ksm: pages_unshared and pages_volatile
[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;
166
167 /* Number of pages ksmd should scan in one batch */
168 static unsigned int ksm_thread_pages_to_scan;
169
170 /* Milliseconds ksmd should sleep between batches */
171 static unsigned int ksm_thread_sleep_millisecs;
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;
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  * We use break_ksm to break COW on a ksm page: it's a stripped down
288  *
289  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
290  *              put_page(page);
291  *
292  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
293  * in case the application has unmapped and remapped mm,addr meanwhile.
294  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
295  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
296  */
297 static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
298 {
299         struct page *page;
300         int ret;
301
302         do {
303                 cond_resched();
304                 page = follow_page(vma, addr, FOLL_GET);
305                 if (!page)
306                         break;
307                 if (PageKsm(page))
308                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
309                                                         FAULT_FLAG_WRITE);
310                 else
311                         ret = VM_FAULT_WRITE;
312                 put_page(page);
313         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS)));
314
315         /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
316 }
317
318 static void __break_cow(struct mm_struct *mm, unsigned long addr)
319 {
320         struct vm_area_struct *vma;
321
322         vma = find_vma(mm, addr);
323         if (!vma || vma->vm_start > addr)
324                 return;
325         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
326                 return;
327         break_ksm(vma, addr);
328 }
329
330 static void break_cow(struct mm_struct *mm, unsigned long addr)
331 {
332         down_read(&mm->mmap_sem);
333         __break_cow(mm, addr);
334         up_read(&mm->mmap_sem);
335 }
336
337 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
338 {
339         struct mm_struct *mm = rmap_item->mm;
340         unsigned long addr = rmap_item->address;
341         struct vm_area_struct *vma;
342         struct page *page;
343
344         down_read(&mm->mmap_sem);
345         vma = find_vma(mm, addr);
346         if (!vma || vma->vm_start > addr)
347                 goto out;
348         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
349                 goto out;
350
351         page = follow_page(vma, addr, FOLL_GET);
352         if (!page)
353                 goto out;
354         if (PageAnon(page)) {
355                 flush_anon_page(vma, page, addr);
356                 flush_dcache_page(page);
357         } else {
358                 put_page(page);
359 out:            page = NULL;
360         }
361         up_read(&mm->mmap_sem);
362         return page;
363 }
364
365 /*
366  * get_ksm_page: checks if the page at the virtual address in rmap_item
367  * is still PageKsm, in which case we can trust the content of the page,
368  * and it returns the gotten page; but NULL if the page has been zapped.
369  */
370 static struct page *get_ksm_page(struct rmap_item *rmap_item)
371 {
372         struct page *page;
373
374         page = get_mergeable_page(rmap_item);
375         if (page && !PageKsm(page)) {
376                 put_page(page);
377                 page = NULL;
378         }
379         return page;
380 }
381
382 /*
383  * Removing rmap_item from stable or unstable tree.
384  * This function will clean the information from the stable/unstable tree.
385  */
386 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
387 {
388         if (in_stable_tree(rmap_item)) {
389                 struct rmap_item *next_item = rmap_item->next;
390
391                 if (rmap_item->address & NODE_FLAG) {
392                         if (next_item) {
393                                 rb_replace_node(&rmap_item->node,
394                                                 &next_item->node,
395                                                 &root_stable_tree);
396                                 next_item->address |= NODE_FLAG;
397                                 ksm_pages_sharing--;
398                         } else {
399                                 rb_erase(&rmap_item->node, &root_stable_tree);
400                                 ksm_pages_shared--;
401                         }
402                 } else {
403                         struct rmap_item *prev_item = rmap_item->prev;
404
405                         BUG_ON(prev_item->next != rmap_item);
406                         prev_item->next = next_item;
407                         if (next_item) {
408                                 BUG_ON(next_item->prev != rmap_item);
409                                 next_item->prev = rmap_item->prev;
410                         }
411                         ksm_pages_sharing--;
412                 }
413
414                 rmap_item->next = NULL;
415
416         } else if (rmap_item->address & NODE_FLAG) {
417                 unsigned char age;
418                 /*
419                  * ksm_thread can and must skip the rb_erase, because
420                  * root_unstable_tree was already reset to RB_ROOT.
421                  * But __ksm_exit has to be careful: do the rb_erase
422                  * if it's interrupting a scan, and this rmap_item was
423                  * inserted by this scan rather than left from before.
424                  *
425                  * Because of the case in which remove_mm_from_lists
426                  * increments seqnr before removing rmaps, unstable_nr
427                  * may even be 2 behind seqnr, but should never be
428                  * further behind.  Yes, I did have trouble with this!
429                  */
430                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
431                 BUG_ON(age > 2);
432                 if (!age)
433                         rb_erase(&rmap_item->node, &root_unstable_tree);
434                 ksm_pages_unshared--;
435         }
436
437         rmap_item->address &= PAGE_MASK;
438
439         cond_resched();         /* we're called from many long loops */
440 }
441
442 static void remove_all_slot_rmap_items(struct mm_slot *mm_slot)
443 {
444         struct rmap_item *rmap_item, *node;
445
446         list_for_each_entry_safe(rmap_item, node, &mm_slot->rmap_list, link) {
447                 remove_rmap_item_from_tree(rmap_item);
448                 list_del(&rmap_item->link);
449                 free_rmap_item(rmap_item);
450         }
451 }
452
453 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
454                                        struct list_head *cur)
455 {
456         struct rmap_item *rmap_item;
457
458         while (cur != &mm_slot->rmap_list) {
459                 rmap_item = list_entry(cur, struct rmap_item, link);
460                 cur = cur->next;
461                 remove_rmap_item_from_tree(rmap_item);
462                 list_del(&rmap_item->link);
463                 free_rmap_item(rmap_item);
464         }
465 }
466
467 /*
468  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
469  * than check every pte of a given vma, the locking doesn't quite work for
470  * that - an rmap_item is assigned to the stable tree after inserting ksm
471  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
472  * rmap_items from parent to child at fork time (so as not to waste time
473  * if exit comes before the next scan reaches it).
474  */
475 static void unmerge_ksm_pages(struct vm_area_struct *vma,
476                               unsigned long start, unsigned long end)
477 {
478         unsigned long addr;
479
480         for (addr = start; addr < end; addr += PAGE_SIZE)
481                 break_ksm(vma, addr);
482 }
483
484 static void unmerge_and_remove_all_rmap_items(void)
485 {
486         struct mm_slot *mm_slot;
487         struct mm_struct *mm;
488         struct vm_area_struct *vma;
489
490         list_for_each_entry(mm_slot, &ksm_mm_head.mm_list, mm_list) {
491                 mm = mm_slot->mm;
492                 down_read(&mm->mmap_sem);
493                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
494                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
495                                 continue;
496                         unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end);
497                 }
498                 remove_all_slot_rmap_items(mm_slot);
499                 up_read(&mm->mmap_sem);
500         }
501
502         spin_lock(&ksm_mmlist_lock);
503         if (ksm_scan.mm_slot != &ksm_mm_head) {
504                 ksm_scan.mm_slot = &ksm_mm_head;
505                 ksm_scan.seqnr++;
506         }
507         spin_unlock(&ksm_mmlist_lock);
508 }
509
510 static void remove_mm_from_lists(struct mm_struct *mm)
511 {
512         struct mm_slot *mm_slot;
513
514         spin_lock(&ksm_mmlist_lock);
515         mm_slot = get_mm_slot(mm);
516
517         /*
518          * This mm_slot is always at the scanning cursor when we're
519          * called from scan_get_next_rmap_item; but it's a special
520          * case when we're called from __ksm_exit.
521          */
522         if (ksm_scan.mm_slot == mm_slot) {
523                 ksm_scan.mm_slot = list_entry(
524                         mm_slot->mm_list.next, struct mm_slot, mm_list);
525                 ksm_scan.address = 0;
526                 ksm_scan.rmap_item = list_entry(
527                         &ksm_scan.mm_slot->rmap_list, struct rmap_item, link);
528                 if (ksm_scan.mm_slot == &ksm_mm_head)
529                         ksm_scan.seqnr++;
530         }
531
532         hlist_del(&mm_slot->link);
533         list_del(&mm_slot->mm_list);
534         spin_unlock(&ksm_mmlist_lock);
535
536         remove_all_slot_rmap_items(mm_slot);
537         free_mm_slot(mm_slot);
538         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
539 }
540
541 static u32 calc_checksum(struct page *page)
542 {
543         u32 checksum;
544         void *addr = kmap_atomic(page, KM_USER0);
545         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
546         kunmap_atomic(addr, KM_USER0);
547         return checksum;
548 }
549
550 static int memcmp_pages(struct page *page1, struct page *page2)
551 {
552         char *addr1, *addr2;
553         int ret;
554
555         addr1 = kmap_atomic(page1, KM_USER0);
556         addr2 = kmap_atomic(page2, KM_USER1);
557         ret = memcmp(addr1, addr2, PAGE_SIZE);
558         kunmap_atomic(addr2, KM_USER1);
559         kunmap_atomic(addr1, KM_USER0);
560         return ret;
561 }
562
563 static inline int pages_identical(struct page *page1, struct page *page2)
564 {
565         return !memcmp_pages(page1, page2);
566 }
567
568 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
569                               pte_t *orig_pte)
570 {
571         struct mm_struct *mm = vma->vm_mm;
572         unsigned long addr;
573         pte_t *ptep;
574         spinlock_t *ptl;
575         int swapped;
576         int err = -EFAULT;
577
578         addr = page_address_in_vma(page, vma);
579         if (addr == -EFAULT)
580                 goto out;
581
582         ptep = page_check_address(page, mm, addr, &ptl, 0);
583         if (!ptep)
584                 goto out;
585
586         if (pte_write(*ptep)) {
587                 pte_t entry;
588
589                 swapped = PageSwapCache(page);
590                 flush_cache_page(vma, addr, page_to_pfn(page));
591                 /*
592                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
593                  * take any lock, therefore the check that we are going to make
594                  * with the pagecount against the mapcount is racey and
595                  * O_DIRECT can happen right after the check.
596                  * So we clear the pte and flush the tlb before the check
597                  * this assure us that no O_DIRECT can happen after the check
598                  * or in the middle of the check.
599                  */
600                 entry = ptep_clear_flush(vma, addr, ptep);
601                 /*
602                  * Check that no O_DIRECT or similar I/O is in progress on the
603                  * page
604                  */
605                 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
606                         set_pte_at_notify(mm, addr, ptep, entry);
607                         goto out_unlock;
608                 }
609                 entry = pte_wrprotect(entry);
610                 set_pte_at_notify(mm, addr, ptep, entry);
611         }
612         *orig_pte = *ptep;
613         err = 0;
614
615 out_unlock:
616         pte_unmap_unlock(ptep, ptl);
617 out:
618         return err;
619 }
620
621 /**
622  * replace_page - replace page in vma by new ksm page
623  * @vma:      vma that holds the pte pointing to oldpage
624  * @oldpage:  the page we are replacing by newpage
625  * @newpage:  the ksm page we replace oldpage by
626  * @orig_pte: the original value of the pte
627  *
628  * Returns 0 on success, -EFAULT on failure.
629  */
630 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
631                         struct page *newpage, pte_t orig_pte)
632 {
633         struct mm_struct *mm = vma->vm_mm;
634         pgd_t *pgd;
635         pud_t *pud;
636         pmd_t *pmd;
637         pte_t *ptep;
638         spinlock_t *ptl;
639         unsigned long addr;
640         pgprot_t prot;
641         int err = -EFAULT;
642
643         prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
644
645         addr = page_address_in_vma(oldpage, vma);
646         if (addr == -EFAULT)
647                 goto out;
648
649         pgd = pgd_offset(mm, addr);
650         if (!pgd_present(*pgd))
651                 goto out;
652
653         pud = pud_offset(pgd, addr);
654         if (!pud_present(*pud))
655                 goto out;
656
657         pmd = pmd_offset(pud, addr);
658         if (!pmd_present(*pmd))
659                 goto out;
660
661         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
662         if (!pte_same(*ptep, orig_pte)) {
663                 pte_unmap_unlock(ptep, ptl);
664                 goto out;
665         }
666
667         get_page(newpage);
668         page_add_ksm_rmap(newpage);
669
670         flush_cache_page(vma, addr, pte_pfn(*ptep));
671         ptep_clear_flush(vma, addr, ptep);
672         set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
673
674         page_remove_rmap(oldpage);
675         put_page(oldpage);
676
677         pte_unmap_unlock(ptep, ptl);
678         err = 0;
679 out:
680         return err;
681 }
682
683 /*
684  * try_to_merge_one_page - take two pages and merge them into one
685  * @vma: the vma that hold the pte pointing into oldpage
686  * @oldpage: the page that we want to replace with newpage
687  * @newpage: the page that we want to map instead of oldpage
688  *
689  * Note:
690  * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
691  * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
692  *
693  * This function returns 0 if the pages were merged, -EFAULT otherwise.
694  */
695 static int try_to_merge_one_page(struct vm_area_struct *vma,
696                                  struct page *oldpage,
697                                  struct page *newpage)
698 {
699         pte_t orig_pte = __pte(0);
700         int err = -EFAULT;
701
702         if (!(vma->vm_flags & VM_MERGEABLE))
703                 goto out;
704
705         if (!PageAnon(oldpage))
706                 goto out;
707
708         get_page(newpage);
709         get_page(oldpage);
710
711         /*
712          * We need the page lock to read a stable PageSwapCache in
713          * write_protect_page().  We use trylock_page() instead of
714          * lock_page() because we don't want to wait here - we
715          * prefer to continue scanning and merging different pages,
716          * then come back to this page when it is unlocked.
717          */
718         if (!trylock_page(oldpage))
719                 goto out_putpage;
720         /*
721          * If this anonymous page is mapped only here, its pte may need
722          * to be write-protected.  If it's mapped elsewhere, all of its
723          * ptes are necessarily already write-protected.  But in either
724          * case, we need to lock and check page_count is not raised.
725          */
726         if (write_protect_page(vma, oldpage, &orig_pte)) {
727                 unlock_page(oldpage);
728                 goto out_putpage;
729         }
730         unlock_page(oldpage);
731
732         if (pages_identical(oldpage, newpage))
733                 err = replace_page(vma, oldpage, newpage, orig_pte);
734
735 out_putpage:
736         put_page(oldpage);
737         put_page(newpage);
738 out:
739         return err;
740 }
741
742 /*
743  * try_to_merge_two_pages - take two identical pages and prepare them
744  * to be merged into one page.
745  *
746  * This function returns 0 if we successfully mapped two identical pages
747  * into one page, -EFAULT otherwise.
748  *
749  * Note that this function allocates a new kernel page: if one of the pages
750  * is already a ksm page, try_to_merge_with_ksm_page should be used.
751  */
752 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
753                                   struct page *page1, struct mm_struct *mm2,
754                                   unsigned long addr2, struct page *page2)
755 {
756         struct vm_area_struct *vma;
757         struct page *kpage;
758         int err = -EFAULT;
759
760         /*
761          * The number of nodes in the stable tree
762          * is the number of kernel pages that we hold.
763          */
764         if (ksm_max_kernel_pages &&
765             ksm_max_kernel_pages <= ksm_pages_shared)
766                 return err;
767
768         kpage = alloc_page(GFP_HIGHUSER);
769         if (!kpage)
770                 return err;
771
772         down_read(&mm1->mmap_sem);
773         vma = find_vma(mm1, addr1);
774         if (!vma || vma->vm_start > addr1) {
775                 put_page(kpage);
776                 up_read(&mm1->mmap_sem);
777                 return err;
778         }
779
780         copy_user_highpage(kpage, page1, addr1, vma);
781         err = try_to_merge_one_page(vma, page1, kpage);
782         up_read(&mm1->mmap_sem);
783
784         if (!err) {
785                 down_read(&mm2->mmap_sem);
786                 vma = find_vma(mm2, addr2);
787                 if (!vma || vma->vm_start > addr2) {
788                         put_page(kpage);
789                         up_read(&mm2->mmap_sem);
790                         break_cow(mm1, addr1);
791                         return -EFAULT;
792                 }
793
794                 err = try_to_merge_one_page(vma, page2, kpage);
795                 up_read(&mm2->mmap_sem);
796
797                 /*
798                  * If the second try_to_merge_one_page failed, we have a
799                  * ksm page with just one pte pointing to it, so break it.
800                  */
801                 if (err)
802                         break_cow(mm1, addr1);
803         }
804
805         put_page(kpage);
806         return err;
807 }
808
809 /*
810  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
811  * but no new kernel page is allocated: kpage must already be a ksm page.
812  */
813 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
814                                       unsigned long addr1,
815                                       struct page *page1,
816                                       struct page *kpage)
817 {
818         struct vm_area_struct *vma;
819         int err = -EFAULT;
820
821         down_read(&mm1->mmap_sem);
822         vma = find_vma(mm1, addr1);
823         if (!vma || vma->vm_start > addr1) {
824                 up_read(&mm1->mmap_sem);
825                 return err;
826         }
827
828         err = try_to_merge_one_page(vma, page1, kpage);
829         up_read(&mm1->mmap_sem);
830
831         return err;
832 }
833
834 /*
835  * stable_tree_search - search page inside the stable tree
836  * @page: the page that we are searching identical pages to.
837  * @page2: pointer into identical page that we are holding inside the stable
838  *         tree that we have found.
839  * @rmap_item: the reverse mapping item
840  *
841  * This function checks if there is a page inside the stable tree
842  * with identical content to the page that we are scanning right now.
843  *
844  * This function return rmap_item pointer to the identical item if found,
845  * NULL otherwise.
846  */
847 static struct rmap_item *stable_tree_search(struct page *page,
848                                             struct page **page2,
849                                             struct rmap_item *rmap_item)
850 {
851         struct rb_node *node = root_stable_tree.rb_node;
852
853         while (node) {
854                 struct rmap_item *tree_rmap_item, *next_rmap_item;
855                 int ret;
856
857                 tree_rmap_item = rb_entry(node, struct rmap_item, node);
858                 while (tree_rmap_item) {
859                         BUG_ON(!in_stable_tree(tree_rmap_item));
860                         cond_resched();
861                         page2[0] = get_ksm_page(tree_rmap_item);
862                         if (page2[0])
863                                 break;
864                         next_rmap_item = tree_rmap_item->next;
865                         remove_rmap_item_from_tree(tree_rmap_item);
866                         tree_rmap_item = next_rmap_item;
867                 }
868                 if (!tree_rmap_item)
869                         return NULL;
870
871                 ret = memcmp_pages(page, page2[0]);
872
873                 if (ret < 0) {
874                         put_page(page2[0]);
875                         node = node->rb_left;
876                 } else if (ret > 0) {
877                         put_page(page2[0]);
878                         node = node->rb_right;
879                 } else {
880                         return tree_rmap_item;
881                 }
882         }
883
884         return NULL;
885 }
886
887 /*
888  * stable_tree_insert - insert rmap_item pointing to new ksm page
889  * into the stable tree.
890  *
891  * @page: the page that we are searching identical page to inside the stable
892  *        tree.
893  * @rmap_item: pointer to the reverse mapping item.
894  *
895  * This function returns rmap_item if success, NULL otherwise.
896  */
897 static struct rmap_item *stable_tree_insert(struct page *page,
898                                             struct rmap_item *rmap_item)
899 {
900         struct rb_node **new = &root_stable_tree.rb_node;
901         struct rb_node *parent = NULL;
902
903         while (*new) {
904                 struct rmap_item *tree_rmap_item, *next_rmap_item;
905                 struct page *tree_page;
906                 int ret;
907
908                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
909                 while (tree_rmap_item) {
910                         BUG_ON(!in_stable_tree(tree_rmap_item));
911                         cond_resched();
912                         tree_page = get_ksm_page(tree_rmap_item);
913                         if (tree_page)
914                                 break;
915                         next_rmap_item = tree_rmap_item->next;
916                         remove_rmap_item_from_tree(tree_rmap_item);
917                         tree_rmap_item = next_rmap_item;
918                 }
919                 if (!tree_rmap_item)
920                         return NULL;
921
922                 ret = memcmp_pages(page, tree_page);
923                 put_page(tree_page);
924
925                 parent = *new;
926                 if (ret < 0)
927                         new = &parent->rb_left;
928                 else if (ret > 0)
929                         new = &parent->rb_right;
930                 else {
931                         /*
932                          * It is not a bug that stable_tree_search() didn't
933                          * find this node: because at that time our page was
934                          * not yet write-protected, so may have changed since.
935                          */
936                         return NULL;
937                 }
938         }
939
940         rmap_item->address |= NODE_FLAG | STABLE_FLAG;
941         rmap_item->next = NULL;
942         rb_link_node(&rmap_item->node, parent, new);
943         rb_insert_color(&rmap_item->node, &root_stable_tree);
944
945         ksm_pages_shared++;
946         return rmap_item;
947 }
948
949 /*
950  * unstable_tree_search_insert - search and insert items into the unstable tree.
951  *
952  * @page: the page that we are going to search for identical page or to insert
953  *        into the unstable tree
954  * @page2: pointer into identical page that was found inside the unstable tree
955  * @rmap_item: the reverse mapping item of page
956  *
957  * This function searches for a page in the unstable tree identical to the
958  * page currently being scanned; and if no identical page is found in the
959  * tree, we insert rmap_item as a new object into the unstable tree.
960  *
961  * This function returns pointer to rmap_item found to be identical
962  * to the currently scanned page, NULL otherwise.
963  *
964  * This function does both searching and inserting, because they share
965  * the same walking algorithm in an rbtree.
966  */
967 static struct rmap_item *unstable_tree_search_insert(struct page *page,
968                                                 struct page **page2,
969                                                 struct rmap_item *rmap_item)
970 {
971         struct rb_node **new = &root_unstable_tree.rb_node;
972         struct rb_node *parent = NULL;
973
974         while (*new) {
975                 struct rmap_item *tree_rmap_item;
976                 int ret;
977
978                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
979                 page2[0] = get_mergeable_page(tree_rmap_item);
980                 if (!page2[0])
981                         return NULL;
982
983                 /*
984                  * Don't substitute an unswappable ksm page
985                  * just for one good swappable forked page.
986                  */
987                 if (page == page2[0]) {
988                         put_page(page2[0]);
989                         return NULL;
990                 }
991
992                 ret = memcmp_pages(page, page2[0]);
993
994                 parent = *new;
995                 if (ret < 0) {
996                         put_page(page2[0]);
997                         new = &parent->rb_left;
998                 } else if (ret > 0) {
999                         put_page(page2[0]);
1000                         new = &parent->rb_right;
1001                 } else {
1002                         return tree_rmap_item;
1003                 }
1004         }
1005
1006         rmap_item->address |= NODE_FLAG;
1007         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1008         rb_link_node(&rmap_item->node, parent, new);
1009         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1010
1011         ksm_pages_unshared++;
1012         return NULL;
1013 }
1014
1015 /*
1016  * stable_tree_append - add another rmap_item to the linked list of
1017  * rmap_items hanging off a given node of the stable tree, all sharing
1018  * the same ksm page.
1019  */
1020 static void stable_tree_append(struct rmap_item *rmap_item,
1021                                struct rmap_item *tree_rmap_item)
1022 {
1023         rmap_item->next = tree_rmap_item->next;
1024         rmap_item->prev = tree_rmap_item;
1025
1026         if (tree_rmap_item->next)
1027                 tree_rmap_item->next->prev = rmap_item;
1028
1029         tree_rmap_item->next = rmap_item;
1030         rmap_item->address |= STABLE_FLAG;
1031
1032         ksm_pages_sharing++;
1033 }
1034
1035 /*
1036  * cmp_and_merge_page - take a page computes its hash value and check if there
1037  * is similar hash value to different page,
1038  * in case we find that there is similar hash to different page we call to
1039  * try_to_merge_two_pages().
1040  *
1041  * @page: the page that we are searching identical page to.
1042  * @rmap_item: the reverse mapping into the virtual address of this page
1043  */
1044 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1045 {
1046         struct page *page2[1];
1047         struct rmap_item *tree_rmap_item;
1048         unsigned int checksum;
1049         int err;
1050
1051         if (in_stable_tree(rmap_item))
1052                 remove_rmap_item_from_tree(rmap_item);
1053
1054         /* We first start with searching the page inside the stable tree */
1055         tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1056         if (tree_rmap_item) {
1057                 if (page == page2[0])                   /* forked */
1058                         err = 0;
1059                 else
1060                         err = try_to_merge_with_ksm_page(rmap_item->mm,
1061                                                          rmap_item->address,
1062                                                          page, page2[0]);
1063                 put_page(page2[0]);
1064
1065                 if (!err) {
1066                         /*
1067                          * The page was successfully merged:
1068                          * add its rmap_item to the stable tree.
1069                          */
1070                         stable_tree_append(rmap_item, tree_rmap_item);
1071                 }
1072                 return;
1073         }
1074
1075         /*
1076          * A ksm page might have got here by fork, but its other
1077          * references have already been removed from the stable tree.
1078          */
1079         if (PageKsm(page))
1080                 break_cow(rmap_item->mm, rmap_item->address);
1081
1082         /*
1083          * In case the hash value of the page was changed from the last time we
1084          * have calculated it, this page to be changed frequely, therefore we
1085          * don't want to insert it to the unstable tree, and we don't want to
1086          * waste our time to search if there is something identical to it there.
1087          */
1088         checksum = calc_checksum(page);
1089         if (rmap_item->oldchecksum != checksum) {
1090                 rmap_item->oldchecksum = checksum;
1091                 return;
1092         }
1093
1094         tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1095         if (tree_rmap_item) {
1096                 err = try_to_merge_two_pages(rmap_item->mm,
1097                                              rmap_item->address, page,
1098                                              tree_rmap_item->mm,
1099                                              tree_rmap_item->address, page2[0]);
1100                 /*
1101                  * As soon as we merge this page, we want to remove the
1102                  * rmap_item of the page we have merged with from the unstable
1103                  * tree, and insert it instead as new node in the stable tree.
1104                  */
1105                 if (!err) {
1106                         rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1107                         tree_rmap_item->address &= ~NODE_FLAG;
1108                         ksm_pages_unshared--;
1109
1110                         /*
1111                          * If we fail to insert the page into the stable tree,
1112                          * we will have 2 virtual addresses that are pointing
1113                          * to a ksm page left outside the stable tree,
1114                          * in which case we need to break_cow on both.
1115                          */
1116                         if (stable_tree_insert(page2[0], tree_rmap_item))
1117                                 stable_tree_append(rmap_item, tree_rmap_item);
1118                         else {
1119                                 break_cow(tree_rmap_item->mm,
1120                                                 tree_rmap_item->address);
1121                                 break_cow(rmap_item->mm, rmap_item->address);
1122                         }
1123                 }
1124
1125                 put_page(page2[0]);
1126         }
1127 }
1128
1129 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1130                                             struct list_head *cur,
1131                                             unsigned long addr)
1132 {
1133         struct rmap_item *rmap_item;
1134
1135         while (cur != &mm_slot->rmap_list) {
1136                 rmap_item = list_entry(cur, struct rmap_item, link);
1137                 if ((rmap_item->address & PAGE_MASK) == addr) {
1138                         if (!in_stable_tree(rmap_item))
1139                                 remove_rmap_item_from_tree(rmap_item);
1140                         return rmap_item;
1141                 }
1142                 if (rmap_item->address > addr)
1143                         break;
1144                 cur = cur->next;
1145                 remove_rmap_item_from_tree(rmap_item);
1146                 list_del(&rmap_item->link);
1147                 free_rmap_item(rmap_item);
1148         }
1149
1150         rmap_item = alloc_rmap_item();
1151         if (rmap_item) {
1152                 /* It has already been zeroed */
1153                 rmap_item->mm = mm_slot->mm;
1154                 rmap_item->address = addr;
1155                 list_add_tail(&rmap_item->link, cur);
1156         }
1157         return rmap_item;
1158 }
1159
1160 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1161 {
1162         struct mm_struct *mm;
1163         struct mm_slot *slot;
1164         struct vm_area_struct *vma;
1165         struct rmap_item *rmap_item;
1166
1167         if (list_empty(&ksm_mm_head.mm_list))
1168                 return NULL;
1169
1170         slot = ksm_scan.mm_slot;
1171         if (slot == &ksm_mm_head) {
1172                 root_unstable_tree = RB_ROOT;
1173
1174                 spin_lock(&ksm_mmlist_lock);
1175                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1176                 ksm_scan.mm_slot = slot;
1177                 spin_unlock(&ksm_mmlist_lock);
1178 next_mm:
1179                 ksm_scan.address = 0;
1180                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1181                                                 struct rmap_item, link);
1182         }
1183
1184         mm = slot->mm;
1185         down_read(&mm->mmap_sem);
1186         for (vma = find_vma(mm, ksm_scan.address); vma; vma = vma->vm_next) {
1187                 if (!(vma->vm_flags & VM_MERGEABLE))
1188                         continue;
1189                 if (ksm_scan.address < vma->vm_start)
1190                         ksm_scan.address = vma->vm_start;
1191                 if (!vma->anon_vma)
1192                         ksm_scan.address = vma->vm_end;
1193
1194                 while (ksm_scan.address < vma->vm_end) {
1195                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1196                         if (*page && PageAnon(*page)) {
1197                                 flush_anon_page(vma, *page, ksm_scan.address);
1198                                 flush_dcache_page(*page);
1199                                 rmap_item = get_next_rmap_item(slot,
1200                                         ksm_scan.rmap_item->link.next,
1201                                         ksm_scan.address);
1202                                 if (rmap_item) {
1203                                         ksm_scan.rmap_item = rmap_item;
1204                                         ksm_scan.address += PAGE_SIZE;
1205                                 } else
1206                                         put_page(*page);
1207                                 up_read(&mm->mmap_sem);
1208                                 return rmap_item;
1209                         }
1210                         if (*page)
1211                                 put_page(*page);
1212                         ksm_scan.address += PAGE_SIZE;
1213                         cond_resched();
1214                 }
1215         }
1216
1217         if (!ksm_scan.address) {
1218                 /*
1219                  * We've completed a full scan of all vmas, holding mmap_sem
1220                  * throughout, and found no VM_MERGEABLE: so do the same as
1221                  * __ksm_exit does to remove this mm from all our lists now.
1222                  */
1223                 remove_mm_from_lists(mm);
1224                 up_read(&mm->mmap_sem);
1225                 slot = ksm_scan.mm_slot;
1226                 if (slot != &ksm_mm_head)
1227                         goto next_mm;
1228                 return NULL;
1229         }
1230
1231         /*
1232          * Nuke all the rmap_items that are above this current rmap:
1233          * because there were no VM_MERGEABLE vmas with such addresses.
1234          */
1235         remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1236         up_read(&mm->mmap_sem);
1237
1238         spin_lock(&ksm_mmlist_lock);
1239         slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1240         ksm_scan.mm_slot = slot;
1241         spin_unlock(&ksm_mmlist_lock);
1242
1243         /* Repeat until we've completed scanning the whole list */
1244         if (slot != &ksm_mm_head)
1245                 goto next_mm;
1246
1247         /*
1248          * Bump seqnr here rather than at top, so that __ksm_exit
1249          * can skip rb_erase on unstable tree until we run again.
1250          */
1251         ksm_scan.seqnr++;
1252         return NULL;
1253 }
1254
1255 /**
1256  * ksm_do_scan  - the ksm scanner main worker function.
1257  * @scan_npages - number of pages we want to scan before we return.
1258  */
1259 static void ksm_do_scan(unsigned int scan_npages)
1260 {
1261         struct rmap_item *rmap_item;
1262         struct page *page;
1263
1264         while (scan_npages--) {
1265                 cond_resched();
1266                 rmap_item = scan_get_next_rmap_item(&page);
1267                 if (!rmap_item)
1268                         return;
1269                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1270                         cmp_and_merge_page(page, rmap_item);
1271                 put_page(page);
1272         }
1273 }
1274
1275 static int ksm_scan_thread(void *nothing)
1276 {
1277         set_user_nice(current, 5);
1278
1279         while (!kthread_should_stop()) {
1280                 if (ksm_run & KSM_RUN_MERGE) {
1281                         mutex_lock(&ksm_thread_mutex);
1282                         ksm_do_scan(ksm_thread_pages_to_scan);
1283                         mutex_unlock(&ksm_thread_mutex);
1284                         schedule_timeout_interruptible(
1285                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1286                 } else {
1287                         wait_event_interruptible(ksm_thread_wait,
1288                                         (ksm_run & KSM_RUN_MERGE) ||
1289                                         kthread_should_stop());
1290                 }
1291         }
1292         return 0;
1293 }
1294
1295 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1296                 unsigned long end, int advice, unsigned long *vm_flags)
1297 {
1298         struct mm_struct *mm = vma->vm_mm;
1299
1300         switch (advice) {
1301         case MADV_MERGEABLE:
1302                 /*
1303                  * Be somewhat over-protective for now!
1304                  */
1305                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1306                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1307                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1308                                  VM_MIXEDMAP  | VM_SAO))
1309                         return 0;               /* just ignore the advice */
1310
1311                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
1312                         if (__ksm_enter(mm) < 0)
1313                                 return -EAGAIN;
1314
1315                 *vm_flags |= VM_MERGEABLE;
1316                 break;
1317
1318         case MADV_UNMERGEABLE:
1319                 if (!(*vm_flags & VM_MERGEABLE))
1320                         return 0;               /* just ignore the advice */
1321
1322                 if (vma->anon_vma)
1323                         unmerge_ksm_pages(vma, start, end);
1324
1325                 *vm_flags &= ~VM_MERGEABLE;
1326                 break;
1327         }
1328
1329         return 0;
1330 }
1331
1332 int __ksm_enter(struct mm_struct *mm)
1333 {
1334         struct mm_slot *mm_slot = alloc_mm_slot();
1335         if (!mm_slot)
1336                 return -ENOMEM;
1337
1338         spin_lock(&ksm_mmlist_lock);
1339         insert_to_mm_slots_hash(mm, mm_slot);
1340         /*
1341          * Insert just behind the scanning cursor, to let the area settle
1342          * down a little; when fork is followed by immediate exec, we don't
1343          * want ksmd to waste time setting up and tearing down an rmap_list.
1344          */
1345         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1346         spin_unlock(&ksm_mmlist_lock);
1347
1348         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1349         return 0;
1350 }
1351
1352 void __ksm_exit(struct mm_struct *mm)
1353 {
1354         /*
1355          * This process is exiting: doesn't hold and doesn't need mmap_sem;
1356          * but we do need to exclude ksmd and other exiters while we modify
1357          * the various lists and trees.
1358          */
1359         mutex_lock(&ksm_thread_mutex);
1360         remove_mm_from_lists(mm);
1361         mutex_unlock(&ksm_thread_mutex);
1362 }
1363
1364 #define KSM_ATTR_RO(_name) \
1365         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1366 #define KSM_ATTR(_name) \
1367         static struct kobj_attribute _name##_attr = \
1368                 __ATTR(_name, 0644, _name##_show, _name##_store)
1369
1370 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1371                                     struct kobj_attribute *attr, char *buf)
1372 {
1373         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1374 }
1375
1376 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1377                                      struct kobj_attribute *attr,
1378                                      const char *buf, size_t count)
1379 {
1380         unsigned long msecs;
1381         int err;
1382
1383         err = strict_strtoul(buf, 10, &msecs);
1384         if (err || msecs > UINT_MAX)
1385                 return -EINVAL;
1386
1387         ksm_thread_sleep_millisecs = msecs;
1388
1389         return count;
1390 }
1391 KSM_ATTR(sleep_millisecs);
1392
1393 static ssize_t pages_to_scan_show(struct kobject *kobj,
1394                                   struct kobj_attribute *attr, char *buf)
1395 {
1396         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1397 }
1398
1399 static ssize_t pages_to_scan_store(struct kobject *kobj,
1400                                    struct kobj_attribute *attr,
1401                                    const char *buf, size_t count)
1402 {
1403         int err;
1404         unsigned long nr_pages;
1405
1406         err = strict_strtoul(buf, 10, &nr_pages);
1407         if (err || nr_pages > UINT_MAX)
1408                 return -EINVAL;
1409
1410         ksm_thread_pages_to_scan = nr_pages;
1411
1412         return count;
1413 }
1414 KSM_ATTR(pages_to_scan);
1415
1416 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1417                         char *buf)
1418 {
1419         return sprintf(buf, "%u\n", ksm_run);
1420 }
1421
1422 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1423                          const char *buf, size_t count)
1424 {
1425         int err;
1426         unsigned long flags;
1427
1428         err = strict_strtoul(buf, 10, &flags);
1429         if (err || flags > UINT_MAX)
1430                 return -EINVAL;
1431         if (flags > KSM_RUN_UNMERGE)
1432                 return -EINVAL;
1433
1434         /*
1435          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1436          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1437          * breaking COW to free the unswappable pages_shared (but leaves
1438          * mm_slots on the list for when ksmd may be set running again).
1439          */
1440
1441         mutex_lock(&ksm_thread_mutex);
1442         if (ksm_run != flags) {
1443                 ksm_run = flags;
1444                 if (flags & KSM_RUN_UNMERGE)
1445                         unmerge_and_remove_all_rmap_items();
1446         }
1447         mutex_unlock(&ksm_thread_mutex);
1448
1449         if (flags & KSM_RUN_MERGE)
1450                 wake_up_interruptible(&ksm_thread_wait);
1451
1452         return count;
1453 }
1454 KSM_ATTR(run);
1455
1456 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1457                                       struct kobj_attribute *attr,
1458                                       const char *buf, size_t count)
1459 {
1460         int err;
1461         unsigned long nr_pages;
1462
1463         err = strict_strtoul(buf, 10, &nr_pages);
1464         if (err)
1465                 return -EINVAL;
1466
1467         ksm_max_kernel_pages = nr_pages;
1468
1469         return count;
1470 }
1471
1472 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1473                                      struct kobj_attribute *attr, char *buf)
1474 {
1475         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1476 }
1477 KSM_ATTR(max_kernel_pages);
1478
1479 static ssize_t pages_shared_show(struct kobject *kobj,
1480                                  struct kobj_attribute *attr, char *buf)
1481 {
1482         return sprintf(buf, "%lu\n", ksm_pages_shared);
1483 }
1484 KSM_ATTR_RO(pages_shared);
1485
1486 static ssize_t pages_sharing_show(struct kobject *kobj,
1487                                   struct kobj_attribute *attr, char *buf)
1488 {
1489         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1490 }
1491 KSM_ATTR_RO(pages_sharing);
1492
1493 static ssize_t pages_unshared_show(struct kobject *kobj,
1494                                    struct kobj_attribute *attr, char *buf)
1495 {
1496         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1497 }
1498 KSM_ATTR_RO(pages_unshared);
1499
1500 static ssize_t pages_volatile_show(struct kobject *kobj,
1501                                    struct kobj_attribute *attr, char *buf)
1502 {
1503         long ksm_pages_volatile;
1504
1505         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1506                                 - ksm_pages_sharing - ksm_pages_unshared;
1507         /*
1508          * It was not worth any locking to calculate that statistic,
1509          * but it might therefore sometimes be negative: conceal that.
1510          */
1511         if (ksm_pages_volatile < 0)
1512                 ksm_pages_volatile = 0;
1513         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1514 }
1515 KSM_ATTR_RO(pages_volatile);
1516
1517 static ssize_t full_scans_show(struct kobject *kobj,
1518                                struct kobj_attribute *attr, char *buf)
1519 {
1520         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1521 }
1522 KSM_ATTR_RO(full_scans);
1523
1524 static struct attribute *ksm_attrs[] = {
1525         &sleep_millisecs_attr.attr,
1526         &pages_to_scan_attr.attr,
1527         &run_attr.attr,
1528         &max_kernel_pages_attr.attr,
1529         &pages_shared_attr.attr,
1530         &pages_sharing_attr.attr,
1531         &pages_unshared_attr.attr,
1532         &pages_volatile_attr.attr,
1533         &full_scans_attr.attr,
1534         NULL,
1535 };
1536
1537 static struct attribute_group ksm_attr_group = {
1538         .attrs = ksm_attrs,
1539         .name = "ksm",
1540 };
1541
1542 static int __init ksm_init(void)
1543 {
1544         struct task_struct *ksm_thread;
1545         int err;
1546
1547         err = ksm_slab_init();
1548         if (err)
1549                 goto out;
1550
1551         err = mm_slots_hash_init();
1552         if (err)
1553                 goto out_free1;
1554
1555         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1556         if (IS_ERR(ksm_thread)) {
1557                 printk(KERN_ERR "ksm: creating kthread failed\n");
1558                 err = PTR_ERR(ksm_thread);
1559                 goto out_free2;
1560         }
1561
1562         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1563         if (err) {
1564                 printk(KERN_ERR "ksm: register sysfs failed\n");
1565                 goto out_free3;
1566         }
1567
1568         return 0;
1569
1570 out_free3:
1571         kthread_stop(ksm_thread);
1572 out_free2:
1573         mm_slots_hash_free();
1574 out_free1:
1575         ksm_slab_free();
1576 out:
1577         return err;
1578 }
1579 module_init(ksm_init)