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/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hashtable.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
40 #include <asm/tlbflush.h>
44 * A few notes about the KSM scanning process,
45 * to make it easier to understand the data structures below:
47 * In order to reduce excessive scanning, KSM sorts the memory pages by their
48 * contents into a data structure that holds pointers to the pages' locations.
50 * Since the contents of the pages may change at any moment, KSM cannot just
51 * insert the pages into a normal sorted tree and expect it to find anything.
52 * Therefore KSM uses two data structures - the stable and the unstable tree.
54 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
55 * by their contents. Because each such page is write-protected, searching on
56 * this tree is fully assured to be working (except when pages are unmapped),
57 * and therefore this tree is called the stable tree.
59 * In addition to the stable tree, KSM uses a second data structure called the
60 * unstable tree: this tree holds pointers to pages which have been found to
61 * be "unchanged for a period of time". The unstable tree sorts these pages
62 * by their contents, but since they are not write-protected, KSM cannot rely
63 * upon the unstable tree to work correctly - the unstable tree is liable to
64 * be corrupted as its contents are modified, and so it is called unstable.
66 * KSM solves this problem by several techniques:
68 * 1) The unstable tree is flushed every time KSM completes scanning all
69 * memory areas, and then the tree is rebuilt again from the beginning.
70 * 2) KSM will only insert into the unstable tree, pages whose hash value
71 * has not changed since the previous scan of all memory areas.
72 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
73 * colors of the nodes and not on their contents, assuring that even when
74 * the tree gets "corrupted" it won't get out of balance, so scanning time
75 * remains the same (also, searching and inserting nodes in an rbtree uses
76 * the same algorithm, so we have no overhead when we flush and rebuild).
77 * 4) KSM never flushes the stable tree, which means that even if it were to
78 * take 10 attempts to find a page in the unstable tree, once it is found,
79 * it is secured in the stable tree. (When we scan a new page, we first
80 * compare it against the stable tree, and then against the unstable tree.)
84 * struct mm_slot - ksm information per mm that is being scanned
85 * @link: link to the mm_slots hash list
86 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
87 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
88 * @mm: the mm that this information is valid for
91 struct hlist_node link;
92 struct list_head mm_list;
93 struct rmap_item *rmap_list;
98 * struct ksm_scan - cursor for scanning
99 * @mm_slot: the current mm_slot we are scanning
100 * @address: the next address inside that to be scanned
101 * @rmap_list: link to the next rmap to be scanned in the rmap_list
102 * @seqnr: count of completed full scans (needed when removing unstable node)
104 * There is only the one ksm_scan instance of this cursor structure.
107 struct mm_slot *mm_slot;
108 unsigned long address;
109 struct rmap_item **rmap_list;
114 * struct stable_node - node of the stable rbtree
115 * @node: rb node of this ksm page in the stable tree
116 * @hlist: hlist head of rmap_items using this ksm page
117 * @kpfn: page frame number of this ksm page
121 struct hlist_head hlist;
126 * struct rmap_item - reverse mapping item for virtual addresses
127 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
128 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
129 * @mm: the memory structure this rmap_item is pointing into
130 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
131 * @oldchecksum: previous checksum of the page at that virtual address
132 * @node: rb node of this rmap_item in the unstable tree
133 * @head: pointer to stable_node heading this list in the stable tree
134 * @hlist: link into hlist of rmap_items hanging off that stable_node
137 struct rmap_item *rmap_list;
138 struct anon_vma *anon_vma; /* when stable */
139 struct mm_struct *mm;
140 unsigned long address; /* + low bits used for flags below */
141 unsigned int oldchecksum; /* when unstable */
143 struct rb_node node; /* when node of unstable tree */
144 struct { /* when listed from stable tree */
145 struct stable_node *head;
146 struct hlist_node hlist;
151 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
152 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
153 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
155 /* The stable and unstable tree heads */
156 static struct rb_root root_stable_tree = RB_ROOT;
157 static struct rb_root root_unstable_tree = RB_ROOT;
159 #define MM_SLOTS_HASH_BITS 10
160 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
162 static struct mm_slot ksm_mm_head = {
163 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165 static struct ksm_scan ksm_scan = {
166 .mm_slot = &ksm_mm_head,
169 static struct kmem_cache *rmap_item_cache;
170 static struct kmem_cache *stable_node_cache;
171 static struct kmem_cache *mm_slot_cache;
173 /* The number of nodes in the stable tree */
174 static unsigned long ksm_pages_shared;
176 /* The number of page slots additionally sharing those nodes */
177 static unsigned long ksm_pages_sharing;
179 /* The number of nodes in the unstable tree */
180 static unsigned long ksm_pages_unshared;
182 /* The number of rmap_items in use: to calculate pages_volatile */
183 static unsigned long ksm_rmap_items;
185 /* Number of pages ksmd should scan in one batch */
186 static unsigned int ksm_thread_pages_to_scan = 100;
188 /* Milliseconds ksmd should sleep between batches */
189 static unsigned int ksm_thread_sleep_millisecs = 20;
191 #define KSM_RUN_STOP 0
192 #define KSM_RUN_MERGE 1
193 #define KSM_RUN_UNMERGE 2
194 static unsigned int ksm_run = KSM_RUN_STOP;
196 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
197 static DEFINE_MUTEX(ksm_thread_mutex);
198 static DEFINE_SPINLOCK(ksm_mmlist_lock);
200 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
201 sizeof(struct __struct), __alignof__(struct __struct),\
204 static int __init ksm_slab_init(void)
206 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
207 if (!rmap_item_cache)
210 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
211 if (!stable_node_cache)
214 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
221 kmem_cache_destroy(stable_node_cache);
223 kmem_cache_destroy(rmap_item_cache);
228 static void __init ksm_slab_free(void)
230 kmem_cache_destroy(mm_slot_cache);
231 kmem_cache_destroy(stable_node_cache);
232 kmem_cache_destroy(rmap_item_cache);
233 mm_slot_cache = NULL;
236 static inline struct rmap_item *alloc_rmap_item(void)
238 struct rmap_item *rmap_item;
240 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
246 static inline void free_rmap_item(struct rmap_item *rmap_item)
249 rmap_item->mm = NULL; /* debug safety */
250 kmem_cache_free(rmap_item_cache, rmap_item);
253 static inline struct stable_node *alloc_stable_node(void)
255 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
258 static inline void free_stable_node(struct stable_node *stable_node)
260 kmem_cache_free(stable_node_cache, stable_node);
263 static inline struct mm_slot *alloc_mm_slot(void)
265 if (!mm_slot_cache) /* initialization failed */
267 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
270 static inline void free_mm_slot(struct mm_slot *mm_slot)
272 kmem_cache_free(mm_slot_cache, mm_slot);
275 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277 struct hlist_node *node;
278 struct mm_slot *slot;
280 hash_for_each_possible(mm_slots_hash, slot, node, link, (unsigned long)mm)
287 static void insert_to_mm_slots_hash(struct mm_struct *mm,
288 struct mm_slot *mm_slot)
291 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
294 static inline int in_stable_tree(struct rmap_item *rmap_item)
296 return rmap_item->address & STABLE_FLAG;
300 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
301 * page tables after it has passed through ksm_exit() - which, if necessary,
302 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
303 * a special flag: they can just back out as soon as mm_users goes to zero.
304 * ksm_test_exit() is used throughout to make this test for exit: in some
305 * places for correctness, in some places just to avoid unnecessary work.
307 static inline bool ksm_test_exit(struct mm_struct *mm)
309 return atomic_read(&mm->mm_users) == 0;
313 * We use break_ksm to break COW on a ksm page: it's a stripped down
315 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
318 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
319 * in case the application has unmapped and remapped mm,addr meanwhile.
320 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
321 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
323 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
330 page = follow_page(vma, addr, FOLL_GET);
331 if (IS_ERR_OR_NULL(page))
334 ret = handle_mm_fault(vma->vm_mm, vma, addr,
337 ret = VM_FAULT_WRITE;
339 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
341 * We must loop because handle_mm_fault() may back out if there's
342 * any difficulty e.g. if pte accessed bit gets updated concurrently.
344 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
345 * COW has been broken, even if the vma does not permit VM_WRITE;
346 * but note that a concurrent fault might break PageKsm for us.
348 * VM_FAULT_SIGBUS could occur if we race with truncation of the
349 * backing file, which also invalidates anonymous pages: that's
350 * okay, that truncation will have unmapped the PageKsm for us.
352 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
353 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
354 * current task has TIF_MEMDIE set, and will be OOM killed on return
355 * to user; and ksmd, having no mm, would never be chosen for that.
357 * But if the mm is in a limited mem_cgroup, then the fault may fail
358 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
359 * even ksmd can fail in this way - though it's usually breaking ksm
360 * just to undo a merge it made a moment before, so unlikely to oom.
362 * That's a pity: we might therefore have more kernel pages allocated
363 * than we're counting as nodes in the stable tree; but ksm_do_scan
364 * will retry to break_cow on each pass, so should recover the page
365 * in due course. The important thing is to not let VM_MERGEABLE
366 * be cleared while any such pages might remain in the area.
368 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
371 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
374 struct vm_area_struct *vma;
375 if (ksm_test_exit(mm))
377 vma = find_vma(mm, addr);
378 if (!vma || vma->vm_start > addr)
380 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
385 static void break_cow(struct rmap_item *rmap_item)
387 struct mm_struct *mm = rmap_item->mm;
388 unsigned long addr = rmap_item->address;
389 struct vm_area_struct *vma;
392 * It is not an accident that whenever we want to break COW
393 * to undo, we also need to drop a reference to the anon_vma.
395 put_anon_vma(rmap_item->anon_vma);
397 down_read(&mm->mmap_sem);
398 vma = find_mergeable_vma(mm, addr);
400 break_ksm(vma, addr);
401 up_read(&mm->mmap_sem);
404 static struct page *page_trans_compound_anon(struct page *page)
406 if (PageTransCompound(page)) {
407 struct page *head = compound_trans_head(page);
409 * head may actually be splitted and freed from under
410 * us but it's ok here.
418 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
420 struct mm_struct *mm = rmap_item->mm;
421 unsigned long addr = rmap_item->address;
422 struct vm_area_struct *vma;
425 down_read(&mm->mmap_sem);
426 vma = find_mergeable_vma(mm, addr);
430 page = follow_page(vma, addr, FOLL_GET);
431 if (IS_ERR_OR_NULL(page))
433 if (PageAnon(page) || page_trans_compound_anon(page)) {
434 flush_anon_page(vma, page, addr);
435 flush_dcache_page(page);
440 up_read(&mm->mmap_sem);
444 static void remove_node_from_stable_tree(struct stable_node *stable_node)
446 struct rmap_item *rmap_item;
447 struct hlist_node *hlist;
449 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
450 if (rmap_item->hlist.next)
454 put_anon_vma(rmap_item->anon_vma);
455 rmap_item->address &= PAGE_MASK;
459 rb_erase(&stable_node->node, &root_stable_tree);
460 free_stable_node(stable_node);
464 * get_ksm_page: checks if the page indicated by the stable node
465 * is still its ksm page, despite having held no reference to it.
466 * In which case we can trust the content of the page, and it
467 * returns the gotten page; but if the page has now been zapped,
468 * remove the stale node from the stable tree and return NULL.
470 * You would expect the stable_node to hold a reference to the ksm page.
471 * But if it increments the page's count, swapping out has to wait for
472 * ksmd to come around again before it can free the page, which may take
473 * seconds or even minutes: much too unresponsive. So instead we use a
474 * "keyhole reference": access to the ksm page from the stable node peeps
475 * out through its keyhole to see if that page still holds the right key,
476 * pointing back to this stable node. This relies on freeing a PageAnon
477 * page to reset its page->mapping to NULL, and relies on no other use of
478 * a page to put something that might look like our key in page->mapping.
480 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
481 * but this is different - made simpler by ksm_thread_mutex being held, but
482 * interesting for assuming that no other use of the struct page could ever
483 * put our expected_mapping into page->mapping (or a field of the union which
484 * coincides with page->mapping). The RCU calls are not for KSM at all, but
485 * to keep the page_count protocol described with page_cache_get_speculative.
487 * Note: it is possible that get_ksm_page() will return NULL one moment,
488 * then page the next, if the page is in between page_freeze_refs() and
489 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
490 * is on its way to being freed; but it is an anomaly to bear in mind.
492 static struct page *get_ksm_page(struct stable_node *stable_node)
495 void *expected_mapping;
497 page = pfn_to_page(stable_node->kpfn);
498 expected_mapping = (void *)stable_node +
499 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
501 if (page->mapping != expected_mapping)
503 if (!get_page_unless_zero(page))
505 if (page->mapping != expected_mapping) {
513 remove_node_from_stable_tree(stable_node);
518 * Removing rmap_item from stable or unstable tree.
519 * This function will clean the information from the stable/unstable tree.
521 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
523 if (rmap_item->address & STABLE_FLAG) {
524 struct stable_node *stable_node;
527 stable_node = rmap_item->head;
528 page = get_ksm_page(stable_node);
533 hlist_del(&rmap_item->hlist);
537 if (stable_node->hlist.first)
542 put_anon_vma(rmap_item->anon_vma);
543 rmap_item->address &= PAGE_MASK;
545 } else if (rmap_item->address & UNSTABLE_FLAG) {
548 * Usually ksmd can and must skip the rb_erase, because
549 * root_unstable_tree was already reset to RB_ROOT.
550 * But be careful when an mm is exiting: do the rb_erase
551 * if this rmap_item was inserted by this scan, rather
552 * than left over from before.
554 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
557 rb_erase(&rmap_item->node, &root_unstable_tree);
559 ksm_pages_unshared--;
560 rmap_item->address &= PAGE_MASK;
563 cond_resched(); /* we're called from many long loops */
566 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
567 struct rmap_item **rmap_list)
570 struct rmap_item *rmap_item = *rmap_list;
571 *rmap_list = rmap_item->rmap_list;
572 remove_rmap_item_from_tree(rmap_item);
573 free_rmap_item(rmap_item);
578 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
579 * than check every pte of a given vma, the locking doesn't quite work for
580 * that - an rmap_item is assigned to the stable tree after inserting ksm
581 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
582 * rmap_items from parent to child at fork time (so as not to waste time
583 * if exit comes before the next scan reaches it).
585 * Similarly, although we'd like to remove rmap_items (so updating counts
586 * and freeing memory) when unmerging an area, it's easier to leave that
587 * to the next pass of ksmd - consider, for example, how ksmd might be
588 * in cmp_and_merge_page on one of the rmap_items we would be removing.
590 static int unmerge_ksm_pages(struct vm_area_struct *vma,
591 unsigned long start, unsigned long end)
596 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
597 if (ksm_test_exit(vma->vm_mm))
599 if (signal_pending(current))
602 err = break_ksm(vma, addr);
609 * Only called through the sysfs control interface:
611 static int unmerge_and_remove_all_rmap_items(void)
613 struct mm_slot *mm_slot;
614 struct mm_struct *mm;
615 struct vm_area_struct *vma;
618 spin_lock(&ksm_mmlist_lock);
619 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
620 struct mm_slot, mm_list);
621 spin_unlock(&ksm_mmlist_lock);
623 for (mm_slot = ksm_scan.mm_slot;
624 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
626 down_read(&mm->mmap_sem);
627 for (vma = mm->mmap; vma; vma = vma->vm_next) {
628 if (ksm_test_exit(mm))
630 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
632 err = unmerge_ksm_pages(vma,
633 vma->vm_start, vma->vm_end);
638 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
640 spin_lock(&ksm_mmlist_lock);
641 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
642 struct mm_slot, mm_list);
643 if (ksm_test_exit(mm)) {
644 hash_del(&mm_slot->link);
645 list_del(&mm_slot->mm_list);
646 spin_unlock(&ksm_mmlist_lock);
648 free_mm_slot(mm_slot);
649 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
650 up_read(&mm->mmap_sem);
653 spin_unlock(&ksm_mmlist_lock);
654 up_read(&mm->mmap_sem);
662 up_read(&mm->mmap_sem);
663 spin_lock(&ksm_mmlist_lock);
664 ksm_scan.mm_slot = &ksm_mm_head;
665 spin_unlock(&ksm_mmlist_lock);
668 #endif /* CONFIG_SYSFS */
670 static u32 calc_checksum(struct page *page)
673 void *addr = kmap_atomic(page);
674 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
679 static int memcmp_pages(struct page *page1, struct page *page2)
684 addr1 = kmap_atomic(page1);
685 addr2 = kmap_atomic(page2);
686 ret = memcmp(addr1, addr2, PAGE_SIZE);
687 kunmap_atomic(addr2);
688 kunmap_atomic(addr1);
692 static inline int pages_identical(struct page *page1, struct page *page2)
694 return !memcmp_pages(page1, page2);
697 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
700 struct mm_struct *mm = vma->vm_mm;
706 unsigned long mmun_start; /* For mmu_notifiers */
707 unsigned long mmun_end; /* For mmu_notifiers */
709 addr = page_address_in_vma(page, vma);
713 BUG_ON(PageTransCompound(page));
716 mmun_end = addr + PAGE_SIZE;
717 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
719 ptep = page_check_address(page, mm, addr, &ptl, 0);
723 if (pte_write(*ptep) || pte_dirty(*ptep)) {
726 swapped = PageSwapCache(page);
727 flush_cache_page(vma, addr, page_to_pfn(page));
729 * Ok this is tricky, when get_user_pages_fast() run it doesn't
730 * take any lock, therefore the check that we are going to make
731 * with the pagecount against the mapcount is racey and
732 * O_DIRECT can happen right after the check.
733 * So we clear the pte and flush the tlb before the check
734 * this assure us that no O_DIRECT can happen after the check
735 * or in the middle of the check.
737 entry = ptep_clear_flush(vma, addr, ptep);
739 * Check that no O_DIRECT or similar I/O is in progress on the
742 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
743 set_pte_at(mm, addr, ptep, entry);
746 if (pte_dirty(entry))
747 set_page_dirty(page);
748 entry = pte_mkclean(pte_wrprotect(entry));
749 set_pte_at_notify(mm, addr, ptep, entry);
755 pte_unmap_unlock(ptep, ptl);
757 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
763 * replace_page - replace page in vma by new ksm page
764 * @vma: vma that holds the pte pointing to page
765 * @page: the page we are replacing by kpage
766 * @kpage: the ksm page we replace page by
767 * @orig_pte: the original value of the pte
769 * Returns 0 on success, -EFAULT on failure.
771 static int replace_page(struct vm_area_struct *vma, struct page *page,
772 struct page *kpage, pte_t orig_pte)
774 struct mm_struct *mm = vma->vm_mm;
780 unsigned long mmun_start; /* For mmu_notifiers */
781 unsigned long mmun_end; /* For mmu_notifiers */
783 addr = page_address_in_vma(page, vma);
787 pmd = mm_find_pmd(mm, addr);
790 BUG_ON(pmd_trans_huge(*pmd));
793 mmun_end = addr + PAGE_SIZE;
794 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
796 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
797 if (!pte_same(*ptep, orig_pte)) {
798 pte_unmap_unlock(ptep, ptl);
803 page_add_anon_rmap(kpage, vma, addr);
805 flush_cache_page(vma, addr, pte_pfn(*ptep));
806 ptep_clear_flush(vma, addr, ptep);
807 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
809 page_remove_rmap(page);
810 if (!page_mapped(page))
811 try_to_free_swap(page);
814 pte_unmap_unlock(ptep, ptl);
817 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
822 static int page_trans_compound_anon_split(struct page *page)
825 struct page *transhuge_head = page_trans_compound_anon(page);
826 if (transhuge_head) {
827 /* Get the reference on the head to split it. */
828 if (get_page_unless_zero(transhuge_head)) {
830 * Recheck we got the reference while the head
831 * was still anonymous.
833 if (PageAnon(transhuge_head))
834 ret = split_huge_page(transhuge_head);
837 * Retry later if split_huge_page run
841 put_page(transhuge_head);
843 /* Retry later if split_huge_page run from under us. */
850 * try_to_merge_one_page - take two pages and merge them into one
851 * @vma: the vma that holds the pte pointing to page
852 * @page: the PageAnon page that we want to replace with kpage
853 * @kpage: the PageKsm page that we want to map instead of page,
854 * or NULL the first time when we want to use page as kpage.
856 * This function returns 0 if the pages were merged, -EFAULT otherwise.
858 static int try_to_merge_one_page(struct vm_area_struct *vma,
859 struct page *page, struct page *kpage)
861 pte_t orig_pte = __pte(0);
864 if (page == kpage) /* ksm page forked */
867 if (!(vma->vm_flags & VM_MERGEABLE))
869 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
871 BUG_ON(PageTransCompound(page));
876 * We need the page lock to read a stable PageSwapCache in
877 * write_protect_page(). We use trylock_page() instead of
878 * lock_page() because we don't want to wait here - we
879 * prefer to continue scanning and merging different pages,
880 * then come back to this page when it is unlocked.
882 if (!trylock_page(page))
885 * If this anonymous page is mapped only here, its pte may need
886 * to be write-protected. If it's mapped elsewhere, all of its
887 * ptes are necessarily already write-protected. But in either
888 * case, we need to lock and check page_count is not raised.
890 if (write_protect_page(vma, page, &orig_pte) == 0) {
893 * While we hold page lock, upgrade page from
894 * PageAnon+anon_vma to PageKsm+NULL stable_node:
895 * stable_tree_insert() will update stable_node.
897 set_page_stable_node(page, NULL);
898 mark_page_accessed(page);
900 } else if (pages_identical(page, kpage))
901 err = replace_page(vma, page, kpage, orig_pte);
904 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
905 munlock_vma_page(page);
906 if (!PageMlocked(kpage)) {
909 mlock_vma_page(kpage);
910 page = kpage; /* for final unlock */
920 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
921 * but no new kernel page is allocated: kpage must already be a ksm page.
923 * This function returns 0 if the pages were merged, -EFAULT otherwise.
925 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
926 struct page *page, struct page *kpage)
928 struct mm_struct *mm = rmap_item->mm;
929 struct vm_area_struct *vma;
932 down_read(&mm->mmap_sem);
933 if (ksm_test_exit(mm))
935 vma = find_vma(mm, rmap_item->address);
936 if (!vma || vma->vm_start > rmap_item->address)
939 err = try_to_merge_one_page(vma, page, kpage);
943 /* Must get reference to anon_vma while still holding mmap_sem */
944 rmap_item->anon_vma = vma->anon_vma;
945 get_anon_vma(vma->anon_vma);
947 up_read(&mm->mmap_sem);
952 * try_to_merge_two_pages - take two identical pages and prepare them
953 * to be merged into one page.
955 * This function returns the kpage if we successfully merged two identical
956 * pages into one ksm page, NULL otherwise.
958 * Note that this function upgrades page to ksm page: if one of the pages
959 * is already a ksm page, try_to_merge_with_ksm_page should be used.
961 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
963 struct rmap_item *tree_rmap_item,
964 struct page *tree_page)
968 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
970 err = try_to_merge_with_ksm_page(tree_rmap_item,
973 * If that fails, we have a ksm page with only one pte
974 * pointing to it: so break it.
977 break_cow(rmap_item);
979 return err ? NULL : page;
983 * stable_tree_search - search for page inside the stable tree
985 * This function checks if there is a page inside the stable tree
986 * with identical content to the page that we are scanning right now.
988 * This function returns the stable tree node of identical content if found,
991 static struct page *stable_tree_search(struct page *page)
993 struct rb_node *node = root_stable_tree.rb_node;
994 struct stable_node *stable_node;
996 stable_node = page_stable_node(page);
997 if (stable_node) { /* ksm page forked */
1003 struct page *tree_page;
1007 stable_node = rb_entry(node, struct stable_node, node);
1008 tree_page = get_ksm_page(stable_node);
1012 ret = memcmp_pages(page, tree_page);
1015 put_page(tree_page);
1016 node = node->rb_left;
1017 } else if (ret > 0) {
1018 put_page(tree_page);
1019 node = node->rb_right;
1028 * stable_tree_insert - insert rmap_item pointing to new ksm page
1029 * into the stable tree.
1031 * This function returns the stable tree node just allocated on success,
1034 static struct stable_node *stable_tree_insert(struct page *kpage)
1036 struct rb_node **new = &root_stable_tree.rb_node;
1037 struct rb_node *parent = NULL;
1038 struct stable_node *stable_node;
1041 struct page *tree_page;
1045 stable_node = rb_entry(*new, struct stable_node, node);
1046 tree_page = get_ksm_page(stable_node);
1050 ret = memcmp_pages(kpage, tree_page);
1051 put_page(tree_page);
1055 new = &parent->rb_left;
1057 new = &parent->rb_right;
1060 * It is not a bug that stable_tree_search() didn't
1061 * find this node: because at that time our page was
1062 * not yet write-protected, so may have changed since.
1068 stable_node = alloc_stable_node();
1072 rb_link_node(&stable_node->node, parent, new);
1073 rb_insert_color(&stable_node->node, &root_stable_tree);
1075 INIT_HLIST_HEAD(&stable_node->hlist);
1077 stable_node->kpfn = page_to_pfn(kpage);
1078 set_page_stable_node(kpage, stable_node);
1084 * unstable_tree_search_insert - search for identical page,
1085 * else insert rmap_item into the unstable tree.
1087 * This function searches for a page in the unstable tree identical to the
1088 * page currently being scanned; and if no identical page is found in the
1089 * tree, we insert rmap_item as a new object into the unstable tree.
1091 * This function returns pointer to rmap_item found to be identical
1092 * to the currently scanned page, NULL otherwise.
1094 * This function does both searching and inserting, because they share
1095 * the same walking algorithm in an rbtree.
1098 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1100 struct page **tree_pagep)
1103 struct rb_node **new = &root_unstable_tree.rb_node;
1104 struct rb_node *parent = NULL;
1107 struct rmap_item *tree_rmap_item;
1108 struct page *tree_page;
1112 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1113 tree_page = get_mergeable_page(tree_rmap_item);
1114 if (IS_ERR_OR_NULL(tree_page))
1118 * Don't substitute a ksm page for a forked page.
1120 if (page == tree_page) {
1121 put_page(tree_page);
1125 ret = memcmp_pages(page, tree_page);
1129 put_page(tree_page);
1130 new = &parent->rb_left;
1131 } else if (ret > 0) {
1132 put_page(tree_page);
1133 new = &parent->rb_right;
1135 *tree_pagep = tree_page;
1136 return tree_rmap_item;
1140 rmap_item->address |= UNSTABLE_FLAG;
1141 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1142 rb_link_node(&rmap_item->node, parent, new);
1143 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1145 ksm_pages_unshared++;
1150 * stable_tree_append - add another rmap_item to the linked list of
1151 * rmap_items hanging off a given node of the stable tree, all sharing
1152 * the same ksm page.
1154 static void stable_tree_append(struct rmap_item *rmap_item,
1155 struct stable_node *stable_node)
1157 rmap_item->head = stable_node;
1158 rmap_item->address |= STABLE_FLAG;
1159 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1161 if (rmap_item->hlist.next)
1162 ksm_pages_sharing++;
1168 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1169 * if not, compare checksum to previous and if it's the same, see if page can
1170 * be inserted into the unstable tree, or merged with a page already there and
1171 * both transferred to the stable tree.
1173 * @page: the page that we are searching identical page to.
1174 * @rmap_item: the reverse mapping into the virtual address of this page
1176 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1178 struct rmap_item *tree_rmap_item;
1179 struct page *tree_page = NULL;
1180 struct stable_node *stable_node;
1182 unsigned int checksum;
1185 remove_rmap_item_from_tree(rmap_item);
1187 /* We first start with searching the page inside the stable tree */
1188 kpage = stable_tree_search(page);
1190 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1193 * The page was successfully merged:
1194 * add its rmap_item to the stable tree.
1197 stable_tree_append(rmap_item, page_stable_node(kpage));
1205 * If the hash value of the page has changed from the last time
1206 * we calculated it, this page is changing frequently: therefore we
1207 * don't want to insert it in the unstable tree, and we don't want
1208 * to waste our time searching for something identical to it there.
1210 checksum = calc_checksum(page);
1211 if (rmap_item->oldchecksum != checksum) {
1212 rmap_item->oldchecksum = checksum;
1217 unstable_tree_search_insert(rmap_item, page, &tree_page);
1218 if (tree_rmap_item) {
1219 kpage = try_to_merge_two_pages(rmap_item, page,
1220 tree_rmap_item, tree_page);
1221 put_page(tree_page);
1223 * As soon as we merge this page, we want to remove the
1224 * rmap_item of the page we have merged with from the unstable
1225 * tree, and insert it instead as new node in the stable tree.
1228 remove_rmap_item_from_tree(tree_rmap_item);
1231 stable_node = stable_tree_insert(kpage);
1233 stable_tree_append(tree_rmap_item, stable_node);
1234 stable_tree_append(rmap_item, stable_node);
1239 * If we fail to insert the page into the stable tree,
1240 * we will have 2 virtual addresses that are pointing
1241 * to a ksm page left outside the stable tree,
1242 * in which case we need to break_cow on both.
1245 break_cow(tree_rmap_item);
1246 break_cow(rmap_item);
1252 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1253 struct rmap_item **rmap_list,
1256 struct rmap_item *rmap_item;
1258 while (*rmap_list) {
1259 rmap_item = *rmap_list;
1260 if ((rmap_item->address & PAGE_MASK) == addr)
1262 if (rmap_item->address > addr)
1264 *rmap_list = rmap_item->rmap_list;
1265 remove_rmap_item_from_tree(rmap_item);
1266 free_rmap_item(rmap_item);
1269 rmap_item = alloc_rmap_item();
1271 /* It has already been zeroed */
1272 rmap_item->mm = mm_slot->mm;
1273 rmap_item->address = addr;
1274 rmap_item->rmap_list = *rmap_list;
1275 *rmap_list = rmap_item;
1280 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1282 struct mm_struct *mm;
1283 struct mm_slot *slot;
1284 struct vm_area_struct *vma;
1285 struct rmap_item *rmap_item;
1287 if (list_empty(&ksm_mm_head.mm_list))
1290 slot = ksm_scan.mm_slot;
1291 if (slot == &ksm_mm_head) {
1293 * A number of pages can hang around indefinitely on per-cpu
1294 * pagevecs, raised page count preventing write_protect_page
1295 * from merging them. Though it doesn't really matter much,
1296 * it is puzzling to see some stuck in pages_volatile until
1297 * other activity jostles them out, and they also prevented
1298 * LTP's KSM test from succeeding deterministically; so drain
1299 * them here (here rather than on entry to ksm_do_scan(),
1300 * so we don't IPI too often when pages_to_scan is set low).
1302 lru_add_drain_all();
1304 root_unstable_tree = RB_ROOT;
1306 spin_lock(&ksm_mmlist_lock);
1307 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1308 ksm_scan.mm_slot = slot;
1309 spin_unlock(&ksm_mmlist_lock);
1311 * Although we tested list_empty() above, a racing __ksm_exit
1312 * of the last mm on the list may have removed it since then.
1314 if (slot == &ksm_mm_head)
1317 ksm_scan.address = 0;
1318 ksm_scan.rmap_list = &slot->rmap_list;
1322 down_read(&mm->mmap_sem);
1323 if (ksm_test_exit(mm))
1326 vma = find_vma(mm, ksm_scan.address);
1328 for (; vma; vma = vma->vm_next) {
1329 if (!(vma->vm_flags & VM_MERGEABLE))
1331 if (ksm_scan.address < vma->vm_start)
1332 ksm_scan.address = vma->vm_start;
1334 ksm_scan.address = vma->vm_end;
1336 while (ksm_scan.address < vma->vm_end) {
1337 if (ksm_test_exit(mm))
1339 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1340 if (IS_ERR_OR_NULL(*page)) {
1341 ksm_scan.address += PAGE_SIZE;
1345 if (PageAnon(*page) ||
1346 page_trans_compound_anon(*page)) {
1347 flush_anon_page(vma, *page, ksm_scan.address);
1348 flush_dcache_page(*page);
1349 rmap_item = get_next_rmap_item(slot,
1350 ksm_scan.rmap_list, ksm_scan.address);
1352 ksm_scan.rmap_list =
1353 &rmap_item->rmap_list;
1354 ksm_scan.address += PAGE_SIZE;
1357 up_read(&mm->mmap_sem);
1361 ksm_scan.address += PAGE_SIZE;
1366 if (ksm_test_exit(mm)) {
1367 ksm_scan.address = 0;
1368 ksm_scan.rmap_list = &slot->rmap_list;
1371 * Nuke all the rmap_items that are above this current rmap:
1372 * because there were no VM_MERGEABLE vmas with such addresses.
1374 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1376 spin_lock(&ksm_mmlist_lock);
1377 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1378 struct mm_slot, mm_list);
1379 if (ksm_scan.address == 0) {
1381 * We've completed a full scan of all vmas, holding mmap_sem
1382 * throughout, and found no VM_MERGEABLE: so do the same as
1383 * __ksm_exit does to remove this mm from all our lists now.
1384 * This applies either when cleaning up after __ksm_exit
1385 * (but beware: we can reach here even before __ksm_exit),
1386 * or when all VM_MERGEABLE areas have been unmapped (and
1387 * mmap_sem then protects against race with MADV_MERGEABLE).
1389 hash_del(&slot->link);
1390 list_del(&slot->mm_list);
1391 spin_unlock(&ksm_mmlist_lock);
1394 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1395 up_read(&mm->mmap_sem);
1398 spin_unlock(&ksm_mmlist_lock);
1399 up_read(&mm->mmap_sem);
1402 /* Repeat until we've completed scanning the whole list */
1403 slot = ksm_scan.mm_slot;
1404 if (slot != &ksm_mm_head)
1412 * ksm_do_scan - the ksm scanner main worker function.
1413 * @scan_npages - number of pages we want to scan before we return.
1415 static void ksm_do_scan(unsigned int scan_npages)
1417 struct rmap_item *rmap_item;
1418 struct page *uninitialized_var(page);
1420 while (scan_npages-- && likely(!freezing(current))) {
1422 rmap_item = scan_get_next_rmap_item(&page);
1425 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1426 cmp_and_merge_page(page, rmap_item);
1431 static int ksmd_should_run(void)
1433 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1436 static int ksm_scan_thread(void *nothing)
1439 set_user_nice(current, 5);
1441 while (!kthread_should_stop()) {
1442 mutex_lock(&ksm_thread_mutex);
1443 if (ksmd_should_run())
1444 ksm_do_scan(ksm_thread_pages_to_scan);
1445 mutex_unlock(&ksm_thread_mutex);
1449 if (ksmd_should_run()) {
1450 schedule_timeout_interruptible(
1451 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1453 wait_event_freezable(ksm_thread_wait,
1454 ksmd_should_run() || kthread_should_stop());
1460 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1461 unsigned long end, int advice, unsigned long *vm_flags)
1463 struct mm_struct *mm = vma->vm_mm;
1467 case MADV_MERGEABLE:
1469 * Be somewhat over-protective for now!
1471 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1472 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1473 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1474 return 0; /* just ignore the advice */
1477 if (*vm_flags & VM_SAO)
1481 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1482 err = __ksm_enter(mm);
1487 *vm_flags |= VM_MERGEABLE;
1490 case MADV_UNMERGEABLE:
1491 if (!(*vm_flags & VM_MERGEABLE))
1492 return 0; /* just ignore the advice */
1494 if (vma->anon_vma) {
1495 err = unmerge_ksm_pages(vma, start, end);
1500 *vm_flags &= ~VM_MERGEABLE;
1507 int __ksm_enter(struct mm_struct *mm)
1509 struct mm_slot *mm_slot;
1512 mm_slot = alloc_mm_slot();
1516 /* Check ksm_run too? Would need tighter locking */
1517 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1519 spin_lock(&ksm_mmlist_lock);
1520 insert_to_mm_slots_hash(mm, mm_slot);
1522 * Insert just behind the scanning cursor, to let the area settle
1523 * down a little; when fork is followed by immediate exec, we don't
1524 * want ksmd to waste time setting up and tearing down an rmap_list.
1526 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1527 spin_unlock(&ksm_mmlist_lock);
1529 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1530 atomic_inc(&mm->mm_count);
1533 wake_up_interruptible(&ksm_thread_wait);
1538 void __ksm_exit(struct mm_struct *mm)
1540 struct mm_slot *mm_slot;
1541 int easy_to_free = 0;
1544 * This process is exiting: if it's straightforward (as is the
1545 * case when ksmd was never running), free mm_slot immediately.
1546 * But if it's at the cursor or has rmap_items linked to it, use
1547 * mmap_sem to synchronize with any break_cows before pagetables
1548 * are freed, and leave the mm_slot on the list for ksmd to free.
1549 * Beware: ksm may already have noticed it exiting and freed the slot.
1552 spin_lock(&ksm_mmlist_lock);
1553 mm_slot = get_mm_slot(mm);
1554 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1555 if (!mm_slot->rmap_list) {
1556 hash_del(&mm_slot->link);
1557 list_del(&mm_slot->mm_list);
1560 list_move(&mm_slot->mm_list,
1561 &ksm_scan.mm_slot->mm_list);
1564 spin_unlock(&ksm_mmlist_lock);
1567 free_mm_slot(mm_slot);
1568 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1570 } else if (mm_slot) {
1571 down_write(&mm->mmap_sem);
1572 up_write(&mm->mmap_sem);
1576 struct page *ksm_does_need_to_copy(struct page *page,
1577 struct vm_area_struct *vma, unsigned long address)
1579 struct page *new_page;
1581 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1583 copy_user_highpage(new_page, page, address, vma);
1585 SetPageDirty(new_page);
1586 __SetPageUptodate(new_page);
1587 __set_page_locked(new_page);
1593 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1594 unsigned long *vm_flags)
1596 struct stable_node *stable_node;
1597 struct rmap_item *rmap_item;
1598 struct hlist_node *hlist;
1599 unsigned int mapcount = page_mapcount(page);
1601 int search_new_forks = 0;
1603 VM_BUG_ON(!PageKsm(page));
1604 VM_BUG_ON(!PageLocked(page));
1606 stable_node = page_stable_node(page);
1610 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1611 struct anon_vma *anon_vma = rmap_item->anon_vma;
1612 struct anon_vma_chain *vmac;
1613 struct vm_area_struct *vma;
1615 anon_vma_lock_read(anon_vma);
1616 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1619 if (rmap_item->address < vma->vm_start ||
1620 rmap_item->address >= vma->vm_end)
1623 * Initially we examine only the vma which covers this
1624 * rmap_item; but later, if there is still work to do,
1625 * we examine covering vmas in other mms: in case they
1626 * were forked from the original since ksmd passed.
1628 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1631 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1634 referenced += page_referenced_one(page, vma,
1635 rmap_item->address, &mapcount, vm_flags);
1636 if (!search_new_forks || !mapcount)
1639 anon_vma_unlock_read(anon_vma);
1643 if (!search_new_forks++)
1649 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1651 struct stable_node *stable_node;
1652 struct hlist_node *hlist;
1653 struct rmap_item *rmap_item;
1654 int ret = SWAP_AGAIN;
1655 int search_new_forks = 0;
1657 VM_BUG_ON(!PageKsm(page));
1658 VM_BUG_ON(!PageLocked(page));
1660 stable_node = page_stable_node(page);
1664 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1665 struct anon_vma *anon_vma = rmap_item->anon_vma;
1666 struct anon_vma_chain *vmac;
1667 struct vm_area_struct *vma;
1669 anon_vma_lock_read(anon_vma);
1670 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1673 if (rmap_item->address < vma->vm_start ||
1674 rmap_item->address >= vma->vm_end)
1677 * Initially we examine only the vma which covers this
1678 * rmap_item; but later, if there is still work to do,
1679 * we examine covering vmas in other mms: in case they
1680 * were forked from the original since ksmd passed.
1682 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1685 ret = try_to_unmap_one(page, vma,
1686 rmap_item->address, flags);
1687 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1688 anon_vma_unlock_read(anon_vma);
1692 anon_vma_unlock_read(anon_vma);
1694 if (!search_new_forks++)
1700 #ifdef CONFIG_MIGRATION
1701 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1702 struct vm_area_struct *, unsigned long, void *), void *arg)
1704 struct stable_node *stable_node;
1705 struct hlist_node *hlist;
1706 struct rmap_item *rmap_item;
1707 int ret = SWAP_AGAIN;
1708 int search_new_forks = 0;
1710 VM_BUG_ON(!PageKsm(page));
1711 VM_BUG_ON(!PageLocked(page));
1713 stable_node = page_stable_node(page);
1717 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1718 struct anon_vma *anon_vma = rmap_item->anon_vma;
1719 struct anon_vma_chain *vmac;
1720 struct vm_area_struct *vma;
1722 anon_vma_lock_read(anon_vma);
1723 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1726 if (rmap_item->address < vma->vm_start ||
1727 rmap_item->address >= vma->vm_end)
1730 * Initially we examine only the vma which covers this
1731 * rmap_item; but later, if there is still work to do,
1732 * we examine covering vmas in other mms: in case they
1733 * were forked from the original since ksmd passed.
1735 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1738 ret = rmap_one(page, vma, rmap_item->address, arg);
1739 if (ret != SWAP_AGAIN) {
1740 anon_vma_unlock_read(anon_vma);
1744 anon_vma_unlock_read(anon_vma);
1746 if (!search_new_forks++)
1752 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1754 struct stable_node *stable_node;
1756 VM_BUG_ON(!PageLocked(oldpage));
1757 VM_BUG_ON(!PageLocked(newpage));
1758 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1760 stable_node = page_stable_node(newpage);
1762 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1763 stable_node->kpfn = page_to_pfn(newpage);
1766 #endif /* CONFIG_MIGRATION */
1768 #ifdef CONFIG_MEMORY_HOTREMOVE
1769 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1770 unsigned long end_pfn)
1772 struct rb_node *node;
1774 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1775 struct stable_node *stable_node;
1777 stable_node = rb_entry(node, struct stable_node, node);
1778 if (stable_node->kpfn >= start_pfn &&
1779 stable_node->kpfn < end_pfn)
1785 static int ksm_memory_callback(struct notifier_block *self,
1786 unsigned long action, void *arg)
1788 struct memory_notify *mn = arg;
1789 struct stable_node *stable_node;
1792 case MEM_GOING_OFFLINE:
1794 * Keep it very simple for now: just lock out ksmd and
1795 * MADV_UNMERGEABLE while any memory is going offline.
1796 * mutex_lock_nested() is necessary because lockdep was alarmed
1797 * that here we take ksm_thread_mutex inside notifier chain
1798 * mutex, and later take notifier chain mutex inside
1799 * ksm_thread_mutex to unlock it. But that's safe because both
1800 * are inside mem_hotplug_mutex.
1802 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1807 * Most of the work is done by page migration; but there might
1808 * be a few stable_nodes left over, still pointing to struct
1809 * pages which have been offlined: prune those from the tree.
1811 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1812 mn->start_pfn + mn->nr_pages)) != NULL)
1813 remove_node_from_stable_tree(stable_node);
1816 case MEM_CANCEL_OFFLINE:
1817 mutex_unlock(&ksm_thread_mutex);
1822 #endif /* CONFIG_MEMORY_HOTREMOVE */
1826 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1829 #define KSM_ATTR_RO(_name) \
1830 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1831 #define KSM_ATTR(_name) \
1832 static struct kobj_attribute _name##_attr = \
1833 __ATTR(_name, 0644, _name##_show, _name##_store)
1835 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1836 struct kobj_attribute *attr, char *buf)
1838 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1841 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1842 struct kobj_attribute *attr,
1843 const char *buf, size_t count)
1845 unsigned long msecs;
1848 err = strict_strtoul(buf, 10, &msecs);
1849 if (err || msecs > UINT_MAX)
1852 ksm_thread_sleep_millisecs = msecs;
1856 KSM_ATTR(sleep_millisecs);
1858 static ssize_t pages_to_scan_show(struct kobject *kobj,
1859 struct kobj_attribute *attr, char *buf)
1861 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1864 static ssize_t pages_to_scan_store(struct kobject *kobj,
1865 struct kobj_attribute *attr,
1866 const char *buf, size_t count)
1869 unsigned long nr_pages;
1871 err = strict_strtoul(buf, 10, &nr_pages);
1872 if (err || nr_pages > UINT_MAX)
1875 ksm_thread_pages_to_scan = nr_pages;
1879 KSM_ATTR(pages_to_scan);
1881 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1884 return sprintf(buf, "%u\n", ksm_run);
1887 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1888 const char *buf, size_t count)
1891 unsigned long flags;
1893 err = strict_strtoul(buf, 10, &flags);
1894 if (err || flags > UINT_MAX)
1896 if (flags > KSM_RUN_UNMERGE)
1900 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1901 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1902 * breaking COW to free the pages_shared (but leaves mm_slots
1903 * on the list for when ksmd may be set running again).
1906 mutex_lock(&ksm_thread_mutex);
1907 if (ksm_run != flags) {
1909 if (flags & KSM_RUN_UNMERGE) {
1910 set_current_oom_origin();
1911 err = unmerge_and_remove_all_rmap_items();
1912 clear_current_oom_origin();
1914 ksm_run = KSM_RUN_STOP;
1919 mutex_unlock(&ksm_thread_mutex);
1921 if (flags & KSM_RUN_MERGE)
1922 wake_up_interruptible(&ksm_thread_wait);
1928 static ssize_t pages_shared_show(struct kobject *kobj,
1929 struct kobj_attribute *attr, char *buf)
1931 return sprintf(buf, "%lu\n", ksm_pages_shared);
1933 KSM_ATTR_RO(pages_shared);
1935 static ssize_t pages_sharing_show(struct kobject *kobj,
1936 struct kobj_attribute *attr, char *buf)
1938 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1940 KSM_ATTR_RO(pages_sharing);
1942 static ssize_t pages_unshared_show(struct kobject *kobj,
1943 struct kobj_attribute *attr, char *buf)
1945 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1947 KSM_ATTR_RO(pages_unshared);
1949 static ssize_t pages_volatile_show(struct kobject *kobj,
1950 struct kobj_attribute *attr, char *buf)
1952 long ksm_pages_volatile;
1954 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1955 - ksm_pages_sharing - ksm_pages_unshared;
1957 * It was not worth any locking to calculate that statistic,
1958 * but it might therefore sometimes be negative: conceal that.
1960 if (ksm_pages_volatile < 0)
1961 ksm_pages_volatile = 0;
1962 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1964 KSM_ATTR_RO(pages_volatile);
1966 static ssize_t full_scans_show(struct kobject *kobj,
1967 struct kobj_attribute *attr, char *buf)
1969 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1971 KSM_ATTR_RO(full_scans);
1973 static struct attribute *ksm_attrs[] = {
1974 &sleep_millisecs_attr.attr,
1975 &pages_to_scan_attr.attr,
1977 &pages_shared_attr.attr,
1978 &pages_sharing_attr.attr,
1979 &pages_unshared_attr.attr,
1980 &pages_volatile_attr.attr,
1981 &full_scans_attr.attr,
1985 static struct attribute_group ksm_attr_group = {
1989 #endif /* CONFIG_SYSFS */
1991 static int __init ksm_init(void)
1993 struct task_struct *ksm_thread;
1996 err = ksm_slab_init();
2000 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2001 if (IS_ERR(ksm_thread)) {
2002 printk(KERN_ERR "ksm: creating kthread failed\n");
2003 err = PTR_ERR(ksm_thread);
2008 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2010 printk(KERN_ERR "ksm: register sysfs failed\n");
2011 kthread_stop(ksm_thread);
2015 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2017 #endif /* CONFIG_SYSFS */
2019 #ifdef CONFIG_MEMORY_HOTREMOVE
2021 * Choose a high priority since the callback takes ksm_thread_mutex:
2022 * later callbacks could only be taking locks which nest within that.
2024 hotplug_memory_notifier(ksm_memory_callback, 100);
2033 module_init(ksm_init)