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/swap.h>
34 #include <linux/ksm.h>
36 #include <asm/tlbflush.h>
40 * A few notes about the KSM scanning process,
41 * to make it easier to understand the data structures below:
43 * In order to reduce excessive scanning, KSM sorts the memory pages by their
44 * contents into a data structure that holds pointers to the pages' locations.
46 * Since the contents of the pages may change at any moment, KSM cannot just
47 * insert the pages into a normal sorted tree and expect it to find anything.
48 * Therefore KSM uses two data structures - the stable and the unstable tree.
50 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
51 * by their contents. Because each such page is write-protected, searching on
52 * this tree is fully assured to be working (except when pages are unmapped),
53 * and therefore this tree is called the stable tree.
55 * In addition to the stable tree, KSM uses a second data structure called the
56 * unstable tree: this tree holds pointers to pages which have been found to
57 * be "unchanged for a period of time". The unstable tree sorts these pages
58 * by their contents, but since they are not write-protected, KSM cannot rely
59 * upon the unstable tree to work correctly - the unstable tree is liable to
60 * be corrupted as its contents are modified, and so it is called unstable.
62 * KSM solves this problem by several techniques:
64 * 1) The unstable tree is flushed every time KSM completes scanning all
65 * memory areas, and then the tree is rebuilt again from the beginning.
66 * 2) KSM will only insert into the unstable tree, pages whose hash value
67 * has not changed since the previous scan of all memory areas.
68 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
69 * colors of the nodes and not on their contents, assuring that even when
70 * the tree gets "corrupted" it won't get out of balance, so scanning time
71 * remains the same (also, searching and inserting nodes in an rbtree uses
72 * the same algorithm, so we have no overhead when we flush and rebuild).
73 * 4) KSM never flushes the stable tree, which means that even if it were to
74 * take 10 attempts to find a page in the unstable tree, once it is found,
75 * it is secured in the stable tree. (When we scan a new page, we first
76 * compare it against the stable tree, and then against the unstable tree.)
80 * struct mm_slot - ksm information per mm that is being scanned
81 * @link: link to the mm_slots hash list
82 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
83 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
84 * @mm: the mm that this information is valid for
87 struct hlist_node link;
88 struct list_head mm_list;
89 struct rmap_item *rmap_list;
94 * struct ksm_scan - cursor for scanning
95 * @mm_slot: the current mm_slot we are scanning
96 * @address: the next address inside that to be scanned
97 * @rmap_list: link to the next rmap to be scanned in the rmap_list
98 * @seqnr: count of completed full scans (needed when removing unstable node)
100 * There is only the one ksm_scan instance of this cursor structure.
103 struct mm_slot *mm_slot;
104 unsigned long address;
105 struct rmap_item **rmap_list;
110 * struct stable_node - node of the stable rbtree
111 * @page: pointer to struct page of the ksm page
112 * @node: rb node of this ksm page in the stable tree
113 * @hlist: hlist head of rmap_items using this ksm page
118 struct hlist_head hlist;
122 * struct rmap_item - reverse mapping item for virtual addresses
123 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
124 * @filler: unused space we're making available in this patch
125 * @mm: the memory structure this rmap_item is pointing into
126 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
127 * @oldchecksum: previous checksum of the page at that virtual address
128 * @node: rb node of this rmap_item in the unstable tree
129 * @head: pointer to stable_node heading this list in the stable tree
130 * @hlist: link into hlist of rmap_items hanging off that stable_node
133 struct rmap_item *rmap_list;
134 unsigned long filler;
135 struct mm_struct *mm;
136 unsigned long address; /* + low bits used for flags below */
137 unsigned int oldchecksum; /* when unstable */
139 struct rb_node node; /* when node of unstable tree */
140 struct { /* when listed from stable tree */
141 struct stable_node *head;
142 struct hlist_node hlist;
147 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
148 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
149 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
151 /* The stable and unstable tree heads */
152 static struct rb_root root_stable_tree = RB_ROOT;
153 static struct rb_root root_unstable_tree = RB_ROOT;
155 #define MM_SLOTS_HASH_HEADS 1024
156 static struct hlist_head *mm_slots_hash;
158 static struct mm_slot ksm_mm_head = {
159 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
161 static struct ksm_scan ksm_scan = {
162 .mm_slot = &ksm_mm_head,
165 static struct kmem_cache *rmap_item_cache;
166 static struct kmem_cache *stable_node_cache;
167 static struct kmem_cache *mm_slot_cache;
169 /* The number of nodes in the stable tree */
170 static unsigned long ksm_pages_shared;
172 /* The number of page slots additionally sharing those nodes */
173 static unsigned long ksm_pages_sharing;
175 /* The number of nodes in the unstable tree */
176 static unsigned long ksm_pages_unshared;
178 /* The number of rmap_items in use: to calculate pages_volatile */
179 static unsigned long ksm_rmap_items;
181 /* Limit on the number of unswappable pages used */
182 static unsigned long ksm_max_kernel_pages;
184 /* Number of pages ksmd should scan in one batch */
185 static unsigned int ksm_thread_pages_to_scan = 100;
187 /* Milliseconds ksmd should sleep between batches */
188 static unsigned int ksm_thread_sleep_millisecs = 20;
190 #define KSM_RUN_STOP 0
191 #define KSM_RUN_MERGE 1
192 #define KSM_RUN_UNMERGE 2
193 static unsigned int ksm_run = KSM_RUN_STOP;
195 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
196 static DEFINE_MUTEX(ksm_thread_mutex);
197 static DEFINE_SPINLOCK(ksm_mmlist_lock);
199 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
200 sizeof(struct __struct), __alignof__(struct __struct),\
203 static int __init ksm_slab_init(void)
205 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
206 if (!rmap_item_cache)
209 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
210 if (!stable_node_cache)
213 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
220 kmem_cache_destroy(stable_node_cache);
222 kmem_cache_destroy(rmap_item_cache);
227 static void __init ksm_slab_free(void)
229 kmem_cache_destroy(mm_slot_cache);
230 kmem_cache_destroy(stable_node_cache);
231 kmem_cache_destroy(rmap_item_cache);
232 mm_slot_cache = NULL;
235 static inline struct rmap_item *alloc_rmap_item(void)
237 struct rmap_item *rmap_item;
239 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
245 static inline void free_rmap_item(struct rmap_item *rmap_item)
248 rmap_item->mm = NULL; /* debug safety */
249 kmem_cache_free(rmap_item_cache, rmap_item);
252 static inline struct stable_node *alloc_stable_node(void)
254 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
257 static inline void free_stable_node(struct stable_node *stable_node)
259 kmem_cache_free(stable_node_cache, stable_node);
262 static inline struct mm_slot *alloc_mm_slot(void)
264 if (!mm_slot_cache) /* initialization failed */
266 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
269 static inline void free_mm_slot(struct mm_slot *mm_slot)
271 kmem_cache_free(mm_slot_cache, mm_slot);
274 static int __init mm_slots_hash_init(void)
276 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
283 static void __init mm_slots_hash_free(void)
285 kfree(mm_slots_hash);
288 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
290 struct mm_slot *mm_slot;
291 struct hlist_head *bucket;
292 struct hlist_node *node;
294 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
295 % MM_SLOTS_HASH_HEADS];
296 hlist_for_each_entry(mm_slot, node, bucket, link) {
297 if (mm == mm_slot->mm)
303 static void insert_to_mm_slots_hash(struct mm_struct *mm,
304 struct mm_slot *mm_slot)
306 struct hlist_head *bucket;
308 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
309 % MM_SLOTS_HASH_HEADS];
311 hlist_add_head(&mm_slot->link, bucket);
314 static inline int in_stable_tree(struct rmap_item *rmap_item)
316 return rmap_item->address & STABLE_FLAG;
320 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
321 * page tables after it has passed through ksm_exit() - which, if necessary,
322 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
323 * a special flag: they can just back out as soon as mm_users goes to zero.
324 * ksm_test_exit() is used throughout to make this test for exit: in some
325 * places for correctness, in some places just to avoid unnecessary work.
327 static inline bool ksm_test_exit(struct mm_struct *mm)
329 return atomic_read(&mm->mm_users) == 0;
333 * We use break_ksm to break COW on a ksm page: it's a stripped down
335 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
338 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
339 * in case the application has unmapped and remapped mm,addr meanwhile.
340 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
341 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
343 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
350 page = follow_page(vma, addr, FOLL_GET);
354 ret = handle_mm_fault(vma->vm_mm, vma, addr,
357 ret = VM_FAULT_WRITE;
359 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
361 * We must loop because handle_mm_fault() may back out if there's
362 * any difficulty e.g. if pte accessed bit gets updated concurrently.
364 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
365 * COW has been broken, even if the vma does not permit VM_WRITE;
366 * but note that a concurrent fault might break PageKsm for us.
368 * VM_FAULT_SIGBUS could occur if we race with truncation of the
369 * backing file, which also invalidates anonymous pages: that's
370 * okay, that truncation will have unmapped the PageKsm for us.
372 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
373 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
374 * current task has TIF_MEMDIE set, and will be OOM killed on return
375 * to user; and ksmd, having no mm, would never be chosen for that.
377 * But if the mm is in a limited mem_cgroup, then the fault may fail
378 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
379 * even ksmd can fail in this way - though it's usually breaking ksm
380 * just to undo a merge it made a moment before, so unlikely to oom.
382 * That's a pity: we might therefore have more kernel pages allocated
383 * than we're counting as nodes in the stable tree; but ksm_do_scan
384 * will retry to break_cow on each pass, so should recover the page
385 * in due course. The important thing is to not let VM_MERGEABLE
386 * be cleared while any such pages might remain in the area.
388 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
391 static void break_cow(struct rmap_item *rmap_item)
393 struct mm_struct *mm = rmap_item->mm;
394 unsigned long addr = rmap_item->address;
395 struct vm_area_struct *vma;
397 down_read(&mm->mmap_sem);
398 if (ksm_test_exit(mm))
400 vma = find_vma(mm, addr);
401 if (!vma || vma->vm_start > addr)
403 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
405 break_ksm(vma, addr);
407 up_read(&mm->mmap_sem);
410 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
412 struct mm_struct *mm = rmap_item->mm;
413 unsigned long addr = rmap_item->address;
414 struct vm_area_struct *vma;
417 down_read(&mm->mmap_sem);
418 if (ksm_test_exit(mm))
420 vma = find_vma(mm, addr);
421 if (!vma || vma->vm_start > addr)
423 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
426 page = follow_page(vma, addr, FOLL_GET);
429 if (PageAnon(page)) {
430 flush_anon_page(vma, page, addr);
431 flush_dcache_page(page);
436 up_read(&mm->mmap_sem);
441 * Removing rmap_item from stable or unstable tree.
442 * This function will clean the information from the stable/unstable tree.
444 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
446 if (rmap_item->address & STABLE_FLAG) {
447 struct stable_node *stable_node;
449 stable_node = rmap_item->head;
450 hlist_del(&rmap_item->hlist);
451 if (stable_node->hlist.first)
454 set_page_stable_node(stable_node->page, NULL);
455 put_page(stable_node->page);
457 rb_erase(&stable_node->node, &root_stable_tree);
458 free_stable_node(stable_node);
462 rmap_item->address &= PAGE_MASK;
464 } else if (rmap_item->address & UNSTABLE_FLAG) {
467 * Usually ksmd can and must skip the rb_erase, because
468 * root_unstable_tree was already reset to RB_ROOT.
469 * But be careful when an mm is exiting: do the rb_erase
470 * if this rmap_item was inserted by this scan, rather
471 * than left over from before.
473 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
476 rb_erase(&rmap_item->node, &root_unstable_tree);
478 ksm_pages_unshared--;
479 rmap_item->address &= PAGE_MASK;
482 cond_resched(); /* we're called from many long loops */
485 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
486 struct rmap_item **rmap_list)
489 struct rmap_item *rmap_item = *rmap_list;
490 *rmap_list = rmap_item->rmap_list;
491 remove_rmap_item_from_tree(rmap_item);
492 free_rmap_item(rmap_item);
497 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
498 * than check every pte of a given vma, the locking doesn't quite work for
499 * that - an rmap_item is assigned to the stable tree after inserting ksm
500 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
501 * rmap_items from parent to child at fork time (so as not to waste time
502 * if exit comes before the next scan reaches it).
504 * Similarly, although we'd like to remove rmap_items (so updating counts
505 * and freeing memory) when unmerging an area, it's easier to leave that
506 * to the next pass of ksmd - consider, for example, how ksmd might be
507 * in cmp_and_merge_page on one of the rmap_items we would be removing.
509 static int unmerge_ksm_pages(struct vm_area_struct *vma,
510 unsigned long start, unsigned long end)
515 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
516 if (ksm_test_exit(vma->vm_mm))
518 if (signal_pending(current))
521 err = break_ksm(vma, addr);
528 * Only called through the sysfs control interface:
530 static int unmerge_and_remove_all_rmap_items(void)
532 struct mm_slot *mm_slot;
533 struct mm_struct *mm;
534 struct vm_area_struct *vma;
537 spin_lock(&ksm_mmlist_lock);
538 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
539 struct mm_slot, mm_list);
540 spin_unlock(&ksm_mmlist_lock);
542 for (mm_slot = ksm_scan.mm_slot;
543 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
545 down_read(&mm->mmap_sem);
546 for (vma = mm->mmap; vma; vma = vma->vm_next) {
547 if (ksm_test_exit(mm))
549 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
551 err = unmerge_ksm_pages(vma,
552 vma->vm_start, vma->vm_end);
557 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
559 spin_lock(&ksm_mmlist_lock);
560 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
561 struct mm_slot, mm_list);
562 if (ksm_test_exit(mm)) {
563 hlist_del(&mm_slot->link);
564 list_del(&mm_slot->mm_list);
565 spin_unlock(&ksm_mmlist_lock);
567 free_mm_slot(mm_slot);
568 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
569 up_read(&mm->mmap_sem);
572 spin_unlock(&ksm_mmlist_lock);
573 up_read(&mm->mmap_sem);
581 up_read(&mm->mmap_sem);
582 spin_lock(&ksm_mmlist_lock);
583 ksm_scan.mm_slot = &ksm_mm_head;
584 spin_unlock(&ksm_mmlist_lock);
587 #endif /* CONFIG_SYSFS */
589 static u32 calc_checksum(struct page *page)
592 void *addr = kmap_atomic(page, KM_USER0);
593 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
594 kunmap_atomic(addr, KM_USER0);
598 static int memcmp_pages(struct page *page1, struct page *page2)
603 addr1 = kmap_atomic(page1, KM_USER0);
604 addr2 = kmap_atomic(page2, KM_USER1);
605 ret = memcmp(addr1, addr2, PAGE_SIZE);
606 kunmap_atomic(addr2, KM_USER1);
607 kunmap_atomic(addr1, KM_USER0);
611 static inline int pages_identical(struct page *page1, struct page *page2)
613 return !memcmp_pages(page1, page2);
616 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
619 struct mm_struct *mm = vma->vm_mm;
626 addr = page_address_in_vma(page, vma);
630 ptep = page_check_address(page, mm, addr, &ptl, 0);
634 if (pte_write(*ptep)) {
637 swapped = PageSwapCache(page);
638 flush_cache_page(vma, addr, page_to_pfn(page));
640 * Ok this is tricky, when get_user_pages_fast() run it doesnt
641 * take any lock, therefore the check that we are going to make
642 * with the pagecount against the mapcount is racey and
643 * O_DIRECT can happen right after the check.
644 * So we clear the pte and flush the tlb before the check
645 * this assure us that no O_DIRECT can happen after the check
646 * or in the middle of the check.
648 entry = ptep_clear_flush(vma, addr, ptep);
650 * Check that no O_DIRECT or similar I/O is in progress on the
653 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
654 set_pte_at_notify(mm, addr, ptep, entry);
657 entry = pte_wrprotect(entry);
658 set_pte_at_notify(mm, addr, ptep, entry);
664 pte_unmap_unlock(ptep, ptl);
670 * replace_page - replace page in vma by new ksm page
671 * @vma: vma that holds the pte pointing to page
672 * @page: the page we are replacing by kpage
673 * @kpage: the ksm page we replace page by
674 * @orig_pte: the original value of the pte
676 * Returns 0 on success, -EFAULT on failure.
678 static int replace_page(struct vm_area_struct *vma, struct page *page,
679 struct page *kpage, pte_t orig_pte)
681 struct mm_struct *mm = vma->vm_mm;
690 addr = page_address_in_vma(page, vma);
694 pgd = pgd_offset(mm, addr);
695 if (!pgd_present(*pgd))
698 pud = pud_offset(pgd, addr);
699 if (!pud_present(*pud))
702 pmd = pmd_offset(pud, addr);
703 if (!pmd_present(*pmd))
706 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
707 if (!pte_same(*ptep, orig_pte)) {
708 pte_unmap_unlock(ptep, ptl);
713 page_add_ksm_rmap(kpage);
715 flush_cache_page(vma, addr, pte_pfn(*ptep));
716 ptep_clear_flush(vma, addr, ptep);
717 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
719 page_remove_rmap(page);
722 pte_unmap_unlock(ptep, ptl);
729 * try_to_merge_one_page - take two pages and merge them into one
730 * @vma: the vma that holds the pte pointing to page
731 * @page: the PageAnon page that we want to replace with kpage
732 * @kpage: the PageKsm page that we want to map instead of page
734 * This function returns 0 if the pages were merged, -EFAULT otherwise.
736 static int try_to_merge_one_page(struct vm_area_struct *vma,
737 struct page *page, struct page *kpage)
739 pte_t orig_pte = __pte(0);
742 if (!(vma->vm_flags & VM_MERGEABLE))
748 * We need the page lock to read a stable PageSwapCache in
749 * write_protect_page(). We use trylock_page() instead of
750 * lock_page() because we don't want to wait here - we
751 * prefer to continue scanning and merging different pages,
752 * then come back to this page when it is unlocked.
754 if (!trylock_page(page))
757 * If this anonymous page is mapped only here, its pte may need
758 * to be write-protected. If it's mapped elsewhere, all of its
759 * ptes are necessarily already write-protected. But in either
760 * case, we need to lock and check page_count is not raised.
762 if (write_protect_page(vma, page, &orig_pte) == 0 &&
763 pages_identical(page, kpage))
764 err = replace_page(vma, page, kpage, orig_pte);
766 if ((vma->vm_flags & VM_LOCKED) && !err)
767 munlock_vma_page(page);
775 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
776 * but no new kernel page is allocated: kpage must already be a ksm page.
778 * This function returns 0 if the pages were merged, -EFAULT otherwise.
780 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
781 struct page *page, struct page *kpage)
783 struct mm_struct *mm = rmap_item->mm;
784 struct vm_area_struct *vma;
787 if (page == kpage) /* ksm page forked */
790 down_read(&mm->mmap_sem);
791 if (ksm_test_exit(mm))
793 vma = find_vma(mm, rmap_item->address);
794 if (!vma || vma->vm_start > rmap_item->address)
797 err = try_to_merge_one_page(vma, page, kpage);
799 up_read(&mm->mmap_sem);
804 * try_to_merge_two_pages - take two identical pages and prepare them
805 * to be merged into one page.
807 * This function returns the kpage if we successfully merged two identical
808 * pages into one ksm page, NULL otherwise.
810 * Note that this function allocates a new kernel page: if one of the pages
811 * is already a ksm page, try_to_merge_with_ksm_page should be used.
813 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
815 struct rmap_item *tree_rmap_item,
816 struct page *tree_page)
818 struct mm_struct *mm = rmap_item->mm;
819 struct vm_area_struct *vma;
824 * The number of nodes in the stable tree
825 * is the number of kernel pages that we hold.
827 if (ksm_max_kernel_pages &&
828 ksm_max_kernel_pages <= ksm_pages_shared)
831 kpage = alloc_page(GFP_HIGHUSER);
835 down_read(&mm->mmap_sem);
836 if (ksm_test_exit(mm))
838 vma = find_vma(mm, rmap_item->address);
839 if (!vma || vma->vm_start > rmap_item->address)
842 copy_user_highpage(kpage, page, rmap_item->address, vma);
844 set_page_stable_node(kpage, NULL); /* mark it PageKsm */
846 err = try_to_merge_one_page(vma, page, kpage);
848 up_read(&mm->mmap_sem);
851 err = try_to_merge_with_ksm_page(tree_rmap_item,
854 * If that fails, we have a ksm page with only one pte
855 * pointing to it: so break it.
858 break_cow(rmap_item);
868 * stable_tree_search - search for page inside the stable tree
870 * This function checks if there is a page inside the stable tree
871 * with identical content to the page that we are scanning right now.
873 * This function returns the stable tree node of identical content if found,
876 static struct stable_node *stable_tree_search(struct page *page)
878 struct rb_node *node = root_stable_tree.rb_node;
879 struct stable_node *stable_node;
881 stable_node = page_stable_node(page);
882 if (stable_node) { /* ksm page forked */
891 stable_node = rb_entry(node, struct stable_node, node);
893 ret = memcmp_pages(page, stable_node->page);
896 node = node->rb_left;
898 node = node->rb_right;
900 get_page(stable_node->page);
909 * stable_tree_insert - insert rmap_item pointing to new ksm page
910 * into the stable tree.
912 * This function returns the stable tree node just allocated on success,
915 static struct stable_node *stable_tree_insert(struct page *kpage)
917 struct rb_node **new = &root_stable_tree.rb_node;
918 struct rb_node *parent = NULL;
919 struct stable_node *stable_node;
925 stable_node = rb_entry(*new, struct stable_node, node);
927 ret = memcmp_pages(kpage, stable_node->page);
931 new = &parent->rb_left;
933 new = &parent->rb_right;
936 * It is not a bug that stable_tree_search() didn't
937 * find this node: because at that time our page was
938 * not yet write-protected, so may have changed since.
944 stable_node = alloc_stable_node();
948 rb_link_node(&stable_node->node, parent, new);
949 rb_insert_color(&stable_node->node, &root_stable_tree);
951 INIT_HLIST_HEAD(&stable_node->hlist);
954 stable_node->page = kpage;
955 set_page_stable_node(kpage, stable_node);
961 * unstable_tree_search_insert - search for identical page,
962 * else insert rmap_item into the unstable tree.
964 * This function searches for a page in the unstable tree identical to the
965 * page currently being scanned; and if no identical page is found in the
966 * tree, we insert rmap_item as a new object into the unstable tree.
968 * This function returns pointer to rmap_item found to be identical
969 * to the currently scanned page, NULL otherwise.
971 * This function does both searching and inserting, because they share
972 * the same walking algorithm in an rbtree.
975 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
977 struct page **tree_pagep)
980 struct rb_node **new = &root_unstable_tree.rb_node;
981 struct rb_node *parent = NULL;
984 struct rmap_item *tree_rmap_item;
985 struct page *tree_page;
989 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
990 tree_page = get_mergeable_page(tree_rmap_item);
995 * Don't substitute a ksm page for a forked page.
997 if (page == tree_page) {
1002 ret = memcmp_pages(page, tree_page);
1006 put_page(tree_page);
1007 new = &parent->rb_left;
1008 } else if (ret > 0) {
1009 put_page(tree_page);
1010 new = &parent->rb_right;
1012 *tree_pagep = tree_page;
1013 return tree_rmap_item;
1017 rmap_item->address |= UNSTABLE_FLAG;
1018 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1019 rb_link_node(&rmap_item->node, parent, new);
1020 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1022 ksm_pages_unshared++;
1027 * stable_tree_append - add another rmap_item to the linked list of
1028 * rmap_items hanging off a given node of the stable tree, all sharing
1029 * the same ksm page.
1031 static void stable_tree_append(struct rmap_item *rmap_item,
1032 struct stable_node *stable_node)
1034 rmap_item->head = stable_node;
1035 rmap_item->address |= STABLE_FLAG;
1036 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1038 if (rmap_item->hlist.next)
1039 ksm_pages_sharing++;
1045 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1046 * if not, compare checksum to previous and if it's the same, see if page can
1047 * be inserted into the unstable tree, or merged with a page already there and
1048 * both transferred to the stable tree.
1050 * @page: the page that we are searching identical page to.
1051 * @rmap_item: the reverse mapping into the virtual address of this page
1053 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1055 struct rmap_item *tree_rmap_item;
1056 struct page *tree_page = NULL;
1057 struct stable_node *stable_node;
1059 unsigned int checksum;
1062 remove_rmap_item_from_tree(rmap_item);
1064 /* We first start with searching the page inside the stable tree */
1065 stable_node = stable_tree_search(page);
1067 kpage = stable_node->page;
1068 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1071 * The page was successfully merged:
1072 * add its rmap_item to the stable tree.
1074 stable_tree_append(rmap_item, stable_node);
1081 * A ksm page might have got here by fork, but its other
1082 * references have already been removed from the stable tree.
1083 * Or it might be left over from a break_ksm which failed
1084 * when the mem_cgroup had reached its limit: try again now.
1087 break_cow(rmap_item);
1090 * In case the hash value of the page was changed from the last time we
1091 * have calculated it, this page to be changed frequely, therefore we
1092 * don't want to insert it to the unstable tree, and we don't want to
1093 * waste our time to search if there is something identical to it there.
1095 checksum = calc_checksum(page);
1096 if (rmap_item->oldchecksum != checksum) {
1097 rmap_item->oldchecksum = checksum;
1102 unstable_tree_search_insert(rmap_item, page, &tree_page);
1103 if (tree_rmap_item) {
1104 kpage = try_to_merge_two_pages(rmap_item, page,
1105 tree_rmap_item, tree_page);
1106 put_page(tree_page);
1108 * As soon as we merge this page, we want to remove the
1109 * rmap_item of the page we have merged with from the unstable
1110 * tree, and insert it instead as new node in the stable tree.
1113 remove_rmap_item_from_tree(tree_rmap_item);
1115 stable_node = stable_tree_insert(kpage);
1117 stable_tree_append(tree_rmap_item, stable_node);
1118 stable_tree_append(rmap_item, stable_node);
1123 * If we fail to insert the page into the stable tree,
1124 * we will have 2 virtual addresses that are pointing
1125 * to a ksm page left outside the stable tree,
1126 * in which case we need to break_cow on both.
1129 break_cow(tree_rmap_item);
1130 break_cow(rmap_item);
1136 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1137 struct rmap_item **rmap_list,
1140 struct rmap_item *rmap_item;
1142 while (*rmap_list) {
1143 rmap_item = *rmap_list;
1144 if ((rmap_item->address & PAGE_MASK) == addr)
1146 if (rmap_item->address > addr)
1148 *rmap_list = rmap_item->rmap_list;
1149 remove_rmap_item_from_tree(rmap_item);
1150 free_rmap_item(rmap_item);
1153 rmap_item = alloc_rmap_item();
1155 /* It has already been zeroed */
1156 rmap_item->mm = mm_slot->mm;
1157 rmap_item->address = addr;
1158 rmap_item->rmap_list = *rmap_list;
1159 *rmap_list = rmap_item;
1164 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1166 struct mm_struct *mm;
1167 struct mm_slot *slot;
1168 struct vm_area_struct *vma;
1169 struct rmap_item *rmap_item;
1171 if (list_empty(&ksm_mm_head.mm_list))
1174 slot = ksm_scan.mm_slot;
1175 if (slot == &ksm_mm_head) {
1176 root_unstable_tree = RB_ROOT;
1178 spin_lock(&ksm_mmlist_lock);
1179 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1180 ksm_scan.mm_slot = slot;
1181 spin_unlock(&ksm_mmlist_lock);
1183 ksm_scan.address = 0;
1184 ksm_scan.rmap_list = &slot->rmap_list;
1188 down_read(&mm->mmap_sem);
1189 if (ksm_test_exit(mm))
1192 vma = find_vma(mm, ksm_scan.address);
1194 for (; vma; vma = vma->vm_next) {
1195 if (!(vma->vm_flags & VM_MERGEABLE))
1197 if (ksm_scan.address < vma->vm_start)
1198 ksm_scan.address = vma->vm_start;
1200 ksm_scan.address = vma->vm_end;
1202 while (ksm_scan.address < vma->vm_end) {
1203 if (ksm_test_exit(mm))
1205 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1206 if (*page && PageAnon(*page)) {
1207 flush_anon_page(vma, *page, ksm_scan.address);
1208 flush_dcache_page(*page);
1209 rmap_item = get_next_rmap_item(slot,
1210 ksm_scan.rmap_list, ksm_scan.address);
1212 ksm_scan.rmap_list =
1213 &rmap_item->rmap_list;
1214 ksm_scan.address += PAGE_SIZE;
1217 up_read(&mm->mmap_sem);
1222 ksm_scan.address += PAGE_SIZE;
1227 if (ksm_test_exit(mm)) {
1228 ksm_scan.address = 0;
1229 ksm_scan.rmap_list = &slot->rmap_list;
1232 * Nuke all the rmap_items that are above this current rmap:
1233 * because there were no VM_MERGEABLE vmas with such addresses.
1235 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1237 spin_lock(&ksm_mmlist_lock);
1238 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1239 struct mm_slot, mm_list);
1240 if (ksm_scan.address == 0) {
1242 * We've completed a full scan of all vmas, holding mmap_sem
1243 * throughout, and found no VM_MERGEABLE: so do the same as
1244 * __ksm_exit does to remove this mm from all our lists now.
1245 * This applies either when cleaning up after __ksm_exit
1246 * (but beware: we can reach here even before __ksm_exit),
1247 * or when all VM_MERGEABLE areas have been unmapped (and
1248 * mmap_sem then protects against race with MADV_MERGEABLE).
1250 hlist_del(&slot->link);
1251 list_del(&slot->mm_list);
1252 spin_unlock(&ksm_mmlist_lock);
1255 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1256 up_read(&mm->mmap_sem);
1259 spin_unlock(&ksm_mmlist_lock);
1260 up_read(&mm->mmap_sem);
1263 /* Repeat until we've completed scanning the whole list */
1264 slot = ksm_scan.mm_slot;
1265 if (slot != &ksm_mm_head)
1273 * ksm_do_scan - the ksm scanner main worker function.
1274 * @scan_npages - number of pages we want to scan before we return.
1276 static void ksm_do_scan(unsigned int scan_npages)
1278 struct rmap_item *rmap_item;
1281 while (scan_npages--) {
1283 rmap_item = scan_get_next_rmap_item(&page);
1286 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1287 cmp_and_merge_page(page, rmap_item);
1288 else if (page_mapcount(page) == 1) {
1290 * Replace now-unshared ksm page by ordinary page.
1292 break_cow(rmap_item);
1293 remove_rmap_item_from_tree(rmap_item);
1294 rmap_item->oldchecksum = calc_checksum(page);
1300 static int ksmd_should_run(void)
1302 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1305 static int ksm_scan_thread(void *nothing)
1307 set_user_nice(current, 5);
1309 while (!kthread_should_stop()) {
1310 mutex_lock(&ksm_thread_mutex);
1311 if (ksmd_should_run())
1312 ksm_do_scan(ksm_thread_pages_to_scan);
1313 mutex_unlock(&ksm_thread_mutex);
1315 if (ksmd_should_run()) {
1316 schedule_timeout_interruptible(
1317 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1319 wait_event_interruptible(ksm_thread_wait,
1320 ksmd_should_run() || kthread_should_stop());
1326 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1327 unsigned long end, int advice, unsigned long *vm_flags)
1329 struct mm_struct *mm = vma->vm_mm;
1333 case MADV_MERGEABLE:
1335 * Be somewhat over-protective for now!
1337 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1338 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1339 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1340 VM_MIXEDMAP | VM_SAO))
1341 return 0; /* just ignore the advice */
1343 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1344 err = __ksm_enter(mm);
1349 *vm_flags |= VM_MERGEABLE;
1352 case MADV_UNMERGEABLE:
1353 if (!(*vm_flags & VM_MERGEABLE))
1354 return 0; /* just ignore the advice */
1356 if (vma->anon_vma) {
1357 err = unmerge_ksm_pages(vma, start, end);
1362 *vm_flags &= ~VM_MERGEABLE;
1369 int __ksm_enter(struct mm_struct *mm)
1371 struct mm_slot *mm_slot;
1374 mm_slot = alloc_mm_slot();
1378 /* Check ksm_run too? Would need tighter locking */
1379 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1381 spin_lock(&ksm_mmlist_lock);
1382 insert_to_mm_slots_hash(mm, mm_slot);
1384 * Insert just behind the scanning cursor, to let the area settle
1385 * down a little; when fork is followed by immediate exec, we don't
1386 * want ksmd to waste time setting up and tearing down an rmap_list.
1388 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1389 spin_unlock(&ksm_mmlist_lock);
1391 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1392 atomic_inc(&mm->mm_count);
1395 wake_up_interruptible(&ksm_thread_wait);
1400 void __ksm_exit(struct mm_struct *mm)
1402 struct mm_slot *mm_slot;
1403 int easy_to_free = 0;
1406 * This process is exiting: if it's straightforward (as is the
1407 * case when ksmd was never running), free mm_slot immediately.
1408 * But if it's at the cursor or has rmap_items linked to it, use
1409 * mmap_sem to synchronize with any break_cows before pagetables
1410 * are freed, and leave the mm_slot on the list for ksmd to free.
1411 * Beware: ksm may already have noticed it exiting and freed the slot.
1414 spin_lock(&ksm_mmlist_lock);
1415 mm_slot = get_mm_slot(mm);
1416 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1417 if (!mm_slot->rmap_list) {
1418 hlist_del(&mm_slot->link);
1419 list_del(&mm_slot->mm_list);
1422 list_move(&mm_slot->mm_list,
1423 &ksm_scan.mm_slot->mm_list);
1426 spin_unlock(&ksm_mmlist_lock);
1429 free_mm_slot(mm_slot);
1430 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1432 } else if (mm_slot) {
1433 down_write(&mm->mmap_sem);
1434 up_write(&mm->mmap_sem);
1440 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1443 #define KSM_ATTR_RO(_name) \
1444 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1445 #define KSM_ATTR(_name) \
1446 static struct kobj_attribute _name##_attr = \
1447 __ATTR(_name, 0644, _name##_show, _name##_store)
1449 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1450 struct kobj_attribute *attr, char *buf)
1452 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1455 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1456 struct kobj_attribute *attr,
1457 const char *buf, size_t count)
1459 unsigned long msecs;
1462 err = strict_strtoul(buf, 10, &msecs);
1463 if (err || msecs > UINT_MAX)
1466 ksm_thread_sleep_millisecs = msecs;
1470 KSM_ATTR(sleep_millisecs);
1472 static ssize_t pages_to_scan_show(struct kobject *kobj,
1473 struct kobj_attribute *attr, char *buf)
1475 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1478 static ssize_t pages_to_scan_store(struct kobject *kobj,
1479 struct kobj_attribute *attr,
1480 const char *buf, size_t count)
1483 unsigned long nr_pages;
1485 err = strict_strtoul(buf, 10, &nr_pages);
1486 if (err || nr_pages > UINT_MAX)
1489 ksm_thread_pages_to_scan = nr_pages;
1493 KSM_ATTR(pages_to_scan);
1495 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1498 return sprintf(buf, "%u\n", ksm_run);
1501 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1502 const char *buf, size_t count)
1505 unsigned long flags;
1507 err = strict_strtoul(buf, 10, &flags);
1508 if (err || flags > UINT_MAX)
1510 if (flags > KSM_RUN_UNMERGE)
1514 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1515 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1516 * breaking COW to free the unswappable pages_shared (but leaves
1517 * mm_slots on the list for when ksmd may be set running again).
1520 mutex_lock(&ksm_thread_mutex);
1521 if (ksm_run != flags) {
1523 if (flags & KSM_RUN_UNMERGE) {
1524 current->flags |= PF_OOM_ORIGIN;
1525 err = unmerge_and_remove_all_rmap_items();
1526 current->flags &= ~PF_OOM_ORIGIN;
1528 ksm_run = KSM_RUN_STOP;
1533 mutex_unlock(&ksm_thread_mutex);
1535 if (flags & KSM_RUN_MERGE)
1536 wake_up_interruptible(&ksm_thread_wait);
1542 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1543 struct kobj_attribute *attr,
1544 const char *buf, size_t count)
1547 unsigned long nr_pages;
1549 err = strict_strtoul(buf, 10, &nr_pages);
1553 ksm_max_kernel_pages = nr_pages;
1558 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1559 struct kobj_attribute *attr, char *buf)
1561 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1563 KSM_ATTR(max_kernel_pages);
1565 static ssize_t pages_shared_show(struct kobject *kobj,
1566 struct kobj_attribute *attr, char *buf)
1568 return sprintf(buf, "%lu\n", ksm_pages_shared);
1570 KSM_ATTR_RO(pages_shared);
1572 static ssize_t pages_sharing_show(struct kobject *kobj,
1573 struct kobj_attribute *attr, char *buf)
1575 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1577 KSM_ATTR_RO(pages_sharing);
1579 static ssize_t pages_unshared_show(struct kobject *kobj,
1580 struct kobj_attribute *attr, char *buf)
1582 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1584 KSM_ATTR_RO(pages_unshared);
1586 static ssize_t pages_volatile_show(struct kobject *kobj,
1587 struct kobj_attribute *attr, char *buf)
1589 long ksm_pages_volatile;
1591 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1592 - ksm_pages_sharing - ksm_pages_unshared;
1594 * It was not worth any locking to calculate that statistic,
1595 * but it might therefore sometimes be negative: conceal that.
1597 if (ksm_pages_volatile < 0)
1598 ksm_pages_volatile = 0;
1599 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1601 KSM_ATTR_RO(pages_volatile);
1603 static ssize_t full_scans_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1606 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1608 KSM_ATTR_RO(full_scans);
1610 static struct attribute *ksm_attrs[] = {
1611 &sleep_millisecs_attr.attr,
1612 &pages_to_scan_attr.attr,
1614 &max_kernel_pages_attr.attr,
1615 &pages_shared_attr.attr,
1616 &pages_sharing_attr.attr,
1617 &pages_unshared_attr.attr,
1618 &pages_volatile_attr.attr,
1619 &full_scans_attr.attr,
1623 static struct attribute_group ksm_attr_group = {
1627 #endif /* CONFIG_SYSFS */
1629 static int __init ksm_init(void)
1631 struct task_struct *ksm_thread;
1634 ksm_max_kernel_pages = totalram_pages / 4;
1636 err = ksm_slab_init();
1640 err = mm_slots_hash_init();
1644 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1645 if (IS_ERR(ksm_thread)) {
1646 printk(KERN_ERR "ksm: creating kthread failed\n");
1647 err = PTR_ERR(ksm_thread);
1652 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1654 printk(KERN_ERR "ksm: register sysfs failed\n");
1655 kthread_stop(ksm_thread);
1659 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
1661 #endif /* CONFIG_SYSFS */
1666 mm_slots_hash_free();
1672 module_init(ksm_init)