HWPOISON: return early on non-LRU pages
[safe/jmp/linux-2.6] / mm / ksm.c
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  *      Izik Eidus
10  *      Andrea Arcangeli
11  *      Chris Wright
12  *      Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/mmu_notifier.h>
33 #include <linux/swap.h>
34 #include <linux/ksm.h>
35
36 #include <asm/tlbflush.h>
37
38 /*
39  * A few notes about the KSM scanning process,
40  * to make it easier to understand the data structures below:
41  *
42  * In order to reduce excessive scanning, KSM sorts the memory pages by their
43  * contents into a data structure that holds pointers to the pages' locations.
44  *
45  * Since the contents of the pages may change at any moment, KSM cannot just
46  * insert the pages into a normal sorted tree and expect it to find anything.
47  * Therefore KSM uses two data structures - the stable and the unstable tree.
48  *
49  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
50  * by their contents.  Because each such page is write-protected, searching on
51  * this tree is fully assured to be working (except when pages are unmapped),
52  * and therefore this tree is called the stable tree.
53  *
54  * In addition to the stable tree, KSM uses a second data structure called the
55  * unstable tree: this tree holds pointers to pages which have been found to
56  * be "unchanged for a period of time".  The unstable tree sorts these pages
57  * by their contents, but since they are not write-protected, KSM cannot rely
58  * upon the unstable tree to work correctly - the unstable tree is liable to
59  * be corrupted as its contents are modified, and so it is called unstable.
60  *
61  * KSM solves this problem by several techniques:
62  *
63  * 1) The unstable tree is flushed every time KSM completes scanning all
64  *    memory areas, and then the tree is rebuilt again from the beginning.
65  * 2) KSM will only insert into the unstable tree, pages whose hash value
66  *    has not changed since the previous scan of all memory areas.
67  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
68  *    colors of the nodes and not on their contents, assuring that even when
69  *    the tree gets "corrupted" it won't get out of balance, so scanning time
70  *    remains the same (also, searching and inserting nodes in an rbtree uses
71  *    the same algorithm, so we have no overhead when we flush and rebuild).
72  * 4) KSM never flushes the stable tree, which means that even if it were to
73  *    take 10 attempts to find a page in the unstable tree, once it is found,
74  *    it is secured in the stable tree.  (When we scan a new page, we first
75  *    compare it against the stable tree, and then against the unstable tree.)
76  */
77
78 /**
79  * struct mm_slot - ksm information per mm that is being scanned
80  * @link: link to the mm_slots hash list
81  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
82  * @rmap_list: head for this mm_slot's list of rmap_items
83  * @mm: the mm that this information is valid for
84  */
85 struct mm_slot {
86         struct hlist_node link;
87         struct list_head mm_list;
88         struct list_head rmap_list;
89         struct mm_struct *mm;
90 };
91
92 /**
93  * struct ksm_scan - cursor for scanning
94  * @mm_slot: the current mm_slot we are scanning
95  * @address: the next address inside that to be scanned
96  * @rmap_item: the current rmap that we are scanning inside the rmap_list
97  * @seqnr: count of completed full scans (needed when removing unstable node)
98  *
99  * There is only the one ksm_scan instance of this cursor structure.
100  */
101 struct ksm_scan {
102         struct mm_slot *mm_slot;
103         unsigned long address;
104         struct rmap_item *rmap_item;
105         unsigned long seqnr;
106 };
107
108 /**
109  * struct rmap_item - reverse mapping item for virtual addresses
110  * @link: link into mm_slot's rmap_list (rmap_list is per mm)
111  * @mm: the memory structure this rmap_item is pointing into
112  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
113  * @oldchecksum: previous checksum of the page at that virtual address
114  * @node: rb_node of this rmap_item in either unstable or stable tree
115  * @next: next rmap_item hanging off the same node of the stable tree
116  * @prev: previous rmap_item hanging off the same node of the stable tree
117  */
118 struct rmap_item {
119         struct list_head link;
120         struct mm_struct *mm;
121         unsigned long address;          /* + low bits used for flags below */
122         union {
123                 unsigned int oldchecksum;               /* when unstable */
124                 struct rmap_item *next;                 /* when stable */
125         };
126         union {
127                 struct rb_node node;                    /* when tree node */
128                 struct rmap_item *prev;                 /* in stable list */
129         };
130 };
131
132 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
133 #define NODE_FLAG       0x100   /* is a node of unstable or stable tree */
134 #define STABLE_FLAG     0x200   /* is a node or list item of stable tree */
135
136 /* The stable and unstable tree heads */
137 static struct rb_root root_stable_tree = RB_ROOT;
138 static struct rb_root root_unstable_tree = RB_ROOT;
139
140 #define MM_SLOTS_HASH_HEADS 1024
141 static struct hlist_head *mm_slots_hash;
142
143 static struct mm_slot ksm_mm_head = {
144         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
145 };
146 static struct ksm_scan ksm_scan = {
147         .mm_slot = &ksm_mm_head,
148 };
149
150 static struct kmem_cache *rmap_item_cache;
151 static struct kmem_cache *mm_slot_cache;
152
153 /* The number of nodes in the stable tree */
154 static unsigned long ksm_pages_shared;
155
156 /* The number of page slots additionally sharing those nodes */
157 static unsigned long ksm_pages_sharing;
158
159 /* The number of nodes in the unstable tree */
160 static unsigned long ksm_pages_unshared;
161
162 /* The number of rmap_items in use: to calculate pages_volatile */
163 static unsigned long ksm_rmap_items;
164
165 /* Limit on the number of unswappable pages used */
166 static unsigned long ksm_max_kernel_pages;
167
168 /* Number of pages ksmd should scan in one batch */
169 static unsigned int ksm_thread_pages_to_scan = 100;
170
171 /* Milliseconds ksmd should sleep between batches */
172 static unsigned int ksm_thread_sleep_millisecs = 20;
173
174 #define KSM_RUN_STOP    0
175 #define KSM_RUN_MERGE   1
176 #define KSM_RUN_UNMERGE 2
177 static unsigned int ksm_run = KSM_RUN_STOP;
178
179 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
180 static DEFINE_MUTEX(ksm_thread_mutex);
181 static DEFINE_SPINLOCK(ksm_mmlist_lock);
182
183 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
184                 sizeof(struct __struct), __alignof__(struct __struct),\
185                 (__flags), NULL)
186
187 static void __init ksm_init_max_kernel_pages(void)
188 {
189         ksm_max_kernel_pages = nr_free_buffer_pages() / 4;
190 }
191
192 static int __init ksm_slab_init(void)
193 {
194         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
195         if (!rmap_item_cache)
196                 goto out;
197
198         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
199         if (!mm_slot_cache)
200                 goto out_free;
201
202         return 0;
203
204 out_free:
205         kmem_cache_destroy(rmap_item_cache);
206 out:
207         return -ENOMEM;
208 }
209
210 static void __init ksm_slab_free(void)
211 {
212         kmem_cache_destroy(mm_slot_cache);
213         kmem_cache_destroy(rmap_item_cache);
214         mm_slot_cache = NULL;
215 }
216
217 static inline struct rmap_item *alloc_rmap_item(void)
218 {
219         struct rmap_item *rmap_item;
220
221         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
222         if (rmap_item)
223                 ksm_rmap_items++;
224         return rmap_item;
225 }
226
227 static inline void free_rmap_item(struct rmap_item *rmap_item)
228 {
229         ksm_rmap_items--;
230         rmap_item->mm = NULL;   /* debug safety */
231         kmem_cache_free(rmap_item_cache, rmap_item);
232 }
233
234 static inline struct mm_slot *alloc_mm_slot(void)
235 {
236         if (!mm_slot_cache)     /* initialization failed */
237                 return NULL;
238         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
239 }
240
241 static inline void free_mm_slot(struct mm_slot *mm_slot)
242 {
243         kmem_cache_free(mm_slot_cache, mm_slot);
244 }
245
246 static int __init mm_slots_hash_init(void)
247 {
248         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
249                                 GFP_KERNEL);
250         if (!mm_slots_hash)
251                 return -ENOMEM;
252         return 0;
253 }
254
255 static void __init mm_slots_hash_free(void)
256 {
257         kfree(mm_slots_hash);
258 }
259
260 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
261 {
262         struct mm_slot *mm_slot;
263         struct hlist_head *bucket;
264         struct hlist_node *node;
265
266         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
267                                 % MM_SLOTS_HASH_HEADS];
268         hlist_for_each_entry(mm_slot, node, bucket, link) {
269                 if (mm == mm_slot->mm)
270                         return mm_slot;
271         }
272         return NULL;
273 }
274
275 static void insert_to_mm_slots_hash(struct mm_struct *mm,
276                                     struct mm_slot *mm_slot)
277 {
278         struct hlist_head *bucket;
279
280         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
281                                 % MM_SLOTS_HASH_HEADS];
282         mm_slot->mm = mm;
283         INIT_LIST_HEAD(&mm_slot->rmap_list);
284         hlist_add_head(&mm_slot->link, bucket);
285 }
286
287 static inline int in_stable_tree(struct rmap_item *rmap_item)
288 {
289         return rmap_item->address & STABLE_FLAG;
290 }
291
292 /*
293  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
294  * page tables after it has passed through ksm_exit() - which, if necessary,
295  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
296  * a special flag: they can just back out as soon as mm_users goes to zero.
297  * ksm_test_exit() is used throughout to make this test for exit: in some
298  * places for correctness, in some places just to avoid unnecessary work.
299  */
300 static inline bool ksm_test_exit(struct mm_struct *mm)
301 {
302         return atomic_read(&mm->mm_users) == 0;
303 }
304
305 /*
306  * We use break_ksm to break COW on a ksm page: it's a stripped down
307  *
308  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
309  *              put_page(page);
310  *
311  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
312  * in case the application has unmapped and remapped mm,addr meanwhile.
313  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
314  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
315  */
316 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
317 {
318         struct page *page;
319         int ret = 0;
320
321         do {
322                 cond_resched();
323                 page = follow_page(vma, addr, FOLL_GET);
324                 if (!page)
325                         break;
326                 if (PageKsm(page))
327                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
328                                                         FAULT_FLAG_WRITE);
329                 else
330                         ret = VM_FAULT_WRITE;
331                 put_page(page);
332         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
333         /*
334          * We must loop because handle_mm_fault() may back out if there's
335          * any difficulty e.g. if pte accessed bit gets updated concurrently.
336          *
337          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
338          * COW has been broken, even if the vma does not permit VM_WRITE;
339          * but note that a concurrent fault might break PageKsm for us.
340          *
341          * VM_FAULT_SIGBUS could occur if we race with truncation of the
342          * backing file, which also invalidates anonymous pages: that's
343          * okay, that truncation will have unmapped the PageKsm for us.
344          *
345          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
346          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
347          * current task has TIF_MEMDIE set, and will be OOM killed on return
348          * to user; and ksmd, having no mm, would never be chosen for that.
349          *
350          * But if the mm is in a limited mem_cgroup, then the fault may fail
351          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
352          * even ksmd can fail in this way - though it's usually breaking ksm
353          * just to undo a merge it made a moment before, so unlikely to oom.
354          *
355          * That's a pity: we might therefore have more kernel pages allocated
356          * than we're counting as nodes in the stable tree; but ksm_do_scan
357          * will retry to break_cow on each pass, so should recover the page
358          * in due course.  The important thing is to not let VM_MERGEABLE
359          * be cleared while any such pages might remain in the area.
360          */
361         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
362 }
363
364 static void break_cow(struct mm_struct *mm, unsigned long addr)
365 {
366         struct vm_area_struct *vma;
367
368         down_read(&mm->mmap_sem);
369         if (ksm_test_exit(mm))
370                 goto out;
371         vma = find_vma(mm, addr);
372         if (!vma || vma->vm_start > addr)
373                 goto out;
374         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
375                 goto out;
376         break_ksm(vma, addr);
377 out:
378         up_read(&mm->mmap_sem);
379 }
380
381 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
382 {
383         struct mm_struct *mm = rmap_item->mm;
384         unsigned long addr = rmap_item->address;
385         struct vm_area_struct *vma;
386         struct page *page;
387
388         down_read(&mm->mmap_sem);
389         if (ksm_test_exit(mm))
390                 goto out;
391         vma = find_vma(mm, addr);
392         if (!vma || vma->vm_start > addr)
393                 goto out;
394         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
395                 goto out;
396
397         page = follow_page(vma, addr, FOLL_GET);
398         if (!page)
399                 goto out;
400         if (PageAnon(page)) {
401                 flush_anon_page(vma, page, addr);
402                 flush_dcache_page(page);
403         } else {
404                 put_page(page);
405 out:            page = NULL;
406         }
407         up_read(&mm->mmap_sem);
408         return page;
409 }
410
411 /*
412  * get_ksm_page: checks if the page at the virtual address in rmap_item
413  * is still PageKsm, in which case we can trust the content of the page,
414  * and it returns the gotten page; but NULL if the page has been zapped.
415  */
416 static struct page *get_ksm_page(struct rmap_item *rmap_item)
417 {
418         struct page *page;
419
420         page = get_mergeable_page(rmap_item);
421         if (page && !PageKsm(page)) {
422                 put_page(page);
423                 page = NULL;
424         }
425         return page;
426 }
427
428 /*
429  * Removing rmap_item from stable or unstable tree.
430  * This function will clean the information from the stable/unstable tree.
431  */
432 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
433 {
434         if (in_stable_tree(rmap_item)) {
435                 struct rmap_item *next_item = rmap_item->next;
436
437                 if (rmap_item->address & NODE_FLAG) {
438                         if (next_item) {
439                                 rb_replace_node(&rmap_item->node,
440                                                 &next_item->node,
441                                                 &root_stable_tree);
442                                 next_item->address |= NODE_FLAG;
443                                 ksm_pages_sharing--;
444                         } else {
445                                 rb_erase(&rmap_item->node, &root_stable_tree);
446                                 ksm_pages_shared--;
447                         }
448                 } else {
449                         struct rmap_item *prev_item = rmap_item->prev;
450
451                         BUG_ON(prev_item->next != rmap_item);
452                         prev_item->next = next_item;
453                         if (next_item) {
454                                 BUG_ON(next_item->prev != rmap_item);
455                                 next_item->prev = rmap_item->prev;
456                         }
457                         ksm_pages_sharing--;
458                 }
459
460                 rmap_item->next = NULL;
461
462         } else if (rmap_item->address & NODE_FLAG) {
463                 unsigned char age;
464                 /*
465                  * Usually ksmd can and must skip the rb_erase, because
466                  * root_unstable_tree was already reset to RB_ROOT.
467                  * But be careful when an mm is exiting: do the rb_erase
468                  * if this rmap_item was inserted by this scan, rather
469                  * than left over from before.
470                  */
471                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
472                 BUG_ON(age > 1);
473                 if (!age)
474                         rb_erase(&rmap_item->node, &root_unstable_tree);
475                 ksm_pages_unshared--;
476         }
477
478         rmap_item->address &= PAGE_MASK;
479
480         cond_resched();         /* we're called from many long loops */
481 }
482
483 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
484                                        struct list_head *cur)
485 {
486         struct rmap_item *rmap_item;
487
488         while (cur != &mm_slot->rmap_list) {
489                 rmap_item = list_entry(cur, struct rmap_item, link);
490                 cur = cur->next;
491                 remove_rmap_item_from_tree(rmap_item);
492                 list_del(&rmap_item->link);
493                 free_rmap_item(rmap_item);
494         }
495 }
496
497 /*
498  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
499  * than check every pte of a given vma, the locking doesn't quite work for
500  * that - an rmap_item is assigned to the stable tree after inserting ksm
501  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
502  * rmap_items from parent to child at fork time (so as not to waste time
503  * if exit comes before the next scan reaches it).
504  *
505  * Similarly, although we'd like to remove rmap_items (so updating counts
506  * and freeing memory) when unmerging an area, it's easier to leave that
507  * to the next pass of ksmd - consider, for example, how ksmd might be
508  * in cmp_and_merge_page on one of the rmap_items we would be removing.
509  */
510 static int unmerge_ksm_pages(struct vm_area_struct *vma,
511                              unsigned long start, unsigned long end)
512 {
513         unsigned long addr;
514         int err = 0;
515
516         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
517                 if (ksm_test_exit(vma->vm_mm))
518                         break;
519                 if (signal_pending(current))
520                         err = -ERESTARTSYS;
521                 else
522                         err = break_ksm(vma, addr);
523         }
524         return err;
525 }
526
527 #ifdef CONFIG_SYSFS
528 /*
529  * Only called through the sysfs control interface:
530  */
531 static int unmerge_and_remove_all_rmap_items(void)
532 {
533         struct mm_slot *mm_slot;
534         struct mm_struct *mm;
535         struct vm_area_struct *vma;
536         int err = 0;
537
538         spin_lock(&ksm_mmlist_lock);
539         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
540                                                 struct mm_slot, mm_list);
541         spin_unlock(&ksm_mmlist_lock);
542
543         for (mm_slot = ksm_scan.mm_slot;
544                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
545                 mm = mm_slot->mm;
546                 down_read(&mm->mmap_sem);
547                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
548                         if (ksm_test_exit(mm))
549                                 break;
550                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
551                                 continue;
552                         err = unmerge_ksm_pages(vma,
553                                                 vma->vm_start, vma->vm_end);
554                         if (err)
555                                 goto error;
556                 }
557
558                 remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
559
560                 spin_lock(&ksm_mmlist_lock);
561                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
562                                                 struct mm_slot, mm_list);
563                 if (ksm_test_exit(mm)) {
564                         hlist_del(&mm_slot->link);
565                         list_del(&mm_slot->mm_list);
566                         spin_unlock(&ksm_mmlist_lock);
567
568                         free_mm_slot(mm_slot);
569                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
570                         up_read(&mm->mmap_sem);
571                         mmdrop(mm);
572                 } else {
573                         spin_unlock(&ksm_mmlist_lock);
574                         up_read(&mm->mmap_sem);
575                 }
576         }
577
578         ksm_scan.seqnr = 0;
579         return 0;
580
581 error:
582         up_read(&mm->mmap_sem);
583         spin_lock(&ksm_mmlist_lock);
584         ksm_scan.mm_slot = &ksm_mm_head;
585         spin_unlock(&ksm_mmlist_lock);
586         return err;
587 }
588 #endif /* CONFIG_SYSFS */
589
590 static u32 calc_checksum(struct page *page)
591 {
592         u32 checksum;
593         void *addr = kmap_atomic(page, KM_USER0);
594         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
595         kunmap_atomic(addr, KM_USER0);
596         return checksum;
597 }
598
599 static int memcmp_pages(struct page *page1, struct page *page2)
600 {
601         char *addr1, *addr2;
602         int ret;
603
604         addr1 = kmap_atomic(page1, KM_USER0);
605         addr2 = kmap_atomic(page2, KM_USER1);
606         ret = memcmp(addr1, addr2, PAGE_SIZE);
607         kunmap_atomic(addr2, KM_USER1);
608         kunmap_atomic(addr1, KM_USER0);
609         return ret;
610 }
611
612 static inline int pages_identical(struct page *page1, struct page *page2)
613 {
614         return !memcmp_pages(page1, page2);
615 }
616
617 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
618                               pte_t *orig_pte)
619 {
620         struct mm_struct *mm = vma->vm_mm;
621         unsigned long addr;
622         pte_t *ptep;
623         spinlock_t *ptl;
624         int swapped;
625         int err = -EFAULT;
626
627         addr = page_address_in_vma(page, vma);
628         if (addr == -EFAULT)
629                 goto out;
630
631         ptep = page_check_address(page, mm, addr, &ptl, 0);
632         if (!ptep)
633                 goto out;
634
635         if (pte_write(*ptep)) {
636                 pte_t entry;
637
638                 swapped = PageSwapCache(page);
639                 flush_cache_page(vma, addr, page_to_pfn(page));
640                 /*
641                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
642                  * take any lock, therefore the check that we are going to make
643                  * with the pagecount against the mapcount is racey and
644                  * O_DIRECT can happen right after the check.
645                  * So we clear the pte and flush the tlb before the check
646                  * this assure us that no O_DIRECT can happen after the check
647                  * or in the middle of the check.
648                  */
649                 entry = ptep_clear_flush(vma, addr, ptep);
650                 /*
651                  * Check that no O_DIRECT or similar I/O is in progress on the
652                  * page
653                  */
654                 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
655                         set_pte_at_notify(mm, addr, ptep, entry);
656                         goto out_unlock;
657                 }
658                 entry = pte_wrprotect(entry);
659                 set_pte_at_notify(mm, addr, ptep, entry);
660         }
661         *orig_pte = *ptep;
662         err = 0;
663
664 out_unlock:
665         pte_unmap_unlock(ptep, ptl);
666 out:
667         return err;
668 }
669
670 /**
671  * replace_page - replace page in vma by new ksm page
672  * @vma:      vma that holds the pte pointing to oldpage
673  * @oldpage:  the page we are replacing by newpage
674  * @newpage:  the ksm page we replace oldpage by
675  * @orig_pte: the original value of the pte
676  *
677  * Returns 0 on success, -EFAULT on failure.
678  */
679 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
680                         struct page *newpage, pte_t orig_pte)
681 {
682         struct mm_struct *mm = vma->vm_mm;
683         pgd_t *pgd;
684         pud_t *pud;
685         pmd_t *pmd;
686         pte_t *ptep;
687         spinlock_t *ptl;
688         unsigned long addr;
689         pgprot_t prot;
690         int err = -EFAULT;
691
692         prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
693
694         addr = page_address_in_vma(oldpage, vma);
695         if (addr == -EFAULT)
696                 goto out;
697
698         pgd = pgd_offset(mm, addr);
699         if (!pgd_present(*pgd))
700                 goto out;
701
702         pud = pud_offset(pgd, addr);
703         if (!pud_present(*pud))
704                 goto out;
705
706         pmd = pmd_offset(pud, addr);
707         if (!pmd_present(*pmd))
708                 goto out;
709
710         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
711         if (!pte_same(*ptep, orig_pte)) {
712                 pte_unmap_unlock(ptep, ptl);
713                 goto out;
714         }
715
716         get_page(newpage);
717         page_add_ksm_rmap(newpage);
718
719         flush_cache_page(vma, addr, pte_pfn(*ptep));
720         ptep_clear_flush(vma, addr, ptep);
721         set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
722
723         page_remove_rmap(oldpage);
724         put_page(oldpage);
725
726         pte_unmap_unlock(ptep, ptl);
727         err = 0;
728 out:
729         return err;
730 }
731
732 /*
733  * try_to_merge_one_page - take two pages and merge them into one
734  * @vma: the vma that hold the pte pointing into oldpage
735  * @oldpage: the page that we want to replace with newpage
736  * @newpage: the page that we want to map instead of oldpage
737  *
738  * Note:
739  * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
740  * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
741  *
742  * This function returns 0 if the pages were merged, -EFAULT otherwise.
743  */
744 static int try_to_merge_one_page(struct vm_area_struct *vma,
745                                  struct page *oldpage,
746                                  struct page *newpage)
747 {
748         pte_t orig_pte = __pte(0);
749         int err = -EFAULT;
750
751         if (!(vma->vm_flags & VM_MERGEABLE))
752                 goto out;
753
754         if (!PageAnon(oldpage))
755                 goto out;
756
757         get_page(newpage);
758         get_page(oldpage);
759
760         /*
761          * We need the page lock to read a stable PageSwapCache in
762          * write_protect_page().  We use trylock_page() instead of
763          * lock_page() because we don't want to wait here - we
764          * prefer to continue scanning and merging different pages,
765          * then come back to this page when it is unlocked.
766          */
767         if (!trylock_page(oldpage))
768                 goto out_putpage;
769         /*
770          * If this anonymous page is mapped only here, its pte may need
771          * to be write-protected.  If it's mapped elsewhere, all of its
772          * ptes are necessarily already write-protected.  But in either
773          * case, we need to lock and check page_count is not raised.
774          */
775         if (write_protect_page(vma, oldpage, &orig_pte)) {
776                 unlock_page(oldpage);
777                 goto out_putpage;
778         }
779         unlock_page(oldpage);
780
781         if (pages_identical(oldpage, newpage))
782                 err = replace_page(vma, oldpage, newpage, orig_pte);
783
784 out_putpage:
785         put_page(oldpage);
786         put_page(newpage);
787 out:
788         return err;
789 }
790
791 /*
792  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
793  * but no new kernel page is allocated: kpage must already be a ksm page.
794  */
795 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
796                                       unsigned long addr1,
797                                       struct page *page1,
798                                       struct page *kpage)
799 {
800         struct vm_area_struct *vma;
801         int err = -EFAULT;
802
803         down_read(&mm1->mmap_sem);
804         if (ksm_test_exit(mm1))
805                 goto out;
806
807         vma = find_vma(mm1, addr1);
808         if (!vma || vma->vm_start > addr1)
809                 goto out;
810
811         err = try_to_merge_one_page(vma, page1, kpage);
812 out:
813         up_read(&mm1->mmap_sem);
814         return err;
815 }
816
817 /*
818  * try_to_merge_two_pages - take two identical pages and prepare them
819  * to be merged into one page.
820  *
821  * This function returns 0 if we successfully mapped two identical pages
822  * into one page, -EFAULT otherwise.
823  *
824  * Note that this function allocates a new kernel page: if one of the pages
825  * is already a ksm page, try_to_merge_with_ksm_page should be used.
826  */
827 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
828                                   struct page *page1, struct mm_struct *mm2,
829                                   unsigned long addr2, struct page *page2)
830 {
831         struct vm_area_struct *vma;
832         struct page *kpage;
833         int err = -EFAULT;
834
835         /*
836          * The number of nodes in the stable tree
837          * is the number of kernel pages that we hold.
838          */
839         if (ksm_max_kernel_pages &&
840             ksm_max_kernel_pages <= ksm_pages_shared)
841                 return err;
842
843         kpage = alloc_page(GFP_HIGHUSER);
844         if (!kpage)
845                 return err;
846
847         down_read(&mm1->mmap_sem);
848         if (ksm_test_exit(mm1)) {
849                 up_read(&mm1->mmap_sem);
850                 goto out;
851         }
852         vma = find_vma(mm1, addr1);
853         if (!vma || vma->vm_start > addr1) {
854                 up_read(&mm1->mmap_sem);
855                 goto out;
856         }
857
858         copy_user_highpage(kpage, page1, addr1, vma);
859         err = try_to_merge_one_page(vma, page1, kpage);
860         up_read(&mm1->mmap_sem);
861
862         if (!err) {
863                 err = try_to_merge_with_ksm_page(mm2, addr2, page2, kpage);
864                 /*
865                  * If that fails, we have a ksm page with only one pte
866                  * pointing to it: so break it.
867                  */
868                 if (err)
869                         break_cow(mm1, addr1);
870         }
871 out:
872         put_page(kpage);
873         return err;
874 }
875
876 /*
877  * stable_tree_search - search page inside the stable tree
878  * @page: the page that we are searching identical pages to.
879  * @page2: pointer into identical page that we are holding inside the stable
880  *         tree that we have found.
881  * @rmap_item: the reverse mapping item
882  *
883  * This function checks if there is a page inside the stable tree
884  * with identical content to the page that we are scanning right now.
885  *
886  * This function return rmap_item pointer to the identical item if found,
887  * NULL otherwise.
888  */
889 static struct rmap_item *stable_tree_search(struct page *page,
890                                             struct page **page2,
891                                             struct rmap_item *rmap_item)
892 {
893         struct rb_node *node = root_stable_tree.rb_node;
894
895         while (node) {
896                 struct rmap_item *tree_rmap_item, *next_rmap_item;
897                 int ret;
898
899                 tree_rmap_item = rb_entry(node, struct rmap_item, node);
900                 while (tree_rmap_item) {
901                         BUG_ON(!in_stable_tree(tree_rmap_item));
902                         cond_resched();
903                         page2[0] = get_ksm_page(tree_rmap_item);
904                         if (page2[0])
905                                 break;
906                         next_rmap_item = tree_rmap_item->next;
907                         remove_rmap_item_from_tree(tree_rmap_item);
908                         tree_rmap_item = next_rmap_item;
909                 }
910                 if (!tree_rmap_item)
911                         return NULL;
912
913                 ret = memcmp_pages(page, page2[0]);
914
915                 if (ret < 0) {
916                         put_page(page2[0]);
917                         node = node->rb_left;
918                 } else if (ret > 0) {
919                         put_page(page2[0]);
920                         node = node->rb_right;
921                 } else {
922                         return tree_rmap_item;
923                 }
924         }
925
926         return NULL;
927 }
928
929 /*
930  * stable_tree_insert - insert rmap_item pointing to new ksm page
931  * into the stable tree.
932  *
933  * @page: the page that we are searching identical page to inside the stable
934  *        tree.
935  * @rmap_item: pointer to the reverse mapping item.
936  *
937  * This function returns rmap_item if success, NULL otherwise.
938  */
939 static struct rmap_item *stable_tree_insert(struct page *page,
940                                             struct rmap_item *rmap_item)
941 {
942         struct rb_node **new = &root_stable_tree.rb_node;
943         struct rb_node *parent = NULL;
944
945         while (*new) {
946                 struct rmap_item *tree_rmap_item, *next_rmap_item;
947                 struct page *tree_page;
948                 int ret;
949
950                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
951                 while (tree_rmap_item) {
952                         BUG_ON(!in_stable_tree(tree_rmap_item));
953                         cond_resched();
954                         tree_page = get_ksm_page(tree_rmap_item);
955                         if (tree_page)
956                                 break;
957                         next_rmap_item = tree_rmap_item->next;
958                         remove_rmap_item_from_tree(tree_rmap_item);
959                         tree_rmap_item = next_rmap_item;
960                 }
961                 if (!tree_rmap_item)
962                         return NULL;
963
964                 ret = memcmp_pages(page, tree_page);
965                 put_page(tree_page);
966
967                 parent = *new;
968                 if (ret < 0)
969                         new = &parent->rb_left;
970                 else if (ret > 0)
971                         new = &parent->rb_right;
972                 else {
973                         /*
974                          * It is not a bug that stable_tree_search() didn't
975                          * find this node: because at that time our page was
976                          * not yet write-protected, so may have changed since.
977                          */
978                         return NULL;
979                 }
980         }
981
982         rmap_item->address |= NODE_FLAG | STABLE_FLAG;
983         rmap_item->next = NULL;
984         rb_link_node(&rmap_item->node, parent, new);
985         rb_insert_color(&rmap_item->node, &root_stable_tree);
986
987         ksm_pages_shared++;
988         return rmap_item;
989 }
990
991 /*
992  * unstable_tree_search_insert - search and insert items into the unstable tree.
993  *
994  * @page: the page that we are going to search for identical page or to insert
995  *        into the unstable tree
996  * @page2: pointer into identical page that was found inside the unstable tree
997  * @rmap_item: the reverse mapping item of page
998  *
999  * This function searches for a page in the unstable tree identical to the
1000  * page currently being scanned; and if no identical page is found in the
1001  * tree, we insert rmap_item as a new object into the unstable tree.
1002  *
1003  * This function returns pointer to rmap_item found to be identical
1004  * to the currently scanned page, NULL otherwise.
1005  *
1006  * This function does both searching and inserting, because they share
1007  * the same walking algorithm in an rbtree.
1008  */
1009 static struct rmap_item *unstable_tree_search_insert(struct page *page,
1010                                                 struct page **page2,
1011                                                 struct rmap_item *rmap_item)
1012 {
1013         struct rb_node **new = &root_unstable_tree.rb_node;
1014         struct rb_node *parent = NULL;
1015
1016         while (*new) {
1017                 struct rmap_item *tree_rmap_item;
1018                 int ret;
1019
1020                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1021                 page2[0] = get_mergeable_page(tree_rmap_item);
1022                 if (!page2[0])
1023                         return NULL;
1024
1025                 /*
1026                  * Don't substitute an unswappable ksm page
1027                  * just for one good swappable forked page.
1028                  */
1029                 if (page == page2[0]) {
1030                         put_page(page2[0]);
1031                         return NULL;
1032                 }
1033
1034                 ret = memcmp_pages(page, page2[0]);
1035
1036                 parent = *new;
1037                 if (ret < 0) {
1038                         put_page(page2[0]);
1039                         new = &parent->rb_left;
1040                 } else if (ret > 0) {
1041                         put_page(page2[0]);
1042                         new = &parent->rb_right;
1043                 } else {
1044                         return tree_rmap_item;
1045                 }
1046         }
1047
1048         rmap_item->address |= NODE_FLAG;
1049         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1050         rb_link_node(&rmap_item->node, parent, new);
1051         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1052
1053         ksm_pages_unshared++;
1054         return NULL;
1055 }
1056
1057 /*
1058  * stable_tree_append - add another rmap_item to the linked list of
1059  * rmap_items hanging off a given node of the stable tree, all sharing
1060  * the same ksm page.
1061  */
1062 static void stable_tree_append(struct rmap_item *rmap_item,
1063                                struct rmap_item *tree_rmap_item)
1064 {
1065         rmap_item->next = tree_rmap_item->next;
1066         rmap_item->prev = tree_rmap_item;
1067
1068         if (tree_rmap_item->next)
1069                 tree_rmap_item->next->prev = rmap_item;
1070
1071         tree_rmap_item->next = rmap_item;
1072         rmap_item->address |= STABLE_FLAG;
1073
1074         ksm_pages_sharing++;
1075 }
1076
1077 /*
1078  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1079  * if not, compare checksum to previous and if it's the same, see if page can
1080  * be inserted into the unstable tree, or merged with a page already there and
1081  * both transferred to the stable tree.
1082  *
1083  * @page: the page that we are searching identical page to.
1084  * @rmap_item: the reverse mapping into the virtual address of this page
1085  */
1086 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1087 {
1088         struct page *page2[1];
1089         struct rmap_item *tree_rmap_item;
1090         unsigned int checksum;
1091         int err;
1092
1093         if (in_stable_tree(rmap_item))
1094                 remove_rmap_item_from_tree(rmap_item);
1095
1096         /* We first start with searching the page inside the stable tree */
1097         tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1098         if (tree_rmap_item) {
1099                 if (page == page2[0])                   /* forked */
1100                         err = 0;
1101                 else
1102                         err = try_to_merge_with_ksm_page(rmap_item->mm,
1103                                                          rmap_item->address,
1104                                                          page, page2[0]);
1105                 put_page(page2[0]);
1106
1107                 if (!err) {
1108                         /*
1109                          * The page was successfully merged:
1110                          * add its rmap_item to the stable tree.
1111                          */
1112                         stable_tree_append(rmap_item, tree_rmap_item);
1113                 }
1114                 return;
1115         }
1116
1117         /*
1118          * A ksm page might have got here by fork, but its other
1119          * references have already been removed from the stable tree.
1120          * Or it might be left over from a break_ksm which failed
1121          * when the mem_cgroup had reached its limit: try again now.
1122          */
1123         if (PageKsm(page))
1124                 break_cow(rmap_item->mm, rmap_item->address);
1125
1126         /*
1127          * In case the hash value of the page was changed from the last time we
1128          * have calculated it, this page to be changed frequely, therefore we
1129          * don't want to insert it to the unstable tree, and we don't want to
1130          * waste our time to search if there is something identical to it there.
1131          */
1132         checksum = calc_checksum(page);
1133         if (rmap_item->oldchecksum != checksum) {
1134                 rmap_item->oldchecksum = checksum;
1135                 return;
1136         }
1137
1138         tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1139         if (tree_rmap_item) {
1140                 err = try_to_merge_two_pages(rmap_item->mm,
1141                                              rmap_item->address, page,
1142                                              tree_rmap_item->mm,
1143                                              tree_rmap_item->address, page2[0]);
1144                 /*
1145                  * As soon as we merge this page, we want to remove the
1146                  * rmap_item of the page we have merged with from the unstable
1147                  * tree, and insert it instead as new node in the stable tree.
1148                  */
1149                 if (!err) {
1150                         rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1151                         tree_rmap_item->address &= ~NODE_FLAG;
1152                         ksm_pages_unshared--;
1153
1154                         /*
1155                          * If we fail to insert the page into the stable tree,
1156                          * we will have 2 virtual addresses that are pointing
1157                          * to a ksm page left outside the stable tree,
1158                          * in which case we need to break_cow on both.
1159                          */
1160                         if (stable_tree_insert(page2[0], tree_rmap_item))
1161                                 stable_tree_append(rmap_item, tree_rmap_item);
1162                         else {
1163                                 break_cow(tree_rmap_item->mm,
1164                                                 tree_rmap_item->address);
1165                                 break_cow(rmap_item->mm, rmap_item->address);
1166                         }
1167                 }
1168
1169                 put_page(page2[0]);
1170         }
1171 }
1172
1173 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1174                                             struct list_head *cur,
1175                                             unsigned long addr)
1176 {
1177         struct rmap_item *rmap_item;
1178
1179         while (cur != &mm_slot->rmap_list) {
1180                 rmap_item = list_entry(cur, struct rmap_item, link);
1181                 if ((rmap_item->address & PAGE_MASK) == addr) {
1182                         if (!in_stable_tree(rmap_item))
1183                                 remove_rmap_item_from_tree(rmap_item);
1184                         return rmap_item;
1185                 }
1186                 if (rmap_item->address > addr)
1187                         break;
1188                 cur = cur->next;
1189                 remove_rmap_item_from_tree(rmap_item);
1190                 list_del(&rmap_item->link);
1191                 free_rmap_item(rmap_item);
1192         }
1193
1194         rmap_item = alloc_rmap_item();
1195         if (rmap_item) {
1196                 /* It has already been zeroed */
1197                 rmap_item->mm = mm_slot->mm;
1198                 rmap_item->address = addr;
1199                 list_add_tail(&rmap_item->link, cur);
1200         }
1201         return rmap_item;
1202 }
1203
1204 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1205 {
1206         struct mm_struct *mm;
1207         struct mm_slot *slot;
1208         struct vm_area_struct *vma;
1209         struct rmap_item *rmap_item;
1210
1211         if (list_empty(&ksm_mm_head.mm_list))
1212                 return NULL;
1213
1214         slot = ksm_scan.mm_slot;
1215         if (slot == &ksm_mm_head) {
1216                 root_unstable_tree = RB_ROOT;
1217
1218                 spin_lock(&ksm_mmlist_lock);
1219                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1220                 ksm_scan.mm_slot = slot;
1221                 spin_unlock(&ksm_mmlist_lock);
1222 next_mm:
1223                 ksm_scan.address = 0;
1224                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1225                                                 struct rmap_item, link);
1226         }
1227
1228         mm = slot->mm;
1229         down_read(&mm->mmap_sem);
1230         if (ksm_test_exit(mm))
1231                 vma = NULL;
1232         else
1233                 vma = find_vma(mm, ksm_scan.address);
1234
1235         for (; vma; vma = vma->vm_next) {
1236                 if (!(vma->vm_flags & VM_MERGEABLE))
1237                         continue;
1238                 if (ksm_scan.address < vma->vm_start)
1239                         ksm_scan.address = vma->vm_start;
1240                 if (!vma->anon_vma)
1241                         ksm_scan.address = vma->vm_end;
1242
1243                 while (ksm_scan.address < vma->vm_end) {
1244                         if (ksm_test_exit(mm))
1245                                 break;
1246                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1247                         if (*page && PageAnon(*page)) {
1248                                 flush_anon_page(vma, *page, ksm_scan.address);
1249                                 flush_dcache_page(*page);
1250                                 rmap_item = get_next_rmap_item(slot,
1251                                         ksm_scan.rmap_item->link.next,
1252                                         ksm_scan.address);
1253                                 if (rmap_item) {
1254                                         ksm_scan.rmap_item = rmap_item;
1255                                         ksm_scan.address += PAGE_SIZE;
1256                                 } else
1257                                         put_page(*page);
1258                                 up_read(&mm->mmap_sem);
1259                                 return rmap_item;
1260                         }
1261                         if (*page)
1262                                 put_page(*page);
1263                         ksm_scan.address += PAGE_SIZE;
1264                         cond_resched();
1265                 }
1266         }
1267
1268         if (ksm_test_exit(mm)) {
1269                 ksm_scan.address = 0;
1270                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1271                                                 struct rmap_item, link);
1272         }
1273         /*
1274          * Nuke all the rmap_items that are above this current rmap:
1275          * because there were no VM_MERGEABLE vmas with such addresses.
1276          */
1277         remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1278
1279         spin_lock(&ksm_mmlist_lock);
1280         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1281                                                 struct mm_slot, mm_list);
1282         if (ksm_scan.address == 0) {
1283                 /*
1284                  * We've completed a full scan of all vmas, holding mmap_sem
1285                  * throughout, and found no VM_MERGEABLE: so do the same as
1286                  * __ksm_exit does to remove this mm from all our lists now.
1287                  * This applies either when cleaning up after __ksm_exit
1288                  * (but beware: we can reach here even before __ksm_exit),
1289                  * or when all VM_MERGEABLE areas have been unmapped (and
1290                  * mmap_sem then protects against race with MADV_MERGEABLE).
1291                  */
1292                 hlist_del(&slot->link);
1293                 list_del(&slot->mm_list);
1294                 spin_unlock(&ksm_mmlist_lock);
1295
1296                 free_mm_slot(slot);
1297                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1298                 up_read(&mm->mmap_sem);
1299                 mmdrop(mm);
1300         } else {
1301                 spin_unlock(&ksm_mmlist_lock);
1302                 up_read(&mm->mmap_sem);
1303         }
1304
1305         /* Repeat until we've completed scanning the whole list */
1306         slot = ksm_scan.mm_slot;
1307         if (slot != &ksm_mm_head)
1308                 goto next_mm;
1309
1310         ksm_scan.seqnr++;
1311         return NULL;
1312 }
1313
1314 /**
1315  * ksm_do_scan  - the ksm scanner main worker function.
1316  * @scan_npages - number of pages we want to scan before we return.
1317  */
1318 static void ksm_do_scan(unsigned int scan_npages)
1319 {
1320         struct rmap_item *rmap_item;
1321         struct page *page;
1322
1323         while (scan_npages--) {
1324                 cond_resched();
1325                 rmap_item = scan_get_next_rmap_item(&page);
1326                 if (!rmap_item)
1327                         return;
1328                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1329                         cmp_and_merge_page(page, rmap_item);
1330                 else if (page_mapcount(page) == 1) {
1331                         /*
1332                          * Replace now-unshared ksm page by ordinary page.
1333                          */
1334                         break_cow(rmap_item->mm, rmap_item->address);
1335                         remove_rmap_item_from_tree(rmap_item);
1336                         rmap_item->oldchecksum = calc_checksum(page);
1337                 }
1338                 put_page(page);
1339         }
1340 }
1341
1342 static int ksmd_should_run(void)
1343 {
1344         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1345 }
1346
1347 static int ksm_scan_thread(void *nothing)
1348 {
1349         set_user_nice(current, 5);
1350
1351         while (!kthread_should_stop()) {
1352                 mutex_lock(&ksm_thread_mutex);
1353                 if (ksmd_should_run())
1354                         ksm_do_scan(ksm_thread_pages_to_scan);
1355                 mutex_unlock(&ksm_thread_mutex);
1356
1357                 if (ksmd_should_run()) {
1358                         schedule_timeout_interruptible(
1359                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1360                 } else {
1361                         wait_event_interruptible(ksm_thread_wait,
1362                                 ksmd_should_run() || kthread_should_stop());
1363                 }
1364         }
1365         return 0;
1366 }
1367
1368 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1369                 unsigned long end, int advice, unsigned long *vm_flags)
1370 {
1371         struct mm_struct *mm = vma->vm_mm;
1372         int err;
1373
1374         switch (advice) {
1375         case MADV_MERGEABLE:
1376                 /*
1377                  * Be somewhat over-protective for now!
1378                  */
1379                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1380                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1381                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1382                                  VM_MIXEDMAP  | VM_SAO))
1383                         return 0;               /* just ignore the advice */
1384
1385                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1386                         err = __ksm_enter(mm);
1387                         if (err)
1388                                 return err;
1389                 }
1390
1391                 *vm_flags |= VM_MERGEABLE;
1392                 break;
1393
1394         case MADV_UNMERGEABLE:
1395                 if (!(*vm_flags & VM_MERGEABLE))
1396                         return 0;               /* just ignore the advice */
1397
1398                 if (vma->anon_vma) {
1399                         err = unmerge_ksm_pages(vma, start, end);
1400                         if (err)
1401                                 return err;
1402                 }
1403
1404                 *vm_flags &= ~VM_MERGEABLE;
1405                 break;
1406         }
1407
1408         return 0;
1409 }
1410
1411 int __ksm_enter(struct mm_struct *mm)
1412 {
1413         struct mm_slot *mm_slot;
1414         int needs_wakeup;
1415
1416         mm_slot = alloc_mm_slot();
1417         if (!mm_slot)
1418                 return -ENOMEM;
1419
1420         /* Check ksm_run too?  Would need tighter locking */
1421         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1422
1423         spin_lock(&ksm_mmlist_lock);
1424         insert_to_mm_slots_hash(mm, mm_slot);
1425         /*
1426          * Insert just behind the scanning cursor, to let the area settle
1427          * down a little; when fork is followed by immediate exec, we don't
1428          * want ksmd to waste time setting up and tearing down an rmap_list.
1429          */
1430         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1431         spin_unlock(&ksm_mmlist_lock);
1432
1433         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1434         atomic_inc(&mm->mm_count);
1435
1436         if (needs_wakeup)
1437                 wake_up_interruptible(&ksm_thread_wait);
1438
1439         return 0;
1440 }
1441
1442 void __ksm_exit(struct mm_struct *mm)
1443 {
1444         struct mm_slot *mm_slot;
1445         int easy_to_free = 0;
1446
1447         /*
1448          * This process is exiting: if it's straightforward (as is the
1449          * case when ksmd was never running), free mm_slot immediately.
1450          * But if it's at the cursor or has rmap_items linked to it, use
1451          * mmap_sem to synchronize with any break_cows before pagetables
1452          * are freed, and leave the mm_slot on the list for ksmd to free.
1453          * Beware: ksm may already have noticed it exiting and freed the slot.
1454          */
1455
1456         spin_lock(&ksm_mmlist_lock);
1457         mm_slot = get_mm_slot(mm);
1458         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1459                 if (list_empty(&mm_slot->rmap_list)) {
1460                         hlist_del(&mm_slot->link);
1461                         list_del(&mm_slot->mm_list);
1462                         easy_to_free = 1;
1463                 } else {
1464                         list_move(&mm_slot->mm_list,
1465                                   &ksm_scan.mm_slot->mm_list);
1466                 }
1467         }
1468         spin_unlock(&ksm_mmlist_lock);
1469
1470         if (easy_to_free) {
1471                 free_mm_slot(mm_slot);
1472                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1473                 mmdrop(mm);
1474         } else if (mm_slot) {
1475                 down_write(&mm->mmap_sem);
1476                 up_write(&mm->mmap_sem);
1477         }
1478 }
1479
1480 #ifdef CONFIG_SYSFS
1481 /*
1482  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1483  */
1484
1485 #define KSM_ATTR_RO(_name) \
1486         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1487 #define KSM_ATTR(_name) \
1488         static struct kobj_attribute _name##_attr = \
1489                 __ATTR(_name, 0644, _name##_show, _name##_store)
1490
1491 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1492                                     struct kobj_attribute *attr, char *buf)
1493 {
1494         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1495 }
1496
1497 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1498                                      struct kobj_attribute *attr,
1499                                      const char *buf, size_t count)
1500 {
1501         unsigned long msecs;
1502         int err;
1503
1504         err = strict_strtoul(buf, 10, &msecs);
1505         if (err || msecs > UINT_MAX)
1506                 return -EINVAL;
1507
1508         ksm_thread_sleep_millisecs = msecs;
1509
1510         return count;
1511 }
1512 KSM_ATTR(sleep_millisecs);
1513
1514 static ssize_t pages_to_scan_show(struct kobject *kobj,
1515                                   struct kobj_attribute *attr, char *buf)
1516 {
1517         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1518 }
1519
1520 static ssize_t pages_to_scan_store(struct kobject *kobj,
1521                                    struct kobj_attribute *attr,
1522                                    const char *buf, size_t count)
1523 {
1524         int err;
1525         unsigned long nr_pages;
1526
1527         err = strict_strtoul(buf, 10, &nr_pages);
1528         if (err || nr_pages > UINT_MAX)
1529                 return -EINVAL;
1530
1531         ksm_thread_pages_to_scan = nr_pages;
1532
1533         return count;
1534 }
1535 KSM_ATTR(pages_to_scan);
1536
1537 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1538                         char *buf)
1539 {
1540         return sprintf(buf, "%u\n", ksm_run);
1541 }
1542
1543 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1544                          const char *buf, size_t count)
1545 {
1546         int err;
1547         unsigned long flags;
1548
1549         err = strict_strtoul(buf, 10, &flags);
1550         if (err || flags > UINT_MAX)
1551                 return -EINVAL;
1552         if (flags > KSM_RUN_UNMERGE)
1553                 return -EINVAL;
1554
1555         /*
1556          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1557          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1558          * breaking COW to free the unswappable pages_shared (but leaves
1559          * mm_slots on the list for when ksmd may be set running again).
1560          */
1561
1562         mutex_lock(&ksm_thread_mutex);
1563         if (ksm_run != flags) {
1564                 ksm_run = flags;
1565                 if (flags & KSM_RUN_UNMERGE) {
1566                         current->flags |= PF_OOM_ORIGIN;
1567                         err = unmerge_and_remove_all_rmap_items();
1568                         current->flags &= ~PF_OOM_ORIGIN;
1569                         if (err) {
1570                                 ksm_run = KSM_RUN_STOP;
1571                                 count = err;
1572                         }
1573                 }
1574         }
1575         mutex_unlock(&ksm_thread_mutex);
1576
1577         if (flags & KSM_RUN_MERGE)
1578                 wake_up_interruptible(&ksm_thread_wait);
1579
1580         return count;
1581 }
1582 KSM_ATTR(run);
1583
1584 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1585                                       struct kobj_attribute *attr,
1586                                       const char *buf, size_t count)
1587 {
1588         int err;
1589         unsigned long nr_pages;
1590
1591         err = strict_strtoul(buf, 10, &nr_pages);
1592         if (err)
1593                 return -EINVAL;
1594
1595         ksm_max_kernel_pages = nr_pages;
1596
1597         return count;
1598 }
1599
1600 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1601                                      struct kobj_attribute *attr, char *buf)
1602 {
1603         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1604 }
1605 KSM_ATTR(max_kernel_pages);
1606
1607 static ssize_t pages_shared_show(struct kobject *kobj,
1608                                  struct kobj_attribute *attr, char *buf)
1609 {
1610         return sprintf(buf, "%lu\n", ksm_pages_shared);
1611 }
1612 KSM_ATTR_RO(pages_shared);
1613
1614 static ssize_t pages_sharing_show(struct kobject *kobj,
1615                                   struct kobj_attribute *attr, char *buf)
1616 {
1617         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1618 }
1619 KSM_ATTR_RO(pages_sharing);
1620
1621 static ssize_t pages_unshared_show(struct kobject *kobj,
1622                                    struct kobj_attribute *attr, char *buf)
1623 {
1624         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1625 }
1626 KSM_ATTR_RO(pages_unshared);
1627
1628 static ssize_t pages_volatile_show(struct kobject *kobj,
1629                                    struct kobj_attribute *attr, char *buf)
1630 {
1631         long ksm_pages_volatile;
1632
1633         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1634                                 - ksm_pages_sharing - ksm_pages_unshared;
1635         /*
1636          * It was not worth any locking to calculate that statistic,
1637          * but it might therefore sometimes be negative: conceal that.
1638          */
1639         if (ksm_pages_volatile < 0)
1640                 ksm_pages_volatile = 0;
1641         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1642 }
1643 KSM_ATTR_RO(pages_volatile);
1644
1645 static ssize_t full_scans_show(struct kobject *kobj,
1646                                struct kobj_attribute *attr, char *buf)
1647 {
1648         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1649 }
1650 KSM_ATTR_RO(full_scans);
1651
1652 static struct attribute *ksm_attrs[] = {
1653         &sleep_millisecs_attr.attr,
1654         &pages_to_scan_attr.attr,
1655         &run_attr.attr,
1656         &max_kernel_pages_attr.attr,
1657         &pages_shared_attr.attr,
1658         &pages_sharing_attr.attr,
1659         &pages_unshared_attr.attr,
1660         &pages_volatile_attr.attr,
1661         &full_scans_attr.attr,
1662         NULL,
1663 };
1664
1665 static struct attribute_group ksm_attr_group = {
1666         .attrs = ksm_attrs,
1667         .name = "ksm",
1668 };
1669 #endif /* CONFIG_SYSFS */
1670
1671 static int __init ksm_init(void)
1672 {
1673         struct task_struct *ksm_thread;
1674         int err;
1675
1676         ksm_init_max_kernel_pages();
1677
1678         err = ksm_slab_init();
1679         if (err)
1680                 goto out;
1681
1682         err = mm_slots_hash_init();
1683         if (err)
1684                 goto out_free1;
1685
1686         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1687         if (IS_ERR(ksm_thread)) {
1688                 printk(KERN_ERR "ksm: creating kthread failed\n");
1689                 err = PTR_ERR(ksm_thread);
1690                 goto out_free2;
1691         }
1692
1693 #ifdef CONFIG_SYSFS
1694         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1695         if (err) {
1696                 printk(KERN_ERR "ksm: register sysfs failed\n");
1697                 kthread_stop(ksm_thread);
1698                 goto out_free2;
1699         }
1700 #endif /* CONFIG_SYSFS */
1701
1702         return 0;
1703
1704 out_free2:
1705         mm_slots_hash_free();
1706 out_free1:
1707         ksm_slab_free();
1708 out:
1709         return err;
1710 }
1711 module_init(ksm_init)