2 * Memory merging support.
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
7 * Copyright (C) 2008-2009 Red Hat, Inc.
14 * This work is licensed under the terms of the GNU GPL, version 2.
17 #include <linux/errno.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>
36 #include <asm/tlbflush.h>
39 * A few notes about the KSM scanning process,
40 * to make it easier to understand the data structures below:
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.
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.
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.
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.
61 * KSM solves this problem by several techniques:
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.)
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
86 struct hlist_node link;
87 struct list_head mm_list;
88 struct list_head rmap_list;
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)
99 * There is only the one ksm_scan instance of this cursor structure.
102 struct mm_slot *mm_slot;
103 unsigned long address;
104 struct rmap_item *rmap_item;
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
119 struct list_head link;
120 struct mm_struct *mm;
121 unsigned long address; /* + low bits used for flags below */
123 unsigned int oldchecksum; /* when unstable */
124 struct rmap_item *next; /* when stable */
127 struct rb_node node; /* when tree node */
128 struct rmap_item *prev; /* in stable list */
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 */
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;
140 #define MM_SLOTS_HASH_HEADS 1024
141 static struct hlist_head *mm_slots_hash;
143 static struct mm_slot ksm_mm_head = {
144 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
146 static struct ksm_scan ksm_scan = {
147 .mm_slot = &ksm_mm_head,
150 static struct kmem_cache *rmap_item_cache;
151 static struct kmem_cache *mm_slot_cache;
153 /* The number of nodes in the stable tree */
154 static unsigned long ksm_pages_shared;
156 /* The number of page slots additionally sharing those nodes */
157 static unsigned long ksm_pages_sharing;
159 /* The number of nodes in the unstable tree */
160 static unsigned long ksm_pages_unshared;
162 /* The number of rmap_items in use: to calculate pages_volatile */
163 static unsigned long ksm_rmap_items;
165 /* Limit on the number of unswappable pages used */
166 static unsigned long ksm_max_kernel_pages = 2000;
168 /* Number of pages ksmd should scan in one batch */
169 static unsigned int ksm_thread_pages_to_scan = 200;
171 /* Milliseconds ksmd should sleep between batches */
172 static unsigned int ksm_thread_sleep_millisecs = 20;
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_MERGE;
179 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
180 static DEFINE_MUTEX(ksm_thread_mutex);
181 static DEFINE_SPINLOCK(ksm_mmlist_lock);
183 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
184 sizeof(struct __struct), __alignof__(struct __struct),\
187 static int __init ksm_slab_init(void)
189 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
190 if (!rmap_item_cache)
193 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
200 kmem_cache_destroy(rmap_item_cache);
205 static void __init ksm_slab_free(void)
207 kmem_cache_destroy(mm_slot_cache);
208 kmem_cache_destroy(rmap_item_cache);
209 mm_slot_cache = NULL;
212 static inline struct rmap_item *alloc_rmap_item(void)
214 struct rmap_item *rmap_item;
216 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
222 static inline void free_rmap_item(struct rmap_item *rmap_item)
225 rmap_item->mm = NULL; /* debug safety */
226 kmem_cache_free(rmap_item_cache, rmap_item);
229 static inline struct mm_slot *alloc_mm_slot(void)
231 if (!mm_slot_cache) /* initialization failed */
233 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
236 static inline void free_mm_slot(struct mm_slot *mm_slot)
238 kmem_cache_free(mm_slot_cache, mm_slot);
241 static int __init mm_slots_hash_init(void)
243 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
250 static void __init mm_slots_hash_free(void)
252 kfree(mm_slots_hash);
255 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
257 struct mm_slot *mm_slot;
258 struct hlist_head *bucket;
259 struct hlist_node *node;
261 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
262 % MM_SLOTS_HASH_HEADS];
263 hlist_for_each_entry(mm_slot, node, bucket, link) {
264 if (mm == mm_slot->mm)
270 static void insert_to_mm_slots_hash(struct mm_struct *mm,
271 struct mm_slot *mm_slot)
273 struct hlist_head *bucket;
275 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
276 % MM_SLOTS_HASH_HEADS];
278 INIT_LIST_HEAD(&mm_slot->rmap_list);
279 hlist_add_head(&mm_slot->link, bucket);
282 static inline int in_stable_tree(struct rmap_item *rmap_item)
284 return rmap_item->address & STABLE_FLAG;
288 * We use break_ksm to break COW on a ksm page: it's a stripped down
290 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
293 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
294 * in case the application has unmapped and remapped mm,addr meanwhile.
295 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
296 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
298 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
305 page = follow_page(vma, addr, FOLL_GET);
309 ret = handle_mm_fault(vma->vm_mm, vma, addr,
312 ret = VM_FAULT_WRITE;
314 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
316 * We must loop because handle_mm_fault() may back out if there's
317 * any difficulty e.g. if pte accessed bit gets updated concurrently.
319 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
320 * COW has been broken, even if the vma does not permit VM_WRITE;
321 * but note that a concurrent fault might break PageKsm for us.
323 * VM_FAULT_SIGBUS could occur if we race with truncation of the
324 * backing file, which also invalidates anonymous pages: that's
325 * okay, that truncation will have unmapped the PageKsm for us.
327 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
328 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
329 * current task has TIF_MEMDIE set, and will be OOM killed on return
330 * to user; and ksmd, having no mm, would never be chosen for that.
332 * But if the mm is in a limited mem_cgroup, then the fault may fail
333 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
334 * even ksmd can fail in this way - though it's usually breaking ksm
335 * just to undo a merge it made a moment before, so unlikely to oom.
337 * That's a pity: we might therefore have more kernel pages allocated
338 * than we're counting as nodes in the stable tree; but ksm_do_scan
339 * will retry to break_cow on each pass, so should recover the page
340 * in due course. The important thing is to not let VM_MERGEABLE
341 * be cleared while any such pages might remain in the area.
343 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
346 static void break_cow(struct mm_struct *mm, unsigned long addr)
348 struct vm_area_struct *vma;
350 down_read(&mm->mmap_sem);
351 if (ksm_test_exit(mm))
353 vma = find_vma(mm, addr);
354 if (!vma || vma->vm_start > addr)
356 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
358 break_ksm(vma, addr);
360 up_read(&mm->mmap_sem);
363 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
365 struct mm_struct *mm = rmap_item->mm;
366 unsigned long addr = rmap_item->address;
367 struct vm_area_struct *vma;
370 down_read(&mm->mmap_sem);
371 if (ksm_test_exit(mm))
373 vma = find_vma(mm, addr);
374 if (!vma || vma->vm_start > addr)
376 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
379 page = follow_page(vma, addr, FOLL_GET);
382 if (PageAnon(page)) {
383 flush_anon_page(vma, page, addr);
384 flush_dcache_page(page);
389 up_read(&mm->mmap_sem);
394 * get_ksm_page: checks if the page at the virtual address in rmap_item
395 * is still PageKsm, in which case we can trust the content of the page,
396 * and it returns the gotten page; but NULL if the page has been zapped.
398 static struct page *get_ksm_page(struct rmap_item *rmap_item)
402 page = get_mergeable_page(rmap_item);
403 if (page && !PageKsm(page)) {
411 * Removing rmap_item from stable or unstable tree.
412 * This function will clean the information from the stable/unstable tree.
414 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
416 if (in_stable_tree(rmap_item)) {
417 struct rmap_item *next_item = rmap_item->next;
419 if (rmap_item->address & NODE_FLAG) {
421 rb_replace_node(&rmap_item->node,
424 next_item->address |= NODE_FLAG;
427 rb_erase(&rmap_item->node, &root_stable_tree);
431 struct rmap_item *prev_item = rmap_item->prev;
433 BUG_ON(prev_item->next != rmap_item);
434 prev_item->next = next_item;
436 BUG_ON(next_item->prev != rmap_item);
437 next_item->prev = rmap_item->prev;
442 rmap_item->next = NULL;
444 } else if (rmap_item->address & NODE_FLAG) {
447 * Usually ksmd can and must skip the rb_erase, because
448 * root_unstable_tree was already reset to RB_ROOT.
449 * But be careful when an mm is exiting: do the rb_erase
450 * if this rmap_item was inserted by this scan, rather
451 * than left over from before.
453 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
456 rb_erase(&rmap_item->node, &root_unstable_tree);
457 ksm_pages_unshared--;
460 rmap_item->address &= PAGE_MASK;
462 cond_resched(); /* we're called from many long loops */
465 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
466 struct list_head *cur)
468 struct rmap_item *rmap_item;
470 while (cur != &mm_slot->rmap_list) {
471 rmap_item = list_entry(cur, struct rmap_item, link);
473 remove_rmap_item_from_tree(rmap_item);
474 list_del(&rmap_item->link);
475 free_rmap_item(rmap_item);
480 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
481 * than check every pte of a given vma, the locking doesn't quite work for
482 * that - an rmap_item is assigned to the stable tree after inserting ksm
483 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
484 * rmap_items from parent to child at fork time (so as not to waste time
485 * if exit comes before the next scan reaches it).
487 * Similarly, although we'd like to remove rmap_items (so updating counts
488 * and freeing memory) when unmerging an area, it's easier to leave that
489 * to the next pass of ksmd - consider, for example, how ksmd might be
490 * in cmp_and_merge_page on one of the rmap_items we would be removing.
492 static int unmerge_ksm_pages(struct vm_area_struct *vma,
493 unsigned long start, unsigned long end)
498 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
499 if (ksm_test_exit(vma->vm_mm))
501 if (signal_pending(current))
504 err = break_ksm(vma, addr);
511 * Only called through the sysfs control interface:
513 static int unmerge_and_remove_all_rmap_items(void)
515 struct mm_slot *mm_slot;
516 struct mm_struct *mm;
517 struct vm_area_struct *vma;
520 spin_lock(&ksm_mmlist_lock);
521 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
522 struct mm_slot, mm_list);
523 spin_unlock(&ksm_mmlist_lock);
525 for (mm_slot = ksm_scan.mm_slot;
526 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
528 down_read(&mm->mmap_sem);
529 for (vma = mm->mmap; vma; vma = vma->vm_next) {
530 if (ksm_test_exit(mm))
532 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
534 err = unmerge_ksm_pages(vma,
535 vma->vm_start, vma->vm_end);
540 remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
542 spin_lock(&ksm_mmlist_lock);
543 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
544 struct mm_slot, mm_list);
545 if (ksm_test_exit(mm)) {
546 hlist_del(&mm_slot->link);
547 list_del(&mm_slot->mm_list);
548 spin_unlock(&ksm_mmlist_lock);
550 free_mm_slot(mm_slot);
551 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
552 up_read(&mm->mmap_sem);
555 spin_unlock(&ksm_mmlist_lock);
556 up_read(&mm->mmap_sem);
564 up_read(&mm->mmap_sem);
565 spin_lock(&ksm_mmlist_lock);
566 ksm_scan.mm_slot = &ksm_mm_head;
567 spin_unlock(&ksm_mmlist_lock);
570 #endif /* CONFIG_SYSFS */
572 static u32 calc_checksum(struct page *page)
575 void *addr = kmap_atomic(page, KM_USER0);
576 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
577 kunmap_atomic(addr, KM_USER0);
581 static int memcmp_pages(struct page *page1, struct page *page2)
586 addr1 = kmap_atomic(page1, KM_USER0);
587 addr2 = kmap_atomic(page2, KM_USER1);
588 ret = memcmp(addr1, addr2, PAGE_SIZE);
589 kunmap_atomic(addr2, KM_USER1);
590 kunmap_atomic(addr1, KM_USER0);
594 static inline int pages_identical(struct page *page1, struct page *page2)
596 return !memcmp_pages(page1, page2);
599 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
602 struct mm_struct *mm = vma->vm_mm;
609 addr = page_address_in_vma(page, vma);
613 ptep = page_check_address(page, mm, addr, &ptl, 0);
617 if (pte_write(*ptep)) {
620 swapped = PageSwapCache(page);
621 flush_cache_page(vma, addr, page_to_pfn(page));
623 * Ok this is tricky, when get_user_pages_fast() run it doesnt
624 * take any lock, therefore the check that we are going to make
625 * with the pagecount against the mapcount is racey and
626 * O_DIRECT can happen right after the check.
627 * So we clear the pte and flush the tlb before the check
628 * this assure us that no O_DIRECT can happen after the check
629 * or in the middle of the check.
631 entry = ptep_clear_flush(vma, addr, ptep);
633 * Check that no O_DIRECT or similar I/O is in progress on the
636 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
637 set_pte_at_notify(mm, addr, ptep, entry);
640 entry = pte_wrprotect(entry);
641 set_pte_at_notify(mm, addr, ptep, entry);
647 pte_unmap_unlock(ptep, ptl);
653 * replace_page - replace page in vma by new ksm page
654 * @vma: vma that holds the pte pointing to oldpage
655 * @oldpage: the page we are replacing by newpage
656 * @newpage: the ksm page we replace oldpage by
657 * @orig_pte: the original value of the pte
659 * Returns 0 on success, -EFAULT on failure.
661 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
662 struct page *newpage, pte_t orig_pte)
664 struct mm_struct *mm = vma->vm_mm;
674 prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
676 addr = page_address_in_vma(oldpage, vma);
680 pgd = pgd_offset(mm, addr);
681 if (!pgd_present(*pgd))
684 pud = pud_offset(pgd, addr);
685 if (!pud_present(*pud))
688 pmd = pmd_offset(pud, addr);
689 if (!pmd_present(*pmd))
692 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
693 if (!pte_same(*ptep, orig_pte)) {
694 pte_unmap_unlock(ptep, ptl);
699 page_add_ksm_rmap(newpage);
701 flush_cache_page(vma, addr, pte_pfn(*ptep));
702 ptep_clear_flush(vma, addr, ptep);
703 set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
705 page_remove_rmap(oldpage);
708 pte_unmap_unlock(ptep, ptl);
715 * try_to_merge_one_page - take two pages and merge them into one
716 * @vma: the vma that hold the pte pointing into oldpage
717 * @oldpage: the page that we want to replace with newpage
718 * @newpage: the page that we want to map instead of oldpage
721 * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
722 * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
724 * This function returns 0 if the pages were merged, -EFAULT otherwise.
726 static int try_to_merge_one_page(struct vm_area_struct *vma,
727 struct page *oldpage,
728 struct page *newpage)
730 pte_t orig_pte = __pte(0);
733 if (!(vma->vm_flags & VM_MERGEABLE))
736 if (!PageAnon(oldpage))
743 * We need the page lock to read a stable PageSwapCache in
744 * write_protect_page(). We use trylock_page() instead of
745 * lock_page() because we don't want to wait here - we
746 * prefer to continue scanning and merging different pages,
747 * then come back to this page when it is unlocked.
749 if (!trylock_page(oldpage))
752 * If this anonymous page is mapped only here, its pte may need
753 * to be write-protected. If it's mapped elsewhere, all of its
754 * ptes are necessarily already write-protected. But in either
755 * case, we need to lock and check page_count is not raised.
757 if (write_protect_page(vma, oldpage, &orig_pte)) {
758 unlock_page(oldpage);
761 unlock_page(oldpage);
763 if (pages_identical(oldpage, newpage))
764 err = replace_page(vma, oldpage, newpage, orig_pte);
774 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
775 * but no new kernel page is allocated: kpage must already be a ksm page.
777 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
782 struct vm_area_struct *vma;
785 down_read(&mm1->mmap_sem);
786 if (ksm_test_exit(mm1))
789 vma = find_vma(mm1, addr1);
790 if (!vma || vma->vm_start > addr1)
793 err = try_to_merge_one_page(vma, page1, kpage);
795 up_read(&mm1->mmap_sem);
800 * try_to_merge_two_pages - take two identical pages and prepare them
801 * to be merged into one page.
803 * This function returns 0 if we successfully mapped two identical pages
804 * into one page, -EFAULT otherwise.
806 * Note that this function allocates a new kernel page: if one of the pages
807 * is already a ksm page, try_to_merge_with_ksm_page should be used.
809 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
810 struct page *page1, struct mm_struct *mm2,
811 unsigned long addr2, struct page *page2)
813 struct vm_area_struct *vma;
818 * The number of nodes in the stable tree
819 * is the number of kernel pages that we hold.
821 if (ksm_max_kernel_pages &&
822 ksm_max_kernel_pages <= ksm_pages_shared)
825 kpage = alloc_page(GFP_HIGHUSER);
829 down_read(&mm1->mmap_sem);
830 if (ksm_test_exit(mm1)) {
831 up_read(&mm1->mmap_sem);
834 vma = find_vma(mm1, addr1);
835 if (!vma || vma->vm_start > addr1) {
836 up_read(&mm1->mmap_sem);
840 copy_user_highpage(kpage, page1, addr1, vma);
841 err = try_to_merge_one_page(vma, page1, kpage);
842 up_read(&mm1->mmap_sem);
845 err = try_to_merge_with_ksm_page(mm2, addr2, page2, kpage);
847 * If that fails, we have a ksm page with only one pte
848 * pointing to it: so break it.
851 break_cow(mm1, addr1);
859 * stable_tree_search - search page inside the stable tree
860 * @page: the page that we are searching identical pages to.
861 * @page2: pointer into identical page that we are holding inside the stable
862 * tree that we have found.
863 * @rmap_item: the reverse mapping item
865 * This function checks if there is a page inside the stable tree
866 * with identical content to the page that we are scanning right now.
868 * This function return rmap_item pointer to the identical item if found,
871 static struct rmap_item *stable_tree_search(struct page *page,
873 struct rmap_item *rmap_item)
875 struct rb_node *node = root_stable_tree.rb_node;
878 struct rmap_item *tree_rmap_item, *next_rmap_item;
881 tree_rmap_item = rb_entry(node, struct rmap_item, node);
882 while (tree_rmap_item) {
883 BUG_ON(!in_stable_tree(tree_rmap_item));
885 page2[0] = get_ksm_page(tree_rmap_item);
888 next_rmap_item = tree_rmap_item->next;
889 remove_rmap_item_from_tree(tree_rmap_item);
890 tree_rmap_item = next_rmap_item;
895 ret = memcmp_pages(page, page2[0]);
899 node = node->rb_left;
900 } else if (ret > 0) {
902 node = node->rb_right;
904 return tree_rmap_item;
912 * stable_tree_insert - insert rmap_item pointing to new ksm page
913 * into the stable tree.
915 * @page: the page that we are searching identical page to inside the stable
917 * @rmap_item: pointer to the reverse mapping item.
919 * This function returns rmap_item if success, NULL otherwise.
921 static struct rmap_item *stable_tree_insert(struct page *page,
922 struct rmap_item *rmap_item)
924 struct rb_node **new = &root_stable_tree.rb_node;
925 struct rb_node *parent = NULL;
928 struct rmap_item *tree_rmap_item, *next_rmap_item;
929 struct page *tree_page;
932 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
933 while (tree_rmap_item) {
934 BUG_ON(!in_stable_tree(tree_rmap_item));
936 tree_page = get_ksm_page(tree_rmap_item);
939 next_rmap_item = tree_rmap_item->next;
940 remove_rmap_item_from_tree(tree_rmap_item);
941 tree_rmap_item = next_rmap_item;
946 ret = memcmp_pages(page, tree_page);
951 new = &parent->rb_left;
953 new = &parent->rb_right;
956 * It is not a bug that stable_tree_search() didn't
957 * find this node: because at that time our page was
958 * not yet write-protected, so may have changed since.
964 rmap_item->address |= NODE_FLAG | STABLE_FLAG;
965 rmap_item->next = NULL;
966 rb_link_node(&rmap_item->node, parent, new);
967 rb_insert_color(&rmap_item->node, &root_stable_tree);
974 * unstable_tree_search_insert - search and insert items into the unstable tree.
976 * @page: the page that we are going to search for identical page or to insert
977 * into the unstable tree
978 * @page2: pointer into identical page that was found inside the unstable tree
979 * @rmap_item: the reverse mapping item of page
981 * This function searches for a page in the unstable tree identical to the
982 * page currently being scanned; and if no identical page is found in the
983 * tree, we insert rmap_item as a new object into the unstable tree.
985 * This function returns pointer to rmap_item found to be identical
986 * to the currently scanned page, NULL otherwise.
988 * This function does both searching and inserting, because they share
989 * the same walking algorithm in an rbtree.
991 static struct rmap_item *unstable_tree_search_insert(struct page *page,
993 struct rmap_item *rmap_item)
995 struct rb_node **new = &root_unstable_tree.rb_node;
996 struct rb_node *parent = NULL;
999 struct rmap_item *tree_rmap_item;
1002 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1003 page2[0] = get_mergeable_page(tree_rmap_item);
1008 * Don't substitute an unswappable ksm page
1009 * just for one good swappable forked page.
1011 if (page == page2[0]) {
1016 ret = memcmp_pages(page, page2[0]);
1021 new = &parent->rb_left;
1022 } else if (ret > 0) {
1024 new = &parent->rb_right;
1026 return tree_rmap_item;
1030 rmap_item->address |= NODE_FLAG;
1031 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1032 rb_link_node(&rmap_item->node, parent, new);
1033 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1035 ksm_pages_unshared++;
1040 * stable_tree_append - add another rmap_item to the linked list of
1041 * rmap_items hanging off a given node of the stable tree, all sharing
1042 * the same ksm page.
1044 static void stable_tree_append(struct rmap_item *rmap_item,
1045 struct rmap_item *tree_rmap_item)
1047 rmap_item->next = tree_rmap_item->next;
1048 rmap_item->prev = tree_rmap_item;
1050 if (tree_rmap_item->next)
1051 tree_rmap_item->next->prev = rmap_item;
1053 tree_rmap_item->next = rmap_item;
1054 rmap_item->address |= STABLE_FLAG;
1056 ksm_pages_sharing++;
1060 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1061 * if not, compare checksum to previous and if it's the same, see if page can
1062 * be inserted into the unstable tree, or merged with a page already there and
1063 * both transferred to the stable tree.
1065 * @page: the page that we are searching identical page to.
1066 * @rmap_item: the reverse mapping into the virtual address of this page
1068 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1070 struct page *page2[1];
1071 struct rmap_item *tree_rmap_item;
1072 unsigned int checksum;
1075 if (in_stable_tree(rmap_item))
1076 remove_rmap_item_from_tree(rmap_item);
1078 /* We first start with searching the page inside the stable tree */
1079 tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1080 if (tree_rmap_item) {
1081 if (page == page2[0]) /* forked */
1084 err = try_to_merge_with_ksm_page(rmap_item->mm,
1091 * The page was successfully merged:
1092 * add its rmap_item to the stable tree.
1094 stable_tree_append(rmap_item, tree_rmap_item);
1100 * A ksm page might have got here by fork, but its other
1101 * references have already been removed from the stable tree.
1102 * Or it might be left over from a break_ksm which failed
1103 * when the mem_cgroup had reached its limit: try again now.
1106 break_cow(rmap_item->mm, rmap_item->address);
1109 * In case the hash value of the page was changed from the last time we
1110 * have calculated it, this page to be changed frequely, therefore we
1111 * don't want to insert it to the unstable tree, and we don't want to
1112 * waste our time to search if there is something identical to it there.
1114 checksum = calc_checksum(page);
1115 if (rmap_item->oldchecksum != checksum) {
1116 rmap_item->oldchecksum = checksum;
1120 tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1121 if (tree_rmap_item) {
1122 err = try_to_merge_two_pages(rmap_item->mm,
1123 rmap_item->address, page,
1125 tree_rmap_item->address, page2[0]);
1127 * As soon as we merge this page, we want to remove the
1128 * rmap_item of the page we have merged with from the unstable
1129 * tree, and insert it instead as new node in the stable tree.
1132 rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1133 tree_rmap_item->address &= ~NODE_FLAG;
1134 ksm_pages_unshared--;
1137 * If we fail to insert the page into the stable tree,
1138 * we will have 2 virtual addresses that are pointing
1139 * to a ksm page left outside the stable tree,
1140 * in which case we need to break_cow on both.
1142 if (stable_tree_insert(page2[0], tree_rmap_item))
1143 stable_tree_append(rmap_item, tree_rmap_item);
1145 break_cow(tree_rmap_item->mm,
1146 tree_rmap_item->address);
1147 break_cow(rmap_item->mm, rmap_item->address);
1155 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1156 struct list_head *cur,
1159 struct rmap_item *rmap_item;
1161 while (cur != &mm_slot->rmap_list) {
1162 rmap_item = list_entry(cur, struct rmap_item, link);
1163 if ((rmap_item->address & PAGE_MASK) == addr) {
1164 if (!in_stable_tree(rmap_item))
1165 remove_rmap_item_from_tree(rmap_item);
1168 if (rmap_item->address > addr)
1171 remove_rmap_item_from_tree(rmap_item);
1172 list_del(&rmap_item->link);
1173 free_rmap_item(rmap_item);
1176 rmap_item = alloc_rmap_item();
1178 /* It has already been zeroed */
1179 rmap_item->mm = mm_slot->mm;
1180 rmap_item->address = addr;
1181 list_add_tail(&rmap_item->link, cur);
1186 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1188 struct mm_struct *mm;
1189 struct mm_slot *slot;
1190 struct vm_area_struct *vma;
1191 struct rmap_item *rmap_item;
1193 if (list_empty(&ksm_mm_head.mm_list))
1196 slot = ksm_scan.mm_slot;
1197 if (slot == &ksm_mm_head) {
1198 root_unstable_tree = RB_ROOT;
1200 spin_lock(&ksm_mmlist_lock);
1201 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1202 ksm_scan.mm_slot = slot;
1203 spin_unlock(&ksm_mmlist_lock);
1205 ksm_scan.address = 0;
1206 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1207 struct rmap_item, link);
1211 down_read(&mm->mmap_sem);
1212 if (ksm_test_exit(mm))
1215 vma = find_vma(mm, ksm_scan.address);
1217 for (; vma; vma = vma->vm_next) {
1218 if (!(vma->vm_flags & VM_MERGEABLE))
1220 if (ksm_scan.address < vma->vm_start)
1221 ksm_scan.address = vma->vm_start;
1223 ksm_scan.address = vma->vm_end;
1225 while (ksm_scan.address < vma->vm_end) {
1226 if (ksm_test_exit(mm))
1228 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1229 if (*page && PageAnon(*page)) {
1230 flush_anon_page(vma, *page, ksm_scan.address);
1231 flush_dcache_page(*page);
1232 rmap_item = get_next_rmap_item(slot,
1233 ksm_scan.rmap_item->link.next,
1236 ksm_scan.rmap_item = rmap_item;
1237 ksm_scan.address += PAGE_SIZE;
1240 up_read(&mm->mmap_sem);
1245 ksm_scan.address += PAGE_SIZE;
1250 if (ksm_test_exit(mm)) {
1251 ksm_scan.address = 0;
1252 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1253 struct rmap_item, link);
1256 * Nuke all the rmap_items that are above this current rmap:
1257 * because there were no VM_MERGEABLE vmas with such addresses.
1259 remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1261 spin_lock(&ksm_mmlist_lock);
1262 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1263 struct mm_slot, mm_list);
1264 if (ksm_scan.address == 0) {
1266 * We've completed a full scan of all vmas, holding mmap_sem
1267 * throughout, and found no VM_MERGEABLE: so do the same as
1268 * __ksm_exit does to remove this mm from all our lists now.
1269 * This applies either when cleaning up after __ksm_exit
1270 * (but beware: we can reach here even before __ksm_exit),
1271 * or when all VM_MERGEABLE areas have been unmapped (and
1272 * mmap_sem then protects against race with MADV_MERGEABLE).
1274 hlist_del(&slot->link);
1275 list_del(&slot->mm_list);
1276 spin_unlock(&ksm_mmlist_lock);
1279 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1280 up_read(&mm->mmap_sem);
1283 spin_unlock(&ksm_mmlist_lock);
1284 up_read(&mm->mmap_sem);
1287 /* Repeat until we've completed scanning the whole list */
1288 slot = ksm_scan.mm_slot;
1289 if (slot != &ksm_mm_head)
1297 * ksm_do_scan - the ksm scanner main worker function.
1298 * @scan_npages - number of pages we want to scan before we return.
1300 static void ksm_do_scan(unsigned int scan_npages)
1302 struct rmap_item *rmap_item;
1305 while (scan_npages--) {
1307 rmap_item = scan_get_next_rmap_item(&page);
1310 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1311 cmp_and_merge_page(page, rmap_item);
1312 else if (page_mapcount(page) == 1) {
1314 * Replace now-unshared ksm page by ordinary page.
1316 break_cow(rmap_item->mm, rmap_item->address);
1317 remove_rmap_item_from_tree(rmap_item);
1318 rmap_item->oldchecksum = calc_checksum(page);
1324 static int ksmd_should_run(void)
1326 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1329 static int ksm_scan_thread(void *nothing)
1331 set_user_nice(current, 5);
1333 while (!kthread_should_stop()) {
1334 mutex_lock(&ksm_thread_mutex);
1335 if (ksmd_should_run())
1336 ksm_do_scan(ksm_thread_pages_to_scan);
1337 mutex_unlock(&ksm_thread_mutex);
1339 if (ksmd_should_run()) {
1340 schedule_timeout_interruptible(
1341 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1343 wait_event_interruptible(ksm_thread_wait,
1344 ksmd_should_run() || kthread_should_stop());
1350 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1351 unsigned long end, int advice, unsigned long *vm_flags)
1353 struct mm_struct *mm = vma->vm_mm;
1357 case MADV_MERGEABLE:
1359 * Be somewhat over-protective for now!
1361 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1362 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1363 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1364 VM_MIXEDMAP | VM_SAO))
1365 return 0; /* just ignore the advice */
1367 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1368 err = __ksm_enter(mm);
1373 *vm_flags |= VM_MERGEABLE;
1376 case MADV_UNMERGEABLE:
1377 if (!(*vm_flags & VM_MERGEABLE))
1378 return 0; /* just ignore the advice */
1380 if (vma->anon_vma) {
1381 err = unmerge_ksm_pages(vma, start, end);
1386 *vm_flags &= ~VM_MERGEABLE;
1393 int __ksm_enter(struct mm_struct *mm)
1395 struct mm_slot *mm_slot;
1398 mm_slot = alloc_mm_slot();
1402 /* Check ksm_run too? Would need tighter locking */
1403 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1405 spin_lock(&ksm_mmlist_lock);
1406 insert_to_mm_slots_hash(mm, mm_slot);
1408 * Insert just behind the scanning cursor, to let the area settle
1409 * down a little; when fork is followed by immediate exec, we don't
1410 * want ksmd to waste time setting up and tearing down an rmap_list.
1412 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1413 spin_unlock(&ksm_mmlist_lock);
1415 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1416 atomic_inc(&mm->mm_count);
1419 wake_up_interruptible(&ksm_thread_wait);
1424 void __ksm_exit(struct mm_struct *mm)
1426 struct mm_slot *mm_slot;
1427 int easy_to_free = 0;
1430 * This process is exiting: if it's straightforward (as is the
1431 * case when ksmd was never running), free mm_slot immediately.
1432 * But if it's at the cursor or has rmap_items linked to it, use
1433 * mmap_sem to synchronize with any break_cows before pagetables
1434 * are freed, and leave the mm_slot on the list for ksmd to free.
1435 * Beware: ksm may already have noticed it exiting and freed the slot.
1438 spin_lock(&ksm_mmlist_lock);
1439 mm_slot = get_mm_slot(mm);
1440 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1441 if (list_empty(&mm_slot->rmap_list)) {
1442 hlist_del(&mm_slot->link);
1443 list_del(&mm_slot->mm_list);
1446 list_move(&mm_slot->mm_list,
1447 &ksm_scan.mm_slot->mm_list);
1450 spin_unlock(&ksm_mmlist_lock);
1453 free_mm_slot(mm_slot);
1454 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1456 } else if (mm_slot) {
1457 down_write(&mm->mmap_sem);
1458 up_write(&mm->mmap_sem);
1464 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1467 #define KSM_ATTR_RO(_name) \
1468 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1469 #define KSM_ATTR(_name) \
1470 static struct kobj_attribute _name##_attr = \
1471 __ATTR(_name, 0644, _name##_show, _name##_store)
1473 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1474 struct kobj_attribute *attr, char *buf)
1476 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1479 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1480 struct kobj_attribute *attr,
1481 const char *buf, size_t count)
1483 unsigned long msecs;
1486 err = strict_strtoul(buf, 10, &msecs);
1487 if (err || msecs > UINT_MAX)
1490 ksm_thread_sleep_millisecs = msecs;
1494 KSM_ATTR(sleep_millisecs);
1496 static ssize_t pages_to_scan_show(struct kobject *kobj,
1497 struct kobj_attribute *attr, char *buf)
1499 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1502 static ssize_t pages_to_scan_store(struct kobject *kobj,
1503 struct kobj_attribute *attr,
1504 const char *buf, size_t count)
1507 unsigned long nr_pages;
1509 err = strict_strtoul(buf, 10, &nr_pages);
1510 if (err || nr_pages > UINT_MAX)
1513 ksm_thread_pages_to_scan = nr_pages;
1517 KSM_ATTR(pages_to_scan);
1519 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1522 return sprintf(buf, "%u\n", ksm_run);
1525 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1526 const char *buf, size_t count)
1529 unsigned long flags;
1531 err = strict_strtoul(buf, 10, &flags);
1532 if (err || flags > UINT_MAX)
1534 if (flags > KSM_RUN_UNMERGE)
1538 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1539 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1540 * breaking COW to free the unswappable pages_shared (but leaves
1541 * mm_slots on the list for when ksmd may be set running again).
1544 mutex_lock(&ksm_thread_mutex);
1545 if (ksm_run != flags) {
1547 if (flags & KSM_RUN_UNMERGE) {
1548 err = unmerge_and_remove_all_rmap_items();
1550 ksm_run = KSM_RUN_STOP;
1555 mutex_unlock(&ksm_thread_mutex);
1557 if (flags & KSM_RUN_MERGE)
1558 wake_up_interruptible(&ksm_thread_wait);
1564 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1565 struct kobj_attribute *attr,
1566 const char *buf, size_t count)
1569 unsigned long nr_pages;
1571 err = strict_strtoul(buf, 10, &nr_pages);
1575 ksm_max_kernel_pages = nr_pages;
1580 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1581 struct kobj_attribute *attr, char *buf)
1583 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1585 KSM_ATTR(max_kernel_pages);
1587 static ssize_t pages_shared_show(struct kobject *kobj,
1588 struct kobj_attribute *attr, char *buf)
1590 return sprintf(buf, "%lu\n", ksm_pages_shared);
1592 KSM_ATTR_RO(pages_shared);
1594 static ssize_t pages_sharing_show(struct kobject *kobj,
1595 struct kobj_attribute *attr, char *buf)
1597 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1599 KSM_ATTR_RO(pages_sharing);
1601 static ssize_t pages_unshared_show(struct kobject *kobj,
1602 struct kobj_attribute *attr, char *buf)
1604 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1606 KSM_ATTR_RO(pages_unshared);
1608 static ssize_t pages_volatile_show(struct kobject *kobj,
1609 struct kobj_attribute *attr, char *buf)
1611 long ksm_pages_volatile;
1613 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1614 - ksm_pages_sharing - ksm_pages_unshared;
1616 * It was not worth any locking to calculate that statistic,
1617 * but it might therefore sometimes be negative: conceal that.
1619 if (ksm_pages_volatile < 0)
1620 ksm_pages_volatile = 0;
1621 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1623 KSM_ATTR_RO(pages_volatile);
1625 static ssize_t full_scans_show(struct kobject *kobj,
1626 struct kobj_attribute *attr, char *buf)
1628 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1630 KSM_ATTR_RO(full_scans);
1632 static struct attribute *ksm_attrs[] = {
1633 &sleep_millisecs_attr.attr,
1634 &pages_to_scan_attr.attr,
1636 &max_kernel_pages_attr.attr,
1637 &pages_shared_attr.attr,
1638 &pages_sharing_attr.attr,
1639 &pages_unshared_attr.attr,
1640 &pages_volatile_attr.attr,
1641 &full_scans_attr.attr,
1645 static struct attribute_group ksm_attr_group = {
1649 #endif /* CONFIG_SYSFS */
1651 static int __init ksm_init(void)
1653 struct task_struct *ksm_thread;
1656 err = ksm_slab_init();
1660 err = mm_slots_hash_init();
1664 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1665 if (IS_ERR(ksm_thread)) {
1666 printk(KERN_ERR "ksm: creating kthread failed\n");
1667 err = PTR_ERR(ksm_thread);
1672 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1674 printk(KERN_ERR "ksm: register sysfs failed\n");
1675 kthread_stop(ksm_thread);
1678 #endif /* CONFIG_SYSFS */
1683 mm_slots_hash_free();
1689 module_init(ksm_init)