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>
35 #include <asm/tlbflush.h>
38 * A few notes about the KSM scanning process,
39 * to make it easier to understand the data structures below:
41 * In order to reduce excessive scanning, KSM sorts the memory pages by their
42 * contents into a data structure that holds pointers to the pages' locations.
44 * Since the contents of the pages may change at any moment, KSM cannot just
45 * insert the pages into a normal sorted tree and expect it to find anything.
46 * Therefore KSM uses two data structures - the stable and the unstable tree.
48 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
49 * by their contents. Because each such page is write-protected, searching on
50 * this tree is fully assured to be working (except when pages are unmapped),
51 * and therefore this tree is called the stable tree.
53 * In addition to the stable tree, KSM uses a second data structure called the
54 * unstable tree: this tree holds pointers to pages which have been found to
55 * be "unchanged for a period of time". The unstable tree sorts these pages
56 * by their contents, but since they are not write-protected, KSM cannot rely
57 * upon the unstable tree to work correctly - the unstable tree is liable to
58 * be corrupted as its contents are modified, and so it is called unstable.
60 * KSM solves this problem by several techniques:
62 * 1) The unstable tree is flushed every time KSM completes scanning all
63 * memory areas, and then the tree is rebuilt again from the beginning.
64 * 2) KSM will only insert into the unstable tree, pages whose hash value
65 * has not changed since the previous scan of all memory areas.
66 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
67 * colors of the nodes and not on their contents, assuring that even when
68 * the tree gets "corrupted" it won't get out of balance, so scanning time
69 * remains the same (also, searching and inserting nodes in an rbtree uses
70 * the same algorithm, so we have no overhead when we flush and rebuild).
71 * 4) KSM never flushes the stable tree, which means that even if it were to
72 * take 10 attempts to find a page in the unstable tree, once it is found,
73 * it is secured in the stable tree. (When we scan a new page, we first
74 * compare it against the stable tree, and then against the unstable tree.)
78 * struct mm_slot - ksm information per mm that is being scanned
79 * @link: link to the mm_slots hash list
80 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
81 * @rmap_list: head for this mm_slot's list of rmap_items
82 * @mm: the mm that this information is valid for
85 struct hlist_node link;
86 struct list_head mm_list;
87 struct list_head rmap_list;
92 * struct ksm_scan - cursor for scanning
93 * @mm_slot: the current mm_slot we are scanning
94 * @address: the next address inside that to be scanned
95 * @rmap_item: the current rmap that we are scanning inside the rmap_list
96 * @seqnr: count of completed full scans (needed when removing unstable node)
98 * There is only the one ksm_scan instance of this cursor structure.
101 struct mm_slot *mm_slot;
102 unsigned long address;
103 struct rmap_item *rmap_item;
108 * struct rmap_item - reverse mapping item for virtual addresses
109 * @link: link into mm_slot's rmap_list (rmap_list is per mm)
110 * @mm: the memory structure this rmap_item is pointing into
111 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
112 * @oldchecksum: previous checksum of the page at that virtual address
113 * @node: rb_node of this rmap_item in either unstable or stable tree
114 * @next: next rmap_item hanging off the same node of the stable tree
115 * @prev: previous rmap_item hanging off the same node of the stable tree
118 struct list_head link;
119 struct mm_struct *mm;
120 unsigned long address; /* + low bits used for flags below */
122 unsigned int oldchecksum; /* when unstable */
123 struct rmap_item *next; /* when stable */
126 struct rb_node node; /* when tree node */
127 struct rmap_item *prev; /* in stable list */
131 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
132 #define NODE_FLAG 0x100 /* is a node of unstable or stable tree */
133 #define STABLE_FLAG 0x200 /* is a node or list item of stable tree */
135 /* The stable and unstable tree heads */
136 static struct rb_root root_stable_tree = RB_ROOT;
137 static struct rb_root root_unstable_tree = RB_ROOT;
139 #define MM_SLOTS_HASH_HEADS 1024
140 static struct hlist_head *mm_slots_hash;
142 static struct mm_slot ksm_mm_head = {
143 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
145 static struct ksm_scan ksm_scan = {
146 .mm_slot = &ksm_mm_head,
149 static struct kmem_cache *rmap_item_cache;
150 static struct kmem_cache *mm_slot_cache;
152 /* The number of nodes in the stable tree */
153 static unsigned long ksm_kernel_pages_allocated;
155 /* The number of page slots sharing those nodes */
156 static unsigned long ksm_pages_shared;
158 /* Limit on the number of unswappable pages used */
159 static unsigned long ksm_max_kernel_pages;
161 /* Number of pages ksmd should scan in one batch */
162 static unsigned int ksm_thread_pages_to_scan;
164 /* Milliseconds ksmd should sleep between batches */
165 static unsigned int ksm_thread_sleep_millisecs;
167 #define KSM_RUN_STOP 0
168 #define KSM_RUN_MERGE 1
169 #define KSM_RUN_UNMERGE 2
170 static unsigned int ksm_run;
172 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
173 static DEFINE_MUTEX(ksm_thread_mutex);
174 static DEFINE_SPINLOCK(ksm_mmlist_lock);
176 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
177 sizeof(struct __struct), __alignof__(struct __struct),\
180 static int __init ksm_slab_init(void)
182 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
183 if (!rmap_item_cache)
186 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
193 kmem_cache_destroy(rmap_item_cache);
198 static void __init ksm_slab_free(void)
200 kmem_cache_destroy(mm_slot_cache);
201 kmem_cache_destroy(rmap_item_cache);
202 mm_slot_cache = NULL;
205 static inline struct rmap_item *alloc_rmap_item(void)
207 return kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
210 static inline void free_rmap_item(struct rmap_item *rmap_item)
212 rmap_item->mm = NULL; /* debug safety */
213 kmem_cache_free(rmap_item_cache, rmap_item);
216 static inline struct mm_slot *alloc_mm_slot(void)
218 if (!mm_slot_cache) /* initialization failed */
220 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
223 static inline void free_mm_slot(struct mm_slot *mm_slot)
225 kmem_cache_free(mm_slot_cache, mm_slot);
228 static int __init mm_slots_hash_init(void)
230 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
237 static void __init mm_slots_hash_free(void)
239 kfree(mm_slots_hash);
242 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
244 struct mm_slot *mm_slot;
245 struct hlist_head *bucket;
246 struct hlist_node *node;
248 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
249 % MM_SLOTS_HASH_HEADS];
250 hlist_for_each_entry(mm_slot, node, bucket, link) {
251 if (mm == mm_slot->mm)
257 static void insert_to_mm_slots_hash(struct mm_struct *mm,
258 struct mm_slot *mm_slot)
260 struct hlist_head *bucket;
262 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
263 % MM_SLOTS_HASH_HEADS];
265 INIT_LIST_HEAD(&mm_slot->rmap_list);
266 hlist_add_head(&mm_slot->link, bucket);
269 static inline int in_stable_tree(struct rmap_item *rmap_item)
271 return rmap_item->address & STABLE_FLAG;
275 * We use break_ksm to break COW on a ksm page: it's a stripped down
277 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
280 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
281 * in case the application has unmapped and remapped mm,addr meanwhile.
282 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
283 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
285 static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
292 page = follow_page(vma, addr, FOLL_GET);
296 ret = handle_mm_fault(vma->vm_mm, vma, addr,
299 ret = VM_FAULT_WRITE;
301 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS)));
303 /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
306 static void __break_cow(struct mm_struct *mm, unsigned long addr)
308 struct vm_area_struct *vma;
310 vma = find_vma(mm, addr);
311 if (!vma || vma->vm_start > addr)
313 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
315 break_ksm(vma, addr);
318 static void break_cow(struct mm_struct *mm, unsigned long addr)
320 down_read(&mm->mmap_sem);
321 __break_cow(mm, addr);
322 up_read(&mm->mmap_sem);
325 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
327 struct mm_struct *mm = rmap_item->mm;
328 unsigned long addr = rmap_item->address;
329 struct vm_area_struct *vma;
332 down_read(&mm->mmap_sem);
333 vma = find_vma(mm, addr);
334 if (!vma || vma->vm_start > addr)
336 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
339 page = follow_page(vma, addr, FOLL_GET);
342 if (PageAnon(page)) {
343 flush_anon_page(vma, page, addr);
344 flush_dcache_page(page);
349 up_read(&mm->mmap_sem);
354 * get_ksm_page: checks if the page at the virtual address in rmap_item
355 * is still PageKsm, in which case we can trust the content of the page,
356 * and it returns the gotten page; but NULL if the page has been zapped.
358 static struct page *get_ksm_page(struct rmap_item *rmap_item)
362 page = get_mergeable_page(rmap_item);
363 if (page && !PageKsm(page)) {
371 * Removing rmap_item from stable or unstable tree.
372 * This function will clean the information from the stable/unstable tree.
374 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
376 if (in_stable_tree(rmap_item)) {
377 struct rmap_item *next_item = rmap_item->next;
379 if (rmap_item->address & NODE_FLAG) {
381 rb_replace_node(&rmap_item->node,
384 next_item->address |= NODE_FLAG;
386 rb_erase(&rmap_item->node, &root_stable_tree);
387 ksm_kernel_pages_allocated--;
390 struct rmap_item *prev_item = rmap_item->prev;
392 BUG_ON(prev_item->next != rmap_item);
393 prev_item->next = next_item;
395 BUG_ON(next_item->prev != rmap_item);
396 next_item->prev = rmap_item->prev;
400 rmap_item->next = NULL;
403 } else if (rmap_item->address & NODE_FLAG) {
406 * ksm_thread can and must skip the rb_erase, because
407 * root_unstable_tree was already reset to RB_ROOT.
408 * But __ksm_exit has to be careful: do the rb_erase
409 * if it's interrupting a scan, and this rmap_item was
410 * inserted by this scan rather than left from before.
412 * Because of the case in which remove_mm_from_lists
413 * increments seqnr before removing rmaps, unstable_nr
414 * may even be 2 behind seqnr, but should never be
415 * further behind. Yes, I did have trouble with this!
417 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
420 rb_erase(&rmap_item->node, &root_unstable_tree);
423 rmap_item->address &= PAGE_MASK;
425 cond_resched(); /* we're called from many long loops */
428 static void remove_all_slot_rmap_items(struct mm_slot *mm_slot)
430 struct rmap_item *rmap_item, *node;
432 list_for_each_entry_safe(rmap_item, node, &mm_slot->rmap_list, link) {
433 remove_rmap_item_from_tree(rmap_item);
434 list_del(&rmap_item->link);
435 free_rmap_item(rmap_item);
439 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
440 struct list_head *cur)
442 struct rmap_item *rmap_item;
444 while (cur != &mm_slot->rmap_list) {
445 rmap_item = list_entry(cur, struct rmap_item, link);
447 remove_rmap_item_from_tree(rmap_item);
448 list_del(&rmap_item->link);
449 free_rmap_item(rmap_item);
454 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
455 * than check every pte of a given vma, the locking doesn't quite work for
456 * that - an rmap_item is assigned to the stable tree after inserting ksm
457 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
458 * rmap_items from parent to child at fork time (so as not to waste time
459 * if exit comes before the next scan reaches it).
461 static void unmerge_ksm_pages(struct vm_area_struct *vma,
462 unsigned long start, unsigned long end)
466 for (addr = start; addr < end; addr += PAGE_SIZE)
467 break_ksm(vma, addr);
470 static void unmerge_and_remove_all_rmap_items(void)
472 struct mm_slot *mm_slot;
473 struct mm_struct *mm;
474 struct vm_area_struct *vma;
476 list_for_each_entry(mm_slot, &ksm_mm_head.mm_list, mm_list) {
478 down_read(&mm->mmap_sem);
479 for (vma = mm->mmap; vma; vma = vma->vm_next) {
480 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
482 unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end);
484 remove_all_slot_rmap_items(mm_slot);
485 up_read(&mm->mmap_sem);
488 spin_lock(&ksm_mmlist_lock);
489 if (ksm_scan.mm_slot != &ksm_mm_head) {
490 ksm_scan.mm_slot = &ksm_mm_head;
493 spin_unlock(&ksm_mmlist_lock);
496 static void remove_mm_from_lists(struct mm_struct *mm)
498 struct mm_slot *mm_slot;
500 spin_lock(&ksm_mmlist_lock);
501 mm_slot = get_mm_slot(mm);
504 * This mm_slot is always at the scanning cursor when we're
505 * called from scan_get_next_rmap_item; but it's a special
506 * case when we're called from __ksm_exit.
508 if (ksm_scan.mm_slot == mm_slot) {
509 ksm_scan.mm_slot = list_entry(
510 mm_slot->mm_list.next, struct mm_slot, mm_list);
511 ksm_scan.address = 0;
512 ksm_scan.rmap_item = list_entry(
513 &ksm_scan.mm_slot->rmap_list, struct rmap_item, link);
514 if (ksm_scan.mm_slot == &ksm_mm_head)
518 hlist_del(&mm_slot->link);
519 list_del(&mm_slot->mm_list);
520 spin_unlock(&ksm_mmlist_lock);
522 remove_all_slot_rmap_items(mm_slot);
523 free_mm_slot(mm_slot);
524 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
527 static u32 calc_checksum(struct page *page)
530 void *addr = kmap_atomic(page, KM_USER0);
531 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
532 kunmap_atomic(addr, KM_USER0);
536 static int memcmp_pages(struct page *page1, struct page *page2)
541 addr1 = kmap_atomic(page1, KM_USER0);
542 addr2 = kmap_atomic(page2, KM_USER1);
543 ret = memcmp(addr1, addr2, PAGE_SIZE);
544 kunmap_atomic(addr2, KM_USER1);
545 kunmap_atomic(addr1, KM_USER0);
549 static inline int pages_identical(struct page *page1, struct page *page2)
551 return !memcmp_pages(page1, page2);
554 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
557 struct mm_struct *mm = vma->vm_mm;
564 addr = page_address_in_vma(page, vma);
568 ptep = page_check_address(page, mm, addr, &ptl, 0);
572 if (pte_write(*ptep)) {
575 swapped = PageSwapCache(page);
576 flush_cache_page(vma, addr, page_to_pfn(page));
578 * Ok this is tricky, when get_user_pages_fast() run it doesnt
579 * take any lock, therefore the check that we are going to make
580 * with the pagecount against the mapcount is racey and
581 * O_DIRECT can happen right after the check.
582 * So we clear the pte and flush the tlb before the check
583 * this assure us that no O_DIRECT can happen after the check
584 * or in the middle of the check.
586 entry = ptep_clear_flush(vma, addr, ptep);
588 * Check that no O_DIRECT or similar I/O is in progress on the
591 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
592 set_pte_at_notify(mm, addr, ptep, entry);
595 entry = pte_wrprotect(entry);
596 set_pte_at_notify(mm, addr, ptep, entry);
602 pte_unmap_unlock(ptep, ptl);
608 * replace_page - replace page in vma by new ksm page
609 * @vma: vma that holds the pte pointing to oldpage
610 * @oldpage: the page we are replacing by newpage
611 * @newpage: the ksm page we replace oldpage by
612 * @orig_pte: the original value of the pte
614 * Returns 0 on success, -EFAULT on failure.
616 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
617 struct page *newpage, pte_t orig_pte)
619 struct mm_struct *mm = vma->vm_mm;
629 prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
631 addr = page_address_in_vma(oldpage, vma);
635 pgd = pgd_offset(mm, addr);
636 if (!pgd_present(*pgd))
639 pud = pud_offset(pgd, addr);
640 if (!pud_present(*pud))
643 pmd = pmd_offset(pud, addr);
644 if (!pmd_present(*pmd))
647 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
648 if (!pte_same(*ptep, orig_pte)) {
649 pte_unmap_unlock(ptep, ptl);
654 page_add_ksm_rmap(newpage);
656 flush_cache_page(vma, addr, pte_pfn(*ptep));
657 ptep_clear_flush(vma, addr, ptep);
658 set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
660 page_remove_rmap(oldpage);
663 pte_unmap_unlock(ptep, ptl);
670 * try_to_merge_one_page - take two pages and merge them into one
671 * @vma: the vma that hold the pte pointing into oldpage
672 * @oldpage: the page that we want to replace with newpage
673 * @newpage: the page that we want to map instead of oldpage
676 * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
677 * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
679 * This function returns 0 if the pages were merged, -EFAULT otherwise.
681 static int try_to_merge_one_page(struct vm_area_struct *vma,
682 struct page *oldpage,
683 struct page *newpage)
685 pte_t orig_pte = __pte(0);
688 if (!(vma->vm_flags & VM_MERGEABLE))
691 if (!PageAnon(oldpage))
698 * We need the page lock to read a stable PageSwapCache in
699 * write_protect_page(). We use trylock_page() instead of
700 * lock_page() because we don't want to wait here - we
701 * prefer to continue scanning and merging different pages,
702 * then come back to this page when it is unlocked.
704 if (!trylock_page(oldpage))
707 * If this anonymous page is mapped only here, its pte may need
708 * to be write-protected. If it's mapped elsewhere, all of its
709 * ptes are necessarily already write-protected. But in either
710 * case, we need to lock and check page_count is not raised.
712 if (write_protect_page(vma, oldpage, &orig_pte)) {
713 unlock_page(oldpage);
716 unlock_page(oldpage);
718 if (pages_identical(oldpage, newpage))
719 err = replace_page(vma, oldpage, newpage, orig_pte);
729 * try_to_merge_two_pages - take two identical pages and prepare them
730 * to be merged into one page.
732 * This function returns 0 if we successfully mapped two identical pages
733 * into one page, -EFAULT otherwise.
735 * Note that this function allocates a new kernel page: if one of the pages
736 * is already a ksm page, try_to_merge_with_ksm_page should be used.
738 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
739 struct page *page1, struct mm_struct *mm2,
740 unsigned long addr2, struct page *page2)
742 struct vm_area_struct *vma;
747 * The number of nodes in the stable tree
748 * is the number of kernel pages that we hold.
750 if (ksm_max_kernel_pages &&
751 ksm_max_kernel_pages <= ksm_kernel_pages_allocated)
754 kpage = alloc_page(GFP_HIGHUSER);
758 down_read(&mm1->mmap_sem);
759 vma = find_vma(mm1, addr1);
760 if (!vma || vma->vm_start > addr1) {
762 up_read(&mm1->mmap_sem);
766 copy_user_highpage(kpage, page1, addr1, vma);
767 err = try_to_merge_one_page(vma, page1, kpage);
768 up_read(&mm1->mmap_sem);
771 down_read(&mm2->mmap_sem);
772 vma = find_vma(mm2, addr2);
773 if (!vma || vma->vm_start > addr2) {
775 up_read(&mm2->mmap_sem);
776 break_cow(mm1, addr1);
780 err = try_to_merge_one_page(vma, page2, kpage);
781 up_read(&mm2->mmap_sem);
784 * If the second try_to_merge_one_page failed, we have a
785 * ksm page with just one pte pointing to it, so break it.
788 break_cow(mm1, addr1);
790 ksm_pages_shared += 2;
798 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
799 * but no new kernel page is allocated: kpage must already be a ksm page.
801 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
806 struct vm_area_struct *vma;
809 down_read(&mm1->mmap_sem);
810 vma = find_vma(mm1, addr1);
811 if (!vma || vma->vm_start > addr1) {
812 up_read(&mm1->mmap_sem);
816 err = try_to_merge_one_page(vma, page1, kpage);
817 up_read(&mm1->mmap_sem);
826 * stable_tree_search - search page inside the stable tree
827 * @page: the page that we are searching identical pages to.
828 * @page2: pointer into identical page that we are holding inside the stable
829 * tree that we have found.
830 * @rmap_item: the reverse mapping item
832 * This function checks if there is a page inside the stable tree
833 * with identical content to the page that we are scanning right now.
835 * This function return rmap_item pointer to the identical item if found,
838 static struct rmap_item *stable_tree_search(struct page *page,
840 struct rmap_item *rmap_item)
842 struct rb_node *node = root_stable_tree.rb_node;
845 struct rmap_item *tree_rmap_item, *next_rmap_item;
848 tree_rmap_item = rb_entry(node, struct rmap_item, node);
849 while (tree_rmap_item) {
850 BUG_ON(!in_stable_tree(tree_rmap_item));
852 page2[0] = get_ksm_page(tree_rmap_item);
855 next_rmap_item = tree_rmap_item->next;
856 remove_rmap_item_from_tree(tree_rmap_item);
857 tree_rmap_item = next_rmap_item;
862 ret = memcmp_pages(page, page2[0]);
866 node = node->rb_left;
867 } else if (ret > 0) {
869 node = node->rb_right;
871 return tree_rmap_item;
879 * stable_tree_insert - insert rmap_item pointing to new ksm page
880 * into the stable tree.
882 * @page: the page that we are searching identical page to inside the stable
884 * @rmap_item: pointer to the reverse mapping item.
886 * This function returns rmap_item if success, NULL otherwise.
888 static struct rmap_item *stable_tree_insert(struct page *page,
889 struct rmap_item *rmap_item)
891 struct rb_node **new = &root_stable_tree.rb_node;
892 struct rb_node *parent = NULL;
895 struct rmap_item *tree_rmap_item, *next_rmap_item;
896 struct page *tree_page;
899 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
900 while (tree_rmap_item) {
901 BUG_ON(!in_stable_tree(tree_rmap_item));
903 tree_page = get_ksm_page(tree_rmap_item);
906 next_rmap_item = tree_rmap_item->next;
907 remove_rmap_item_from_tree(tree_rmap_item);
908 tree_rmap_item = next_rmap_item;
913 ret = memcmp_pages(page, tree_page);
918 new = &parent->rb_left;
920 new = &parent->rb_right;
923 * It is not a bug that stable_tree_search() didn't
924 * find this node: because at that time our page was
925 * not yet write-protected, so may have changed since.
931 ksm_kernel_pages_allocated++;
933 rmap_item->address |= NODE_FLAG | STABLE_FLAG;
934 rmap_item->next = NULL;
935 rb_link_node(&rmap_item->node, parent, new);
936 rb_insert_color(&rmap_item->node, &root_stable_tree);
942 * unstable_tree_search_insert - search and insert items into the unstable tree.
944 * @page: the page that we are going to search for identical page or to insert
945 * into the unstable tree
946 * @page2: pointer into identical page that was found inside the unstable tree
947 * @rmap_item: the reverse mapping item of page
949 * This function searches for a page in the unstable tree identical to the
950 * page currently being scanned; and if no identical page is found in the
951 * tree, we insert rmap_item as a new object into the unstable tree.
953 * This function returns pointer to rmap_item found to be identical
954 * to the currently scanned page, NULL otherwise.
956 * This function does both searching and inserting, because they share
957 * the same walking algorithm in an rbtree.
959 static struct rmap_item *unstable_tree_search_insert(struct page *page,
961 struct rmap_item *rmap_item)
963 struct rb_node **new = &root_unstable_tree.rb_node;
964 struct rb_node *parent = NULL;
967 struct rmap_item *tree_rmap_item;
970 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
971 page2[0] = get_mergeable_page(tree_rmap_item);
976 * Don't substitute an unswappable ksm page
977 * just for one good swappable forked page.
979 if (page == page2[0]) {
984 ret = memcmp_pages(page, page2[0]);
989 new = &parent->rb_left;
990 } else if (ret > 0) {
992 new = &parent->rb_right;
994 return tree_rmap_item;
998 rmap_item->address |= NODE_FLAG;
999 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1000 rb_link_node(&rmap_item->node, parent, new);
1001 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1007 * stable_tree_append - add another rmap_item to the linked list of
1008 * rmap_items hanging off a given node of the stable tree, all sharing
1009 * the same ksm page.
1011 static void stable_tree_append(struct rmap_item *rmap_item,
1012 struct rmap_item *tree_rmap_item)
1014 rmap_item->next = tree_rmap_item->next;
1015 rmap_item->prev = tree_rmap_item;
1017 if (tree_rmap_item->next)
1018 tree_rmap_item->next->prev = rmap_item;
1020 tree_rmap_item->next = rmap_item;
1021 rmap_item->address |= STABLE_FLAG;
1025 * cmp_and_merge_page - take a page computes its hash value and check if there
1026 * is similar hash value to different page,
1027 * in case we find that there is similar hash to different page we call to
1028 * try_to_merge_two_pages().
1030 * @page: the page that we are searching identical page to.
1031 * @rmap_item: the reverse mapping into the virtual address of this page
1033 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1035 struct page *page2[1];
1036 struct rmap_item *tree_rmap_item;
1037 unsigned int checksum;
1040 if (in_stable_tree(rmap_item))
1041 remove_rmap_item_from_tree(rmap_item);
1043 /* We first start with searching the page inside the stable tree */
1044 tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1045 if (tree_rmap_item) {
1046 if (page == page2[0]) { /* forked */
1050 err = try_to_merge_with_ksm_page(rmap_item->mm,
1057 * The page was successfully merged:
1058 * add its rmap_item to the stable tree.
1060 stable_tree_append(rmap_item, tree_rmap_item);
1066 * A ksm page might have got here by fork, but its other
1067 * references have already been removed from the stable tree.
1070 break_cow(rmap_item->mm, rmap_item->address);
1073 * In case the hash value of the page was changed from the last time we
1074 * have calculated it, this page to be changed frequely, therefore we
1075 * don't want to insert it to the unstable tree, and we don't want to
1076 * waste our time to search if there is something identical to it there.
1078 checksum = calc_checksum(page);
1079 if (rmap_item->oldchecksum != checksum) {
1080 rmap_item->oldchecksum = checksum;
1084 tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1085 if (tree_rmap_item) {
1086 err = try_to_merge_two_pages(rmap_item->mm,
1087 rmap_item->address, page,
1089 tree_rmap_item->address, page2[0]);
1091 * As soon as we merge this page, we want to remove the
1092 * rmap_item of the page we have merged with from the unstable
1093 * tree, and insert it instead as new node in the stable tree.
1096 rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1097 tree_rmap_item->address &= ~NODE_FLAG;
1099 * If we fail to insert the page into the stable tree,
1100 * we will have 2 virtual addresses that are pointing
1101 * to a ksm page left outside the stable tree,
1102 * in which case we need to break_cow on both.
1104 if (stable_tree_insert(page2[0], tree_rmap_item))
1105 stable_tree_append(rmap_item, tree_rmap_item);
1107 break_cow(tree_rmap_item->mm,
1108 tree_rmap_item->address);
1109 break_cow(rmap_item->mm, rmap_item->address);
1110 ksm_pages_shared -= 2;
1118 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1119 struct list_head *cur,
1122 struct rmap_item *rmap_item;
1124 while (cur != &mm_slot->rmap_list) {
1125 rmap_item = list_entry(cur, struct rmap_item, link);
1126 if ((rmap_item->address & PAGE_MASK) == addr) {
1127 if (!in_stable_tree(rmap_item))
1128 remove_rmap_item_from_tree(rmap_item);
1131 if (rmap_item->address > addr)
1134 remove_rmap_item_from_tree(rmap_item);
1135 list_del(&rmap_item->link);
1136 free_rmap_item(rmap_item);
1139 rmap_item = alloc_rmap_item();
1141 /* It has already been zeroed */
1142 rmap_item->mm = mm_slot->mm;
1143 rmap_item->address = addr;
1144 list_add_tail(&rmap_item->link, cur);
1149 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1151 struct mm_struct *mm;
1152 struct mm_slot *slot;
1153 struct vm_area_struct *vma;
1154 struct rmap_item *rmap_item;
1156 if (list_empty(&ksm_mm_head.mm_list))
1159 slot = ksm_scan.mm_slot;
1160 if (slot == &ksm_mm_head) {
1161 root_unstable_tree = RB_ROOT;
1163 spin_lock(&ksm_mmlist_lock);
1164 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1165 ksm_scan.mm_slot = slot;
1166 spin_unlock(&ksm_mmlist_lock);
1168 ksm_scan.address = 0;
1169 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1170 struct rmap_item, link);
1174 down_read(&mm->mmap_sem);
1175 for (vma = find_vma(mm, ksm_scan.address); vma; vma = vma->vm_next) {
1176 if (!(vma->vm_flags & VM_MERGEABLE))
1178 if (ksm_scan.address < vma->vm_start)
1179 ksm_scan.address = vma->vm_start;
1181 ksm_scan.address = vma->vm_end;
1183 while (ksm_scan.address < vma->vm_end) {
1184 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1185 if (*page && PageAnon(*page)) {
1186 flush_anon_page(vma, *page, ksm_scan.address);
1187 flush_dcache_page(*page);
1188 rmap_item = get_next_rmap_item(slot,
1189 ksm_scan.rmap_item->link.next,
1192 ksm_scan.rmap_item = rmap_item;
1193 ksm_scan.address += PAGE_SIZE;
1196 up_read(&mm->mmap_sem);
1201 ksm_scan.address += PAGE_SIZE;
1206 if (!ksm_scan.address) {
1208 * We've completed a full scan of all vmas, holding mmap_sem
1209 * throughout, and found no VM_MERGEABLE: so do the same as
1210 * __ksm_exit does to remove this mm from all our lists now.
1212 remove_mm_from_lists(mm);
1213 up_read(&mm->mmap_sem);
1214 slot = ksm_scan.mm_slot;
1215 if (slot != &ksm_mm_head)
1221 * Nuke all the rmap_items that are above this current rmap:
1222 * because there were no VM_MERGEABLE vmas with such addresses.
1224 remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1225 up_read(&mm->mmap_sem);
1227 spin_lock(&ksm_mmlist_lock);
1228 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1229 ksm_scan.mm_slot = slot;
1230 spin_unlock(&ksm_mmlist_lock);
1232 /* Repeat until we've completed scanning the whole list */
1233 if (slot != &ksm_mm_head)
1237 * Bump seqnr here rather than at top, so that __ksm_exit
1238 * can skip rb_erase on unstable tree until we run again.
1245 * ksm_do_scan - the ksm scanner main worker function.
1246 * @scan_npages - number of pages we want to scan before we return.
1248 static void ksm_do_scan(unsigned int scan_npages)
1250 struct rmap_item *rmap_item;
1253 while (scan_npages--) {
1255 rmap_item = scan_get_next_rmap_item(&page);
1258 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1259 cmp_and_merge_page(page, rmap_item);
1264 static int ksm_scan_thread(void *nothing)
1266 set_user_nice(current, 0);
1268 while (!kthread_should_stop()) {
1269 if (ksm_run & KSM_RUN_MERGE) {
1270 mutex_lock(&ksm_thread_mutex);
1271 ksm_do_scan(ksm_thread_pages_to_scan);
1272 mutex_unlock(&ksm_thread_mutex);
1273 schedule_timeout_interruptible(
1274 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1276 wait_event_interruptible(ksm_thread_wait,
1277 (ksm_run & KSM_RUN_MERGE) ||
1278 kthread_should_stop());
1284 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1285 unsigned long end, int advice, unsigned long *vm_flags)
1287 struct mm_struct *mm = vma->vm_mm;
1290 case MADV_MERGEABLE:
1292 * Be somewhat over-protective for now!
1294 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1295 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1296 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1297 VM_MIXEDMAP | VM_SAO))
1298 return 0; /* just ignore the advice */
1300 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
1301 if (__ksm_enter(mm) < 0)
1304 *vm_flags |= VM_MERGEABLE;
1307 case MADV_UNMERGEABLE:
1308 if (!(*vm_flags & VM_MERGEABLE))
1309 return 0; /* just ignore the advice */
1312 unmerge_ksm_pages(vma, start, end);
1314 *vm_flags &= ~VM_MERGEABLE;
1321 int __ksm_enter(struct mm_struct *mm)
1323 struct mm_slot *mm_slot = alloc_mm_slot();
1327 spin_lock(&ksm_mmlist_lock);
1328 insert_to_mm_slots_hash(mm, mm_slot);
1330 * Insert just behind the scanning cursor, to let the area settle
1331 * down a little; when fork is followed by immediate exec, we don't
1332 * want ksmd to waste time setting up and tearing down an rmap_list.
1334 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1335 spin_unlock(&ksm_mmlist_lock);
1337 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1341 void __ksm_exit(struct mm_struct *mm)
1344 * This process is exiting: doesn't hold and doesn't need mmap_sem;
1345 * but we do need to exclude ksmd and other exiters while we modify
1346 * the various lists and trees.
1348 mutex_lock(&ksm_thread_mutex);
1349 remove_mm_from_lists(mm);
1350 mutex_unlock(&ksm_thread_mutex);
1353 #define KSM_ATTR_RO(_name) \
1354 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1355 #define KSM_ATTR(_name) \
1356 static struct kobj_attribute _name##_attr = \
1357 __ATTR(_name, 0644, _name##_show, _name##_store)
1359 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1360 struct kobj_attribute *attr, char *buf)
1362 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1365 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1366 struct kobj_attribute *attr,
1367 const char *buf, size_t count)
1369 unsigned long msecs;
1372 err = strict_strtoul(buf, 10, &msecs);
1373 if (err || msecs > UINT_MAX)
1376 ksm_thread_sleep_millisecs = msecs;
1380 KSM_ATTR(sleep_millisecs);
1382 static ssize_t pages_to_scan_show(struct kobject *kobj,
1383 struct kobj_attribute *attr, char *buf)
1385 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1388 static ssize_t pages_to_scan_store(struct kobject *kobj,
1389 struct kobj_attribute *attr,
1390 const char *buf, size_t count)
1393 unsigned long nr_pages;
1395 err = strict_strtoul(buf, 10, &nr_pages);
1396 if (err || nr_pages > UINT_MAX)
1399 ksm_thread_pages_to_scan = nr_pages;
1403 KSM_ATTR(pages_to_scan);
1405 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1408 return sprintf(buf, "%u\n", ksm_run);
1411 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1412 const char *buf, size_t count)
1415 unsigned long flags;
1417 err = strict_strtoul(buf, 10, &flags);
1418 if (err || flags > UINT_MAX)
1420 if (flags > KSM_RUN_UNMERGE)
1424 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1425 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1426 * breaking COW to free the kernel_pages_allocated (but leaves
1427 * mm_slots on the list for when ksmd may be set running again).
1430 mutex_lock(&ksm_thread_mutex);
1431 if (ksm_run != flags) {
1433 if (flags & KSM_RUN_UNMERGE)
1434 unmerge_and_remove_all_rmap_items();
1436 mutex_unlock(&ksm_thread_mutex);
1438 if (flags & KSM_RUN_MERGE)
1439 wake_up_interruptible(&ksm_thread_wait);
1445 static ssize_t pages_shared_show(struct kobject *kobj,
1446 struct kobj_attribute *attr, char *buf)
1448 return sprintf(buf, "%lu\n",
1449 ksm_pages_shared - ksm_kernel_pages_allocated);
1451 KSM_ATTR_RO(pages_shared);
1453 static ssize_t kernel_pages_allocated_show(struct kobject *kobj,
1454 struct kobj_attribute *attr,
1457 return sprintf(buf, "%lu\n", ksm_kernel_pages_allocated);
1459 KSM_ATTR_RO(kernel_pages_allocated);
1461 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1462 struct kobj_attribute *attr,
1463 const char *buf, size_t count)
1466 unsigned long nr_pages;
1468 err = strict_strtoul(buf, 10, &nr_pages);
1472 ksm_max_kernel_pages = nr_pages;
1477 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1478 struct kobj_attribute *attr, char *buf)
1480 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1482 KSM_ATTR(max_kernel_pages);
1484 static struct attribute *ksm_attrs[] = {
1485 &sleep_millisecs_attr.attr,
1486 &pages_to_scan_attr.attr,
1488 &pages_shared_attr.attr,
1489 &kernel_pages_allocated_attr.attr,
1490 &max_kernel_pages_attr.attr,
1494 static struct attribute_group ksm_attr_group = {
1499 static int __init ksm_init(void)
1501 struct task_struct *ksm_thread;
1504 err = ksm_slab_init();
1508 err = mm_slots_hash_init();
1512 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1513 if (IS_ERR(ksm_thread)) {
1514 printk(KERN_ERR "ksm: creating kthread failed\n");
1515 err = PTR_ERR(ksm_thread);
1519 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1521 printk(KERN_ERR "ksm: register sysfs failed\n");
1528 kthread_stop(ksm_thread);
1530 mm_slots_hash_free();
1536 module_init(ksm_init)