thp: freeze khugepaged and ksmd
[linux-3.10.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hash.h>
37 #include <linux/freezer.h>
38
39 #include <asm/tlbflush.h>
40 #include "internal.h"
41
42 /*
43  * A few notes about the KSM scanning process,
44  * to make it easier to understand the data structures below:
45  *
46  * In order to reduce excessive scanning, KSM sorts the memory pages by their
47  * contents into a data structure that holds pointers to the pages' locations.
48  *
49  * Since the contents of the pages may change at any moment, KSM cannot just
50  * insert the pages into a normal sorted tree and expect it to find anything.
51  * Therefore KSM uses two data structures - the stable and the unstable tree.
52  *
53  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
54  * by their contents.  Because each such page is write-protected, searching on
55  * this tree is fully assured to be working (except when pages are unmapped),
56  * and therefore this tree is called the stable tree.
57  *
58  * In addition to the stable tree, KSM uses a second data structure called the
59  * unstable tree: this tree holds pointers to pages which have been found to
60  * be "unchanged for a period of time".  The unstable tree sorts these pages
61  * by their contents, but since they are not write-protected, KSM cannot rely
62  * upon the unstable tree to work correctly - the unstable tree is liable to
63  * be corrupted as its contents are modified, and so it is called unstable.
64  *
65  * KSM solves this problem by several techniques:
66  *
67  * 1) The unstable tree is flushed every time KSM completes scanning all
68  *    memory areas, and then the tree is rebuilt again from the beginning.
69  * 2) KSM will only insert into the unstable tree, pages whose hash value
70  *    has not changed since the previous scan of all memory areas.
71  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
72  *    colors of the nodes and not on their contents, assuring that even when
73  *    the tree gets "corrupted" it won't get out of balance, so scanning time
74  *    remains the same (also, searching and inserting nodes in an rbtree uses
75  *    the same algorithm, so we have no overhead when we flush and rebuild).
76  * 4) KSM never flushes the stable tree, which means that even if it were to
77  *    take 10 attempts to find a page in the unstable tree, once it is found,
78  *    it is secured in the stable tree.  (When we scan a new page, we first
79  *    compare it against the stable tree, and then against the unstable tree.)
80  */
81
82 /**
83  * struct mm_slot - ksm information per mm that is being scanned
84  * @link: link to the mm_slots hash list
85  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
86  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
87  * @mm: the mm that this information is valid for
88  */
89 struct mm_slot {
90         struct hlist_node link;
91         struct list_head mm_list;
92         struct rmap_item *rmap_list;
93         struct mm_struct *mm;
94 };
95
96 /**
97  * struct ksm_scan - cursor for scanning
98  * @mm_slot: the current mm_slot we are scanning
99  * @address: the next address inside that to be scanned
100  * @rmap_list: link to the next rmap to be scanned in the rmap_list
101  * @seqnr: count of completed full scans (needed when removing unstable node)
102  *
103  * There is only the one ksm_scan instance of this cursor structure.
104  */
105 struct ksm_scan {
106         struct mm_slot *mm_slot;
107         unsigned long address;
108         struct rmap_item **rmap_list;
109         unsigned long seqnr;
110 };
111
112 /**
113  * struct stable_node - node of the stable rbtree
114  * @node: rb node of this ksm page in the stable tree
115  * @hlist: hlist head of rmap_items using this ksm page
116  * @kpfn: page frame number of this ksm page
117  */
118 struct stable_node {
119         struct rb_node node;
120         struct hlist_head hlist;
121         unsigned long kpfn;
122 };
123
124 /**
125  * struct rmap_item - reverse mapping item for virtual addresses
126  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
127  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
128  * @mm: the memory structure this rmap_item is pointing into
129  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
130  * @oldchecksum: previous checksum of the page at that virtual address
131  * @node: rb node of this rmap_item in the unstable tree
132  * @head: pointer to stable_node heading this list in the stable tree
133  * @hlist: link into hlist of rmap_items hanging off that stable_node
134  */
135 struct rmap_item {
136         struct rmap_item *rmap_list;
137         struct anon_vma *anon_vma;      /* when stable */
138         struct mm_struct *mm;
139         unsigned long address;          /* + low bits used for flags below */
140         unsigned int oldchecksum;       /* when unstable */
141         union {
142                 struct rb_node node;    /* when node of unstable tree */
143                 struct {                /* when listed from stable tree */
144                         struct stable_node *head;
145                         struct hlist_node hlist;
146                 };
147         };
148 };
149
150 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
151 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
152 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
153
154 /* The stable and unstable tree heads */
155 static struct rb_root root_stable_tree = RB_ROOT;
156 static struct rb_root root_unstable_tree = RB_ROOT;
157
158 #define MM_SLOTS_HASH_SHIFT 10
159 #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
160 static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
161
162 static struct mm_slot ksm_mm_head = {
163         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
164 };
165 static struct ksm_scan ksm_scan = {
166         .mm_slot = &ksm_mm_head,
167 };
168
169 static struct kmem_cache *rmap_item_cache;
170 static struct kmem_cache *stable_node_cache;
171 static struct kmem_cache *mm_slot_cache;
172
173 /* The number of nodes in the stable tree */
174 static unsigned long ksm_pages_shared;
175
176 /* The number of page slots additionally sharing those nodes */
177 static unsigned long ksm_pages_sharing;
178
179 /* The number of nodes in the unstable tree */
180 static unsigned long ksm_pages_unshared;
181
182 /* The number of rmap_items in use: to calculate pages_volatile */
183 static unsigned long ksm_rmap_items;
184
185 /* Number of pages ksmd should scan in one batch */
186 static unsigned int ksm_thread_pages_to_scan = 100;
187
188 /* Milliseconds ksmd should sleep between batches */
189 static unsigned int ksm_thread_sleep_millisecs = 20;
190
191 #define KSM_RUN_STOP    0
192 #define KSM_RUN_MERGE   1
193 #define KSM_RUN_UNMERGE 2
194 static unsigned int ksm_run = KSM_RUN_STOP;
195
196 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
197 static DEFINE_MUTEX(ksm_thread_mutex);
198 static DEFINE_SPINLOCK(ksm_mmlist_lock);
199
200 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
201                 sizeof(struct __struct), __alignof__(struct __struct),\
202                 (__flags), NULL)
203
204 static int __init ksm_slab_init(void)
205 {
206         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
207         if (!rmap_item_cache)
208                 goto out;
209
210         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
211         if (!stable_node_cache)
212                 goto out_free1;
213
214         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
215         if (!mm_slot_cache)
216                 goto out_free2;
217
218         return 0;
219
220 out_free2:
221         kmem_cache_destroy(stable_node_cache);
222 out_free1:
223         kmem_cache_destroy(rmap_item_cache);
224 out:
225         return -ENOMEM;
226 }
227
228 static void __init ksm_slab_free(void)
229 {
230         kmem_cache_destroy(mm_slot_cache);
231         kmem_cache_destroy(stable_node_cache);
232         kmem_cache_destroy(rmap_item_cache);
233         mm_slot_cache = NULL;
234 }
235
236 static inline struct rmap_item *alloc_rmap_item(void)
237 {
238         struct rmap_item *rmap_item;
239
240         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
241         if (rmap_item)
242                 ksm_rmap_items++;
243         return rmap_item;
244 }
245
246 static inline void free_rmap_item(struct rmap_item *rmap_item)
247 {
248         ksm_rmap_items--;
249         rmap_item->mm = NULL;   /* debug safety */
250         kmem_cache_free(rmap_item_cache, rmap_item);
251 }
252
253 static inline struct stable_node *alloc_stable_node(void)
254 {
255         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
256 }
257
258 static inline void free_stable_node(struct stable_node *stable_node)
259 {
260         kmem_cache_free(stable_node_cache, stable_node);
261 }
262
263 static inline struct mm_slot *alloc_mm_slot(void)
264 {
265         if (!mm_slot_cache)     /* initialization failed */
266                 return NULL;
267         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
268 }
269
270 static inline void free_mm_slot(struct mm_slot *mm_slot)
271 {
272         kmem_cache_free(mm_slot_cache, mm_slot);
273 }
274
275 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
276 {
277         struct mm_slot *mm_slot;
278         struct hlist_head *bucket;
279         struct hlist_node *node;
280
281         bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
282         hlist_for_each_entry(mm_slot, node, bucket, link) {
283                 if (mm == mm_slot->mm)
284                         return mm_slot;
285         }
286         return NULL;
287 }
288
289 static void insert_to_mm_slots_hash(struct mm_struct *mm,
290                                     struct mm_slot *mm_slot)
291 {
292         struct hlist_head *bucket;
293
294         bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
295         mm_slot->mm = mm;
296         hlist_add_head(&mm_slot->link, bucket);
297 }
298
299 static inline int in_stable_tree(struct rmap_item *rmap_item)
300 {
301         return rmap_item->address & STABLE_FLAG;
302 }
303
304 static void hold_anon_vma(struct rmap_item *rmap_item,
305                           struct anon_vma *anon_vma)
306 {
307         rmap_item->anon_vma = anon_vma;
308         get_anon_vma(anon_vma);
309 }
310
311 static void ksm_drop_anon_vma(struct rmap_item *rmap_item)
312 {
313         struct anon_vma *anon_vma = rmap_item->anon_vma;
314
315         drop_anon_vma(anon_vma);
316 }
317
318 /*
319  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
320  * page tables after it has passed through ksm_exit() - which, if necessary,
321  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
322  * a special flag: they can just back out as soon as mm_users goes to zero.
323  * ksm_test_exit() is used throughout to make this test for exit: in some
324  * places for correctness, in some places just to avoid unnecessary work.
325  */
326 static inline bool ksm_test_exit(struct mm_struct *mm)
327 {
328         return atomic_read(&mm->mm_users) == 0;
329 }
330
331 /*
332  * We use break_ksm to break COW on a ksm page: it's a stripped down
333  *
334  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
335  *              put_page(page);
336  *
337  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
338  * in case the application has unmapped and remapped mm,addr meanwhile.
339  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
340  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
341  */
342 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
343 {
344         struct page *page;
345         int ret = 0;
346
347         do {
348                 cond_resched();
349                 page = follow_page(vma, addr, FOLL_GET);
350                 if (IS_ERR_OR_NULL(page))
351                         break;
352                 if (PageKsm(page))
353                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
354                                                         FAULT_FLAG_WRITE);
355                 else
356                         ret = VM_FAULT_WRITE;
357                 put_page(page);
358         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
359         /*
360          * We must loop because handle_mm_fault() may back out if there's
361          * any difficulty e.g. if pte accessed bit gets updated concurrently.
362          *
363          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
364          * COW has been broken, even if the vma does not permit VM_WRITE;
365          * but note that a concurrent fault might break PageKsm for us.
366          *
367          * VM_FAULT_SIGBUS could occur if we race with truncation of the
368          * backing file, which also invalidates anonymous pages: that's
369          * okay, that truncation will have unmapped the PageKsm for us.
370          *
371          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
372          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
373          * current task has TIF_MEMDIE set, and will be OOM killed on return
374          * to user; and ksmd, having no mm, would never be chosen for that.
375          *
376          * But if the mm is in a limited mem_cgroup, then the fault may fail
377          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
378          * even ksmd can fail in this way - though it's usually breaking ksm
379          * just to undo a merge it made a moment before, so unlikely to oom.
380          *
381          * That's a pity: we might therefore have more kernel pages allocated
382          * than we're counting as nodes in the stable tree; but ksm_do_scan
383          * will retry to break_cow on each pass, so should recover the page
384          * in due course.  The important thing is to not let VM_MERGEABLE
385          * be cleared while any such pages might remain in the area.
386          */
387         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
388 }
389
390 static void break_cow(struct rmap_item *rmap_item)
391 {
392         struct mm_struct *mm = rmap_item->mm;
393         unsigned long addr = rmap_item->address;
394         struct vm_area_struct *vma;
395
396         /*
397          * It is not an accident that whenever we want to break COW
398          * to undo, we also need to drop a reference to the anon_vma.
399          */
400         ksm_drop_anon_vma(rmap_item);
401
402         down_read(&mm->mmap_sem);
403         if (ksm_test_exit(mm))
404                 goto out;
405         vma = find_vma(mm, addr);
406         if (!vma || vma->vm_start > addr)
407                 goto out;
408         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
409                 goto out;
410         break_ksm(vma, addr);
411 out:
412         up_read(&mm->mmap_sem);
413 }
414
415 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
416 {
417         struct mm_struct *mm = rmap_item->mm;
418         unsigned long addr = rmap_item->address;
419         struct vm_area_struct *vma;
420         struct page *page;
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
431         page = follow_page(vma, addr, FOLL_GET);
432         if (IS_ERR_OR_NULL(page))
433                 goto out;
434         if (PageAnon(page) && !PageTransCompound(page)) {
435                 flush_anon_page(vma, page, addr);
436                 flush_dcache_page(page);
437         } else {
438                 put_page(page);
439 out:            page = NULL;
440         }
441         up_read(&mm->mmap_sem);
442         return page;
443 }
444
445 static void remove_node_from_stable_tree(struct stable_node *stable_node)
446 {
447         struct rmap_item *rmap_item;
448         struct hlist_node *hlist;
449
450         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
451                 if (rmap_item->hlist.next)
452                         ksm_pages_sharing--;
453                 else
454                         ksm_pages_shared--;
455                 ksm_drop_anon_vma(rmap_item);
456                 rmap_item->address &= PAGE_MASK;
457                 cond_resched();
458         }
459
460         rb_erase(&stable_node->node, &root_stable_tree);
461         free_stable_node(stable_node);
462 }
463
464 /*
465  * get_ksm_page: checks if the page indicated by the stable node
466  * is still its ksm page, despite having held no reference to it.
467  * In which case we can trust the content of the page, and it
468  * returns the gotten page; but if the page has now been zapped,
469  * remove the stale node from the stable tree and return NULL.
470  *
471  * You would expect the stable_node to hold a reference to the ksm page.
472  * But if it increments the page's count, swapping out has to wait for
473  * ksmd to come around again before it can free the page, which may take
474  * seconds or even minutes: much too unresponsive.  So instead we use a
475  * "keyhole reference": access to the ksm page from the stable node peeps
476  * out through its keyhole to see if that page still holds the right key,
477  * pointing back to this stable node.  This relies on freeing a PageAnon
478  * page to reset its page->mapping to NULL, and relies on no other use of
479  * a page to put something that might look like our key in page->mapping.
480  *
481  * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
482  * but this is different - made simpler by ksm_thread_mutex being held, but
483  * interesting for assuming that no other use of the struct page could ever
484  * put our expected_mapping into page->mapping (or a field of the union which
485  * coincides with page->mapping).  The RCU calls are not for KSM at all, but
486  * to keep the page_count protocol described with page_cache_get_speculative.
487  *
488  * Note: it is possible that get_ksm_page() will return NULL one moment,
489  * then page the next, if the page is in between page_freeze_refs() and
490  * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
491  * is on its way to being freed; but it is an anomaly to bear in mind.
492  */
493 static struct page *get_ksm_page(struct stable_node *stable_node)
494 {
495         struct page *page;
496         void *expected_mapping;
497
498         page = pfn_to_page(stable_node->kpfn);
499         expected_mapping = (void *)stable_node +
500                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
501         rcu_read_lock();
502         if (page->mapping != expected_mapping)
503                 goto stale;
504         if (!get_page_unless_zero(page))
505                 goto stale;
506         if (page->mapping != expected_mapping) {
507                 put_page(page);
508                 goto stale;
509         }
510         rcu_read_unlock();
511         return page;
512 stale:
513         rcu_read_unlock();
514         remove_node_from_stable_tree(stable_node);
515         return NULL;
516 }
517
518 /*
519  * Removing rmap_item from stable or unstable tree.
520  * This function will clean the information from the stable/unstable tree.
521  */
522 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
523 {
524         if (rmap_item->address & STABLE_FLAG) {
525                 struct stable_node *stable_node;
526                 struct page *page;
527
528                 stable_node = rmap_item->head;
529                 page = get_ksm_page(stable_node);
530                 if (!page)
531                         goto out;
532
533                 lock_page(page);
534                 hlist_del(&rmap_item->hlist);
535                 unlock_page(page);
536                 put_page(page);
537
538                 if (stable_node->hlist.first)
539                         ksm_pages_sharing--;
540                 else
541                         ksm_pages_shared--;
542
543                 ksm_drop_anon_vma(rmap_item);
544                 rmap_item->address &= PAGE_MASK;
545
546         } else if (rmap_item->address & UNSTABLE_FLAG) {
547                 unsigned char age;
548                 /*
549                  * Usually ksmd can and must skip the rb_erase, because
550                  * root_unstable_tree was already reset to RB_ROOT.
551                  * But be careful when an mm is exiting: do the rb_erase
552                  * if this rmap_item was inserted by this scan, rather
553                  * than left over from before.
554                  */
555                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
556                 BUG_ON(age > 1);
557                 if (!age)
558                         rb_erase(&rmap_item->node, &root_unstable_tree);
559
560                 ksm_pages_unshared--;
561                 rmap_item->address &= PAGE_MASK;
562         }
563 out:
564         cond_resched();         /* we're called from many long loops */
565 }
566
567 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
568                                        struct rmap_item **rmap_list)
569 {
570         while (*rmap_list) {
571                 struct rmap_item *rmap_item = *rmap_list;
572                 *rmap_list = rmap_item->rmap_list;
573                 remove_rmap_item_from_tree(rmap_item);
574                 free_rmap_item(rmap_item);
575         }
576 }
577
578 /*
579  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
580  * than check every pte of a given vma, the locking doesn't quite work for
581  * that - an rmap_item is assigned to the stable tree after inserting ksm
582  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
583  * rmap_items from parent to child at fork time (so as not to waste time
584  * if exit comes before the next scan reaches it).
585  *
586  * Similarly, although we'd like to remove rmap_items (so updating counts
587  * and freeing memory) when unmerging an area, it's easier to leave that
588  * to the next pass of ksmd - consider, for example, how ksmd might be
589  * in cmp_and_merge_page on one of the rmap_items we would be removing.
590  */
591 static int unmerge_ksm_pages(struct vm_area_struct *vma,
592                              unsigned long start, unsigned long end)
593 {
594         unsigned long addr;
595         int err = 0;
596
597         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
598                 if (ksm_test_exit(vma->vm_mm))
599                         break;
600                 if (signal_pending(current))
601                         err = -ERESTARTSYS;
602                 else
603                         err = break_ksm(vma, addr);
604         }
605         return err;
606 }
607
608 #ifdef CONFIG_SYSFS
609 /*
610  * Only called through the sysfs control interface:
611  */
612 static int unmerge_and_remove_all_rmap_items(void)
613 {
614         struct mm_slot *mm_slot;
615         struct mm_struct *mm;
616         struct vm_area_struct *vma;
617         int err = 0;
618
619         spin_lock(&ksm_mmlist_lock);
620         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
621                                                 struct mm_slot, mm_list);
622         spin_unlock(&ksm_mmlist_lock);
623
624         for (mm_slot = ksm_scan.mm_slot;
625                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
626                 mm = mm_slot->mm;
627                 down_read(&mm->mmap_sem);
628                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
629                         if (ksm_test_exit(mm))
630                                 break;
631                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
632                                 continue;
633                         err = unmerge_ksm_pages(vma,
634                                                 vma->vm_start, vma->vm_end);
635                         if (err)
636                                 goto error;
637                 }
638
639                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
640
641                 spin_lock(&ksm_mmlist_lock);
642                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
643                                                 struct mm_slot, mm_list);
644                 if (ksm_test_exit(mm)) {
645                         hlist_del(&mm_slot->link);
646                         list_del(&mm_slot->mm_list);
647                         spin_unlock(&ksm_mmlist_lock);
648
649                         free_mm_slot(mm_slot);
650                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
651                         up_read(&mm->mmap_sem);
652                         mmdrop(mm);
653                 } else {
654                         spin_unlock(&ksm_mmlist_lock);
655                         up_read(&mm->mmap_sem);
656                 }
657         }
658
659         ksm_scan.seqnr = 0;
660         return 0;
661
662 error:
663         up_read(&mm->mmap_sem);
664         spin_lock(&ksm_mmlist_lock);
665         ksm_scan.mm_slot = &ksm_mm_head;
666         spin_unlock(&ksm_mmlist_lock);
667         return err;
668 }
669 #endif /* CONFIG_SYSFS */
670
671 static u32 calc_checksum(struct page *page)
672 {
673         u32 checksum;
674         void *addr = kmap_atomic(page, KM_USER0);
675         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
676         kunmap_atomic(addr, KM_USER0);
677         return checksum;
678 }
679
680 static int memcmp_pages(struct page *page1, struct page *page2)
681 {
682         char *addr1, *addr2;
683         int ret;
684
685         addr1 = kmap_atomic(page1, KM_USER0);
686         addr2 = kmap_atomic(page2, KM_USER1);
687         ret = memcmp(addr1, addr2, PAGE_SIZE);
688         kunmap_atomic(addr2, KM_USER1);
689         kunmap_atomic(addr1, KM_USER0);
690         return ret;
691 }
692
693 static inline int pages_identical(struct page *page1, struct page *page2)
694 {
695         return !memcmp_pages(page1, page2);
696 }
697
698 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
699                               pte_t *orig_pte)
700 {
701         struct mm_struct *mm = vma->vm_mm;
702         unsigned long addr;
703         pte_t *ptep;
704         spinlock_t *ptl;
705         int swapped;
706         int err = -EFAULT;
707
708         addr = page_address_in_vma(page, vma);
709         if (addr == -EFAULT)
710                 goto out;
711
712         ptep = page_check_address(page, mm, addr, &ptl, 0);
713         if (!ptep)
714                 goto out;
715
716         if (pte_write(*ptep) || pte_dirty(*ptep)) {
717                 pte_t entry;
718
719                 swapped = PageSwapCache(page);
720                 flush_cache_page(vma, addr, page_to_pfn(page));
721                 /*
722                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
723                  * take any lock, therefore the check that we are going to make
724                  * with the pagecount against the mapcount is racey and
725                  * O_DIRECT can happen right after the check.
726                  * So we clear the pte and flush the tlb before the check
727                  * this assure us that no O_DIRECT can happen after the check
728                  * or in the middle of the check.
729                  */
730                 entry = ptep_clear_flush(vma, addr, ptep);
731                 /*
732                  * Check that no O_DIRECT or similar I/O is in progress on the
733                  * page
734                  */
735                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
736                         set_pte_at(mm, addr, ptep, entry);
737                         goto out_unlock;
738                 }
739                 if (pte_dirty(entry))
740                         set_page_dirty(page);
741                 entry = pte_mkclean(pte_wrprotect(entry));
742                 set_pte_at_notify(mm, addr, ptep, entry);
743         }
744         *orig_pte = *ptep;
745         err = 0;
746
747 out_unlock:
748         pte_unmap_unlock(ptep, ptl);
749 out:
750         return err;
751 }
752
753 /**
754  * replace_page - replace page in vma by new ksm page
755  * @vma:      vma that holds the pte pointing to page
756  * @page:     the page we are replacing by kpage
757  * @kpage:    the ksm page we replace page by
758  * @orig_pte: the original value of the pte
759  *
760  * Returns 0 on success, -EFAULT on failure.
761  */
762 static int replace_page(struct vm_area_struct *vma, struct page *page,
763                         struct page *kpage, pte_t orig_pte)
764 {
765         struct mm_struct *mm = vma->vm_mm;
766         pgd_t *pgd;
767         pud_t *pud;
768         pmd_t *pmd;
769         pte_t *ptep;
770         spinlock_t *ptl;
771         unsigned long addr;
772         int err = -EFAULT;
773
774         addr = page_address_in_vma(page, vma);
775         if (addr == -EFAULT)
776                 goto out;
777
778         pgd = pgd_offset(mm, addr);
779         if (!pgd_present(*pgd))
780                 goto out;
781
782         pud = pud_offset(pgd, addr);
783         if (!pud_present(*pud))
784                 goto out;
785
786         pmd = pmd_offset(pud, addr);
787         if (!pmd_present(*pmd))
788                 goto out;
789
790         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
791         if (!pte_same(*ptep, orig_pte)) {
792                 pte_unmap_unlock(ptep, ptl);
793                 goto out;
794         }
795
796         get_page(kpage);
797         page_add_anon_rmap(kpage, vma, addr);
798
799         flush_cache_page(vma, addr, pte_pfn(*ptep));
800         ptep_clear_flush(vma, addr, ptep);
801         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
802
803         page_remove_rmap(page);
804         if (!page_mapped(page))
805                 try_to_free_swap(page);
806         put_page(page);
807
808         pte_unmap_unlock(ptep, ptl);
809         err = 0;
810 out:
811         return err;
812 }
813
814 /*
815  * try_to_merge_one_page - take two pages and merge them into one
816  * @vma: the vma that holds the pte pointing to page
817  * @page: the PageAnon page that we want to replace with kpage
818  * @kpage: the PageKsm page that we want to map instead of page,
819  *         or NULL the first time when we want to use page as kpage.
820  *
821  * This function returns 0 if the pages were merged, -EFAULT otherwise.
822  */
823 static int try_to_merge_one_page(struct vm_area_struct *vma,
824                                  struct page *page, struct page *kpage)
825 {
826         pte_t orig_pte = __pte(0);
827         int err = -EFAULT;
828
829         if (page == kpage)                      /* ksm page forked */
830                 return 0;
831
832         if (!(vma->vm_flags & VM_MERGEABLE))
833                 goto out;
834         if (!PageAnon(page))
835                 goto out;
836
837         /*
838          * We need the page lock to read a stable PageSwapCache in
839          * write_protect_page().  We use trylock_page() instead of
840          * lock_page() because we don't want to wait here - we
841          * prefer to continue scanning and merging different pages,
842          * then come back to this page when it is unlocked.
843          */
844         if (!trylock_page(page))
845                 goto out;
846         /*
847          * If this anonymous page is mapped only here, its pte may need
848          * to be write-protected.  If it's mapped elsewhere, all of its
849          * ptes are necessarily already write-protected.  But in either
850          * case, we need to lock and check page_count is not raised.
851          */
852         if (write_protect_page(vma, page, &orig_pte) == 0) {
853                 if (!kpage) {
854                         /*
855                          * While we hold page lock, upgrade page from
856                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
857                          * stable_tree_insert() will update stable_node.
858                          */
859                         set_page_stable_node(page, NULL);
860                         mark_page_accessed(page);
861                         err = 0;
862                 } else if (pages_identical(page, kpage))
863                         err = replace_page(vma, page, kpage, orig_pte);
864         }
865
866         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
867                 munlock_vma_page(page);
868                 if (!PageMlocked(kpage)) {
869                         unlock_page(page);
870                         lock_page(kpage);
871                         mlock_vma_page(kpage);
872                         page = kpage;           /* for final unlock */
873                 }
874         }
875
876         unlock_page(page);
877 out:
878         return err;
879 }
880
881 /*
882  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
883  * but no new kernel page is allocated: kpage must already be a ksm page.
884  *
885  * This function returns 0 if the pages were merged, -EFAULT otherwise.
886  */
887 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
888                                       struct page *page, struct page *kpage)
889 {
890         struct mm_struct *mm = rmap_item->mm;
891         struct vm_area_struct *vma;
892         int err = -EFAULT;
893
894         down_read(&mm->mmap_sem);
895         if (ksm_test_exit(mm))
896                 goto out;
897         vma = find_vma(mm, rmap_item->address);
898         if (!vma || vma->vm_start > rmap_item->address)
899                 goto out;
900
901         err = try_to_merge_one_page(vma, page, kpage);
902         if (err)
903                 goto out;
904
905         /* Must get reference to anon_vma while still holding mmap_sem */
906         hold_anon_vma(rmap_item, vma->anon_vma);
907 out:
908         up_read(&mm->mmap_sem);
909         return err;
910 }
911
912 /*
913  * try_to_merge_two_pages - take two identical pages and prepare them
914  * to be merged into one page.
915  *
916  * This function returns the kpage if we successfully merged two identical
917  * pages into one ksm page, NULL otherwise.
918  *
919  * Note that this function upgrades page to ksm page: if one of the pages
920  * is already a ksm page, try_to_merge_with_ksm_page should be used.
921  */
922 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
923                                            struct page *page,
924                                            struct rmap_item *tree_rmap_item,
925                                            struct page *tree_page)
926 {
927         int err;
928
929         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
930         if (!err) {
931                 err = try_to_merge_with_ksm_page(tree_rmap_item,
932                                                         tree_page, page);
933                 /*
934                  * If that fails, we have a ksm page with only one pte
935                  * pointing to it: so break it.
936                  */
937                 if (err)
938                         break_cow(rmap_item);
939         }
940         return err ? NULL : page;
941 }
942
943 /*
944  * stable_tree_search - search for page inside the stable tree
945  *
946  * This function checks if there is a page inside the stable tree
947  * with identical content to the page that we are scanning right now.
948  *
949  * This function returns the stable tree node of identical content if found,
950  * NULL otherwise.
951  */
952 static struct page *stable_tree_search(struct page *page)
953 {
954         struct rb_node *node = root_stable_tree.rb_node;
955         struct stable_node *stable_node;
956
957         stable_node = page_stable_node(page);
958         if (stable_node) {                      /* ksm page forked */
959                 get_page(page);
960                 return page;
961         }
962
963         while (node) {
964                 struct page *tree_page;
965                 int ret;
966
967                 cond_resched();
968                 stable_node = rb_entry(node, struct stable_node, node);
969                 tree_page = get_ksm_page(stable_node);
970                 if (!tree_page)
971                         return NULL;
972
973                 ret = memcmp_pages(page, tree_page);
974
975                 if (ret < 0) {
976                         put_page(tree_page);
977                         node = node->rb_left;
978                 } else if (ret > 0) {
979                         put_page(tree_page);
980                         node = node->rb_right;
981                 } else
982                         return tree_page;
983         }
984
985         return NULL;
986 }
987
988 /*
989  * stable_tree_insert - insert rmap_item pointing to new ksm page
990  * into the stable tree.
991  *
992  * This function returns the stable tree node just allocated on success,
993  * NULL otherwise.
994  */
995 static struct stable_node *stable_tree_insert(struct page *kpage)
996 {
997         struct rb_node **new = &root_stable_tree.rb_node;
998         struct rb_node *parent = NULL;
999         struct stable_node *stable_node;
1000
1001         while (*new) {
1002                 struct page *tree_page;
1003                 int ret;
1004
1005                 cond_resched();
1006                 stable_node = rb_entry(*new, struct stable_node, node);
1007                 tree_page = get_ksm_page(stable_node);
1008                 if (!tree_page)
1009                         return NULL;
1010
1011                 ret = memcmp_pages(kpage, tree_page);
1012                 put_page(tree_page);
1013
1014                 parent = *new;
1015                 if (ret < 0)
1016                         new = &parent->rb_left;
1017                 else if (ret > 0)
1018                         new = &parent->rb_right;
1019                 else {
1020                         /*
1021                          * It is not a bug that stable_tree_search() didn't
1022                          * find this node: because at that time our page was
1023                          * not yet write-protected, so may have changed since.
1024                          */
1025                         return NULL;
1026                 }
1027         }
1028
1029         stable_node = alloc_stable_node();
1030         if (!stable_node)
1031                 return NULL;
1032
1033         rb_link_node(&stable_node->node, parent, new);
1034         rb_insert_color(&stable_node->node, &root_stable_tree);
1035
1036         INIT_HLIST_HEAD(&stable_node->hlist);
1037
1038         stable_node->kpfn = page_to_pfn(kpage);
1039         set_page_stable_node(kpage, stable_node);
1040
1041         return stable_node;
1042 }
1043
1044 /*
1045  * unstable_tree_search_insert - search for identical page,
1046  * else insert rmap_item into the unstable tree.
1047  *
1048  * This function searches for a page in the unstable tree identical to the
1049  * page currently being scanned; and if no identical page is found in the
1050  * tree, we insert rmap_item as a new object into the unstable tree.
1051  *
1052  * This function returns pointer to rmap_item found to be identical
1053  * to the currently scanned page, NULL otherwise.
1054  *
1055  * This function does both searching and inserting, because they share
1056  * the same walking algorithm in an rbtree.
1057  */
1058 static
1059 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1060                                               struct page *page,
1061                                               struct page **tree_pagep)
1062
1063 {
1064         struct rb_node **new = &root_unstable_tree.rb_node;
1065         struct rb_node *parent = NULL;
1066
1067         while (*new) {
1068                 struct rmap_item *tree_rmap_item;
1069                 struct page *tree_page;
1070                 int ret;
1071
1072                 cond_resched();
1073                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1074                 tree_page = get_mergeable_page(tree_rmap_item);
1075                 if (IS_ERR_OR_NULL(tree_page))
1076                         return NULL;
1077
1078                 /*
1079                  * Don't substitute a ksm page for a forked page.
1080                  */
1081                 if (page == tree_page) {
1082                         put_page(tree_page);
1083                         return NULL;
1084                 }
1085
1086                 ret = memcmp_pages(page, tree_page);
1087
1088                 parent = *new;
1089                 if (ret < 0) {
1090                         put_page(tree_page);
1091                         new = &parent->rb_left;
1092                 } else if (ret > 0) {
1093                         put_page(tree_page);
1094                         new = &parent->rb_right;
1095                 } else {
1096                         *tree_pagep = tree_page;
1097                         return tree_rmap_item;
1098                 }
1099         }
1100
1101         rmap_item->address |= UNSTABLE_FLAG;
1102         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1103         rb_link_node(&rmap_item->node, parent, new);
1104         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1105
1106         ksm_pages_unshared++;
1107         return NULL;
1108 }
1109
1110 /*
1111  * stable_tree_append - add another rmap_item to the linked list of
1112  * rmap_items hanging off a given node of the stable tree, all sharing
1113  * the same ksm page.
1114  */
1115 static void stable_tree_append(struct rmap_item *rmap_item,
1116                                struct stable_node *stable_node)
1117 {
1118         rmap_item->head = stable_node;
1119         rmap_item->address |= STABLE_FLAG;
1120         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1121
1122         if (rmap_item->hlist.next)
1123                 ksm_pages_sharing++;
1124         else
1125                 ksm_pages_shared++;
1126 }
1127
1128 /*
1129  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1130  * if not, compare checksum to previous and if it's the same, see if page can
1131  * be inserted into the unstable tree, or merged with a page already there and
1132  * both transferred to the stable tree.
1133  *
1134  * @page: the page that we are searching identical page to.
1135  * @rmap_item: the reverse mapping into the virtual address of this page
1136  */
1137 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1138 {
1139         struct rmap_item *tree_rmap_item;
1140         struct page *tree_page = NULL;
1141         struct stable_node *stable_node;
1142         struct page *kpage;
1143         unsigned int checksum;
1144         int err;
1145
1146         remove_rmap_item_from_tree(rmap_item);
1147
1148         /* We first start with searching the page inside the stable tree */
1149         kpage = stable_tree_search(page);
1150         if (kpage) {
1151                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1152                 if (!err) {
1153                         /*
1154                          * The page was successfully merged:
1155                          * add its rmap_item to the stable tree.
1156                          */
1157                         lock_page(kpage);
1158                         stable_tree_append(rmap_item, page_stable_node(kpage));
1159                         unlock_page(kpage);
1160                 }
1161                 put_page(kpage);
1162                 return;
1163         }
1164
1165         /*
1166          * If the hash value of the page has changed from the last time
1167          * we calculated it, this page is changing frequently: therefore we
1168          * don't want to insert it in the unstable tree, and we don't want
1169          * to waste our time searching for something identical to it there.
1170          */
1171         checksum = calc_checksum(page);
1172         if (rmap_item->oldchecksum != checksum) {
1173                 rmap_item->oldchecksum = checksum;
1174                 return;
1175         }
1176
1177         tree_rmap_item =
1178                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1179         if (tree_rmap_item) {
1180                 kpage = try_to_merge_two_pages(rmap_item, page,
1181                                                 tree_rmap_item, tree_page);
1182                 put_page(tree_page);
1183                 /*
1184                  * As soon as we merge this page, we want to remove the
1185                  * rmap_item of the page we have merged with from the unstable
1186                  * tree, and insert it instead as new node in the stable tree.
1187                  */
1188                 if (kpage) {
1189                         remove_rmap_item_from_tree(tree_rmap_item);
1190
1191                         lock_page(kpage);
1192                         stable_node = stable_tree_insert(kpage);
1193                         if (stable_node) {
1194                                 stable_tree_append(tree_rmap_item, stable_node);
1195                                 stable_tree_append(rmap_item, stable_node);
1196                         }
1197                         unlock_page(kpage);
1198
1199                         /*
1200                          * If we fail to insert the page into the stable tree,
1201                          * we will have 2 virtual addresses that are pointing
1202                          * to a ksm page left outside the stable tree,
1203                          * in which case we need to break_cow on both.
1204                          */
1205                         if (!stable_node) {
1206                                 break_cow(tree_rmap_item);
1207                                 break_cow(rmap_item);
1208                         }
1209                 }
1210         }
1211 }
1212
1213 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1214                                             struct rmap_item **rmap_list,
1215                                             unsigned long addr)
1216 {
1217         struct rmap_item *rmap_item;
1218
1219         while (*rmap_list) {
1220                 rmap_item = *rmap_list;
1221                 if ((rmap_item->address & PAGE_MASK) == addr)
1222                         return rmap_item;
1223                 if (rmap_item->address > addr)
1224                         break;
1225                 *rmap_list = rmap_item->rmap_list;
1226                 remove_rmap_item_from_tree(rmap_item);
1227                 free_rmap_item(rmap_item);
1228         }
1229
1230         rmap_item = alloc_rmap_item();
1231         if (rmap_item) {
1232                 /* It has already been zeroed */
1233                 rmap_item->mm = mm_slot->mm;
1234                 rmap_item->address = addr;
1235                 rmap_item->rmap_list = *rmap_list;
1236                 *rmap_list = rmap_item;
1237         }
1238         return rmap_item;
1239 }
1240
1241 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1242 {
1243         struct mm_struct *mm;
1244         struct mm_slot *slot;
1245         struct vm_area_struct *vma;
1246         struct rmap_item *rmap_item;
1247
1248         if (list_empty(&ksm_mm_head.mm_list))
1249                 return NULL;
1250
1251         slot = ksm_scan.mm_slot;
1252         if (slot == &ksm_mm_head) {
1253                 root_unstable_tree = RB_ROOT;
1254
1255                 spin_lock(&ksm_mmlist_lock);
1256                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1257                 ksm_scan.mm_slot = slot;
1258                 spin_unlock(&ksm_mmlist_lock);
1259 next_mm:
1260                 ksm_scan.address = 0;
1261                 ksm_scan.rmap_list = &slot->rmap_list;
1262         }
1263
1264         mm = slot->mm;
1265         down_read(&mm->mmap_sem);
1266         if (ksm_test_exit(mm))
1267                 vma = NULL;
1268         else
1269                 vma = find_vma(mm, ksm_scan.address);
1270
1271         for (; vma; vma = vma->vm_next) {
1272                 if (!(vma->vm_flags & VM_MERGEABLE))
1273                         continue;
1274                 if (ksm_scan.address < vma->vm_start)
1275                         ksm_scan.address = vma->vm_start;
1276                 if (!vma->anon_vma)
1277                         ksm_scan.address = vma->vm_end;
1278
1279                 while (ksm_scan.address < vma->vm_end) {
1280                         if (ksm_test_exit(mm))
1281                                 break;
1282                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1283                         if (IS_ERR_OR_NULL(*page)) {
1284                                 ksm_scan.address += PAGE_SIZE;
1285                                 cond_resched();
1286                                 continue;
1287                         }
1288                         if (PageTransCompound(*page)) {
1289                                 put_page(*page);
1290                                 ksm_scan.address &= HPAGE_PMD_MASK;
1291                                 ksm_scan.address += HPAGE_PMD_SIZE;
1292                                 cond_resched();
1293                                 continue;
1294                         }
1295                         if (PageAnon(*page)) {
1296                                 flush_anon_page(vma, *page, ksm_scan.address);
1297                                 flush_dcache_page(*page);
1298                                 rmap_item = get_next_rmap_item(slot,
1299                                         ksm_scan.rmap_list, ksm_scan.address);
1300                                 if (rmap_item) {
1301                                         ksm_scan.rmap_list =
1302                                                         &rmap_item->rmap_list;
1303                                         ksm_scan.address += PAGE_SIZE;
1304                                 } else
1305                                         put_page(*page);
1306                                 up_read(&mm->mmap_sem);
1307                                 return rmap_item;
1308                         }
1309                         put_page(*page);
1310                         ksm_scan.address += PAGE_SIZE;
1311                         cond_resched();
1312                 }
1313         }
1314
1315         if (ksm_test_exit(mm)) {
1316                 ksm_scan.address = 0;
1317                 ksm_scan.rmap_list = &slot->rmap_list;
1318         }
1319         /*
1320          * Nuke all the rmap_items that are above this current rmap:
1321          * because there were no VM_MERGEABLE vmas with such addresses.
1322          */
1323         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1324
1325         spin_lock(&ksm_mmlist_lock);
1326         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1327                                                 struct mm_slot, mm_list);
1328         if (ksm_scan.address == 0) {
1329                 /*
1330                  * We've completed a full scan of all vmas, holding mmap_sem
1331                  * throughout, and found no VM_MERGEABLE: so do the same as
1332                  * __ksm_exit does to remove this mm from all our lists now.
1333                  * This applies either when cleaning up after __ksm_exit
1334                  * (but beware: we can reach here even before __ksm_exit),
1335                  * or when all VM_MERGEABLE areas have been unmapped (and
1336                  * mmap_sem then protects against race with MADV_MERGEABLE).
1337                  */
1338                 hlist_del(&slot->link);
1339                 list_del(&slot->mm_list);
1340                 spin_unlock(&ksm_mmlist_lock);
1341
1342                 free_mm_slot(slot);
1343                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1344                 up_read(&mm->mmap_sem);
1345                 mmdrop(mm);
1346         } else {
1347                 spin_unlock(&ksm_mmlist_lock);
1348                 up_read(&mm->mmap_sem);
1349         }
1350
1351         /* Repeat until we've completed scanning the whole list */
1352         slot = ksm_scan.mm_slot;
1353         if (slot != &ksm_mm_head)
1354                 goto next_mm;
1355
1356         ksm_scan.seqnr++;
1357         return NULL;
1358 }
1359
1360 /**
1361  * ksm_do_scan  - the ksm scanner main worker function.
1362  * @scan_npages - number of pages we want to scan before we return.
1363  */
1364 static void ksm_do_scan(unsigned int scan_npages)
1365 {
1366         struct rmap_item *rmap_item;
1367         struct page *uninitialized_var(page);
1368
1369         while (scan_npages-- && likely(!freezing(current))) {
1370                 cond_resched();
1371                 rmap_item = scan_get_next_rmap_item(&page);
1372                 if (!rmap_item)
1373                         return;
1374                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1375                         cmp_and_merge_page(page, rmap_item);
1376                 put_page(page);
1377         }
1378 }
1379
1380 static int ksmd_should_run(void)
1381 {
1382         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1383 }
1384
1385 static int ksm_scan_thread(void *nothing)
1386 {
1387         set_freezable();
1388         set_user_nice(current, 5);
1389
1390         while (!kthread_should_stop()) {
1391                 mutex_lock(&ksm_thread_mutex);
1392                 if (ksmd_should_run())
1393                         ksm_do_scan(ksm_thread_pages_to_scan);
1394                 mutex_unlock(&ksm_thread_mutex);
1395
1396                 try_to_freeze();
1397
1398                 if (ksmd_should_run()) {
1399                         schedule_timeout_interruptible(
1400                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1401                 } else {
1402                         wait_event_freezable(ksm_thread_wait,
1403                                 ksmd_should_run() || kthread_should_stop());
1404                 }
1405         }
1406         return 0;
1407 }
1408
1409 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1410                 unsigned long end, int advice, unsigned long *vm_flags)
1411 {
1412         struct mm_struct *mm = vma->vm_mm;
1413         int err;
1414
1415         switch (advice) {
1416         case MADV_MERGEABLE:
1417                 /*
1418                  * Be somewhat over-protective for now!
1419                  */
1420                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1421                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1422                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1423                                  VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1424                         return 0;               /* just ignore the advice */
1425
1426                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1427                         err = __ksm_enter(mm);
1428                         if (err)
1429                                 return err;
1430                 }
1431
1432                 *vm_flags |= VM_MERGEABLE;
1433                 break;
1434
1435         case MADV_UNMERGEABLE:
1436                 if (!(*vm_flags & VM_MERGEABLE))
1437                         return 0;               /* just ignore the advice */
1438
1439                 if (vma->anon_vma) {
1440                         err = unmerge_ksm_pages(vma, start, end);
1441                         if (err)
1442                                 return err;
1443                 }
1444
1445                 *vm_flags &= ~VM_MERGEABLE;
1446                 break;
1447         }
1448
1449         return 0;
1450 }
1451
1452 int __ksm_enter(struct mm_struct *mm)
1453 {
1454         struct mm_slot *mm_slot;
1455         int needs_wakeup;
1456
1457         mm_slot = alloc_mm_slot();
1458         if (!mm_slot)
1459                 return -ENOMEM;
1460
1461         /* Check ksm_run too?  Would need tighter locking */
1462         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1463
1464         spin_lock(&ksm_mmlist_lock);
1465         insert_to_mm_slots_hash(mm, mm_slot);
1466         /*
1467          * Insert just behind the scanning cursor, to let the area settle
1468          * down a little; when fork is followed by immediate exec, we don't
1469          * want ksmd to waste time setting up and tearing down an rmap_list.
1470          */
1471         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1472         spin_unlock(&ksm_mmlist_lock);
1473
1474         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1475         atomic_inc(&mm->mm_count);
1476
1477         if (needs_wakeup)
1478                 wake_up_interruptible(&ksm_thread_wait);
1479
1480         return 0;
1481 }
1482
1483 void __ksm_exit(struct mm_struct *mm)
1484 {
1485         struct mm_slot *mm_slot;
1486         int easy_to_free = 0;
1487
1488         /*
1489          * This process is exiting: if it's straightforward (as is the
1490          * case when ksmd was never running), free mm_slot immediately.
1491          * But if it's at the cursor or has rmap_items linked to it, use
1492          * mmap_sem to synchronize with any break_cows before pagetables
1493          * are freed, and leave the mm_slot on the list for ksmd to free.
1494          * Beware: ksm may already have noticed it exiting and freed the slot.
1495          */
1496
1497         spin_lock(&ksm_mmlist_lock);
1498         mm_slot = get_mm_slot(mm);
1499         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1500                 if (!mm_slot->rmap_list) {
1501                         hlist_del(&mm_slot->link);
1502                         list_del(&mm_slot->mm_list);
1503                         easy_to_free = 1;
1504                 } else {
1505                         list_move(&mm_slot->mm_list,
1506                                   &ksm_scan.mm_slot->mm_list);
1507                 }
1508         }
1509         spin_unlock(&ksm_mmlist_lock);
1510
1511         if (easy_to_free) {
1512                 free_mm_slot(mm_slot);
1513                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1514                 mmdrop(mm);
1515         } else if (mm_slot) {
1516                 down_write(&mm->mmap_sem);
1517                 up_write(&mm->mmap_sem);
1518         }
1519 }
1520
1521 struct page *ksm_does_need_to_copy(struct page *page,
1522                         struct vm_area_struct *vma, unsigned long address)
1523 {
1524         struct page *new_page;
1525
1526         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1527         if (new_page) {
1528                 copy_user_highpage(new_page, page, address, vma);
1529
1530                 SetPageDirty(new_page);
1531                 __SetPageUptodate(new_page);
1532                 SetPageSwapBacked(new_page);
1533                 __set_page_locked(new_page);
1534
1535                 if (page_evictable(new_page, vma))
1536                         lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1537                 else
1538                         add_page_to_unevictable_list(new_page);
1539         }
1540
1541         return new_page;
1542 }
1543
1544 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1545                         unsigned long *vm_flags)
1546 {
1547         struct stable_node *stable_node;
1548         struct rmap_item *rmap_item;
1549         struct hlist_node *hlist;
1550         unsigned int mapcount = page_mapcount(page);
1551         int referenced = 0;
1552         int search_new_forks = 0;
1553
1554         VM_BUG_ON(!PageKsm(page));
1555         VM_BUG_ON(!PageLocked(page));
1556
1557         stable_node = page_stable_node(page);
1558         if (!stable_node)
1559                 return 0;
1560 again:
1561         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1562                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1563                 struct anon_vma_chain *vmac;
1564                 struct vm_area_struct *vma;
1565
1566                 anon_vma_lock(anon_vma);
1567                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1568                         vma = vmac->vma;
1569                         if (rmap_item->address < vma->vm_start ||
1570                             rmap_item->address >= vma->vm_end)
1571                                 continue;
1572                         /*
1573                          * Initially we examine only the vma which covers this
1574                          * rmap_item; but later, if there is still work to do,
1575                          * we examine covering vmas in other mms: in case they
1576                          * were forked from the original since ksmd passed.
1577                          */
1578                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1579                                 continue;
1580
1581                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1582                                 continue;
1583
1584                         referenced += page_referenced_one(page, vma,
1585                                 rmap_item->address, &mapcount, vm_flags);
1586                         if (!search_new_forks || !mapcount)
1587                                 break;
1588                 }
1589                 anon_vma_unlock(anon_vma);
1590                 if (!mapcount)
1591                         goto out;
1592         }
1593         if (!search_new_forks++)
1594                 goto again;
1595 out:
1596         return referenced;
1597 }
1598
1599 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1600 {
1601         struct stable_node *stable_node;
1602         struct hlist_node *hlist;
1603         struct rmap_item *rmap_item;
1604         int ret = SWAP_AGAIN;
1605         int search_new_forks = 0;
1606
1607         VM_BUG_ON(!PageKsm(page));
1608         VM_BUG_ON(!PageLocked(page));
1609
1610         stable_node = page_stable_node(page);
1611         if (!stable_node)
1612                 return SWAP_FAIL;
1613 again:
1614         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1615                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1616                 struct anon_vma_chain *vmac;
1617                 struct vm_area_struct *vma;
1618
1619                 anon_vma_lock(anon_vma);
1620                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1621                         vma = vmac->vma;
1622                         if (rmap_item->address < vma->vm_start ||
1623                             rmap_item->address >= vma->vm_end)
1624                                 continue;
1625                         /*
1626                          * Initially we examine only the vma which covers this
1627                          * rmap_item; but later, if there is still work to do,
1628                          * we examine covering vmas in other mms: in case they
1629                          * were forked from the original since ksmd passed.
1630                          */
1631                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1632                                 continue;
1633
1634                         ret = try_to_unmap_one(page, vma,
1635                                         rmap_item->address, flags);
1636                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1637                                 anon_vma_unlock(anon_vma);
1638                                 goto out;
1639                         }
1640                 }
1641                 anon_vma_unlock(anon_vma);
1642         }
1643         if (!search_new_forks++)
1644                 goto again;
1645 out:
1646         return ret;
1647 }
1648
1649 #ifdef CONFIG_MIGRATION
1650 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1651                   struct vm_area_struct *, unsigned long, void *), void *arg)
1652 {
1653         struct stable_node *stable_node;
1654         struct hlist_node *hlist;
1655         struct rmap_item *rmap_item;
1656         int ret = SWAP_AGAIN;
1657         int search_new_forks = 0;
1658
1659         VM_BUG_ON(!PageKsm(page));
1660         VM_BUG_ON(!PageLocked(page));
1661
1662         stable_node = page_stable_node(page);
1663         if (!stable_node)
1664                 return ret;
1665 again:
1666         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1667                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1668                 struct anon_vma_chain *vmac;
1669                 struct vm_area_struct *vma;
1670
1671                 anon_vma_lock(anon_vma);
1672                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1673                         vma = vmac->vma;
1674                         if (rmap_item->address < vma->vm_start ||
1675                             rmap_item->address >= vma->vm_end)
1676                                 continue;
1677                         /*
1678                          * Initially we examine only the vma which covers this
1679                          * rmap_item; but later, if there is still work to do,
1680                          * we examine covering vmas in other mms: in case they
1681                          * were forked from the original since ksmd passed.
1682                          */
1683                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1684                                 continue;
1685
1686                         ret = rmap_one(page, vma, rmap_item->address, arg);
1687                         if (ret != SWAP_AGAIN) {
1688                                 anon_vma_unlock(anon_vma);
1689                                 goto out;
1690                         }
1691                 }
1692                 anon_vma_unlock(anon_vma);
1693         }
1694         if (!search_new_forks++)
1695                 goto again;
1696 out:
1697         return ret;
1698 }
1699
1700 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1701 {
1702         struct stable_node *stable_node;
1703
1704         VM_BUG_ON(!PageLocked(oldpage));
1705         VM_BUG_ON(!PageLocked(newpage));
1706         VM_BUG_ON(newpage->mapping != oldpage->mapping);
1707
1708         stable_node = page_stable_node(newpage);
1709         if (stable_node) {
1710                 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1711                 stable_node->kpfn = page_to_pfn(newpage);
1712         }
1713 }
1714 #endif /* CONFIG_MIGRATION */
1715
1716 #ifdef CONFIG_MEMORY_HOTREMOVE
1717 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1718                                                  unsigned long end_pfn)
1719 {
1720         struct rb_node *node;
1721
1722         for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1723                 struct stable_node *stable_node;
1724
1725                 stable_node = rb_entry(node, struct stable_node, node);
1726                 if (stable_node->kpfn >= start_pfn &&
1727                     stable_node->kpfn < end_pfn)
1728                         return stable_node;
1729         }
1730         return NULL;
1731 }
1732
1733 static int ksm_memory_callback(struct notifier_block *self,
1734                                unsigned long action, void *arg)
1735 {
1736         struct memory_notify *mn = arg;
1737         struct stable_node *stable_node;
1738
1739         switch (action) {
1740         case MEM_GOING_OFFLINE:
1741                 /*
1742                  * Keep it very simple for now: just lock out ksmd and
1743                  * MADV_UNMERGEABLE while any memory is going offline.
1744                  * mutex_lock_nested() is necessary because lockdep was alarmed
1745                  * that here we take ksm_thread_mutex inside notifier chain
1746                  * mutex, and later take notifier chain mutex inside
1747                  * ksm_thread_mutex to unlock it.   But that's safe because both
1748                  * are inside mem_hotplug_mutex.
1749                  */
1750                 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1751                 break;
1752
1753         case MEM_OFFLINE:
1754                 /*
1755                  * Most of the work is done by page migration; but there might
1756                  * be a few stable_nodes left over, still pointing to struct
1757                  * pages which have been offlined: prune those from the tree.
1758                  */
1759                 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1760                                         mn->start_pfn + mn->nr_pages)) != NULL)
1761                         remove_node_from_stable_tree(stable_node);
1762                 /* fallthrough */
1763
1764         case MEM_CANCEL_OFFLINE:
1765                 mutex_unlock(&ksm_thread_mutex);
1766                 break;
1767         }
1768         return NOTIFY_OK;
1769 }
1770 #endif /* CONFIG_MEMORY_HOTREMOVE */
1771
1772 #ifdef CONFIG_SYSFS
1773 /*
1774  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1775  */
1776
1777 #define KSM_ATTR_RO(_name) \
1778         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1779 #define KSM_ATTR(_name) \
1780         static struct kobj_attribute _name##_attr = \
1781                 __ATTR(_name, 0644, _name##_show, _name##_store)
1782
1783 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1784                                     struct kobj_attribute *attr, char *buf)
1785 {
1786         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1787 }
1788
1789 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1790                                      struct kobj_attribute *attr,
1791                                      const char *buf, size_t count)
1792 {
1793         unsigned long msecs;
1794         int err;
1795
1796         err = strict_strtoul(buf, 10, &msecs);
1797         if (err || msecs > UINT_MAX)
1798                 return -EINVAL;
1799
1800         ksm_thread_sleep_millisecs = msecs;
1801
1802         return count;
1803 }
1804 KSM_ATTR(sleep_millisecs);
1805
1806 static ssize_t pages_to_scan_show(struct kobject *kobj,
1807                                   struct kobj_attribute *attr, char *buf)
1808 {
1809         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1810 }
1811
1812 static ssize_t pages_to_scan_store(struct kobject *kobj,
1813                                    struct kobj_attribute *attr,
1814                                    const char *buf, size_t count)
1815 {
1816         int err;
1817         unsigned long nr_pages;
1818
1819         err = strict_strtoul(buf, 10, &nr_pages);
1820         if (err || nr_pages > UINT_MAX)
1821                 return -EINVAL;
1822
1823         ksm_thread_pages_to_scan = nr_pages;
1824
1825         return count;
1826 }
1827 KSM_ATTR(pages_to_scan);
1828
1829 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1830                         char *buf)
1831 {
1832         return sprintf(buf, "%u\n", ksm_run);
1833 }
1834
1835 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1836                          const char *buf, size_t count)
1837 {
1838         int err;
1839         unsigned long flags;
1840
1841         err = strict_strtoul(buf, 10, &flags);
1842         if (err || flags > UINT_MAX)
1843                 return -EINVAL;
1844         if (flags > KSM_RUN_UNMERGE)
1845                 return -EINVAL;
1846
1847         /*
1848          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1849          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1850          * breaking COW to free the pages_shared (but leaves mm_slots
1851          * on the list for when ksmd may be set running again).
1852          */
1853
1854         mutex_lock(&ksm_thread_mutex);
1855         if (ksm_run != flags) {
1856                 ksm_run = flags;
1857                 if (flags & KSM_RUN_UNMERGE) {
1858                         current->flags |= PF_OOM_ORIGIN;
1859                         err = unmerge_and_remove_all_rmap_items();
1860                         current->flags &= ~PF_OOM_ORIGIN;
1861                         if (err) {
1862                                 ksm_run = KSM_RUN_STOP;
1863                                 count = err;
1864                         }
1865                 }
1866         }
1867         mutex_unlock(&ksm_thread_mutex);
1868
1869         if (flags & KSM_RUN_MERGE)
1870                 wake_up_interruptible(&ksm_thread_wait);
1871
1872         return count;
1873 }
1874 KSM_ATTR(run);
1875
1876 static ssize_t pages_shared_show(struct kobject *kobj,
1877                                  struct kobj_attribute *attr, char *buf)
1878 {
1879         return sprintf(buf, "%lu\n", ksm_pages_shared);
1880 }
1881 KSM_ATTR_RO(pages_shared);
1882
1883 static ssize_t pages_sharing_show(struct kobject *kobj,
1884                                   struct kobj_attribute *attr, char *buf)
1885 {
1886         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1887 }
1888 KSM_ATTR_RO(pages_sharing);
1889
1890 static ssize_t pages_unshared_show(struct kobject *kobj,
1891                                    struct kobj_attribute *attr, char *buf)
1892 {
1893         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1894 }
1895 KSM_ATTR_RO(pages_unshared);
1896
1897 static ssize_t pages_volatile_show(struct kobject *kobj,
1898                                    struct kobj_attribute *attr, char *buf)
1899 {
1900         long ksm_pages_volatile;
1901
1902         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1903                                 - ksm_pages_sharing - ksm_pages_unshared;
1904         /*
1905          * It was not worth any locking to calculate that statistic,
1906          * but it might therefore sometimes be negative: conceal that.
1907          */
1908         if (ksm_pages_volatile < 0)
1909                 ksm_pages_volatile = 0;
1910         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1911 }
1912 KSM_ATTR_RO(pages_volatile);
1913
1914 static ssize_t full_scans_show(struct kobject *kobj,
1915                                struct kobj_attribute *attr, char *buf)
1916 {
1917         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1918 }
1919 KSM_ATTR_RO(full_scans);
1920
1921 static struct attribute *ksm_attrs[] = {
1922         &sleep_millisecs_attr.attr,
1923         &pages_to_scan_attr.attr,
1924         &run_attr.attr,
1925         &pages_shared_attr.attr,
1926         &pages_sharing_attr.attr,
1927         &pages_unshared_attr.attr,
1928         &pages_volatile_attr.attr,
1929         &full_scans_attr.attr,
1930         NULL,
1931 };
1932
1933 static struct attribute_group ksm_attr_group = {
1934         .attrs = ksm_attrs,
1935         .name = "ksm",
1936 };
1937 #endif /* CONFIG_SYSFS */
1938
1939 static int __init ksm_init(void)
1940 {
1941         struct task_struct *ksm_thread;
1942         int err;
1943
1944         err = ksm_slab_init();
1945         if (err)
1946                 goto out;
1947
1948         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1949         if (IS_ERR(ksm_thread)) {
1950                 printk(KERN_ERR "ksm: creating kthread failed\n");
1951                 err = PTR_ERR(ksm_thread);
1952                 goto out_free;
1953         }
1954
1955 #ifdef CONFIG_SYSFS
1956         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1957         if (err) {
1958                 printk(KERN_ERR "ksm: register sysfs failed\n");
1959                 kthread_stop(ksm_thread);
1960                 goto out_free;
1961         }
1962 #else
1963         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
1964
1965 #endif /* CONFIG_SYSFS */
1966
1967 #ifdef CONFIG_MEMORY_HOTREMOVE
1968         /*
1969          * Choose a high priority since the callback takes ksm_thread_mutex:
1970          * later callbacks could only be taking locks which nest within that.
1971          */
1972         hotplug_memory_notifier(ksm_memory_callback, 100);
1973 #endif
1974         return 0;
1975
1976 out_free:
1977         ksm_slab_free();
1978 out:
1979         return err;
1980 }
1981 module_init(ksm_init)