ksm: hold anon_vma in rmap_item
[linux-2.6.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/mmu_notifier.h>
33 #include <linux/swap.h>
34 #include <linux/ksm.h>
35
36 #include <asm/tlbflush.h>
37 #include "internal.h"
38
39 /*
40  * A few notes about the KSM scanning process,
41  * to make it easier to understand the data structures below:
42  *
43  * In order to reduce excessive scanning, KSM sorts the memory pages by their
44  * contents into a data structure that holds pointers to the pages' locations.
45  *
46  * Since the contents of the pages may change at any moment, KSM cannot just
47  * insert the pages into a normal sorted tree and expect it to find anything.
48  * Therefore KSM uses two data structures - the stable and the unstable tree.
49  *
50  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
51  * by their contents.  Because each such page is write-protected, searching on
52  * this tree is fully assured to be working (except when pages are unmapped),
53  * and therefore this tree is called the stable tree.
54  *
55  * In addition to the stable tree, KSM uses a second data structure called the
56  * unstable tree: this tree holds pointers to pages which have been found to
57  * be "unchanged for a period of time".  The unstable tree sorts these pages
58  * by their contents, but since they are not write-protected, KSM cannot rely
59  * upon the unstable tree to work correctly - the unstable tree is liable to
60  * be corrupted as its contents are modified, and so it is called unstable.
61  *
62  * KSM solves this problem by several techniques:
63  *
64  * 1) The unstable tree is flushed every time KSM completes scanning all
65  *    memory areas, and then the tree is rebuilt again from the beginning.
66  * 2) KSM will only insert into the unstable tree, pages whose hash value
67  *    has not changed since the previous scan of all memory areas.
68  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
69  *    colors of the nodes and not on their contents, assuring that even when
70  *    the tree gets "corrupted" it won't get out of balance, so scanning time
71  *    remains the same (also, searching and inserting nodes in an rbtree uses
72  *    the same algorithm, so we have no overhead when we flush and rebuild).
73  * 4) KSM never flushes the stable tree, which means that even if it were to
74  *    take 10 attempts to find a page in the unstable tree, once it is found,
75  *    it is secured in the stable tree.  (When we scan a new page, we first
76  *    compare it against the stable tree, and then against the unstable tree.)
77  */
78
79 /**
80  * struct mm_slot - ksm information per mm that is being scanned
81  * @link: link to the mm_slots hash list
82  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
83  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
84  * @mm: the mm that this information is valid for
85  */
86 struct mm_slot {
87         struct hlist_node link;
88         struct list_head mm_list;
89         struct rmap_item *rmap_list;
90         struct mm_struct *mm;
91 };
92
93 /**
94  * struct ksm_scan - cursor for scanning
95  * @mm_slot: the current mm_slot we are scanning
96  * @address: the next address inside that to be scanned
97  * @rmap_list: link to the next rmap to be scanned in the rmap_list
98  * @seqnr: count of completed full scans (needed when removing unstable node)
99  *
100  * There is only the one ksm_scan instance of this cursor structure.
101  */
102 struct ksm_scan {
103         struct mm_slot *mm_slot;
104         unsigned long address;
105         struct rmap_item **rmap_list;
106         unsigned long seqnr;
107 };
108
109 /**
110  * struct stable_node - node of the stable rbtree
111  * @page: pointer to struct page of the ksm page
112  * @node: rb node of this ksm page in the stable tree
113  * @hlist: hlist head of rmap_items using this ksm page
114  */
115 struct stable_node {
116         struct page *page;
117         struct rb_node node;
118         struct hlist_head hlist;
119 };
120
121 /**
122  * struct rmap_item - reverse mapping item for virtual addresses
123  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
124  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
125  * @mm: the memory structure this rmap_item is pointing into
126  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
127  * @oldchecksum: previous checksum of the page at that virtual address
128  * @node: rb node of this rmap_item in the unstable tree
129  * @head: pointer to stable_node heading this list in the stable tree
130  * @hlist: link into hlist of rmap_items hanging off that stable_node
131  */
132 struct rmap_item {
133         struct rmap_item *rmap_list;
134         struct anon_vma *anon_vma;      /* when stable */
135         struct mm_struct *mm;
136         unsigned long address;          /* + low bits used for flags below */
137         unsigned int oldchecksum;       /* when unstable */
138         union {
139                 struct rb_node node;    /* when node of unstable tree */
140                 struct {                /* when listed from stable tree */
141                         struct stable_node *head;
142                         struct hlist_node hlist;
143                 };
144         };
145 };
146
147 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
148 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
149 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
150
151 /* The stable and unstable tree heads */
152 static struct rb_root root_stable_tree = RB_ROOT;
153 static struct rb_root root_unstable_tree = RB_ROOT;
154
155 #define MM_SLOTS_HASH_HEADS 1024
156 static struct hlist_head *mm_slots_hash;
157
158 static struct mm_slot ksm_mm_head = {
159         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
160 };
161 static struct ksm_scan ksm_scan = {
162         .mm_slot = &ksm_mm_head,
163 };
164
165 static struct kmem_cache *rmap_item_cache;
166 static struct kmem_cache *stable_node_cache;
167 static struct kmem_cache *mm_slot_cache;
168
169 /* The number of nodes in the stable tree */
170 static unsigned long ksm_pages_shared;
171
172 /* The number of page slots additionally sharing those nodes */
173 static unsigned long ksm_pages_sharing;
174
175 /* The number of nodes in the unstable tree */
176 static unsigned long ksm_pages_unshared;
177
178 /* The number of rmap_items in use: to calculate pages_volatile */
179 static unsigned long ksm_rmap_items;
180
181 /* Limit on the number of unswappable pages used */
182 static unsigned long ksm_max_kernel_pages;
183
184 /* Number of pages ksmd should scan in one batch */
185 static unsigned int ksm_thread_pages_to_scan = 100;
186
187 /* Milliseconds ksmd should sleep between batches */
188 static unsigned int ksm_thread_sleep_millisecs = 20;
189
190 #define KSM_RUN_STOP    0
191 #define KSM_RUN_MERGE   1
192 #define KSM_RUN_UNMERGE 2
193 static unsigned int ksm_run = KSM_RUN_STOP;
194
195 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
196 static DEFINE_MUTEX(ksm_thread_mutex);
197 static DEFINE_SPINLOCK(ksm_mmlist_lock);
198
199 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
200                 sizeof(struct __struct), __alignof__(struct __struct),\
201                 (__flags), NULL)
202
203 static int __init ksm_slab_init(void)
204 {
205         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
206         if (!rmap_item_cache)
207                 goto out;
208
209         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
210         if (!stable_node_cache)
211                 goto out_free1;
212
213         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
214         if (!mm_slot_cache)
215                 goto out_free2;
216
217         return 0;
218
219 out_free2:
220         kmem_cache_destroy(stable_node_cache);
221 out_free1:
222         kmem_cache_destroy(rmap_item_cache);
223 out:
224         return -ENOMEM;
225 }
226
227 static void __init ksm_slab_free(void)
228 {
229         kmem_cache_destroy(mm_slot_cache);
230         kmem_cache_destroy(stable_node_cache);
231         kmem_cache_destroy(rmap_item_cache);
232         mm_slot_cache = NULL;
233 }
234
235 static inline struct rmap_item *alloc_rmap_item(void)
236 {
237         struct rmap_item *rmap_item;
238
239         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
240         if (rmap_item)
241                 ksm_rmap_items++;
242         return rmap_item;
243 }
244
245 static inline void free_rmap_item(struct rmap_item *rmap_item)
246 {
247         ksm_rmap_items--;
248         rmap_item->mm = NULL;   /* debug safety */
249         kmem_cache_free(rmap_item_cache, rmap_item);
250 }
251
252 static inline struct stable_node *alloc_stable_node(void)
253 {
254         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
255 }
256
257 static inline void free_stable_node(struct stable_node *stable_node)
258 {
259         kmem_cache_free(stable_node_cache, stable_node);
260 }
261
262 static inline struct mm_slot *alloc_mm_slot(void)
263 {
264         if (!mm_slot_cache)     /* initialization failed */
265                 return NULL;
266         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
267 }
268
269 static inline void free_mm_slot(struct mm_slot *mm_slot)
270 {
271         kmem_cache_free(mm_slot_cache, mm_slot);
272 }
273
274 static int __init mm_slots_hash_init(void)
275 {
276         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
277                                 GFP_KERNEL);
278         if (!mm_slots_hash)
279                 return -ENOMEM;
280         return 0;
281 }
282
283 static void __init mm_slots_hash_free(void)
284 {
285         kfree(mm_slots_hash);
286 }
287
288 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
289 {
290         struct mm_slot *mm_slot;
291         struct hlist_head *bucket;
292         struct hlist_node *node;
293
294         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
295                                 % MM_SLOTS_HASH_HEADS];
296         hlist_for_each_entry(mm_slot, node, bucket, link) {
297                 if (mm == mm_slot->mm)
298                         return mm_slot;
299         }
300         return NULL;
301 }
302
303 static void insert_to_mm_slots_hash(struct mm_struct *mm,
304                                     struct mm_slot *mm_slot)
305 {
306         struct hlist_head *bucket;
307
308         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
309                                 % MM_SLOTS_HASH_HEADS];
310         mm_slot->mm = mm;
311         hlist_add_head(&mm_slot->link, bucket);
312 }
313
314 static inline int in_stable_tree(struct rmap_item *rmap_item)
315 {
316         return rmap_item->address & STABLE_FLAG;
317 }
318
319 static void hold_anon_vma(struct rmap_item *rmap_item,
320                           struct anon_vma *anon_vma)
321 {
322         rmap_item->anon_vma = anon_vma;
323         atomic_inc(&anon_vma->ksm_refcount);
324 }
325
326 static void drop_anon_vma(struct rmap_item *rmap_item)
327 {
328         struct anon_vma *anon_vma = rmap_item->anon_vma;
329
330         if (atomic_dec_and_lock(&anon_vma->ksm_refcount, &anon_vma->lock)) {
331                 int empty = list_empty(&anon_vma->head);
332                 spin_unlock(&anon_vma->lock);
333                 if (empty)
334                         anon_vma_free(anon_vma);
335         }
336 }
337
338 /*
339  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
340  * page tables after it has passed through ksm_exit() - which, if necessary,
341  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
342  * a special flag: they can just back out as soon as mm_users goes to zero.
343  * ksm_test_exit() is used throughout to make this test for exit: in some
344  * places for correctness, in some places just to avoid unnecessary work.
345  */
346 static inline bool ksm_test_exit(struct mm_struct *mm)
347 {
348         return atomic_read(&mm->mm_users) == 0;
349 }
350
351 /*
352  * We use break_ksm to break COW on a ksm page: it's a stripped down
353  *
354  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
355  *              put_page(page);
356  *
357  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
358  * in case the application has unmapped and remapped mm,addr meanwhile.
359  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
360  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
361  */
362 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
363 {
364         struct page *page;
365         int ret = 0;
366
367         do {
368                 cond_resched();
369                 page = follow_page(vma, addr, FOLL_GET);
370                 if (!page)
371                         break;
372                 if (PageKsm(page))
373                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
374                                                         FAULT_FLAG_WRITE);
375                 else
376                         ret = VM_FAULT_WRITE;
377                 put_page(page);
378         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
379         /*
380          * We must loop because handle_mm_fault() may back out if there's
381          * any difficulty e.g. if pte accessed bit gets updated concurrently.
382          *
383          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
384          * COW has been broken, even if the vma does not permit VM_WRITE;
385          * but note that a concurrent fault might break PageKsm for us.
386          *
387          * VM_FAULT_SIGBUS could occur if we race with truncation of the
388          * backing file, which also invalidates anonymous pages: that's
389          * okay, that truncation will have unmapped the PageKsm for us.
390          *
391          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
392          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
393          * current task has TIF_MEMDIE set, and will be OOM killed on return
394          * to user; and ksmd, having no mm, would never be chosen for that.
395          *
396          * But if the mm is in a limited mem_cgroup, then the fault may fail
397          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
398          * even ksmd can fail in this way - though it's usually breaking ksm
399          * just to undo a merge it made a moment before, so unlikely to oom.
400          *
401          * That's a pity: we might therefore have more kernel pages allocated
402          * than we're counting as nodes in the stable tree; but ksm_do_scan
403          * will retry to break_cow on each pass, so should recover the page
404          * in due course.  The important thing is to not let VM_MERGEABLE
405          * be cleared while any such pages might remain in the area.
406          */
407         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
408 }
409
410 static void break_cow(struct rmap_item *rmap_item)
411 {
412         struct mm_struct *mm = rmap_item->mm;
413         unsigned long addr = rmap_item->address;
414         struct vm_area_struct *vma;
415
416         down_read(&mm->mmap_sem);
417         if (ksm_test_exit(mm))
418                 goto out;
419         vma = find_vma(mm, addr);
420         if (!vma || vma->vm_start > addr)
421                 goto out;
422         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
423                 goto out;
424         break_ksm(vma, addr);
425 out:
426         up_read(&mm->mmap_sem);
427 }
428
429 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
430 {
431         struct mm_struct *mm = rmap_item->mm;
432         unsigned long addr = rmap_item->address;
433         struct vm_area_struct *vma;
434         struct page *page;
435
436         down_read(&mm->mmap_sem);
437         if (ksm_test_exit(mm))
438                 goto out;
439         vma = find_vma(mm, addr);
440         if (!vma || vma->vm_start > addr)
441                 goto out;
442         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
443                 goto out;
444
445         page = follow_page(vma, addr, FOLL_GET);
446         if (!page)
447                 goto out;
448         if (PageAnon(page)) {
449                 flush_anon_page(vma, page, addr);
450                 flush_dcache_page(page);
451         } else {
452                 put_page(page);
453 out:            page = NULL;
454         }
455         up_read(&mm->mmap_sem);
456         return page;
457 }
458
459 /*
460  * Removing rmap_item from stable or unstable tree.
461  * This function will clean the information from the stable/unstable tree.
462  */
463 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
464 {
465         if (rmap_item->address & STABLE_FLAG) {
466                 struct stable_node *stable_node;
467                 struct page *page;
468
469                 stable_node = rmap_item->head;
470                 page = stable_node->page;
471                 lock_page(page);
472
473                 hlist_del(&rmap_item->hlist);
474                 if (stable_node->hlist.first) {
475                         unlock_page(page);
476                         ksm_pages_sharing--;
477                 } else {
478                         set_page_stable_node(page, NULL);
479                         unlock_page(page);
480                         put_page(page);
481
482                         rb_erase(&stable_node->node, &root_stable_tree);
483                         free_stable_node(stable_node);
484                         ksm_pages_shared--;
485                 }
486
487                 drop_anon_vma(rmap_item);
488                 rmap_item->address &= PAGE_MASK;
489
490         } else if (rmap_item->address & UNSTABLE_FLAG) {
491                 unsigned char age;
492                 /*
493                  * Usually ksmd can and must skip the rb_erase, because
494                  * root_unstable_tree was already reset to RB_ROOT.
495                  * But be careful when an mm is exiting: do the rb_erase
496                  * if this rmap_item was inserted by this scan, rather
497                  * than left over from before.
498                  */
499                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
500                 BUG_ON(age > 1);
501                 if (!age)
502                         rb_erase(&rmap_item->node, &root_unstable_tree);
503
504                 ksm_pages_unshared--;
505                 rmap_item->address &= PAGE_MASK;
506         }
507
508         cond_resched();         /* we're called from many long loops */
509 }
510
511 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
512                                        struct rmap_item **rmap_list)
513 {
514         while (*rmap_list) {
515                 struct rmap_item *rmap_item = *rmap_list;
516                 *rmap_list = rmap_item->rmap_list;
517                 remove_rmap_item_from_tree(rmap_item);
518                 free_rmap_item(rmap_item);
519         }
520 }
521
522 /*
523  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
524  * than check every pte of a given vma, the locking doesn't quite work for
525  * that - an rmap_item is assigned to the stable tree after inserting ksm
526  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
527  * rmap_items from parent to child at fork time (so as not to waste time
528  * if exit comes before the next scan reaches it).
529  *
530  * Similarly, although we'd like to remove rmap_items (so updating counts
531  * and freeing memory) when unmerging an area, it's easier to leave that
532  * to the next pass of ksmd - consider, for example, how ksmd might be
533  * in cmp_and_merge_page on one of the rmap_items we would be removing.
534  */
535 static int unmerge_ksm_pages(struct vm_area_struct *vma,
536                              unsigned long start, unsigned long end)
537 {
538         unsigned long addr;
539         int err = 0;
540
541         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
542                 if (ksm_test_exit(vma->vm_mm))
543                         break;
544                 if (signal_pending(current))
545                         err = -ERESTARTSYS;
546                 else
547                         err = break_ksm(vma, addr);
548         }
549         return err;
550 }
551
552 #ifdef CONFIG_SYSFS
553 /*
554  * Only called through the sysfs control interface:
555  */
556 static int unmerge_and_remove_all_rmap_items(void)
557 {
558         struct mm_slot *mm_slot;
559         struct mm_struct *mm;
560         struct vm_area_struct *vma;
561         int err = 0;
562
563         spin_lock(&ksm_mmlist_lock);
564         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
565                                                 struct mm_slot, mm_list);
566         spin_unlock(&ksm_mmlist_lock);
567
568         for (mm_slot = ksm_scan.mm_slot;
569                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
570                 mm = mm_slot->mm;
571                 down_read(&mm->mmap_sem);
572                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
573                         if (ksm_test_exit(mm))
574                                 break;
575                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
576                                 continue;
577                         err = unmerge_ksm_pages(vma,
578                                                 vma->vm_start, vma->vm_end);
579                         if (err)
580                                 goto error;
581                 }
582
583                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
584
585                 spin_lock(&ksm_mmlist_lock);
586                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
587                                                 struct mm_slot, mm_list);
588                 if (ksm_test_exit(mm)) {
589                         hlist_del(&mm_slot->link);
590                         list_del(&mm_slot->mm_list);
591                         spin_unlock(&ksm_mmlist_lock);
592
593                         free_mm_slot(mm_slot);
594                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
595                         up_read(&mm->mmap_sem);
596                         mmdrop(mm);
597                 } else {
598                         spin_unlock(&ksm_mmlist_lock);
599                         up_read(&mm->mmap_sem);
600                 }
601         }
602
603         ksm_scan.seqnr = 0;
604         return 0;
605
606 error:
607         up_read(&mm->mmap_sem);
608         spin_lock(&ksm_mmlist_lock);
609         ksm_scan.mm_slot = &ksm_mm_head;
610         spin_unlock(&ksm_mmlist_lock);
611         return err;
612 }
613 #endif /* CONFIG_SYSFS */
614
615 static u32 calc_checksum(struct page *page)
616 {
617         u32 checksum;
618         void *addr = kmap_atomic(page, KM_USER0);
619         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
620         kunmap_atomic(addr, KM_USER0);
621         return checksum;
622 }
623
624 static int memcmp_pages(struct page *page1, struct page *page2)
625 {
626         char *addr1, *addr2;
627         int ret;
628
629         addr1 = kmap_atomic(page1, KM_USER0);
630         addr2 = kmap_atomic(page2, KM_USER1);
631         ret = memcmp(addr1, addr2, PAGE_SIZE);
632         kunmap_atomic(addr2, KM_USER1);
633         kunmap_atomic(addr1, KM_USER0);
634         return ret;
635 }
636
637 static inline int pages_identical(struct page *page1, struct page *page2)
638 {
639         return !memcmp_pages(page1, page2);
640 }
641
642 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
643                               pte_t *orig_pte)
644 {
645         struct mm_struct *mm = vma->vm_mm;
646         unsigned long addr;
647         pte_t *ptep;
648         spinlock_t *ptl;
649         int swapped;
650         int err = -EFAULT;
651
652         addr = page_address_in_vma(page, vma);
653         if (addr == -EFAULT)
654                 goto out;
655
656         ptep = page_check_address(page, mm, addr, &ptl, 0);
657         if (!ptep)
658                 goto out;
659
660         if (pte_write(*ptep)) {
661                 pte_t entry;
662
663                 swapped = PageSwapCache(page);
664                 flush_cache_page(vma, addr, page_to_pfn(page));
665                 /*
666                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
667                  * take any lock, therefore the check that we are going to make
668                  * with the pagecount against the mapcount is racey and
669                  * O_DIRECT can happen right after the check.
670                  * So we clear the pte and flush the tlb before the check
671                  * this assure us that no O_DIRECT can happen after the check
672                  * or in the middle of the check.
673                  */
674                 entry = ptep_clear_flush(vma, addr, ptep);
675                 /*
676                  * Check that no O_DIRECT or similar I/O is in progress on the
677                  * page
678                  */
679                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
680                         set_pte_at_notify(mm, addr, ptep, entry);
681                         goto out_unlock;
682                 }
683                 entry = pte_wrprotect(entry);
684                 set_pte_at_notify(mm, addr, ptep, entry);
685         }
686         *orig_pte = *ptep;
687         err = 0;
688
689 out_unlock:
690         pte_unmap_unlock(ptep, ptl);
691 out:
692         return err;
693 }
694
695 /**
696  * replace_page - replace page in vma by new ksm page
697  * @vma:      vma that holds the pte pointing to page
698  * @page:     the page we are replacing by kpage
699  * @kpage:    the ksm page we replace page by
700  * @orig_pte: the original value of the pte
701  *
702  * Returns 0 on success, -EFAULT on failure.
703  */
704 static int replace_page(struct vm_area_struct *vma, struct page *page,
705                         struct page *kpage, pte_t orig_pte)
706 {
707         struct mm_struct *mm = vma->vm_mm;
708         pgd_t *pgd;
709         pud_t *pud;
710         pmd_t *pmd;
711         pte_t *ptep;
712         spinlock_t *ptl;
713         unsigned long addr;
714         int err = -EFAULT;
715
716         addr = page_address_in_vma(page, vma);
717         if (addr == -EFAULT)
718                 goto out;
719
720         pgd = pgd_offset(mm, addr);
721         if (!pgd_present(*pgd))
722                 goto out;
723
724         pud = pud_offset(pgd, addr);
725         if (!pud_present(*pud))
726                 goto out;
727
728         pmd = pmd_offset(pud, addr);
729         if (!pmd_present(*pmd))
730                 goto out;
731
732         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
733         if (!pte_same(*ptep, orig_pte)) {
734                 pte_unmap_unlock(ptep, ptl);
735                 goto out;
736         }
737
738         get_page(kpage);
739         page_add_anon_rmap(kpage, vma, addr);
740
741         flush_cache_page(vma, addr, pte_pfn(*ptep));
742         ptep_clear_flush(vma, addr, ptep);
743         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
744
745         page_remove_rmap(page);
746         put_page(page);
747
748         pte_unmap_unlock(ptep, ptl);
749         err = 0;
750 out:
751         return err;
752 }
753
754 /*
755  * try_to_merge_one_page - take two pages and merge them into one
756  * @vma: the vma that holds the pte pointing to page
757  * @page: the PageAnon page that we want to replace with kpage
758  * @kpage: the PageKsm page that we want to map instead of page
759  *
760  * This function returns 0 if the pages were merged, -EFAULT otherwise.
761  */
762 static int try_to_merge_one_page(struct vm_area_struct *vma,
763                                  struct page *page, struct page *kpage)
764 {
765         pte_t orig_pte = __pte(0);
766         int err = -EFAULT;
767
768         if (page == kpage)                      /* ksm page forked */
769                 return 0;
770
771         if (!(vma->vm_flags & VM_MERGEABLE))
772                 goto out;
773         if (!PageAnon(page))
774                 goto out;
775
776         /*
777          * We need the page lock to read a stable PageSwapCache in
778          * write_protect_page().  We use trylock_page() instead of
779          * lock_page() because we don't want to wait here - we
780          * prefer to continue scanning and merging different pages,
781          * then come back to this page when it is unlocked.
782          */
783         if (!trylock_page(page))
784                 goto out;
785         /*
786          * If this anonymous page is mapped only here, its pte may need
787          * to be write-protected.  If it's mapped elsewhere, all of its
788          * ptes are necessarily already write-protected.  But in either
789          * case, we need to lock and check page_count is not raised.
790          */
791         if (write_protect_page(vma, page, &orig_pte) == 0 &&
792             pages_identical(page, kpage))
793                 err = replace_page(vma, page, kpage, orig_pte);
794
795         if ((vma->vm_flags & VM_LOCKED) && !err) {
796                 munlock_vma_page(page);
797                 if (!PageMlocked(kpage)) {
798                         unlock_page(page);
799                         lru_add_drain();
800                         lock_page(kpage);
801                         mlock_vma_page(kpage);
802                         page = kpage;           /* for final unlock */
803                 }
804         }
805
806         unlock_page(page);
807 out:
808         return err;
809 }
810
811 /*
812  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
813  * but no new kernel page is allocated: kpage must already be a ksm page.
814  *
815  * This function returns 0 if the pages were merged, -EFAULT otherwise.
816  */
817 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
818                                       struct page *page, struct page *kpage)
819 {
820         struct mm_struct *mm = rmap_item->mm;
821         struct vm_area_struct *vma;
822         int err = -EFAULT;
823
824         down_read(&mm->mmap_sem);
825         if (ksm_test_exit(mm))
826                 goto out;
827         vma = find_vma(mm, rmap_item->address);
828         if (!vma || vma->vm_start > rmap_item->address)
829                 goto out;
830
831         err = try_to_merge_one_page(vma, page, kpage);
832         if (err)
833                 goto out;
834
835         /* Must get reference to anon_vma while still holding mmap_sem */
836         hold_anon_vma(rmap_item, vma->anon_vma);
837 out:
838         up_read(&mm->mmap_sem);
839         return err;
840 }
841
842 /*
843  * try_to_merge_two_pages - take two identical pages and prepare them
844  * to be merged into one page.
845  *
846  * This function returns the kpage if we successfully merged two identical
847  * pages into one ksm page, NULL otherwise.
848  *
849  * Note that this function allocates a new kernel page: if one of the pages
850  * is already a ksm page, try_to_merge_with_ksm_page should be used.
851  */
852 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
853                                            struct page *page,
854                                            struct rmap_item *tree_rmap_item,
855                                            struct page *tree_page)
856 {
857         struct mm_struct *mm = rmap_item->mm;
858         struct vm_area_struct *vma;
859         struct page *kpage;
860         int err = -EFAULT;
861
862         /*
863          * The number of nodes in the stable tree
864          * is the number of kernel pages that we hold.
865          */
866         if (ksm_max_kernel_pages &&
867             ksm_max_kernel_pages <= ksm_pages_shared)
868                 return NULL;
869
870         kpage = alloc_page(GFP_HIGHUSER);
871         if (!kpage)
872                 return NULL;
873
874         down_read(&mm->mmap_sem);
875         if (ksm_test_exit(mm))
876                 goto up;
877         vma = find_vma(mm, rmap_item->address);
878         if (!vma || vma->vm_start > rmap_item->address)
879                 goto up;
880
881         copy_user_highpage(kpage, page, rmap_item->address, vma);
882
883         SetPageDirty(kpage);
884         __SetPageUptodate(kpage);
885         SetPageSwapBacked(kpage);
886         set_page_stable_node(kpage, NULL);      /* mark it PageKsm */
887         lru_cache_add_lru(kpage, LRU_ACTIVE_ANON);
888
889         err = try_to_merge_one_page(vma, page, kpage);
890         if (err)
891                 goto up;
892
893         /* Must get reference to anon_vma while still holding mmap_sem */
894         hold_anon_vma(rmap_item, vma->anon_vma);
895 up:
896         up_read(&mm->mmap_sem);
897
898         if (!err) {
899                 err = try_to_merge_with_ksm_page(tree_rmap_item,
900                                                         tree_page, kpage);
901                 /*
902                  * If that fails, we have a ksm page with only one pte
903                  * pointing to it: so break it.
904                  */
905                 if (err) {
906                         drop_anon_vma(rmap_item);
907                         break_cow(rmap_item);
908                 }
909         }
910         if (err) {
911                 put_page(kpage);
912                 kpage = NULL;
913         }
914         return kpage;
915 }
916
917 /*
918  * stable_tree_search - search for page inside the stable tree
919  *
920  * This function checks if there is a page inside the stable tree
921  * with identical content to the page that we are scanning right now.
922  *
923  * This function returns the stable tree node of identical content if found,
924  * NULL otherwise.
925  */
926 static struct stable_node *stable_tree_search(struct page *page)
927 {
928         struct rb_node *node = root_stable_tree.rb_node;
929         struct stable_node *stable_node;
930
931         stable_node = page_stable_node(page);
932         if (stable_node) {                      /* ksm page forked */
933                 get_page(page);
934                 return stable_node;
935         }
936
937         while (node) {
938                 int ret;
939
940                 cond_resched();
941                 stable_node = rb_entry(node, struct stable_node, node);
942
943                 ret = memcmp_pages(page, stable_node->page);
944
945                 if (ret < 0)
946                         node = node->rb_left;
947                 else if (ret > 0)
948                         node = node->rb_right;
949                 else {
950                         get_page(stable_node->page);
951                         return stable_node;
952                 }
953         }
954
955         return NULL;
956 }
957
958 /*
959  * stable_tree_insert - insert rmap_item pointing to new ksm page
960  * into the stable tree.
961  *
962  * This function returns the stable tree node just allocated on success,
963  * NULL otherwise.
964  */
965 static struct stable_node *stable_tree_insert(struct page *kpage)
966 {
967         struct rb_node **new = &root_stable_tree.rb_node;
968         struct rb_node *parent = NULL;
969         struct stable_node *stable_node;
970
971         while (*new) {
972                 int ret;
973
974                 cond_resched();
975                 stable_node = rb_entry(*new, struct stable_node, node);
976
977                 ret = memcmp_pages(kpage, stable_node->page);
978
979                 parent = *new;
980                 if (ret < 0)
981                         new = &parent->rb_left;
982                 else if (ret > 0)
983                         new = &parent->rb_right;
984                 else {
985                         /*
986                          * It is not a bug that stable_tree_search() didn't
987                          * find this node: because at that time our page was
988                          * not yet write-protected, so may have changed since.
989                          */
990                         return NULL;
991                 }
992         }
993
994         stable_node = alloc_stable_node();
995         if (!stable_node)
996                 return NULL;
997
998         rb_link_node(&stable_node->node, parent, new);
999         rb_insert_color(&stable_node->node, &root_stable_tree);
1000
1001         INIT_HLIST_HEAD(&stable_node->hlist);
1002
1003         get_page(kpage);
1004         stable_node->page = kpage;
1005         set_page_stable_node(kpage, stable_node);
1006
1007         return stable_node;
1008 }
1009
1010 /*
1011  * unstable_tree_search_insert - search for identical page,
1012  * else insert rmap_item into the unstable tree.
1013  *
1014  * This function searches for a page in the unstable tree identical to the
1015  * page currently being scanned; and if no identical page is found in the
1016  * tree, we insert rmap_item as a new object into the unstable tree.
1017  *
1018  * This function returns pointer to rmap_item found to be identical
1019  * to the currently scanned page, NULL otherwise.
1020  *
1021  * This function does both searching and inserting, because they share
1022  * the same walking algorithm in an rbtree.
1023  */
1024 static
1025 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1026                                               struct page *page,
1027                                               struct page **tree_pagep)
1028
1029 {
1030         struct rb_node **new = &root_unstable_tree.rb_node;
1031         struct rb_node *parent = NULL;
1032
1033         while (*new) {
1034                 struct rmap_item *tree_rmap_item;
1035                 struct page *tree_page;
1036                 int ret;
1037
1038                 cond_resched();
1039                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1040                 tree_page = get_mergeable_page(tree_rmap_item);
1041                 if (!tree_page)
1042                         return NULL;
1043
1044                 /*
1045                  * Don't substitute a ksm page for a forked page.
1046                  */
1047                 if (page == tree_page) {
1048                         put_page(tree_page);
1049                         return NULL;
1050                 }
1051
1052                 ret = memcmp_pages(page, tree_page);
1053
1054                 parent = *new;
1055                 if (ret < 0) {
1056                         put_page(tree_page);
1057                         new = &parent->rb_left;
1058                 } else if (ret > 0) {
1059                         put_page(tree_page);
1060                         new = &parent->rb_right;
1061                 } else {
1062                         *tree_pagep = tree_page;
1063                         return tree_rmap_item;
1064                 }
1065         }
1066
1067         rmap_item->address |= UNSTABLE_FLAG;
1068         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1069         rb_link_node(&rmap_item->node, parent, new);
1070         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1071
1072         ksm_pages_unshared++;
1073         return NULL;
1074 }
1075
1076 /*
1077  * stable_tree_append - add another rmap_item to the linked list of
1078  * rmap_items hanging off a given node of the stable tree, all sharing
1079  * the same ksm page.
1080  */
1081 static void stable_tree_append(struct rmap_item *rmap_item,
1082                                struct stable_node *stable_node)
1083 {
1084         rmap_item->head = stable_node;
1085         rmap_item->address |= STABLE_FLAG;
1086         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1087
1088         if (rmap_item->hlist.next)
1089                 ksm_pages_sharing++;
1090         else
1091                 ksm_pages_shared++;
1092 }
1093
1094 /*
1095  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1096  * if not, compare checksum to previous and if it's the same, see if page can
1097  * be inserted into the unstable tree, or merged with a page already there and
1098  * both transferred to the stable tree.
1099  *
1100  * @page: the page that we are searching identical page to.
1101  * @rmap_item: the reverse mapping into the virtual address of this page
1102  */
1103 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1104 {
1105         struct rmap_item *tree_rmap_item;
1106         struct page *tree_page = NULL;
1107         struct stable_node *stable_node;
1108         struct page *kpage;
1109         unsigned int checksum;
1110         int err;
1111
1112         remove_rmap_item_from_tree(rmap_item);
1113
1114         /* We first start with searching the page inside the stable tree */
1115         stable_node = stable_tree_search(page);
1116         if (stable_node) {
1117                 kpage = stable_node->page;
1118                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1119                 if (!err) {
1120                         /*
1121                          * The page was successfully merged:
1122                          * add its rmap_item to the stable tree.
1123                          */
1124                         lock_page(kpage);
1125                         stable_tree_append(rmap_item, stable_node);
1126                         unlock_page(kpage);
1127                 }
1128                 put_page(kpage);
1129                 return;
1130         }
1131
1132         /*
1133          * A ksm page might have got here by fork, but its other
1134          * references have already been removed from the stable tree.
1135          * Or it might be left over from a break_ksm which failed
1136          * when the mem_cgroup had reached its limit: try again now.
1137          */
1138         if (PageKsm(page))
1139                 break_cow(rmap_item);
1140
1141         /*
1142          * In case the hash value of the page was changed from the last time we
1143          * have calculated it, this page to be changed frequely, therefore we
1144          * don't want to insert it to the unstable tree, and we don't want to
1145          * waste our time to search if there is something identical to it there.
1146          */
1147         checksum = calc_checksum(page);
1148         if (rmap_item->oldchecksum != checksum) {
1149                 rmap_item->oldchecksum = checksum;
1150                 return;
1151         }
1152
1153         tree_rmap_item =
1154                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1155         if (tree_rmap_item) {
1156                 kpage = try_to_merge_two_pages(rmap_item, page,
1157                                                 tree_rmap_item, tree_page);
1158                 put_page(tree_page);
1159                 /*
1160                  * As soon as we merge this page, we want to remove the
1161                  * rmap_item of the page we have merged with from the unstable
1162                  * tree, and insert it instead as new node in the stable tree.
1163                  */
1164                 if (kpage) {
1165                         remove_rmap_item_from_tree(tree_rmap_item);
1166
1167                         lock_page(kpage);
1168                         stable_node = stable_tree_insert(kpage);
1169                         if (stable_node) {
1170                                 stable_tree_append(tree_rmap_item, stable_node);
1171                                 stable_tree_append(rmap_item, stable_node);
1172                         }
1173                         unlock_page(kpage);
1174                         put_page(kpage);
1175
1176                         /*
1177                          * If we fail to insert the page into the stable tree,
1178                          * we will have 2 virtual addresses that are pointing
1179                          * to a ksm page left outside the stable tree,
1180                          * in which case we need to break_cow on both.
1181                          */
1182                         if (!stable_node) {
1183                                 drop_anon_vma(tree_rmap_item);
1184                                 break_cow(tree_rmap_item);
1185                                 drop_anon_vma(rmap_item);
1186                                 break_cow(rmap_item);
1187                         }
1188                 }
1189         }
1190 }
1191
1192 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1193                                             struct rmap_item **rmap_list,
1194                                             unsigned long addr)
1195 {
1196         struct rmap_item *rmap_item;
1197
1198         while (*rmap_list) {
1199                 rmap_item = *rmap_list;
1200                 if ((rmap_item->address & PAGE_MASK) == addr)
1201                         return rmap_item;
1202                 if (rmap_item->address > addr)
1203                         break;
1204                 *rmap_list = rmap_item->rmap_list;
1205                 remove_rmap_item_from_tree(rmap_item);
1206                 free_rmap_item(rmap_item);
1207         }
1208
1209         rmap_item = alloc_rmap_item();
1210         if (rmap_item) {
1211                 /* It has already been zeroed */
1212                 rmap_item->mm = mm_slot->mm;
1213                 rmap_item->address = addr;
1214                 rmap_item->rmap_list = *rmap_list;
1215                 *rmap_list = rmap_item;
1216         }
1217         return rmap_item;
1218 }
1219
1220 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1221 {
1222         struct mm_struct *mm;
1223         struct mm_slot *slot;
1224         struct vm_area_struct *vma;
1225         struct rmap_item *rmap_item;
1226
1227         if (list_empty(&ksm_mm_head.mm_list))
1228                 return NULL;
1229
1230         slot = ksm_scan.mm_slot;
1231         if (slot == &ksm_mm_head) {
1232                 root_unstable_tree = RB_ROOT;
1233
1234                 spin_lock(&ksm_mmlist_lock);
1235                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1236                 ksm_scan.mm_slot = slot;
1237                 spin_unlock(&ksm_mmlist_lock);
1238 next_mm:
1239                 ksm_scan.address = 0;
1240                 ksm_scan.rmap_list = &slot->rmap_list;
1241         }
1242
1243         mm = slot->mm;
1244         down_read(&mm->mmap_sem);
1245         if (ksm_test_exit(mm))
1246                 vma = NULL;
1247         else
1248                 vma = find_vma(mm, ksm_scan.address);
1249
1250         for (; vma; vma = vma->vm_next) {
1251                 if (!(vma->vm_flags & VM_MERGEABLE))
1252                         continue;
1253                 if (ksm_scan.address < vma->vm_start)
1254                         ksm_scan.address = vma->vm_start;
1255                 if (!vma->anon_vma)
1256                         ksm_scan.address = vma->vm_end;
1257
1258                 while (ksm_scan.address < vma->vm_end) {
1259                         if (ksm_test_exit(mm))
1260                                 break;
1261                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1262                         if (*page && PageAnon(*page)) {
1263                                 flush_anon_page(vma, *page, ksm_scan.address);
1264                                 flush_dcache_page(*page);
1265                                 rmap_item = get_next_rmap_item(slot,
1266                                         ksm_scan.rmap_list, ksm_scan.address);
1267                                 if (rmap_item) {
1268                                         ksm_scan.rmap_list =
1269                                                         &rmap_item->rmap_list;
1270                                         ksm_scan.address += PAGE_SIZE;
1271                                 } else
1272                                         put_page(*page);
1273                                 up_read(&mm->mmap_sem);
1274                                 return rmap_item;
1275                         }
1276                         if (*page)
1277                                 put_page(*page);
1278                         ksm_scan.address += PAGE_SIZE;
1279                         cond_resched();
1280                 }
1281         }
1282
1283         if (ksm_test_exit(mm)) {
1284                 ksm_scan.address = 0;
1285                 ksm_scan.rmap_list = &slot->rmap_list;
1286         }
1287         /*
1288          * Nuke all the rmap_items that are above this current rmap:
1289          * because there were no VM_MERGEABLE vmas with such addresses.
1290          */
1291         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1292
1293         spin_lock(&ksm_mmlist_lock);
1294         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1295                                                 struct mm_slot, mm_list);
1296         if (ksm_scan.address == 0) {
1297                 /*
1298                  * We've completed a full scan of all vmas, holding mmap_sem
1299                  * throughout, and found no VM_MERGEABLE: so do the same as
1300                  * __ksm_exit does to remove this mm from all our lists now.
1301                  * This applies either when cleaning up after __ksm_exit
1302                  * (but beware: we can reach here even before __ksm_exit),
1303                  * or when all VM_MERGEABLE areas have been unmapped (and
1304                  * mmap_sem then protects against race with MADV_MERGEABLE).
1305                  */
1306                 hlist_del(&slot->link);
1307                 list_del(&slot->mm_list);
1308                 spin_unlock(&ksm_mmlist_lock);
1309
1310                 free_mm_slot(slot);
1311                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1312                 up_read(&mm->mmap_sem);
1313                 mmdrop(mm);
1314         } else {
1315                 spin_unlock(&ksm_mmlist_lock);
1316                 up_read(&mm->mmap_sem);
1317         }
1318
1319         /* Repeat until we've completed scanning the whole list */
1320         slot = ksm_scan.mm_slot;
1321         if (slot != &ksm_mm_head)
1322                 goto next_mm;
1323
1324         ksm_scan.seqnr++;
1325         return NULL;
1326 }
1327
1328 /**
1329  * ksm_do_scan  - the ksm scanner main worker function.
1330  * @scan_npages - number of pages we want to scan before we return.
1331  */
1332 static void ksm_do_scan(unsigned int scan_npages)
1333 {
1334         struct rmap_item *rmap_item;
1335         struct page *page;
1336
1337         while (scan_npages--) {
1338                 cond_resched();
1339                 rmap_item = scan_get_next_rmap_item(&page);
1340                 if (!rmap_item)
1341                         return;
1342                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1343                         cmp_and_merge_page(page, rmap_item);
1344                 put_page(page);
1345         }
1346 }
1347
1348 static int ksmd_should_run(void)
1349 {
1350         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1351 }
1352
1353 static int ksm_scan_thread(void *nothing)
1354 {
1355         set_user_nice(current, 5);
1356
1357         while (!kthread_should_stop()) {
1358                 mutex_lock(&ksm_thread_mutex);
1359                 if (ksmd_should_run())
1360                         ksm_do_scan(ksm_thread_pages_to_scan);
1361                 mutex_unlock(&ksm_thread_mutex);
1362
1363                 if (ksmd_should_run()) {
1364                         schedule_timeout_interruptible(
1365                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1366                 } else {
1367                         wait_event_interruptible(ksm_thread_wait,
1368                                 ksmd_should_run() || kthread_should_stop());
1369                 }
1370         }
1371         return 0;
1372 }
1373
1374 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1375                 unsigned long end, int advice, unsigned long *vm_flags)
1376 {
1377         struct mm_struct *mm = vma->vm_mm;
1378         int err;
1379
1380         switch (advice) {
1381         case MADV_MERGEABLE:
1382                 /*
1383                  * Be somewhat over-protective for now!
1384                  */
1385                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1386                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1387                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1388                                  VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1389                         return 0;               /* just ignore the advice */
1390
1391                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1392                         err = __ksm_enter(mm);
1393                         if (err)
1394                                 return err;
1395                 }
1396
1397                 *vm_flags |= VM_MERGEABLE;
1398                 break;
1399
1400         case MADV_UNMERGEABLE:
1401                 if (!(*vm_flags & VM_MERGEABLE))
1402                         return 0;               /* just ignore the advice */
1403
1404                 if (vma->anon_vma) {
1405                         err = unmerge_ksm_pages(vma, start, end);
1406                         if (err)
1407                                 return err;
1408                 }
1409
1410                 *vm_flags &= ~VM_MERGEABLE;
1411                 break;
1412         }
1413
1414         return 0;
1415 }
1416
1417 int __ksm_enter(struct mm_struct *mm)
1418 {
1419         struct mm_slot *mm_slot;
1420         int needs_wakeup;
1421
1422         mm_slot = alloc_mm_slot();
1423         if (!mm_slot)
1424                 return -ENOMEM;
1425
1426         /* Check ksm_run too?  Would need tighter locking */
1427         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1428
1429         spin_lock(&ksm_mmlist_lock);
1430         insert_to_mm_slots_hash(mm, mm_slot);
1431         /*
1432          * Insert just behind the scanning cursor, to let the area settle
1433          * down a little; when fork is followed by immediate exec, we don't
1434          * want ksmd to waste time setting up and tearing down an rmap_list.
1435          */
1436         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1437         spin_unlock(&ksm_mmlist_lock);
1438
1439         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1440         atomic_inc(&mm->mm_count);
1441
1442         if (needs_wakeup)
1443                 wake_up_interruptible(&ksm_thread_wait);
1444
1445         return 0;
1446 }
1447
1448 void __ksm_exit(struct mm_struct *mm)
1449 {
1450         struct mm_slot *mm_slot;
1451         int easy_to_free = 0;
1452
1453         /*
1454          * This process is exiting: if it's straightforward (as is the
1455          * case when ksmd was never running), free mm_slot immediately.
1456          * But if it's at the cursor or has rmap_items linked to it, use
1457          * mmap_sem to synchronize with any break_cows before pagetables
1458          * are freed, and leave the mm_slot on the list for ksmd to free.
1459          * Beware: ksm may already have noticed it exiting and freed the slot.
1460          */
1461
1462         spin_lock(&ksm_mmlist_lock);
1463         mm_slot = get_mm_slot(mm);
1464         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1465                 if (!mm_slot->rmap_list) {
1466                         hlist_del(&mm_slot->link);
1467                         list_del(&mm_slot->mm_list);
1468                         easy_to_free = 1;
1469                 } else {
1470                         list_move(&mm_slot->mm_list,
1471                                   &ksm_scan.mm_slot->mm_list);
1472                 }
1473         }
1474         spin_unlock(&ksm_mmlist_lock);
1475
1476         if (easy_to_free) {
1477                 free_mm_slot(mm_slot);
1478                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1479                 mmdrop(mm);
1480         } else if (mm_slot) {
1481                 down_write(&mm->mmap_sem);
1482                 up_write(&mm->mmap_sem);
1483         }
1484 }
1485
1486 struct page *ksm_does_need_to_copy(struct page *page,
1487                         struct vm_area_struct *vma, unsigned long address)
1488 {
1489         struct page *new_page;
1490
1491         unlock_page(page);      /* any racers will COW it, not modify it */
1492
1493         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1494         if (new_page) {
1495                 copy_user_highpage(new_page, page, address, vma);
1496
1497                 SetPageDirty(new_page);
1498                 __SetPageUptodate(new_page);
1499                 SetPageSwapBacked(new_page);
1500                 __set_page_locked(new_page);
1501
1502                 if (page_evictable(new_page, vma))
1503                         lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1504                 else
1505                         add_page_to_unevictable_list(new_page);
1506         }
1507
1508         page_cache_release(page);
1509         return new_page;
1510 }
1511
1512 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1513                         unsigned long *vm_flags)
1514 {
1515         struct stable_node *stable_node;
1516         struct rmap_item *rmap_item;
1517         struct hlist_node *hlist;
1518         unsigned int mapcount = page_mapcount(page);
1519         int referenced = 0;
1520         int search_new_forks = 0;
1521
1522         VM_BUG_ON(!PageKsm(page));
1523         VM_BUG_ON(!PageLocked(page));
1524
1525         stable_node = page_stable_node(page);
1526         if (!stable_node)
1527                 return 0;
1528 again:
1529         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1530                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1531                 struct vm_area_struct *vma;
1532
1533                 spin_lock(&anon_vma->lock);
1534                 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1535                         if (rmap_item->address < vma->vm_start ||
1536                             rmap_item->address >= vma->vm_end)
1537                                 continue;
1538                         /*
1539                          * Initially we examine only the vma which covers this
1540                          * rmap_item; but later, if there is still work to do,
1541                          * we examine covering vmas in other mms: in case they
1542                          * were forked from the original since ksmd passed.
1543                          */
1544                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1545                                 continue;
1546
1547                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1548                                 continue;
1549
1550                         referenced += page_referenced_one(page, vma,
1551                                 rmap_item->address, &mapcount, vm_flags);
1552                         if (!search_new_forks || !mapcount)
1553                                 break;
1554                 }
1555                 spin_unlock(&anon_vma->lock);
1556                 if (!mapcount)
1557                         goto out;
1558         }
1559         if (!search_new_forks++)
1560                 goto again;
1561 out:
1562         return referenced;
1563 }
1564
1565 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1566 {
1567         struct stable_node *stable_node;
1568         struct hlist_node *hlist;
1569         struct rmap_item *rmap_item;
1570         int ret = SWAP_AGAIN;
1571         int search_new_forks = 0;
1572
1573         VM_BUG_ON(!PageKsm(page));
1574         VM_BUG_ON(!PageLocked(page));
1575
1576         stable_node = page_stable_node(page);
1577         if (!stable_node)
1578                 return SWAP_FAIL;
1579 again:
1580         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1581                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1582                 struct vm_area_struct *vma;
1583
1584                 spin_lock(&anon_vma->lock);
1585                 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1586                         if (rmap_item->address < vma->vm_start ||
1587                             rmap_item->address >= vma->vm_end)
1588                                 continue;
1589                         /*
1590                          * Initially we examine only the vma which covers this
1591                          * rmap_item; but later, if there is still work to do,
1592                          * we examine covering vmas in other mms: in case they
1593                          * were forked from the original since ksmd passed.
1594                          */
1595                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1596                                 continue;
1597
1598                         ret = try_to_unmap_one(page, vma,
1599                                         rmap_item->address, flags);
1600                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1601                                 spin_unlock(&anon_vma->lock);
1602                                 goto out;
1603                         }
1604                 }
1605                 spin_unlock(&anon_vma->lock);
1606         }
1607         if (!search_new_forks++)
1608                 goto again;
1609 out:
1610         return ret;
1611 }
1612
1613 #ifdef CONFIG_SYSFS
1614 /*
1615  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1616  */
1617
1618 #define KSM_ATTR_RO(_name) \
1619         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1620 #define KSM_ATTR(_name) \
1621         static struct kobj_attribute _name##_attr = \
1622                 __ATTR(_name, 0644, _name##_show, _name##_store)
1623
1624 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1625                                     struct kobj_attribute *attr, char *buf)
1626 {
1627         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1628 }
1629
1630 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1631                                      struct kobj_attribute *attr,
1632                                      const char *buf, size_t count)
1633 {
1634         unsigned long msecs;
1635         int err;
1636
1637         err = strict_strtoul(buf, 10, &msecs);
1638         if (err || msecs > UINT_MAX)
1639                 return -EINVAL;
1640
1641         ksm_thread_sleep_millisecs = msecs;
1642
1643         return count;
1644 }
1645 KSM_ATTR(sleep_millisecs);
1646
1647 static ssize_t pages_to_scan_show(struct kobject *kobj,
1648                                   struct kobj_attribute *attr, char *buf)
1649 {
1650         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1651 }
1652
1653 static ssize_t pages_to_scan_store(struct kobject *kobj,
1654                                    struct kobj_attribute *attr,
1655                                    const char *buf, size_t count)
1656 {
1657         int err;
1658         unsigned long nr_pages;
1659
1660         err = strict_strtoul(buf, 10, &nr_pages);
1661         if (err || nr_pages > UINT_MAX)
1662                 return -EINVAL;
1663
1664         ksm_thread_pages_to_scan = nr_pages;
1665
1666         return count;
1667 }
1668 KSM_ATTR(pages_to_scan);
1669
1670 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1671                         char *buf)
1672 {
1673         return sprintf(buf, "%u\n", ksm_run);
1674 }
1675
1676 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1677                          const char *buf, size_t count)
1678 {
1679         int err;
1680         unsigned long flags;
1681
1682         err = strict_strtoul(buf, 10, &flags);
1683         if (err || flags > UINT_MAX)
1684                 return -EINVAL;
1685         if (flags > KSM_RUN_UNMERGE)
1686                 return -EINVAL;
1687
1688         /*
1689          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1690          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1691          * breaking COW to free the unswappable pages_shared (but leaves
1692          * mm_slots on the list for when ksmd may be set running again).
1693          */
1694
1695         mutex_lock(&ksm_thread_mutex);
1696         if (ksm_run != flags) {
1697                 ksm_run = flags;
1698                 if (flags & KSM_RUN_UNMERGE) {
1699                         current->flags |= PF_OOM_ORIGIN;
1700                         err = unmerge_and_remove_all_rmap_items();
1701                         current->flags &= ~PF_OOM_ORIGIN;
1702                         if (err) {
1703                                 ksm_run = KSM_RUN_STOP;
1704                                 count = err;
1705                         }
1706                 }
1707         }
1708         mutex_unlock(&ksm_thread_mutex);
1709
1710         if (flags & KSM_RUN_MERGE)
1711                 wake_up_interruptible(&ksm_thread_wait);
1712
1713         return count;
1714 }
1715 KSM_ATTR(run);
1716
1717 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1718                                       struct kobj_attribute *attr,
1719                                       const char *buf, size_t count)
1720 {
1721         int err;
1722         unsigned long nr_pages;
1723
1724         err = strict_strtoul(buf, 10, &nr_pages);
1725         if (err)
1726                 return -EINVAL;
1727
1728         ksm_max_kernel_pages = nr_pages;
1729
1730         return count;
1731 }
1732
1733 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1734                                      struct kobj_attribute *attr, char *buf)
1735 {
1736         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1737 }
1738 KSM_ATTR(max_kernel_pages);
1739
1740 static ssize_t pages_shared_show(struct kobject *kobj,
1741                                  struct kobj_attribute *attr, char *buf)
1742 {
1743         return sprintf(buf, "%lu\n", ksm_pages_shared);
1744 }
1745 KSM_ATTR_RO(pages_shared);
1746
1747 static ssize_t pages_sharing_show(struct kobject *kobj,
1748                                   struct kobj_attribute *attr, char *buf)
1749 {
1750         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1751 }
1752 KSM_ATTR_RO(pages_sharing);
1753
1754 static ssize_t pages_unshared_show(struct kobject *kobj,
1755                                    struct kobj_attribute *attr, char *buf)
1756 {
1757         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1758 }
1759 KSM_ATTR_RO(pages_unshared);
1760
1761 static ssize_t pages_volatile_show(struct kobject *kobj,
1762                                    struct kobj_attribute *attr, char *buf)
1763 {
1764         long ksm_pages_volatile;
1765
1766         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1767                                 - ksm_pages_sharing - ksm_pages_unshared;
1768         /*
1769          * It was not worth any locking to calculate that statistic,
1770          * but it might therefore sometimes be negative: conceal that.
1771          */
1772         if (ksm_pages_volatile < 0)
1773                 ksm_pages_volatile = 0;
1774         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1775 }
1776 KSM_ATTR_RO(pages_volatile);
1777
1778 static ssize_t full_scans_show(struct kobject *kobj,
1779                                struct kobj_attribute *attr, char *buf)
1780 {
1781         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1782 }
1783 KSM_ATTR_RO(full_scans);
1784
1785 static struct attribute *ksm_attrs[] = {
1786         &sleep_millisecs_attr.attr,
1787         &pages_to_scan_attr.attr,
1788         &run_attr.attr,
1789         &max_kernel_pages_attr.attr,
1790         &pages_shared_attr.attr,
1791         &pages_sharing_attr.attr,
1792         &pages_unshared_attr.attr,
1793         &pages_volatile_attr.attr,
1794         &full_scans_attr.attr,
1795         NULL,
1796 };
1797
1798 static struct attribute_group ksm_attr_group = {
1799         .attrs = ksm_attrs,
1800         .name = "ksm",
1801 };
1802 #endif /* CONFIG_SYSFS */
1803
1804 static int __init ksm_init(void)
1805 {
1806         struct task_struct *ksm_thread;
1807         int err;
1808
1809         ksm_max_kernel_pages = totalram_pages / 4;
1810
1811         err = ksm_slab_init();
1812         if (err)
1813                 goto out;
1814
1815         err = mm_slots_hash_init();
1816         if (err)
1817                 goto out_free1;
1818
1819         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1820         if (IS_ERR(ksm_thread)) {
1821                 printk(KERN_ERR "ksm: creating kthread failed\n");
1822                 err = PTR_ERR(ksm_thread);
1823                 goto out_free2;
1824         }
1825
1826 #ifdef CONFIG_SYSFS
1827         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1828         if (err) {
1829                 printk(KERN_ERR "ksm: register sysfs failed\n");
1830                 kthread_stop(ksm_thread);
1831                 goto out_free2;
1832         }
1833 #else
1834         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
1835
1836 #endif /* CONFIG_SYSFS */
1837
1838         return 0;
1839
1840 out_free2:
1841         mm_slots_hash_free();
1842 out_free1:
1843         ksm_slab_free();
1844 out:
1845         return err;
1846 }
1847 module_init(ksm_init)