ksm: unmerge is an origin of OOMs
[linux-3.10.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/mmu_notifier.h>
33 #include <linux/ksm.h>
34
35 #include <asm/tlbflush.h>
36
37 /*
38  * A few notes about the KSM scanning process,
39  * to make it easier to understand the data structures below:
40  *
41  * In order to reduce excessive scanning, KSM sorts the memory pages by their
42  * contents into a data structure that holds pointers to the pages' locations.
43  *
44  * Since the contents of the pages may change at any moment, KSM cannot just
45  * insert the pages into a normal sorted tree and expect it to find anything.
46  * Therefore KSM uses two data structures - the stable and the unstable tree.
47  *
48  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
49  * by their contents.  Because each such page is write-protected, searching on
50  * this tree is fully assured to be working (except when pages are unmapped),
51  * and therefore this tree is called the stable tree.
52  *
53  * In addition to the stable tree, KSM uses a second data structure called the
54  * unstable tree: this tree holds pointers to pages which have been found to
55  * be "unchanged for a period of time".  The unstable tree sorts these pages
56  * by their contents, but since they are not write-protected, KSM cannot rely
57  * upon the unstable tree to work correctly - the unstable tree is liable to
58  * be corrupted as its contents are modified, and so it is called unstable.
59  *
60  * KSM solves this problem by several techniques:
61  *
62  * 1) The unstable tree is flushed every time KSM completes scanning all
63  *    memory areas, and then the tree is rebuilt again from the beginning.
64  * 2) KSM will only insert into the unstable tree, pages whose hash value
65  *    has not changed since the previous scan of all memory areas.
66  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
67  *    colors of the nodes and not on their contents, assuring that even when
68  *    the tree gets "corrupted" it won't get out of balance, so scanning time
69  *    remains the same (also, searching and inserting nodes in an rbtree uses
70  *    the same algorithm, so we have no overhead when we flush and rebuild).
71  * 4) KSM never flushes the stable tree, which means that even if it were to
72  *    take 10 attempts to find a page in the unstable tree, once it is found,
73  *    it is secured in the stable tree.  (When we scan a new page, we first
74  *    compare it against the stable tree, and then against the unstable tree.)
75  */
76
77 /**
78  * struct mm_slot - ksm information per mm that is being scanned
79  * @link: link to the mm_slots hash list
80  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
81  * @rmap_list: head for this mm_slot's list of rmap_items
82  * @mm: the mm that this information is valid for
83  */
84 struct mm_slot {
85         struct hlist_node link;
86         struct list_head mm_list;
87         struct list_head rmap_list;
88         struct mm_struct *mm;
89 };
90
91 /**
92  * struct ksm_scan - cursor for scanning
93  * @mm_slot: the current mm_slot we are scanning
94  * @address: the next address inside that to be scanned
95  * @rmap_item: the current rmap that we are scanning inside the rmap_list
96  * @seqnr: count of completed full scans (needed when removing unstable node)
97  *
98  * There is only the one ksm_scan instance of this cursor structure.
99  */
100 struct ksm_scan {
101         struct mm_slot *mm_slot;
102         unsigned long address;
103         struct rmap_item *rmap_item;
104         unsigned long seqnr;
105 };
106
107 /**
108  * struct rmap_item - reverse mapping item for virtual addresses
109  * @link: link into mm_slot's rmap_list (rmap_list is per mm)
110  * @mm: the memory structure this rmap_item is pointing into
111  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
112  * @oldchecksum: previous checksum of the page at that virtual address
113  * @node: rb_node of this rmap_item in either unstable or stable tree
114  * @next: next rmap_item hanging off the same node of the stable tree
115  * @prev: previous rmap_item hanging off the same node of the stable tree
116  */
117 struct rmap_item {
118         struct list_head link;
119         struct mm_struct *mm;
120         unsigned long address;          /* + low bits used for flags below */
121         union {
122                 unsigned int oldchecksum;               /* when unstable */
123                 struct rmap_item *next;                 /* when stable */
124         };
125         union {
126                 struct rb_node node;                    /* when tree node */
127                 struct rmap_item *prev;                 /* in stable list */
128         };
129 };
130
131 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
132 #define NODE_FLAG       0x100   /* is a node of unstable or stable tree */
133 #define STABLE_FLAG     0x200   /* is a node or list item of stable tree */
134
135 /* The stable and unstable tree heads */
136 static struct rb_root root_stable_tree = RB_ROOT;
137 static struct rb_root root_unstable_tree = RB_ROOT;
138
139 #define MM_SLOTS_HASH_HEADS 1024
140 static struct hlist_head *mm_slots_hash;
141
142 static struct mm_slot ksm_mm_head = {
143         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
144 };
145 static struct ksm_scan ksm_scan = {
146         .mm_slot = &ksm_mm_head,
147 };
148
149 static struct kmem_cache *rmap_item_cache;
150 static struct kmem_cache *mm_slot_cache;
151
152 /* The number of nodes in the stable tree */
153 static unsigned long ksm_pages_shared;
154
155 /* The number of page slots additionally sharing those nodes */
156 static unsigned long ksm_pages_sharing;
157
158 /* The number of nodes in the unstable tree */
159 static unsigned long ksm_pages_unshared;
160
161 /* The number of rmap_items in use: to calculate pages_volatile */
162 static unsigned long ksm_rmap_items;
163
164 /* Limit on the number of unswappable pages used */
165 static unsigned long ksm_max_kernel_pages = 2000;
166
167 /* Number of pages ksmd should scan in one batch */
168 static unsigned int ksm_thread_pages_to_scan = 200;
169
170 /* Milliseconds ksmd should sleep between batches */
171 static unsigned int ksm_thread_sleep_millisecs = 20;
172
173 #define KSM_RUN_STOP    0
174 #define KSM_RUN_MERGE   1
175 #define KSM_RUN_UNMERGE 2
176 static unsigned int ksm_run = KSM_RUN_MERGE;
177
178 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
179 static DEFINE_MUTEX(ksm_thread_mutex);
180 static DEFINE_SPINLOCK(ksm_mmlist_lock);
181
182 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
183                 sizeof(struct __struct), __alignof__(struct __struct),\
184                 (__flags), NULL)
185
186 static int __init ksm_slab_init(void)
187 {
188         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
189         if (!rmap_item_cache)
190                 goto out;
191
192         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
193         if (!mm_slot_cache)
194                 goto out_free;
195
196         return 0;
197
198 out_free:
199         kmem_cache_destroy(rmap_item_cache);
200 out:
201         return -ENOMEM;
202 }
203
204 static void __init ksm_slab_free(void)
205 {
206         kmem_cache_destroy(mm_slot_cache);
207         kmem_cache_destroy(rmap_item_cache);
208         mm_slot_cache = NULL;
209 }
210
211 static inline struct rmap_item *alloc_rmap_item(void)
212 {
213         struct rmap_item *rmap_item;
214
215         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
216         if (rmap_item)
217                 ksm_rmap_items++;
218         return rmap_item;
219 }
220
221 static inline void free_rmap_item(struct rmap_item *rmap_item)
222 {
223         ksm_rmap_items--;
224         rmap_item->mm = NULL;   /* debug safety */
225         kmem_cache_free(rmap_item_cache, rmap_item);
226 }
227
228 static inline struct mm_slot *alloc_mm_slot(void)
229 {
230         if (!mm_slot_cache)     /* initialization failed */
231                 return NULL;
232         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
233 }
234
235 static inline void free_mm_slot(struct mm_slot *mm_slot)
236 {
237         kmem_cache_free(mm_slot_cache, mm_slot);
238 }
239
240 static int __init mm_slots_hash_init(void)
241 {
242         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
243                                 GFP_KERNEL);
244         if (!mm_slots_hash)
245                 return -ENOMEM;
246         return 0;
247 }
248
249 static void __init mm_slots_hash_free(void)
250 {
251         kfree(mm_slots_hash);
252 }
253
254 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
255 {
256         struct mm_slot *mm_slot;
257         struct hlist_head *bucket;
258         struct hlist_node *node;
259
260         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
261                                 % MM_SLOTS_HASH_HEADS];
262         hlist_for_each_entry(mm_slot, node, bucket, link) {
263                 if (mm == mm_slot->mm)
264                         return mm_slot;
265         }
266         return NULL;
267 }
268
269 static void insert_to_mm_slots_hash(struct mm_struct *mm,
270                                     struct mm_slot *mm_slot)
271 {
272         struct hlist_head *bucket;
273
274         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
275                                 % MM_SLOTS_HASH_HEADS];
276         mm_slot->mm = mm;
277         INIT_LIST_HEAD(&mm_slot->rmap_list);
278         hlist_add_head(&mm_slot->link, bucket);
279 }
280
281 static inline int in_stable_tree(struct rmap_item *rmap_item)
282 {
283         return rmap_item->address & STABLE_FLAG;
284 }
285
286 /*
287  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
288  * page tables after it has passed through ksm_exit() - which, if necessary,
289  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
290  * a special flag: they can just back out as soon as mm_users goes to zero.
291  * ksm_test_exit() is used throughout to make this test for exit: in some
292  * places for correctness, in some places just to avoid unnecessary work.
293  */
294 static inline bool ksm_test_exit(struct mm_struct *mm)
295 {
296         return atomic_read(&mm->mm_users) == 0;
297 }
298
299 /*
300  * We use break_ksm to break COW on a ksm page: it's a stripped down
301  *
302  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
303  *              put_page(page);
304  *
305  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
306  * in case the application has unmapped and remapped mm,addr meanwhile.
307  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
308  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
309  */
310 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
311 {
312         struct page *page;
313         int ret = 0;
314
315         do {
316                 cond_resched();
317                 page = follow_page(vma, addr, FOLL_GET);
318                 if (!page)
319                         break;
320                 if (PageKsm(page))
321                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
322                                                         FAULT_FLAG_WRITE);
323                 else
324                         ret = VM_FAULT_WRITE;
325                 put_page(page);
326         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
327         /*
328          * We must loop because handle_mm_fault() may back out if there's
329          * any difficulty e.g. if pte accessed bit gets updated concurrently.
330          *
331          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
332          * COW has been broken, even if the vma does not permit VM_WRITE;
333          * but note that a concurrent fault might break PageKsm for us.
334          *
335          * VM_FAULT_SIGBUS could occur if we race with truncation of the
336          * backing file, which also invalidates anonymous pages: that's
337          * okay, that truncation will have unmapped the PageKsm for us.
338          *
339          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
340          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
341          * current task has TIF_MEMDIE set, and will be OOM killed on return
342          * to user; and ksmd, having no mm, would never be chosen for that.
343          *
344          * But if the mm is in a limited mem_cgroup, then the fault may fail
345          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
346          * even ksmd can fail in this way - though it's usually breaking ksm
347          * just to undo a merge it made a moment before, so unlikely to oom.
348          *
349          * That's a pity: we might therefore have more kernel pages allocated
350          * than we're counting as nodes in the stable tree; but ksm_do_scan
351          * will retry to break_cow on each pass, so should recover the page
352          * in due course.  The important thing is to not let VM_MERGEABLE
353          * be cleared while any such pages might remain in the area.
354          */
355         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
356 }
357
358 static void break_cow(struct mm_struct *mm, unsigned long addr)
359 {
360         struct vm_area_struct *vma;
361
362         down_read(&mm->mmap_sem);
363         if (ksm_test_exit(mm))
364                 goto out;
365         vma = find_vma(mm, addr);
366         if (!vma || vma->vm_start > addr)
367                 goto out;
368         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
369                 goto out;
370         break_ksm(vma, addr);
371 out:
372         up_read(&mm->mmap_sem);
373 }
374
375 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
376 {
377         struct mm_struct *mm = rmap_item->mm;
378         unsigned long addr = rmap_item->address;
379         struct vm_area_struct *vma;
380         struct page *page;
381
382         down_read(&mm->mmap_sem);
383         if (ksm_test_exit(mm))
384                 goto out;
385         vma = find_vma(mm, addr);
386         if (!vma || vma->vm_start > addr)
387                 goto out;
388         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
389                 goto out;
390
391         page = follow_page(vma, addr, FOLL_GET);
392         if (!page)
393                 goto out;
394         if (PageAnon(page)) {
395                 flush_anon_page(vma, page, addr);
396                 flush_dcache_page(page);
397         } else {
398                 put_page(page);
399 out:            page = NULL;
400         }
401         up_read(&mm->mmap_sem);
402         return page;
403 }
404
405 /*
406  * get_ksm_page: checks if the page at the virtual address in rmap_item
407  * is still PageKsm, in which case we can trust the content of the page,
408  * and it returns the gotten page; but NULL if the page has been zapped.
409  */
410 static struct page *get_ksm_page(struct rmap_item *rmap_item)
411 {
412         struct page *page;
413
414         page = get_mergeable_page(rmap_item);
415         if (page && !PageKsm(page)) {
416                 put_page(page);
417                 page = NULL;
418         }
419         return page;
420 }
421
422 /*
423  * Removing rmap_item from stable or unstable tree.
424  * This function will clean the information from the stable/unstable tree.
425  */
426 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
427 {
428         if (in_stable_tree(rmap_item)) {
429                 struct rmap_item *next_item = rmap_item->next;
430
431                 if (rmap_item->address & NODE_FLAG) {
432                         if (next_item) {
433                                 rb_replace_node(&rmap_item->node,
434                                                 &next_item->node,
435                                                 &root_stable_tree);
436                                 next_item->address |= NODE_FLAG;
437                                 ksm_pages_sharing--;
438                         } else {
439                                 rb_erase(&rmap_item->node, &root_stable_tree);
440                                 ksm_pages_shared--;
441                         }
442                 } else {
443                         struct rmap_item *prev_item = rmap_item->prev;
444
445                         BUG_ON(prev_item->next != rmap_item);
446                         prev_item->next = next_item;
447                         if (next_item) {
448                                 BUG_ON(next_item->prev != rmap_item);
449                                 next_item->prev = rmap_item->prev;
450                         }
451                         ksm_pages_sharing--;
452                 }
453
454                 rmap_item->next = NULL;
455
456         } else if (rmap_item->address & NODE_FLAG) {
457                 unsigned char age;
458                 /*
459                  * Usually ksmd can and must skip the rb_erase, because
460                  * root_unstable_tree was already reset to RB_ROOT.
461                  * But be careful when an mm is exiting: do the rb_erase
462                  * if this rmap_item was inserted by this scan, rather
463                  * than left over from before.
464                  */
465                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
466                 BUG_ON(age > 1);
467                 if (!age)
468                         rb_erase(&rmap_item->node, &root_unstable_tree);
469                 ksm_pages_unshared--;
470         }
471
472         rmap_item->address &= PAGE_MASK;
473
474         cond_resched();         /* we're called from many long loops */
475 }
476
477 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
478                                        struct list_head *cur)
479 {
480         struct rmap_item *rmap_item;
481
482         while (cur != &mm_slot->rmap_list) {
483                 rmap_item = list_entry(cur, struct rmap_item, link);
484                 cur = cur->next;
485                 remove_rmap_item_from_tree(rmap_item);
486                 list_del(&rmap_item->link);
487                 free_rmap_item(rmap_item);
488         }
489 }
490
491 /*
492  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
493  * than check every pte of a given vma, the locking doesn't quite work for
494  * that - an rmap_item is assigned to the stable tree after inserting ksm
495  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
496  * rmap_items from parent to child at fork time (so as not to waste time
497  * if exit comes before the next scan reaches it).
498  *
499  * Similarly, although we'd like to remove rmap_items (so updating counts
500  * and freeing memory) when unmerging an area, it's easier to leave that
501  * to the next pass of ksmd - consider, for example, how ksmd might be
502  * in cmp_and_merge_page on one of the rmap_items we would be removing.
503  */
504 static int unmerge_ksm_pages(struct vm_area_struct *vma,
505                              unsigned long start, unsigned long end)
506 {
507         unsigned long addr;
508         int err = 0;
509
510         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
511                 if (ksm_test_exit(vma->vm_mm))
512                         break;
513                 if (signal_pending(current))
514                         err = -ERESTARTSYS;
515                 else
516                         err = break_ksm(vma, addr);
517         }
518         return err;
519 }
520
521 #ifdef CONFIG_SYSFS
522 /*
523  * Only called through the sysfs control interface:
524  */
525 static int unmerge_and_remove_all_rmap_items(void)
526 {
527         struct mm_slot *mm_slot;
528         struct mm_struct *mm;
529         struct vm_area_struct *vma;
530         int err = 0;
531
532         spin_lock(&ksm_mmlist_lock);
533         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
534                                                 struct mm_slot, mm_list);
535         spin_unlock(&ksm_mmlist_lock);
536
537         for (mm_slot = ksm_scan.mm_slot;
538                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
539                 mm = mm_slot->mm;
540                 down_read(&mm->mmap_sem);
541                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
542                         if (ksm_test_exit(mm))
543                                 break;
544                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
545                                 continue;
546                         err = unmerge_ksm_pages(vma,
547                                                 vma->vm_start, vma->vm_end);
548                         if (err)
549                                 goto error;
550                 }
551
552                 remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
553
554                 spin_lock(&ksm_mmlist_lock);
555                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
556                                                 struct mm_slot, mm_list);
557                 if (ksm_test_exit(mm)) {
558                         hlist_del(&mm_slot->link);
559                         list_del(&mm_slot->mm_list);
560                         spin_unlock(&ksm_mmlist_lock);
561
562                         free_mm_slot(mm_slot);
563                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
564                         up_read(&mm->mmap_sem);
565                         mmdrop(mm);
566                 } else {
567                         spin_unlock(&ksm_mmlist_lock);
568                         up_read(&mm->mmap_sem);
569                 }
570         }
571
572         ksm_scan.seqnr = 0;
573         return 0;
574
575 error:
576         up_read(&mm->mmap_sem);
577         spin_lock(&ksm_mmlist_lock);
578         ksm_scan.mm_slot = &ksm_mm_head;
579         spin_unlock(&ksm_mmlist_lock);
580         return err;
581 }
582 #endif /* CONFIG_SYSFS */
583
584 static u32 calc_checksum(struct page *page)
585 {
586         u32 checksum;
587         void *addr = kmap_atomic(page, KM_USER0);
588         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
589         kunmap_atomic(addr, KM_USER0);
590         return checksum;
591 }
592
593 static int memcmp_pages(struct page *page1, struct page *page2)
594 {
595         char *addr1, *addr2;
596         int ret;
597
598         addr1 = kmap_atomic(page1, KM_USER0);
599         addr2 = kmap_atomic(page2, KM_USER1);
600         ret = memcmp(addr1, addr2, PAGE_SIZE);
601         kunmap_atomic(addr2, KM_USER1);
602         kunmap_atomic(addr1, KM_USER0);
603         return ret;
604 }
605
606 static inline int pages_identical(struct page *page1, struct page *page2)
607 {
608         return !memcmp_pages(page1, page2);
609 }
610
611 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
612                               pte_t *orig_pte)
613 {
614         struct mm_struct *mm = vma->vm_mm;
615         unsigned long addr;
616         pte_t *ptep;
617         spinlock_t *ptl;
618         int swapped;
619         int err = -EFAULT;
620
621         addr = page_address_in_vma(page, vma);
622         if (addr == -EFAULT)
623                 goto out;
624
625         ptep = page_check_address(page, mm, addr, &ptl, 0);
626         if (!ptep)
627                 goto out;
628
629         if (pte_write(*ptep)) {
630                 pte_t entry;
631
632                 swapped = PageSwapCache(page);
633                 flush_cache_page(vma, addr, page_to_pfn(page));
634                 /*
635                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
636                  * take any lock, therefore the check that we are going to make
637                  * with the pagecount against the mapcount is racey and
638                  * O_DIRECT can happen right after the check.
639                  * So we clear the pte and flush the tlb before the check
640                  * this assure us that no O_DIRECT can happen after the check
641                  * or in the middle of the check.
642                  */
643                 entry = ptep_clear_flush(vma, addr, ptep);
644                 /*
645                  * Check that no O_DIRECT or similar I/O is in progress on the
646                  * page
647                  */
648                 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
649                         set_pte_at_notify(mm, addr, ptep, entry);
650                         goto out_unlock;
651                 }
652                 entry = pte_wrprotect(entry);
653                 set_pte_at_notify(mm, addr, ptep, entry);
654         }
655         *orig_pte = *ptep;
656         err = 0;
657
658 out_unlock:
659         pte_unmap_unlock(ptep, ptl);
660 out:
661         return err;
662 }
663
664 /**
665  * replace_page - replace page in vma by new ksm page
666  * @vma:      vma that holds the pte pointing to oldpage
667  * @oldpage:  the page we are replacing by newpage
668  * @newpage:  the ksm page we replace oldpage by
669  * @orig_pte: the original value of the pte
670  *
671  * Returns 0 on success, -EFAULT on failure.
672  */
673 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
674                         struct page *newpage, pte_t orig_pte)
675 {
676         struct mm_struct *mm = vma->vm_mm;
677         pgd_t *pgd;
678         pud_t *pud;
679         pmd_t *pmd;
680         pte_t *ptep;
681         spinlock_t *ptl;
682         unsigned long addr;
683         pgprot_t prot;
684         int err = -EFAULT;
685
686         prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
687
688         addr = page_address_in_vma(oldpage, vma);
689         if (addr == -EFAULT)
690                 goto out;
691
692         pgd = pgd_offset(mm, addr);
693         if (!pgd_present(*pgd))
694                 goto out;
695
696         pud = pud_offset(pgd, addr);
697         if (!pud_present(*pud))
698                 goto out;
699
700         pmd = pmd_offset(pud, addr);
701         if (!pmd_present(*pmd))
702                 goto out;
703
704         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
705         if (!pte_same(*ptep, orig_pte)) {
706                 pte_unmap_unlock(ptep, ptl);
707                 goto out;
708         }
709
710         get_page(newpage);
711         page_add_ksm_rmap(newpage);
712
713         flush_cache_page(vma, addr, pte_pfn(*ptep));
714         ptep_clear_flush(vma, addr, ptep);
715         set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
716
717         page_remove_rmap(oldpage);
718         put_page(oldpage);
719
720         pte_unmap_unlock(ptep, ptl);
721         err = 0;
722 out:
723         return err;
724 }
725
726 /*
727  * try_to_merge_one_page - take two pages and merge them into one
728  * @vma: the vma that hold the pte pointing into oldpage
729  * @oldpage: the page that we want to replace with newpage
730  * @newpage: the page that we want to map instead of oldpage
731  *
732  * Note:
733  * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
734  * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
735  *
736  * This function returns 0 if the pages were merged, -EFAULT otherwise.
737  */
738 static int try_to_merge_one_page(struct vm_area_struct *vma,
739                                  struct page *oldpage,
740                                  struct page *newpage)
741 {
742         pte_t orig_pte = __pte(0);
743         int err = -EFAULT;
744
745         if (!(vma->vm_flags & VM_MERGEABLE))
746                 goto out;
747
748         if (!PageAnon(oldpage))
749                 goto out;
750
751         get_page(newpage);
752         get_page(oldpage);
753
754         /*
755          * We need the page lock to read a stable PageSwapCache in
756          * write_protect_page().  We use trylock_page() instead of
757          * lock_page() because we don't want to wait here - we
758          * prefer to continue scanning and merging different pages,
759          * then come back to this page when it is unlocked.
760          */
761         if (!trylock_page(oldpage))
762                 goto out_putpage;
763         /*
764          * If this anonymous page is mapped only here, its pte may need
765          * to be write-protected.  If it's mapped elsewhere, all of its
766          * ptes are necessarily already write-protected.  But in either
767          * case, we need to lock and check page_count is not raised.
768          */
769         if (write_protect_page(vma, oldpage, &orig_pte)) {
770                 unlock_page(oldpage);
771                 goto out_putpage;
772         }
773         unlock_page(oldpage);
774
775         if (pages_identical(oldpage, newpage))
776                 err = replace_page(vma, oldpage, newpage, orig_pte);
777
778 out_putpage:
779         put_page(oldpage);
780         put_page(newpage);
781 out:
782         return err;
783 }
784
785 /*
786  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
787  * but no new kernel page is allocated: kpage must already be a ksm page.
788  */
789 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
790                                       unsigned long addr1,
791                                       struct page *page1,
792                                       struct page *kpage)
793 {
794         struct vm_area_struct *vma;
795         int err = -EFAULT;
796
797         down_read(&mm1->mmap_sem);
798         if (ksm_test_exit(mm1))
799                 goto out;
800
801         vma = find_vma(mm1, addr1);
802         if (!vma || vma->vm_start > addr1)
803                 goto out;
804
805         err = try_to_merge_one_page(vma, page1, kpage);
806 out:
807         up_read(&mm1->mmap_sem);
808         return err;
809 }
810
811 /*
812  * try_to_merge_two_pages - take two identical pages and prepare them
813  * to be merged into one page.
814  *
815  * This function returns 0 if we successfully mapped two identical pages
816  * into one page, -EFAULT otherwise.
817  *
818  * Note that this function allocates a new kernel page: if one of the pages
819  * is already a ksm page, try_to_merge_with_ksm_page should be used.
820  */
821 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
822                                   struct page *page1, struct mm_struct *mm2,
823                                   unsigned long addr2, struct page *page2)
824 {
825         struct vm_area_struct *vma;
826         struct page *kpage;
827         int err = -EFAULT;
828
829         /*
830          * The number of nodes in the stable tree
831          * is the number of kernel pages that we hold.
832          */
833         if (ksm_max_kernel_pages &&
834             ksm_max_kernel_pages <= ksm_pages_shared)
835                 return err;
836
837         kpage = alloc_page(GFP_HIGHUSER);
838         if (!kpage)
839                 return err;
840
841         down_read(&mm1->mmap_sem);
842         if (ksm_test_exit(mm1)) {
843                 up_read(&mm1->mmap_sem);
844                 goto out;
845         }
846         vma = find_vma(mm1, addr1);
847         if (!vma || vma->vm_start > addr1) {
848                 up_read(&mm1->mmap_sem);
849                 goto out;
850         }
851
852         copy_user_highpage(kpage, page1, addr1, vma);
853         err = try_to_merge_one_page(vma, page1, kpage);
854         up_read(&mm1->mmap_sem);
855
856         if (!err) {
857                 err = try_to_merge_with_ksm_page(mm2, addr2, page2, kpage);
858                 /*
859                  * If that fails, we have a ksm page with only one pte
860                  * pointing to it: so break it.
861                  */
862                 if (err)
863                         break_cow(mm1, addr1);
864         }
865 out:
866         put_page(kpage);
867         return err;
868 }
869
870 /*
871  * stable_tree_search - search page inside the stable tree
872  * @page: the page that we are searching identical pages to.
873  * @page2: pointer into identical page that we are holding inside the stable
874  *         tree that we have found.
875  * @rmap_item: the reverse mapping item
876  *
877  * This function checks if there is a page inside the stable tree
878  * with identical content to the page that we are scanning right now.
879  *
880  * This function return rmap_item pointer to the identical item if found,
881  * NULL otherwise.
882  */
883 static struct rmap_item *stable_tree_search(struct page *page,
884                                             struct page **page2,
885                                             struct rmap_item *rmap_item)
886 {
887         struct rb_node *node = root_stable_tree.rb_node;
888
889         while (node) {
890                 struct rmap_item *tree_rmap_item, *next_rmap_item;
891                 int ret;
892
893                 tree_rmap_item = rb_entry(node, struct rmap_item, node);
894                 while (tree_rmap_item) {
895                         BUG_ON(!in_stable_tree(tree_rmap_item));
896                         cond_resched();
897                         page2[0] = get_ksm_page(tree_rmap_item);
898                         if (page2[0])
899                                 break;
900                         next_rmap_item = tree_rmap_item->next;
901                         remove_rmap_item_from_tree(tree_rmap_item);
902                         tree_rmap_item = next_rmap_item;
903                 }
904                 if (!tree_rmap_item)
905                         return NULL;
906
907                 ret = memcmp_pages(page, page2[0]);
908
909                 if (ret < 0) {
910                         put_page(page2[0]);
911                         node = node->rb_left;
912                 } else if (ret > 0) {
913                         put_page(page2[0]);
914                         node = node->rb_right;
915                 } else {
916                         return tree_rmap_item;
917                 }
918         }
919
920         return NULL;
921 }
922
923 /*
924  * stable_tree_insert - insert rmap_item pointing to new ksm page
925  * into the stable tree.
926  *
927  * @page: the page that we are searching identical page to inside the stable
928  *        tree.
929  * @rmap_item: pointer to the reverse mapping item.
930  *
931  * This function returns rmap_item if success, NULL otherwise.
932  */
933 static struct rmap_item *stable_tree_insert(struct page *page,
934                                             struct rmap_item *rmap_item)
935 {
936         struct rb_node **new = &root_stable_tree.rb_node;
937         struct rb_node *parent = NULL;
938
939         while (*new) {
940                 struct rmap_item *tree_rmap_item, *next_rmap_item;
941                 struct page *tree_page;
942                 int ret;
943
944                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
945                 while (tree_rmap_item) {
946                         BUG_ON(!in_stable_tree(tree_rmap_item));
947                         cond_resched();
948                         tree_page = get_ksm_page(tree_rmap_item);
949                         if (tree_page)
950                                 break;
951                         next_rmap_item = tree_rmap_item->next;
952                         remove_rmap_item_from_tree(tree_rmap_item);
953                         tree_rmap_item = next_rmap_item;
954                 }
955                 if (!tree_rmap_item)
956                         return NULL;
957
958                 ret = memcmp_pages(page, tree_page);
959                 put_page(tree_page);
960
961                 parent = *new;
962                 if (ret < 0)
963                         new = &parent->rb_left;
964                 else if (ret > 0)
965                         new = &parent->rb_right;
966                 else {
967                         /*
968                          * It is not a bug that stable_tree_search() didn't
969                          * find this node: because at that time our page was
970                          * not yet write-protected, so may have changed since.
971                          */
972                         return NULL;
973                 }
974         }
975
976         rmap_item->address |= NODE_FLAG | STABLE_FLAG;
977         rmap_item->next = NULL;
978         rb_link_node(&rmap_item->node, parent, new);
979         rb_insert_color(&rmap_item->node, &root_stable_tree);
980
981         ksm_pages_shared++;
982         return rmap_item;
983 }
984
985 /*
986  * unstable_tree_search_insert - search and insert items into the unstable tree.
987  *
988  * @page: the page that we are going to search for identical page or to insert
989  *        into the unstable tree
990  * @page2: pointer into identical page that was found inside the unstable tree
991  * @rmap_item: the reverse mapping item of page
992  *
993  * This function searches for a page in the unstable tree identical to the
994  * page currently being scanned; and if no identical page is found in the
995  * tree, we insert rmap_item as a new object into the unstable tree.
996  *
997  * This function returns pointer to rmap_item found to be identical
998  * to the currently scanned page, NULL otherwise.
999  *
1000  * This function does both searching and inserting, because they share
1001  * the same walking algorithm in an rbtree.
1002  */
1003 static struct rmap_item *unstable_tree_search_insert(struct page *page,
1004                                                 struct page **page2,
1005                                                 struct rmap_item *rmap_item)
1006 {
1007         struct rb_node **new = &root_unstable_tree.rb_node;
1008         struct rb_node *parent = NULL;
1009
1010         while (*new) {
1011                 struct rmap_item *tree_rmap_item;
1012                 int ret;
1013
1014                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1015                 page2[0] = get_mergeable_page(tree_rmap_item);
1016                 if (!page2[0])
1017                         return NULL;
1018
1019                 /*
1020                  * Don't substitute an unswappable ksm page
1021                  * just for one good swappable forked page.
1022                  */
1023                 if (page == page2[0]) {
1024                         put_page(page2[0]);
1025                         return NULL;
1026                 }
1027
1028                 ret = memcmp_pages(page, page2[0]);
1029
1030                 parent = *new;
1031                 if (ret < 0) {
1032                         put_page(page2[0]);
1033                         new = &parent->rb_left;
1034                 } else if (ret > 0) {
1035                         put_page(page2[0]);
1036                         new = &parent->rb_right;
1037                 } else {
1038                         return tree_rmap_item;
1039                 }
1040         }
1041
1042         rmap_item->address |= NODE_FLAG;
1043         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1044         rb_link_node(&rmap_item->node, parent, new);
1045         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1046
1047         ksm_pages_unshared++;
1048         return NULL;
1049 }
1050
1051 /*
1052  * stable_tree_append - add another rmap_item to the linked list of
1053  * rmap_items hanging off a given node of the stable tree, all sharing
1054  * the same ksm page.
1055  */
1056 static void stable_tree_append(struct rmap_item *rmap_item,
1057                                struct rmap_item *tree_rmap_item)
1058 {
1059         rmap_item->next = tree_rmap_item->next;
1060         rmap_item->prev = tree_rmap_item;
1061
1062         if (tree_rmap_item->next)
1063                 tree_rmap_item->next->prev = rmap_item;
1064
1065         tree_rmap_item->next = rmap_item;
1066         rmap_item->address |= STABLE_FLAG;
1067
1068         ksm_pages_sharing++;
1069 }
1070
1071 /*
1072  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1073  * if not, compare checksum to previous and if it's the same, see if page can
1074  * be inserted into the unstable tree, or merged with a page already there and
1075  * both transferred to the stable tree.
1076  *
1077  * @page: the page that we are searching identical page to.
1078  * @rmap_item: the reverse mapping into the virtual address of this page
1079  */
1080 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1081 {
1082         struct page *page2[1];
1083         struct rmap_item *tree_rmap_item;
1084         unsigned int checksum;
1085         int err;
1086
1087         if (in_stable_tree(rmap_item))
1088                 remove_rmap_item_from_tree(rmap_item);
1089
1090         /* We first start with searching the page inside the stable tree */
1091         tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1092         if (tree_rmap_item) {
1093                 if (page == page2[0])                   /* forked */
1094                         err = 0;
1095                 else
1096                         err = try_to_merge_with_ksm_page(rmap_item->mm,
1097                                                          rmap_item->address,
1098                                                          page, page2[0]);
1099                 put_page(page2[0]);
1100
1101                 if (!err) {
1102                         /*
1103                          * The page was successfully merged:
1104                          * add its rmap_item to the stable tree.
1105                          */
1106                         stable_tree_append(rmap_item, tree_rmap_item);
1107                 }
1108                 return;
1109         }
1110
1111         /*
1112          * A ksm page might have got here by fork, but its other
1113          * references have already been removed from the stable tree.
1114          * Or it might be left over from a break_ksm which failed
1115          * when the mem_cgroup had reached its limit: try again now.
1116          */
1117         if (PageKsm(page))
1118                 break_cow(rmap_item->mm, rmap_item->address);
1119
1120         /*
1121          * In case the hash value of the page was changed from the last time we
1122          * have calculated it, this page to be changed frequely, therefore we
1123          * don't want to insert it to the unstable tree, and we don't want to
1124          * waste our time to search if there is something identical to it there.
1125          */
1126         checksum = calc_checksum(page);
1127         if (rmap_item->oldchecksum != checksum) {
1128                 rmap_item->oldchecksum = checksum;
1129                 return;
1130         }
1131
1132         tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1133         if (tree_rmap_item) {
1134                 err = try_to_merge_two_pages(rmap_item->mm,
1135                                              rmap_item->address, page,
1136                                              tree_rmap_item->mm,
1137                                              tree_rmap_item->address, page2[0]);
1138                 /*
1139                  * As soon as we merge this page, we want to remove the
1140                  * rmap_item of the page we have merged with from the unstable
1141                  * tree, and insert it instead as new node in the stable tree.
1142                  */
1143                 if (!err) {
1144                         rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1145                         tree_rmap_item->address &= ~NODE_FLAG;
1146                         ksm_pages_unshared--;
1147
1148                         /*
1149                          * If we fail to insert the page into the stable tree,
1150                          * we will have 2 virtual addresses that are pointing
1151                          * to a ksm page left outside the stable tree,
1152                          * in which case we need to break_cow on both.
1153                          */
1154                         if (stable_tree_insert(page2[0], tree_rmap_item))
1155                                 stable_tree_append(rmap_item, tree_rmap_item);
1156                         else {
1157                                 break_cow(tree_rmap_item->mm,
1158                                                 tree_rmap_item->address);
1159                                 break_cow(rmap_item->mm, rmap_item->address);
1160                         }
1161                 }
1162
1163                 put_page(page2[0]);
1164         }
1165 }
1166
1167 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1168                                             struct list_head *cur,
1169                                             unsigned long addr)
1170 {
1171         struct rmap_item *rmap_item;
1172
1173         while (cur != &mm_slot->rmap_list) {
1174                 rmap_item = list_entry(cur, struct rmap_item, link);
1175                 if ((rmap_item->address & PAGE_MASK) == addr) {
1176                         if (!in_stable_tree(rmap_item))
1177                                 remove_rmap_item_from_tree(rmap_item);
1178                         return rmap_item;
1179                 }
1180                 if (rmap_item->address > addr)
1181                         break;
1182                 cur = cur->next;
1183                 remove_rmap_item_from_tree(rmap_item);
1184                 list_del(&rmap_item->link);
1185                 free_rmap_item(rmap_item);
1186         }
1187
1188         rmap_item = alloc_rmap_item();
1189         if (rmap_item) {
1190                 /* It has already been zeroed */
1191                 rmap_item->mm = mm_slot->mm;
1192                 rmap_item->address = addr;
1193                 list_add_tail(&rmap_item->link, cur);
1194         }
1195         return rmap_item;
1196 }
1197
1198 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1199 {
1200         struct mm_struct *mm;
1201         struct mm_slot *slot;
1202         struct vm_area_struct *vma;
1203         struct rmap_item *rmap_item;
1204
1205         if (list_empty(&ksm_mm_head.mm_list))
1206                 return NULL;
1207
1208         slot = ksm_scan.mm_slot;
1209         if (slot == &ksm_mm_head) {
1210                 root_unstable_tree = RB_ROOT;
1211
1212                 spin_lock(&ksm_mmlist_lock);
1213                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1214                 ksm_scan.mm_slot = slot;
1215                 spin_unlock(&ksm_mmlist_lock);
1216 next_mm:
1217                 ksm_scan.address = 0;
1218                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1219                                                 struct rmap_item, link);
1220         }
1221
1222         mm = slot->mm;
1223         down_read(&mm->mmap_sem);
1224         if (ksm_test_exit(mm))
1225                 vma = NULL;
1226         else
1227                 vma = find_vma(mm, ksm_scan.address);
1228
1229         for (; vma; vma = vma->vm_next) {
1230                 if (!(vma->vm_flags & VM_MERGEABLE))
1231                         continue;
1232                 if (ksm_scan.address < vma->vm_start)
1233                         ksm_scan.address = vma->vm_start;
1234                 if (!vma->anon_vma)
1235                         ksm_scan.address = vma->vm_end;
1236
1237                 while (ksm_scan.address < vma->vm_end) {
1238                         if (ksm_test_exit(mm))
1239                                 break;
1240                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1241                         if (*page && PageAnon(*page)) {
1242                                 flush_anon_page(vma, *page, ksm_scan.address);
1243                                 flush_dcache_page(*page);
1244                                 rmap_item = get_next_rmap_item(slot,
1245                                         ksm_scan.rmap_item->link.next,
1246                                         ksm_scan.address);
1247                                 if (rmap_item) {
1248                                         ksm_scan.rmap_item = rmap_item;
1249                                         ksm_scan.address += PAGE_SIZE;
1250                                 } else
1251                                         put_page(*page);
1252                                 up_read(&mm->mmap_sem);
1253                                 return rmap_item;
1254                         }
1255                         if (*page)
1256                                 put_page(*page);
1257                         ksm_scan.address += PAGE_SIZE;
1258                         cond_resched();
1259                 }
1260         }
1261
1262         if (ksm_test_exit(mm)) {
1263                 ksm_scan.address = 0;
1264                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1265                                                 struct rmap_item, link);
1266         }
1267         /*
1268          * Nuke all the rmap_items that are above this current rmap:
1269          * because there were no VM_MERGEABLE vmas with such addresses.
1270          */
1271         remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1272
1273         spin_lock(&ksm_mmlist_lock);
1274         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1275                                                 struct mm_slot, mm_list);
1276         if (ksm_scan.address == 0) {
1277                 /*
1278                  * We've completed a full scan of all vmas, holding mmap_sem
1279                  * throughout, and found no VM_MERGEABLE: so do the same as
1280                  * __ksm_exit does to remove this mm from all our lists now.
1281                  * This applies either when cleaning up after __ksm_exit
1282                  * (but beware: we can reach here even before __ksm_exit),
1283                  * or when all VM_MERGEABLE areas have been unmapped (and
1284                  * mmap_sem then protects against race with MADV_MERGEABLE).
1285                  */
1286                 hlist_del(&slot->link);
1287                 list_del(&slot->mm_list);
1288                 spin_unlock(&ksm_mmlist_lock);
1289
1290                 free_mm_slot(slot);
1291                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1292                 up_read(&mm->mmap_sem);
1293                 mmdrop(mm);
1294         } else {
1295                 spin_unlock(&ksm_mmlist_lock);
1296                 up_read(&mm->mmap_sem);
1297         }
1298
1299         /* Repeat until we've completed scanning the whole list */
1300         slot = ksm_scan.mm_slot;
1301         if (slot != &ksm_mm_head)
1302                 goto next_mm;
1303
1304         ksm_scan.seqnr++;
1305         return NULL;
1306 }
1307
1308 /**
1309  * ksm_do_scan  - the ksm scanner main worker function.
1310  * @scan_npages - number of pages we want to scan before we return.
1311  */
1312 static void ksm_do_scan(unsigned int scan_npages)
1313 {
1314         struct rmap_item *rmap_item;
1315         struct page *page;
1316
1317         while (scan_npages--) {
1318                 cond_resched();
1319                 rmap_item = scan_get_next_rmap_item(&page);
1320                 if (!rmap_item)
1321                         return;
1322                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1323                         cmp_and_merge_page(page, rmap_item);
1324                 else if (page_mapcount(page) == 1) {
1325                         /*
1326                          * Replace now-unshared ksm page by ordinary page.
1327                          */
1328                         break_cow(rmap_item->mm, rmap_item->address);
1329                         remove_rmap_item_from_tree(rmap_item);
1330                         rmap_item->oldchecksum = calc_checksum(page);
1331                 }
1332                 put_page(page);
1333         }
1334 }
1335
1336 static int ksmd_should_run(void)
1337 {
1338         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1339 }
1340
1341 static int ksm_scan_thread(void *nothing)
1342 {
1343         set_user_nice(current, 5);
1344
1345         while (!kthread_should_stop()) {
1346                 mutex_lock(&ksm_thread_mutex);
1347                 if (ksmd_should_run())
1348                         ksm_do_scan(ksm_thread_pages_to_scan);
1349                 mutex_unlock(&ksm_thread_mutex);
1350
1351                 if (ksmd_should_run()) {
1352                         schedule_timeout_interruptible(
1353                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1354                 } else {
1355                         wait_event_interruptible(ksm_thread_wait,
1356                                 ksmd_should_run() || kthread_should_stop());
1357                 }
1358         }
1359         return 0;
1360 }
1361
1362 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1363                 unsigned long end, int advice, unsigned long *vm_flags)
1364 {
1365         struct mm_struct *mm = vma->vm_mm;
1366         int err;
1367
1368         switch (advice) {
1369         case MADV_MERGEABLE:
1370                 /*
1371                  * Be somewhat over-protective for now!
1372                  */
1373                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1374                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1375                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1376                                  VM_MIXEDMAP  | VM_SAO))
1377                         return 0;               /* just ignore the advice */
1378
1379                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1380                         err = __ksm_enter(mm);
1381                         if (err)
1382                                 return err;
1383                 }
1384
1385                 *vm_flags |= VM_MERGEABLE;
1386                 break;
1387
1388         case MADV_UNMERGEABLE:
1389                 if (!(*vm_flags & VM_MERGEABLE))
1390                         return 0;               /* just ignore the advice */
1391
1392                 if (vma->anon_vma) {
1393                         err = unmerge_ksm_pages(vma, start, end);
1394                         if (err)
1395                                 return err;
1396                 }
1397
1398                 *vm_flags &= ~VM_MERGEABLE;
1399                 break;
1400         }
1401
1402         return 0;
1403 }
1404
1405 int __ksm_enter(struct mm_struct *mm)
1406 {
1407         struct mm_slot *mm_slot;
1408         int needs_wakeup;
1409
1410         mm_slot = alloc_mm_slot();
1411         if (!mm_slot)
1412                 return -ENOMEM;
1413
1414         /* Check ksm_run too?  Would need tighter locking */
1415         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1416
1417         spin_lock(&ksm_mmlist_lock);
1418         insert_to_mm_slots_hash(mm, mm_slot);
1419         /*
1420          * Insert just behind the scanning cursor, to let the area settle
1421          * down a little; when fork is followed by immediate exec, we don't
1422          * want ksmd to waste time setting up and tearing down an rmap_list.
1423          */
1424         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1425         spin_unlock(&ksm_mmlist_lock);
1426
1427         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1428         atomic_inc(&mm->mm_count);
1429
1430         if (needs_wakeup)
1431                 wake_up_interruptible(&ksm_thread_wait);
1432
1433         return 0;
1434 }
1435
1436 void __ksm_exit(struct mm_struct *mm)
1437 {
1438         struct mm_slot *mm_slot;
1439         int easy_to_free = 0;
1440
1441         /*
1442          * This process is exiting: if it's straightforward (as is the
1443          * case when ksmd was never running), free mm_slot immediately.
1444          * But if it's at the cursor or has rmap_items linked to it, use
1445          * mmap_sem to synchronize with any break_cows before pagetables
1446          * are freed, and leave the mm_slot on the list for ksmd to free.
1447          * Beware: ksm may already have noticed it exiting and freed the slot.
1448          */
1449
1450         spin_lock(&ksm_mmlist_lock);
1451         mm_slot = get_mm_slot(mm);
1452         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1453                 if (list_empty(&mm_slot->rmap_list)) {
1454                         hlist_del(&mm_slot->link);
1455                         list_del(&mm_slot->mm_list);
1456                         easy_to_free = 1;
1457                 } else {
1458                         list_move(&mm_slot->mm_list,
1459                                   &ksm_scan.mm_slot->mm_list);
1460                 }
1461         }
1462         spin_unlock(&ksm_mmlist_lock);
1463
1464         if (easy_to_free) {
1465                 free_mm_slot(mm_slot);
1466                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1467                 mmdrop(mm);
1468         } else if (mm_slot) {
1469                 down_write(&mm->mmap_sem);
1470                 up_write(&mm->mmap_sem);
1471         }
1472 }
1473
1474 #ifdef CONFIG_SYSFS
1475 /*
1476  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1477  */
1478
1479 #define KSM_ATTR_RO(_name) \
1480         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1481 #define KSM_ATTR(_name) \
1482         static struct kobj_attribute _name##_attr = \
1483                 __ATTR(_name, 0644, _name##_show, _name##_store)
1484
1485 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1486                                     struct kobj_attribute *attr, char *buf)
1487 {
1488         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1489 }
1490
1491 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1492                                      struct kobj_attribute *attr,
1493                                      const char *buf, size_t count)
1494 {
1495         unsigned long msecs;
1496         int err;
1497
1498         err = strict_strtoul(buf, 10, &msecs);
1499         if (err || msecs > UINT_MAX)
1500                 return -EINVAL;
1501
1502         ksm_thread_sleep_millisecs = msecs;
1503
1504         return count;
1505 }
1506 KSM_ATTR(sleep_millisecs);
1507
1508 static ssize_t pages_to_scan_show(struct kobject *kobj,
1509                                   struct kobj_attribute *attr, char *buf)
1510 {
1511         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1512 }
1513
1514 static ssize_t pages_to_scan_store(struct kobject *kobj,
1515                                    struct kobj_attribute *attr,
1516                                    const char *buf, size_t count)
1517 {
1518         int err;
1519         unsigned long nr_pages;
1520
1521         err = strict_strtoul(buf, 10, &nr_pages);
1522         if (err || nr_pages > UINT_MAX)
1523                 return -EINVAL;
1524
1525         ksm_thread_pages_to_scan = nr_pages;
1526
1527         return count;
1528 }
1529 KSM_ATTR(pages_to_scan);
1530
1531 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1532                         char *buf)
1533 {
1534         return sprintf(buf, "%u\n", ksm_run);
1535 }
1536
1537 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1538                          const char *buf, size_t count)
1539 {
1540         int err;
1541         unsigned long flags;
1542
1543         err = strict_strtoul(buf, 10, &flags);
1544         if (err || flags > UINT_MAX)
1545                 return -EINVAL;
1546         if (flags > KSM_RUN_UNMERGE)
1547                 return -EINVAL;
1548
1549         /*
1550          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1551          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1552          * breaking COW to free the unswappable pages_shared (but leaves
1553          * mm_slots on the list for when ksmd may be set running again).
1554          */
1555
1556         mutex_lock(&ksm_thread_mutex);
1557         if (ksm_run != flags) {
1558                 ksm_run = flags;
1559                 if (flags & KSM_RUN_UNMERGE) {
1560                         current->flags |= PF_OOM_ORIGIN;
1561                         err = unmerge_and_remove_all_rmap_items();
1562                         current->flags &= ~PF_OOM_ORIGIN;
1563                         if (err) {
1564                                 ksm_run = KSM_RUN_STOP;
1565                                 count = err;
1566                         }
1567                 }
1568         }
1569         mutex_unlock(&ksm_thread_mutex);
1570
1571         if (flags & KSM_RUN_MERGE)
1572                 wake_up_interruptible(&ksm_thread_wait);
1573
1574         return count;
1575 }
1576 KSM_ATTR(run);
1577
1578 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1579                                       struct kobj_attribute *attr,
1580                                       const char *buf, size_t count)
1581 {
1582         int err;
1583         unsigned long nr_pages;
1584
1585         err = strict_strtoul(buf, 10, &nr_pages);
1586         if (err)
1587                 return -EINVAL;
1588
1589         ksm_max_kernel_pages = nr_pages;
1590
1591         return count;
1592 }
1593
1594 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1595                                      struct kobj_attribute *attr, char *buf)
1596 {
1597         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1598 }
1599 KSM_ATTR(max_kernel_pages);
1600
1601 static ssize_t pages_shared_show(struct kobject *kobj,
1602                                  struct kobj_attribute *attr, char *buf)
1603 {
1604         return sprintf(buf, "%lu\n", ksm_pages_shared);
1605 }
1606 KSM_ATTR_RO(pages_shared);
1607
1608 static ssize_t pages_sharing_show(struct kobject *kobj,
1609                                   struct kobj_attribute *attr, char *buf)
1610 {
1611         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1612 }
1613 KSM_ATTR_RO(pages_sharing);
1614
1615 static ssize_t pages_unshared_show(struct kobject *kobj,
1616                                    struct kobj_attribute *attr, char *buf)
1617 {
1618         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1619 }
1620 KSM_ATTR_RO(pages_unshared);
1621
1622 static ssize_t pages_volatile_show(struct kobject *kobj,
1623                                    struct kobj_attribute *attr, char *buf)
1624 {
1625         long ksm_pages_volatile;
1626
1627         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1628                                 - ksm_pages_sharing - ksm_pages_unshared;
1629         /*
1630          * It was not worth any locking to calculate that statistic,
1631          * but it might therefore sometimes be negative: conceal that.
1632          */
1633         if (ksm_pages_volatile < 0)
1634                 ksm_pages_volatile = 0;
1635         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1636 }
1637 KSM_ATTR_RO(pages_volatile);
1638
1639 static ssize_t full_scans_show(struct kobject *kobj,
1640                                struct kobj_attribute *attr, char *buf)
1641 {
1642         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1643 }
1644 KSM_ATTR_RO(full_scans);
1645
1646 static struct attribute *ksm_attrs[] = {
1647         &sleep_millisecs_attr.attr,
1648         &pages_to_scan_attr.attr,
1649         &run_attr.attr,
1650         &max_kernel_pages_attr.attr,
1651         &pages_shared_attr.attr,
1652         &pages_sharing_attr.attr,
1653         &pages_unshared_attr.attr,
1654         &pages_volatile_attr.attr,
1655         &full_scans_attr.attr,
1656         NULL,
1657 };
1658
1659 static struct attribute_group ksm_attr_group = {
1660         .attrs = ksm_attrs,
1661         .name = "ksm",
1662 };
1663 #endif /* CONFIG_SYSFS */
1664
1665 static int __init ksm_init(void)
1666 {
1667         struct task_struct *ksm_thread;
1668         int err;
1669
1670         err = ksm_slab_init();
1671         if (err)
1672                 goto out;
1673
1674         err = mm_slots_hash_init();
1675         if (err)
1676                 goto out_free1;
1677
1678         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1679         if (IS_ERR(ksm_thread)) {
1680                 printk(KERN_ERR "ksm: creating kthread failed\n");
1681                 err = PTR_ERR(ksm_thread);
1682                 goto out_free2;
1683         }
1684
1685 #ifdef CONFIG_SYSFS
1686         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1687         if (err) {
1688                 printk(KERN_ERR "ksm: register sysfs failed\n");
1689                 kthread_stop(ksm_thread);
1690                 goto out_free2;
1691         }
1692 #endif /* CONFIG_SYSFS */
1693
1694         return 0;
1695
1696 out_free2:
1697         mm_slots_hash_free();
1698 out_free1:
1699         ksm_slab_free();
1700 out:
1701         return err;
1702 }
1703 module_init(ksm_init)