ksm: take keyhole reference to page
[linux-2.6.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/mmu_notifier.h>
33 #include <linux/swap.h>
34 #include <linux/ksm.h>
35
36 #include <asm/tlbflush.h>
37 #include "internal.h"
38
39 /*
40  * A few notes about the KSM scanning process,
41  * to make it easier to understand the data structures below:
42  *
43  * In order to reduce excessive scanning, KSM sorts the memory pages by their
44  * contents into a data structure that holds pointers to the pages' locations.
45  *
46  * Since the contents of the pages may change at any moment, KSM cannot just
47  * insert the pages into a normal sorted tree and expect it to find anything.
48  * Therefore KSM uses two data structures - the stable and the unstable tree.
49  *
50  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
51  * by their contents.  Because each such page is write-protected, searching on
52  * this tree is fully assured to be working (except when pages are unmapped),
53  * and therefore this tree is called the stable tree.
54  *
55  * In addition to the stable tree, KSM uses a second data structure called the
56  * unstable tree: this tree holds pointers to pages which have been found to
57  * be "unchanged for a period of time".  The unstable tree sorts these pages
58  * by their contents, but since they are not write-protected, KSM cannot rely
59  * upon the unstable tree to work correctly - the unstable tree is liable to
60  * be corrupted as its contents are modified, and so it is called unstable.
61  *
62  * KSM solves this problem by several techniques:
63  *
64  * 1) The unstable tree is flushed every time KSM completes scanning all
65  *    memory areas, and then the tree is rebuilt again from the beginning.
66  * 2) KSM will only insert into the unstable tree, pages whose hash value
67  *    has not changed since the previous scan of all memory areas.
68  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
69  *    colors of the nodes and not on their contents, assuring that even when
70  *    the tree gets "corrupted" it won't get out of balance, so scanning time
71  *    remains the same (also, searching and inserting nodes in an rbtree uses
72  *    the same algorithm, so we have no overhead when we flush and rebuild).
73  * 4) KSM never flushes the stable tree, which means that even if it were to
74  *    take 10 attempts to find a page in the unstable tree, once it is found,
75  *    it is secured in the stable tree.  (When we scan a new page, we first
76  *    compare it against the stable tree, and then against the unstable tree.)
77  */
78
79 /**
80  * struct mm_slot - ksm information per mm that is being scanned
81  * @link: link to the mm_slots hash list
82  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
83  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
84  * @mm: the mm that this information is valid for
85  */
86 struct mm_slot {
87         struct hlist_node link;
88         struct list_head mm_list;
89         struct rmap_item *rmap_list;
90         struct mm_struct *mm;
91 };
92
93 /**
94  * struct ksm_scan - cursor for scanning
95  * @mm_slot: the current mm_slot we are scanning
96  * @address: the next address inside that to be scanned
97  * @rmap_list: link to the next rmap to be scanned in the rmap_list
98  * @seqnr: count of completed full scans (needed when removing unstable node)
99  *
100  * There is only the one ksm_scan instance of this cursor structure.
101  */
102 struct ksm_scan {
103         struct mm_slot *mm_slot;
104         unsigned long address;
105         struct rmap_item **rmap_list;
106         unsigned long seqnr;
107 };
108
109 /**
110  * struct stable_node - node of the stable rbtree
111  * @page: pointer to struct page of the ksm page
112  * @node: rb node of this ksm page in the stable tree
113  * @hlist: hlist head of rmap_items using this ksm page
114  */
115 struct stable_node {
116         struct page *page;
117         struct rb_node node;
118         struct hlist_head hlist;
119 };
120
121 /**
122  * struct rmap_item - reverse mapping item for virtual addresses
123  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
124  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
125  * @mm: the memory structure this rmap_item is pointing into
126  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
127  * @oldchecksum: previous checksum of the page at that virtual address
128  * @node: rb node of this rmap_item in the unstable tree
129  * @head: pointer to stable_node heading this list in the stable tree
130  * @hlist: link into hlist of rmap_items hanging off that stable_node
131  */
132 struct rmap_item {
133         struct rmap_item *rmap_list;
134         struct anon_vma *anon_vma;      /* when stable */
135         struct mm_struct *mm;
136         unsigned long address;          /* + low bits used for flags below */
137         unsigned int oldchecksum;       /* when unstable */
138         union {
139                 struct rb_node node;    /* when node of unstable tree */
140                 struct {                /* when listed from stable tree */
141                         struct stable_node *head;
142                         struct hlist_node hlist;
143                 };
144         };
145 };
146
147 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
148 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
149 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
150
151 /* The stable and unstable tree heads */
152 static struct rb_root root_stable_tree = RB_ROOT;
153 static struct rb_root root_unstable_tree = RB_ROOT;
154
155 #define MM_SLOTS_HASH_HEADS 1024
156 static struct hlist_head *mm_slots_hash;
157
158 static struct mm_slot ksm_mm_head = {
159         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
160 };
161 static struct ksm_scan ksm_scan = {
162         .mm_slot = &ksm_mm_head,
163 };
164
165 static struct kmem_cache *rmap_item_cache;
166 static struct kmem_cache *stable_node_cache;
167 static struct kmem_cache *mm_slot_cache;
168
169 /* The number of nodes in the stable tree */
170 static unsigned long ksm_pages_shared;
171
172 /* The number of page slots additionally sharing those nodes */
173 static unsigned long ksm_pages_sharing;
174
175 /* The number of nodes in the unstable tree */
176 static unsigned long ksm_pages_unshared;
177
178 /* The number of rmap_items in use: to calculate pages_volatile */
179 static unsigned long ksm_rmap_items;
180
181 /* Limit on the number of unswappable pages used */
182 static unsigned long ksm_max_kernel_pages;
183
184 /* Number of pages ksmd should scan in one batch */
185 static unsigned int ksm_thread_pages_to_scan = 100;
186
187 /* Milliseconds ksmd should sleep between batches */
188 static unsigned int ksm_thread_sleep_millisecs = 20;
189
190 #define KSM_RUN_STOP    0
191 #define KSM_RUN_MERGE   1
192 #define KSM_RUN_UNMERGE 2
193 static unsigned int ksm_run = KSM_RUN_STOP;
194
195 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
196 static DEFINE_MUTEX(ksm_thread_mutex);
197 static DEFINE_SPINLOCK(ksm_mmlist_lock);
198
199 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
200                 sizeof(struct __struct), __alignof__(struct __struct),\
201                 (__flags), NULL)
202
203 static int __init ksm_slab_init(void)
204 {
205         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
206         if (!rmap_item_cache)
207                 goto out;
208
209         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
210         if (!stable_node_cache)
211                 goto out_free1;
212
213         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
214         if (!mm_slot_cache)
215                 goto out_free2;
216
217         return 0;
218
219 out_free2:
220         kmem_cache_destroy(stable_node_cache);
221 out_free1:
222         kmem_cache_destroy(rmap_item_cache);
223 out:
224         return -ENOMEM;
225 }
226
227 static void __init ksm_slab_free(void)
228 {
229         kmem_cache_destroy(mm_slot_cache);
230         kmem_cache_destroy(stable_node_cache);
231         kmem_cache_destroy(rmap_item_cache);
232         mm_slot_cache = NULL;
233 }
234
235 static inline struct rmap_item *alloc_rmap_item(void)
236 {
237         struct rmap_item *rmap_item;
238
239         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
240         if (rmap_item)
241                 ksm_rmap_items++;
242         return rmap_item;
243 }
244
245 static inline void free_rmap_item(struct rmap_item *rmap_item)
246 {
247         ksm_rmap_items--;
248         rmap_item->mm = NULL;   /* debug safety */
249         kmem_cache_free(rmap_item_cache, rmap_item);
250 }
251
252 static inline struct stable_node *alloc_stable_node(void)
253 {
254         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
255 }
256
257 static inline void free_stable_node(struct stable_node *stable_node)
258 {
259         kmem_cache_free(stable_node_cache, stable_node);
260 }
261
262 static inline struct mm_slot *alloc_mm_slot(void)
263 {
264         if (!mm_slot_cache)     /* initialization failed */
265                 return NULL;
266         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
267 }
268
269 static inline void free_mm_slot(struct mm_slot *mm_slot)
270 {
271         kmem_cache_free(mm_slot_cache, mm_slot);
272 }
273
274 static int __init mm_slots_hash_init(void)
275 {
276         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
277                                 GFP_KERNEL);
278         if (!mm_slots_hash)
279                 return -ENOMEM;
280         return 0;
281 }
282
283 static void __init mm_slots_hash_free(void)
284 {
285         kfree(mm_slots_hash);
286 }
287
288 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
289 {
290         struct mm_slot *mm_slot;
291         struct hlist_head *bucket;
292         struct hlist_node *node;
293
294         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
295                                 % MM_SLOTS_HASH_HEADS];
296         hlist_for_each_entry(mm_slot, node, bucket, link) {
297                 if (mm == mm_slot->mm)
298                         return mm_slot;
299         }
300         return NULL;
301 }
302
303 static void insert_to_mm_slots_hash(struct mm_struct *mm,
304                                     struct mm_slot *mm_slot)
305 {
306         struct hlist_head *bucket;
307
308         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
309                                 % MM_SLOTS_HASH_HEADS];
310         mm_slot->mm = mm;
311         hlist_add_head(&mm_slot->link, bucket);
312 }
313
314 static inline int in_stable_tree(struct rmap_item *rmap_item)
315 {
316         return rmap_item->address & STABLE_FLAG;
317 }
318
319 static void hold_anon_vma(struct rmap_item *rmap_item,
320                           struct anon_vma *anon_vma)
321 {
322         rmap_item->anon_vma = anon_vma;
323         atomic_inc(&anon_vma->ksm_refcount);
324 }
325
326 static void drop_anon_vma(struct rmap_item *rmap_item)
327 {
328         struct anon_vma *anon_vma = rmap_item->anon_vma;
329
330         if (atomic_dec_and_lock(&anon_vma->ksm_refcount, &anon_vma->lock)) {
331                 int empty = list_empty(&anon_vma->head);
332                 spin_unlock(&anon_vma->lock);
333                 if (empty)
334                         anon_vma_free(anon_vma);
335         }
336 }
337
338 /*
339  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
340  * page tables after it has passed through ksm_exit() - which, if necessary,
341  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
342  * a special flag: they can just back out as soon as mm_users goes to zero.
343  * ksm_test_exit() is used throughout to make this test for exit: in some
344  * places for correctness, in some places just to avoid unnecessary work.
345  */
346 static inline bool ksm_test_exit(struct mm_struct *mm)
347 {
348         return atomic_read(&mm->mm_users) == 0;
349 }
350
351 /*
352  * We use break_ksm to break COW on a ksm page: it's a stripped down
353  *
354  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
355  *              put_page(page);
356  *
357  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
358  * in case the application has unmapped and remapped mm,addr meanwhile.
359  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
360  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
361  */
362 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
363 {
364         struct page *page;
365         int ret = 0;
366
367         do {
368                 cond_resched();
369                 page = follow_page(vma, addr, FOLL_GET);
370                 if (!page)
371                         break;
372                 if (PageKsm(page))
373                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
374                                                         FAULT_FLAG_WRITE);
375                 else
376                         ret = VM_FAULT_WRITE;
377                 put_page(page);
378         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
379         /*
380          * We must loop because handle_mm_fault() may back out if there's
381          * any difficulty e.g. if pte accessed bit gets updated concurrently.
382          *
383          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
384          * COW has been broken, even if the vma does not permit VM_WRITE;
385          * but note that a concurrent fault might break PageKsm for us.
386          *
387          * VM_FAULT_SIGBUS could occur if we race with truncation of the
388          * backing file, which also invalidates anonymous pages: that's
389          * okay, that truncation will have unmapped the PageKsm for us.
390          *
391          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
392          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
393          * current task has TIF_MEMDIE set, and will be OOM killed on return
394          * to user; and ksmd, having no mm, would never be chosen for that.
395          *
396          * But if the mm is in a limited mem_cgroup, then the fault may fail
397          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
398          * even ksmd can fail in this way - though it's usually breaking ksm
399          * just to undo a merge it made a moment before, so unlikely to oom.
400          *
401          * That's a pity: we might therefore have more kernel pages allocated
402          * than we're counting as nodes in the stable tree; but ksm_do_scan
403          * will retry to break_cow on each pass, so should recover the page
404          * in due course.  The important thing is to not let VM_MERGEABLE
405          * be cleared while any such pages might remain in the area.
406          */
407         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
408 }
409
410 static void break_cow(struct rmap_item *rmap_item)
411 {
412         struct mm_struct *mm = rmap_item->mm;
413         unsigned long addr = rmap_item->address;
414         struct vm_area_struct *vma;
415
416         /*
417          * It is not an accident that whenever we want to break COW
418          * to undo, we also need to drop a reference to the anon_vma.
419          */
420         drop_anon_vma(rmap_item);
421
422         down_read(&mm->mmap_sem);
423         if (ksm_test_exit(mm))
424                 goto out;
425         vma = find_vma(mm, addr);
426         if (!vma || vma->vm_start > addr)
427                 goto out;
428         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
429                 goto out;
430         break_ksm(vma, addr);
431 out:
432         up_read(&mm->mmap_sem);
433 }
434
435 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
436 {
437         struct mm_struct *mm = rmap_item->mm;
438         unsigned long addr = rmap_item->address;
439         struct vm_area_struct *vma;
440         struct page *page;
441
442         down_read(&mm->mmap_sem);
443         if (ksm_test_exit(mm))
444                 goto out;
445         vma = find_vma(mm, addr);
446         if (!vma || vma->vm_start > addr)
447                 goto out;
448         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
449                 goto out;
450
451         page = follow_page(vma, addr, FOLL_GET);
452         if (!page)
453                 goto out;
454         if (PageAnon(page)) {
455                 flush_anon_page(vma, page, addr);
456                 flush_dcache_page(page);
457         } else {
458                 put_page(page);
459 out:            page = NULL;
460         }
461         up_read(&mm->mmap_sem);
462         return page;
463 }
464
465 static void remove_node_from_stable_tree(struct stable_node *stable_node)
466 {
467         struct rmap_item *rmap_item;
468         struct hlist_node *hlist;
469
470         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
471                 if (rmap_item->hlist.next)
472                         ksm_pages_sharing--;
473                 else
474                         ksm_pages_shared--;
475                 drop_anon_vma(rmap_item);
476                 rmap_item->address &= PAGE_MASK;
477                 cond_resched();
478         }
479
480         rb_erase(&stable_node->node, &root_stable_tree);
481         free_stable_node(stable_node);
482 }
483
484 /*
485  * get_ksm_page: checks if the page indicated by the stable node
486  * is still its ksm page, despite having held no reference to it.
487  * In which case we can trust the content of the page, and it
488  * returns the gotten page; but if the page has now been zapped,
489  * remove the stale node from the stable tree and return NULL.
490  *
491  * You would expect the stable_node to hold a reference to the ksm page.
492  * But if it increments the page's count, swapping out has to wait for
493  * ksmd to come around again before it can free the page, which may take
494  * seconds or even minutes: much too unresponsive.  So instead we use a
495  * "keyhole reference": access to the ksm page from the stable node peeps
496  * out through its keyhole to see if that page still holds the right key,
497  * pointing back to this stable node.  This relies on freeing a PageAnon
498  * page to reset its page->mapping to NULL, and relies on no other use of
499  * a page to put something that might look like our key in page->mapping.
500  *
501  * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
502  * but this is different - made simpler by ksm_thread_mutex being held, but
503  * interesting for assuming that no other use of the struct page could ever
504  * put our expected_mapping into page->mapping (or a field of the union which
505  * coincides with page->mapping).  The RCU calls are not for KSM at all, but
506  * to keep the page_count protocol described with page_cache_get_speculative.
507  *
508  * Note: it is possible that get_ksm_page() will return NULL one moment,
509  * then page the next, if the page is in between page_freeze_refs() and
510  * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
511  * is on its way to being freed; but it is an anomaly to bear in mind.
512  */
513 static struct page *get_ksm_page(struct stable_node *stable_node)
514 {
515         struct page *page;
516         void *expected_mapping;
517
518         page = stable_node->page;
519         expected_mapping = (void *)stable_node +
520                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
521         rcu_read_lock();
522         if (page->mapping != expected_mapping)
523                 goto stale;
524         if (!get_page_unless_zero(page))
525                 goto stale;
526         if (page->mapping != expected_mapping) {
527                 put_page(page);
528                 goto stale;
529         }
530         rcu_read_unlock();
531         return page;
532 stale:
533         rcu_read_unlock();
534         remove_node_from_stable_tree(stable_node);
535         return NULL;
536 }
537
538 /*
539  * Removing rmap_item from stable or unstable tree.
540  * This function will clean the information from the stable/unstable tree.
541  */
542 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
543 {
544         if (rmap_item->address & STABLE_FLAG) {
545                 struct stable_node *stable_node;
546                 struct page *page;
547
548                 stable_node = rmap_item->head;
549                 page = get_ksm_page(stable_node);
550                 if (!page)
551                         goto out;
552
553                 lock_page(page);
554                 hlist_del(&rmap_item->hlist);
555                 unlock_page(page);
556                 put_page(page);
557
558                 if (stable_node->hlist.first)
559                         ksm_pages_sharing--;
560                 else
561                         ksm_pages_shared--;
562
563                 drop_anon_vma(rmap_item);
564                 rmap_item->address &= PAGE_MASK;
565
566         } else if (rmap_item->address & UNSTABLE_FLAG) {
567                 unsigned char age;
568                 /*
569                  * Usually ksmd can and must skip the rb_erase, because
570                  * root_unstable_tree was already reset to RB_ROOT.
571                  * But be careful when an mm is exiting: do the rb_erase
572                  * if this rmap_item was inserted by this scan, rather
573                  * than left over from before.
574                  */
575                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
576                 BUG_ON(age > 1);
577                 if (!age)
578                         rb_erase(&rmap_item->node, &root_unstable_tree);
579
580                 ksm_pages_unshared--;
581                 rmap_item->address &= PAGE_MASK;
582         }
583 out:
584         cond_resched();         /* we're called from many long loops */
585 }
586
587 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
588                                        struct rmap_item **rmap_list)
589 {
590         while (*rmap_list) {
591                 struct rmap_item *rmap_item = *rmap_list;
592                 *rmap_list = rmap_item->rmap_list;
593                 remove_rmap_item_from_tree(rmap_item);
594                 free_rmap_item(rmap_item);
595         }
596 }
597
598 /*
599  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
600  * than check every pte of a given vma, the locking doesn't quite work for
601  * that - an rmap_item is assigned to the stable tree after inserting ksm
602  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
603  * rmap_items from parent to child at fork time (so as not to waste time
604  * if exit comes before the next scan reaches it).
605  *
606  * Similarly, although we'd like to remove rmap_items (so updating counts
607  * and freeing memory) when unmerging an area, it's easier to leave that
608  * to the next pass of ksmd - consider, for example, how ksmd might be
609  * in cmp_and_merge_page on one of the rmap_items we would be removing.
610  */
611 static int unmerge_ksm_pages(struct vm_area_struct *vma,
612                              unsigned long start, unsigned long end)
613 {
614         unsigned long addr;
615         int err = 0;
616
617         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
618                 if (ksm_test_exit(vma->vm_mm))
619                         break;
620                 if (signal_pending(current))
621                         err = -ERESTARTSYS;
622                 else
623                         err = break_ksm(vma, addr);
624         }
625         return err;
626 }
627
628 #ifdef CONFIG_SYSFS
629 /*
630  * Only called through the sysfs control interface:
631  */
632 static int unmerge_and_remove_all_rmap_items(void)
633 {
634         struct mm_slot *mm_slot;
635         struct mm_struct *mm;
636         struct vm_area_struct *vma;
637         int err = 0;
638
639         spin_lock(&ksm_mmlist_lock);
640         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
641                                                 struct mm_slot, mm_list);
642         spin_unlock(&ksm_mmlist_lock);
643
644         for (mm_slot = ksm_scan.mm_slot;
645                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
646                 mm = mm_slot->mm;
647                 down_read(&mm->mmap_sem);
648                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
649                         if (ksm_test_exit(mm))
650                                 break;
651                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
652                                 continue;
653                         err = unmerge_ksm_pages(vma,
654                                                 vma->vm_start, vma->vm_end);
655                         if (err)
656                                 goto error;
657                 }
658
659                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
660
661                 spin_lock(&ksm_mmlist_lock);
662                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
663                                                 struct mm_slot, mm_list);
664                 if (ksm_test_exit(mm)) {
665                         hlist_del(&mm_slot->link);
666                         list_del(&mm_slot->mm_list);
667                         spin_unlock(&ksm_mmlist_lock);
668
669                         free_mm_slot(mm_slot);
670                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
671                         up_read(&mm->mmap_sem);
672                         mmdrop(mm);
673                 } else {
674                         spin_unlock(&ksm_mmlist_lock);
675                         up_read(&mm->mmap_sem);
676                 }
677         }
678
679         ksm_scan.seqnr = 0;
680         return 0;
681
682 error:
683         up_read(&mm->mmap_sem);
684         spin_lock(&ksm_mmlist_lock);
685         ksm_scan.mm_slot = &ksm_mm_head;
686         spin_unlock(&ksm_mmlist_lock);
687         return err;
688 }
689 #endif /* CONFIG_SYSFS */
690
691 static u32 calc_checksum(struct page *page)
692 {
693         u32 checksum;
694         void *addr = kmap_atomic(page, KM_USER0);
695         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
696         kunmap_atomic(addr, KM_USER0);
697         return checksum;
698 }
699
700 static int memcmp_pages(struct page *page1, struct page *page2)
701 {
702         char *addr1, *addr2;
703         int ret;
704
705         addr1 = kmap_atomic(page1, KM_USER0);
706         addr2 = kmap_atomic(page2, KM_USER1);
707         ret = memcmp(addr1, addr2, PAGE_SIZE);
708         kunmap_atomic(addr2, KM_USER1);
709         kunmap_atomic(addr1, KM_USER0);
710         return ret;
711 }
712
713 static inline int pages_identical(struct page *page1, struct page *page2)
714 {
715         return !memcmp_pages(page1, page2);
716 }
717
718 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
719                               pte_t *orig_pte)
720 {
721         struct mm_struct *mm = vma->vm_mm;
722         unsigned long addr;
723         pte_t *ptep;
724         spinlock_t *ptl;
725         int swapped;
726         int err = -EFAULT;
727
728         addr = page_address_in_vma(page, vma);
729         if (addr == -EFAULT)
730                 goto out;
731
732         ptep = page_check_address(page, mm, addr, &ptl, 0);
733         if (!ptep)
734                 goto out;
735
736         if (pte_write(*ptep)) {
737                 pte_t entry;
738
739                 swapped = PageSwapCache(page);
740                 flush_cache_page(vma, addr, page_to_pfn(page));
741                 /*
742                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
743                  * take any lock, therefore the check that we are going to make
744                  * with the pagecount against the mapcount is racey and
745                  * O_DIRECT can happen right after the check.
746                  * So we clear the pte and flush the tlb before the check
747                  * this assure us that no O_DIRECT can happen after the check
748                  * or in the middle of the check.
749                  */
750                 entry = ptep_clear_flush(vma, addr, ptep);
751                 /*
752                  * Check that no O_DIRECT or similar I/O is in progress on the
753                  * page
754                  */
755                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
756                         set_pte_at_notify(mm, addr, ptep, entry);
757                         goto out_unlock;
758                 }
759                 entry = pte_wrprotect(entry);
760                 set_pte_at_notify(mm, addr, ptep, entry);
761         }
762         *orig_pte = *ptep;
763         err = 0;
764
765 out_unlock:
766         pte_unmap_unlock(ptep, ptl);
767 out:
768         return err;
769 }
770
771 /**
772  * replace_page - replace page in vma by new ksm page
773  * @vma:      vma that holds the pte pointing to page
774  * @page:     the page we are replacing by kpage
775  * @kpage:    the ksm page we replace page by
776  * @orig_pte: the original value of the pte
777  *
778  * Returns 0 on success, -EFAULT on failure.
779  */
780 static int replace_page(struct vm_area_struct *vma, struct page *page,
781                         struct page *kpage, pte_t orig_pte)
782 {
783         struct mm_struct *mm = vma->vm_mm;
784         pgd_t *pgd;
785         pud_t *pud;
786         pmd_t *pmd;
787         pte_t *ptep;
788         spinlock_t *ptl;
789         unsigned long addr;
790         int err = -EFAULT;
791
792         addr = page_address_in_vma(page, vma);
793         if (addr == -EFAULT)
794                 goto out;
795
796         pgd = pgd_offset(mm, addr);
797         if (!pgd_present(*pgd))
798                 goto out;
799
800         pud = pud_offset(pgd, addr);
801         if (!pud_present(*pud))
802                 goto out;
803
804         pmd = pmd_offset(pud, addr);
805         if (!pmd_present(*pmd))
806                 goto out;
807
808         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
809         if (!pte_same(*ptep, orig_pte)) {
810                 pte_unmap_unlock(ptep, ptl);
811                 goto out;
812         }
813
814         get_page(kpage);
815         page_add_anon_rmap(kpage, vma, addr);
816
817         flush_cache_page(vma, addr, pte_pfn(*ptep));
818         ptep_clear_flush(vma, addr, ptep);
819         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
820
821         page_remove_rmap(page);
822         put_page(page);
823
824         pte_unmap_unlock(ptep, ptl);
825         err = 0;
826 out:
827         return err;
828 }
829
830 /*
831  * try_to_merge_one_page - take two pages and merge them into one
832  * @vma: the vma that holds the pte pointing to page
833  * @page: the PageAnon page that we want to replace with kpage
834  * @kpage: the PageKsm page that we want to map instead of page
835  *
836  * This function returns 0 if the pages were merged, -EFAULT otherwise.
837  */
838 static int try_to_merge_one_page(struct vm_area_struct *vma,
839                                  struct page *page, struct page *kpage)
840 {
841         pte_t orig_pte = __pte(0);
842         int err = -EFAULT;
843
844         if (page == kpage)                      /* ksm page forked */
845                 return 0;
846
847         if (!(vma->vm_flags & VM_MERGEABLE))
848                 goto out;
849         if (!PageAnon(page))
850                 goto out;
851
852         /*
853          * We need the page lock to read a stable PageSwapCache in
854          * write_protect_page().  We use trylock_page() instead of
855          * lock_page() because we don't want to wait here - we
856          * prefer to continue scanning and merging different pages,
857          * then come back to this page when it is unlocked.
858          */
859         if (!trylock_page(page))
860                 goto out;
861         /*
862          * If this anonymous page is mapped only here, its pte may need
863          * to be write-protected.  If it's mapped elsewhere, all of its
864          * ptes are necessarily already write-protected.  But in either
865          * case, we need to lock and check page_count is not raised.
866          */
867         if (write_protect_page(vma, page, &orig_pte) == 0 &&
868             pages_identical(page, kpage))
869                 err = replace_page(vma, page, kpage, orig_pte);
870
871         if ((vma->vm_flags & VM_LOCKED) && !err) {
872                 munlock_vma_page(page);
873                 if (!PageMlocked(kpage)) {
874                         unlock_page(page);
875                         lru_add_drain();
876                         lock_page(kpage);
877                         mlock_vma_page(kpage);
878                         page = kpage;           /* for final unlock */
879                 }
880         }
881
882         unlock_page(page);
883 out:
884         return err;
885 }
886
887 /*
888  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
889  * but no new kernel page is allocated: kpage must already be a ksm page.
890  *
891  * This function returns 0 if the pages were merged, -EFAULT otherwise.
892  */
893 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
894                                       struct page *page, struct page *kpage)
895 {
896         struct mm_struct *mm = rmap_item->mm;
897         struct vm_area_struct *vma;
898         int err = -EFAULT;
899
900         down_read(&mm->mmap_sem);
901         if (ksm_test_exit(mm))
902                 goto out;
903         vma = find_vma(mm, rmap_item->address);
904         if (!vma || vma->vm_start > rmap_item->address)
905                 goto out;
906
907         err = try_to_merge_one_page(vma, page, kpage);
908         if (err)
909                 goto out;
910
911         /* Must get reference to anon_vma while still holding mmap_sem */
912         hold_anon_vma(rmap_item, vma->anon_vma);
913 out:
914         up_read(&mm->mmap_sem);
915         return err;
916 }
917
918 /*
919  * try_to_merge_two_pages - take two identical pages and prepare them
920  * to be merged into one page.
921  *
922  * This function returns the kpage if we successfully merged two identical
923  * pages into one ksm page, NULL otherwise.
924  *
925  * Note that this function allocates a new kernel page: if one of the pages
926  * is already a ksm page, try_to_merge_with_ksm_page should be used.
927  */
928 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
929                                            struct page *page,
930                                            struct rmap_item *tree_rmap_item,
931                                            struct page *tree_page)
932 {
933         struct mm_struct *mm = rmap_item->mm;
934         struct vm_area_struct *vma;
935         struct page *kpage;
936         int err = -EFAULT;
937
938         /*
939          * The number of nodes in the stable tree
940          * is the number of kernel pages that we hold.
941          */
942         if (ksm_max_kernel_pages &&
943             ksm_max_kernel_pages <= ksm_pages_shared)
944                 return NULL;
945
946         kpage = alloc_page(GFP_HIGHUSER);
947         if (!kpage)
948                 return NULL;
949
950         down_read(&mm->mmap_sem);
951         if (ksm_test_exit(mm))
952                 goto up;
953         vma = find_vma(mm, rmap_item->address);
954         if (!vma || vma->vm_start > rmap_item->address)
955                 goto up;
956
957         copy_user_highpage(kpage, page, rmap_item->address, vma);
958
959         SetPageDirty(kpage);
960         __SetPageUptodate(kpage);
961         SetPageSwapBacked(kpage);
962         set_page_stable_node(kpage, NULL);      /* mark it PageKsm */
963         lru_cache_add_lru(kpage, LRU_ACTIVE_ANON);
964
965         err = try_to_merge_one_page(vma, page, kpage);
966         if (err)
967                 goto up;
968
969         /* Must get reference to anon_vma while still holding mmap_sem */
970         hold_anon_vma(rmap_item, vma->anon_vma);
971 up:
972         up_read(&mm->mmap_sem);
973
974         if (!err) {
975                 err = try_to_merge_with_ksm_page(tree_rmap_item,
976                                                         tree_page, kpage);
977                 /*
978                  * If that fails, we have a ksm page with only one pte
979                  * pointing to it: so break it.
980                  */
981                 if (err)
982                         break_cow(rmap_item);
983         }
984         if (err) {
985                 put_page(kpage);
986                 kpage = NULL;
987         }
988         return kpage;
989 }
990
991 /*
992  * stable_tree_search - search for page inside the stable tree
993  *
994  * This function checks if there is a page inside the stable tree
995  * with identical content to the page that we are scanning right now.
996  *
997  * This function returns the stable tree node of identical content if found,
998  * NULL otherwise.
999  */
1000 static struct stable_node *stable_tree_search(struct page *page)
1001 {
1002         struct rb_node *node = root_stable_tree.rb_node;
1003         struct stable_node *stable_node;
1004
1005         stable_node = page_stable_node(page);
1006         if (stable_node) {                      /* ksm page forked */
1007                 get_page(page);
1008                 return stable_node;
1009         }
1010
1011         while (node) {
1012                 struct page *tree_page;
1013                 int ret;
1014
1015                 cond_resched();
1016                 stable_node = rb_entry(node, struct stable_node, node);
1017                 tree_page = get_ksm_page(stable_node);
1018                 if (!tree_page)
1019                         return NULL;
1020
1021                 ret = memcmp_pages(page, tree_page);
1022
1023                 if (ret < 0) {
1024                         put_page(tree_page);
1025                         node = node->rb_left;
1026                 } else if (ret > 0) {
1027                         put_page(tree_page);
1028                         node = node->rb_right;
1029                 } else
1030                         return stable_node;
1031         }
1032
1033         return NULL;
1034 }
1035
1036 /*
1037  * stable_tree_insert - insert rmap_item pointing to new ksm page
1038  * into the stable tree.
1039  *
1040  * This function returns the stable tree node just allocated on success,
1041  * NULL otherwise.
1042  */
1043 static struct stable_node *stable_tree_insert(struct page *kpage)
1044 {
1045         struct rb_node **new = &root_stable_tree.rb_node;
1046         struct rb_node *parent = NULL;
1047         struct stable_node *stable_node;
1048
1049         while (*new) {
1050                 struct page *tree_page;
1051                 int ret;
1052
1053                 cond_resched();
1054                 stable_node = rb_entry(*new, struct stable_node, node);
1055                 tree_page = get_ksm_page(stable_node);
1056                 if (!tree_page)
1057                         return NULL;
1058
1059                 ret = memcmp_pages(kpage, tree_page);
1060                 put_page(tree_page);
1061
1062                 parent = *new;
1063                 if (ret < 0)
1064                         new = &parent->rb_left;
1065                 else if (ret > 0)
1066                         new = &parent->rb_right;
1067                 else {
1068                         /*
1069                          * It is not a bug that stable_tree_search() didn't
1070                          * find this node: because at that time our page was
1071                          * not yet write-protected, so may have changed since.
1072                          */
1073                         return NULL;
1074                 }
1075         }
1076
1077         stable_node = alloc_stable_node();
1078         if (!stable_node)
1079                 return NULL;
1080
1081         rb_link_node(&stable_node->node, parent, new);
1082         rb_insert_color(&stable_node->node, &root_stable_tree);
1083
1084         INIT_HLIST_HEAD(&stable_node->hlist);
1085
1086         stable_node->page = kpage;
1087         set_page_stable_node(kpage, stable_node);
1088
1089         return stable_node;
1090 }
1091
1092 /*
1093  * unstable_tree_search_insert - search for identical page,
1094  * else insert rmap_item into the unstable tree.
1095  *
1096  * This function searches for a page in the unstable tree identical to the
1097  * page currently being scanned; and if no identical page is found in the
1098  * tree, we insert rmap_item as a new object into the unstable tree.
1099  *
1100  * This function returns pointer to rmap_item found to be identical
1101  * to the currently scanned page, NULL otherwise.
1102  *
1103  * This function does both searching and inserting, because they share
1104  * the same walking algorithm in an rbtree.
1105  */
1106 static
1107 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1108                                               struct page *page,
1109                                               struct page **tree_pagep)
1110
1111 {
1112         struct rb_node **new = &root_unstable_tree.rb_node;
1113         struct rb_node *parent = NULL;
1114
1115         while (*new) {
1116                 struct rmap_item *tree_rmap_item;
1117                 struct page *tree_page;
1118                 int ret;
1119
1120                 cond_resched();
1121                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1122                 tree_page = get_mergeable_page(tree_rmap_item);
1123                 if (!tree_page)
1124                         return NULL;
1125
1126                 /*
1127                  * Don't substitute a ksm page for a forked page.
1128                  */
1129                 if (page == tree_page) {
1130                         put_page(tree_page);
1131                         return NULL;
1132                 }
1133
1134                 ret = memcmp_pages(page, tree_page);
1135
1136                 parent = *new;
1137                 if (ret < 0) {
1138                         put_page(tree_page);
1139                         new = &parent->rb_left;
1140                 } else if (ret > 0) {
1141                         put_page(tree_page);
1142                         new = &parent->rb_right;
1143                 } else {
1144                         *tree_pagep = tree_page;
1145                         return tree_rmap_item;
1146                 }
1147         }
1148
1149         rmap_item->address |= UNSTABLE_FLAG;
1150         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1151         rb_link_node(&rmap_item->node, parent, new);
1152         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1153
1154         ksm_pages_unshared++;
1155         return NULL;
1156 }
1157
1158 /*
1159  * stable_tree_append - add another rmap_item to the linked list of
1160  * rmap_items hanging off a given node of the stable tree, all sharing
1161  * the same ksm page.
1162  */
1163 static void stable_tree_append(struct rmap_item *rmap_item,
1164                                struct stable_node *stable_node)
1165 {
1166         rmap_item->head = stable_node;
1167         rmap_item->address |= STABLE_FLAG;
1168         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1169
1170         if (rmap_item->hlist.next)
1171                 ksm_pages_sharing++;
1172         else
1173                 ksm_pages_shared++;
1174 }
1175
1176 /*
1177  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1178  * if not, compare checksum to previous and if it's the same, see if page can
1179  * be inserted into the unstable tree, or merged with a page already there and
1180  * both transferred to the stable tree.
1181  *
1182  * @page: the page that we are searching identical page to.
1183  * @rmap_item: the reverse mapping into the virtual address of this page
1184  */
1185 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1186 {
1187         struct rmap_item *tree_rmap_item;
1188         struct page *tree_page = NULL;
1189         struct stable_node *stable_node;
1190         struct page *kpage;
1191         unsigned int checksum;
1192         int err;
1193
1194         remove_rmap_item_from_tree(rmap_item);
1195
1196         /* We first start with searching the page inside the stable tree */
1197         stable_node = stable_tree_search(page);
1198         if (stable_node) {
1199                 kpage = stable_node->page;
1200                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1201                 if (!err) {
1202                         /*
1203                          * The page was successfully merged:
1204                          * add its rmap_item to the stable tree.
1205                          */
1206                         lock_page(kpage);
1207                         stable_tree_append(rmap_item, stable_node);
1208                         unlock_page(kpage);
1209                 }
1210                 put_page(kpage);
1211                 return;
1212         }
1213
1214         /*
1215          * If the hash value of the page has changed from the last time
1216          * we calculated it, this page is changing frequently: therefore we
1217          * don't want to insert it in the unstable tree, and we don't want
1218          * to waste our time searching for something identical to it there.
1219          */
1220         checksum = calc_checksum(page);
1221         if (rmap_item->oldchecksum != checksum) {
1222                 rmap_item->oldchecksum = checksum;
1223                 return;
1224         }
1225
1226         tree_rmap_item =
1227                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1228         if (tree_rmap_item) {
1229                 kpage = try_to_merge_two_pages(rmap_item, page,
1230                                                 tree_rmap_item, tree_page);
1231                 put_page(tree_page);
1232                 /*
1233                  * As soon as we merge this page, we want to remove the
1234                  * rmap_item of the page we have merged with from the unstable
1235                  * tree, and insert it instead as new node in the stable tree.
1236                  */
1237                 if (kpage) {
1238                         remove_rmap_item_from_tree(tree_rmap_item);
1239
1240                         lock_page(kpage);
1241                         stable_node = stable_tree_insert(kpage);
1242                         if (stable_node) {
1243                                 stable_tree_append(tree_rmap_item, stable_node);
1244                                 stable_tree_append(rmap_item, stable_node);
1245                         }
1246                         unlock_page(kpage);
1247                         put_page(kpage);
1248
1249                         /*
1250                          * If we fail to insert the page into the stable tree,
1251                          * we will have 2 virtual addresses that are pointing
1252                          * to a ksm page left outside the stable tree,
1253                          * in which case we need to break_cow on both.
1254                          */
1255                         if (!stable_node) {
1256                                 break_cow(tree_rmap_item);
1257                                 break_cow(rmap_item);
1258                         }
1259                 }
1260         }
1261 }
1262
1263 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1264                                             struct rmap_item **rmap_list,
1265                                             unsigned long addr)
1266 {
1267         struct rmap_item *rmap_item;
1268
1269         while (*rmap_list) {
1270                 rmap_item = *rmap_list;
1271                 if ((rmap_item->address & PAGE_MASK) == addr)
1272                         return rmap_item;
1273                 if (rmap_item->address > addr)
1274                         break;
1275                 *rmap_list = rmap_item->rmap_list;
1276                 remove_rmap_item_from_tree(rmap_item);
1277                 free_rmap_item(rmap_item);
1278         }
1279
1280         rmap_item = alloc_rmap_item();
1281         if (rmap_item) {
1282                 /* It has already been zeroed */
1283                 rmap_item->mm = mm_slot->mm;
1284                 rmap_item->address = addr;
1285                 rmap_item->rmap_list = *rmap_list;
1286                 *rmap_list = rmap_item;
1287         }
1288         return rmap_item;
1289 }
1290
1291 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1292 {
1293         struct mm_struct *mm;
1294         struct mm_slot *slot;
1295         struct vm_area_struct *vma;
1296         struct rmap_item *rmap_item;
1297
1298         if (list_empty(&ksm_mm_head.mm_list))
1299                 return NULL;
1300
1301         slot = ksm_scan.mm_slot;
1302         if (slot == &ksm_mm_head) {
1303                 root_unstable_tree = RB_ROOT;
1304
1305                 spin_lock(&ksm_mmlist_lock);
1306                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1307                 ksm_scan.mm_slot = slot;
1308                 spin_unlock(&ksm_mmlist_lock);
1309 next_mm:
1310                 ksm_scan.address = 0;
1311                 ksm_scan.rmap_list = &slot->rmap_list;
1312         }
1313
1314         mm = slot->mm;
1315         down_read(&mm->mmap_sem);
1316         if (ksm_test_exit(mm))
1317                 vma = NULL;
1318         else
1319                 vma = find_vma(mm, ksm_scan.address);
1320
1321         for (; vma; vma = vma->vm_next) {
1322                 if (!(vma->vm_flags & VM_MERGEABLE))
1323                         continue;
1324                 if (ksm_scan.address < vma->vm_start)
1325                         ksm_scan.address = vma->vm_start;
1326                 if (!vma->anon_vma)
1327                         ksm_scan.address = vma->vm_end;
1328
1329                 while (ksm_scan.address < vma->vm_end) {
1330                         if (ksm_test_exit(mm))
1331                                 break;
1332                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1333                         if (*page && PageAnon(*page)) {
1334                                 flush_anon_page(vma, *page, ksm_scan.address);
1335                                 flush_dcache_page(*page);
1336                                 rmap_item = get_next_rmap_item(slot,
1337                                         ksm_scan.rmap_list, ksm_scan.address);
1338                                 if (rmap_item) {
1339                                         ksm_scan.rmap_list =
1340                                                         &rmap_item->rmap_list;
1341                                         ksm_scan.address += PAGE_SIZE;
1342                                 } else
1343                                         put_page(*page);
1344                                 up_read(&mm->mmap_sem);
1345                                 return rmap_item;
1346                         }
1347                         if (*page)
1348                                 put_page(*page);
1349                         ksm_scan.address += PAGE_SIZE;
1350                         cond_resched();
1351                 }
1352         }
1353
1354         if (ksm_test_exit(mm)) {
1355                 ksm_scan.address = 0;
1356                 ksm_scan.rmap_list = &slot->rmap_list;
1357         }
1358         /*
1359          * Nuke all the rmap_items that are above this current rmap:
1360          * because there were no VM_MERGEABLE vmas with such addresses.
1361          */
1362         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1363
1364         spin_lock(&ksm_mmlist_lock);
1365         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1366                                                 struct mm_slot, mm_list);
1367         if (ksm_scan.address == 0) {
1368                 /*
1369                  * We've completed a full scan of all vmas, holding mmap_sem
1370                  * throughout, and found no VM_MERGEABLE: so do the same as
1371                  * __ksm_exit does to remove this mm from all our lists now.
1372                  * This applies either when cleaning up after __ksm_exit
1373                  * (but beware: we can reach here even before __ksm_exit),
1374                  * or when all VM_MERGEABLE areas have been unmapped (and
1375                  * mmap_sem then protects against race with MADV_MERGEABLE).
1376                  */
1377                 hlist_del(&slot->link);
1378                 list_del(&slot->mm_list);
1379                 spin_unlock(&ksm_mmlist_lock);
1380
1381                 free_mm_slot(slot);
1382                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1383                 up_read(&mm->mmap_sem);
1384                 mmdrop(mm);
1385         } else {
1386                 spin_unlock(&ksm_mmlist_lock);
1387                 up_read(&mm->mmap_sem);
1388         }
1389
1390         /* Repeat until we've completed scanning the whole list */
1391         slot = ksm_scan.mm_slot;
1392         if (slot != &ksm_mm_head)
1393                 goto next_mm;
1394
1395         ksm_scan.seqnr++;
1396         return NULL;
1397 }
1398
1399 /**
1400  * ksm_do_scan  - the ksm scanner main worker function.
1401  * @scan_npages - number of pages we want to scan before we return.
1402  */
1403 static void ksm_do_scan(unsigned int scan_npages)
1404 {
1405         struct rmap_item *rmap_item;
1406         struct page *page;
1407
1408         while (scan_npages--) {
1409                 cond_resched();
1410                 rmap_item = scan_get_next_rmap_item(&page);
1411                 if (!rmap_item)
1412                         return;
1413                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1414                         cmp_and_merge_page(page, rmap_item);
1415                 put_page(page);
1416         }
1417 }
1418
1419 static int ksmd_should_run(void)
1420 {
1421         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1422 }
1423
1424 static int ksm_scan_thread(void *nothing)
1425 {
1426         set_user_nice(current, 5);
1427
1428         while (!kthread_should_stop()) {
1429                 mutex_lock(&ksm_thread_mutex);
1430                 if (ksmd_should_run())
1431                         ksm_do_scan(ksm_thread_pages_to_scan);
1432                 mutex_unlock(&ksm_thread_mutex);
1433
1434                 if (ksmd_should_run()) {
1435                         schedule_timeout_interruptible(
1436                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1437                 } else {
1438                         wait_event_interruptible(ksm_thread_wait,
1439                                 ksmd_should_run() || kthread_should_stop());
1440                 }
1441         }
1442         return 0;
1443 }
1444
1445 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1446                 unsigned long end, int advice, unsigned long *vm_flags)
1447 {
1448         struct mm_struct *mm = vma->vm_mm;
1449         int err;
1450
1451         switch (advice) {
1452         case MADV_MERGEABLE:
1453                 /*
1454                  * Be somewhat over-protective for now!
1455                  */
1456                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1457                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1458                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1459                                  VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1460                         return 0;               /* just ignore the advice */
1461
1462                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1463                         err = __ksm_enter(mm);
1464                         if (err)
1465                                 return err;
1466                 }
1467
1468                 *vm_flags |= VM_MERGEABLE;
1469                 break;
1470
1471         case MADV_UNMERGEABLE:
1472                 if (!(*vm_flags & VM_MERGEABLE))
1473                         return 0;               /* just ignore the advice */
1474
1475                 if (vma->anon_vma) {
1476                         err = unmerge_ksm_pages(vma, start, end);
1477                         if (err)
1478                                 return err;
1479                 }
1480
1481                 *vm_flags &= ~VM_MERGEABLE;
1482                 break;
1483         }
1484
1485         return 0;
1486 }
1487
1488 int __ksm_enter(struct mm_struct *mm)
1489 {
1490         struct mm_slot *mm_slot;
1491         int needs_wakeup;
1492
1493         mm_slot = alloc_mm_slot();
1494         if (!mm_slot)
1495                 return -ENOMEM;
1496
1497         /* Check ksm_run too?  Would need tighter locking */
1498         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1499
1500         spin_lock(&ksm_mmlist_lock);
1501         insert_to_mm_slots_hash(mm, mm_slot);
1502         /*
1503          * Insert just behind the scanning cursor, to let the area settle
1504          * down a little; when fork is followed by immediate exec, we don't
1505          * want ksmd to waste time setting up and tearing down an rmap_list.
1506          */
1507         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1508         spin_unlock(&ksm_mmlist_lock);
1509
1510         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1511         atomic_inc(&mm->mm_count);
1512
1513         if (needs_wakeup)
1514                 wake_up_interruptible(&ksm_thread_wait);
1515
1516         return 0;
1517 }
1518
1519 void __ksm_exit(struct mm_struct *mm)
1520 {
1521         struct mm_slot *mm_slot;
1522         int easy_to_free = 0;
1523
1524         /*
1525          * This process is exiting: if it's straightforward (as is the
1526          * case when ksmd was never running), free mm_slot immediately.
1527          * But if it's at the cursor or has rmap_items linked to it, use
1528          * mmap_sem to synchronize with any break_cows before pagetables
1529          * are freed, and leave the mm_slot on the list for ksmd to free.
1530          * Beware: ksm may already have noticed it exiting and freed the slot.
1531          */
1532
1533         spin_lock(&ksm_mmlist_lock);
1534         mm_slot = get_mm_slot(mm);
1535         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1536                 if (!mm_slot->rmap_list) {
1537                         hlist_del(&mm_slot->link);
1538                         list_del(&mm_slot->mm_list);
1539                         easy_to_free = 1;
1540                 } else {
1541                         list_move(&mm_slot->mm_list,
1542                                   &ksm_scan.mm_slot->mm_list);
1543                 }
1544         }
1545         spin_unlock(&ksm_mmlist_lock);
1546
1547         if (easy_to_free) {
1548                 free_mm_slot(mm_slot);
1549                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1550                 mmdrop(mm);
1551         } else if (mm_slot) {
1552                 down_write(&mm->mmap_sem);
1553                 up_write(&mm->mmap_sem);
1554         }
1555 }
1556
1557 struct page *ksm_does_need_to_copy(struct page *page,
1558                         struct vm_area_struct *vma, unsigned long address)
1559 {
1560         struct page *new_page;
1561
1562         unlock_page(page);      /* any racers will COW it, not modify it */
1563
1564         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1565         if (new_page) {
1566                 copy_user_highpage(new_page, page, address, vma);
1567
1568                 SetPageDirty(new_page);
1569                 __SetPageUptodate(new_page);
1570                 SetPageSwapBacked(new_page);
1571                 __set_page_locked(new_page);
1572
1573                 if (page_evictable(new_page, vma))
1574                         lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1575                 else
1576                         add_page_to_unevictable_list(new_page);
1577         }
1578
1579         page_cache_release(page);
1580         return new_page;
1581 }
1582
1583 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1584                         unsigned long *vm_flags)
1585 {
1586         struct stable_node *stable_node;
1587         struct rmap_item *rmap_item;
1588         struct hlist_node *hlist;
1589         unsigned int mapcount = page_mapcount(page);
1590         int referenced = 0;
1591         int search_new_forks = 0;
1592
1593         VM_BUG_ON(!PageKsm(page));
1594         VM_BUG_ON(!PageLocked(page));
1595
1596         stable_node = page_stable_node(page);
1597         if (!stable_node)
1598                 return 0;
1599 again:
1600         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1601                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1602                 struct vm_area_struct *vma;
1603
1604                 spin_lock(&anon_vma->lock);
1605                 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1606                         if (rmap_item->address < vma->vm_start ||
1607                             rmap_item->address >= vma->vm_end)
1608                                 continue;
1609                         /*
1610                          * Initially we examine only the vma which covers this
1611                          * rmap_item; but later, if there is still work to do,
1612                          * we examine covering vmas in other mms: in case they
1613                          * were forked from the original since ksmd passed.
1614                          */
1615                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1616                                 continue;
1617
1618                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1619                                 continue;
1620
1621                         referenced += page_referenced_one(page, vma,
1622                                 rmap_item->address, &mapcount, vm_flags);
1623                         if (!search_new_forks || !mapcount)
1624                                 break;
1625                 }
1626                 spin_unlock(&anon_vma->lock);
1627                 if (!mapcount)
1628                         goto out;
1629         }
1630         if (!search_new_forks++)
1631                 goto again;
1632 out:
1633         return referenced;
1634 }
1635
1636 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1637 {
1638         struct stable_node *stable_node;
1639         struct hlist_node *hlist;
1640         struct rmap_item *rmap_item;
1641         int ret = SWAP_AGAIN;
1642         int search_new_forks = 0;
1643
1644         VM_BUG_ON(!PageKsm(page));
1645         VM_BUG_ON(!PageLocked(page));
1646
1647         stable_node = page_stable_node(page);
1648         if (!stable_node)
1649                 return SWAP_FAIL;
1650 again:
1651         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1652                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1653                 struct vm_area_struct *vma;
1654
1655                 spin_lock(&anon_vma->lock);
1656                 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1657                         if (rmap_item->address < vma->vm_start ||
1658                             rmap_item->address >= vma->vm_end)
1659                                 continue;
1660                         /*
1661                          * Initially we examine only the vma which covers this
1662                          * rmap_item; but later, if there is still work to do,
1663                          * we examine covering vmas in other mms: in case they
1664                          * were forked from the original since ksmd passed.
1665                          */
1666                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1667                                 continue;
1668
1669                         ret = try_to_unmap_one(page, vma,
1670                                         rmap_item->address, flags);
1671                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1672                                 spin_unlock(&anon_vma->lock);
1673                                 goto out;
1674                         }
1675                 }
1676                 spin_unlock(&anon_vma->lock);
1677         }
1678         if (!search_new_forks++)
1679                 goto again;
1680 out:
1681         return ret;
1682 }
1683
1684 #ifdef CONFIG_SYSFS
1685 /*
1686  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1687  */
1688
1689 #define KSM_ATTR_RO(_name) \
1690         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1691 #define KSM_ATTR(_name) \
1692         static struct kobj_attribute _name##_attr = \
1693                 __ATTR(_name, 0644, _name##_show, _name##_store)
1694
1695 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1696                                     struct kobj_attribute *attr, char *buf)
1697 {
1698         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1699 }
1700
1701 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1702                                      struct kobj_attribute *attr,
1703                                      const char *buf, size_t count)
1704 {
1705         unsigned long msecs;
1706         int err;
1707
1708         err = strict_strtoul(buf, 10, &msecs);
1709         if (err || msecs > UINT_MAX)
1710                 return -EINVAL;
1711
1712         ksm_thread_sleep_millisecs = msecs;
1713
1714         return count;
1715 }
1716 KSM_ATTR(sleep_millisecs);
1717
1718 static ssize_t pages_to_scan_show(struct kobject *kobj,
1719                                   struct kobj_attribute *attr, char *buf)
1720 {
1721         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1722 }
1723
1724 static ssize_t pages_to_scan_store(struct kobject *kobj,
1725                                    struct kobj_attribute *attr,
1726                                    const char *buf, size_t count)
1727 {
1728         int err;
1729         unsigned long nr_pages;
1730
1731         err = strict_strtoul(buf, 10, &nr_pages);
1732         if (err || nr_pages > UINT_MAX)
1733                 return -EINVAL;
1734
1735         ksm_thread_pages_to_scan = nr_pages;
1736
1737         return count;
1738 }
1739 KSM_ATTR(pages_to_scan);
1740
1741 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1742                         char *buf)
1743 {
1744         return sprintf(buf, "%u\n", ksm_run);
1745 }
1746
1747 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1748                          const char *buf, size_t count)
1749 {
1750         int err;
1751         unsigned long flags;
1752
1753         err = strict_strtoul(buf, 10, &flags);
1754         if (err || flags > UINT_MAX)
1755                 return -EINVAL;
1756         if (flags > KSM_RUN_UNMERGE)
1757                 return -EINVAL;
1758
1759         /*
1760          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1761          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1762          * breaking COW to free the unswappable pages_shared (but leaves
1763          * mm_slots on the list for when ksmd may be set running again).
1764          */
1765
1766         mutex_lock(&ksm_thread_mutex);
1767         if (ksm_run != flags) {
1768                 ksm_run = flags;
1769                 if (flags & KSM_RUN_UNMERGE) {
1770                         current->flags |= PF_OOM_ORIGIN;
1771                         err = unmerge_and_remove_all_rmap_items();
1772                         current->flags &= ~PF_OOM_ORIGIN;
1773                         if (err) {
1774                                 ksm_run = KSM_RUN_STOP;
1775                                 count = err;
1776                         }
1777                 }
1778         }
1779         mutex_unlock(&ksm_thread_mutex);
1780
1781         if (flags & KSM_RUN_MERGE)
1782                 wake_up_interruptible(&ksm_thread_wait);
1783
1784         return count;
1785 }
1786 KSM_ATTR(run);
1787
1788 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1789                                       struct kobj_attribute *attr,
1790                                       const char *buf, size_t count)
1791 {
1792         int err;
1793         unsigned long nr_pages;
1794
1795         err = strict_strtoul(buf, 10, &nr_pages);
1796         if (err)
1797                 return -EINVAL;
1798
1799         ksm_max_kernel_pages = nr_pages;
1800
1801         return count;
1802 }
1803
1804 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1805                                      struct kobj_attribute *attr, char *buf)
1806 {
1807         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1808 }
1809 KSM_ATTR(max_kernel_pages);
1810
1811 static ssize_t pages_shared_show(struct kobject *kobj,
1812                                  struct kobj_attribute *attr, char *buf)
1813 {
1814         return sprintf(buf, "%lu\n", ksm_pages_shared);
1815 }
1816 KSM_ATTR_RO(pages_shared);
1817
1818 static ssize_t pages_sharing_show(struct kobject *kobj,
1819                                   struct kobj_attribute *attr, char *buf)
1820 {
1821         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1822 }
1823 KSM_ATTR_RO(pages_sharing);
1824
1825 static ssize_t pages_unshared_show(struct kobject *kobj,
1826                                    struct kobj_attribute *attr, char *buf)
1827 {
1828         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1829 }
1830 KSM_ATTR_RO(pages_unshared);
1831
1832 static ssize_t pages_volatile_show(struct kobject *kobj,
1833                                    struct kobj_attribute *attr, char *buf)
1834 {
1835         long ksm_pages_volatile;
1836
1837         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1838                                 - ksm_pages_sharing - ksm_pages_unshared;
1839         /*
1840          * It was not worth any locking to calculate that statistic,
1841          * but it might therefore sometimes be negative: conceal that.
1842          */
1843         if (ksm_pages_volatile < 0)
1844                 ksm_pages_volatile = 0;
1845         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1846 }
1847 KSM_ATTR_RO(pages_volatile);
1848
1849 static ssize_t full_scans_show(struct kobject *kobj,
1850                                struct kobj_attribute *attr, char *buf)
1851 {
1852         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1853 }
1854 KSM_ATTR_RO(full_scans);
1855
1856 static struct attribute *ksm_attrs[] = {
1857         &sleep_millisecs_attr.attr,
1858         &pages_to_scan_attr.attr,
1859         &run_attr.attr,
1860         &max_kernel_pages_attr.attr,
1861         &pages_shared_attr.attr,
1862         &pages_sharing_attr.attr,
1863         &pages_unshared_attr.attr,
1864         &pages_volatile_attr.attr,
1865         &full_scans_attr.attr,
1866         NULL,
1867 };
1868
1869 static struct attribute_group ksm_attr_group = {
1870         .attrs = ksm_attrs,
1871         .name = "ksm",
1872 };
1873 #endif /* CONFIG_SYSFS */
1874
1875 static int __init ksm_init(void)
1876 {
1877         struct task_struct *ksm_thread;
1878         int err;
1879
1880         ksm_max_kernel_pages = totalram_pages / 4;
1881
1882         err = ksm_slab_init();
1883         if (err)
1884                 goto out;
1885
1886         err = mm_slots_hash_init();
1887         if (err)
1888                 goto out_free1;
1889
1890         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1891         if (IS_ERR(ksm_thread)) {
1892                 printk(KERN_ERR "ksm: creating kthread failed\n");
1893                 err = PTR_ERR(ksm_thread);
1894                 goto out_free2;
1895         }
1896
1897 #ifdef CONFIG_SYSFS
1898         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1899         if (err) {
1900                 printk(KERN_ERR "ksm: register sysfs failed\n");
1901                 kthread_stop(ksm_thread);
1902                 goto out_free2;
1903         }
1904 #else
1905         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
1906
1907 #endif /* CONFIG_SYSFS */
1908
1909         return 0;
1910
1911 out_free2:
1912         mm_slots_hash_free();
1913 out_free1:
1914         ksm_slab_free();
1915 out:
1916         return err;
1917 }
1918 module_init(ksm_init)