Merge branch 'drm-core-next' of git://git.kernel.org/pub/scm/linux/kernel/git/airlied...
[pandora-kernel.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hash.h>
37
38 #include <asm/tlbflush.h>
39 #include "internal.h"
40
41 /*
42  * A few notes about the KSM scanning process,
43  * to make it easier to understand the data structures below:
44  *
45  * In order to reduce excessive scanning, KSM sorts the memory pages by their
46  * contents into a data structure that holds pointers to the pages' locations.
47  *
48  * Since the contents of the pages may change at any moment, KSM cannot just
49  * insert the pages into a normal sorted tree and expect it to find anything.
50  * Therefore KSM uses two data structures - the stable and the unstable tree.
51  *
52  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
53  * by their contents.  Because each such page is write-protected, searching on
54  * this tree is fully assured to be working (except when pages are unmapped),
55  * and therefore this tree is called the stable tree.
56  *
57  * In addition to the stable tree, KSM uses a second data structure called the
58  * unstable tree: this tree holds pointers to pages which have been found to
59  * be "unchanged for a period of time".  The unstable tree sorts these pages
60  * by their contents, but since they are not write-protected, KSM cannot rely
61  * upon the unstable tree to work correctly - the unstable tree is liable to
62  * be corrupted as its contents are modified, and so it is called unstable.
63  *
64  * KSM solves this problem by several techniques:
65  *
66  * 1) The unstable tree is flushed every time KSM completes scanning all
67  *    memory areas, and then the tree is rebuilt again from the beginning.
68  * 2) KSM will only insert into the unstable tree, pages whose hash value
69  *    has not changed since the previous scan of all memory areas.
70  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
71  *    colors of the nodes and not on their contents, assuring that even when
72  *    the tree gets "corrupted" it won't get out of balance, so scanning time
73  *    remains the same (also, searching and inserting nodes in an rbtree uses
74  *    the same algorithm, so we have no overhead when we flush and rebuild).
75  * 4) KSM never flushes the stable tree, which means that even if it were to
76  *    take 10 attempts to find a page in the unstable tree, once it is found,
77  *    it is secured in the stable tree.  (When we scan a new page, we first
78  *    compare it against the stable tree, and then against the unstable tree.)
79  */
80
81 /**
82  * struct mm_slot - ksm information per mm that is being scanned
83  * @link: link to the mm_slots hash list
84  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
85  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
86  * @mm: the mm that this information is valid for
87  */
88 struct mm_slot {
89         struct hlist_node link;
90         struct list_head mm_list;
91         struct rmap_item *rmap_list;
92         struct mm_struct *mm;
93 };
94
95 /**
96  * struct ksm_scan - cursor for scanning
97  * @mm_slot: the current mm_slot we are scanning
98  * @address: the next address inside that to be scanned
99  * @rmap_list: link to the next rmap to be scanned in the rmap_list
100  * @seqnr: count of completed full scans (needed when removing unstable node)
101  *
102  * There is only the one ksm_scan instance of this cursor structure.
103  */
104 struct ksm_scan {
105         struct mm_slot *mm_slot;
106         unsigned long address;
107         struct rmap_item **rmap_list;
108         unsigned long seqnr;
109 };
110
111 /**
112  * struct stable_node - node of the stable rbtree
113  * @node: rb node of this ksm page in the stable tree
114  * @hlist: hlist head of rmap_items using this ksm page
115  * @kpfn: page frame number of this ksm page
116  */
117 struct stable_node {
118         struct rb_node node;
119         struct hlist_head hlist;
120         unsigned long kpfn;
121 };
122
123 /**
124  * struct rmap_item - reverse mapping item for virtual addresses
125  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
126  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
127  * @mm: the memory structure this rmap_item is pointing into
128  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
129  * @oldchecksum: previous checksum of the page at that virtual address
130  * @node: rb node of this rmap_item in the unstable tree
131  * @head: pointer to stable_node heading this list in the stable tree
132  * @hlist: link into hlist of rmap_items hanging off that stable_node
133  */
134 struct rmap_item {
135         struct rmap_item *rmap_list;
136         struct anon_vma *anon_vma;      /* when stable */
137         struct mm_struct *mm;
138         unsigned long address;          /* + low bits used for flags below */
139         unsigned int oldchecksum;       /* when unstable */
140         union {
141                 struct rb_node node;    /* when node of unstable tree */
142                 struct {                /* when listed from stable tree */
143                         struct stable_node *head;
144                         struct hlist_node hlist;
145                 };
146         };
147 };
148
149 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
150 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
151 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
152
153 /* The stable and unstable tree heads */
154 static struct rb_root root_stable_tree = RB_ROOT;
155 static struct rb_root root_unstable_tree = RB_ROOT;
156
157 #define MM_SLOTS_HASH_SHIFT 10
158 #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
159 static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
160
161 static struct mm_slot ksm_mm_head = {
162         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
163 };
164 static struct ksm_scan ksm_scan = {
165         .mm_slot = &ksm_mm_head,
166 };
167
168 static struct kmem_cache *rmap_item_cache;
169 static struct kmem_cache *stable_node_cache;
170 static struct kmem_cache *mm_slot_cache;
171
172 /* The number of nodes in the stable tree */
173 static unsigned long ksm_pages_shared;
174
175 /* The number of page slots additionally sharing those nodes */
176 static unsigned long ksm_pages_sharing;
177
178 /* The number of nodes in the unstable tree */
179 static unsigned long ksm_pages_unshared;
180
181 /* The number of rmap_items in use: to calculate pages_volatile */
182 static unsigned long ksm_rmap_items;
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 struct mm_slot *get_mm_slot(struct mm_struct *mm)
275 {
276         struct mm_slot *mm_slot;
277         struct hlist_head *bucket;
278         struct hlist_node *node;
279
280         bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
281         hlist_for_each_entry(mm_slot, node, bucket, link) {
282                 if (mm == mm_slot->mm)
283                         return mm_slot;
284         }
285         return NULL;
286 }
287
288 static void insert_to_mm_slots_hash(struct mm_struct *mm,
289                                     struct mm_slot *mm_slot)
290 {
291         struct hlist_head *bucket;
292
293         bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
294         mm_slot->mm = mm;
295         hlist_add_head(&mm_slot->link, bucket);
296 }
297
298 static inline int in_stable_tree(struct rmap_item *rmap_item)
299 {
300         return rmap_item->address & STABLE_FLAG;
301 }
302
303 static void hold_anon_vma(struct rmap_item *rmap_item,
304                           struct anon_vma *anon_vma)
305 {
306         rmap_item->anon_vma = anon_vma;
307         get_anon_vma(anon_vma);
308 }
309
310 static void ksm_drop_anon_vma(struct rmap_item *rmap_item)
311 {
312         struct anon_vma *anon_vma = rmap_item->anon_vma;
313
314         drop_anon_vma(anon_vma);
315 }
316
317 /*
318  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
319  * page tables after it has passed through ksm_exit() - which, if necessary,
320  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
321  * a special flag: they can just back out as soon as mm_users goes to zero.
322  * ksm_test_exit() is used throughout to make this test for exit: in some
323  * places for correctness, in some places just to avoid unnecessary work.
324  */
325 static inline bool ksm_test_exit(struct mm_struct *mm)
326 {
327         return atomic_read(&mm->mm_users) == 0;
328 }
329
330 /*
331  * We use break_ksm to break COW on a ksm page: it's a stripped down
332  *
333  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
334  *              put_page(page);
335  *
336  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
337  * in case the application has unmapped and remapped mm,addr meanwhile.
338  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
339  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
340  */
341 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
342 {
343         struct page *page;
344         int ret = 0;
345
346         do {
347                 cond_resched();
348                 page = follow_page(vma, addr, FOLL_GET);
349                 if (IS_ERR_OR_NULL(page))
350                         break;
351                 if (PageKsm(page))
352                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
353                                                         FAULT_FLAG_WRITE);
354                 else
355                         ret = VM_FAULT_WRITE;
356                 put_page(page);
357         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
358         /*
359          * We must loop because handle_mm_fault() may back out if there's
360          * any difficulty e.g. if pte accessed bit gets updated concurrently.
361          *
362          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
363          * COW has been broken, even if the vma does not permit VM_WRITE;
364          * but note that a concurrent fault might break PageKsm for us.
365          *
366          * VM_FAULT_SIGBUS could occur if we race with truncation of the
367          * backing file, which also invalidates anonymous pages: that's
368          * okay, that truncation will have unmapped the PageKsm for us.
369          *
370          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
371          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
372          * current task has TIF_MEMDIE set, and will be OOM killed on return
373          * to user; and ksmd, having no mm, would never be chosen for that.
374          *
375          * But if the mm is in a limited mem_cgroup, then the fault may fail
376          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
377          * even ksmd can fail in this way - though it's usually breaking ksm
378          * just to undo a merge it made a moment before, so unlikely to oom.
379          *
380          * That's a pity: we might therefore have more kernel pages allocated
381          * than we're counting as nodes in the stable tree; but ksm_do_scan
382          * will retry to break_cow on each pass, so should recover the page
383          * in due course.  The important thing is to not let VM_MERGEABLE
384          * be cleared while any such pages might remain in the area.
385          */
386         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
387 }
388
389 static void break_cow(struct rmap_item *rmap_item)
390 {
391         struct mm_struct *mm = rmap_item->mm;
392         unsigned long addr = rmap_item->address;
393         struct vm_area_struct *vma;
394
395         /*
396          * It is not an accident that whenever we want to break COW
397          * to undo, we also need to drop a reference to the anon_vma.
398          */
399         ksm_drop_anon_vma(rmap_item);
400
401         down_read(&mm->mmap_sem);
402         if (ksm_test_exit(mm))
403                 goto out;
404         vma = find_vma(mm, addr);
405         if (!vma || vma->vm_start > addr)
406                 goto out;
407         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
408                 goto out;
409         break_ksm(vma, addr);
410 out:
411         up_read(&mm->mmap_sem);
412 }
413
414 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
415 {
416         struct mm_struct *mm = rmap_item->mm;
417         unsigned long addr = rmap_item->address;
418         struct vm_area_struct *vma;
419         struct page *page;
420
421         down_read(&mm->mmap_sem);
422         if (ksm_test_exit(mm))
423                 goto out;
424         vma = find_vma(mm, addr);
425         if (!vma || vma->vm_start > addr)
426                 goto out;
427         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
428                 goto out;
429
430         page = follow_page(vma, addr, FOLL_GET);
431         if (IS_ERR_OR_NULL(page))
432                 goto out;
433         if (PageAnon(page)) {
434                 flush_anon_page(vma, page, addr);
435                 flush_dcache_page(page);
436         } else {
437                 put_page(page);
438 out:            page = NULL;
439         }
440         up_read(&mm->mmap_sem);
441         return page;
442 }
443
444 static void remove_node_from_stable_tree(struct stable_node *stable_node)
445 {
446         struct rmap_item *rmap_item;
447         struct hlist_node *hlist;
448
449         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
450                 if (rmap_item->hlist.next)
451                         ksm_pages_sharing--;
452                 else
453                         ksm_pages_shared--;
454                 ksm_drop_anon_vma(rmap_item);
455                 rmap_item->address &= PAGE_MASK;
456                 cond_resched();
457         }
458
459         rb_erase(&stable_node->node, &root_stable_tree);
460         free_stable_node(stable_node);
461 }
462
463 /*
464  * get_ksm_page: checks if the page indicated by the stable node
465  * is still its ksm page, despite having held no reference to it.
466  * In which case we can trust the content of the page, and it
467  * returns the gotten page; but if the page has now been zapped,
468  * remove the stale node from the stable tree and return NULL.
469  *
470  * You would expect the stable_node to hold a reference to the ksm page.
471  * But if it increments the page's count, swapping out has to wait for
472  * ksmd to come around again before it can free the page, which may take
473  * seconds or even minutes: much too unresponsive.  So instead we use a
474  * "keyhole reference": access to the ksm page from the stable node peeps
475  * out through its keyhole to see if that page still holds the right key,
476  * pointing back to this stable node.  This relies on freeing a PageAnon
477  * page to reset its page->mapping to NULL, and relies on no other use of
478  * a page to put something that might look like our key in page->mapping.
479  *
480  * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
481  * but this is different - made simpler by ksm_thread_mutex being held, but
482  * interesting for assuming that no other use of the struct page could ever
483  * put our expected_mapping into page->mapping (or a field of the union which
484  * coincides with page->mapping).  The RCU calls are not for KSM at all, but
485  * to keep the page_count protocol described with page_cache_get_speculative.
486  *
487  * Note: it is possible that get_ksm_page() will return NULL one moment,
488  * then page the next, if the page is in between page_freeze_refs() and
489  * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
490  * is on its way to being freed; but it is an anomaly to bear in mind.
491  */
492 static struct page *get_ksm_page(struct stable_node *stable_node)
493 {
494         struct page *page;
495         void *expected_mapping;
496
497         page = pfn_to_page(stable_node->kpfn);
498         expected_mapping = (void *)stable_node +
499                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
500         rcu_read_lock();
501         if (page->mapping != expected_mapping)
502                 goto stale;
503         if (!get_page_unless_zero(page))
504                 goto stale;
505         if (page->mapping != expected_mapping) {
506                 put_page(page);
507                 goto stale;
508         }
509         rcu_read_unlock();
510         return page;
511 stale:
512         rcu_read_unlock();
513         remove_node_from_stable_tree(stable_node);
514         return NULL;
515 }
516
517 /*
518  * Removing rmap_item from stable or unstable tree.
519  * This function will clean the information from the stable/unstable tree.
520  */
521 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
522 {
523         if (rmap_item->address & STABLE_FLAG) {
524                 struct stable_node *stable_node;
525                 struct page *page;
526
527                 stable_node = rmap_item->head;
528                 page = get_ksm_page(stable_node);
529                 if (!page)
530                         goto out;
531
532                 lock_page(page);
533                 hlist_del(&rmap_item->hlist);
534                 unlock_page(page);
535                 put_page(page);
536
537                 if (stable_node->hlist.first)
538                         ksm_pages_sharing--;
539                 else
540                         ksm_pages_shared--;
541
542                 ksm_drop_anon_vma(rmap_item);
543                 rmap_item->address &= PAGE_MASK;
544
545         } else if (rmap_item->address & UNSTABLE_FLAG) {
546                 unsigned char age;
547                 /*
548                  * Usually ksmd can and must skip the rb_erase, because
549                  * root_unstable_tree was already reset to RB_ROOT.
550                  * But be careful when an mm is exiting: do the rb_erase
551                  * if this rmap_item was inserted by this scan, rather
552                  * than left over from before.
553                  */
554                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
555                 BUG_ON(age > 1);
556                 if (!age)
557                         rb_erase(&rmap_item->node, &root_unstable_tree);
558
559                 ksm_pages_unshared--;
560                 rmap_item->address &= PAGE_MASK;
561         }
562 out:
563         cond_resched();         /* we're called from many long loops */
564 }
565
566 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
567                                        struct rmap_item **rmap_list)
568 {
569         while (*rmap_list) {
570                 struct rmap_item *rmap_item = *rmap_list;
571                 *rmap_list = rmap_item->rmap_list;
572                 remove_rmap_item_from_tree(rmap_item);
573                 free_rmap_item(rmap_item);
574         }
575 }
576
577 /*
578  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
579  * than check every pte of a given vma, the locking doesn't quite work for
580  * that - an rmap_item is assigned to the stable tree after inserting ksm
581  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
582  * rmap_items from parent to child at fork time (so as not to waste time
583  * if exit comes before the next scan reaches it).
584  *
585  * Similarly, although we'd like to remove rmap_items (so updating counts
586  * and freeing memory) when unmerging an area, it's easier to leave that
587  * to the next pass of ksmd - consider, for example, how ksmd might be
588  * in cmp_and_merge_page on one of the rmap_items we would be removing.
589  */
590 static int unmerge_ksm_pages(struct vm_area_struct *vma,
591                              unsigned long start, unsigned long end)
592 {
593         unsigned long addr;
594         int err = 0;
595
596         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
597                 if (ksm_test_exit(vma->vm_mm))
598                         break;
599                 if (signal_pending(current))
600                         err = -ERESTARTSYS;
601                 else
602                         err = break_ksm(vma, addr);
603         }
604         return err;
605 }
606
607 #ifdef CONFIG_SYSFS
608 /*
609  * Only called through the sysfs control interface:
610  */
611 static int unmerge_and_remove_all_rmap_items(void)
612 {
613         struct mm_slot *mm_slot;
614         struct mm_struct *mm;
615         struct vm_area_struct *vma;
616         int err = 0;
617
618         spin_lock(&ksm_mmlist_lock);
619         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
620                                                 struct mm_slot, mm_list);
621         spin_unlock(&ksm_mmlist_lock);
622
623         for (mm_slot = ksm_scan.mm_slot;
624                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
625                 mm = mm_slot->mm;
626                 down_read(&mm->mmap_sem);
627                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
628                         if (ksm_test_exit(mm))
629                                 break;
630                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
631                                 continue;
632                         err = unmerge_ksm_pages(vma,
633                                                 vma->vm_start, vma->vm_end);
634                         if (err)
635                                 goto error;
636                 }
637
638                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
639
640                 spin_lock(&ksm_mmlist_lock);
641                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
642                                                 struct mm_slot, mm_list);
643                 if (ksm_test_exit(mm)) {
644                         hlist_del(&mm_slot->link);
645                         list_del(&mm_slot->mm_list);
646                         spin_unlock(&ksm_mmlist_lock);
647
648                         free_mm_slot(mm_slot);
649                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
650                         up_read(&mm->mmap_sem);
651                         mmdrop(mm);
652                 } else {
653                         spin_unlock(&ksm_mmlist_lock);
654                         up_read(&mm->mmap_sem);
655                 }
656         }
657
658         ksm_scan.seqnr = 0;
659         return 0;
660
661 error:
662         up_read(&mm->mmap_sem);
663         spin_lock(&ksm_mmlist_lock);
664         ksm_scan.mm_slot = &ksm_mm_head;
665         spin_unlock(&ksm_mmlist_lock);
666         return err;
667 }
668 #endif /* CONFIG_SYSFS */
669
670 static u32 calc_checksum(struct page *page)
671 {
672         u32 checksum;
673         void *addr = kmap_atomic(page, KM_USER0);
674         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
675         kunmap_atomic(addr, KM_USER0);
676         return checksum;
677 }
678
679 static int memcmp_pages(struct page *page1, struct page *page2)
680 {
681         char *addr1, *addr2;
682         int ret;
683
684         addr1 = kmap_atomic(page1, KM_USER0);
685         addr2 = kmap_atomic(page2, KM_USER1);
686         ret = memcmp(addr1, addr2, PAGE_SIZE);
687         kunmap_atomic(addr2, KM_USER1);
688         kunmap_atomic(addr1, KM_USER0);
689         return ret;
690 }
691
692 static inline int pages_identical(struct page *page1, struct page *page2)
693 {
694         return !memcmp_pages(page1, page2);
695 }
696
697 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
698                               pte_t *orig_pte)
699 {
700         struct mm_struct *mm = vma->vm_mm;
701         unsigned long addr;
702         pte_t *ptep;
703         spinlock_t *ptl;
704         int swapped;
705         int err = -EFAULT;
706
707         addr = page_address_in_vma(page, vma);
708         if (addr == -EFAULT)
709                 goto out;
710
711         ptep = page_check_address(page, mm, addr, &ptl, 0);
712         if (!ptep)
713                 goto out;
714
715         if (pte_write(*ptep)) {
716                 pte_t entry;
717
718                 swapped = PageSwapCache(page);
719                 flush_cache_page(vma, addr, page_to_pfn(page));
720                 /*
721                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
722                  * take any lock, therefore the check that we are going to make
723                  * with the pagecount against the mapcount is racey and
724                  * O_DIRECT can happen right after the check.
725                  * So we clear the pte and flush the tlb before the check
726                  * this assure us that no O_DIRECT can happen after the check
727                  * or in the middle of the check.
728                  */
729                 entry = ptep_clear_flush(vma, addr, ptep);
730                 /*
731                  * Check that no O_DIRECT or similar I/O is in progress on the
732                  * page
733                  */
734                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
735                         set_pte_at(mm, addr, ptep, entry);
736                         goto out_unlock;
737                 }
738                 entry = pte_wrprotect(entry);
739                 set_pte_at_notify(mm, addr, ptep, entry);
740         }
741         *orig_pte = *ptep;
742         err = 0;
743
744 out_unlock:
745         pte_unmap_unlock(ptep, ptl);
746 out:
747         return err;
748 }
749
750 /**
751  * replace_page - replace page in vma by new ksm page
752  * @vma:      vma that holds the pte pointing to page
753  * @page:     the page we are replacing by kpage
754  * @kpage:    the ksm page we replace page by
755  * @orig_pte: the original value of the pte
756  *
757  * Returns 0 on success, -EFAULT on failure.
758  */
759 static int replace_page(struct vm_area_struct *vma, struct page *page,
760                         struct page *kpage, pte_t orig_pte)
761 {
762         struct mm_struct *mm = vma->vm_mm;
763         pgd_t *pgd;
764         pud_t *pud;
765         pmd_t *pmd;
766         pte_t *ptep;
767         spinlock_t *ptl;
768         unsigned long addr;
769         int err = -EFAULT;
770
771         addr = page_address_in_vma(page, vma);
772         if (addr == -EFAULT)
773                 goto out;
774
775         pgd = pgd_offset(mm, addr);
776         if (!pgd_present(*pgd))
777                 goto out;
778
779         pud = pud_offset(pgd, addr);
780         if (!pud_present(*pud))
781                 goto out;
782
783         pmd = pmd_offset(pud, addr);
784         if (!pmd_present(*pmd))
785                 goto out;
786
787         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
788         if (!pte_same(*ptep, orig_pte)) {
789                 pte_unmap_unlock(ptep, ptl);
790                 goto out;
791         }
792
793         get_page(kpage);
794         page_add_anon_rmap(kpage, vma, addr);
795
796         flush_cache_page(vma, addr, pte_pfn(*ptep));
797         ptep_clear_flush(vma, addr, ptep);
798         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
799
800         page_remove_rmap(page);
801         put_page(page);
802
803         pte_unmap_unlock(ptep, ptl);
804         err = 0;
805 out:
806         return err;
807 }
808
809 /*
810  * try_to_merge_one_page - take two pages and merge them into one
811  * @vma: the vma that holds the pte pointing to page
812  * @page: the PageAnon page that we want to replace with kpage
813  * @kpage: the PageKsm page that we want to map instead of page,
814  *         or NULL the first time when we want to use page as kpage.
815  *
816  * This function returns 0 if the pages were merged, -EFAULT otherwise.
817  */
818 static int try_to_merge_one_page(struct vm_area_struct *vma,
819                                  struct page *page, struct page *kpage)
820 {
821         pte_t orig_pte = __pte(0);
822         int err = -EFAULT;
823
824         if (page == kpage)                      /* ksm page forked */
825                 return 0;
826
827         if (!(vma->vm_flags & VM_MERGEABLE))
828                 goto out;
829         if (!PageAnon(page))
830                 goto out;
831
832         /*
833          * We need the page lock to read a stable PageSwapCache in
834          * write_protect_page().  We use trylock_page() instead of
835          * lock_page() because we don't want to wait here - we
836          * prefer to continue scanning and merging different pages,
837          * then come back to this page when it is unlocked.
838          */
839         if (!trylock_page(page))
840                 goto out;
841         /*
842          * If this anonymous page is mapped only here, its pte may need
843          * to be write-protected.  If it's mapped elsewhere, all of its
844          * ptes are necessarily already write-protected.  But in either
845          * case, we need to lock and check page_count is not raised.
846          */
847         if (write_protect_page(vma, page, &orig_pte) == 0) {
848                 if (!kpage) {
849                         /*
850                          * While we hold page lock, upgrade page from
851                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
852                          * stable_tree_insert() will update stable_node.
853                          */
854                         set_page_stable_node(page, NULL);
855                         mark_page_accessed(page);
856                         err = 0;
857                 } else if (pages_identical(page, kpage))
858                         err = replace_page(vma, page, kpage, orig_pte);
859         }
860
861         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
862                 munlock_vma_page(page);
863                 if (!PageMlocked(kpage)) {
864                         unlock_page(page);
865                         lock_page(kpage);
866                         mlock_vma_page(kpage);
867                         page = kpage;           /* for final unlock */
868                 }
869         }
870
871         unlock_page(page);
872 out:
873         return err;
874 }
875
876 /*
877  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
878  * but no new kernel page is allocated: kpage must already be a ksm page.
879  *
880  * This function returns 0 if the pages were merged, -EFAULT otherwise.
881  */
882 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
883                                       struct page *page, struct page *kpage)
884 {
885         struct mm_struct *mm = rmap_item->mm;
886         struct vm_area_struct *vma;
887         int err = -EFAULT;
888
889         down_read(&mm->mmap_sem);
890         if (ksm_test_exit(mm))
891                 goto out;
892         vma = find_vma(mm, rmap_item->address);
893         if (!vma || vma->vm_start > rmap_item->address)
894                 goto out;
895
896         err = try_to_merge_one_page(vma, page, kpage);
897         if (err)
898                 goto out;
899
900         /* Must get reference to anon_vma while still holding mmap_sem */
901         hold_anon_vma(rmap_item, vma->anon_vma);
902 out:
903         up_read(&mm->mmap_sem);
904         return err;
905 }
906
907 /*
908  * try_to_merge_two_pages - take two identical pages and prepare them
909  * to be merged into one page.
910  *
911  * This function returns the kpage if we successfully merged two identical
912  * pages into one ksm page, NULL otherwise.
913  *
914  * Note that this function upgrades page to ksm page: if one of the pages
915  * is already a ksm page, try_to_merge_with_ksm_page should be used.
916  */
917 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
918                                            struct page *page,
919                                            struct rmap_item *tree_rmap_item,
920                                            struct page *tree_page)
921 {
922         int err;
923
924         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
925         if (!err) {
926                 err = try_to_merge_with_ksm_page(tree_rmap_item,
927                                                         tree_page, page);
928                 /*
929                  * If that fails, we have a ksm page with only one pte
930                  * pointing to it: so break it.
931                  */
932                 if (err)
933                         break_cow(rmap_item);
934         }
935         return err ? NULL : page;
936 }
937
938 /*
939  * stable_tree_search - search for page inside the stable tree
940  *
941  * This function checks if there is a page inside the stable tree
942  * with identical content to the page that we are scanning right now.
943  *
944  * This function returns the stable tree node of identical content if found,
945  * NULL otherwise.
946  */
947 static struct page *stable_tree_search(struct page *page)
948 {
949         struct rb_node *node = root_stable_tree.rb_node;
950         struct stable_node *stable_node;
951
952         stable_node = page_stable_node(page);
953         if (stable_node) {                      /* ksm page forked */
954                 get_page(page);
955                 return page;
956         }
957
958         while (node) {
959                 struct page *tree_page;
960                 int ret;
961
962                 cond_resched();
963                 stable_node = rb_entry(node, struct stable_node, node);
964                 tree_page = get_ksm_page(stable_node);
965                 if (!tree_page)
966                         return NULL;
967
968                 ret = memcmp_pages(page, tree_page);
969
970                 if (ret < 0) {
971                         put_page(tree_page);
972                         node = node->rb_left;
973                 } else if (ret > 0) {
974                         put_page(tree_page);
975                         node = node->rb_right;
976                 } else
977                         return tree_page;
978         }
979
980         return NULL;
981 }
982
983 /*
984  * stable_tree_insert - insert rmap_item pointing to new ksm page
985  * into the stable tree.
986  *
987  * This function returns the stable tree node just allocated on success,
988  * NULL otherwise.
989  */
990 static struct stable_node *stable_tree_insert(struct page *kpage)
991 {
992         struct rb_node **new = &root_stable_tree.rb_node;
993         struct rb_node *parent = NULL;
994         struct stable_node *stable_node;
995
996         while (*new) {
997                 struct page *tree_page;
998                 int ret;
999
1000                 cond_resched();
1001                 stable_node = rb_entry(*new, struct stable_node, node);
1002                 tree_page = get_ksm_page(stable_node);
1003                 if (!tree_page)
1004                         return NULL;
1005
1006                 ret = memcmp_pages(kpage, tree_page);
1007                 put_page(tree_page);
1008
1009                 parent = *new;
1010                 if (ret < 0)
1011                         new = &parent->rb_left;
1012                 else if (ret > 0)
1013                         new = &parent->rb_right;
1014                 else {
1015                         /*
1016                          * It is not a bug that stable_tree_search() didn't
1017                          * find this node: because at that time our page was
1018                          * not yet write-protected, so may have changed since.
1019                          */
1020                         return NULL;
1021                 }
1022         }
1023
1024         stable_node = alloc_stable_node();
1025         if (!stable_node)
1026                 return NULL;
1027
1028         rb_link_node(&stable_node->node, parent, new);
1029         rb_insert_color(&stable_node->node, &root_stable_tree);
1030
1031         INIT_HLIST_HEAD(&stable_node->hlist);
1032
1033         stable_node->kpfn = page_to_pfn(kpage);
1034         set_page_stable_node(kpage, stable_node);
1035
1036         return stable_node;
1037 }
1038
1039 /*
1040  * unstable_tree_search_insert - search for identical page,
1041  * else insert rmap_item into the unstable tree.
1042  *
1043  * This function searches for a page in the unstable tree identical to the
1044  * page currently being scanned; and if no identical page is found in the
1045  * tree, we insert rmap_item as a new object into the unstable tree.
1046  *
1047  * This function returns pointer to rmap_item found to be identical
1048  * to the currently scanned page, NULL otherwise.
1049  *
1050  * This function does both searching and inserting, because they share
1051  * the same walking algorithm in an rbtree.
1052  */
1053 static
1054 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1055                                               struct page *page,
1056                                               struct page **tree_pagep)
1057
1058 {
1059         struct rb_node **new = &root_unstable_tree.rb_node;
1060         struct rb_node *parent = NULL;
1061
1062         while (*new) {
1063                 struct rmap_item *tree_rmap_item;
1064                 struct page *tree_page;
1065                 int ret;
1066
1067                 cond_resched();
1068                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1069                 tree_page = get_mergeable_page(tree_rmap_item);
1070                 if (IS_ERR_OR_NULL(tree_page))
1071                         return NULL;
1072
1073                 /*
1074                  * Don't substitute a ksm page for a forked page.
1075                  */
1076                 if (page == tree_page) {
1077                         put_page(tree_page);
1078                         return NULL;
1079                 }
1080
1081                 ret = memcmp_pages(page, tree_page);
1082
1083                 parent = *new;
1084                 if (ret < 0) {
1085                         put_page(tree_page);
1086                         new = &parent->rb_left;
1087                 } else if (ret > 0) {
1088                         put_page(tree_page);
1089                         new = &parent->rb_right;
1090                 } else {
1091                         *tree_pagep = tree_page;
1092                         return tree_rmap_item;
1093                 }
1094         }
1095
1096         rmap_item->address |= UNSTABLE_FLAG;
1097         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1098         rb_link_node(&rmap_item->node, parent, new);
1099         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1100
1101         ksm_pages_unshared++;
1102         return NULL;
1103 }
1104
1105 /*
1106  * stable_tree_append - add another rmap_item to the linked list of
1107  * rmap_items hanging off a given node of the stable tree, all sharing
1108  * the same ksm page.
1109  */
1110 static void stable_tree_append(struct rmap_item *rmap_item,
1111                                struct stable_node *stable_node)
1112 {
1113         rmap_item->head = stable_node;
1114         rmap_item->address |= STABLE_FLAG;
1115         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1116
1117         if (rmap_item->hlist.next)
1118                 ksm_pages_sharing++;
1119         else
1120                 ksm_pages_shared++;
1121 }
1122
1123 /*
1124  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1125  * if not, compare checksum to previous and if it's the same, see if page can
1126  * be inserted into the unstable tree, or merged with a page already there and
1127  * both transferred to the stable tree.
1128  *
1129  * @page: the page that we are searching identical page to.
1130  * @rmap_item: the reverse mapping into the virtual address of this page
1131  */
1132 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1133 {
1134         struct rmap_item *tree_rmap_item;
1135         struct page *tree_page = NULL;
1136         struct stable_node *stable_node;
1137         struct page *kpage;
1138         unsigned int checksum;
1139         int err;
1140
1141         remove_rmap_item_from_tree(rmap_item);
1142
1143         /* We first start with searching the page inside the stable tree */
1144         kpage = stable_tree_search(page);
1145         if (kpage) {
1146                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1147                 if (!err) {
1148                         /*
1149                          * The page was successfully merged:
1150                          * add its rmap_item to the stable tree.
1151                          */
1152                         lock_page(kpage);
1153                         stable_tree_append(rmap_item, page_stable_node(kpage));
1154                         unlock_page(kpage);
1155                 }
1156                 put_page(kpage);
1157                 return;
1158         }
1159
1160         /*
1161          * If the hash value of the page has changed from the last time
1162          * we calculated it, this page is changing frequently: therefore we
1163          * don't want to insert it in the unstable tree, and we don't want
1164          * to waste our time searching for something identical to it there.
1165          */
1166         checksum = calc_checksum(page);
1167         if (rmap_item->oldchecksum != checksum) {
1168                 rmap_item->oldchecksum = checksum;
1169                 return;
1170         }
1171
1172         tree_rmap_item =
1173                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1174         if (tree_rmap_item) {
1175                 kpage = try_to_merge_two_pages(rmap_item, page,
1176                                                 tree_rmap_item, tree_page);
1177                 put_page(tree_page);
1178                 /*
1179                  * As soon as we merge this page, we want to remove the
1180                  * rmap_item of the page we have merged with from the unstable
1181                  * tree, and insert it instead as new node in the stable tree.
1182                  */
1183                 if (kpage) {
1184                         remove_rmap_item_from_tree(tree_rmap_item);
1185
1186                         lock_page(kpage);
1187                         stable_node = stable_tree_insert(kpage);
1188                         if (stable_node) {
1189                                 stable_tree_append(tree_rmap_item, stable_node);
1190                                 stable_tree_append(rmap_item, stable_node);
1191                         }
1192                         unlock_page(kpage);
1193
1194                         /*
1195                          * If we fail to insert the page into the stable tree,
1196                          * we will have 2 virtual addresses that are pointing
1197                          * to a ksm page left outside the stable tree,
1198                          * in which case we need to break_cow on both.
1199                          */
1200                         if (!stable_node) {
1201                                 break_cow(tree_rmap_item);
1202                                 break_cow(rmap_item);
1203                         }
1204                 }
1205         }
1206 }
1207
1208 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1209                                             struct rmap_item **rmap_list,
1210                                             unsigned long addr)
1211 {
1212         struct rmap_item *rmap_item;
1213
1214         while (*rmap_list) {
1215                 rmap_item = *rmap_list;
1216                 if ((rmap_item->address & PAGE_MASK) == addr)
1217                         return rmap_item;
1218                 if (rmap_item->address > addr)
1219                         break;
1220                 *rmap_list = rmap_item->rmap_list;
1221                 remove_rmap_item_from_tree(rmap_item);
1222                 free_rmap_item(rmap_item);
1223         }
1224
1225         rmap_item = alloc_rmap_item();
1226         if (rmap_item) {
1227                 /* It has already been zeroed */
1228                 rmap_item->mm = mm_slot->mm;
1229                 rmap_item->address = addr;
1230                 rmap_item->rmap_list = *rmap_list;
1231                 *rmap_list = rmap_item;
1232         }
1233         return rmap_item;
1234 }
1235
1236 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1237 {
1238         struct mm_struct *mm;
1239         struct mm_slot *slot;
1240         struct vm_area_struct *vma;
1241         struct rmap_item *rmap_item;
1242
1243         if (list_empty(&ksm_mm_head.mm_list))
1244                 return NULL;
1245
1246         slot = ksm_scan.mm_slot;
1247         if (slot == &ksm_mm_head) {
1248                 root_unstable_tree = RB_ROOT;
1249
1250                 spin_lock(&ksm_mmlist_lock);
1251                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1252                 ksm_scan.mm_slot = slot;
1253                 spin_unlock(&ksm_mmlist_lock);
1254 next_mm:
1255                 ksm_scan.address = 0;
1256                 ksm_scan.rmap_list = &slot->rmap_list;
1257         }
1258
1259         mm = slot->mm;
1260         down_read(&mm->mmap_sem);
1261         if (ksm_test_exit(mm))
1262                 vma = NULL;
1263         else
1264                 vma = find_vma(mm, ksm_scan.address);
1265
1266         for (; vma; vma = vma->vm_next) {
1267                 if (!(vma->vm_flags & VM_MERGEABLE))
1268                         continue;
1269                 if (ksm_scan.address < vma->vm_start)
1270                         ksm_scan.address = vma->vm_start;
1271                 if (!vma->anon_vma)
1272                         ksm_scan.address = vma->vm_end;
1273
1274                 while (ksm_scan.address < vma->vm_end) {
1275                         if (ksm_test_exit(mm))
1276                                 break;
1277                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1278                         if (!IS_ERR_OR_NULL(*page) && PageAnon(*page)) {
1279                                 flush_anon_page(vma, *page, ksm_scan.address);
1280                                 flush_dcache_page(*page);
1281                                 rmap_item = get_next_rmap_item(slot,
1282                                         ksm_scan.rmap_list, ksm_scan.address);
1283                                 if (rmap_item) {
1284                                         ksm_scan.rmap_list =
1285                                                         &rmap_item->rmap_list;
1286                                         ksm_scan.address += PAGE_SIZE;
1287                                 } else
1288                                         put_page(*page);
1289                                 up_read(&mm->mmap_sem);
1290                                 return rmap_item;
1291                         }
1292                         if (!IS_ERR_OR_NULL(*page))
1293                                 put_page(*page);
1294                         ksm_scan.address += PAGE_SIZE;
1295                         cond_resched();
1296                 }
1297         }
1298
1299         if (ksm_test_exit(mm)) {
1300                 ksm_scan.address = 0;
1301                 ksm_scan.rmap_list = &slot->rmap_list;
1302         }
1303         /*
1304          * Nuke all the rmap_items that are above this current rmap:
1305          * because there were no VM_MERGEABLE vmas with such addresses.
1306          */
1307         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1308
1309         spin_lock(&ksm_mmlist_lock);
1310         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1311                                                 struct mm_slot, mm_list);
1312         if (ksm_scan.address == 0) {
1313                 /*
1314                  * We've completed a full scan of all vmas, holding mmap_sem
1315                  * throughout, and found no VM_MERGEABLE: so do the same as
1316                  * __ksm_exit does to remove this mm from all our lists now.
1317                  * This applies either when cleaning up after __ksm_exit
1318                  * (but beware: we can reach here even before __ksm_exit),
1319                  * or when all VM_MERGEABLE areas have been unmapped (and
1320                  * mmap_sem then protects against race with MADV_MERGEABLE).
1321                  */
1322                 hlist_del(&slot->link);
1323                 list_del(&slot->mm_list);
1324                 spin_unlock(&ksm_mmlist_lock);
1325
1326                 free_mm_slot(slot);
1327                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1328                 up_read(&mm->mmap_sem);
1329                 mmdrop(mm);
1330         } else {
1331                 spin_unlock(&ksm_mmlist_lock);
1332                 up_read(&mm->mmap_sem);
1333         }
1334
1335         /* Repeat until we've completed scanning the whole list */
1336         slot = ksm_scan.mm_slot;
1337         if (slot != &ksm_mm_head)
1338                 goto next_mm;
1339
1340         ksm_scan.seqnr++;
1341         return NULL;
1342 }
1343
1344 /**
1345  * ksm_do_scan  - the ksm scanner main worker function.
1346  * @scan_npages - number of pages we want to scan before we return.
1347  */
1348 static void ksm_do_scan(unsigned int scan_npages)
1349 {
1350         struct rmap_item *rmap_item;
1351         struct page *uninitialized_var(page);
1352
1353         while (scan_npages--) {
1354                 cond_resched();
1355                 rmap_item = scan_get_next_rmap_item(&page);
1356                 if (!rmap_item)
1357                         return;
1358                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1359                         cmp_and_merge_page(page, rmap_item);
1360                 put_page(page);
1361         }
1362 }
1363
1364 static int ksmd_should_run(void)
1365 {
1366         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1367 }
1368
1369 static int ksm_scan_thread(void *nothing)
1370 {
1371         set_user_nice(current, 5);
1372
1373         while (!kthread_should_stop()) {
1374                 mutex_lock(&ksm_thread_mutex);
1375                 if (ksmd_should_run())
1376                         ksm_do_scan(ksm_thread_pages_to_scan);
1377                 mutex_unlock(&ksm_thread_mutex);
1378
1379                 if (ksmd_should_run()) {
1380                         schedule_timeout_interruptible(
1381                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1382                 } else {
1383                         wait_event_interruptible(ksm_thread_wait,
1384                                 ksmd_should_run() || kthread_should_stop());
1385                 }
1386         }
1387         return 0;
1388 }
1389
1390 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1391                 unsigned long end, int advice, unsigned long *vm_flags)
1392 {
1393         struct mm_struct *mm = vma->vm_mm;
1394         int err;
1395
1396         switch (advice) {
1397         case MADV_MERGEABLE:
1398                 /*
1399                  * Be somewhat over-protective for now!
1400                  */
1401                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1402                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1403                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1404                                  VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1405                         return 0;               /* just ignore the advice */
1406
1407                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1408                         err = __ksm_enter(mm);
1409                         if (err)
1410                                 return err;
1411                 }
1412
1413                 *vm_flags |= VM_MERGEABLE;
1414                 break;
1415
1416         case MADV_UNMERGEABLE:
1417                 if (!(*vm_flags & VM_MERGEABLE))
1418                         return 0;               /* just ignore the advice */
1419
1420                 if (vma->anon_vma) {
1421                         err = unmerge_ksm_pages(vma, start, end);
1422                         if (err)
1423                                 return err;
1424                 }
1425
1426                 *vm_flags &= ~VM_MERGEABLE;
1427                 break;
1428         }
1429
1430         return 0;
1431 }
1432
1433 int __ksm_enter(struct mm_struct *mm)
1434 {
1435         struct mm_slot *mm_slot;
1436         int needs_wakeup;
1437
1438         mm_slot = alloc_mm_slot();
1439         if (!mm_slot)
1440                 return -ENOMEM;
1441
1442         /* Check ksm_run too?  Would need tighter locking */
1443         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1444
1445         spin_lock(&ksm_mmlist_lock);
1446         insert_to_mm_slots_hash(mm, mm_slot);
1447         /*
1448          * Insert just behind the scanning cursor, to let the area settle
1449          * down a little; when fork is followed by immediate exec, we don't
1450          * want ksmd to waste time setting up and tearing down an rmap_list.
1451          */
1452         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1453         spin_unlock(&ksm_mmlist_lock);
1454
1455         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1456         atomic_inc(&mm->mm_count);
1457
1458         if (needs_wakeup)
1459                 wake_up_interruptible(&ksm_thread_wait);
1460
1461         return 0;
1462 }
1463
1464 void __ksm_exit(struct mm_struct *mm)
1465 {
1466         struct mm_slot *mm_slot;
1467         int easy_to_free = 0;
1468
1469         /*
1470          * This process is exiting: if it's straightforward (as is the
1471          * case when ksmd was never running), free mm_slot immediately.
1472          * But if it's at the cursor or has rmap_items linked to it, use
1473          * mmap_sem to synchronize with any break_cows before pagetables
1474          * are freed, and leave the mm_slot on the list for ksmd to free.
1475          * Beware: ksm may already have noticed it exiting and freed the slot.
1476          */
1477
1478         spin_lock(&ksm_mmlist_lock);
1479         mm_slot = get_mm_slot(mm);
1480         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1481                 if (!mm_slot->rmap_list) {
1482                         hlist_del(&mm_slot->link);
1483                         list_del(&mm_slot->mm_list);
1484                         easy_to_free = 1;
1485                 } else {
1486                         list_move(&mm_slot->mm_list,
1487                                   &ksm_scan.mm_slot->mm_list);
1488                 }
1489         }
1490         spin_unlock(&ksm_mmlist_lock);
1491
1492         if (easy_to_free) {
1493                 free_mm_slot(mm_slot);
1494                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1495                 mmdrop(mm);
1496         } else if (mm_slot) {
1497                 down_write(&mm->mmap_sem);
1498                 up_write(&mm->mmap_sem);
1499         }
1500 }
1501
1502 struct page *ksm_does_need_to_copy(struct page *page,
1503                         struct vm_area_struct *vma, unsigned long address)
1504 {
1505         struct page *new_page;
1506
1507         unlock_page(page);      /* any racers will COW it, not modify it */
1508
1509         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1510         if (new_page) {
1511                 copy_user_highpage(new_page, page, address, vma);
1512
1513                 SetPageDirty(new_page);
1514                 __SetPageUptodate(new_page);
1515                 SetPageSwapBacked(new_page);
1516                 __set_page_locked(new_page);
1517
1518                 if (page_evictable(new_page, vma))
1519                         lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1520                 else
1521                         add_page_to_unevictable_list(new_page);
1522         }
1523
1524         page_cache_release(page);
1525         return new_page;
1526 }
1527
1528 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1529                         unsigned long *vm_flags)
1530 {
1531         struct stable_node *stable_node;
1532         struct rmap_item *rmap_item;
1533         struct hlist_node *hlist;
1534         unsigned int mapcount = page_mapcount(page);
1535         int referenced = 0;
1536         int search_new_forks = 0;
1537
1538         VM_BUG_ON(!PageKsm(page));
1539         VM_BUG_ON(!PageLocked(page));
1540
1541         stable_node = page_stable_node(page);
1542         if (!stable_node)
1543                 return 0;
1544 again:
1545         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1546                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1547                 struct anon_vma_chain *vmac;
1548                 struct vm_area_struct *vma;
1549
1550                 anon_vma_lock(anon_vma);
1551                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1552                         vma = vmac->vma;
1553                         if (rmap_item->address < vma->vm_start ||
1554                             rmap_item->address >= vma->vm_end)
1555                                 continue;
1556                         /*
1557                          * Initially we examine only the vma which covers this
1558                          * rmap_item; but later, if there is still work to do,
1559                          * we examine covering vmas in other mms: in case they
1560                          * were forked from the original since ksmd passed.
1561                          */
1562                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1563                                 continue;
1564
1565                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1566                                 continue;
1567
1568                         referenced += page_referenced_one(page, vma,
1569                                 rmap_item->address, &mapcount, vm_flags);
1570                         if (!search_new_forks || !mapcount)
1571                                 break;
1572                 }
1573                 anon_vma_unlock(anon_vma);
1574                 if (!mapcount)
1575                         goto out;
1576         }
1577         if (!search_new_forks++)
1578                 goto again;
1579 out:
1580         return referenced;
1581 }
1582
1583 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1584 {
1585         struct stable_node *stable_node;
1586         struct hlist_node *hlist;
1587         struct rmap_item *rmap_item;
1588         int ret = SWAP_AGAIN;
1589         int search_new_forks = 0;
1590
1591         VM_BUG_ON(!PageKsm(page));
1592         VM_BUG_ON(!PageLocked(page));
1593
1594         stable_node = page_stable_node(page);
1595         if (!stable_node)
1596                 return SWAP_FAIL;
1597 again:
1598         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1599                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1600                 struct anon_vma_chain *vmac;
1601                 struct vm_area_struct *vma;
1602
1603                 anon_vma_lock(anon_vma);
1604                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1605                         vma = vmac->vma;
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                         ret = try_to_unmap_one(page, vma,
1619                                         rmap_item->address, flags);
1620                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1621                                 anon_vma_unlock(anon_vma);
1622                                 goto out;
1623                         }
1624                 }
1625                 anon_vma_unlock(anon_vma);
1626         }
1627         if (!search_new_forks++)
1628                 goto again;
1629 out:
1630         return ret;
1631 }
1632
1633 #ifdef CONFIG_MIGRATION
1634 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1635                   struct vm_area_struct *, unsigned long, void *), void *arg)
1636 {
1637         struct stable_node *stable_node;
1638         struct hlist_node *hlist;
1639         struct rmap_item *rmap_item;
1640         int ret = SWAP_AGAIN;
1641         int search_new_forks = 0;
1642
1643         VM_BUG_ON(!PageKsm(page));
1644         VM_BUG_ON(!PageLocked(page));
1645
1646         stable_node = page_stable_node(page);
1647         if (!stable_node)
1648                 return ret;
1649 again:
1650         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1651                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1652                 struct anon_vma_chain *vmac;
1653                 struct vm_area_struct *vma;
1654
1655                 anon_vma_lock(anon_vma);
1656                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1657                         vma = vmac->vma;
1658                         if (rmap_item->address < vma->vm_start ||
1659                             rmap_item->address >= vma->vm_end)
1660                                 continue;
1661                         /*
1662                          * Initially we examine only the vma which covers this
1663                          * rmap_item; but later, if there is still work to do,
1664                          * we examine covering vmas in other mms: in case they
1665                          * were forked from the original since ksmd passed.
1666                          */
1667                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1668                                 continue;
1669
1670                         ret = rmap_one(page, vma, rmap_item->address, arg);
1671                         if (ret != SWAP_AGAIN) {
1672                                 anon_vma_unlock(anon_vma);
1673                                 goto out;
1674                         }
1675                 }
1676                 anon_vma_unlock(anon_vma);
1677         }
1678         if (!search_new_forks++)
1679                 goto again;
1680 out:
1681         return ret;
1682 }
1683
1684 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1685 {
1686         struct stable_node *stable_node;
1687
1688         VM_BUG_ON(!PageLocked(oldpage));
1689         VM_BUG_ON(!PageLocked(newpage));
1690         VM_BUG_ON(newpage->mapping != oldpage->mapping);
1691
1692         stable_node = page_stable_node(newpage);
1693         if (stable_node) {
1694                 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1695                 stable_node->kpfn = page_to_pfn(newpage);
1696         }
1697 }
1698 #endif /* CONFIG_MIGRATION */
1699
1700 #ifdef CONFIG_MEMORY_HOTREMOVE
1701 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1702                                                  unsigned long end_pfn)
1703 {
1704         struct rb_node *node;
1705
1706         for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1707                 struct stable_node *stable_node;
1708
1709                 stable_node = rb_entry(node, struct stable_node, node);
1710                 if (stable_node->kpfn >= start_pfn &&
1711                     stable_node->kpfn < end_pfn)
1712                         return stable_node;
1713         }
1714         return NULL;
1715 }
1716
1717 static int ksm_memory_callback(struct notifier_block *self,
1718                                unsigned long action, void *arg)
1719 {
1720         struct memory_notify *mn = arg;
1721         struct stable_node *stable_node;
1722
1723         switch (action) {
1724         case MEM_GOING_OFFLINE:
1725                 /*
1726                  * Keep it very simple for now: just lock out ksmd and
1727                  * MADV_UNMERGEABLE while any memory is going offline.
1728                  */
1729                 mutex_lock(&ksm_thread_mutex);
1730                 break;
1731
1732         case MEM_OFFLINE:
1733                 /*
1734                  * Most of the work is done by page migration; but there might
1735                  * be a few stable_nodes left over, still pointing to struct
1736                  * pages which have been offlined: prune those from the tree.
1737                  */
1738                 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1739                                         mn->start_pfn + mn->nr_pages)) != NULL)
1740                         remove_node_from_stable_tree(stable_node);
1741                 /* fallthrough */
1742
1743         case MEM_CANCEL_OFFLINE:
1744                 mutex_unlock(&ksm_thread_mutex);
1745                 break;
1746         }
1747         return NOTIFY_OK;
1748 }
1749 #endif /* CONFIG_MEMORY_HOTREMOVE */
1750
1751 #ifdef CONFIG_SYSFS
1752 /*
1753  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1754  */
1755
1756 #define KSM_ATTR_RO(_name) \
1757         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1758 #define KSM_ATTR(_name) \
1759         static struct kobj_attribute _name##_attr = \
1760                 __ATTR(_name, 0644, _name##_show, _name##_store)
1761
1762 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1763                                     struct kobj_attribute *attr, char *buf)
1764 {
1765         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1766 }
1767
1768 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1769                                      struct kobj_attribute *attr,
1770                                      const char *buf, size_t count)
1771 {
1772         unsigned long msecs;
1773         int err;
1774
1775         err = strict_strtoul(buf, 10, &msecs);
1776         if (err || msecs > UINT_MAX)
1777                 return -EINVAL;
1778
1779         ksm_thread_sleep_millisecs = msecs;
1780
1781         return count;
1782 }
1783 KSM_ATTR(sleep_millisecs);
1784
1785 static ssize_t pages_to_scan_show(struct kobject *kobj,
1786                                   struct kobj_attribute *attr, char *buf)
1787 {
1788         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1789 }
1790
1791 static ssize_t pages_to_scan_store(struct kobject *kobj,
1792                                    struct kobj_attribute *attr,
1793                                    const char *buf, size_t count)
1794 {
1795         int err;
1796         unsigned long nr_pages;
1797
1798         err = strict_strtoul(buf, 10, &nr_pages);
1799         if (err || nr_pages > UINT_MAX)
1800                 return -EINVAL;
1801
1802         ksm_thread_pages_to_scan = nr_pages;
1803
1804         return count;
1805 }
1806 KSM_ATTR(pages_to_scan);
1807
1808 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1809                         char *buf)
1810 {
1811         return sprintf(buf, "%u\n", ksm_run);
1812 }
1813
1814 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1815                          const char *buf, size_t count)
1816 {
1817         int err;
1818         unsigned long flags;
1819
1820         err = strict_strtoul(buf, 10, &flags);
1821         if (err || flags > UINT_MAX)
1822                 return -EINVAL;
1823         if (flags > KSM_RUN_UNMERGE)
1824                 return -EINVAL;
1825
1826         /*
1827          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1828          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1829          * breaking COW to free the pages_shared (but leaves mm_slots
1830          * on the list for when ksmd may be set running again).
1831          */
1832
1833         mutex_lock(&ksm_thread_mutex);
1834         if (ksm_run != flags) {
1835                 ksm_run = flags;
1836                 if (flags & KSM_RUN_UNMERGE) {
1837                         current->flags |= PF_OOM_ORIGIN;
1838                         err = unmerge_and_remove_all_rmap_items();
1839                         current->flags &= ~PF_OOM_ORIGIN;
1840                         if (err) {
1841                                 ksm_run = KSM_RUN_STOP;
1842                                 count = err;
1843                         }
1844                 }
1845         }
1846         mutex_unlock(&ksm_thread_mutex);
1847
1848         if (flags & KSM_RUN_MERGE)
1849                 wake_up_interruptible(&ksm_thread_wait);
1850
1851         return count;
1852 }
1853 KSM_ATTR(run);
1854
1855 static ssize_t pages_shared_show(struct kobject *kobj,
1856                                  struct kobj_attribute *attr, char *buf)
1857 {
1858         return sprintf(buf, "%lu\n", ksm_pages_shared);
1859 }
1860 KSM_ATTR_RO(pages_shared);
1861
1862 static ssize_t pages_sharing_show(struct kobject *kobj,
1863                                   struct kobj_attribute *attr, char *buf)
1864 {
1865         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1866 }
1867 KSM_ATTR_RO(pages_sharing);
1868
1869 static ssize_t pages_unshared_show(struct kobject *kobj,
1870                                    struct kobj_attribute *attr, char *buf)
1871 {
1872         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1873 }
1874 KSM_ATTR_RO(pages_unshared);
1875
1876 static ssize_t pages_volatile_show(struct kobject *kobj,
1877                                    struct kobj_attribute *attr, char *buf)
1878 {
1879         long ksm_pages_volatile;
1880
1881         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1882                                 - ksm_pages_sharing - ksm_pages_unshared;
1883         /*
1884          * It was not worth any locking to calculate that statistic,
1885          * but it might therefore sometimes be negative: conceal that.
1886          */
1887         if (ksm_pages_volatile < 0)
1888                 ksm_pages_volatile = 0;
1889         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1890 }
1891 KSM_ATTR_RO(pages_volatile);
1892
1893 static ssize_t full_scans_show(struct kobject *kobj,
1894                                struct kobj_attribute *attr, char *buf)
1895 {
1896         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1897 }
1898 KSM_ATTR_RO(full_scans);
1899
1900 static struct attribute *ksm_attrs[] = {
1901         &sleep_millisecs_attr.attr,
1902         &pages_to_scan_attr.attr,
1903         &run_attr.attr,
1904         &pages_shared_attr.attr,
1905         &pages_sharing_attr.attr,
1906         &pages_unshared_attr.attr,
1907         &pages_volatile_attr.attr,
1908         &full_scans_attr.attr,
1909         NULL,
1910 };
1911
1912 static struct attribute_group ksm_attr_group = {
1913         .attrs = ksm_attrs,
1914         .name = "ksm",
1915 };
1916 #endif /* CONFIG_SYSFS */
1917
1918 static int __init ksm_init(void)
1919 {
1920         struct task_struct *ksm_thread;
1921         int err;
1922
1923         err = ksm_slab_init();
1924         if (err)
1925                 goto out;
1926
1927         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1928         if (IS_ERR(ksm_thread)) {
1929                 printk(KERN_ERR "ksm: creating kthread failed\n");
1930                 err = PTR_ERR(ksm_thread);
1931                 goto out_free;
1932         }
1933
1934 #ifdef CONFIG_SYSFS
1935         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1936         if (err) {
1937                 printk(KERN_ERR "ksm: register sysfs failed\n");
1938                 kthread_stop(ksm_thread);
1939                 goto out_free;
1940         }
1941 #else
1942         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
1943
1944 #endif /* CONFIG_SYSFS */
1945
1946 #ifdef CONFIG_MEMORY_HOTREMOVE
1947         /*
1948          * Choose a high priority since the callback takes ksm_thread_mutex:
1949          * later callbacks could only be taking locks which nest within that.
1950          */
1951         hotplug_memory_notifier(ksm_memory_callback, 100);
1952 #endif
1953         return 0;
1954
1955 out_free:
1956         ksm_slab_free();
1957 out:
1958         return err;
1959 }
1960 module_init(ksm_init)