ksmd to come around again before it can free the page, which may take
487  * seconds or even minutes: much too unresponsive.  So instead we use a
488  * "keyhole reference": access to the ksm page from the stable node peeps
489  * out through its keyhole to see if that page still holds the right key,
490  * pointing back to this stable node.  This relies on freeing a PageAnon
491  * page to reset its page->mapping to NULL, and relies on no other use of
492  * a page to put something that might look like our key in page->mapping.
493  *
494  * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
495  * but this is different - made simpler by ksm_thread_mutex being held, but
496  * interesting for assuming that no other use of the struct page could ever
497  * put our expected_mapping into page->mapping (or a field of the union which
498  * coincides with page->mapping).  The RCU calls are not for KSM at all, but
499  * to keep the page_count protocol described with page_cache_get_speculative.
500  *
501  * Note: it is possible that get_ksm_page() will return NULL one moment,
502  * then page the next, if the page is in between page_freeze_refs() and
503  * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
504  * is on its way to being freed; but it is an anomaly to bear in mind.
505  */
506 static struct page *get_ksm_page(struct stable_node *stable_node)
507 {
508         struct page *page;
509         void *expected_mapping;
510
511         page = pfn_to_page(stable_node->kpfn);
512         expected_mapping = (void *)stable_node +
513                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
514         rcu_read_lock();
515         if (page->mapping != expected_mapping)
516                 goto stale;
517         if (!get_page_unless_zero(page))
518                 goto stale;
519         if (page->mapping != expected_mapping) {
520                 put_page(page);
521                 goto stale;
522         }
523         rcu_read_unlock();
524         return page;
525 stale:
526         rcu_read_unlock();
527         remove_node_from_stable_tree(stable_node);
528         return NULL;
529 }
530
531 /*
532  * Removing rmap_item from stable or unstable tree.
533  * This function will clean the information from the stable/unstable tree.
534  */
535 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
536 {
537         if (rmap_item->address & STABLE_FLAG) {
538                 struct stable_node *stable_node;
539                 struct page *page;
540
541                 stable_node = rmap_item->head;
542                 page = get_ksm_page(stable_node);
543                 if (!page)
544                         goto out;
545
546                 lock_page(page);
547                 hlist_del(&rmap_item->hlist);
548                 unlock_page(page);
549                 put_page(page);
550
551                 if (stable_node->hlist.first)
552                         ksm_pages_sharing--;
553                 else
554                         ksm_pages_shared--;
555
556                 ksm_drop_anon_vma(rmap_item);
557                 rmap_item->address &= PAGE_MASK;
558
559         } else if (rmap_item->address & UNSTABLE_FLAG) {
560                 unsigned char age;
561                 /*
562                  * Usually ksmd can and must skip the rb_erase, because
563                  * root_unstable_tree was already reset to RB_ROOT.
564                  * But be careful when an mm is exiting: do the rb_erase
565                  * if this rmap_item was inserted by this scan, rather
566                  * than left over from before.
567                  */
568                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
569                 BUG_ON(age > 1);
570                 if (!age)
571                         rb_erase(&rmap_item->node, &root_unstable_tree);
572
573                 ksm_pages_unshared--;
574                 rmap_item->address &= PAGE_MASK;
575         }
576 out:
577         cond_resched();         /* we're called from many long loops */
578 }
579
580 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
581                                        struct rmap_item **rmap_list)
582 {
583         while (*rmap_list) {
584                 struct rmap_item *rmap_item = *rmap_list;
585                 *rmap_list = rmap_item->rmap_list;
586                 remove_rmap_item_from_tree(rmap_item);
587                 free_rmap_item(rmap_item);
588         }
589 }
590
591 /*
592  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
593  * than check every pte of a given vma, the locking doesn't quite work for
594  * that - an rmap_item is assigned to the stable tree after inserting ksm
595  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
596  * rmap_items from parent to child at fork time (so as not to waste time
597  * if exit comes before the next scan reaches it).
598  *
599  * Similarly, although we'd like to remove rmap_items (so updating counts
600  * and freeing memory) when unmerging an area, it's easier to leave that
601  * to the next pass of ksmd - consider, for example, how ksmd might be
602  * in cmp_and_merge_page on one of the rmap_items we would be removing.
603  */
604 static int unmerge_ksm_pages(struct vm_area_struct *vma,
605                              unsigned long start, unsigned long end)
606 {
607         unsigned long addr;
608         int err = 0;
609
610         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
611                 if (ksm_test_exit(vma->vm_mm))
612                         break;
613                 if (signal_pending(current))
614                         err = -ERESTARTSYS;
615                 else
616                         err = break_ksm(vma, addr);
617         }
618         return err;
619 }
620
621 #ifdef CONFIG_SYSFS
622 /*
623  * Only called through the sysfs control interface:
624  */
625 static int unmerge_and_remove_all_rmap_items(void)
626 {
627         struct mm_slot *mm_slot;
628         struct mm_struct *mm;
629         struct vm_area_struct *vma;
630         int err = 0;
631
632         spin_lock(&ksm_mmlist_lock);
633         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
634                                                 struct mm_slot, mm_list);
635         spin_unlock(&ksm_mmlist_lock);
636
637         for (mm_slot = ksm_scan.mm_slot;
638                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
639                 mm = mm_slot->mm;
640                 down_read(&mm->mmap_sem);
641                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
642                         if (ksm_test_exit(mm))
643                                 break;
644                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
645                                 continue;
646                         err = unmerge_ksm_pages(vma,
647                                                 vma->vm_start, vma->vm_end);
648                         if (err)
649                                 goto error;
650                 }
651
652                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
653
654                 spin_lock(&ksm_mmlist_lock);
655                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
656                                                 struct mm_slot, mm_list);
657                 if (ksm_test_exit(mm)) {
658                         hlist_del(&mm_slot->link);
659                         list_del(&mm_slot->mm_list);
660                         spin_unlock(&ksm_mmlist_lock);
661
662                         free_mm_slot(mm_slot);
663                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
664                         up_read(&mm->mmap_sem);
665                         mmdrop(mm);
666                 } else {
667                         spin_unlock(&ksm_mmlist_lock);
668                         up_read(&mm->mmap_sem);
669                 }
670         }
671
672         ksm_scan.seqnr = 0;
673         return 0;
674
675 error:
676         up_read(&mm->mmap_sem);
677         spin_lock(&ksm_mmlist_lock);
678         ksm_scan.mm_slot = &ksm_mm_head;
679         spin_unlock(&ksm_mmlist_lock);
680         return err;
681 }
682 #endif /* CONFIG_SYSFS */
683
684 static u32 calc_checksum(struct page *page)
685 {
686         u32 checksum;
687         void *addr = kmap_atomic(page, KM_USER0);
688         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
689         kunmap_atomic(addr, KM_USER0);
690         return checksum;
691 }
692
693 static int memcmp_pages(struct page *page1, struct page *page2)
694 {
695         char *addr1, *addr2;
696         int ret;
697
698         addr1 = kmap_atomic(page1, KM_USER0);
699         addr2 = kmap_atomic(page2, KM_USER1);
700         ret = memcmp(addr1, addr2, PAGE_SIZE);
701         kunmap_atomic(addr2, KM_USER1);
702         kunmap_atomic(addr1, KM_USER0);
703         return ret;
704 }
705
706 static inline int pages_identical(struct page *page1, struct page *page2)
707 {
708         return !memcmp_pages(page1, page2);
709 }
710
711 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
712                               pte_t *orig_pte)
713 {
714         struct mm_struct *mm = vma->vm_mm;
715         unsigned long addr;
716         pte_t *ptep;
717         spinlock_t *ptl;
718         int swapped;
719         int err = -EFAULT;
720
721         addr = page_address_in_vma(page, vma);
722         if (addr == -EFAULT)
723                 goto out;
724
725         ptep = page_check_address(page, mm, addr, &ptl, 0);
726         if (!ptep)
727                 goto out;
728
729         if (pte_write(*ptep)) {
730                 pte_t entry;
731
732                 swapped = PageSwapCache(page);
733                 flush_cache_page(vma, addr, page_to_pfn(page));
734                 /*
735                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
736                  * take any lock, therefore the check that we are going to make
737                  * with the pagecount against the mapcount is racey and
738                  * O_DIRECT can happen right after the check.
739                  * So we clear the pte and flush the tlb before the check
740                  * this assure us that no O_DIRECT can happen after the check
741                  * or in the middle of the check.
742                  */
743                 entry = ptep_clear_flush(vma, addr, ptep);
744                 /*
745                  * Check that no O_DIRECT or similar I/O is in progress on the
746                  * page
747                  */
748                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
749                         set_pte_at(mm, addr, ptep, entry);
750                         goto out_unlock;
751                 }
752                 entry = pte_wrprotect(entry);
753                 set_pte_at_notify(mm, addr, ptep, entry);
754         }
755         *orig_pte = *ptep;
756         err = 0;
757
758 out_unlock:
759         pte_unmap_unlock(ptep, ptl);
760 out:
761         return err;
762 }
763
764 /**
765  * replace_page - replace page in vma by new ksm page
766  * @vma:      vma that holds the pte pointing to page
767  * @page:     the page we are replacing by kpage
768  * @kpage:    the ksm page we replace page by
769  * @orig_pte: the original value of the pte
770  *
771  * Returns 0 on success, -EFAULT on failure.
772  */
773 static int replace_page(struct vm_area_struct *vma, struct page *page,
774                         struct page *kpage, pte_t orig_pte)
775 {
776         struct mm_struct *mm = vma->vm_mm;
777         pgd_t *pgd;
778         pud_t *pud;
779         pmd_t *pmd;
780         pte_t *ptep;
781         spinlock_t *ptl;
782         unsigned long addr;
783         int err = -EFAULT;
784
785         addr = page_address_in_vma(page, vma);
786         if (addr == -EFAULT)
787                 goto out;
788
789         pgd = pgd_offset(mm, addr);
790         if (!pgd_present(*pgd))
791                 goto out;
792
793         pud = pud_offset(pgd, addr);
794         if (!pud_present(*pud))
795                 goto out;
796
797         pmd = pmd_offset(pud, addr);
798         if (!pmd_present(*pmd))
799                 goto out;
800
801         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
802         if (!pte_same(*ptep, orig_pte)) {
803                 pte_unmap_unlock(ptep, ptl);
804                 goto out;
805         }
806
807         get_page(kpage);
808         page_add_anon_rmap(kpage, vma, addr);
809
810         flush_cache_page(vma, addr, pte_pfn(*ptep));
811         ptep_clear_flush(vma, addr, ptep);
812         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
813
814         page_remove_rmap(page);
815         put_page(page);
816
817         pte_unmap_unlock(ptep, ptl);
818         err = 0;
819 out:
820         return err;
821 }
822
823 /*
824  * try_to_merge_one_page - take two pages and merge them into one
825  * @vma: the vma that holds the pte pointing to page
826  * @page: the PageAnon page that we want to replace with kpage
827  * @kpage: the PageKsm page that we want to map instead of page,
828  *         or NULL the first time when we want to use page as kpage.
829  *
830  * This function returns 0 if the pages were merged, -EFAULT otherwise.
831  */
832 static int try_to_merge_one_page(struct vm_area_struct *vma,
833                                  struct page *page, struct page *kpage)
834 {
835         pte_t orig_pte = __pte(0);
836         int err = -EFAULT;
837
838         if (page == kpage)                      /* ksm page forked */
839                 return 0;
840
841         if (!(vma->vm_flags & VM_MERGEABLE))
842                 goto out;
843         if (!PageAnon(page))
844                 goto out;
845
846         /*
847          * We need the page lock to read a stable PageSwapCache in
848          * write_protect_page().  We use trylock_page() instead of
849          * lock_page() because we don't want to wait here - we
850          * prefer to continue scanning and merging different pages,
851          * then come back to this page when it is unlocked.
852          */
853         if (!trylock_page(page))
854                 goto out;
855         /*
856          * If this anonymous page is mapped only here, its pte may need
857          * to be write-protected.  If it's mapped elsewhere, all of its
858          * ptes are necessarily already write-protected.  But in either
859          * case, we need to lock and check page_count is not raised.
860          */
861         if (write_protect_page(vma, page, &orig_pte) == 0) {
862                 if (!kpage) {
863                         /*
864                          * While we hold page lock, upgrade page from
865                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
866                          * stable_tree_insert() will update stable_node.
867                          */
868                         set_page_stable_node(page, NULL);
869                         mark_page_accessed(page);
870                         err = 0;
871                 } else if (pages_identical(page, kpage))
872                         err = replace_page(vma, page, kpage, orig_pte);
873         }
874
875         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
876                 munlock_vma_page(page);
877                 if (!PageMlocked(kpage)) {
878                         unlock_page(page);
879                         lock_page(kpage);
880                         mlock_vma_page(kpage);
881                         page = kpage;           /* for final unlock */
882                 }
883         }
884
885         unlock_page(page);
886 out:
887         return err;
888 }
889
890 /*
891  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
892  * but no new kernel page is allocated: kpage must already be a ksm page.
893  *
894  * This function returns 0 if the pages were merged, -EFAULT otherwise.
895  */
896 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
897                                       struct page *page, struct page *kpage)
898 {
899         struct mm_struct *mm = rmap_item->mm;
900         struct vm_area_struct *vma;
901         int err = -EFAULT;
902
903         down_read(&mm->mmap_sem);
904         if (ksm_test_exit(mm))
905                 goto out;
906         vma = find_vma(mm, rmap_item->address);
907         if (!vma || vma->vm_start > rmap_item->address)
908                 goto out;
909
910         err = try_to_merge_one_page(vma, page, kpage);
911         if (err)
912                 goto out;
913
914         /* Must get reference to anon_vma while still holding mmap_sem */
915         hold_anon_vma(rmap_item, vma->anon_vma);
916 out:
917         up_read(&mm->mmap_sem);
918         return err;
919 }
920
921 /*
922  * try_to_merge_two_pages - take two identical pages and prepare them
923  * to be merged into one page.
924  *
925  * This function returns the kpage if we successfully merged two identical
926  * pages into one ksm page, NULL otherwise.
927  *
928  * Note that this function upgrades page to ksm page: if one of the pages
929  * is already a ksm page, try_to_merge_with_ksm_page should be used.
930  */
931 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
932                                            struct page *page,
933                                            struct rmap_item *tree_rmap_item,
934                                            struct page *tree_page)
935 {
936         int err;
937
938         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
939         if (!err) {
940                 err = try_to_merge_with_ksm_page(tree_rmap_item,
941                                                         tree_page, page);
942                 /*
943                  * If that fails, we have a ksm page with only one pte
944                  * pointing to it: so break it.
945                  */
946                 if (err)
947                         break_cow(rmap_item);
948         }
949         return err ? NULL : page;
950 }
951
952 /*
953  * stable_tree_search - search for page inside the stable tree
954  *
955  * This function checks if there is a page inside the stable tree
956  * with identical content to the page that we are scanning right now.
957  *
958  * This function returns the stable tree node of identical content if found,
959  * NULL otherwise.
960  */
961 static struct page *stable_tree_search(struct page *page)
962 {
963         struct rb_node *node = root_stable_tree.rb_node;
964         struct stable_node *stable_node;
965
966         stable_node = page_stable_node(page);
967         if (stable_node) {                      /* ksm page forked */
968                 get_page(page);
969                 return page;
970         }
971
972         while (node) {
973                 struct page *tree_page;
974                 int ret;
975
976                 cond_resched();
977                 stable_node = rb_entry(node, struct stable_node, node);
978                 tree_page = get_ksm_page(stable_node);
979                 if (!tree_page)
980                         return NULL;
981
982                 ret = memcmp_pages(page, tree_page);
983
984                 if (ret < 0) {
985                         put_page(tree_page);
986                         node = node->rb_left;
987                 } else if (ret > 0) {
988                         put_page(tree_page);
989                         node = node->rb_right;
990                 } else
991                         return tree_page;
992         }
993
994         return NULL;
995 }
996
997 /*
998  * stable_tree_insert - insert rmap_item pointing to new ksm page
999  * into the stable tree.
1000  *
1001  * This function returns the stable tree node just allocated on success,
1002  * NULL otherwise.
1003  */
1004 static struct stable_node *stable_tree_insert(struct page *kpage)
1005 {
1006         struct rb_node **new = &root_stable_tree.rb_node;
1007         struct rb_node *parent = NULL;
1008         struct stable_node *stable_node;
1009
1010         while (*new) {
1011                 struct page *tree_page;
1012                 int ret;
1013
1014                 cond_resched();
1015                 stable_node = rb_entry(*new, struct stable_node, node);
1016                 tree_page = get_ksm_page(stable_node);
1017                 if (!tree_page)
1018                         return NULL;
1019
1020                 ret = memcmp_pages(kpage, tree_page);
1021                 put_page(tree_page);
1022
1023                 parent = *new;
1024                 if (ret < 0)
1025                         new = &parent->rb_left;
1026                 else if (ret > 0)
1027                         new = &parent->rb_right;
1028                 else {
1029                         /*
1030                          * It is not a bug that stable_tree_search() didn't
1031                          * find this node: because at that time our page was
1032                          * not yet write-protected, so may have changed since.
1033                          */
1034                         return NULL;
1035                 }
1036         }
1037
1038         stable_node = alloc_stable_node();
1039         if (!stable_node)
1040                 return NULL;
1041
1042         rb_link_node(&stable_node->node, parent, new);
1043         rb_insert_color(&stable_node->node, &root_stable_tree);
1044
1045         INIT_HLIST_HEAD(&stable_node->hlist);
1046
1047         stable_node->kpfn = page_to_pfn(kpage);
1048         set_page_stable_node(kpage, stable_node);
1049
1050         return stable_node;
1051 }
1052
1053 /*
1054  * unstable_tree_search_insert - search for identical page,
1055  * else insert rmap_item into the unstable tree.
1056  *
1057  * This function searches for a page in the unstable tree identical to the
1058  * page currently being scanned; and if no identical page is found in the
1059  * tree, we insert rmap_item as a new object into the unstable tree.
1060  *
1061  * This function returns pointer to rmap_item found to be identical
1062  * to the currently scanned page, NULL otherwise.
1063  *
1064  * This function does both searching and inserting, because they share
1065  * the same walking algorithm in an rbtree.
1066  */
1067 static
1068 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1069                                               struct page *page,
1070                                               struct page **tree_pagep)
1071
1072 {
1073         struct rb_node **new = &root_unstable_tree.rb_node;
1074         struct rb_node *parent = NULL;
1075
1076         while (*new) {
1077                 struct rmap_item *tree_rmap_item;
1078                 struct page *tree_page;
1079                 int ret;
1080
1081                 cond_resched();
1082                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1083                 tree_page = get_mergeable_page(tree_rmap_item);
1084                 if (IS_ERR_OR_NULL(tree_page))
1085                         return NULL;
1086
1087                 /*
1088                  * Don't substitute a ksm page for a forked page.
1089                  */
1090                 if (page == tree_page) {
1091                         put_page(tree_page);
1092                         return NULL;
1093                 }
1094
1095                 ret = memcmp_pages(page, tree_page);
1096
1097                 parent = *new;
1098                 if (ret < 0) {
1099                         put_page(tree_page);
1100                         new = &parent->rb_left;
1101                 } else if (ret > 0) {
1102                         put_page(tree_page);
1103                         new = &parent->rb_right;
1104                 } else {
1105                         *tree_pagep = tree_page;
1106                         return tree_rmap_item;
1107                 }
1108         }
1109
1110         rmap_item->address |= UNSTABLE_FLAG;
1111         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1112         rb_link_node(&rmap_item->node, parent, new);
1113         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1114
1115         ksm_pages_unshared++;
1116         return NULL;
1117 }
1118
1119 /*
1120  * stable_tree_append - add another rmap_item to the linked list of
1121  * rmap_items hanging off a given node of the stable tree, all sharing
1122  * the same ksm page.
1123  */
1124 static void stable_tree_append(struct rmap_item *rmap_item,
1125                                struct stable_node *stable_node)
1126 {
1127         rmap_item->head = stable_node;
1128         rmap_item->address |= STABLE_FLAG;
1129         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1130
1131         if (rmap_item->hlist.next)
1132                 ksm_pages_sharing++;
1133         else
1134                 ksm_pages_shared++;
1135 }
1136
1137 /*
1138  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1139  * if not, compare checksum to previous and if it's the same, see if page can
1140  * be inserted into the unstable tree, or merged with a page already there and
1141  * both transferred to the stable tree.
1142  *
1143  * @page: the page that we are searching identical page to.
1144  * @rmap_item: the reverse mapping into the virtual address of this page
1145  */
1146 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1147 {
1148         struct rmap_item *tree_rmap_item;
1149         struct page *tree_page = NULL;
1150         struct stable_node *stable_node;
1151         struct page *kpage;
1152         unsigned int checksum;
1153         int err;
1154
1155         remove_rmap_item_from_tree(rmap_item);
1156
1157         /* We first start with searching the page inside the stable tree */
1158         kpage = stable_tree_search(page);
1159         if (kpage) {
1160                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1161                 if (!err) {
1162                         /*
1163                          * The page was successfully merged:
1164                          * add its rmap_item to the stable tree.
1165                          */
1166                         lock_page(kpage);
1167                         stable_tree_append(rmap_item, page_stable_node(kpage));
1168                         unlock_page(kpage);
1169                 }
1170                 put_page(kpage);
1171                 return;
1172         }
1173
1174         /*
1175          * If the hash value of the page has changed from the last time
1176          * we calculated it, this page is changing frequently: therefore we
1177          * don't want to insert it in the unstable tree, and we don't want
1178          * to waste our time searching for something identical to it there.
1179          */
1180         checksum = calc_checksum(page);
1181         if (rmap_item->oldchecksum != checksum) {
1182                 rmap_item->oldchecksum = checksum;
1183                 return;
1184         }
1185
1186         tree_rmap_item =
1187                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1188         if (tree_rmap_item) {
1189                 kpage = try_to_merge_two_pages(rmap_item, page,
1190                                                 tree_rmap_item, tree_page);
1191                 put_page(tree_page);
1192                 /*
1193                  * As soon as we merge this page, we want to remove the
1194                  * rmap_item of the page we have merged with from the unstable
1195                  * tree, and insert it instead as new node in the stable tree.
1196                  */
1197                 if (kpage) {
1198                         remove_rmap_item_from_tree(tree_rmap_item);
1199
1200                         lock_page(kpage);
1201                         stable_node = stable_tree_insert(kpage);
1202                         if (stable_node) {
1203                                 stable_tree_append(tree_rmap_item, stable_node);
1204                                 stable_tree_append(rmap_item, stable_node);
1205                         }
1206                         unlock_page(kpage);
1207
1208                         /*
1209                          * If we fail to insert the page into the stable tree,
1210                          * we will have 2 virtual addresses that are pointing
1211                          * to a ksm page left outside the stable tree,
1212                          * in which case we need to break_cow on both.
1213                          */
1214                         if (!stable_node) {
1215                                 break_cow(tree_rmap_item);
1216                                 break_cow(rmap_item);
1217                         }
1218                 }
1219         }
1220 }
1221
1222 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1223                                             struct rmap_item **rmap_list,
1224                                             unsigned long addr)
1225 {
1226         struct rmap_item *rmap_item;
1227
1228         while (*rmap_list) {
1229                 rmap_item = *rmap_list;
1230                 if ((rmap_item->address & PAGE_MASK) == addr)
1231                         return rmap_item;
1232                 if (rmap_item->address > addr)
1233                         break;
1234                 *rmap_list = rmap_item->rmap_list;
1235                 remove_rmap_item_from_tree(rmap_item);
1236                 free_rmap_item(rmap_item);
1237         }
1238
1239         rmap_item = alloc_rmap_item();
1240         if (rmap_item) {
1241                 /* It has already been zeroed */
1242                 rmap_item->mm = mm_slot->mm;
1243                 rmap_item->address = addr;
1244                 rmap_item->rmap_list = *rmap_list;
1245                 *rmap_list = rmap_item;
1246         }
1247         return rmap_item;
1248 }
1249
1250 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1251 {
1252         struct mm_struct *mm;
1253         struct mm_slot *slot;
1254         struct vm_area_struct *vma;
1255         struct rmap_item *rmap_item;
1256
1257         if (list_empty(&ksm_mm_head.mm_list))
1258                 return NULL;
1259
1260         slot = ksm_scan.mm_slot;
1261         if (slot == &ksm_mm_head) {
1262                 root_unstable_tree = RB_ROOT;
1263
1264                 spin_lock(&ksm_mmlist_lock);
1265                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1266                 ksm_scan.mm_slot = slot;
1267                 spin_unlock(&ksm_mmlist_lock);
1268 next_mm:
1269                 ksm_scan.address = 0;
1270                 ksm_scan.rmap_list = &slot->rmap_list;
1271         }
1272
1273         mm = slot->mm;
1274         down_read(&mm->mmap_sem);
1275         if (ksm_test_exit(mm))
1276                 vma = NULL;
1277         else
1278                 vma = find_vma(mm, ksm_scan.address);
1279
1280         for (; vma; vma = vma->vm_next) {
1281                 if (!(vma->vm_flags & VM_MERGEABLE))
1282                         continue;
1283                 if (ksm_scan.address < vma->vm_start)
1284                         ksm_scan.address = vma->vm_start;
1285                 if (!vma->anon_vma)
1286                         ksm_scan.address = vma->vm_end;
1287
1288                 while (ksm_scan.address < vma->vm_end) {
1289                         if (ksm_test_exit(mm))
1290                                 break;
1291                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1292                         if (!IS_ERR_OR_NULL(*page) && PageAnon(*page)) {
1293                                 flush_anon_page(vma, *page, ksm_scan.address);
1294                                 flush_dcache_page(*page);
1295                                 rmap_item = get_next_rmap_item(slot,
1296                                         ksm_scan.rmap_list, ksm_scan.address);
1297                                 if (rmap_item) {
1298                                         ksm_scan.rmap_list =
1299                                                         &rmap_item->rmap_list;
1300                                         ksm_scan.address += PAGE_SIZE;
1301                                 } else
1302                                         put_page(*page);
1303                                 up_read(&mm->mmap_sem);
1304                                 return rmap_item;
1305                         }
1306                         if (!IS_ERR_OR_NULL(*page))
1307                                 put_page(*page);
1308                         ksm_scan.address += PAGE_SIZE;
1309                         cond_resched();
1310                 }
1311         }
1312
1313         if (ksm_test_exit(mm)) {
1314                 ksm_scan.address = 0;
1315                 ksm_scan.rmap_list = &slot->rmap_list;
1316         }
1317         /*
1318          * Nuke all the rmap_items that are above this current rmap:
1319          * because there were no VM_MERGEABLE vmas with such addresses.
1320          */
1321         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1322
1323         spin_lock(&ksm_mmlist_lock);
1324         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1325                                                 struct mm_slot, mm_list);
1326         if (ksm_scan.address == 0) {
1327                 /*
1328                  * We've completed a full scan of all vmas, holding mmap_sem
1329                  * throughout, and found no VM_MERGEABLE: so do the same as
1330                  * __ksm_exit does to remove this mm from all our lists now.
1331                  * This applies either when cleaning up after __ksm_exit
1332                  * (but beware: we can reach here even before __ksm_exit),
1333                  * or when all VM_MERGEABLE areas have been unmapped (and
1334                  * mmap_sem then protects against race with MADV_MERGEABLE).
1335                  */
1336                 hlist_del(&slot->link);
1337                 list_del(&slot->mm_list);
1338                 spin_unlock(&ksm_mmlist_lock);
1339
1340                 free_mm_slot(slot);
1341                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1342                 up_read(&mm->mmap_sem);
1343                 mmdrop(mm);
1344         } else {
1345                 spin_unlock(&ksm_mmlist_lock);
1346                 up_read(&mm->mmap_sem);
1347         }
1348
1349         /* Repeat until we've completed scanning the whole list */
1350         slot = ksm_scan.mm_slot;
1351         if (slot != &ksm_mm_head)
1352                 goto next_mm;
1353
1354         ksm_scan.seqnr++;
1355         return NULL;
1356 }
1357
1358 /**
1359  * ksm_do_scan  - the ksm scanner main worker function.
1360  * @scan_npages - number of pages we want to scan before we return.
1361  */
1362 static void ksm_do_scan(unsigned int scan_npages)
1363 {
1364         struct rmap_item *rmap_item;
1365         struct page *uninitialized_var(page);
1366
1367         while (scan_npages--) {
1368                 cond_resched();
1369                 rmap_item = scan_get_next_rmap_item(&page);
1370                 if (!rmap_item)
1371                         return;
1372                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1373                         cmp_and_merge_page(page, rmap_item);
1374                 put_page(page);
1375         }
1376 }
1377
1378 static int ksmd_should_run(void)
1379 {
1380         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1381 }
1382
1383 static int ksm_scan_thread(void *nothing)
1384 {
1385         set_user_nice(current, 5);
1386
1387         while (!kthread_should_stop()) {
1388                 mutex_lock(&ksm_thread_mutex);
1389                 if (ksmd_should_run())
1390                         ksm_do_scan(ksm_thread_pages_to_scan);
1391                 mutex_unlock(&ksm_thread_mutex);
1392
1393                 if (ksmd_should_run()) {
1394                         schedule_timeout_interruptible(
1395                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1396                 } else {
1397                         wait_event_interruptible(ksm_thread_wait,
1398                                 ksmd_should_run() || kthread_should_stop());
1399                 }
1400         }
1401         return 0;
1402 }
1403
1404 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1405                 unsigned long end, int advice, unsigned long *vm_flags)
1406 {
1407         struct mm_struct *mm = vma->vm_mm;
1408         int err;
1409
1410         switch (advice) {
1411         case MADV_MERGEABLE:
1412                 /*
1413                  * Be somewhat over-protective for now!
1414                  */
1415                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1416                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1417                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1418                                  VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1419                         return 0;               /* just ignore the advice */
1420
1421                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1422                         err = __ksm_enter(mm);
1423                         if (err)
1424                                 return err;
1425                 }
1426
1427                 *vm_flags |= VM_MERGEABLE;
1428                 break;
1429
1430         case MADV_UNMERGEABLE:
1431                 if (!(*vm_flags & VM_MERGEABLE))
1432                         return 0;               /* just ignore the advice */
1433
1434                 if (vma->anon_vma) {
1435                         err = unmerge_ksm_pages(vma, start, end);
1436                         if (err)
1437                                 return err;
1438                 }
1439
1440                 *vm_flags &= ~VM_MERGEABLE;
1441                 break;
1442         }
1443
1444         return 0;
1445 }
1446
1447 int __ksm_enter(struct mm_struct *mm)
1448 {
1449         struct mm_slot *mm_slot;
1450         int needs_wakeup;
1451
1452         mm_slot = alloc_mm_slot();
1453         if (!mm_slot)
1454                 return -ENOMEM;
1455
1456         /* Check ksm_run too?  Would need tighter locking */
1457         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1458
1459         spin_lock(&ksm_mmlist_lock);
1460         insert_to_mm_slots_hash(mm, mm_slot);
1461         /*
1462          * Insert just behind the scanning cursor, to let the area settle
1463          * down a little; when fork is followed by immediate exec, we don't
1464          * want ksmd to waste time setting up and tearing down an rmap_list.
1465          */
1466         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1467         spin_unlock(&ksm_mmlist_lock);
1468
1469         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1470         atomic_inc(&mm->mm_count);
1471
1472         if (needs_wakeup)
1473                 wake_up_interruptible(&ksm_thread_wait);
1474
1475         return 0;
1476 }
1477
1478 void __ksm_exit(struct mm_struct *mm)
1479 {
1480         struct mm_slot *mm_slot;
1481         int easy_to_free = 0;
1482
1483         /*
1484          * This process is exiting: if it's straightforward (as is the
1485          * case when ksmd was never running), free mm_slot immediately.
1486          * But if it's at the cursor or has rmap_items linked to it, use
1487          * mmap_sem to synchronize with any break_cows before pagetables
1488          * are freed, and leave the mm_slot on the list for ksmd to free.
1489          * Beware: ksm may already have noticed it exiting and freed the slot.
1490          */
1491
1492         spin_lock(&ksm_mmlist_lock);
1493         mm_slot = get_mm_slot(mm);
1494         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1495                 if (!mm_slot->rmap_list) {
1496                         hlist_del(&mm_slot->link);
1497                         list_del(&mm_slot->mm_list);
1498                         easy_to_free = 1;
1499                 } else {
1500                         list_move(&mm_slot->mm_list,
1501                                   &ksm_scan.mm_slot->mm_list);
1502                 }
1503         }
1504         spin_unlock(&ksm_mmlist_lock);
1505
1506         if (easy_to_free) {
1507                 free_mm_slot(mm_slot);
1508                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1509                 mmdrop(mm);
1510         } else if (mm_slot) {
1511                 down_write(&mm->mmap_sem);
1512                 up_write(&mm->mmap_sem);
1513         }
1514 }
1515
1516 struct page *ksm_does_need_to_copy(struct page *page,
1517                         struct vm_area_struct *vma, unsigned long address)
1518 {
1519         struct page *new_page;
1520
1521         unlock_page(page);      /* any racers will COW it, not modify it */
1522
1523         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1524         if (new_page) {
1525                 copy_user_highpage(new_page, page, address, vma);
1526
1527                 SetPageDirty(new_page);
1528                 __SetPageUptodate(new_page);
1529                 SetPageSwapBacked(new_page);
1530                 __set_page_locked(new_page);
1531
1532                 if (page_evictable(new_page, vma))
1533                         lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1534                 else
1535                         add_page_to_unevictable_list(new_page);
1536         }
1537
1538         page_cache_release(page);
1539         return new_page;
1540 }
1541
1542 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1543                         unsigned long *vm_flags)
1544 {
1545         struct stable_node *stable_node;
1546         struct rmap_item *rmap_item;
1547         struct hlist_node *hlist;
1548         unsigned int mapcount = page_mapcount(page);
1549         int referenced = 0;
1550         int search_new_forks = 0;
1551
1552         VM_BUG_ON(!PageKsm(page));
1553         VM_BUG_ON(!PageLocked(page));
1554
1555         stable_node = page_stable_node(page);
1556         if (!stable_node)
1557                 return 0;
1558 again:
1559         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1560                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1561                 struct anon_vma_chain *vmac;
1562                 struct vm_area_struct *vma;
1563
1564                 anon_vma_lock(anon_vma);
1565                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1566                         vma = vmac->vma;
1567                         if (rmap_item->address < vma->vm_start ||
1568                             rmap_item->address >= vma->vm_end)
1569                                 continue;
1570                         /*
1571                          * Initially we examine only the vma which covers this
1572                          * rmap_item; but later, if there is still work to do,
1573                          * we examine covering vmas in other mms: in case they
1574                          * were forked from the original since ksmd passed.
1575                          */
1576                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1577                                 continue;
1578
1579                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1580                                 continue;
1581
1582                         referenced += page_referenced_one(page, vma,
1583                                 rmap_item->address, &mapcount, vm_flags);
1584                         if (!search_new_forks || !mapcount)
1585                                 break;
1586                 }
1587                 anon_vma_unlock(anon_vma);
1588                 if (!mapcount)
1589                         goto out;
1590         }
1591         if (!search_new_forks++)
1592                 goto again;
1593 out:
1594         return referenced;
1595 }
1596
1597 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1598 {
1599         struct stable_node *stable_node;
1600         struct hlist_node *hlist;
1601         struct rmap_item *rmap_item;
1602         int ret = SWAP_AGAIN;
1603         int search_new_forks = 0;
1604
1605         VM_BUG_ON(!PageKsm(page));
1606         VM_BUG_ON(!PageLocked(page));
1607
1608         stable_node = page_stable_node(page);
1609         if (!stable_node)
1610                 return SWAP_FAIL;
1611 again:
1612         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1613                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1614                 struct anon_vma_chain *vmac;
1615                 struct vm_area_struct *vma;
1616
1617                 anon_vma_lock(anon_vma);
1618                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1619                         vma = vmac->vma;
1620                         if (rmap_item->address < vma->vm_start ||
1621                             rmap_item->address >= vma->vm_end)
1622                                 continue;
1623                         /*
1624                          * Initially we examine only the vma which covers this
1625                          * rmap_item; but later, if there is still work to do,
1626                          * we examine covering vmas in other mms: in case they
1627                          * were forked from the original since ksmd passed.
1628                          */
1629                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1630                                 continue;
1631
1632                         ret = try_to_unmap_one(page, vma,
1633                                         rmap_item->address, flags);
1634                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1635                                 anon_vma_unlock(anon_vma);
1636                                 goto out;
1637                         }
1638                 }
1639                 anon_vma_unlock(anon_vma);
1640         }
1641         if (!search_new_forks++)
1642                 goto again;
1643 out:
1644         return ret;
1645 }
1646
1647 #ifdef CONFIG_MIGRATION
1648 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1649                   struct vm_area_struct *, unsigned long, void *), void *arg)
1650 {
1651         struct stable_node *stable_node;
1652         struct hlist_node *hlist;
1653         struct rmap_item *rmap_item;
1654         int ret = SWAP_AGAIN;
1655         int search_new_forks = 0;
1656
1657         VM_BUG_ON(!PageKsm(page));
1658         VM_BUG_ON(!PageLocked(page));
1659
1660         stable_node = page_stable_node(page);
1661         if (!stable_node)
1662                 return ret;
1663 again:
1664         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1665                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1666                 struct anon_vma_chain *vmac;
1667                 struct vm_area_struct *vma;
1668
1669                 anon_vma_lock(anon_vma);
1670                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1671                         vma = vmac->vma;
1672                         if (rmap_item->address < vma->vm_start ||
1673                             rmap_item->address >= vma->vm_end)
1674                                 continue;
1675                         /*
1676                          * Initially we examine only the vma which covers this
1677                          * rmap_item; but later, if there is still work to do,
1678                          * we examine covering vmas in other mms: in case they
1679                          * were forked from the original since ksmd passed.
1680                          */
1681                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1682                                 continue;
1683
1684                         ret = rmap_one(page, vma, rmap_item->address, arg);
1685                         if (ret != SWAP_AGAIN) {
1686                                 anon_vma_unlock(anon_vma);
1687                                 goto out;
1688                         }
1689                 }
1690                 anon_vma_unlock(anon_vma);
1691         }
1692         if (!search_new_forks++)
1693                 goto again;
1694 out:
1695         return ret;
1696 }
1697
1698 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1699 {
1700         struct stable_node *stable_node;
1701
1702         VM_BUG_ON(!PageLocked(oldpage));
1703         VM_BUG_ON(!PageLocked(newpage));
1704         VM_BUG_ON(newpage->mapping != oldpage->mapping);
1705
1706         stable_node = page_stable_node(newpage);
1707         if (stable_node) {
1708                 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1709                 stable_node->kpfn = page_to_pfn(newpage);
1710         }
1711 }
1712 #endif /* CONFIG_MIGRATION */
1713
1714 #ifdef CONFIG_MEMORY_HOTREMOVE
1715 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1716                                                  unsigned long end_pfn)
1717 {
1718         struct rb_node *node;
1719
1720         for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1721                 struct stable_node *stable_node;
1722
1723                 stable_node = rb_entry(node, struct stable_node, node);
1724                 if (stable_node->kpfn >= start_pfn &&
1725                     stable_node->kpfn < end_pfn)
1726                         return stable_node;
1727         }
1728         return NULL;
1729 }
1730
1731 static int ksm_memory_callback(struct notifier_block *self,
1732                                unsigned long action, void *arg)
1733 {
1734         struct memory_notify *mn = arg;
1735         struct stable_node *stable_node;
1736
1737         switch (action) {
1738         case MEM_GOING_OFFLINE:
1739                 /*
1740                  * Keep it very simple for now: just lock out ksmd and
1741                  * MADV_UNMERGEABLE while any memory is going offline.
1742                  */
1743                 mutex_lock(&ksm_thread_mutex);
1744                 break;
1745
1746         case MEM_OFFLINE:
1747                 /*
1748                  * Most of the work is done by page migration; but there might
1749                  * be a few stable_nodes left over, still pointing to struct
1750                  * pages which have been offlined: prune those from the tree.
1751                  */
1752                 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1753                                         mn->start_pfn + mn->nr_pages)) != NULL)
1754                         remove_node_from_stable_tree(stable_node);
1755                 /* fallthrough */
1756
1757         case MEM_CANCEL_OFFLINE:
1758                 mutex_unlock(&ksm_thread_mutex);
1759                 break;
1760         }
1761         return NOTIFY_OK;
1762 }
1763 #endif /* CONFIG_MEMORY_HOTREMOVE */
1764
1765 #ifdef CONFIG_SYSFS
1766 /*
1767  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1768  */
1769
1770 #define KSM_ATTR_RO(_name) \
1771         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1772 #define KSM_ATTR(_name) \
1773         static struct kobj_attribute _name##_attr = \
1774                 __ATTR(_name, 0644, _name##_show, _name##_store)
1775
1776 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1777                                     struct kobj_attribute *attr, char *buf)
1778 {
1779         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1780 }
1781
1782 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1783                                      struct kobj_attribute *attr,
1784                                      const char *buf, size_t count)
1785 {
1786         unsigned long msecs;
1787         int err;
1788
1789         err = strict_strtoul(buf, 10, &msecs);
1790         if (err || msecs > UINT_MAX)
1791                 return -EINVAL;
1792
1793         ksm_thread_sleep_millisecs = msecs;
1794
1795         return count;
1796 }
1797 KSM_ATTR(sleep_millisecs);
1798
1799 static ssize_t pages_to_scan_show(struct kobject *kobj,
1800                                   struct kobj_attribute *attr, char *buf)
1801 {
1802         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1803 }
1804
1805 static ssize_t pages_to_scan_store(struct kobject *kobj,
1806                                    struct kobj_attribute *attr,
1807                                    const char *buf, size_t count)
1808 {
1809         int err;
1810         unsigned long nr_pages;
1811
1812         err = strict_strtoul(buf, 10, &nr_pages);
1813         if (err || nr_pages > UINT_MAX)
1814                 return -EINVAL;
1815
1816         ksm_thread_pages_to_scan = nr_pages;
1817
1818         return count;
1819 }
1820 KSM_ATTR(pages_to_scan);
1821
1822 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1823                         char *buf)
1824 {
1825         return sprintf(buf, "%u\n", ksm_run);
1826 }
1827
1828 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1829                          const char *buf, size_t count)
1830 {
1831         int err;
1832         unsigned long flags;
1833
1834         err = strict_strtoul(buf, 10, &flags);
1835         if (err || flags > UINT_MAX)
1836                 return -EINVAL;
1837         if (flags > KSM_RUN_UNMERGE)
1838                 return -EINVAL;
1839
1840         /*
1841          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1842          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1843          * breaking COW to free the pages_shared (but leaves mm_slots
1844          * on the list for when ksmd may be set running again).
1845          */
1846
1847         mutex_lock(&ksm_thread_mutex);
1848         if (ksm_run != flags) {
1849                 ksm_run = flags;
1850                 if (flags & KSM_RUN_UNMERGE) {
1851                         current->flags |= PF_OOM_ORIGIN;
1852                         err = unmerge_and_remove_all_rmap_items();
1853                         current->flags &= ~PF_OOM_ORIGIN;
1854                         if (err) {
1855                                 ksm_run = KSM_RUN_STOP;
1856                                 count = err;
1857                         }
1858                 }
1859         }
1860         mutex_unlock(&ksm_thread_mutex);
1861
1862         if (flags & KSM_RUN_MERGE)
1863                 wake_up_interruptible(&ksm_thread_wait);
1864
1865         return count;
1866 }
1867 KSM_ATTR(run);
1868
1869 static ssize_t pages_shared_show(struct kobject *kobj,
1870                                  struct kobj_attribute *attr, char *buf)
1871 {
1872         return sprintf(buf, "%lu\n", ksm_pages_shared);
1873 }
1874 KSM_ATTR_RO(pages_shared);
1875
1876 static ssize_t pages_sharing_show(struct kobject *kobj,
1877                                   struct kobj_attribute *attr, char *buf)
1878 {
1879         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1880 }
1881 KSM_ATTR_RO(pages_sharing);
1882
1883 static ssize_t pages_unshared_show(struct kobject *kobj,
1884                                    struct kobj_attribute *attr, char *buf)
1885 {
1886         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1887 }
1888 KSM_ATTR_RO(pages_unshared);
1889
1890 static ssize_t pages_volatile_show(struct kobject *kobj,
1891                                    struct kobj_attribute *attr, char *buf)
1892 {
1893         long ksm_pages_volatile;
1894
1895         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1896                                 - ksm_pages_sharing - ksm_pages_unshared;
1897         /*
1898          * It was not worth any locking to calculate that statistic,
1899          * but it might therefore sometimes be negative: conceal that.
1900          */
1901         if (ksm_pages_volatile < 0)
1902                 ksm_pages_volatile = 0;
1903         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1904 }
1905 KSM_ATTR_RO(pages_volatile);
1906
1907 static ssize_t full_scans_show(struct kobject *kobj,
1908                                struct kobj_attribute *attr, char *buf)
1909 {
1910         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1911 }
1912 KSM_ATTR_RO(full_scans);
1913
1914 static struct attribute *ksm_attrs[] = {
1915         &sleep_millisecs_attr.attr,
1916         &pages_to_scan_attr.attr,
1917         &run_attr.attr,
1918         &pages_shared_attr.attr,
1919         &pages_sharing_attr.attr,
1920         &pages_unshared_attr.attr,
1921         &pages_volatile_attr.attr,
1922         &full_scans_attr.attr,
1923         NULL,
1924 };
1925
1926 static struct attribute_group ksm_attr_group = {
1927         .attrs = ksm_attrs,
1928         .name = "ksm",
1929 };
1930 #endif /* CONFIG_SYSFS */
1931
1932 static int __init ksm_init(void)
1933 {
1934         struct task_struct *ksm_thread;
1935         int err;
1936
1937         err = ksm_slab_init();
1938         if (err)
1939                 goto out;
1940
1941         err = mm_slots_hash_init();
1942         if (err)
1943                 goto out_free1;
1944
1945         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1946         if (IS_ERR(ksm_thread)) {
1947                 printk(KERN_ERR "ksm: creating kthread failed\n");
1948                 err = PTR_ERR(ksm_thread);
1949                 goto out_free2;
1950         }
1951
1952 #ifdef CONFIG_SYSFS
1953         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1954         if (err) {
1955                 printk(KERN_ERR "ksm: register sysfs failed\n");
1956                 kthread_stop(ksm_thread);
1957                 goto out_free2;
1958         }
1959 #else
1960         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
1961
1962 #endif /* CONFIG_SYSFS */
1963
1964 #ifdef CONFIG_MEMORY_HOTREMOVE
1965         /*
1966          * Choose a high priority since the callback takes ksm_thread_mutex:
1967          * later callbacks could only be taking locks which nest within that.
1968          */
1969         hotplug_memory_notifier(ksm_memory_callback, 100);
1970 #endif
1971         return 0;
1972
1973 out_free2:
1974         mm_slots_hash_free();
1975 out_free1:
1976         ksm_slab_free();
1977 out:
1978         return err;
1979 }
1980 module_init(ksm_init)