ksm: rename kernel_pages_allocated
[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 sharing those nodes */
156 static unsigned long ksm_pages_sharing;
157
158 /* Limit on the number of unswappable pages used */
159 static unsigned long ksm_max_kernel_pages;
160
161 /* Number of pages ksmd should scan in one batch */
162 static unsigned int ksm_thread_pages_to_scan;
163
164 /* Milliseconds ksmd should sleep between batches */
165 static unsigned int ksm_thread_sleep_millisecs;
166
167 #define KSM_RUN_STOP    0
168 #define KSM_RUN_MERGE   1
169 #define KSM_RUN_UNMERGE 2
170 static unsigned int ksm_run;
171
172 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
173 static DEFINE_MUTEX(ksm_thread_mutex);
174 static DEFINE_SPINLOCK(ksm_mmlist_lock);
175
176 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
177                 sizeof(struct __struct), __alignof__(struct __struct),\
178                 (__flags), NULL)
179
180 static int __init ksm_slab_init(void)
181 {
182         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
183         if (!rmap_item_cache)
184                 goto out;
185
186         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
187         if (!mm_slot_cache)
188                 goto out_free;
189
190         return 0;
191
192 out_free:
193         kmem_cache_destroy(rmap_item_cache);
194 out:
195         return -ENOMEM;
196 }
197
198 static void __init ksm_slab_free(void)
199 {
200         kmem_cache_destroy(mm_slot_cache);
201         kmem_cache_destroy(rmap_item_cache);
202         mm_slot_cache = NULL;
203 }
204
205 static inline struct rmap_item *alloc_rmap_item(void)
206 {
207         return kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
208 }
209
210 static inline void free_rmap_item(struct rmap_item *rmap_item)
211 {
212         rmap_item->mm = NULL;   /* debug safety */
213         kmem_cache_free(rmap_item_cache, rmap_item);
214 }
215
216 static inline struct mm_slot *alloc_mm_slot(void)
217 {
218         if (!mm_slot_cache)     /* initialization failed */
219                 return NULL;
220         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
221 }
222
223 static inline void free_mm_slot(struct mm_slot *mm_slot)
224 {
225         kmem_cache_free(mm_slot_cache, mm_slot);
226 }
227
228 static int __init mm_slots_hash_init(void)
229 {
230         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
231                                 GFP_KERNEL);
232         if (!mm_slots_hash)
233                 return -ENOMEM;
234         return 0;
235 }
236
237 static void __init mm_slots_hash_free(void)
238 {
239         kfree(mm_slots_hash);
240 }
241
242 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
243 {
244         struct mm_slot *mm_slot;
245         struct hlist_head *bucket;
246         struct hlist_node *node;
247
248         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
249                                 % MM_SLOTS_HASH_HEADS];
250         hlist_for_each_entry(mm_slot, node, bucket, link) {
251                 if (mm == mm_slot->mm)
252                         return mm_slot;
253         }
254         return NULL;
255 }
256
257 static void insert_to_mm_slots_hash(struct mm_struct *mm,
258                                     struct mm_slot *mm_slot)
259 {
260         struct hlist_head *bucket;
261
262         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
263                                 % MM_SLOTS_HASH_HEADS];
264         mm_slot->mm = mm;
265         INIT_LIST_HEAD(&mm_slot->rmap_list);
266         hlist_add_head(&mm_slot->link, bucket);
267 }
268
269 static inline int in_stable_tree(struct rmap_item *rmap_item)
270 {
271         return rmap_item->address & STABLE_FLAG;
272 }
273
274 /*
275  * We use break_ksm to break COW on a ksm page: it's a stripped down
276  *
277  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
278  *              put_page(page);
279  *
280  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
281  * in case the application has unmapped and remapped mm,addr meanwhile.
282  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
283  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
284  */
285 static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
286 {
287         struct page *page;
288         int ret;
289
290         do {
291                 cond_resched();
292                 page = follow_page(vma, addr, FOLL_GET);
293                 if (!page)
294                         break;
295                 if (PageKsm(page))
296                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
297                                                         FAULT_FLAG_WRITE);
298                 else
299                         ret = VM_FAULT_WRITE;
300                 put_page(page);
301         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS)));
302
303         /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
304 }
305
306 static void __break_cow(struct mm_struct *mm, unsigned long addr)
307 {
308         struct vm_area_struct *vma;
309
310         vma = find_vma(mm, addr);
311         if (!vma || vma->vm_start > addr)
312                 return;
313         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
314                 return;
315         break_ksm(vma, addr);
316 }
317
318 static void break_cow(struct mm_struct *mm, unsigned long addr)
319 {
320         down_read(&mm->mmap_sem);
321         __break_cow(mm, addr);
322         up_read(&mm->mmap_sem);
323 }
324
325 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
326 {
327         struct mm_struct *mm = rmap_item->mm;
328         unsigned long addr = rmap_item->address;
329         struct vm_area_struct *vma;
330         struct page *page;
331
332         down_read(&mm->mmap_sem);
333         vma = find_vma(mm, addr);
334         if (!vma || vma->vm_start > addr)
335                 goto out;
336         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
337                 goto out;
338
339         page = follow_page(vma, addr, FOLL_GET);
340         if (!page)
341                 goto out;
342         if (PageAnon(page)) {
343                 flush_anon_page(vma, page, addr);
344                 flush_dcache_page(page);
345         } else {
346                 put_page(page);
347 out:            page = NULL;
348         }
349         up_read(&mm->mmap_sem);
350         return page;
351 }
352
353 /*
354  * get_ksm_page: checks if the page at the virtual address in rmap_item
355  * is still PageKsm, in which case we can trust the content of the page,
356  * and it returns the gotten page; but NULL if the page has been zapped.
357  */
358 static struct page *get_ksm_page(struct rmap_item *rmap_item)
359 {
360         struct page *page;
361
362         page = get_mergeable_page(rmap_item);
363         if (page && !PageKsm(page)) {
364                 put_page(page);
365                 page = NULL;
366         }
367         return page;
368 }
369
370 /*
371  * Removing rmap_item from stable or unstable tree.
372  * This function will clean the information from the stable/unstable tree.
373  */
374 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
375 {
376         if (in_stable_tree(rmap_item)) {
377                 struct rmap_item *next_item = rmap_item->next;
378
379                 if (rmap_item->address & NODE_FLAG) {
380                         if (next_item) {
381                                 rb_replace_node(&rmap_item->node,
382                                                 &next_item->node,
383                                                 &root_stable_tree);
384                                 next_item->address |= NODE_FLAG;
385                         } else {
386                                 rb_erase(&rmap_item->node, &root_stable_tree);
387                                 ksm_pages_shared--;
388                         }
389                 } else {
390                         struct rmap_item *prev_item = rmap_item->prev;
391
392                         BUG_ON(prev_item->next != rmap_item);
393                         prev_item->next = next_item;
394                         if (next_item) {
395                                 BUG_ON(next_item->prev != rmap_item);
396                                 next_item->prev = rmap_item->prev;
397                         }
398                 }
399
400                 rmap_item->next = NULL;
401                 ksm_pages_sharing--;
402
403         } else if (rmap_item->address & NODE_FLAG) {
404                 unsigned char age;
405                 /*
406                  * ksm_thread can and must skip the rb_erase, because
407                  * root_unstable_tree was already reset to RB_ROOT.
408                  * But __ksm_exit has to be careful: do the rb_erase
409                  * if it's interrupting a scan, and this rmap_item was
410                  * inserted by this scan rather than left from before.
411                  *
412                  * Because of the case in which remove_mm_from_lists
413                  * increments seqnr before removing rmaps, unstable_nr
414                  * may even be 2 behind seqnr, but should never be
415                  * further behind.  Yes, I did have trouble with this!
416                  */
417                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
418                 BUG_ON(age > 2);
419                 if (!age)
420                         rb_erase(&rmap_item->node, &root_unstable_tree);
421         }
422
423         rmap_item->address &= PAGE_MASK;
424
425         cond_resched();         /* we're called from many long loops */
426 }
427
428 static void remove_all_slot_rmap_items(struct mm_slot *mm_slot)
429 {
430         struct rmap_item *rmap_item, *node;
431
432         list_for_each_entry_safe(rmap_item, node, &mm_slot->rmap_list, link) {
433                 remove_rmap_item_from_tree(rmap_item);
434                 list_del(&rmap_item->link);
435                 free_rmap_item(rmap_item);
436         }
437 }
438
439 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
440                                        struct list_head *cur)
441 {
442         struct rmap_item *rmap_item;
443
444         while (cur != &mm_slot->rmap_list) {
445                 rmap_item = list_entry(cur, struct rmap_item, link);
446                 cur = cur->next;
447                 remove_rmap_item_from_tree(rmap_item);
448                 list_del(&rmap_item->link);
449                 free_rmap_item(rmap_item);
450         }
451 }
452
453 /*
454  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
455  * than check every pte of a given vma, the locking doesn't quite work for
456  * that - an rmap_item is assigned to the stable tree after inserting ksm
457  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
458  * rmap_items from parent to child at fork time (so as not to waste time
459  * if exit comes before the next scan reaches it).
460  */
461 static void unmerge_ksm_pages(struct vm_area_struct *vma,
462                               unsigned long start, unsigned long end)
463 {
464         unsigned long addr;
465
466         for (addr = start; addr < end; addr += PAGE_SIZE)
467                 break_ksm(vma, addr);
468 }
469
470 static void unmerge_and_remove_all_rmap_items(void)
471 {
472         struct mm_slot *mm_slot;
473         struct mm_struct *mm;
474         struct vm_area_struct *vma;
475
476         list_for_each_entry(mm_slot, &ksm_mm_head.mm_list, mm_list) {
477                 mm = mm_slot->mm;
478                 down_read(&mm->mmap_sem);
479                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
480                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
481                                 continue;
482                         unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end);
483                 }
484                 remove_all_slot_rmap_items(mm_slot);
485                 up_read(&mm->mmap_sem);
486         }
487
488         spin_lock(&ksm_mmlist_lock);
489         if (ksm_scan.mm_slot != &ksm_mm_head) {
490                 ksm_scan.mm_slot = &ksm_mm_head;
491                 ksm_scan.seqnr++;
492         }
493         spin_unlock(&ksm_mmlist_lock);
494 }
495
496 static void remove_mm_from_lists(struct mm_struct *mm)
497 {
498         struct mm_slot *mm_slot;
499
500         spin_lock(&ksm_mmlist_lock);
501         mm_slot = get_mm_slot(mm);
502
503         /*
504          * This mm_slot is always at the scanning cursor when we're
505          * called from scan_get_next_rmap_item; but it's a special
506          * case when we're called from __ksm_exit.
507          */
508         if (ksm_scan.mm_slot == mm_slot) {
509                 ksm_scan.mm_slot = list_entry(
510                         mm_slot->mm_list.next, struct mm_slot, mm_list);
511                 ksm_scan.address = 0;
512                 ksm_scan.rmap_item = list_entry(
513                         &ksm_scan.mm_slot->rmap_list, struct rmap_item, link);
514                 if (ksm_scan.mm_slot == &ksm_mm_head)
515                         ksm_scan.seqnr++;
516         }
517
518         hlist_del(&mm_slot->link);
519         list_del(&mm_slot->mm_list);
520         spin_unlock(&ksm_mmlist_lock);
521
522         remove_all_slot_rmap_items(mm_slot);
523         free_mm_slot(mm_slot);
524         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
525 }
526
527 static u32 calc_checksum(struct page *page)
528 {
529         u32 checksum;
530         void *addr = kmap_atomic(page, KM_USER0);
531         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
532         kunmap_atomic(addr, KM_USER0);
533         return checksum;
534 }
535
536 static int memcmp_pages(struct page *page1, struct page *page2)
537 {
538         char *addr1, *addr2;
539         int ret;
540
541         addr1 = kmap_atomic(page1, KM_USER0);
542         addr2 = kmap_atomic(page2, KM_USER1);
543         ret = memcmp(addr1, addr2, PAGE_SIZE);
544         kunmap_atomic(addr2, KM_USER1);
545         kunmap_atomic(addr1, KM_USER0);
546         return ret;
547 }
548
549 static inline int pages_identical(struct page *page1, struct page *page2)
550 {
551         return !memcmp_pages(page1, page2);
552 }
553
554 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
555                               pte_t *orig_pte)
556 {
557         struct mm_struct *mm = vma->vm_mm;
558         unsigned long addr;
559         pte_t *ptep;
560         spinlock_t *ptl;
561         int swapped;
562         int err = -EFAULT;
563
564         addr = page_address_in_vma(page, vma);
565         if (addr == -EFAULT)
566                 goto out;
567
568         ptep = page_check_address(page, mm, addr, &ptl, 0);
569         if (!ptep)
570                 goto out;
571
572         if (pte_write(*ptep)) {
573                 pte_t entry;
574
575                 swapped = PageSwapCache(page);
576                 flush_cache_page(vma, addr, page_to_pfn(page));
577                 /*
578                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
579                  * take any lock, therefore the check that we are going to make
580                  * with the pagecount against the mapcount is racey and
581                  * O_DIRECT can happen right after the check.
582                  * So we clear the pte and flush the tlb before the check
583                  * this assure us that no O_DIRECT can happen after the check
584                  * or in the middle of the check.
585                  */
586                 entry = ptep_clear_flush(vma, addr, ptep);
587                 /*
588                  * Check that no O_DIRECT or similar I/O is in progress on the
589                  * page
590                  */
591                 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
592                         set_pte_at_notify(mm, addr, ptep, entry);
593                         goto out_unlock;
594                 }
595                 entry = pte_wrprotect(entry);
596                 set_pte_at_notify(mm, addr, ptep, entry);
597         }
598         *orig_pte = *ptep;
599         err = 0;
600
601 out_unlock:
602         pte_unmap_unlock(ptep, ptl);
603 out:
604         return err;
605 }
606
607 /**
608  * replace_page - replace page in vma by new ksm page
609  * @vma:      vma that holds the pte pointing to oldpage
610  * @oldpage:  the page we are replacing by newpage
611  * @newpage:  the ksm page we replace oldpage by
612  * @orig_pte: the original value of the pte
613  *
614  * Returns 0 on success, -EFAULT on failure.
615  */
616 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
617                         struct page *newpage, pte_t orig_pte)
618 {
619         struct mm_struct *mm = vma->vm_mm;
620         pgd_t *pgd;
621         pud_t *pud;
622         pmd_t *pmd;
623         pte_t *ptep;
624         spinlock_t *ptl;
625         unsigned long addr;
626         pgprot_t prot;
627         int err = -EFAULT;
628
629         prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
630
631         addr = page_address_in_vma(oldpage, vma);
632         if (addr == -EFAULT)
633                 goto out;
634
635         pgd = pgd_offset(mm, addr);
636         if (!pgd_present(*pgd))
637                 goto out;
638
639         pud = pud_offset(pgd, addr);
640         if (!pud_present(*pud))
641                 goto out;
642
643         pmd = pmd_offset(pud, addr);
644         if (!pmd_present(*pmd))
645                 goto out;
646
647         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
648         if (!pte_same(*ptep, orig_pte)) {
649                 pte_unmap_unlock(ptep, ptl);
650                 goto out;
651         }
652
653         get_page(newpage);
654         page_add_ksm_rmap(newpage);
655
656         flush_cache_page(vma, addr, pte_pfn(*ptep));
657         ptep_clear_flush(vma, addr, ptep);
658         set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
659
660         page_remove_rmap(oldpage);
661         put_page(oldpage);
662
663         pte_unmap_unlock(ptep, ptl);
664         err = 0;
665 out:
666         return err;
667 }
668
669 /*
670  * try_to_merge_one_page - take two pages and merge them into one
671  * @vma: the vma that hold the pte pointing into oldpage
672  * @oldpage: the page that we want to replace with newpage
673  * @newpage: the page that we want to map instead of oldpage
674  *
675  * Note:
676  * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
677  * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
678  *
679  * This function returns 0 if the pages were merged, -EFAULT otherwise.
680  */
681 static int try_to_merge_one_page(struct vm_area_struct *vma,
682                                  struct page *oldpage,
683                                  struct page *newpage)
684 {
685         pte_t orig_pte = __pte(0);
686         int err = -EFAULT;
687
688         if (!(vma->vm_flags & VM_MERGEABLE))
689                 goto out;
690
691         if (!PageAnon(oldpage))
692                 goto out;
693
694         get_page(newpage);
695         get_page(oldpage);
696
697         /*
698          * We need the page lock to read a stable PageSwapCache in
699          * write_protect_page().  We use trylock_page() instead of
700          * lock_page() because we don't want to wait here - we
701          * prefer to continue scanning and merging different pages,
702          * then come back to this page when it is unlocked.
703          */
704         if (!trylock_page(oldpage))
705                 goto out_putpage;
706         /*
707          * If this anonymous page is mapped only here, its pte may need
708          * to be write-protected.  If it's mapped elsewhere, all of its
709          * ptes are necessarily already write-protected.  But in either
710          * case, we need to lock and check page_count is not raised.
711          */
712         if (write_protect_page(vma, oldpage, &orig_pte)) {
713                 unlock_page(oldpage);
714                 goto out_putpage;
715         }
716         unlock_page(oldpage);
717
718         if (pages_identical(oldpage, newpage))
719                 err = replace_page(vma, oldpage, newpage, orig_pte);
720
721 out_putpage:
722         put_page(oldpage);
723         put_page(newpage);
724 out:
725         return err;
726 }
727
728 /*
729  * try_to_merge_two_pages - take two identical pages and prepare them
730  * to be merged into one page.
731  *
732  * This function returns 0 if we successfully mapped two identical pages
733  * into one page, -EFAULT otherwise.
734  *
735  * Note that this function allocates a new kernel page: if one of the pages
736  * is already a ksm page, try_to_merge_with_ksm_page should be used.
737  */
738 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
739                                   struct page *page1, struct mm_struct *mm2,
740                                   unsigned long addr2, struct page *page2)
741 {
742         struct vm_area_struct *vma;
743         struct page *kpage;
744         int err = -EFAULT;
745
746         /*
747          * The number of nodes in the stable tree
748          * is the number of kernel pages that we hold.
749          */
750         if (ksm_max_kernel_pages &&
751             ksm_max_kernel_pages <= ksm_pages_shared)
752                 return err;
753
754         kpage = alloc_page(GFP_HIGHUSER);
755         if (!kpage)
756                 return err;
757
758         down_read(&mm1->mmap_sem);
759         vma = find_vma(mm1, addr1);
760         if (!vma || vma->vm_start > addr1) {
761                 put_page(kpage);
762                 up_read(&mm1->mmap_sem);
763                 return err;
764         }
765
766         copy_user_highpage(kpage, page1, addr1, vma);
767         err = try_to_merge_one_page(vma, page1, kpage);
768         up_read(&mm1->mmap_sem);
769
770         if (!err) {
771                 down_read(&mm2->mmap_sem);
772                 vma = find_vma(mm2, addr2);
773                 if (!vma || vma->vm_start > addr2) {
774                         put_page(kpage);
775                         up_read(&mm2->mmap_sem);
776                         break_cow(mm1, addr1);
777                         return -EFAULT;
778                 }
779
780                 err = try_to_merge_one_page(vma, page2, kpage);
781                 up_read(&mm2->mmap_sem);
782
783                 /*
784                  * If the second try_to_merge_one_page failed, we have a
785                  * ksm page with just one pte pointing to it, so break it.
786                  */
787                 if (err)
788                         break_cow(mm1, addr1);
789                 else
790                         ksm_pages_sharing += 2;
791         }
792
793         put_page(kpage);
794         return err;
795 }
796
797 /*
798  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
799  * but no new kernel page is allocated: kpage must already be a ksm page.
800  */
801 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
802                                       unsigned long addr1,
803                                       struct page *page1,
804                                       struct page *kpage)
805 {
806         struct vm_area_struct *vma;
807         int err = -EFAULT;
808
809         down_read(&mm1->mmap_sem);
810         vma = find_vma(mm1, addr1);
811         if (!vma || vma->vm_start > addr1) {
812                 up_read(&mm1->mmap_sem);
813                 return err;
814         }
815
816         err = try_to_merge_one_page(vma, page1, kpage);
817         up_read(&mm1->mmap_sem);
818
819         if (!err)
820                 ksm_pages_sharing++;
821
822         return err;
823 }
824
825 /*
826  * stable_tree_search - search page inside the stable tree
827  * @page: the page that we are searching identical pages to.
828  * @page2: pointer into identical page that we are holding inside the stable
829  *         tree that we have found.
830  * @rmap_item: the reverse mapping item
831  *
832  * This function checks if there is a page inside the stable tree
833  * with identical content to the page that we are scanning right now.
834  *
835  * This function return rmap_item pointer to the identical item if found,
836  * NULL otherwise.
837  */
838 static struct rmap_item *stable_tree_search(struct page *page,
839                                             struct page **page2,
840                                             struct rmap_item *rmap_item)
841 {
842         struct rb_node *node = root_stable_tree.rb_node;
843
844         while (node) {
845                 struct rmap_item *tree_rmap_item, *next_rmap_item;
846                 int ret;
847
848                 tree_rmap_item = rb_entry(node, struct rmap_item, node);
849                 while (tree_rmap_item) {
850                         BUG_ON(!in_stable_tree(tree_rmap_item));
851                         cond_resched();
852                         page2[0] = get_ksm_page(tree_rmap_item);
853                         if (page2[0])
854                                 break;
855                         next_rmap_item = tree_rmap_item->next;
856                         remove_rmap_item_from_tree(tree_rmap_item);
857                         tree_rmap_item = next_rmap_item;
858                 }
859                 if (!tree_rmap_item)
860                         return NULL;
861
862                 ret = memcmp_pages(page, page2[0]);
863
864                 if (ret < 0) {
865                         put_page(page2[0]);
866                         node = node->rb_left;
867                 } else if (ret > 0) {
868                         put_page(page2[0]);
869                         node = node->rb_right;
870                 } else {
871                         return tree_rmap_item;
872                 }
873         }
874
875         return NULL;
876 }
877
878 /*
879  * stable_tree_insert - insert rmap_item pointing to new ksm page
880  * into the stable tree.
881  *
882  * @page: the page that we are searching identical page to inside the stable
883  *        tree.
884  * @rmap_item: pointer to the reverse mapping item.
885  *
886  * This function returns rmap_item if success, NULL otherwise.
887  */
888 static struct rmap_item *stable_tree_insert(struct page *page,
889                                             struct rmap_item *rmap_item)
890 {
891         struct rb_node **new = &root_stable_tree.rb_node;
892         struct rb_node *parent = NULL;
893
894         while (*new) {
895                 struct rmap_item *tree_rmap_item, *next_rmap_item;
896                 struct page *tree_page;
897                 int ret;
898
899                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
900                 while (tree_rmap_item) {
901                         BUG_ON(!in_stable_tree(tree_rmap_item));
902                         cond_resched();
903                         tree_page = get_ksm_page(tree_rmap_item);
904                         if (tree_page)
905                                 break;
906                         next_rmap_item = tree_rmap_item->next;
907                         remove_rmap_item_from_tree(tree_rmap_item);
908                         tree_rmap_item = next_rmap_item;
909                 }
910                 if (!tree_rmap_item)
911                         return NULL;
912
913                 ret = memcmp_pages(page, tree_page);
914                 put_page(tree_page);
915
916                 parent = *new;
917                 if (ret < 0)
918                         new = &parent->rb_left;
919                 else if (ret > 0)
920                         new = &parent->rb_right;
921                 else {
922                         /*
923                          * It is not a bug that stable_tree_search() didn't
924                          * find this node: because at that time our page was
925                          * not yet write-protected, so may have changed since.
926                          */
927                         return NULL;
928                 }
929         }
930
931         ksm_pages_shared++;
932
933         rmap_item->address |= NODE_FLAG | STABLE_FLAG;
934         rmap_item->next = NULL;
935         rb_link_node(&rmap_item->node, parent, new);
936         rb_insert_color(&rmap_item->node, &root_stable_tree);
937
938         return rmap_item;
939 }
940
941 /*
942  * unstable_tree_search_insert - search and insert items into the unstable tree.
943  *
944  * @page: the page that we are going to search for identical page or to insert
945  *        into the unstable tree
946  * @page2: pointer into identical page that was found inside the unstable tree
947  * @rmap_item: the reverse mapping item of page
948  *
949  * This function searches for a page in the unstable tree identical to the
950  * page currently being scanned; and if no identical page is found in the
951  * tree, we insert rmap_item as a new object into the unstable tree.
952  *
953  * This function returns pointer to rmap_item found to be identical
954  * to the currently scanned page, NULL otherwise.
955  *
956  * This function does both searching and inserting, because they share
957  * the same walking algorithm in an rbtree.
958  */
959 static struct rmap_item *unstable_tree_search_insert(struct page *page,
960                                                 struct page **page2,
961                                                 struct rmap_item *rmap_item)
962 {
963         struct rb_node **new = &root_unstable_tree.rb_node;
964         struct rb_node *parent = NULL;
965
966         while (*new) {
967                 struct rmap_item *tree_rmap_item;
968                 int ret;
969
970                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
971                 page2[0] = get_mergeable_page(tree_rmap_item);
972                 if (!page2[0])
973                         return NULL;
974
975                 /*
976                  * Don't substitute an unswappable ksm page
977                  * just for one good swappable forked page.
978                  */
979                 if (page == page2[0]) {
980                         put_page(page2[0]);
981                         return NULL;
982                 }
983
984                 ret = memcmp_pages(page, page2[0]);
985
986                 parent = *new;
987                 if (ret < 0) {
988                         put_page(page2[0]);
989                         new = &parent->rb_left;
990                 } else if (ret > 0) {
991                         put_page(page2[0]);
992                         new = &parent->rb_right;
993                 } else {
994                         return tree_rmap_item;
995                 }
996         }
997
998         rmap_item->address |= NODE_FLAG;
999         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1000         rb_link_node(&rmap_item->node, parent, new);
1001         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1002
1003         return NULL;
1004 }
1005
1006 /*
1007  * stable_tree_append - add another rmap_item to the linked list of
1008  * rmap_items hanging off a given node of the stable tree, all sharing
1009  * the same ksm page.
1010  */
1011 static void stable_tree_append(struct rmap_item *rmap_item,
1012                                struct rmap_item *tree_rmap_item)
1013 {
1014         rmap_item->next = tree_rmap_item->next;
1015         rmap_item->prev = tree_rmap_item;
1016
1017         if (tree_rmap_item->next)
1018                 tree_rmap_item->next->prev = rmap_item;
1019
1020         tree_rmap_item->next = rmap_item;
1021         rmap_item->address |= STABLE_FLAG;
1022 }
1023
1024 /*
1025  * cmp_and_merge_page - take a page computes its hash value and check if there
1026  * is similar hash value to different page,
1027  * in case we find that there is similar hash to different page we call to
1028  * try_to_merge_two_pages().
1029  *
1030  * @page: the page that we are searching identical page to.
1031  * @rmap_item: the reverse mapping into the virtual address of this page
1032  */
1033 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1034 {
1035         struct page *page2[1];
1036         struct rmap_item *tree_rmap_item;
1037         unsigned int checksum;
1038         int err;
1039
1040         if (in_stable_tree(rmap_item))
1041                 remove_rmap_item_from_tree(rmap_item);
1042
1043         /* We first start with searching the page inside the stable tree */
1044         tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1045         if (tree_rmap_item) {
1046                 if (page == page2[0]) {                 /* forked */
1047                         ksm_pages_sharing++;
1048                         err = 0;
1049                 } else
1050                         err = try_to_merge_with_ksm_page(rmap_item->mm,
1051                                                          rmap_item->address,
1052                                                          page, page2[0]);
1053                 put_page(page2[0]);
1054
1055                 if (!err) {
1056                         /*
1057                          * The page was successfully merged:
1058                          * add its rmap_item to the stable tree.
1059                          */
1060                         stable_tree_append(rmap_item, tree_rmap_item);
1061                 }
1062                 return;
1063         }
1064
1065         /*
1066          * A ksm page might have got here by fork, but its other
1067          * references have already been removed from the stable tree.
1068          */
1069         if (PageKsm(page))
1070                 break_cow(rmap_item->mm, rmap_item->address);
1071
1072         /*
1073          * In case the hash value of the page was changed from the last time we
1074          * have calculated it, this page to be changed frequely, therefore we
1075          * don't want to insert it to the unstable tree, and we don't want to
1076          * waste our time to search if there is something identical to it there.
1077          */
1078         checksum = calc_checksum(page);
1079         if (rmap_item->oldchecksum != checksum) {
1080                 rmap_item->oldchecksum = checksum;
1081                 return;
1082         }
1083
1084         tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1085         if (tree_rmap_item) {
1086                 err = try_to_merge_two_pages(rmap_item->mm,
1087                                              rmap_item->address, page,
1088                                              tree_rmap_item->mm,
1089                                              tree_rmap_item->address, page2[0]);
1090                 /*
1091                  * As soon as we merge this page, we want to remove the
1092                  * rmap_item of the page we have merged with from the unstable
1093                  * tree, and insert it instead as new node in the stable tree.
1094                  */
1095                 if (!err) {
1096                         rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1097                         tree_rmap_item->address &= ~NODE_FLAG;
1098                         /*
1099                          * If we fail to insert the page into the stable tree,
1100                          * we will have 2 virtual addresses that are pointing
1101                          * to a ksm page left outside the stable tree,
1102                          * in which case we need to break_cow on both.
1103                          */
1104                         if (stable_tree_insert(page2[0], tree_rmap_item))
1105                                 stable_tree_append(rmap_item, tree_rmap_item);
1106                         else {
1107                                 break_cow(tree_rmap_item->mm,
1108                                                 tree_rmap_item->address);
1109                                 break_cow(rmap_item->mm, rmap_item->address);
1110                                 ksm_pages_sharing -= 2;
1111                         }
1112                 }
1113
1114                 put_page(page2[0]);
1115         }
1116 }
1117
1118 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1119                                             struct list_head *cur,
1120                                             unsigned long addr)
1121 {
1122         struct rmap_item *rmap_item;
1123
1124         while (cur != &mm_slot->rmap_list) {
1125                 rmap_item = list_entry(cur, struct rmap_item, link);
1126                 if ((rmap_item->address & PAGE_MASK) == addr) {
1127                         if (!in_stable_tree(rmap_item))
1128                                 remove_rmap_item_from_tree(rmap_item);
1129                         return rmap_item;
1130                 }
1131                 if (rmap_item->address > addr)
1132                         break;
1133                 cur = cur->next;
1134                 remove_rmap_item_from_tree(rmap_item);
1135                 list_del(&rmap_item->link);
1136                 free_rmap_item(rmap_item);
1137         }
1138
1139         rmap_item = alloc_rmap_item();
1140         if (rmap_item) {
1141                 /* It has already been zeroed */
1142                 rmap_item->mm = mm_slot->mm;
1143                 rmap_item->address = addr;
1144                 list_add_tail(&rmap_item->link, cur);
1145         }
1146         return rmap_item;
1147 }
1148
1149 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1150 {
1151         struct mm_struct *mm;
1152         struct mm_slot *slot;
1153         struct vm_area_struct *vma;
1154         struct rmap_item *rmap_item;
1155
1156         if (list_empty(&ksm_mm_head.mm_list))
1157                 return NULL;
1158
1159         slot = ksm_scan.mm_slot;
1160         if (slot == &ksm_mm_head) {
1161                 root_unstable_tree = RB_ROOT;
1162
1163                 spin_lock(&ksm_mmlist_lock);
1164                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1165                 ksm_scan.mm_slot = slot;
1166                 spin_unlock(&ksm_mmlist_lock);
1167 next_mm:
1168                 ksm_scan.address = 0;
1169                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1170                                                 struct rmap_item, link);
1171         }
1172
1173         mm = slot->mm;
1174         down_read(&mm->mmap_sem);
1175         for (vma = find_vma(mm, ksm_scan.address); vma; vma = vma->vm_next) {
1176                 if (!(vma->vm_flags & VM_MERGEABLE))
1177                         continue;
1178                 if (ksm_scan.address < vma->vm_start)
1179                         ksm_scan.address = vma->vm_start;
1180                 if (!vma->anon_vma)
1181                         ksm_scan.address = vma->vm_end;
1182
1183                 while (ksm_scan.address < vma->vm_end) {
1184                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1185                         if (*page && PageAnon(*page)) {
1186                                 flush_anon_page(vma, *page, ksm_scan.address);
1187                                 flush_dcache_page(*page);
1188                                 rmap_item = get_next_rmap_item(slot,
1189                                         ksm_scan.rmap_item->link.next,
1190                                         ksm_scan.address);
1191                                 if (rmap_item) {
1192                                         ksm_scan.rmap_item = rmap_item;
1193                                         ksm_scan.address += PAGE_SIZE;
1194                                 } else
1195                                         put_page(*page);
1196                                 up_read(&mm->mmap_sem);
1197                                 return rmap_item;
1198                         }
1199                         if (*page)
1200                                 put_page(*page);
1201                         ksm_scan.address += PAGE_SIZE;
1202                         cond_resched();
1203                 }
1204         }
1205
1206         if (!ksm_scan.address) {
1207                 /*
1208                  * We've completed a full scan of all vmas, holding mmap_sem
1209                  * throughout, and found no VM_MERGEABLE: so do the same as
1210                  * __ksm_exit does to remove this mm from all our lists now.
1211                  */
1212                 remove_mm_from_lists(mm);
1213                 up_read(&mm->mmap_sem);
1214                 slot = ksm_scan.mm_slot;
1215                 if (slot != &ksm_mm_head)
1216                         goto next_mm;
1217                 return NULL;
1218         }
1219
1220         /*
1221          * Nuke all the rmap_items that are above this current rmap:
1222          * because there were no VM_MERGEABLE vmas with such addresses.
1223          */
1224         remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1225         up_read(&mm->mmap_sem);
1226
1227         spin_lock(&ksm_mmlist_lock);
1228         slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1229         ksm_scan.mm_slot = slot;
1230         spin_unlock(&ksm_mmlist_lock);
1231
1232         /* Repeat until we've completed scanning the whole list */
1233         if (slot != &ksm_mm_head)
1234                 goto next_mm;
1235
1236         /*
1237          * Bump seqnr here rather than at top, so that __ksm_exit
1238          * can skip rb_erase on unstable tree until we run again.
1239          */
1240         ksm_scan.seqnr++;
1241         return NULL;
1242 }
1243
1244 /**
1245  * ksm_do_scan  - the ksm scanner main worker function.
1246  * @scan_npages - number of pages we want to scan before we return.
1247  */
1248 static void ksm_do_scan(unsigned int scan_npages)
1249 {
1250         struct rmap_item *rmap_item;
1251         struct page *page;
1252
1253         while (scan_npages--) {
1254                 cond_resched();
1255                 rmap_item = scan_get_next_rmap_item(&page);
1256                 if (!rmap_item)
1257                         return;
1258                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1259                         cmp_and_merge_page(page, rmap_item);
1260                 put_page(page);
1261         }
1262 }
1263
1264 static int ksm_scan_thread(void *nothing)
1265 {
1266         set_user_nice(current, 5);
1267
1268         while (!kthread_should_stop()) {
1269                 if (ksm_run & KSM_RUN_MERGE) {
1270                         mutex_lock(&ksm_thread_mutex);
1271                         ksm_do_scan(ksm_thread_pages_to_scan);
1272                         mutex_unlock(&ksm_thread_mutex);
1273                         schedule_timeout_interruptible(
1274                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1275                 } else {
1276                         wait_event_interruptible(ksm_thread_wait,
1277                                         (ksm_run & KSM_RUN_MERGE) ||
1278                                         kthread_should_stop());
1279                 }
1280         }
1281         return 0;
1282 }
1283
1284 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1285                 unsigned long end, int advice, unsigned long *vm_flags)
1286 {
1287         struct mm_struct *mm = vma->vm_mm;
1288
1289         switch (advice) {
1290         case MADV_MERGEABLE:
1291                 /*
1292                  * Be somewhat over-protective for now!
1293                  */
1294                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1295                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1296                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1297                                  VM_MIXEDMAP  | VM_SAO))
1298                         return 0;               /* just ignore the advice */
1299
1300                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
1301                         if (__ksm_enter(mm) < 0)
1302                                 return -EAGAIN;
1303
1304                 *vm_flags |= VM_MERGEABLE;
1305                 break;
1306
1307         case MADV_UNMERGEABLE:
1308                 if (!(*vm_flags & VM_MERGEABLE))
1309                         return 0;               /* just ignore the advice */
1310
1311                 if (vma->anon_vma)
1312                         unmerge_ksm_pages(vma, start, end);
1313
1314                 *vm_flags &= ~VM_MERGEABLE;
1315                 break;
1316         }
1317
1318         return 0;
1319 }
1320
1321 int __ksm_enter(struct mm_struct *mm)
1322 {
1323         struct mm_slot *mm_slot = alloc_mm_slot();
1324         if (!mm_slot)
1325                 return -ENOMEM;
1326
1327         spin_lock(&ksm_mmlist_lock);
1328         insert_to_mm_slots_hash(mm, mm_slot);
1329         /*
1330          * Insert just behind the scanning cursor, to let the area settle
1331          * down a little; when fork is followed by immediate exec, we don't
1332          * want ksmd to waste time setting up and tearing down an rmap_list.
1333          */
1334         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1335         spin_unlock(&ksm_mmlist_lock);
1336
1337         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1338         return 0;
1339 }
1340
1341 void __ksm_exit(struct mm_struct *mm)
1342 {
1343         /*
1344          * This process is exiting: doesn't hold and doesn't need mmap_sem;
1345          * but we do need to exclude ksmd and other exiters while we modify
1346          * the various lists and trees.
1347          */
1348         mutex_lock(&ksm_thread_mutex);
1349         remove_mm_from_lists(mm);
1350         mutex_unlock(&ksm_thread_mutex);
1351 }
1352
1353 #define KSM_ATTR_RO(_name) \
1354         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1355 #define KSM_ATTR(_name) \
1356         static struct kobj_attribute _name##_attr = \
1357                 __ATTR(_name, 0644, _name##_show, _name##_store)
1358
1359 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1360                                     struct kobj_attribute *attr, char *buf)
1361 {
1362         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1363 }
1364
1365 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1366                                      struct kobj_attribute *attr,
1367                                      const char *buf, size_t count)
1368 {
1369         unsigned long msecs;
1370         int err;
1371
1372         err = strict_strtoul(buf, 10, &msecs);
1373         if (err || msecs > UINT_MAX)
1374                 return -EINVAL;
1375
1376         ksm_thread_sleep_millisecs = msecs;
1377
1378         return count;
1379 }
1380 KSM_ATTR(sleep_millisecs);
1381
1382 static ssize_t pages_to_scan_show(struct kobject *kobj,
1383                                   struct kobj_attribute *attr, char *buf)
1384 {
1385         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1386 }
1387
1388 static ssize_t pages_to_scan_store(struct kobject *kobj,
1389                                    struct kobj_attribute *attr,
1390                                    const char *buf, size_t count)
1391 {
1392         int err;
1393         unsigned long nr_pages;
1394
1395         err = strict_strtoul(buf, 10, &nr_pages);
1396         if (err || nr_pages > UINT_MAX)
1397                 return -EINVAL;
1398
1399         ksm_thread_pages_to_scan = nr_pages;
1400
1401         return count;
1402 }
1403 KSM_ATTR(pages_to_scan);
1404
1405 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1406                         char *buf)
1407 {
1408         return sprintf(buf, "%u\n", ksm_run);
1409 }
1410
1411 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1412                          const char *buf, size_t count)
1413 {
1414         int err;
1415         unsigned long flags;
1416
1417         err = strict_strtoul(buf, 10, &flags);
1418         if (err || flags > UINT_MAX)
1419                 return -EINVAL;
1420         if (flags > KSM_RUN_UNMERGE)
1421                 return -EINVAL;
1422
1423         /*
1424          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1425          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1426          * breaking COW to free the unswappable pages_shared (but leaves
1427          * mm_slots on the list for when ksmd may be set running again).
1428          */
1429
1430         mutex_lock(&ksm_thread_mutex);
1431         if (ksm_run != flags) {
1432                 ksm_run = flags;
1433                 if (flags & KSM_RUN_UNMERGE)
1434                         unmerge_and_remove_all_rmap_items();
1435         }
1436         mutex_unlock(&ksm_thread_mutex);
1437
1438         if (flags & KSM_RUN_MERGE)
1439                 wake_up_interruptible(&ksm_thread_wait);
1440
1441         return count;
1442 }
1443 KSM_ATTR(run);
1444
1445 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1446                                       struct kobj_attribute *attr,
1447                                       const char *buf, size_t count)
1448 {
1449         int err;
1450         unsigned long nr_pages;
1451
1452         err = strict_strtoul(buf, 10, &nr_pages);
1453         if (err)
1454                 return -EINVAL;
1455
1456         ksm_max_kernel_pages = nr_pages;
1457
1458         return count;
1459 }
1460
1461 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1462                                      struct kobj_attribute *attr, char *buf)
1463 {
1464         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1465 }
1466 KSM_ATTR(max_kernel_pages);
1467
1468 static ssize_t pages_shared_show(struct kobject *kobj,
1469                                  struct kobj_attribute *attr, char *buf)
1470 {
1471         return sprintf(buf, "%lu\n", ksm_pages_shared);
1472 }
1473 KSM_ATTR_RO(pages_shared);
1474
1475 static ssize_t pages_sharing_show(struct kobject *kobj,
1476                                   struct kobj_attribute *attr, char *buf)
1477 {
1478         return sprintf(buf, "%lu\n",
1479                         ksm_pages_sharing - ksm_pages_shared);
1480 }
1481 KSM_ATTR_RO(pages_sharing);
1482
1483 static struct attribute *ksm_attrs[] = {
1484         &sleep_millisecs_attr.attr,
1485         &pages_to_scan_attr.attr,
1486         &run_attr.attr,
1487         &max_kernel_pages_attr.attr,
1488         &pages_shared_attr.attr,
1489         &pages_sharing_attr.attr,
1490         NULL,
1491 };
1492
1493 static struct attribute_group ksm_attr_group = {
1494         .attrs = ksm_attrs,
1495         .name = "ksm",
1496 };
1497
1498 static int __init ksm_init(void)
1499 {
1500         struct task_struct *ksm_thread;
1501         int err;
1502
1503         err = ksm_slab_init();
1504         if (err)
1505                 goto out;
1506
1507         err = mm_slots_hash_init();
1508         if (err)
1509                 goto out_free1;
1510
1511         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1512         if (IS_ERR(ksm_thread)) {
1513                 printk(KERN_ERR "ksm: creating kthread failed\n");
1514                 err = PTR_ERR(ksm_thread);
1515                 goto out_free2;
1516         }
1517
1518         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1519         if (err) {
1520                 printk(KERN_ERR "ksm: register sysfs failed\n");
1521                 goto out_free3;
1522         }
1523
1524         return 0;
1525
1526 out_free3:
1527         kthread_stop(ksm_thread);
1528 out_free2:
1529         mm_slots_hash_free();
1530 out_free1:
1531         ksm_slab_free();
1532 out:
1533         return err;
1534 }
1535 module_init(ksm_init)