7e4d255dadc0609c51e36951e6c22f046ccba234
[linux-2.6.git] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/mmu_notifier.h>
33 #include <linux/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 int break_ksm(struct vm_area_struct *vma, unsigned long addr)
298 {
299         struct page *page;
300         int ret = 0;
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 | VM_FAULT_OOM)));
314         /*
315          * We must loop because handle_mm_fault() may back out if there's
316          * any difficulty e.g. if pte accessed bit gets updated concurrently.
317          *
318          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
319          * COW has been broken, even if the vma does not permit VM_WRITE;
320          * but note that a concurrent fault might break PageKsm for us.
321          *
322          * VM_FAULT_SIGBUS could occur if we race with truncation of the
323          * backing file, which also invalidates anonymous pages: that's
324          * okay, that truncation will have unmapped the PageKsm for us.
325          *
326          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
327          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
328          * current task has TIF_MEMDIE set, and will be OOM killed on return
329          * to user; and ksmd, having no mm, would never be chosen for that.
330          *
331          * But if the mm is in a limited mem_cgroup, then the fault may fail
332          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
333          * even ksmd can fail in this way - though it's usually breaking ksm
334          * just to undo a merge it made a moment before, so unlikely to oom.
335          *
336          * That's a pity: we might therefore have more kernel pages allocated
337          * than we're counting as nodes in the stable tree; but ksm_do_scan
338          * will retry to break_cow on each pass, so should recover the page
339          * in due course.  The important thing is to not let VM_MERGEABLE
340          * be cleared while any such pages might remain in the area.
341          */
342         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
343 }
344
345 static void break_cow(struct mm_struct *mm, unsigned long addr)
346 {
347         struct vm_area_struct *vma;
348
349         down_read(&mm->mmap_sem);
350         vma = find_vma(mm, addr);
351         if (!vma || vma->vm_start > addr)
352                 goto out;
353         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
354                 goto out;
355         break_ksm(vma, addr);
356 out:
357         up_read(&mm->mmap_sem);
358 }
359
360 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
361 {
362         struct mm_struct *mm = rmap_item->mm;
363         unsigned long addr = rmap_item->address;
364         struct vm_area_struct *vma;
365         struct page *page;
366
367         down_read(&mm->mmap_sem);
368         vma = find_vma(mm, addr);
369         if (!vma || vma->vm_start > addr)
370                 goto out;
371         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
372                 goto out;
373
374         page = follow_page(vma, addr, FOLL_GET);
375         if (!page)
376                 goto out;
377         if (PageAnon(page)) {
378                 flush_anon_page(vma, page, addr);
379                 flush_dcache_page(page);
380         } else {
381                 put_page(page);
382 out:            page = NULL;
383         }
384         up_read(&mm->mmap_sem);
385         return page;
386 }
387
388 /*
389  * get_ksm_page: checks if the page at the virtual address in rmap_item
390  * is still PageKsm, in which case we can trust the content of the page,
391  * and it returns the gotten page; but NULL if the page has been zapped.
392  */
393 static struct page *get_ksm_page(struct rmap_item *rmap_item)
394 {
395         struct page *page;
396
397         page = get_mergeable_page(rmap_item);
398         if (page && !PageKsm(page)) {
399                 put_page(page);
400                 page = NULL;
401         }
402         return page;
403 }
404
405 /*
406  * Removing rmap_item from stable or unstable tree.
407  * This function will clean the information from the stable/unstable tree.
408  */
409 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
410 {
411         if (in_stable_tree(rmap_item)) {
412                 struct rmap_item *next_item = rmap_item->next;
413
414                 if (rmap_item->address & NODE_FLAG) {
415                         if (next_item) {
416                                 rb_replace_node(&rmap_item->node,
417                                                 &next_item->node,
418                                                 &root_stable_tree);
419                                 next_item->address |= NODE_FLAG;
420                                 ksm_pages_sharing--;
421                         } else {
422                                 rb_erase(&rmap_item->node, &root_stable_tree);
423                                 ksm_pages_shared--;
424                         }
425                 } else {
426                         struct rmap_item *prev_item = rmap_item->prev;
427
428                         BUG_ON(prev_item->next != rmap_item);
429                         prev_item->next = next_item;
430                         if (next_item) {
431                                 BUG_ON(next_item->prev != rmap_item);
432                                 next_item->prev = rmap_item->prev;
433                         }
434                         ksm_pages_sharing--;
435                 }
436
437                 rmap_item->next = NULL;
438
439         } else if (rmap_item->address & NODE_FLAG) {
440                 unsigned char age;
441                 /*
442                  * ksm_thread can and must skip the rb_erase, because
443                  * root_unstable_tree was already reset to RB_ROOT.
444                  * But __ksm_exit has to be careful: do the rb_erase
445                  * if it's interrupting a scan, and this rmap_item was
446                  * inserted by this scan rather than left from before.
447                  */
448                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
449                 BUG_ON(age > 1);
450                 if (!age)
451                         rb_erase(&rmap_item->node, &root_unstable_tree);
452                 ksm_pages_unshared--;
453         }
454
455         rmap_item->address &= PAGE_MASK;
456
457         cond_resched();         /* we're called from many long loops */
458 }
459
460 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
461                                        struct list_head *cur)
462 {
463         struct rmap_item *rmap_item;
464
465         while (cur != &mm_slot->rmap_list) {
466                 rmap_item = list_entry(cur, struct rmap_item, link);
467                 cur = cur->next;
468                 remove_rmap_item_from_tree(rmap_item);
469                 list_del(&rmap_item->link);
470                 free_rmap_item(rmap_item);
471         }
472 }
473
474 /*
475  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
476  * than check every pte of a given vma, the locking doesn't quite work for
477  * that - an rmap_item is assigned to the stable tree after inserting ksm
478  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
479  * rmap_items from parent to child at fork time (so as not to waste time
480  * if exit comes before the next scan reaches it).
481  *
482  * Similarly, although we'd like to remove rmap_items (so updating counts
483  * and freeing memory) when unmerging an area, it's easier to leave that
484  * to the next pass of ksmd - consider, for example, how ksmd might be
485  * in cmp_and_merge_page on one of the rmap_items we would be removing.
486  */
487 static int unmerge_ksm_pages(struct vm_area_struct *vma,
488                              unsigned long start, unsigned long end)
489 {
490         unsigned long addr;
491         int err = 0;
492
493         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
494                 if (signal_pending(current))
495                         err = -ERESTARTSYS;
496                 else
497                         err = break_ksm(vma, addr);
498         }
499         return err;
500 }
501
502 static int unmerge_and_remove_all_rmap_items(void)
503 {
504         struct mm_slot *mm_slot;
505         struct mm_struct *mm;
506         struct vm_area_struct *vma;
507         int err = 0;
508
509         spin_lock(&ksm_mmlist_lock);
510         mm_slot = list_entry(ksm_mm_head.mm_list.next,
511                                                 struct mm_slot, mm_list);
512         spin_unlock(&ksm_mmlist_lock);
513
514         while (mm_slot != &ksm_mm_head) {
515                 mm = mm_slot->mm;
516                 down_read(&mm->mmap_sem);
517                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
518                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
519                                 continue;
520                         err = unmerge_ksm_pages(vma,
521                                                 vma->vm_start, vma->vm_end);
522                         if (err) {
523                                 up_read(&mm->mmap_sem);
524                                 goto out;
525                         }
526                 }
527                 remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
528                 up_read(&mm->mmap_sem);
529
530                 spin_lock(&ksm_mmlist_lock);
531                 mm_slot = list_entry(mm_slot->mm_list.next,
532                                                 struct mm_slot, mm_list);
533                 spin_unlock(&ksm_mmlist_lock);
534         }
535
536         ksm_scan.seqnr = 0;
537 out:
538         spin_lock(&ksm_mmlist_lock);
539         ksm_scan.mm_slot = &ksm_mm_head;
540         spin_unlock(&ksm_mmlist_lock);
541         return err;
542 }
543
544 static u32 calc_checksum(struct page *page)
545 {
546         u32 checksum;
547         void *addr = kmap_atomic(page, KM_USER0);
548         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
549         kunmap_atomic(addr, KM_USER0);
550         return checksum;
551 }
552
553 static int memcmp_pages(struct page *page1, struct page *page2)
554 {
555         char *addr1, *addr2;
556         int ret;
557
558         addr1 = kmap_atomic(page1, KM_USER0);
559         addr2 = kmap_atomic(page2, KM_USER1);
560         ret = memcmp(addr1, addr2, PAGE_SIZE);
561         kunmap_atomic(addr2, KM_USER1);
562         kunmap_atomic(addr1, KM_USER0);
563         return ret;
564 }
565
566 static inline int pages_identical(struct page *page1, struct page *page2)
567 {
568         return !memcmp_pages(page1, page2);
569 }
570
571 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
572                               pte_t *orig_pte)
573 {
574         struct mm_struct *mm = vma->vm_mm;
575         unsigned long addr;
576         pte_t *ptep;
577         spinlock_t *ptl;
578         int swapped;
579         int err = -EFAULT;
580
581         addr = page_address_in_vma(page, vma);
582         if (addr == -EFAULT)
583                 goto out;
584
585         ptep = page_check_address(page, mm, addr, &ptl, 0);
586         if (!ptep)
587                 goto out;
588
589         if (pte_write(*ptep)) {
590                 pte_t entry;
591
592                 swapped = PageSwapCache(page);
593                 flush_cache_page(vma, addr, page_to_pfn(page));
594                 /*
595                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
596                  * take any lock, therefore the check that we are going to make
597                  * with the pagecount against the mapcount is racey and
598                  * O_DIRECT can happen right after the check.
599                  * So we clear the pte and flush the tlb before the check
600                  * this assure us that no O_DIRECT can happen after the check
601                  * or in the middle of the check.
602                  */
603                 entry = ptep_clear_flush(vma, addr, ptep);
604                 /*
605                  * Check that no O_DIRECT or similar I/O is in progress on the
606                  * page
607                  */
608                 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
609                         set_pte_at_notify(mm, addr, ptep, entry);
610                         goto out_unlock;
611                 }
612                 entry = pte_wrprotect(entry);
613                 set_pte_at_notify(mm, addr, ptep, entry);
614         }
615         *orig_pte = *ptep;
616         err = 0;
617
618 out_unlock:
619         pte_unmap_unlock(ptep, ptl);
620 out:
621         return err;
622 }
623
624 /**
625  * replace_page - replace page in vma by new ksm page
626  * @vma:      vma that holds the pte pointing to oldpage
627  * @oldpage:  the page we are replacing by newpage
628  * @newpage:  the ksm page we replace oldpage by
629  * @orig_pte: the original value of the pte
630  *
631  * Returns 0 on success, -EFAULT on failure.
632  */
633 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
634                         struct page *newpage, pte_t orig_pte)
635 {
636         struct mm_struct *mm = vma->vm_mm;
637         pgd_t *pgd;
638         pud_t *pud;
639         pmd_t *pmd;
640         pte_t *ptep;
641         spinlock_t *ptl;
642         unsigned long addr;
643         pgprot_t prot;
644         int err = -EFAULT;
645
646         prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
647
648         addr = page_address_in_vma(oldpage, vma);
649         if (addr == -EFAULT)
650                 goto out;
651
652         pgd = pgd_offset(mm, addr);
653         if (!pgd_present(*pgd))
654                 goto out;
655
656         pud = pud_offset(pgd, addr);
657         if (!pud_present(*pud))
658                 goto out;
659
660         pmd = pmd_offset(pud, addr);
661         if (!pmd_present(*pmd))
662                 goto out;
663
664         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
665         if (!pte_same(*ptep, orig_pte)) {
666                 pte_unmap_unlock(ptep, ptl);
667                 goto out;
668         }
669
670         get_page(newpage);
671         page_add_ksm_rmap(newpage);
672
673         flush_cache_page(vma, addr, pte_pfn(*ptep));
674         ptep_clear_flush(vma, addr, ptep);
675         set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
676
677         page_remove_rmap(oldpage);
678         put_page(oldpage);
679
680         pte_unmap_unlock(ptep, ptl);
681         err = 0;
682 out:
683         return err;
684 }
685
686 /*
687  * try_to_merge_one_page - take two pages and merge them into one
688  * @vma: the vma that hold the pte pointing into oldpage
689  * @oldpage: the page that we want to replace with newpage
690  * @newpage: the page that we want to map instead of oldpage
691  *
692  * Note:
693  * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
694  * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
695  *
696  * This function returns 0 if the pages were merged, -EFAULT otherwise.
697  */
698 static int try_to_merge_one_page(struct vm_area_struct *vma,
699                                  struct page *oldpage,
700                                  struct page *newpage)
701 {
702         pte_t orig_pte = __pte(0);
703         int err = -EFAULT;
704
705         if (!(vma->vm_flags & VM_MERGEABLE))
706                 goto out;
707
708         if (!PageAnon(oldpage))
709                 goto out;
710
711         get_page(newpage);
712         get_page(oldpage);
713
714         /*
715          * We need the page lock to read a stable PageSwapCache in
716          * write_protect_page().  We use trylock_page() instead of
717          * lock_page() because we don't want to wait here - we
718          * prefer to continue scanning and merging different pages,
719          * then come back to this page when it is unlocked.
720          */
721         if (!trylock_page(oldpage))
722                 goto out_putpage;
723         /*
724          * If this anonymous page is mapped only here, its pte may need
725          * to be write-protected.  If it's mapped elsewhere, all of its
726          * ptes are necessarily already write-protected.  But in either
727          * case, we need to lock and check page_count is not raised.
728          */
729         if (write_protect_page(vma, oldpage, &orig_pte)) {
730                 unlock_page(oldpage);
731                 goto out_putpage;
732         }
733         unlock_page(oldpage);
734
735         if (pages_identical(oldpage, newpage))
736                 err = replace_page(vma, oldpage, newpage, orig_pte);
737
738 out_putpage:
739         put_page(oldpage);
740         put_page(newpage);
741 out:
742         return err;
743 }
744
745 /*
746  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
747  * but no new kernel page is allocated: kpage must already be a ksm page.
748  */
749 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
750                                       unsigned long addr1,
751                                       struct page *page1,
752                                       struct page *kpage)
753 {
754         struct vm_area_struct *vma;
755         int err = -EFAULT;
756
757         down_read(&mm1->mmap_sem);
758         vma = find_vma(mm1, addr1);
759         if (!vma || vma->vm_start > addr1)
760                 goto out;
761
762         err = try_to_merge_one_page(vma, page1, kpage);
763 out:
764         up_read(&mm1->mmap_sem);
765         return err;
766 }
767
768 /*
769  * try_to_merge_two_pages - take two identical pages and prepare them
770  * to be merged into one page.
771  *
772  * This function returns 0 if we successfully mapped two identical pages
773  * into one page, -EFAULT otherwise.
774  *
775  * Note that this function allocates a new kernel page: if one of the pages
776  * is already a ksm page, try_to_merge_with_ksm_page should be used.
777  */
778 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
779                                   struct page *page1, struct mm_struct *mm2,
780                                   unsigned long addr2, struct page *page2)
781 {
782         struct vm_area_struct *vma;
783         struct page *kpage;
784         int err = -EFAULT;
785
786         /*
787          * The number of nodes in the stable tree
788          * is the number of kernel pages that we hold.
789          */
790         if (ksm_max_kernel_pages &&
791             ksm_max_kernel_pages <= ksm_pages_shared)
792                 return err;
793
794         kpage = alloc_page(GFP_HIGHUSER);
795         if (!kpage)
796                 return err;
797
798         down_read(&mm1->mmap_sem);
799         vma = find_vma(mm1, addr1);
800         if (!vma || vma->vm_start > addr1) {
801                 up_read(&mm1->mmap_sem);
802                 goto out;
803         }
804
805         copy_user_highpage(kpage, page1, addr1, vma);
806         err = try_to_merge_one_page(vma, page1, kpage);
807         up_read(&mm1->mmap_sem);
808
809         if (!err) {
810                 err = try_to_merge_with_ksm_page(mm2, addr2, page2, kpage);
811                 /*
812                  * If that fails, we have a ksm page with only one pte
813                  * pointing to it: so break it.
814                  */
815                 if (err)
816                         break_cow(mm1, addr1);
817         }
818 out:
819         put_page(kpage);
820         return err;
821 }
822
823 /*
824  * stable_tree_search - search page inside the stable tree
825  * @page: the page that we are searching identical pages to.
826  * @page2: pointer into identical page that we are holding inside the stable
827  *         tree that we have found.
828  * @rmap_item: the reverse mapping item
829  *
830  * This function checks if there is a page inside the stable tree
831  * with identical content to the page that we are scanning right now.
832  *
833  * This function return rmap_item pointer to the identical item if found,
834  * NULL otherwise.
835  */
836 static struct rmap_item *stable_tree_search(struct page *page,
837                                             struct page **page2,
838                                             struct rmap_item *rmap_item)
839 {
840         struct rb_node *node = root_stable_tree.rb_node;
841
842         while (node) {
843                 struct rmap_item *tree_rmap_item, *next_rmap_item;
844                 int ret;
845
846                 tree_rmap_item = rb_entry(node, struct rmap_item, node);
847                 while (tree_rmap_item) {
848                         BUG_ON(!in_stable_tree(tree_rmap_item));
849                         cond_resched();
850                         page2[0] = get_ksm_page(tree_rmap_item);
851                         if (page2[0])
852                                 break;
853                         next_rmap_item = tree_rmap_item->next;
854                         remove_rmap_item_from_tree(tree_rmap_item);
855                         tree_rmap_item = next_rmap_item;
856                 }
857                 if (!tree_rmap_item)
858                         return NULL;
859
860                 ret = memcmp_pages(page, page2[0]);
861
862                 if (ret < 0) {
863                         put_page(page2[0]);
864                         node = node->rb_left;
865                 } else if (ret > 0) {
866                         put_page(page2[0]);
867                         node = node->rb_right;
868                 } else {
869                         return tree_rmap_item;
870                 }
871         }
872
873         return NULL;
874 }
875
876 /*
877  * stable_tree_insert - insert rmap_item pointing to new ksm page
878  * into the stable tree.
879  *
880  * @page: the page that we are searching identical page to inside the stable
881  *        tree.
882  * @rmap_item: pointer to the reverse mapping item.
883  *
884  * This function returns rmap_item if success, NULL otherwise.
885  */
886 static struct rmap_item *stable_tree_insert(struct page *page,
887                                             struct rmap_item *rmap_item)
888 {
889         struct rb_node **new = &root_stable_tree.rb_node;
890         struct rb_node *parent = NULL;
891
892         while (*new) {
893                 struct rmap_item *tree_rmap_item, *next_rmap_item;
894                 struct page *tree_page;
895                 int ret;
896
897                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
898                 while (tree_rmap_item) {
899                         BUG_ON(!in_stable_tree(tree_rmap_item));
900                         cond_resched();
901                         tree_page = get_ksm_page(tree_rmap_item);
902                         if (tree_page)
903                                 break;
904                         next_rmap_item = tree_rmap_item->next;
905                         remove_rmap_item_from_tree(tree_rmap_item);
906                         tree_rmap_item = next_rmap_item;
907                 }
908                 if (!tree_rmap_item)
909                         return NULL;
910
911                 ret = memcmp_pages(page, tree_page);
912                 put_page(tree_page);
913
914                 parent = *new;
915                 if (ret < 0)
916                         new = &parent->rb_left;
917                 else if (ret > 0)
918                         new = &parent->rb_right;
919                 else {
920                         /*
921                          * It is not a bug that stable_tree_search() didn't
922                          * find this node: because at that time our page was
923                          * not yet write-protected, so may have changed since.
924                          */
925                         return NULL;
926                 }
927         }
928
929         rmap_item->address |= NODE_FLAG | STABLE_FLAG;
930         rmap_item->next = NULL;
931         rb_link_node(&rmap_item->node, parent, new);
932         rb_insert_color(&rmap_item->node, &root_stable_tree);
933
934         ksm_pages_shared++;
935         return rmap_item;
936 }
937
938 /*
939  * unstable_tree_search_insert - search and insert items into the unstable tree.
940  *
941  * @page: the page that we are going to search for identical page or to insert
942  *        into the unstable tree
943  * @page2: pointer into identical page that was found inside the unstable tree
944  * @rmap_item: the reverse mapping item of page
945  *
946  * This function searches for a page in the unstable tree identical to the
947  * page currently being scanned; and if no identical page is found in the
948  * tree, we insert rmap_item as a new object into the unstable tree.
949  *
950  * This function returns pointer to rmap_item found to be identical
951  * to the currently scanned page, NULL otherwise.
952  *
953  * This function does both searching and inserting, because they share
954  * the same walking algorithm in an rbtree.
955  */
956 static struct rmap_item *unstable_tree_search_insert(struct page *page,
957                                                 struct page **page2,
958                                                 struct rmap_item *rmap_item)
959 {
960         struct rb_node **new = &root_unstable_tree.rb_node;
961         struct rb_node *parent = NULL;
962
963         while (*new) {
964                 struct rmap_item *tree_rmap_item;
965                 int ret;
966
967                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
968                 page2[0] = get_mergeable_page(tree_rmap_item);
969                 if (!page2[0])
970                         return NULL;
971
972                 /*
973                  * Don't substitute an unswappable ksm page
974                  * just for one good swappable forked page.
975                  */
976                 if (page == page2[0]) {
977                         put_page(page2[0]);
978                         return NULL;
979                 }
980
981                 ret = memcmp_pages(page, page2[0]);
982
983                 parent = *new;
984                 if (ret < 0) {
985                         put_page(page2[0]);
986                         new = &parent->rb_left;
987                 } else if (ret > 0) {
988                         put_page(page2[0]);
989                         new = &parent->rb_right;
990                 } else {
991                         return tree_rmap_item;
992                 }
993         }
994
995         rmap_item->address |= NODE_FLAG;
996         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
997         rb_link_node(&rmap_item->node, parent, new);
998         rb_insert_color(&rmap_item->node, &root_unstable_tree);
999
1000         ksm_pages_unshared++;
1001         return NULL;
1002 }
1003
1004 /*
1005  * stable_tree_append - add another rmap_item to the linked list of
1006  * rmap_items hanging off a given node of the stable tree, all sharing
1007  * the same ksm page.
1008  */
1009 static void stable_tree_append(struct rmap_item *rmap_item,
1010                                struct rmap_item *tree_rmap_item)
1011 {
1012         rmap_item->next = tree_rmap_item->next;
1013         rmap_item->prev = tree_rmap_item;
1014
1015         if (tree_rmap_item->next)
1016                 tree_rmap_item->next->prev = rmap_item;
1017
1018         tree_rmap_item->next = rmap_item;
1019         rmap_item->address |= STABLE_FLAG;
1020
1021         ksm_pages_sharing++;
1022 }
1023
1024 /*
1025  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1026  * if not, compare checksum to previous and if it's the same, see if page can
1027  * be inserted into the unstable tree, or merged with a page already there and
1028  * both transferred to the stable tree.
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                         err = 0;
1048                 else
1049                         err = try_to_merge_with_ksm_page(rmap_item->mm,
1050                                                          rmap_item->address,
1051                                                          page, page2[0]);
1052                 put_page(page2[0]);
1053
1054                 if (!err) {
1055                         /*
1056                          * The page was successfully merged:
1057                          * add its rmap_item to the stable tree.
1058                          */
1059                         stable_tree_append(rmap_item, tree_rmap_item);
1060                 }
1061                 return;
1062         }
1063
1064         /*
1065          * A ksm page might have got here by fork, but its other
1066          * references have already been removed from the stable tree.
1067          * Or it might be left over from a break_ksm which failed
1068          * when the mem_cgroup had reached its limit: try again now.
1069          */
1070         if (PageKsm(page))
1071                 break_cow(rmap_item->mm, rmap_item->address);
1072
1073         /*
1074          * In case the hash value of the page was changed from the last time we
1075          * have calculated it, this page to be changed frequely, therefore we
1076          * don't want to insert it to the unstable tree, and we don't want to
1077          * waste our time to search if there is something identical to it there.
1078          */
1079         checksum = calc_checksum(page);
1080         if (rmap_item->oldchecksum != checksum) {
1081                 rmap_item->oldchecksum = checksum;
1082                 return;
1083         }
1084
1085         tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1086         if (tree_rmap_item) {
1087                 err = try_to_merge_two_pages(rmap_item->mm,
1088                                              rmap_item->address, page,
1089                                              tree_rmap_item->mm,
1090                                              tree_rmap_item->address, page2[0]);
1091                 /*
1092                  * As soon as we merge this page, we want to remove the
1093                  * rmap_item of the page we have merged with from the unstable
1094                  * tree, and insert it instead as new node in the stable tree.
1095                  */
1096                 if (!err) {
1097                         rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1098                         tree_rmap_item->address &= ~NODE_FLAG;
1099                         ksm_pages_unshared--;
1100
1101                         /*
1102                          * If we fail to insert the page into the stable tree,
1103                          * we will have 2 virtual addresses that are pointing
1104                          * to a ksm page left outside the stable tree,
1105                          * in which case we need to break_cow on both.
1106                          */
1107                         if (stable_tree_insert(page2[0], tree_rmap_item))
1108                                 stable_tree_append(rmap_item, tree_rmap_item);
1109                         else {
1110                                 break_cow(tree_rmap_item->mm,
1111                                                 tree_rmap_item->address);
1112                                 break_cow(rmap_item->mm, rmap_item->address);
1113                         }
1114                 }
1115
1116                 put_page(page2[0]);
1117         }
1118 }
1119
1120 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1121                                             struct list_head *cur,
1122                                             unsigned long addr)
1123 {
1124         struct rmap_item *rmap_item;
1125
1126         while (cur != &mm_slot->rmap_list) {
1127                 rmap_item = list_entry(cur, struct rmap_item, link);
1128                 if ((rmap_item->address & PAGE_MASK) == addr) {
1129                         if (!in_stable_tree(rmap_item))
1130                                 remove_rmap_item_from_tree(rmap_item);
1131                         return rmap_item;
1132                 }
1133                 if (rmap_item->address > addr)
1134                         break;
1135                 cur = cur->next;
1136                 remove_rmap_item_from_tree(rmap_item);
1137                 list_del(&rmap_item->link);
1138                 free_rmap_item(rmap_item);
1139         }
1140
1141         rmap_item = alloc_rmap_item();
1142         if (rmap_item) {
1143                 /* It has already been zeroed */
1144                 rmap_item->mm = mm_slot->mm;
1145                 rmap_item->address = addr;
1146                 list_add_tail(&rmap_item->link, cur);
1147         }
1148         return rmap_item;
1149 }
1150
1151 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1152 {
1153         struct mm_struct *mm;
1154         struct mm_slot *slot;
1155         struct vm_area_struct *vma;
1156         struct rmap_item *rmap_item;
1157
1158         if (list_empty(&ksm_mm_head.mm_list))
1159                 return NULL;
1160
1161         slot = ksm_scan.mm_slot;
1162         if (slot == &ksm_mm_head) {
1163                 root_unstable_tree = RB_ROOT;
1164
1165                 spin_lock(&ksm_mmlist_lock);
1166                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1167                 ksm_scan.mm_slot = slot;
1168                 spin_unlock(&ksm_mmlist_lock);
1169 next_mm:
1170                 ksm_scan.address = 0;
1171                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1172                                                 struct rmap_item, link);
1173         }
1174
1175         mm = slot->mm;
1176         down_read(&mm->mmap_sem);
1177         for (vma = find_vma(mm, ksm_scan.address); vma; vma = vma->vm_next) {
1178                 if (!(vma->vm_flags & VM_MERGEABLE))
1179                         continue;
1180                 if (ksm_scan.address < vma->vm_start)
1181                         ksm_scan.address = vma->vm_start;
1182                 if (!vma->anon_vma)
1183                         ksm_scan.address = vma->vm_end;
1184
1185                 while (ksm_scan.address < vma->vm_end) {
1186                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1187                         if (*page && PageAnon(*page)) {
1188                                 flush_anon_page(vma, *page, ksm_scan.address);
1189                                 flush_dcache_page(*page);
1190                                 rmap_item = get_next_rmap_item(slot,
1191                                         ksm_scan.rmap_item->link.next,
1192                                         ksm_scan.address);
1193                                 if (rmap_item) {
1194                                         ksm_scan.rmap_item = rmap_item;
1195                                         ksm_scan.address += PAGE_SIZE;
1196                                 } else
1197                                         put_page(*page);
1198                                 up_read(&mm->mmap_sem);
1199                                 return rmap_item;
1200                         }
1201                         if (*page)
1202                                 put_page(*page);
1203                         ksm_scan.address += PAGE_SIZE;
1204                         cond_resched();
1205                 }
1206         }
1207
1208         /*
1209          * Nuke all the rmap_items that are above this current rmap:
1210          * because there were no VM_MERGEABLE vmas with such addresses.
1211          */
1212         remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1213
1214         spin_lock(&ksm_mmlist_lock);
1215         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1216                                                 struct mm_slot, mm_list);
1217         if (ksm_scan.address == 0) {
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                 hlist_del(&slot->link);
1224                 list_del(&slot->mm_list);
1225                 free_mm_slot(slot);
1226                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1227         }
1228         spin_unlock(&ksm_mmlist_lock);
1229         up_read(&mm->mmap_sem);
1230
1231         /* Repeat until we've completed scanning the whole list */
1232         slot = ksm_scan.mm_slot;
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                 else if (page_mapcount(page) == 1) {
1261                         /*
1262                          * Replace now-unshared ksm page by ordinary page.
1263                          */
1264                         break_cow(rmap_item->mm, rmap_item->address);
1265                         remove_rmap_item_from_tree(rmap_item);
1266                         rmap_item->oldchecksum = calc_checksum(page);
1267                 }
1268                 put_page(page);
1269         }
1270 }
1271
1272 static int ksmd_should_run(void)
1273 {
1274         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1275 }
1276
1277 static int ksm_scan_thread(void *nothing)
1278 {
1279         set_user_nice(current, 5);
1280
1281         while (!kthread_should_stop()) {
1282                 mutex_lock(&ksm_thread_mutex);
1283                 if (ksmd_should_run())
1284                         ksm_do_scan(ksm_thread_pages_to_scan);
1285                 mutex_unlock(&ksm_thread_mutex);
1286
1287                 if (ksmd_should_run()) {
1288                         schedule_timeout_interruptible(
1289                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1290                 } else {
1291                         wait_event_interruptible(ksm_thread_wait,
1292                                 ksmd_should_run() || kthread_should_stop());
1293                 }
1294         }
1295         return 0;
1296 }
1297
1298 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1299                 unsigned long end, int advice, unsigned long *vm_flags)
1300 {
1301         struct mm_struct *mm = vma->vm_mm;
1302         int err;
1303
1304         switch (advice) {
1305         case MADV_MERGEABLE:
1306                 /*
1307                  * Be somewhat over-protective for now!
1308                  */
1309                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1310                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1311                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1312                                  VM_MIXEDMAP  | VM_SAO))
1313                         return 0;               /* just ignore the advice */
1314
1315                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1316                         err = __ksm_enter(mm);
1317                         if (err)
1318                                 return err;
1319                 }
1320
1321                 *vm_flags |= VM_MERGEABLE;
1322                 break;
1323
1324         case MADV_UNMERGEABLE:
1325                 if (!(*vm_flags & VM_MERGEABLE))
1326                         return 0;               /* just ignore the advice */
1327
1328                 if (vma->anon_vma) {
1329                         err = unmerge_ksm_pages(vma, start, end);
1330                         if (err)
1331                                 return err;
1332                 }
1333
1334                 *vm_flags &= ~VM_MERGEABLE;
1335                 break;
1336         }
1337
1338         return 0;
1339 }
1340
1341 int __ksm_enter(struct mm_struct *mm)
1342 {
1343         struct mm_slot *mm_slot;
1344         int needs_wakeup;
1345
1346         mm_slot = alloc_mm_slot();
1347         if (!mm_slot)
1348                 return -ENOMEM;
1349
1350         /* Check ksm_run too?  Would need tighter locking */
1351         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1352
1353         spin_lock(&ksm_mmlist_lock);
1354         insert_to_mm_slots_hash(mm, mm_slot);
1355         /*
1356          * Insert just behind the scanning cursor, to let the area settle
1357          * down a little; when fork is followed by immediate exec, we don't
1358          * want ksmd to waste time setting up and tearing down an rmap_list.
1359          */
1360         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1361         spin_unlock(&ksm_mmlist_lock);
1362
1363         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1364
1365         if (needs_wakeup)
1366                 wake_up_interruptible(&ksm_thread_wait);
1367
1368         return 0;
1369 }
1370
1371 void __ksm_exit(struct mm_struct *mm)
1372 {
1373         struct mm_slot *mm_slot;
1374
1375         /*
1376          * This process is exiting: doesn't hold and doesn't need mmap_sem;
1377          * but we do need to exclude ksmd and other exiters while we modify
1378          * the various lists and trees.
1379          */
1380         mutex_lock(&ksm_thread_mutex);
1381         spin_lock(&ksm_mmlist_lock);
1382         mm_slot = get_mm_slot(mm);
1383         if (!list_empty(&mm_slot->rmap_list)) {
1384                 spin_unlock(&ksm_mmlist_lock);
1385                 remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
1386                 spin_lock(&ksm_mmlist_lock);
1387         }
1388
1389         if (ksm_scan.mm_slot == mm_slot) {
1390                 ksm_scan.mm_slot = list_entry(
1391                         mm_slot->mm_list.next, struct mm_slot, mm_list);
1392                 ksm_scan.address = 0;
1393                 ksm_scan.rmap_item = list_entry(
1394                         &ksm_scan.mm_slot->rmap_list, struct rmap_item, link);
1395                 if (ksm_scan.mm_slot == &ksm_mm_head)
1396                         ksm_scan.seqnr++;
1397         }
1398
1399         hlist_del(&mm_slot->link);
1400         list_del(&mm_slot->mm_list);
1401         spin_unlock(&ksm_mmlist_lock);
1402
1403         free_mm_slot(mm_slot);
1404         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1405         mutex_unlock(&ksm_thread_mutex);
1406 }
1407
1408 #define KSM_ATTR_RO(_name) \
1409         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1410 #define KSM_ATTR(_name) \
1411         static struct kobj_attribute _name##_attr = \
1412                 __ATTR(_name, 0644, _name##_show, _name##_store)
1413
1414 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1415                                     struct kobj_attribute *attr, char *buf)
1416 {
1417         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1418 }
1419
1420 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1421                                      struct kobj_attribute *attr,
1422                                      const char *buf, size_t count)
1423 {
1424         unsigned long msecs;
1425         int err;
1426
1427         err = strict_strtoul(buf, 10, &msecs);
1428         if (err || msecs > UINT_MAX)
1429                 return -EINVAL;
1430
1431         ksm_thread_sleep_millisecs = msecs;
1432
1433         return count;
1434 }
1435 KSM_ATTR(sleep_millisecs);
1436
1437 static ssize_t pages_to_scan_show(struct kobject *kobj,
1438                                   struct kobj_attribute *attr, char *buf)
1439 {
1440         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1441 }
1442
1443 static ssize_t pages_to_scan_store(struct kobject *kobj,
1444                                    struct kobj_attribute *attr,
1445                                    const char *buf, size_t count)
1446 {
1447         int err;
1448         unsigned long nr_pages;
1449
1450         err = strict_strtoul(buf, 10, &nr_pages);
1451         if (err || nr_pages > UINT_MAX)
1452                 return -EINVAL;
1453
1454         ksm_thread_pages_to_scan = nr_pages;
1455
1456         return count;
1457 }
1458 KSM_ATTR(pages_to_scan);
1459
1460 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1461                         char *buf)
1462 {
1463         return sprintf(buf, "%u\n", ksm_run);
1464 }
1465
1466 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1467                          const char *buf, size_t count)
1468 {
1469         int err;
1470         unsigned long flags;
1471
1472         err = strict_strtoul(buf, 10, &flags);
1473         if (err || flags > UINT_MAX)
1474                 return -EINVAL;
1475         if (flags > KSM_RUN_UNMERGE)
1476                 return -EINVAL;
1477
1478         /*
1479          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1480          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1481          * breaking COW to free the unswappable pages_shared (but leaves
1482          * mm_slots on the list for when ksmd may be set running again).
1483          */
1484
1485         mutex_lock(&ksm_thread_mutex);
1486         if (ksm_run != flags) {
1487                 ksm_run = flags;
1488                 if (flags & KSM_RUN_UNMERGE) {
1489                         err = unmerge_and_remove_all_rmap_items();
1490                         if (err) {
1491                                 ksm_run = KSM_RUN_STOP;
1492                                 count = err;
1493                         }
1494                 }
1495         }
1496         mutex_unlock(&ksm_thread_mutex);
1497
1498         if (flags & KSM_RUN_MERGE)
1499                 wake_up_interruptible(&ksm_thread_wait);
1500
1501         return count;
1502 }
1503 KSM_ATTR(run);
1504
1505 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1506                                       struct kobj_attribute *attr,
1507                                       const char *buf, size_t count)
1508 {
1509         int err;
1510         unsigned long nr_pages;
1511
1512         err = strict_strtoul(buf, 10, &nr_pages);
1513         if (err)
1514                 return -EINVAL;
1515
1516         ksm_max_kernel_pages = nr_pages;
1517
1518         return count;
1519 }
1520
1521 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1522                                      struct kobj_attribute *attr, char *buf)
1523 {
1524         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1525 }
1526 KSM_ATTR(max_kernel_pages);
1527
1528 static ssize_t pages_shared_show(struct kobject *kobj,
1529                                  struct kobj_attribute *attr, char *buf)
1530 {
1531         return sprintf(buf, "%lu\n", ksm_pages_shared);
1532 }
1533 KSM_ATTR_RO(pages_shared);
1534
1535 static ssize_t pages_sharing_show(struct kobject *kobj,
1536                                   struct kobj_attribute *attr, char *buf)
1537 {
1538         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1539 }
1540 KSM_ATTR_RO(pages_sharing);
1541
1542 static ssize_t pages_unshared_show(struct kobject *kobj,
1543                                    struct kobj_attribute *attr, char *buf)
1544 {
1545         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1546 }
1547 KSM_ATTR_RO(pages_unshared);
1548
1549 static ssize_t pages_volatile_show(struct kobject *kobj,
1550                                    struct kobj_attribute *attr, char *buf)
1551 {
1552         long ksm_pages_volatile;
1553
1554         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1555                                 - ksm_pages_sharing - ksm_pages_unshared;
1556         /*
1557          * It was not worth any locking to calculate that statistic,
1558          * but it might therefore sometimes be negative: conceal that.
1559          */
1560         if (ksm_pages_volatile < 0)
1561                 ksm_pages_volatile = 0;
1562         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1563 }
1564 KSM_ATTR_RO(pages_volatile);
1565
1566 static ssize_t full_scans_show(struct kobject *kobj,
1567                                struct kobj_attribute *attr, char *buf)
1568 {
1569         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1570 }
1571 KSM_ATTR_RO(full_scans);
1572
1573 static struct attribute *ksm_attrs[] = {
1574         &sleep_millisecs_attr.attr,
1575         &pages_to_scan_attr.attr,
1576         &run_attr.attr,
1577         &max_kernel_pages_attr.attr,
1578         &pages_shared_attr.attr,
1579         &pages_sharing_attr.attr,
1580         &pages_unshared_attr.attr,
1581         &pages_volatile_attr.attr,
1582         &full_scans_attr.attr,
1583         NULL,
1584 };
1585
1586 static struct attribute_group ksm_attr_group = {
1587         .attrs = ksm_attrs,
1588         .name = "ksm",
1589 };
1590
1591 static int __init ksm_init(void)
1592 {
1593         struct task_struct *ksm_thread;
1594         int err;
1595
1596         err = ksm_slab_init();
1597         if (err)
1598                 goto out;
1599
1600         err = mm_slots_hash_init();
1601         if (err)
1602                 goto out_free1;
1603
1604         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1605         if (IS_ERR(ksm_thread)) {
1606                 printk(KERN_ERR "ksm: creating kthread failed\n");
1607                 err = PTR_ERR(ksm_thread);
1608                 goto out_free2;
1609         }
1610
1611         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1612         if (err) {
1613                 printk(KERN_ERR "ksm: register sysfs failed\n");
1614                 goto out_free3;
1615         }
1616
1617         return 0;
1618
1619 out_free3:
1620         kthread_stop(ksm_thread);
1621 out_free2:
1622         mm_slots_hash_free();
1623 out_free1:
1624         ksm_slab_free();
1625 out:
1626         return err;
1627 }
1628 module_init(ksm_init)