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>
39 * A few notes about the KSM scanning process,
40 * to make it easier to understand the data structures below:
42 * In order to reduce excessive scanning, KSM sorts the memory pages by their
43 * contents into a data structure that holds pointers to the pages' locations.
45 * Since the contents of the pages may change at any moment, KSM cannot just
46 * insert the pages into a normal sorted tree and expect it to find anything.
47 * Therefore KSM uses two data structures - the stable and the unstable tree.
49 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
50 * by their contents. Because each such page is write-protected, searching on
51 * this tree is fully assured to be working (except when pages are unmapped),
52 * and therefore this tree is called the stable tree.
54 * In addition to the stable tree, KSM uses a second data structure called the
55 * unstable tree: this tree holds pointers to pages which have been found to
56 * be "unchanged for a period of time". The unstable tree sorts these pages
57 * by their contents, but since they are not write-protected, KSM cannot rely
58 * upon the unstable tree to work correctly - the unstable tree is liable to
59 * be corrupted as its contents are modified, and so it is called unstable.
61 * KSM solves this problem by several techniques:
63 * 1) The unstable tree is flushed every time KSM completes scanning all
64 * memory areas, and then the tree is rebuilt again from the beginning.
65 * 2) KSM will only insert into the unstable tree, pages whose hash value
66 * has not changed since the previous scan of all memory areas.
67 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
68 * colors of the nodes and not on their contents, assuring that even when
69 * the tree gets "corrupted" it won't get out of balance, so scanning time
70 * remains the same (also, searching and inserting nodes in an rbtree uses
71 * the same algorithm, so we have no overhead when we flush and rebuild).
72 * 4) KSM never flushes the stable tree, which means that even if it were to
73 * take 10 attempts to find a page in the unstable tree, once it is found,
74 * it is secured in the stable tree. (When we scan a new page, we first
75 * compare it against the stable tree, and then against the unstable tree.)
79 * struct mm_slot - ksm information per mm that is being scanned
80 * @link: link to the mm_slots hash list
81 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
82 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
83 * @mm: the mm that this information is valid for
86 struct hlist_node link;
87 struct list_head mm_list;
88 struct rmap_item *rmap_list;
93 * struct ksm_scan - cursor for scanning
94 * @mm_slot: the current mm_slot we are scanning
95 * @address: the next address inside that to be scanned
96 * @rmap_list: link to the next rmap to be scanned in the rmap_list
97 * @seqnr: count of completed full scans (needed when removing unstable node)
99 * There is only the one ksm_scan instance of this cursor structure.
102 struct mm_slot *mm_slot;
103 unsigned long address;
104 struct rmap_item **rmap_list;
109 * struct rmap_item - reverse mapping item for virtual addresses
110 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
111 * @filler: unused space we're making available in this patch
112 * @mm: the memory structure this rmap_item is pointing into
113 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
114 * @oldchecksum: previous checksum of the page at that virtual address
115 * @node: rb_node of this rmap_item in either unstable or stable tree
116 * @next: next rmap_item hanging off the same node of the stable tree
117 * @prev: previous rmap_item hanging off the same node of the stable tree
120 struct rmap_item *rmap_list;
121 unsigned long filler;
122 struct mm_struct *mm;
123 unsigned long address; /* + low bits used for flags below */
125 unsigned int oldchecksum; /* when unstable */
126 struct rmap_item *next; /* when stable */
129 struct rb_node node; /* when tree node */
130 struct rmap_item *prev; /* in stable list */
134 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
135 #define NODE_FLAG 0x100 /* is a node of unstable or stable tree */
136 #define STABLE_FLAG 0x200 /* is a node or list item of stable tree */
138 /* The stable and unstable tree heads */
139 static struct rb_root root_stable_tree = RB_ROOT;
140 static struct rb_root root_unstable_tree = RB_ROOT;
142 #define MM_SLOTS_HASH_HEADS 1024
143 static struct hlist_head *mm_slots_hash;
145 static struct mm_slot ksm_mm_head = {
146 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
148 static struct ksm_scan ksm_scan = {
149 .mm_slot = &ksm_mm_head,
152 static struct kmem_cache *rmap_item_cache;
153 static struct kmem_cache *mm_slot_cache;
155 /* The number of nodes in the stable tree */
156 static unsigned long ksm_pages_shared;
158 /* The number of page slots additionally sharing those nodes */
159 static unsigned long ksm_pages_sharing;
161 /* The number of nodes in the unstable tree */
162 static unsigned long ksm_pages_unshared;
164 /* The number of rmap_items in use: to calculate pages_volatile */
165 static unsigned long ksm_rmap_items;
167 /* Limit on the number of unswappable pages used */
168 static unsigned long ksm_max_kernel_pages;
170 /* Number of pages ksmd should scan in one batch */
171 static unsigned int ksm_thread_pages_to_scan = 100;
173 /* Milliseconds ksmd should sleep between batches */
174 static unsigned int ksm_thread_sleep_millisecs = 20;
176 #define KSM_RUN_STOP 0
177 #define KSM_RUN_MERGE 1
178 #define KSM_RUN_UNMERGE 2
179 static unsigned int ksm_run = KSM_RUN_STOP;
181 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
182 static DEFINE_MUTEX(ksm_thread_mutex);
183 static DEFINE_SPINLOCK(ksm_mmlist_lock);
185 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
186 sizeof(struct __struct), __alignof__(struct __struct),\
189 static int __init ksm_slab_init(void)
191 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
192 if (!rmap_item_cache)
195 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
202 kmem_cache_destroy(rmap_item_cache);
207 static void __init ksm_slab_free(void)
209 kmem_cache_destroy(mm_slot_cache);
210 kmem_cache_destroy(rmap_item_cache);
211 mm_slot_cache = NULL;
214 static inline struct rmap_item *alloc_rmap_item(void)
216 struct rmap_item *rmap_item;
218 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
224 static inline void free_rmap_item(struct rmap_item *rmap_item)
227 rmap_item->mm = NULL; /* debug safety */
228 kmem_cache_free(rmap_item_cache, rmap_item);
231 static inline struct mm_slot *alloc_mm_slot(void)
233 if (!mm_slot_cache) /* initialization failed */
235 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
238 static inline void free_mm_slot(struct mm_slot *mm_slot)
240 kmem_cache_free(mm_slot_cache, mm_slot);
243 static int __init mm_slots_hash_init(void)
245 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
252 static void __init mm_slots_hash_free(void)
254 kfree(mm_slots_hash);
257 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
259 struct mm_slot *mm_slot;
260 struct hlist_head *bucket;
261 struct hlist_node *node;
263 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
264 % MM_SLOTS_HASH_HEADS];
265 hlist_for_each_entry(mm_slot, node, bucket, link) {
266 if (mm == mm_slot->mm)
272 static void insert_to_mm_slots_hash(struct mm_struct *mm,
273 struct mm_slot *mm_slot)
275 struct hlist_head *bucket;
277 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
278 % MM_SLOTS_HASH_HEADS];
280 hlist_add_head(&mm_slot->link, bucket);
283 static inline int in_stable_tree(struct rmap_item *rmap_item)
285 return rmap_item->address & STABLE_FLAG;
289 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
290 * page tables after it has passed through ksm_exit() - which, if necessary,
291 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
292 * a special flag: they can just back out as soon as mm_users goes to zero.
293 * ksm_test_exit() is used throughout to make this test for exit: in some
294 * places for correctness, in some places just to avoid unnecessary work.
296 static inline bool ksm_test_exit(struct mm_struct *mm)
298 return atomic_read(&mm->mm_users) == 0;
302 * We use break_ksm to break COW on a ksm page: it's a stripped down
304 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
307 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
308 * in case the application has unmapped and remapped mm,addr meanwhile.
309 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
310 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
312 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
319 page = follow_page(vma, addr, FOLL_GET);
323 ret = handle_mm_fault(vma->vm_mm, vma, addr,
326 ret = VM_FAULT_WRITE;
328 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
330 * We must loop because handle_mm_fault() may back out if there's
331 * any difficulty e.g. if pte accessed bit gets updated concurrently.
333 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
334 * COW has been broken, even if the vma does not permit VM_WRITE;
335 * but note that a concurrent fault might break PageKsm for us.
337 * VM_FAULT_SIGBUS could occur if we race with truncation of the
338 * backing file, which also invalidates anonymous pages: that's
339 * okay, that truncation will have unmapped the PageKsm for us.
341 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
342 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
343 * current task has TIF_MEMDIE set, and will be OOM killed on return
344 * to user; and ksmd, having no mm, would never be chosen for that.
346 * But if the mm is in a limited mem_cgroup, then the fault may fail
347 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
348 * even ksmd can fail in this way - though it's usually breaking ksm
349 * just to undo a merge it made a moment before, so unlikely to oom.
351 * That's a pity: we might therefore have more kernel pages allocated
352 * than we're counting as nodes in the stable tree; but ksm_do_scan
353 * will retry to break_cow on each pass, so should recover the page
354 * in due course. The important thing is to not let VM_MERGEABLE
355 * be cleared while any such pages might remain in the area.
357 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
360 static void break_cow(struct rmap_item *rmap_item)
362 struct mm_struct *mm = rmap_item->mm;
363 unsigned long addr = rmap_item->address;
364 struct vm_area_struct *vma;
366 down_read(&mm->mmap_sem);
367 if (ksm_test_exit(mm))
369 vma = find_vma(mm, addr);
370 if (!vma || vma->vm_start > addr)
372 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
374 break_ksm(vma, addr);
376 up_read(&mm->mmap_sem);
379 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
381 struct mm_struct *mm = rmap_item->mm;
382 unsigned long addr = rmap_item->address;
383 struct vm_area_struct *vma;
386 down_read(&mm->mmap_sem);
387 if (ksm_test_exit(mm))
389 vma = find_vma(mm, addr);
390 if (!vma || vma->vm_start > addr)
392 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
395 page = follow_page(vma, addr, FOLL_GET);
398 if (PageAnon(page)) {
399 flush_anon_page(vma, page, addr);
400 flush_dcache_page(page);
405 up_read(&mm->mmap_sem);
410 * get_ksm_page: checks if the page at the virtual address in rmap_item
411 * is still PageKsm, in which case we can trust the content of the page,
412 * and it returns the gotten page; but NULL if the page has been zapped.
414 static struct page *get_ksm_page(struct rmap_item *rmap_item)
418 page = get_mergeable_page(rmap_item);
419 if (page && !PageKsm(page)) {
427 * Removing rmap_item from stable or unstable tree.
428 * This function will clean the information from the stable/unstable tree.
430 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
432 if (in_stable_tree(rmap_item)) {
433 struct rmap_item *next_item = rmap_item->next;
435 if (rmap_item->address & NODE_FLAG) {
437 rb_replace_node(&rmap_item->node,
440 next_item->address |= NODE_FLAG;
443 rb_erase(&rmap_item->node, &root_stable_tree);
447 struct rmap_item *prev_item = rmap_item->prev;
449 BUG_ON(prev_item->next != rmap_item);
450 prev_item->next = next_item;
452 BUG_ON(next_item->prev != rmap_item);
453 next_item->prev = rmap_item->prev;
458 rmap_item->next = NULL;
459 rmap_item->address &= PAGE_MASK;
461 } else if (rmap_item->address & NODE_FLAG) {
464 * Usually ksmd can and must skip the rb_erase, because
465 * root_unstable_tree was already reset to RB_ROOT.
466 * But be careful when an mm is exiting: do the rb_erase
467 * if this rmap_item was inserted by this scan, rather
468 * than left over from before.
470 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
473 rb_erase(&rmap_item->node, &root_unstable_tree);
475 ksm_pages_unshared--;
476 rmap_item->address &= PAGE_MASK;
479 cond_resched(); /* we're called from many long loops */
482 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
483 struct rmap_item **rmap_list)
486 struct rmap_item *rmap_item = *rmap_list;
487 *rmap_list = rmap_item->rmap_list;
488 remove_rmap_item_from_tree(rmap_item);
489 free_rmap_item(rmap_item);
494 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
495 * than check every pte of a given vma, the locking doesn't quite work for
496 * that - an rmap_item is assigned to the stable tree after inserting ksm
497 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
498 * rmap_items from parent to child at fork time (so as not to waste time
499 * if exit comes before the next scan reaches it).
501 * Similarly, although we'd like to remove rmap_items (so updating counts
502 * and freeing memory) when unmerging an area, it's easier to leave that
503 * to the next pass of ksmd - consider, for example, how ksmd might be
504 * in cmp_and_merge_page on one of the rmap_items we would be removing.
506 static int unmerge_ksm_pages(struct vm_area_struct *vma,
507 unsigned long start, unsigned long end)
512 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
513 if (ksm_test_exit(vma->vm_mm))
515 if (signal_pending(current))
518 err = break_ksm(vma, addr);
525 * Only called through the sysfs control interface:
527 static int unmerge_and_remove_all_rmap_items(void)
529 struct mm_slot *mm_slot;
530 struct mm_struct *mm;
531 struct vm_area_struct *vma;
534 spin_lock(&ksm_mmlist_lock);
535 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
536 struct mm_slot, mm_list);
537 spin_unlock(&ksm_mmlist_lock);
539 for (mm_slot = ksm_scan.mm_slot;
540 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
542 down_read(&mm->mmap_sem);
543 for (vma = mm->mmap; vma; vma = vma->vm_next) {
544 if (ksm_test_exit(mm))
546 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
548 err = unmerge_ksm_pages(vma,
549 vma->vm_start, vma->vm_end);
554 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
556 spin_lock(&ksm_mmlist_lock);
557 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
558 struct mm_slot, mm_list);
559 if (ksm_test_exit(mm)) {
560 hlist_del(&mm_slot->link);
561 list_del(&mm_slot->mm_list);
562 spin_unlock(&ksm_mmlist_lock);
564 free_mm_slot(mm_slot);
565 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
566 up_read(&mm->mmap_sem);
569 spin_unlock(&ksm_mmlist_lock);
570 up_read(&mm->mmap_sem);
578 up_read(&mm->mmap_sem);
579 spin_lock(&ksm_mmlist_lock);
580 ksm_scan.mm_slot = &ksm_mm_head;
581 spin_unlock(&ksm_mmlist_lock);
584 #endif /* CONFIG_SYSFS */
586 static u32 calc_checksum(struct page *page)
589 void *addr = kmap_atomic(page, KM_USER0);
590 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
591 kunmap_atomic(addr, KM_USER0);
595 static int memcmp_pages(struct page *page1, struct page *page2)
600 addr1 = kmap_atomic(page1, KM_USER0);
601 addr2 = kmap_atomic(page2, KM_USER1);
602 ret = memcmp(addr1, addr2, PAGE_SIZE);
603 kunmap_atomic(addr2, KM_USER1);
604 kunmap_atomic(addr1, KM_USER0);
608 static inline int pages_identical(struct page *page1, struct page *page2)
610 return !memcmp_pages(page1, page2);
613 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
616 struct mm_struct *mm = vma->vm_mm;
623 addr = page_address_in_vma(page, vma);
627 ptep = page_check_address(page, mm, addr, &ptl, 0);
631 if (pte_write(*ptep)) {
634 swapped = PageSwapCache(page);
635 flush_cache_page(vma, addr, page_to_pfn(page));
637 * Ok this is tricky, when get_user_pages_fast() run it doesnt
638 * take any lock, therefore the check that we are going to make
639 * with the pagecount against the mapcount is racey and
640 * O_DIRECT can happen right after the check.
641 * So we clear the pte and flush the tlb before the check
642 * this assure us that no O_DIRECT can happen after the check
643 * or in the middle of the check.
645 entry = ptep_clear_flush(vma, addr, ptep);
647 * Check that no O_DIRECT or similar I/O is in progress on the
650 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
651 set_pte_at_notify(mm, addr, ptep, entry);
654 entry = pte_wrprotect(entry);
655 set_pte_at_notify(mm, addr, ptep, entry);
661 pte_unmap_unlock(ptep, ptl);
667 * replace_page - replace page in vma by new ksm page
668 * @vma: vma that holds the pte pointing to page
669 * @page: the page we are replacing by kpage
670 * @kpage: the ksm page we replace page by
671 * @orig_pte: the original value of the pte
673 * Returns 0 on success, -EFAULT on failure.
675 static int replace_page(struct vm_area_struct *vma, struct page *page,
676 struct page *kpage, pte_t orig_pte)
678 struct mm_struct *mm = vma->vm_mm;
687 addr = page_address_in_vma(page, vma);
691 pgd = pgd_offset(mm, addr);
692 if (!pgd_present(*pgd))
695 pud = pud_offset(pgd, addr);
696 if (!pud_present(*pud))
699 pmd = pmd_offset(pud, addr);
700 if (!pmd_present(*pmd))
703 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
704 if (!pte_same(*ptep, orig_pte)) {
705 pte_unmap_unlock(ptep, ptl);
710 page_add_ksm_rmap(kpage);
712 flush_cache_page(vma, addr, pte_pfn(*ptep));
713 ptep_clear_flush(vma, addr, ptep);
714 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
716 page_remove_rmap(page);
719 pte_unmap_unlock(ptep, ptl);
726 * try_to_merge_one_page - take two pages and merge them into one
727 * @vma: the vma that holds the pte pointing to page
728 * @page: the PageAnon page that we want to replace with kpage
729 * @kpage: the PageKsm page (or newly allocated page which page_add_ksm_rmap
730 * will make PageKsm) that we want to map instead of page
732 * This function returns 0 if the pages were merged, -EFAULT otherwise.
734 static int try_to_merge_one_page(struct vm_area_struct *vma,
735 struct page *page, struct page *kpage)
737 pte_t orig_pte = __pte(0);
740 if (!(vma->vm_flags & VM_MERGEABLE))
746 * We need the page lock to read a stable PageSwapCache in
747 * write_protect_page(). We use trylock_page() instead of
748 * lock_page() because we don't want to wait here - we
749 * prefer to continue scanning and merging different pages,
750 * then come back to this page when it is unlocked.
752 if (!trylock_page(page))
755 * If this anonymous page is mapped only here, its pte may need
756 * to be write-protected. If it's mapped elsewhere, all of its
757 * ptes are necessarily already write-protected. But in either
758 * case, we need to lock and check page_count is not raised.
760 if (write_protect_page(vma, page, &orig_pte) == 0 &&
761 pages_identical(page, kpage))
762 err = replace_page(vma, page, kpage, orig_pte);
770 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
771 * but no new kernel page is allocated: kpage must already be a ksm page.
773 * This function returns 0 if the pages were merged, -EFAULT otherwise.
775 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
776 struct page *page, struct page *kpage)
778 struct mm_struct *mm = rmap_item->mm;
779 struct vm_area_struct *vma;
782 down_read(&mm->mmap_sem);
783 if (ksm_test_exit(mm))
785 vma = find_vma(mm, rmap_item->address);
786 if (!vma || vma->vm_start > rmap_item->address)
789 err = try_to_merge_one_page(vma, page, kpage);
791 up_read(&mm->mmap_sem);
796 * try_to_merge_two_pages - take two identical pages and prepare them
797 * to be merged into one page.
799 * This function returns the kpage if we successfully merged two identical
800 * pages into one ksm page, NULL otherwise.
802 * Note that this function allocates a new kernel page: if one of the pages
803 * is already a ksm page, try_to_merge_with_ksm_page should be used.
805 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
807 struct rmap_item *tree_rmap_item,
808 struct page *tree_page)
810 struct mm_struct *mm = rmap_item->mm;
811 struct vm_area_struct *vma;
816 * The number of nodes in the stable tree
817 * is the number of kernel pages that we hold.
819 if (ksm_max_kernel_pages &&
820 ksm_max_kernel_pages <= ksm_pages_shared)
823 kpage = alloc_page(GFP_HIGHUSER);
827 down_read(&mm->mmap_sem);
828 if (ksm_test_exit(mm))
830 vma = find_vma(mm, rmap_item->address);
831 if (!vma || vma->vm_start > rmap_item->address)
834 copy_user_highpage(kpage, page, rmap_item->address, vma);
835 err = try_to_merge_one_page(vma, page, kpage);
837 up_read(&mm->mmap_sem);
840 err = try_to_merge_with_ksm_page(tree_rmap_item,
843 * If that fails, we have a ksm page with only one pte
844 * pointing to it: so break it.
847 break_cow(rmap_item);
857 * stable_tree_search - search for page inside the stable tree
859 * This function checks if there is a page inside the stable tree
860 * with identical content to the page that we are scanning right now.
862 * This function return rmap_item pointer to the identical item if found,
865 static struct rmap_item *stable_tree_search(struct page *page,
866 struct page **tree_pagep)
868 struct rb_node *node = root_stable_tree.rb_node;
871 struct rmap_item *tree_rmap_item, *next_rmap_item;
872 struct page *tree_page;
875 tree_rmap_item = rb_entry(node, struct rmap_item, node);
876 while (tree_rmap_item) {
877 BUG_ON(!in_stable_tree(tree_rmap_item));
879 tree_page = get_ksm_page(tree_rmap_item);
882 next_rmap_item = tree_rmap_item->next;
883 remove_rmap_item_from_tree(tree_rmap_item);
884 tree_rmap_item = next_rmap_item;
889 ret = memcmp_pages(page, tree_page);
893 node = node->rb_left;
894 } else if (ret > 0) {
896 node = node->rb_right;
898 *tree_pagep = tree_page;
899 return tree_rmap_item;
907 * stable_tree_insert - insert rmap_item pointing to new ksm page
908 * into the stable tree.
910 * This function returns rmap_item if success, NULL otherwise.
912 static struct rmap_item *stable_tree_insert(struct page *kpage,
913 struct rmap_item *rmap_item)
915 struct rb_node **new = &root_stable_tree.rb_node;
916 struct rb_node *parent = NULL;
919 struct rmap_item *tree_rmap_item, *next_rmap_item;
920 struct page *tree_page;
923 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
924 while (tree_rmap_item) {
925 BUG_ON(!in_stable_tree(tree_rmap_item));
927 tree_page = get_ksm_page(tree_rmap_item);
930 next_rmap_item = tree_rmap_item->next;
931 remove_rmap_item_from_tree(tree_rmap_item);
932 tree_rmap_item = next_rmap_item;
937 ret = memcmp_pages(kpage, tree_page);
942 new = &parent->rb_left;
944 new = &parent->rb_right;
947 * It is not a bug that stable_tree_search() didn't
948 * find this node: because at that time our page was
949 * not yet write-protected, so may have changed since.
955 rmap_item->address |= NODE_FLAG | STABLE_FLAG;
956 rmap_item->next = NULL;
957 rb_link_node(&rmap_item->node, parent, new);
958 rb_insert_color(&rmap_item->node, &root_stable_tree);
965 * unstable_tree_search_insert - search for identical page,
966 * else insert rmap_item into the unstable tree.
968 * This function searches for a page in the unstable tree identical to the
969 * page currently being scanned; and if no identical page is found in the
970 * tree, we insert rmap_item as a new object into the unstable tree.
972 * This function returns pointer to rmap_item found to be identical
973 * to the currently scanned page, NULL otherwise.
975 * This function does both searching and inserting, because they share
976 * the same walking algorithm in an rbtree.
979 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
981 struct page **tree_pagep)
984 struct rb_node **new = &root_unstable_tree.rb_node;
985 struct rb_node *parent = NULL;
988 struct rmap_item *tree_rmap_item;
989 struct page *tree_page;
993 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
994 tree_page = get_mergeable_page(tree_rmap_item);
999 * Don't substitute a ksm page for a forked page.
1001 if (page == tree_page) {
1002 put_page(tree_page);
1006 ret = memcmp_pages(page, tree_page);
1010 put_page(tree_page);
1011 new = &parent->rb_left;
1012 } else if (ret > 0) {
1013 put_page(tree_page);
1014 new = &parent->rb_right;
1016 *tree_pagep = tree_page;
1017 return tree_rmap_item;
1021 rmap_item->address |= NODE_FLAG;
1022 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1023 rb_link_node(&rmap_item->node, parent, new);
1024 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1026 ksm_pages_unshared++;
1031 * stable_tree_append - add another rmap_item to the linked list of
1032 * rmap_items hanging off a given node of the stable tree, all sharing
1033 * the same ksm page.
1035 static void stable_tree_append(struct rmap_item *rmap_item,
1036 struct rmap_item *tree_rmap_item)
1038 rmap_item->next = tree_rmap_item->next;
1039 rmap_item->prev = tree_rmap_item;
1041 if (tree_rmap_item->next)
1042 tree_rmap_item->next->prev = rmap_item;
1044 tree_rmap_item->next = rmap_item;
1045 rmap_item->address |= STABLE_FLAG;
1047 ksm_pages_sharing++;
1051 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1052 * if not, compare checksum to previous and if it's the same, see if page can
1053 * be inserted into the unstable tree, or merged with a page already there and
1054 * both transferred to the stable tree.
1056 * @page: the page that we are searching identical page to.
1057 * @rmap_item: the reverse mapping into the virtual address of this page
1059 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1061 struct rmap_item *tree_rmap_item;
1062 struct page *tree_page = NULL;
1064 unsigned int checksum;
1067 remove_rmap_item_from_tree(rmap_item);
1069 /* We first start with searching the page inside the stable tree */
1070 tree_rmap_item = stable_tree_search(page, &tree_page);
1071 if (tree_rmap_item) {
1073 if (page == kpage) /* forked */
1076 err = try_to_merge_with_ksm_page(rmap_item,
1080 * The page was successfully merged:
1081 * add its rmap_item to the stable tree.
1083 stable_tree_append(rmap_item, tree_rmap_item);
1090 * A ksm page might have got here by fork, but its other
1091 * references have already been removed from the stable tree.
1092 * Or it might be left over from a break_ksm which failed
1093 * when the mem_cgroup had reached its limit: try again now.
1096 break_cow(rmap_item);
1099 * In case the hash value of the page was changed from the last time we
1100 * have calculated it, this page to be changed frequely, therefore we
1101 * don't want to insert it to the unstable tree, and we don't want to
1102 * waste our time to search if there is something identical to it there.
1104 checksum = calc_checksum(page);
1105 if (rmap_item->oldchecksum != checksum) {
1106 rmap_item->oldchecksum = checksum;
1111 unstable_tree_search_insert(rmap_item, page, &tree_page);
1112 if (tree_rmap_item) {
1113 kpage = try_to_merge_two_pages(rmap_item, page,
1114 tree_rmap_item, tree_page);
1115 put_page(tree_page);
1117 * As soon as we merge this page, we want to remove the
1118 * rmap_item of the page we have merged with from the unstable
1119 * tree, and insert it instead as new node in the stable tree.
1122 remove_rmap_item_from_tree(tree_rmap_item);
1125 * If we fail to insert the page into the stable tree,
1126 * we will have 2 virtual addresses that are pointing
1127 * to a ksm page left outside the stable tree,
1128 * in which case we need to break_cow on both.
1130 if (stable_tree_insert(kpage, tree_rmap_item))
1131 stable_tree_append(rmap_item, tree_rmap_item);
1133 break_cow(tree_rmap_item);
1134 break_cow(rmap_item);
1141 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1142 struct rmap_item **rmap_list,
1145 struct rmap_item *rmap_item;
1147 while (*rmap_list) {
1148 rmap_item = *rmap_list;
1149 if ((rmap_item->address & PAGE_MASK) == addr)
1151 if (rmap_item->address > addr)
1153 *rmap_list = rmap_item->rmap_list;
1154 remove_rmap_item_from_tree(rmap_item);
1155 free_rmap_item(rmap_item);
1158 rmap_item = alloc_rmap_item();
1160 /* It has already been zeroed */
1161 rmap_item->mm = mm_slot->mm;
1162 rmap_item->address = addr;
1163 rmap_item->rmap_list = *rmap_list;
1164 *rmap_list = rmap_item;
1169 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1171 struct mm_struct *mm;
1172 struct mm_slot *slot;
1173 struct vm_area_struct *vma;
1174 struct rmap_item *rmap_item;
1176 if (list_empty(&ksm_mm_head.mm_list))
1179 slot = ksm_scan.mm_slot;
1180 if (slot == &ksm_mm_head) {
1181 root_unstable_tree = RB_ROOT;
1183 spin_lock(&ksm_mmlist_lock);
1184 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1185 ksm_scan.mm_slot = slot;
1186 spin_unlock(&ksm_mmlist_lock);
1188 ksm_scan.address = 0;
1189 ksm_scan.rmap_list = &slot->rmap_list;
1193 down_read(&mm->mmap_sem);
1194 if (ksm_test_exit(mm))
1197 vma = find_vma(mm, ksm_scan.address);
1199 for (; vma; vma = vma->vm_next) {
1200 if (!(vma->vm_flags & VM_MERGEABLE))
1202 if (ksm_scan.address < vma->vm_start)
1203 ksm_scan.address = vma->vm_start;
1205 ksm_scan.address = vma->vm_end;
1207 while (ksm_scan.address < vma->vm_end) {
1208 if (ksm_test_exit(mm))
1210 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1211 if (*page && PageAnon(*page)) {
1212 flush_anon_page(vma, *page, ksm_scan.address);
1213 flush_dcache_page(*page);
1214 rmap_item = get_next_rmap_item(slot,
1215 ksm_scan.rmap_list, ksm_scan.address);
1217 ksm_scan.rmap_list =
1218 &rmap_item->rmap_list;
1219 ksm_scan.address += PAGE_SIZE;
1222 up_read(&mm->mmap_sem);
1227 ksm_scan.address += PAGE_SIZE;
1232 if (ksm_test_exit(mm)) {
1233 ksm_scan.address = 0;
1234 ksm_scan.rmap_list = &slot->rmap_list;
1237 * Nuke all the rmap_items that are above this current rmap:
1238 * because there were no VM_MERGEABLE vmas with such addresses.
1240 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1242 spin_lock(&ksm_mmlist_lock);
1243 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1244 struct mm_slot, mm_list);
1245 if (ksm_scan.address == 0) {
1247 * We've completed a full scan of all vmas, holding mmap_sem
1248 * throughout, and found no VM_MERGEABLE: so do the same as
1249 * __ksm_exit does to remove this mm from all our lists now.
1250 * This applies either when cleaning up after __ksm_exit
1251 * (but beware: we can reach here even before __ksm_exit),
1252 * or when all VM_MERGEABLE areas have been unmapped (and
1253 * mmap_sem then protects against race with MADV_MERGEABLE).
1255 hlist_del(&slot->link);
1256 list_del(&slot->mm_list);
1257 spin_unlock(&ksm_mmlist_lock);
1260 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1261 up_read(&mm->mmap_sem);
1264 spin_unlock(&ksm_mmlist_lock);
1265 up_read(&mm->mmap_sem);
1268 /* Repeat until we've completed scanning the whole list */
1269 slot = ksm_scan.mm_slot;
1270 if (slot != &ksm_mm_head)
1278 * ksm_do_scan - the ksm scanner main worker function.
1279 * @scan_npages - number of pages we want to scan before we return.
1281 static void ksm_do_scan(unsigned int scan_npages)
1283 struct rmap_item *rmap_item;
1286 while (scan_npages--) {
1288 rmap_item = scan_get_next_rmap_item(&page);
1291 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1292 cmp_and_merge_page(page, rmap_item);
1293 else if (page_mapcount(page) == 1) {
1295 * Replace now-unshared ksm page by ordinary page.
1297 break_cow(rmap_item);
1298 remove_rmap_item_from_tree(rmap_item);
1299 rmap_item->oldchecksum = calc_checksum(page);
1305 static int ksmd_should_run(void)
1307 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1310 static int ksm_scan_thread(void *nothing)
1312 set_user_nice(current, 5);
1314 while (!kthread_should_stop()) {
1315 mutex_lock(&ksm_thread_mutex);
1316 if (ksmd_should_run())
1317 ksm_do_scan(ksm_thread_pages_to_scan);
1318 mutex_unlock(&ksm_thread_mutex);
1320 if (ksmd_should_run()) {
1321 schedule_timeout_interruptible(
1322 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1324 wait_event_interruptible(ksm_thread_wait,
1325 ksmd_should_run() || kthread_should_stop());
1331 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1332 unsigned long end, int advice, unsigned long *vm_flags)
1334 struct mm_struct *mm = vma->vm_mm;
1338 case MADV_MERGEABLE:
1340 * Be somewhat over-protective for now!
1342 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1343 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1344 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1345 VM_MIXEDMAP | VM_SAO))
1346 return 0; /* just ignore the advice */
1348 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1349 err = __ksm_enter(mm);
1354 *vm_flags |= VM_MERGEABLE;
1357 case MADV_UNMERGEABLE:
1358 if (!(*vm_flags & VM_MERGEABLE))
1359 return 0; /* just ignore the advice */
1361 if (vma->anon_vma) {
1362 err = unmerge_ksm_pages(vma, start, end);
1367 *vm_flags &= ~VM_MERGEABLE;
1374 int __ksm_enter(struct mm_struct *mm)
1376 struct mm_slot *mm_slot;
1379 mm_slot = alloc_mm_slot();
1383 /* Check ksm_run too? Would need tighter locking */
1384 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1386 spin_lock(&ksm_mmlist_lock);
1387 insert_to_mm_slots_hash(mm, mm_slot);
1389 * Insert just behind the scanning cursor, to let the area settle
1390 * down a little; when fork is followed by immediate exec, we don't
1391 * want ksmd to waste time setting up and tearing down an rmap_list.
1393 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1394 spin_unlock(&ksm_mmlist_lock);
1396 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1397 atomic_inc(&mm->mm_count);
1400 wake_up_interruptible(&ksm_thread_wait);
1405 void __ksm_exit(struct mm_struct *mm)
1407 struct mm_slot *mm_slot;
1408 int easy_to_free = 0;
1411 * This process is exiting: if it's straightforward (as is the
1412 * case when ksmd was never running), free mm_slot immediately.
1413 * But if it's at the cursor or has rmap_items linked to it, use
1414 * mmap_sem to synchronize with any break_cows before pagetables
1415 * are freed, and leave the mm_slot on the list for ksmd to free.
1416 * Beware: ksm may already have noticed it exiting and freed the slot.
1419 spin_lock(&ksm_mmlist_lock);
1420 mm_slot = get_mm_slot(mm);
1421 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1422 if (!mm_slot->rmap_list) {
1423 hlist_del(&mm_slot->link);
1424 list_del(&mm_slot->mm_list);
1427 list_move(&mm_slot->mm_list,
1428 &ksm_scan.mm_slot->mm_list);
1431 spin_unlock(&ksm_mmlist_lock);
1434 free_mm_slot(mm_slot);
1435 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1437 } else if (mm_slot) {
1438 down_write(&mm->mmap_sem);
1439 up_write(&mm->mmap_sem);
1445 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1448 #define KSM_ATTR_RO(_name) \
1449 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1450 #define KSM_ATTR(_name) \
1451 static struct kobj_attribute _name##_attr = \
1452 __ATTR(_name, 0644, _name##_show, _name##_store)
1454 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1455 struct kobj_attribute *attr, char *buf)
1457 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1460 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1461 struct kobj_attribute *attr,
1462 const char *buf, size_t count)
1464 unsigned long msecs;
1467 err = strict_strtoul(buf, 10, &msecs);
1468 if (err || msecs > UINT_MAX)
1471 ksm_thread_sleep_millisecs = msecs;
1475 KSM_ATTR(sleep_millisecs);
1477 static ssize_t pages_to_scan_show(struct kobject *kobj,
1478 struct kobj_attribute *attr, char *buf)
1480 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1483 static ssize_t pages_to_scan_store(struct kobject *kobj,
1484 struct kobj_attribute *attr,
1485 const char *buf, size_t count)
1488 unsigned long nr_pages;
1490 err = strict_strtoul(buf, 10, &nr_pages);
1491 if (err || nr_pages > UINT_MAX)
1494 ksm_thread_pages_to_scan = nr_pages;
1498 KSM_ATTR(pages_to_scan);
1500 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1503 return sprintf(buf, "%u\n", ksm_run);
1506 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1507 const char *buf, size_t count)
1510 unsigned long flags;
1512 err = strict_strtoul(buf, 10, &flags);
1513 if (err || flags > UINT_MAX)
1515 if (flags > KSM_RUN_UNMERGE)
1519 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1520 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1521 * breaking COW to free the unswappable pages_shared (but leaves
1522 * mm_slots on the list for when ksmd may be set running again).
1525 mutex_lock(&ksm_thread_mutex);
1526 if (ksm_run != flags) {
1528 if (flags & KSM_RUN_UNMERGE) {
1529 current->flags |= PF_OOM_ORIGIN;
1530 err = unmerge_and_remove_all_rmap_items();
1531 current->flags &= ~PF_OOM_ORIGIN;
1533 ksm_run = KSM_RUN_STOP;
1538 mutex_unlock(&ksm_thread_mutex);
1540 if (flags & KSM_RUN_MERGE)
1541 wake_up_interruptible(&ksm_thread_wait);
1547 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1548 struct kobj_attribute *attr,
1549 const char *buf, size_t count)
1552 unsigned long nr_pages;
1554 err = strict_strtoul(buf, 10, &nr_pages);
1558 ksm_max_kernel_pages = nr_pages;
1563 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1564 struct kobj_attribute *attr, char *buf)
1566 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1568 KSM_ATTR(max_kernel_pages);
1570 static ssize_t pages_shared_show(struct kobject *kobj,
1571 struct kobj_attribute *attr, char *buf)
1573 return sprintf(buf, "%lu\n", ksm_pages_shared);
1575 KSM_ATTR_RO(pages_shared);
1577 static ssize_t pages_sharing_show(struct kobject *kobj,
1578 struct kobj_attribute *attr, char *buf)
1580 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1582 KSM_ATTR_RO(pages_sharing);
1584 static ssize_t pages_unshared_show(struct kobject *kobj,
1585 struct kobj_attribute *attr, char *buf)
1587 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1589 KSM_ATTR_RO(pages_unshared);
1591 static ssize_t pages_volatile_show(struct kobject *kobj,
1592 struct kobj_attribute *attr, char *buf)
1594 long ksm_pages_volatile;
1596 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1597 - ksm_pages_sharing - ksm_pages_unshared;
1599 * It was not worth any locking to calculate that statistic,
1600 * but it might therefore sometimes be negative: conceal that.
1602 if (ksm_pages_volatile < 0)
1603 ksm_pages_volatile = 0;
1604 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1606 KSM_ATTR_RO(pages_volatile);
1608 static ssize_t full_scans_show(struct kobject *kobj,
1609 struct kobj_attribute *attr, char *buf)
1611 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1613 KSM_ATTR_RO(full_scans);
1615 static struct attribute *ksm_attrs[] = {
1616 &sleep_millisecs_attr.attr,
1617 &pages_to_scan_attr.attr,
1619 &max_kernel_pages_attr.attr,
1620 &pages_shared_attr.attr,
1621 &pages_sharing_attr.attr,
1622 &pages_unshared_attr.attr,
1623 &pages_volatile_attr.attr,
1624 &full_scans_attr.attr,
1628 static struct attribute_group ksm_attr_group = {
1632 #endif /* CONFIG_SYSFS */
1634 static int __init ksm_init(void)
1636 struct task_struct *ksm_thread;
1639 ksm_max_kernel_pages = totalram_pages / 4;
1641 err = ksm_slab_init();
1645 err = mm_slots_hash_init();
1649 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1650 if (IS_ERR(ksm_thread)) {
1651 printk(KERN_ERR "ksm: creating kthread failed\n");
1652 err = PTR_ERR(ksm_thread);
1657 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1659 printk(KERN_ERR "ksm: register sysfs failed\n");
1660 kthread_stop(ksm_thread);
1664 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
1666 #endif /* CONFIG_SYSFS */
1671 mm_slots_hash_free();
1677 module_init(ksm_init)