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 Red Hat, Inc.
13 * This work is licensed under the terms of the GNU GPL, version 2.
16 #include <linux/errno.h>
19 #include <linux/mman.h>
20 #include <linux/sched.h>
21 #include <linux/rwsem.h>
22 #include <linux/pagemap.h>
23 #include <linux/rmap.h>
24 #include <linux/spinlock.h>
25 #include <linux/jhash.h>
26 #include <linux/delay.h>
27 #include <linux/kthread.h>
28 #include <linux/wait.h>
29 #include <linux/slab.h>
30 #include <linux/rbtree.h>
31 #include <linux/mmu_notifier.h>
32 #include <linux/ksm.h>
34 #include <asm/tlbflush.h>
37 * A few notes about the KSM scanning process,
38 * to make it easier to understand the data structures below:
40 * In order to reduce excessive scanning, KSM sorts the memory pages by their
41 * contents into a data structure that holds pointers to the pages' locations.
43 * Since the contents of the pages may change at any moment, KSM cannot just
44 * insert the pages into a normal sorted tree and expect it to find anything.
45 * Therefore KSM uses two data structures - the stable and the unstable tree.
47 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
48 * by their contents. Because each such page is write-protected, searching on
49 * this tree is fully assured to be working (except when pages are unmapped),
50 * and therefore this tree is called the stable tree.
52 * In addition to the stable tree, KSM uses a second data structure called the
53 * unstable tree: this tree holds pointers to pages which have been found to
54 * be "unchanged for a period of time". The unstable tree sorts these pages
55 * by their contents, but since they are not write-protected, KSM cannot rely
56 * upon the unstable tree to work correctly - the unstable tree is liable to
57 * be corrupted as its contents are modified, and so it is called unstable.
59 * KSM solves this problem by several techniques:
61 * 1) The unstable tree is flushed every time KSM completes scanning all
62 * memory areas, and then the tree is rebuilt again from the beginning.
63 * 2) KSM will only insert into the unstable tree, pages whose hash value
64 * has not changed since the previous scan of all memory areas.
65 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
66 * colors of the nodes and not on their contents, assuring that even when
67 * the tree gets "corrupted" it won't get out of balance, so scanning time
68 * remains the same (also, searching and inserting nodes in an rbtree uses
69 * the same algorithm, so we have no overhead when we flush and rebuild).
70 * 4) KSM never flushes the stable tree, which means that even if it were to
71 * take 10 attempts to find a page in the unstable tree, once it is found,
72 * it is secured in the stable tree. (When we scan a new page, we first
73 * compare it against the stable tree, and then against the unstable tree.)
77 * struct mm_slot - ksm information per mm that is being scanned
78 * @link: link to the mm_slots hash list
79 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
80 * @rmap_list: head for this mm_slot's list of rmap_items
81 * @mm: the mm that this information is valid for
84 struct hlist_node link;
85 struct list_head mm_list;
86 struct list_head rmap_list;
91 * struct ksm_scan - cursor for scanning
92 * @mm_slot: the current mm_slot we are scanning
93 * @address: the next address inside that to be scanned
94 * @rmap_item: the current rmap that we are scanning inside the rmap_list
95 * @seqnr: count of completed full scans (needed when removing unstable node)
97 * There is only the one ksm_scan instance of this cursor structure.
100 struct mm_slot *mm_slot;
101 unsigned long address;
102 struct rmap_item *rmap_item;
107 * struct rmap_item - reverse mapping item for virtual addresses
108 * @link: link into mm_slot's rmap_list (rmap_list is per mm)
109 * @mm: the memory structure this rmap_item is pointing into
110 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
111 * @oldchecksum: previous checksum of the page at that virtual address
112 * @node: rb_node of this rmap_item in either unstable or stable tree
113 * @next: next rmap_item hanging off the same node of the stable tree
114 * @prev: previous rmap_item hanging off the same node of the stable tree
117 struct list_head link;
118 struct mm_struct *mm;
119 unsigned long address; /* + low bits used for flags below */
121 unsigned int oldchecksum; /* when unstable */
122 struct rmap_item *next; /* when stable */
125 struct rb_node node; /* when tree node */
126 struct rmap_item *prev; /* in stable list */
130 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
131 #define NODE_FLAG 0x100 /* is a node of unstable or stable tree */
132 #define STABLE_FLAG 0x200 /* is a node or list item of stable tree */
134 /* The stable and unstable tree heads */
135 static struct rb_root root_stable_tree = RB_ROOT;
136 static struct rb_root root_unstable_tree = RB_ROOT;
138 #define MM_SLOTS_HASH_HEADS 1024
139 static struct hlist_head *mm_slots_hash;
141 static struct mm_slot ksm_mm_head = {
142 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
144 static struct ksm_scan ksm_scan = {
145 .mm_slot = &ksm_mm_head,
148 static struct kmem_cache *rmap_item_cache;
149 static struct kmem_cache *mm_slot_cache;
151 /* The number of nodes in the stable tree */
152 static unsigned long ksm_kernel_pages_allocated;
154 /* The number of page slots sharing those nodes */
155 static unsigned long ksm_pages_shared;
157 /* Limit on the number of unswappable pages used */
158 static unsigned long ksm_max_kernel_pages;
160 /* Number of pages ksmd should scan in one batch */
161 static unsigned int ksm_thread_pages_to_scan;
163 /* Milliseconds ksmd should sleep between batches */
164 static unsigned int ksm_thread_sleep_millisecs;
166 #define KSM_RUN_STOP 0
167 #define KSM_RUN_MERGE 1
168 #define KSM_RUN_UNMERGE 2
169 static unsigned int ksm_run;
171 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
172 static DEFINE_MUTEX(ksm_thread_mutex);
173 static DEFINE_SPINLOCK(ksm_mmlist_lock);
175 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
176 sizeof(struct __struct), __alignof__(struct __struct),\
179 static int __init ksm_slab_init(void)
181 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
182 if (!rmap_item_cache)
185 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
192 kmem_cache_destroy(rmap_item_cache);
197 static void __init ksm_slab_free(void)
199 kmem_cache_destroy(mm_slot_cache);
200 kmem_cache_destroy(rmap_item_cache);
201 mm_slot_cache = NULL;
204 static inline struct rmap_item *alloc_rmap_item(void)
206 return kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
209 static inline void free_rmap_item(struct rmap_item *rmap_item)
211 rmap_item->mm = NULL; /* debug safety */
212 kmem_cache_free(rmap_item_cache, rmap_item);
215 static inline struct mm_slot *alloc_mm_slot(void)
217 if (!mm_slot_cache) /* initialization failed */
219 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
222 static inline void free_mm_slot(struct mm_slot *mm_slot)
224 kmem_cache_free(mm_slot_cache, mm_slot);
227 static int __init mm_slots_hash_init(void)
229 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
236 static void __init mm_slots_hash_free(void)
238 kfree(mm_slots_hash);
241 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
243 struct mm_slot *mm_slot;
244 struct hlist_head *bucket;
245 struct hlist_node *node;
247 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
248 % MM_SLOTS_HASH_HEADS];
249 hlist_for_each_entry(mm_slot, node, bucket, link) {
250 if (mm == mm_slot->mm)
256 static void insert_to_mm_slots_hash(struct mm_struct *mm,
257 struct mm_slot *mm_slot)
259 struct hlist_head *bucket;
261 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
262 % MM_SLOTS_HASH_HEADS];
264 INIT_LIST_HEAD(&mm_slot->rmap_list);
265 hlist_add_head(&mm_slot->link, bucket);
268 static inline int in_stable_tree(struct rmap_item *rmap_item)
270 return rmap_item->address & STABLE_FLAG;
274 * We use break_ksm to break COW on a ksm page: it's a stripped down
276 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
279 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
280 * in case the application has unmapped and remapped mm,addr meanwhile.
281 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
282 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
284 static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
291 page = follow_page(vma, addr, FOLL_GET);
295 ret = handle_mm_fault(vma->vm_mm, vma, addr,
298 ret = VM_FAULT_WRITE;
300 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS)));
302 /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
305 static void __break_cow(struct mm_struct *mm, unsigned long addr)
307 struct vm_area_struct *vma;
309 vma = find_vma(mm, addr);
310 if (!vma || vma->vm_start > addr)
312 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
314 break_ksm(vma, addr);
317 static void break_cow(struct mm_struct *mm, unsigned long addr)
319 down_read(&mm->mmap_sem);
320 __break_cow(mm, addr);
321 up_read(&mm->mmap_sem);
324 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
326 struct mm_struct *mm = rmap_item->mm;
327 unsigned long addr = rmap_item->address;
328 struct vm_area_struct *vma;
331 down_read(&mm->mmap_sem);
332 vma = find_vma(mm, addr);
333 if (!vma || vma->vm_start > addr)
335 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
338 page = follow_page(vma, addr, FOLL_GET);
341 if (PageAnon(page)) {
342 flush_anon_page(vma, page, addr);
343 flush_dcache_page(page);
348 up_read(&mm->mmap_sem);
353 * get_ksm_page: checks if the page at the virtual address in rmap_item
354 * is still PageKsm, in which case we can trust the content of the page,
355 * and it returns the gotten page; but NULL if the page has been zapped.
357 static struct page *get_ksm_page(struct rmap_item *rmap_item)
361 page = get_mergeable_page(rmap_item);
362 if (page && !PageKsm(page)) {
370 * Removing rmap_item from stable or unstable tree.
371 * This function will clean the information from the stable/unstable tree.
373 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
375 if (in_stable_tree(rmap_item)) {
376 struct rmap_item *next_item = rmap_item->next;
378 if (rmap_item->address & NODE_FLAG) {
380 rb_replace_node(&rmap_item->node,
383 next_item->address |= NODE_FLAG;
385 rb_erase(&rmap_item->node, &root_stable_tree);
386 ksm_kernel_pages_allocated--;
389 struct rmap_item *prev_item = rmap_item->prev;
391 BUG_ON(prev_item->next != rmap_item);
392 prev_item->next = next_item;
394 BUG_ON(next_item->prev != rmap_item);
395 next_item->prev = rmap_item->prev;
399 rmap_item->next = NULL;
402 } else if (rmap_item->address & NODE_FLAG) {
405 * ksm_thread can and must skip the rb_erase, because
406 * root_unstable_tree was already reset to RB_ROOT.
407 * But __ksm_exit has to be careful: do the rb_erase
408 * if it's interrupting a scan, and this rmap_item was
409 * inserted by this scan rather than left from before.
411 * Because of the case in which remove_mm_from_lists
412 * increments seqnr before removing rmaps, unstable_nr
413 * may even be 2 behind seqnr, but should never be
414 * further behind. Yes, I did have trouble with this!
416 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
419 rb_erase(&rmap_item->node, &root_unstable_tree);
422 rmap_item->address &= PAGE_MASK;
424 cond_resched(); /* we're called from many long loops */
427 static void remove_all_slot_rmap_items(struct mm_slot *mm_slot)
429 struct rmap_item *rmap_item, *node;
431 list_for_each_entry_safe(rmap_item, node, &mm_slot->rmap_list, link) {
432 remove_rmap_item_from_tree(rmap_item);
433 list_del(&rmap_item->link);
434 free_rmap_item(rmap_item);
438 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
439 struct list_head *cur)
441 struct rmap_item *rmap_item;
443 while (cur != &mm_slot->rmap_list) {
444 rmap_item = list_entry(cur, struct rmap_item, link);
446 remove_rmap_item_from_tree(rmap_item);
447 list_del(&rmap_item->link);
448 free_rmap_item(rmap_item);
453 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
454 * than check every pte of a given vma, the locking doesn't quite work for
455 * that - an rmap_item is assigned to the stable tree after inserting ksm
456 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
457 * rmap_items from parent to child at fork time (so as not to waste time
458 * if exit comes before the next scan reaches it).
460 static void unmerge_ksm_pages(struct vm_area_struct *vma,
461 unsigned long start, unsigned long end)
465 for (addr = start; addr < end; addr += PAGE_SIZE)
466 break_ksm(vma, addr);
469 static void unmerge_and_remove_all_rmap_items(void)
471 struct mm_slot *mm_slot;
472 struct mm_struct *mm;
473 struct vm_area_struct *vma;
475 list_for_each_entry(mm_slot, &ksm_mm_head.mm_list, mm_list) {
477 down_read(&mm->mmap_sem);
478 for (vma = mm->mmap; vma; vma = vma->vm_next) {
479 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
481 unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end);
483 remove_all_slot_rmap_items(mm_slot);
484 up_read(&mm->mmap_sem);
487 spin_lock(&ksm_mmlist_lock);
488 if (ksm_scan.mm_slot != &ksm_mm_head) {
489 ksm_scan.mm_slot = &ksm_mm_head;
492 spin_unlock(&ksm_mmlist_lock);
495 static void remove_mm_from_lists(struct mm_struct *mm)
497 struct mm_slot *mm_slot;
499 spin_lock(&ksm_mmlist_lock);
500 mm_slot = get_mm_slot(mm);
503 * This mm_slot is always at the scanning cursor when we're
504 * called from scan_get_next_rmap_item; but it's a special
505 * case when we're called from __ksm_exit.
507 if (ksm_scan.mm_slot == mm_slot) {
508 ksm_scan.mm_slot = list_entry(
509 mm_slot->mm_list.next, struct mm_slot, mm_list);
510 ksm_scan.address = 0;
511 ksm_scan.rmap_item = list_entry(
512 &ksm_scan.mm_slot->rmap_list, struct rmap_item, link);
513 if (ksm_scan.mm_slot == &ksm_mm_head)
517 hlist_del(&mm_slot->link);
518 list_del(&mm_slot->mm_list);
519 spin_unlock(&ksm_mmlist_lock);
521 remove_all_slot_rmap_items(mm_slot);
522 free_mm_slot(mm_slot);
523 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
526 static u32 calc_checksum(struct page *page)
529 void *addr = kmap_atomic(page, KM_USER0);
530 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
531 kunmap_atomic(addr, KM_USER0);
535 static int memcmp_pages(struct page *page1, struct page *page2)
540 addr1 = kmap_atomic(page1, KM_USER0);
541 addr2 = kmap_atomic(page2, KM_USER1);
542 ret = memcmp(addr1, addr2, PAGE_SIZE);
543 kunmap_atomic(addr2, KM_USER1);
544 kunmap_atomic(addr1, KM_USER0);
548 static inline int pages_identical(struct page *page1, struct page *page2)
550 return !memcmp_pages(page1, page2);
553 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
556 struct mm_struct *mm = vma->vm_mm;
563 addr = page_address_in_vma(page, vma);
567 ptep = page_check_address(page, mm, addr, &ptl, 0);
571 if (pte_write(*ptep)) {
574 swapped = PageSwapCache(page);
575 flush_cache_page(vma, addr, page_to_pfn(page));
577 * Ok this is tricky, when get_user_pages_fast() run it doesnt
578 * take any lock, therefore the check that we are going to make
579 * with the pagecount against the mapcount is racey and
580 * O_DIRECT can happen right after the check.
581 * So we clear the pte and flush the tlb before the check
582 * this assure us that no O_DIRECT can happen after the check
583 * or in the middle of the check.
585 entry = ptep_clear_flush(vma, addr, ptep);
587 * Check that no O_DIRECT or similar I/O is in progress on the
590 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
591 set_pte_at_notify(mm, addr, ptep, entry);
594 entry = pte_wrprotect(entry);
595 set_pte_at_notify(mm, addr, ptep, entry);
601 pte_unmap_unlock(ptep, ptl);
607 * replace_page - replace page in vma by new ksm page
608 * @vma: vma that holds the pte pointing to oldpage
609 * @oldpage: the page we are replacing by newpage
610 * @newpage: the ksm page we replace oldpage by
611 * @orig_pte: the original value of the pte
613 * Returns 0 on success, -EFAULT on failure.
615 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
616 struct page *newpage, pte_t orig_pte)
618 struct mm_struct *mm = vma->vm_mm;
628 prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
630 addr = page_address_in_vma(oldpage, vma);
634 pgd = pgd_offset(mm, addr);
635 if (!pgd_present(*pgd))
638 pud = pud_offset(pgd, addr);
639 if (!pud_present(*pud))
642 pmd = pmd_offset(pud, addr);
643 if (!pmd_present(*pmd))
646 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
647 if (!pte_same(*ptep, orig_pte)) {
648 pte_unmap_unlock(ptep, ptl);
653 page_add_ksm_rmap(newpage);
655 flush_cache_page(vma, addr, pte_pfn(*ptep));
656 ptep_clear_flush(vma, addr, ptep);
657 set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
659 page_remove_rmap(oldpage);
662 pte_unmap_unlock(ptep, ptl);
669 * try_to_merge_one_page - take two pages and merge them into one
670 * @vma: the vma that hold the pte pointing into oldpage
671 * @oldpage: the page that we want to replace with newpage
672 * @newpage: the page that we want to map instead of oldpage
675 * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
676 * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
678 * This function returns 0 if the pages were merged, -EFAULT otherwise.
680 static int try_to_merge_one_page(struct vm_area_struct *vma,
681 struct page *oldpage,
682 struct page *newpage)
684 pte_t orig_pte = __pte(0);
687 if (!(vma->vm_flags & VM_MERGEABLE))
690 if (!PageAnon(oldpage))
697 * We need the page lock to read a stable PageSwapCache in
698 * write_protect_page(). We use trylock_page() instead of
699 * lock_page() because we don't want to wait here - we
700 * prefer to continue scanning and merging different pages,
701 * then come back to this page when it is unlocked.
703 if (!trylock_page(oldpage))
706 * If this anonymous page is mapped only here, its pte may need
707 * to be write-protected. If it's mapped elsewhere, all of its
708 * ptes are necessarily already write-protected. But in either
709 * case, we need to lock and check page_count is not raised.
711 if (write_protect_page(vma, oldpage, &orig_pte)) {
712 unlock_page(oldpage);
715 unlock_page(oldpage);
717 if (pages_identical(oldpage, newpage))
718 err = replace_page(vma, oldpage, newpage, orig_pte);
728 * try_to_merge_two_pages - take two identical pages and prepare them
729 * to be merged into one page.
731 * This function returns 0 if we successfully mapped two identical pages
732 * into one page, -EFAULT otherwise.
734 * Note that this function allocates a new kernel page: if one of the pages
735 * is already a ksm page, try_to_merge_with_ksm_page should be used.
737 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
738 struct page *page1, struct mm_struct *mm2,
739 unsigned long addr2, struct page *page2)
741 struct vm_area_struct *vma;
746 * The number of nodes in the stable tree
747 * is the number of kernel pages that we hold.
749 if (ksm_max_kernel_pages &&
750 ksm_max_kernel_pages <= ksm_kernel_pages_allocated)
753 kpage = alloc_page(GFP_HIGHUSER);
757 down_read(&mm1->mmap_sem);
758 vma = find_vma(mm1, addr1);
759 if (!vma || vma->vm_start > addr1) {
761 up_read(&mm1->mmap_sem);
765 copy_user_highpage(kpage, page1, addr1, vma);
766 err = try_to_merge_one_page(vma, page1, kpage);
767 up_read(&mm1->mmap_sem);
770 down_read(&mm2->mmap_sem);
771 vma = find_vma(mm2, addr2);
772 if (!vma || vma->vm_start > addr2) {
774 up_read(&mm2->mmap_sem);
775 break_cow(mm1, addr1);
779 err = try_to_merge_one_page(vma, page2, kpage);
780 up_read(&mm2->mmap_sem);
783 * If the second try_to_merge_one_page failed, we have a
784 * ksm page with just one pte pointing to it, so break it.
787 break_cow(mm1, addr1);
789 ksm_pages_shared += 2;
797 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
798 * but no new kernel page is allocated: kpage must already be a ksm page.
800 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
805 struct vm_area_struct *vma;
808 down_read(&mm1->mmap_sem);
809 vma = find_vma(mm1, addr1);
810 if (!vma || vma->vm_start > addr1) {
811 up_read(&mm1->mmap_sem);
815 err = try_to_merge_one_page(vma, page1, kpage);
816 up_read(&mm1->mmap_sem);
825 * stable_tree_search - search page inside the stable tree
826 * @page: the page that we are searching identical pages to.
827 * @page2: pointer into identical page that we are holding inside the stable
828 * tree that we have found.
829 * @rmap_item: the reverse mapping item
831 * This function checks if there is a page inside the stable tree
832 * with identical content to the page that we are scanning right now.
834 * This function return rmap_item pointer to the identical item if found,
837 static struct rmap_item *stable_tree_search(struct page *page,
839 struct rmap_item *rmap_item)
841 struct rb_node *node = root_stable_tree.rb_node;
844 struct rmap_item *tree_rmap_item, *next_rmap_item;
847 tree_rmap_item = rb_entry(node, struct rmap_item, node);
848 while (tree_rmap_item) {
849 BUG_ON(!in_stable_tree(tree_rmap_item));
851 page2[0] = get_ksm_page(tree_rmap_item);
854 next_rmap_item = tree_rmap_item->next;
855 remove_rmap_item_from_tree(tree_rmap_item);
856 tree_rmap_item = next_rmap_item;
861 ret = memcmp_pages(page, page2[0]);
865 node = node->rb_left;
866 } else if (ret > 0) {
868 node = node->rb_right;
870 return tree_rmap_item;
878 * stable_tree_insert - insert rmap_item pointing to new ksm page
879 * into the stable tree.
881 * @page: the page that we are searching identical page to inside the stable
883 * @rmap_item: pointer to the reverse mapping item.
885 * This function returns rmap_item if success, NULL otherwise.
887 static struct rmap_item *stable_tree_insert(struct page *page,
888 struct rmap_item *rmap_item)
890 struct rb_node **new = &root_stable_tree.rb_node;
891 struct rb_node *parent = NULL;
894 struct rmap_item *tree_rmap_item, *next_rmap_item;
895 struct page *tree_page;
898 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
899 while (tree_rmap_item) {
900 BUG_ON(!in_stable_tree(tree_rmap_item));
902 tree_page = get_ksm_page(tree_rmap_item);
905 next_rmap_item = tree_rmap_item->next;
906 remove_rmap_item_from_tree(tree_rmap_item);
907 tree_rmap_item = next_rmap_item;
912 ret = memcmp_pages(page, tree_page);
917 new = &parent->rb_left;
919 new = &parent->rb_right;
922 * It is not a bug that stable_tree_search() didn't
923 * find this node: because at that time our page was
924 * not yet write-protected, so may have changed since.
930 ksm_kernel_pages_allocated++;
932 rmap_item->address |= NODE_FLAG | STABLE_FLAG;
933 rmap_item->next = NULL;
934 rb_link_node(&rmap_item->node, parent, new);
935 rb_insert_color(&rmap_item->node, &root_stable_tree);
941 * unstable_tree_search_insert - search and insert items into the unstable tree.
943 * @page: the page that we are going to search for identical page or to insert
944 * into the unstable tree
945 * @page2: pointer into identical page that was found inside the unstable tree
946 * @rmap_item: the reverse mapping item of page
948 * This function searches for a page in the unstable tree identical to the
949 * page currently being scanned; and if no identical page is found in the
950 * tree, we insert rmap_item as a new object into the unstable tree.
952 * This function returns pointer to rmap_item found to be identical
953 * to the currently scanned page, NULL otherwise.
955 * This function does both searching and inserting, because they share
956 * the same walking algorithm in an rbtree.
958 static struct rmap_item *unstable_tree_search_insert(struct page *page,
960 struct rmap_item *rmap_item)
962 struct rb_node **new = &root_unstable_tree.rb_node;
963 struct rb_node *parent = NULL;
966 struct rmap_item *tree_rmap_item;
969 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
970 page2[0] = get_mergeable_page(tree_rmap_item);
975 * Don't substitute an unswappable ksm page
976 * just for one good swappable forked page.
978 if (page == page2[0]) {
983 ret = memcmp_pages(page, page2[0]);
988 new = &parent->rb_left;
989 } else if (ret > 0) {
991 new = &parent->rb_right;
993 return tree_rmap_item;
997 rmap_item->address |= NODE_FLAG;
998 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
999 rb_link_node(&rmap_item->node, parent, new);
1000 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1006 * stable_tree_append - add another rmap_item to the linked list of
1007 * rmap_items hanging off a given node of the stable tree, all sharing
1008 * the same ksm page.
1010 static void stable_tree_append(struct rmap_item *rmap_item,
1011 struct rmap_item *tree_rmap_item)
1013 rmap_item->next = tree_rmap_item->next;
1014 rmap_item->prev = tree_rmap_item;
1016 if (tree_rmap_item->next)
1017 tree_rmap_item->next->prev = rmap_item;
1019 tree_rmap_item->next = rmap_item;
1020 rmap_item->address |= STABLE_FLAG;
1024 * cmp_and_merge_page - take a page computes its hash value and check if there
1025 * is similar hash value to different page,
1026 * in case we find that there is similar hash to different page we call to
1027 * try_to_merge_two_pages().
1029 * @page: the page that we are searching identical page to.
1030 * @rmap_item: the reverse mapping into the virtual address of this page
1032 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1034 struct page *page2[1];
1035 struct rmap_item *tree_rmap_item;
1036 unsigned int checksum;
1039 if (in_stable_tree(rmap_item))
1040 remove_rmap_item_from_tree(rmap_item);
1042 /* We first start with searching the page inside the stable tree */
1043 tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1044 if (tree_rmap_item) {
1045 if (page == page2[0]) { /* forked */
1049 err = try_to_merge_with_ksm_page(rmap_item->mm,
1056 * The page was successfully merged:
1057 * add its rmap_item to the stable tree.
1059 stable_tree_append(rmap_item, tree_rmap_item);
1065 * A ksm page might have got here by fork, but its other
1066 * references have already been removed from the stable tree.
1069 break_cow(rmap_item->mm, rmap_item->address);
1072 * In case the hash value of the page was changed from the last time we
1073 * have calculated it, this page to be changed frequely, therefore we
1074 * don't want to insert it to the unstable tree, and we don't want to
1075 * waste our time to search if there is something identical to it there.
1077 checksum = calc_checksum(page);
1078 if (rmap_item->oldchecksum != checksum) {
1079 rmap_item->oldchecksum = checksum;
1083 tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1084 if (tree_rmap_item) {
1085 err = try_to_merge_two_pages(rmap_item->mm,
1086 rmap_item->address, page,
1088 tree_rmap_item->address, page2[0]);
1090 * As soon as we merge this page, we want to remove the
1091 * rmap_item of the page we have merged with from the unstable
1092 * tree, and insert it instead as new node in the stable tree.
1095 rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1096 tree_rmap_item->address &= ~NODE_FLAG;
1098 * If we fail to insert the page into the stable tree,
1099 * we will have 2 virtual addresses that are pointing
1100 * to a ksm page left outside the stable tree,
1101 * in which case we need to break_cow on both.
1103 if (stable_tree_insert(page2[0], tree_rmap_item))
1104 stable_tree_append(rmap_item, tree_rmap_item);
1106 break_cow(tree_rmap_item->mm,
1107 tree_rmap_item->address);
1108 break_cow(rmap_item->mm, rmap_item->address);
1109 ksm_pages_shared -= 2;
1117 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1118 struct list_head *cur,
1121 struct rmap_item *rmap_item;
1123 while (cur != &mm_slot->rmap_list) {
1124 rmap_item = list_entry(cur, struct rmap_item, link);
1125 if ((rmap_item->address & PAGE_MASK) == addr) {
1126 if (!in_stable_tree(rmap_item))
1127 remove_rmap_item_from_tree(rmap_item);
1130 if (rmap_item->address > addr)
1133 remove_rmap_item_from_tree(rmap_item);
1134 list_del(&rmap_item->link);
1135 free_rmap_item(rmap_item);
1138 rmap_item = alloc_rmap_item();
1140 /* It has already been zeroed */
1141 rmap_item->mm = mm_slot->mm;
1142 rmap_item->address = addr;
1143 list_add_tail(&rmap_item->link, cur);
1148 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1150 struct mm_struct *mm;
1151 struct mm_slot *slot;
1152 struct vm_area_struct *vma;
1153 struct rmap_item *rmap_item;
1155 if (list_empty(&ksm_mm_head.mm_list))
1158 slot = ksm_scan.mm_slot;
1159 if (slot == &ksm_mm_head) {
1160 root_unstable_tree = RB_ROOT;
1162 spin_lock(&ksm_mmlist_lock);
1163 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1164 ksm_scan.mm_slot = slot;
1165 spin_unlock(&ksm_mmlist_lock);
1167 ksm_scan.address = 0;
1168 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1169 struct rmap_item, link);
1173 down_read(&mm->mmap_sem);
1174 for (vma = find_vma(mm, ksm_scan.address); vma; vma = vma->vm_next) {
1175 if (!(vma->vm_flags & VM_MERGEABLE))
1177 if (ksm_scan.address < vma->vm_start)
1178 ksm_scan.address = vma->vm_start;
1180 ksm_scan.address = vma->vm_end;
1182 while (ksm_scan.address < vma->vm_end) {
1183 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1184 if (*page && PageAnon(*page)) {
1185 flush_anon_page(vma, *page, ksm_scan.address);
1186 flush_dcache_page(*page);
1187 rmap_item = get_next_rmap_item(slot,
1188 ksm_scan.rmap_item->link.next,
1191 ksm_scan.rmap_item = rmap_item;
1192 ksm_scan.address += PAGE_SIZE;
1195 up_read(&mm->mmap_sem);
1200 ksm_scan.address += PAGE_SIZE;
1205 if (!ksm_scan.address) {
1207 * We've completed a full scan of all vmas, holding mmap_sem
1208 * throughout, and found no VM_MERGEABLE: so do the same as
1209 * __ksm_exit does to remove this mm from all our lists now.
1211 remove_mm_from_lists(mm);
1212 up_read(&mm->mmap_sem);
1213 slot = ksm_scan.mm_slot;
1214 if (slot != &ksm_mm_head)
1220 * Nuke all the rmap_items that are above this current rmap:
1221 * because there were no VM_MERGEABLE vmas with such addresses.
1223 remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1224 up_read(&mm->mmap_sem);
1226 spin_lock(&ksm_mmlist_lock);
1227 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1228 ksm_scan.mm_slot = slot;
1229 spin_unlock(&ksm_mmlist_lock);
1231 /* Repeat until we've completed scanning the whole list */
1232 if (slot != &ksm_mm_head)
1236 * Bump seqnr here rather than at top, so that __ksm_exit
1237 * can skip rb_erase on unstable tree until we run again.
1244 * ksm_do_scan - the ksm scanner main worker function.
1245 * @scan_npages - number of pages we want to scan before we return.
1247 static void ksm_do_scan(unsigned int scan_npages)
1249 struct rmap_item *rmap_item;
1252 while (scan_npages--) {
1254 rmap_item = scan_get_next_rmap_item(&page);
1257 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1258 cmp_and_merge_page(page, rmap_item);
1263 static int ksm_scan_thread(void *nothing)
1265 set_user_nice(current, 0);
1267 while (!kthread_should_stop()) {
1268 if (ksm_run & KSM_RUN_MERGE) {
1269 mutex_lock(&ksm_thread_mutex);
1270 ksm_do_scan(ksm_thread_pages_to_scan);
1271 mutex_unlock(&ksm_thread_mutex);
1272 schedule_timeout_interruptible(
1273 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1275 wait_event_interruptible(ksm_thread_wait,
1276 (ksm_run & KSM_RUN_MERGE) ||
1277 kthread_should_stop());
1283 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1284 unsigned long end, int advice, unsigned long *vm_flags)
1286 struct mm_struct *mm = vma->vm_mm;
1289 case MADV_MERGEABLE:
1291 * Be somewhat over-protective for now!
1293 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1294 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1295 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1296 VM_MIXEDMAP | VM_SAO))
1297 return 0; /* just ignore the advice */
1299 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
1300 if (__ksm_enter(mm) < 0)
1303 *vm_flags |= VM_MERGEABLE;
1306 case MADV_UNMERGEABLE:
1307 if (!(*vm_flags & VM_MERGEABLE))
1308 return 0; /* just ignore the advice */
1311 unmerge_ksm_pages(vma, start, end);
1313 *vm_flags &= ~VM_MERGEABLE;
1320 int __ksm_enter(struct mm_struct *mm)
1322 struct mm_slot *mm_slot = alloc_mm_slot();
1326 spin_lock(&ksm_mmlist_lock);
1327 insert_to_mm_slots_hash(mm, mm_slot);
1329 * Insert just behind the scanning cursor, to let the area settle
1330 * down a little; when fork is followed by immediate exec, we don't
1331 * want ksmd to waste time setting up and tearing down an rmap_list.
1333 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1334 spin_unlock(&ksm_mmlist_lock);
1336 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1340 void __ksm_exit(struct mm_struct *mm)
1343 * This process is exiting: doesn't hold and doesn't need mmap_sem;
1344 * but we do need to exclude ksmd and other exiters while we modify
1345 * the various lists and trees.
1347 mutex_lock(&ksm_thread_mutex);
1348 remove_mm_from_lists(mm);
1349 mutex_unlock(&ksm_thread_mutex);
1352 #define KSM_ATTR_RO(_name) \
1353 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1354 #define KSM_ATTR(_name) \
1355 static struct kobj_attribute _name##_attr = \
1356 __ATTR(_name, 0644, _name##_show, _name##_store)
1358 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1359 struct kobj_attribute *attr, char *buf)
1361 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1364 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1365 struct kobj_attribute *attr,
1366 const char *buf, size_t count)
1368 unsigned long msecs;
1371 err = strict_strtoul(buf, 10, &msecs);
1372 if (err || msecs > UINT_MAX)
1375 ksm_thread_sleep_millisecs = msecs;
1379 KSM_ATTR(sleep_millisecs);
1381 static ssize_t pages_to_scan_show(struct kobject *kobj,
1382 struct kobj_attribute *attr, char *buf)
1384 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1387 static ssize_t pages_to_scan_store(struct kobject *kobj,
1388 struct kobj_attribute *attr,
1389 const char *buf, size_t count)
1392 unsigned long nr_pages;
1394 err = strict_strtoul(buf, 10, &nr_pages);
1395 if (err || nr_pages > UINT_MAX)
1398 ksm_thread_pages_to_scan = nr_pages;
1402 KSM_ATTR(pages_to_scan);
1404 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1407 return sprintf(buf, "%u\n", ksm_run);
1410 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1411 const char *buf, size_t count)
1414 unsigned long flags;
1416 err = strict_strtoul(buf, 10, &flags);
1417 if (err || flags > UINT_MAX)
1419 if (flags > KSM_RUN_UNMERGE)
1423 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1424 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1425 * breaking COW to free the kernel_pages_allocated (but leaves
1426 * mm_slots on the list for when ksmd may be set running again).
1429 mutex_lock(&ksm_thread_mutex);
1430 if (ksm_run != flags) {
1432 if (flags & KSM_RUN_UNMERGE)
1433 unmerge_and_remove_all_rmap_items();
1435 mutex_unlock(&ksm_thread_mutex);
1437 if (flags & KSM_RUN_MERGE)
1438 wake_up_interruptible(&ksm_thread_wait);
1444 static ssize_t pages_shared_show(struct kobject *kobj,
1445 struct kobj_attribute *attr, char *buf)
1447 return sprintf(buf, "%lu\n",
1448 ksm_pages_shared - ksm_kernel_pages_allocated);
1450 KSM_ATTR_RO(pages_shared);
1452 static ssize_t kernel_pages_allocated_show(struct kobject *kobj,
1453 struct kobj_attribute *attr,
1456 return sprintf(buf, "%lu\n", ksm_kernel_pages_allocated);
1458 KSM_ATTR_RO(kernel_pages_allocated);
1460 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1461 struct kobj_attribute *attr,
1462 const char *buf, size_t count)
1465 unsigned long nr_pages;
1467 err = strict_strtoul(buf, 10, &nr_pages);
1471 ksm_max_kernel_pages = nr_pages;
1476 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1477 struct kobj_attribute *attr, char *buf)
1479 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1481 KSM_ATTR(max_kernel_pages);
1483 static struct attribute *ksm_attrs[] = {
1484 &sleep_millisecs_attr.attr,
1485 &pages_to_scan_attr.attr,
1487 &pages_shared_attr.attr,
1488 &kernel_pages_allocated_attr.attr,
1489 &max_kernel_pages_attr.attr,
1493 static struct attribute_group ksm_attr_group = {
1498 static int __init ksm_init(void)
1500 struct task_struct *ksm_thread;
1503 err = ksm_slab_init();
1507 err = mm_slots_hash_init();
1511 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1512 if (IS_ERR(ksm_thread)) {
1513 printk(KERN_ERR "ksm: creating kthread failed\n");
1514 err = PTR_ERR(ksm_thread);
1518 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1520 printk(KERN_ERR "ksm: register sysfs failed\n");
1527 kthread_stop(ksm_thread);
1529 mm_slots_hash_free();
1535 module_init(ksm_init)