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>
39 #include <linux/numa.h>
41 #include <asm/tlbflush.h>
46 #define DO_NUMA(x) do { (x); } while (0)
49 #define DO_NUMA(x) do { } while (0)
53 * A few notes about the KSM scanning process,
54 * to make it easier to understand the data structures below:
56 * In order to reduce excessive scanning, KSM sorts the memory pages by their
57 * contents into a data structure that holds pointers to the pages' locations.
59 * Since the contents of the pages may change at any moment, KSM cannot just
60 * insert the pages into a normal sorted tree and expect it to find anything.
61 * Therefore KSM uses two data structures - the stable and the unstable tree.
63 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64 * by their contents. Because each such page is write-protected, searching on
65 * this tree is fully assured to be working (except when pages are unmapped),
66 * and therefore this tree is called the stable tree.
68 * In addition to the stable tree, KSM uses a second data structure called the
69 * unstable tree: this tree holds pointers to pages which have been found to
70 * be "unchanged for a period of time". The unstable tree sorts these pages
71 * by their contents, but since they are not write-protected, KSM cannot rely
72 * upon the unstable tree to work correctly - the unstable tree is liable to
73 * be corrupted as its contents are modified, and so it is called unstable.
75 * KSM solves this problem by several techniques:
77 * 1) The unstable tree is flushed every time KSM completes scanning all
78 * memory areas, and then the tree is rebuilt again from the beginning.
79 * 2) KSM will only insert into the unstable tree, pages whose hash value
80 * has not changed since the previous scan of all memory areas.
81 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82 * colors of the nodes and not on their contents, assuring that even when
83 * the tree gets "corrupted" it won't get out of balance, so scanning time
84 * remains the same (also, searching and inserting nodes in an rbtree uses
85 * the same algorithm, so we have no overhead when we flush and rebuild).
86 * 4) KSM never flushes the stable tree, which means that even if it were to
87 * take 10 attempts to find a page in the unstable tree, once it is found,
88 * it is secured in the stable tree. (When we scan a new page, we first
89 * compare it against the stable tree, and then against the unstable tree.)
93 * struct mm_slot - ksm information per mm that is being scanned
94 * @link: link to the mm_slots hash list
95 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
96 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
97 * @mm: the mm that this information is valid for
100 struct hlist_node link;
101 struct list_head mm_list;
102 struct rmap_item *rmap_list;
103 struct mm_struct *mm;
107 * struct ksm_scan - cursor for scanning
108 * @mm_slot: the current mm_slot we are scanning
109 * @address: the next address inside that to be scanned
110 * @rmap_list: link to the next rmap to be scanned in the rmap_list
111 * @seqnr: count of completed full scans (needed when removing unstable node)
113 * There is only the one ksm_scan instance of this cursor structure.
116 struct mm_slot *mm_slot;
117 unsigned long address;
118 struct rmap_item **rmap_list;
123 * struct stable_node - node of the stable rbtree
124 * @node: rb node of this ksm page in the stable tree
125 * @hlist: hlist head of rmap_items using this ksm page
126 * @kpfn: page frame number of this ksm page
130 struct hlist_head hlist;
135 * struct rmap_item - reverse mapping item for virtual addresses
136 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
137 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
138 * @mm: the memory structure this rmap_item is pointing into
139 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
140 * @oldchecksum: previous checksum of the page at that virtual address
141 * @nid: NUMA node id of unstable tree in which linked (may not match page)
142 * @node: rb node of this rmap_item in the unstable tree
143 * @head: pointer to stable_node heading this list in the stable tree
144 * @hlist: link into hlist of rmap_items hanging off that stable_node
147 struct rmap_item *rmap_list;
148 struct anon_vma *anon_vma; /* when stable */
149 struct mm_struct *mm;
150 unsigned long address; /* + low bits used for flags below */
151 unsigned int oldchecksum; /* when unstable */
156 struct rb_node node; /* when node of unstable tree */
157 struct { /* when listed from stable tree */
158 struct stable_node *head;
159 struct hlist_node hlist;
164 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
165 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
166 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
168 /* The stable and unstable tree heads */
169 static struct rb_root root_unstable_tree[MAX_NUMNODES];
170 static struct rb_root root_stable_tree[MAX_NUMNODES];
172 #define MM_SLOTS_HASH_BITS 10
173 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
175 static struct mm_slot ksm_mm_head = {
176 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
178 static struct ksm_scan ksm_scan = {
179 .mm_slot = &ksm_mm_head,
182 static struct kmem_cache *rmap_item_cache;
183 static struct kmem_cache *stable_node_cache;
184 static struct kmem_cache *mm_slot_cache;
186 /* The number of nodes in the stable tree */
187 static unsigned long ksm_pages_shared;
189 /* The number of page slots additionally sharing those nodes */
190 static unsigned long ksm_pages_sharing;
192 /* The number of nodes in the unstable tree */
193 static unsigned long ksm_pages_unshared;
195 /* The number of rmap_items in use: to calculate pages_volatile */
196 static unsigned long ksm_rmap_items;
198 /* Number of pages ksmd should scan in one batch */
199 static unsigned int ksm_thread_pages_to_scan = 100;
201 /* Milliseconds ksmd should sleep between batches */
202 static unsigned int ksm_thread_sleep_millisecs = 20;
205 /* Zeroed when merging across nodes is not allowed */
206 static unsigned int ksm_merge_across_nodes = 1;
208 #define ksm_merge_across_nodes 1U
211 #define KSM_RUN_STOP 0
212 #define KSM_RUN_MERGE 1
213 #define KSM_RUN_UNMERGE 2
214 static unsigned int ksm_run = KSM_RUN_STOP;
216 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
217 static DEFINE_MUTEX(ksm_thread_mutex);
218 static DEFINE_SPINLOCK(ksm_mmlist_lock);
220 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
221 sizeof(struct __struct), __alignof__(struct __struct),\
224 static int __init ksm_slab_init(void)
226 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
227 if (!rmap_item_cache)
230 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
231 if (!stable_node_cache)
234 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
241 kmem_cache_destroy(stable_node_cache);
243 kmem_cache_destroy(rmap_item_cache);
248 static void __init ksm_slab_free(void)
250 kmem_cache_destroy(mm_slot_cache);
251 kmem_cache_destroy(stable_node_cache);
252 kmem_cache_destroy(rmap_item_cache);
253 mm_slot_cache = NULL;
256 static inline struct rmap_item *alloc_rmap_item(void)
258 struct rmap_item *rmap_item;
260 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
266 static inline void free_rmap_item(struct rmap_item *rmap_item)
269 rmap_item->mm = NULL; /* debug safety */
270 kmem_cache_free(rmap_item_cache, rmap_item);
273 static inline struct stable_node *alloc_stable_node(void)
275 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
278 static inline void free_stable_node(struct stable_node *stable_node)
280 kmem_cache_free(stable_node_cache, stable_node);
283 static inline struct mm_slot *alloc_mm_slot(void)
285 if (!mm_slot_cache) /* initialization failed */
287 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
290 static inline void free_mm_slot(struct mm_slot *mm_slot)
292 kmem_cache_free(mm_slot_cache, mm_slot);
295 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
297 struct hlist_node *node;
298 struct mm_slot *slot;
300 hash_for_each_possible(mm_slots_hash, slot, node, link, (unsigned long)mm)
307 static void insert_to_mm_slots_hash(struct mm_struct *mm,
308 struct mm_slot *mm_slot)
311 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
314 static inline int in_stable_tree(struct rmap_item *rmap_item)
316 return rmap_item->address & STABLE_FLAG;
320 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
321 * page tables after it has passed through ksm_exit() - which, if necessary,
322 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
323 * a special flag: they can just back out as soon as mm_users goes to zero.
324 * ksm_test_exit() is used throughout to make this test for exit: in some
325 * places for correctness, in some places just to avoid unnecessary work.
327 static inline bool ksm_test_exit(struct mm_struct *mm)
329 return atomic_read(&mm->mm_users) == 0;
333 * We use break_ksm to break COW on a ksm page: it's a stripped down
335 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
338 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
339 * in case the application has unmapped and remapped mm,addr meanwhile.
340 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
341 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
343 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
350 page = follow_page(vma, addr, FOLL_GET);
351 if (IS_ERR_OR_NULL(page))
354 ret = handle_mm_fault(vma->vm_mm, vma, addr,
357 ret = VM_FAULT_WRITE;
359 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
361 * We must loop because handle_mm_fault() may back out if there's
362 * any difficulty e.g. if pte accessed bit gets updated concurrently.
364 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
365 * COW has been broken, even if the vma does not permit VM_WRITE;
366 * but note that a concurrent fault might break PageKsm for us.
368 * VM_FAULT_SIGBUS could occur if we race with truncation of the
369 * backing file, which also invalidates anonymous pages: that's
370 * okay, that truncation will have unmapped the PageKsm for us.
372 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
373 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
374 * current task has TIF_MEMDIE set, and will be OOM killed on return
375 * to user; and ksmd, having no mm, would never be chosen for that.
377 * But if the mm is in a limited mem_cgroup, then the fault may fail
378 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
379 * even ksmd can fail in this way - though it's usually breaking ksm
380 * just to undo a merge it made a moment before, so unlikely to oom.
382 * That's a pity: we might therefore have more kernel pages allocated
383 * than we're counting as nodes in the stable tree; but ksm_do_scan
384 * will retry to break_cow on each pass, so should recover the page
385 * in due course. The important thing is to not let VM_MERGEABLE
386 * be cleared while any such pages might remain in the area.
388 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
391 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
394 struct vm_area_struct *vma;
395 if (ksm_test_exit(mm))
397 vma = find_vma(mm, addr);
398 if (!vma || vma->vm_start > addr)
400 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
405 static void break_cow(struct rmap_item *rmap_item)
407 struct mm_struct *mm = rmap_item->mm;
408 unsigned long addr = rmap_item->address;
409 struct vm_area_struct *vma;
412 * It is not an accident that whenever we want to break COW
413 * to undo, we also need to drop a reference to the anon_vma.
415 put_anon_vma(rmap_item->anon_vma);
417 down_read(&mm->mmap_sem);
418 vma = find_mergeable_vma(mm, addr);
420 break_ksm(vma, addr);
421 up_read(&mm->mmap_sem);
424 static struct page *page_trans_compound_anon(struct page *page)
426 if (PageTransCompound(page)) {
427 struct page *head = compound_trans_head(page);
429 * head may actually be splitted and freed from under
430 * us but it's ok here.
438 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
440 struct mm_struct *mm = rmap_item->mm;
441 unsigned long addr = rmap_item->address;
442 struct vm_area_struct *vma;
445 down_read(&mm->mmap_sem);
446 vma = find_mergeable_vma(mm, addr);
450 page = follow_page(vma, addr, FOLL_GET);
451 if (IS_ERR_OR_NULL(page))
453 if (PageAnon(page) || page_trans_compound_anon(page)) {
454 flush_anon_page(vma, page, addr);
455 flush_dcache_page(page);
460 up_read(&mm->mmap_sem);
465 * This helper is used for getting right index into array of tree roots.
466 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
467 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
468 * every node has its own stable and unstable tree.
470 static inline int get_kpfn_nid(unsigned long kpfn)
472 return ksm_merge_across_nodes ? 0 : pfn_to_nid(kpfn);
475 static void remove_node_from_stable_tree(struct stable_node *stable_node)
477 struct rmap_item *rmap_item;
478 struct hlist_node *hlist;
481 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
482 if (rmap_item->hlist.next)
486 put_anon_vma(rmap_item->anon_vma);
487 rmap_item->address &= PAGE_MASK;
491 nid = get_kpfn_nid(stable_node->kpfn);
492 rb_erase(&stable_node->node, &root_stable_tree[nid]);
493 free_stable_node(stable_node);
497 * get_ksm_page: checks if the page indicated by the stable node
498 * is still its ksm page, despite having held no reference to it.
499 * In which case we can trust the content of the page, and it
500 * returns the gotten page; but if the page has now been zapped,
501 * remove the stale node from the stable tree and return NULL.
502 * But beware, the stable node's page might be being migrated.
504 * You would expect the stable_node to hold a reference to the ksm page.
505 * But if it increments the page's count, swapping out has to wait for
506 * ksmd to come around again before it can free the page, which may take
507 * seconds or even minutes: much too unresponsive. So instead we use a
508 * "keyhole reference": access to the ksm page from the stable node peeps
509 * out through its keyhole to see if that page still holds the right key,
510 * pointing back to this stable node. This relies on freeing a PageAnon
511 * page to reset its page->mapping to NULL, and relies on no other use of
512 * a page to put something that might look like our key in page->mapping.
513 * is on its way to being freed; but it is an anomaly to bear in mind.
515 static struct page *get_ksm_page(struct stable_node *stable_node, bool locked)
518 void *expected_mapping;
521 expected_mapping = (void *)stable_node +
522 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
524 kpfn = ACCESS_ONCE(stable_node->kpfn);
525 page = pfn_to_page(kpfn);
528 * page is computed from kpfn, so on most architectures reading
529 * page->mapping is naturally ordered after reading node->kpfn,
530 * but on Alpha we need to be more careful.
532 smp_read_barrier_depends();
533 if (ACCESS_ONCE(page->mapping) != expected_mapping)
537 * We cannot do anything with the page while its refcount is 0.
538 * Usually 0 means free, or tail of a higher-order page: in which
539 * case this node is no longer referenced, and should be freed;
540 * however, it might mean that the page is under page_freeze_refs().
541 * The __remove_mapping() case is easy, again the node is now stale;
542 * but if page is swapcache in migrate_page_move_mapping(), it might
543 * still be our page, in which case it's essential to keep the node.
545 while (!get_page_unless_zero(page)) {
547 * Another check for page->mapping != expected_mapping would
548 * work here too. We have chosen the !PageSwapCache test to
549 * optimize the common case, when the page is or is about to
550 * be freed: PageSwapCache is cleared (under spin_lock_irq)
551 * in the freeze_refs section of __remove_mapping(); but Anon
552 * page->mapping reset to NULL later, in free_pages_prepare().
554 if (!PageSwapCache(page))
559 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
566 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
576 * We come here from above when page->mapping or !PageSwapCache
577 * suggests that the node is stale; but it might be under migration.
578 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
579 * before checking whether node->kpfn has been changed.
582 if (ACCESS_ONCE(stable_node->kpfn) != kpfn)
584 remove_node_from_stable_tree(stable_node);
589 * Removing rmap_item from stable or unstable tree.
590 * This function will clean the information from the stable/unstable tree.
592 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
594 if (rmap_item->address & STABLE_FLAG) {
595 struct stable_node *stable_node;
598 stable_node = rmap_item->head;
599 page = get_ksm_page(stable_node, true);
603 hlist_del(&rmap_item->hlist);
607 if (stable_node->hlist.first)
612 put_anon_vma(rmap_item->anon_vma);
613 rmap_item->address &= PAGE_MASK;
615 } else if (rmap_item->address & UNSTABLE_FLAG) {
618 * Usually ksmd can and must skip the rb_erase, because
619 * root_unstable_tree was already reset to RB_ROOT.
620 * But be careful when an mm is exiting: do the rb_erase
621 * if this rmap_item was inserted by this scan, rather
622 * than left over from before.
624 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
627 rb_erase(&rmap_item->node,
628 &root_unstable_tree[NUMA(rmap_item->nid)]);
629 ksm_pages_unshared--;
630 rmap_item->address &= PAGE_MASK;
633 cond_resched(); /* we're called from many long loops */
636 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
637 struct rmap_item **rmap_list)
640 struct rmap_item *rmap_item = *rmap_list;
641 *rmap_list = rmap_item->rmap_list;
642 remove_rmap_item_from_tree(rmap_item);
643 free_rmap_item(rmap_item);
648 * Though it's very tempting to unmerge rmap_items from stable tree rather
649 * than check every pte of a given vma, the locking doesn't quite work for
650 * that - an rmap_item is assigned to the stable tree after inserting ksm
651 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
652 * rmap_items from parent to child at fork time (so as not to waste time
653 * if exit comes before the next scan reaches it).
655 * Similarly, although we'd like to remove rmap_items (so updating counts
656 * and freeing memory) when unmerging an area, it's easier to leave that
657 * to the next pass of ksmd - consider, for example, how ksmd might be
658 * in cmp_and_merge_page on one of the rmap_items we would be removing.
660 static int unmerge_ksm_pages(struct vm_area_struct *vma,
661 unsigned long start, unsigned long end)
666 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
667 if (ksm_test_exit(vma->vm_mm))
669 if (signal_pending(current))
672 err = break_ksm(vma, addr);
679 * Only called through the sysfs control interface:
681 static int remove_stable_node(struct stable_node *stable_node)
686 page = get_ksm_page(stable_node, true);
689 * get_ksm_page did remove_node_from_stable_tree itself.
694 if (WARN_ON_ONCE(page_mapped(page)))
698 * This page might be in a pagevec waiting to be freed,
699 * or it might be PageSwapCache (perhaps under writeback),
700 * or it might have been removed from swapcache a moment ago.
702 set_page_stable_node(page, NULL);
703 remove_node_from_stable_tree(stable_node);
712 static int remove_all_stable_nodes(void)
714 struct stable_node *stable_node;
718 for (nid = 0; nid < nr_node_ids; nid++) {
719 while (root_stable_tree[nid].rb_node) {
720 stable_node = rb_entry(root_stable_tree[nid].rb_node,
721 struct stable_node, node);
722 if (remove_stable_node(stable_node)) {
724 break; /* proceed to next nid */
732 static int unmerge_and_remove_all_rmap_items(void)
734 struct mm_slot *mm_slot;
735 struct mm_struct *mm;
736 struct vm_area_struct *vma;
739 spin_lock(&ksm_mmlist_lock);
740 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
741 struct mm_slot, mm_list);
742 spin_unlock(&ksm_mmlist_lock);
744 for (mm_slot = ksm_scan.mm_slot;
745 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
747 down_read(&mm->mmap_sem);
748 for (vma = mm->mmap; vma; vma = vma->vm_next) {
749 if (ksm_test_exit(mm))
751 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
753 err = unmerge_ksm_pages(vma,
754 vma->vm_start, vma->vm_end);
759 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
761 spin_lock(&ksm_mmlist_lock);
762 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
763 struct mm_slot, mm_list);
764 if (ksm_test_exit(mm)) {
765 hash_del(&mm_slot->link);
766 list_del(&mm_slot->mm_list);
767 spin_unlock(&ksm_mmlist_lock);
769 free_mm_slot(mm_slot);
770 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
771 up_read(&mm->mmap_sem);
774 spin_unlock(&ksm_mmlist_lock);
775 up_read(&mm->mmap_sem);
779 /* Clean up stable nodes, but don't worry if some are still busy */
780 remove_all_stable_nodes();
785 up_read(&mm->mmap_sem);
786 spin_lock(&ksm_mmlist_lock);
787 ksm_scan.mm_slot = &ksm_mm_head;
788 spin_unlock(&ksm_mmlist_lock);
791 #endif /* CONFIG_SYSFS */
793 static u32 calc_checksum(struct page *page)
796 void *addr = kmap_atomic(page);
797 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
802 static int memcmp_pages(struct page *page1, struct page *page2)
807 addr1 = kmap_atomic(page1);
808 addr2 = kmap_atomic(page2);
809 ret = memcmp(addr1, addr2, PAGE_SIZE);
810 kunmap_atomic(addr2);
811 kunmap_atomic(addr1);
815 static inline int pages_identical(struct page *page1, struct page *page2)
817 return !memcmp_pages(page1, page2);
820 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
823 struct mm_struct *mm = vma->vm_mm;
829 unsigned long mmun_start; /* For mmu_notifiers */
830 unsigned long mmun_end; /* For mmu_notifiers */
832 addr = page_address_in_vma(page, vma);
836 BUG_ON(PageTransCompound(page));
839 mmun_end = addr + PAGE_SIZE;
840 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
842 ptep = page_check_address(page, mm, addr, &ptl, 0);
846 if (pte_write(*ptep) || pte_dirty(*ptep)) {
849 swapped = PageSwapCache(page);
850 flush_cache_page(vma, addr, page_to_pfn(page));
852 * Ok this is tricky, when get_user_pages_fast() run it doesn't
853 * take any lock, therefore the check that we are going to make
854 * with the pagecount against the mapcount is racey and
855 * O_DIRECT can happen right after the check.
856 * So we clear the pte and flush the tlb before the check
857 * this assure us that no O_DIRECT can happen after the check
858 * or in the middle of the check.
860 entry = ptep_clear_flush(vma, addr, ptep);
862 * Check that no O_DIRECT or similar I/O is in progress on the
865 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
866 set_pte_at(mm, addr, ptep, entry);
869 if (pte_dirty(entry))
870 set_page_dirty(page);
871 entry = pte_mkclean(pte_wrprotect(entry));
872 set_pte_at_notify(mm, addr, ptep, entry);
878 pte_unmap_unlock(ptep, ptl);
880 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
886 * replace_page - replace page in vma by new ksm page
887 * @vma: vma that holds the pte pointing to page
888 * @page: the page we are replacing by kpage
889 * @kpage: the ksm page we replace page by
890 * @orig_pte: the original value of the pte
892 * Returns 0 on success, -EFAULT on failure.
894 static int replace_page(struct vm_area_struct *vma, struct page *page,
895 struct page *kpage, pte_t orig_pte)
897 struct mm_struct *mm = vma->vm_mm;
903 unsigned long mmun_start; /* For mmu_notifiers */
904 unsigned long mmun_end; /* For mmu_notifiers */
906 addr = page_address_in_vma(page, vma);
910 pmd = mm_find_pmd(mm, addr);
913 BUG_ON(pmd_trans_huge(*pmd));
916 mmun_end = addr + PAGE_SIZE;
917 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
919 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
920 if (!pte_same(*ptep, orig_pte)) {
921 pte_unmap_unlock(ptep, ptl);
926 page_add_anon_rmap(kpage, vma, addr);
928 flush_cache_page(vma, addr, pte_pfn(*ptep));
929 ptep_clear_flush(vma, addr, ptep);
930 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
932 page_remove_rmap(page);
933 if (!page_mapped(page))
934 try_to_free_swap(page);
937 pte_unmap_unlock(ptep, ptl);
940 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
945 static int page_trans_compound_anon_split(struct page *page)
948 struct page *transhuge_head = page_trans_compound_anon(page);
949 if (transhuge_head) {
950 /* Get the reference on the head to split it. */
951 if (get_page_unless_zero(transhuge_head)) {
953 * Recheck we got the reference while the head
954 * was still anonymous.
956 if (PageAnon(transhuge_head))
957 ret = split_huge_page(transhuge_head);
960 * Retry later if split_huge_page run
964 put_page(transhuge_head);
966 /* Retry later if split_huge_page run from under us. */
973 * try_to_merge_one_page - take two pages and merge them into one
974 * @vma: the vma that holds the pte pointing to page
975 * @page: the PageAnon page that we want to replace with kpage
976 * @kpage: the PageKsm page that we want to map instead of page,
977 * or NULL the first time when we want to use page as kpage.
979 * This function returns 0 if the pages were merged, -EFAULT otherwise.
981 static int try_to_merge_one_page(struct vm_area_struct *vma,
982 struct page *page, struct page *kpage)
984 pte_t orig_pte = __pte(0);
987 if (page == kpage) /* ksm page forked */
990 if (!(vma->vm_flags & VM_MERGEABLE))
992 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
994 BUG_ON(PageTransCompound(page));
999 * We need the page lock to read a stable PageSwapCache in
1000 * write_protect_page(). We use trylock_page() instead of
1001 * lock_page() because we don't want to wait here - we
1002 * prefer to continue scanning and merging different pages,
1003 * then come back to this page when it is unlocked.
1005 if (!trylock_page(page))
1008 * If this anonymous page is mapped only here, its pte may need
1009 * to be write-protected. If it's mapped elsewhere, all of its
1010 * ptes are necessarily already write-protected. But in either
1011 * case, we need to lock and check page_count is not raised.
1013 if (write_protect_page(vma, page, &orig_pte) == 0) {
1016 * While we hold page lock, upgrade page from
1017 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1018 * stable_tree_insert() will update stable_node.
1020 set_page_stable_node(page, NULL);
1021 mark_page_accessed(page);
1023 } else if (pages_identical(page, kpage))
1024 err = replace_page(vma, page, kpage, orig_pte);
1027 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1028 munlock_vma_page(page);
1029 if (!PageMlocked(kpage)) {
1032 mlock_vma_page(kpage);
1033 page = kpage; /* for final unlock */
1043 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1044 * but no new kernel page is allocated: kpage must already be a ksm page.
1046 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1048 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1049 struct page *page, struct page *kpage)
1051 struct mm_struct *mm = rmap_item->mm;
1052 struct vm_area_struct *vma;
1055 down_read(&mm->mmap_sem);
1056 if (ksm_test_exit(mm))
1058 vma = find_vma(mm, rmap_item->address);
1059 if (!vma || vma->vm_start > rmap_item->address)
1062 err = try_to_merge_one_page(vma, page, kpage);
1066 /* Must get reference to anon_vma while still holding mmap_sem */
1067 rmap_item->anon_vma = vma->anon_vma;
1068 get_anon_vma(vma->anon_vma);
1070 up_read(&mm->mmap_sem);
1075 * try_to_merge_two_pages - take two identical pages and prepare them
1076 * to be merged into one page.
1078 * This function returns the kpage if we successfully merged two identical
1079 * pages into one ksm page, NULL otherwise.
1081 * Note that this function upgrades page to ksm page: if one of the pages
1082 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1084 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1086 struct rmap_item *tree_rmap_item,
1087 struct page *tree_page)
1091 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1093 err = try_to_merge_with_ksm_page(tree_rmap_item,
1096 * If that fails, we have a ksm page with only one pte
1097 * pointing to it: so break it.
1100 break_cow(rmap_item);
1102 return err ? NULL : page;
1106 * stable_tree_search - search for page inside the stable tree
1108 * This function checks if there is a page inside the stable tree
1109 * with identical content to the page that we are scanning right now.
1111 * This function returns the stable tree node of identical content if found,
1114 static struct page *stable_tree_search(struct page *page)
1116 struct rb_node *node;
1117 struct stable_node *stable_node;
1120 stable_node = page_stable_node(page);
1121 if (stable_node) { /* ksm page forked */
1126 nid = get_kpfn_nid(page_to_pfn(page));
1127 node = root_stable_tree[nid].rb_node;
1130 struct page *tree_page;
1134 stable_node = rb_entry(node, struct stable_node, node);
1135 tree_page = get_ksm_page(stable_node, false);
1139 ret = memcmp_pages(page, tree_page);
1140 put_page(tree_page);
1143 node = node->rb_left;
1145 node = node->rb_right;
1148 * Lock and unlock the stable_node's page (which
1149 * might already have been migrated) so that page
1150 * migration is sure to notice its raised count.
1151 * It would be more elegant to return stable_node
1152 * than kpage, but that involves more changes.
1154 tree_page = get_ksm_page(stable_node, true);
1156 unlock_page(tree_page);
1165 * stable_tree_insert - insert stable tree node pointing to new ksm page
1166 * into the stable tree.
1168 * This function returns the stable tree node just allocated on success,
1171 static struct stable_node *stable_tree_insert(struct page *kpage)
1175 struct rb_node **new;
1176 struct rb_node *parent = NULL;
1177 struct stable_node *stable_node;
1179 kpfn = page_to_pfn(kpage);
1180 nid = get_kpfn_nid(kpfn);
1181 new = &root_stable_tree[nid].rb_node;
1184 struct page *tree_page;
1188 stable_node = rb_entry(*new, struct stable_node, node);
1189 tree_page = get_ksm_page(stable_node, false);
1193 ret = memcmp_pages(kpage, tree_page);
1194 put_page(tree_page);
1198 new = &parent->rb_left;
1200 new = &parent->rb_right;
1203 * It is not a bug that stable_tree_search() didn't
1204 * find this node: because at that time our page was
1205 * not yet write-protected, so may have changed since.
1211 stable_node = alloc_stable_node();
1215 INIT_HLIST_HEAD(&stable_node->hlist);
1216 stable_node->kpfn = kpfn;
1217 set_page_stable_node(kpage, stable_node);
1218 rb_link_node(&stable_node->node, parent, new);
1219 rb_insert_color(&stable_node->node, &root_stable_tree[nid]);
1225 * unstable_tree_search_insert - search for identical page,
1226 * else insert rmap_item into the unstable tree.
1228 * This function searches for a page in the unstable tree identical to the
1229 * page currently being scanned; and if no identical page is found in the
1230 * tree, we insert rmap_item as a new object into the unstable tree.
1232 * This function returns pointer to rmap_item found to be identical
1233 * to the currently scanned page, NULL otherwise.
1235 * This function does both searching and inserting, because they share
1236 * the same walking algorithm in an rbtree.
1239 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1241 struct page **tree_pagep)
1243 struct rb_node **new;
1244 struct rb_root *root;
1245 struct rb_node *parent = NULL;
1248 nid = get_kpfn_nid(page_to_pfn(page));
1249 root = &root_unstable_tree[nid];
1250 new = &root->rb_node;
1253 struct rmap_item *tree_rmap_item;
1254 struct page *tree_page;
1258 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1259 tree_page = get_mergeable_page(tree_rmap_item);
1260 if (IS_ERR_OR_NULL(tree_page))
1264 * Don't substitute a ksm page for a forked page.
1266 if (page == tree_page) {
1267 put_page(tree_page);
1272 * If tree_page has been migrated to another NUMA node, it
1273 * will be flushed out and put into the right unstable tree
1274 * next time: only merge with it if merge_across_nodes.
1276 if (!ksm_merge_across_nodes && page_to_nid(tree_page) != nid) {
1277 put_page(tree_page);
1281 ret = memcmp_pages(page, tree_page);
1285 put_page(tree_page);
1286 new = &parent->rb_left;
1287 } else if (ret > 0) {
1288 put_page(tree_page);
1289 new = &parent->rb_right;
1291 *tree_pagep = tree_page;
1292 return tree_rmap_item;
1296 rmap_item->address |= UNSTABLE_FLAG;
1297 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1298 DO_NUMA(rmap_item->nid = nid);
1299 rb_link_node(&rmap_item->node, parent, new);
1300 rb_insert_color(&rmap_item->node, root);
1302 ksm_pages_unshared++;
1307 * stable_tree_append - add another rmap_item to the linked list of
1308 * rmap_items hanging off a given node of the stable tree, all sharing
1309 * the same ksm page.
1311 static void stable_tree_append(struct rmap_item *rmap_item,
1312 struct stable_node *stable_node)
1315 * Usually rmap_item->nid is already set correctly,
1316 * but it may be wrong after switching merge_across_nodes.
1318 DO_NUMA(rmap_item->nid = get_kpfn_nid(stable_node->kpfn));
1319 rmap_item->head = stable_node;
1320 rmap_item->address |= STABLE_FLAG;
1321 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1323 if (rmap_item->hlist.next)
1324 ksm_pages_sharing++;
1330 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1331 * if not, compare checksum to previous and if it's the same, see if page can
1332 * be inserted into the unstable tree, or merged with a page already there and
1333 * both transferred to the stable tree.
1335 * @page: the page that we are searching identical page to.
1336 * @rmap_item: the reverse mapping into the virtual address of this page
1338 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1340 struct rmap_item *tree_rmap_item;
1341 struct page *tree_page = NULL;
1342 struct stable_node *stable_node;
1344 unsigned int checksum;
1347 remove_rmap_item_from_tree(rmap_item);
1349 /* We first start with searching the page inside the stable tree */
1350 kpage = stable_tree_search(page);
1352 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1355 * The page was successfully merged:
1356 * add its rmap_item to the stable tree.
1359 stable_tree_append(rmap_item, page_stable_node(kpage));
1367 * If the hash value of the page has changed from the last time
1368 * we calculated it, this page is changing frequently: therefore we
1369 * don't want to insert it in the unstable tree, and we don't want
1370 * to waste our time searching for something identical to it there.
1372 checksum = calc_checksum(page);
1373 if (rmap_item->oldchecksum != checksum) {
1374 rmap_item->oldchecksum = checksum;
1379 unstable_tree_search_insert(rmap_item, page, &tree_page);
1380 if (tree_rmap_item) {
1381 kpage = try_to_merge_two_pages(rmap_item, page,
1382 tree_rmap_item, tree_page);
1383 put_page(tree_page);
1385 * As soon as we merge this page, we want to remove the
1386 * rmap_item of the page we have merged with from the unstable
1387 * tree, and insert it instead as new node in the stable tree.
1390 remove_rmap_item_from_tree(tree_rmap_item);
1393 stable_node = stable_tree_insert(kpage);
1395 stable_tree_append(tree_rmap_item, stable_node);
1396 stable_tree_append(rmap_item, stable_node);
1401 * If we fail to insert the page into the stable tree,
1402 * we will have 2 virtual addresses that are pointing
1403 * to a ksm page left outside the stable tree,
1404 * in which case we need to break_cow on both.
1407 break_cow(tree_rmap_item);
1408 break_cow(rmap_item);
1414 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1415 struct rmap_item **rmap_list,
1418 struct rmap_item *rmap_item;
1420 while (*rmap_list) {
1421 rmap_item = *rmap_list;
1422 if ((rmap_item->address & PAGE_MASK) == addr)
1424 if (rmap_item->address > addr)
1426 *rmap_list = rmap_item->rmap_list;
1427 remove_rmap_item_from_tree(rmap_item);
1428 free_rmap_item(rmap_item);
1431 rmap_item = alloc_rmap_item();
1433 /* It has already been zeroed */
1434 rmap_item->mm = mm_slot->mm;
1435 rmap_item->address = addr;
1436 rmap_item->rmap_list = *rmap_list;
1437 *rmap_list = rmap_item;
1442 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1444 struct mm_struct *mm;
1445 struct mm_slot *slot;
1446 struct vm_area_struct *vma;
1447 struct rmap_item *rmap_item;
1450 if (list_empty(&ksm_mm_head.mm_list))
1453 slot = ksm_scan.mm_slot;
1454 if (slot == &ksm_mm_head) {
1456 * A number of pages can hang around indefinitely on per-cpu
1457 * pagevecs, raised page count preventing write_protect_page
1458 * from merging them. Though it doesn't really matter much,
1459 * it is puzzling to see some stuck in pages_volatile until
1460 * other activity jostles them out, and they also prevented
1461 * LTP's KSM test from succeeding deterministically; so drain
1462 * them here (here rather than on entry to ksm_do_scan(),
1463 * so we don't IPI too often when pages_to_scan is set low).
1465 lru_add_drain_all();
1467 for (nid = 0; nid < nr_node_ids; nid++)
1468 root_unstable_tree[nid] = RB_ROOT;
1470 spin_lock(&ksm_mmlist_lock);
1471 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1472 ksm_scan.mm_slot = slot;
1473 spin_unlock(&ksm_mmlist_lock);
1475 * Although we tested list_empty() above, a racing __ksm_exit
1476 * of the last mm on the list may have removed it since then.
1478 if (slot == &ksm_mm_head)
1481 ksm_scan.address = 0;
1482 ksm_scan.rmap_list = &slot->rmap_list;
1486 down_read(&mm->mmap_sem);
1487 if (ksm_test_exit(mm))
1490 vma = find_vma(mm, ksm_scan.address);
1492 for (; vma; vma = vma->vm_next) {
1493 if (!(vma->vm_flags & VM_MERGEABLE))
1495 if (ksm_scan.address < vma->vm_start)
1496 ksm_scan.address = vma->vm_start;
1498 ksm_scan.address = vma->vm_end;
1500 while (ksm_scan.address < vma->vm_end) {
1501 if (ksm_test_exit(mm))
1503 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1504 if (IS_ERR_OR_NULL(*page)) {
1505 ksm_scan.address += PAGE_SIZE;
1509 if (PageAnon(*page) ||
1510 page_trans_compound_anon(*page)) {
1511 flush_anon_page(vma, *page, ksm_scan.address);
1512 flush_dcache_page(*page);
1513 rmap_item = get_next_rmap_item(slot,
1514 ksm_scan.rmap_list, ksm_scan.address);
1516 ksm_scan.rmap_list =
1517 &rmap_item->rmap_list;
1518 ksm_scan.address += PAGE_SIZE;
1521 up_read(&mm->mmap_sem);
1525 ksm_scan.address += PAGE_SIZE;
1530 if (ksm_test_exit(mm)) {
1531 ksm_scan.address = 0;
1532 ksm_scan.rmap_list = &slot->rmap_list;
1535 * Nuke all the rmap_items that are above this current rmap:
1536 * because there were no VM_MERGEABLE vmas with such addresses.
1538 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1540 spin_lock(&ksm_mmlist_lock);
1541 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1542 struct mm_slot, mm_list);
1543 if (ksm_scan.address == 0) {
1545 * We've completed a full scan of all vmas, holding mmap_sem
1546 * throughout, and found no VM_MERGEABLE: so do the same as
1547 * __ksm_exit does to remove this mm from all our lists now.
1548 * This applies either when cleaning up after __ksm_exit
1549 * (but beware: we can reach here even before __ksm_exit),
1550 * or when all VM_MERGEABLE areas have been unmapped (and
1551 * mmap_sem then protects against race with MADV_MERGEABLE).
1553 hash_del(&slot->link);
1554 list_del(&slot->mm_list);
1555 spin_unlock(&ksm_mmlist_lock);
1558 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1559 up_read(&mm->mmap_sem);
1562 spin_unlock(&ksm_mmlist_lock);
1563 up_read(&mm->mmap_sem);
1566 /* Repeat until we've completed scanning the whole list */
1567 slot = ksm_scan.mm_slot;
1568 if (slot != &ksm_mm_head)
1576 * ksm_do_scan - the ksm scanner main worker function.
1577 * @scan_npages - number of pages we want to scan before we return.
1579 static void ksm_do_scan(unsigned int scan_npages)
1581 struct rmap_item *rmap_item;
1582 struct page *uninitialized_var(page);
1584 while (scan_npages-- && likely(!freezing(current))) {
1586 rmap_item = scan_get_next_rmap_item(&page);
1589 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1590 cmp_and_merge_page(page, rmap_item);
1595 static int ksmd_should_run(void)
1597 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1600 static int ksm_scan_thread(void *nothing)
1603 set_user_nice(current, 5);
1605 while (!kthread_should_stop()) {
1606 mutex_lock(&ksm_thread_mutex);
1607 if (ksmd_should_run())
1608 ksm_do_scan(ksm_thread_pages_to_scan);
1609 mutex_unlock(&ksm_thread_mutex);
1613 if (ksmd_should_run()) {
1614 schedule_timeout_interruptible(
1615 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1617 wait_event_freezable(ksm_thread_wait,
1618 ksmd_should_run() || kthread_should_stop());
1624 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1625 unsigned long end, int advice, unsigned long *vm_flags)
1627 struct mm_struct *mm = vma->vm_mm;
1631 case MADV_MERGEABLE:
1633 * Be somewhat over-protective for now!
1635 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1636 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1637 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1638 return 0; /* just ignore the advice */
1641 if (*vm_flags & VM_SAO)
1645 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1646 err = __ksm_enter(mm);
1651 *vm_flags |= VM_MERGEABLE;
1654 case MADV_UNMERGEABLE:
1655 if (!(*vm_flags & VM_MERGEABLE))
1656 return 0; /* just ignore the advice */
1658 if (vma->anon_vma) {
1659 err = unmerge_ksm_pages(vma, start, end);
1664 *vm_flags &= ~VM_MERGEABLE;
1671 int __ksm_enter(struct mm_struct *mm)
1673 struct mm_slot *mm_slot;
1676 mm_slot = alloc_mm_slot();
1680 /* Check ksm_run too? Would need tighter locking */
1681 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1683 spin_lock(&ksm_mmlist_lock);
1684 insert_to_mm_slots_hash(mm, mm_slot);
1686 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1687 * insert just behind the scanning cursor, to let the area settle
1688 * down a little; when fork is followed by immediate exec, we don't
1689 * want ksmd to waste time setting up and tearing down an rmap_list.
1691 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1692 * scanning cursor, otherwise KSM pages in newly forked mms will be
1693 * missed: then we might as well insert at the end of the list.
1695 if (ksm_run & KSM_RUN_UNMERGE)
1696 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1698 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1699 spin_unlock(&ksm_mmlist_lock);
1701 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1702 atomic_inc(&mm->mm_count);
1705 wake_up_interruptible(&ksm_thread_wait);
1710 void __ksm_exit(struct mm_struct *mm)
1712 struct mm_slot *mm_slot;
1713 int easy_to_free = 0;
1716 * This process is exiting: if it's straightforward (as is the
1717 * case when ksmd was never running), free mm_slot immediately.
1718 * But if it's at the cursor or has rmap_items linked to it, use
1719 * mmap_sem to synchronize with any break_cows before pagetables
1720 * are freed, and leave the mm_slot on the list for ksmd to free.
1721 * Beware: ksm may already have noticed it exiting and freed the slot.
1724 spin_lock(&ksm_mmlist_lock);
1725 mm_slot = get_mm_slot(mm);
1726 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1727 if (!mm_slot->rmap_list) {
1728 hash_del(&mm_slot->link);
1729 list_del(&mm_slot->mm_list);
1732 list_move(&mm_slot->mm_list,
1733 &ksm_scan.mm_slot->mm_list);
1736 spin_unlock(&ksm_mmlist_lock);
1739 free_mm_slot(mm_slot);
1740 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1742 } else if (mm_slot) {
1743 down_write(&mm->mmap_sem);
1744 up_write(&mm->mmap_sem);
1748 struct page *ksm_might_need_to_copy(struct page *page,
1749 struct vm_area_struct *vma, unsigned long address)
1751 struct anon_vma *anon_vma = page_anon_vma(page);
1752 struct page *new_page;
1754 if (PageKsm(page)) {
1755 if (page_stable_node(page) &&
1756 !(ksm_run & KSM_RUN_UNMERGE))
1757 return page; /* no need to copy it */
1758 } else if (!anon_vma) {
1759 return page; /* no need to copy it */
1760 } else if (anon_vma->root == vma->anon_vma->root &&
1761 page->index == linear_page_index(vma, address)) {
1762 return page; /* still no need to copy it */
1764 if (!PageUptodate(page))
1765 return page; /* let do_swap_page report the error */
1767 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1769 copy_user_highpage(new_page, page, address, vma);
1771 SetPageDirty(new_page);
1772 __SetPageUptodate(new_page);
1773 __set_page_locked(new_page);
1779 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1780 unsigned long *vm_flags)
1782 struct stable_node *stable_node;
1783 struct rmap_item *rmap_item;
1784 struct hlist_node *hlist;
1785 unsigned int mapcount = page_mapcount(page);
1787 int search_new_forks = 0;
1789 VM_BUG_ON(!PageKsm(page));
1790 VM_BUG_ON(!PageLocked(page));
1792 stable_node = page_stable_node(page);
1796 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1797 struct anon_vma *anon_vma = rmap_item->anon_vma;
1798 struct anon_vma_chain *vmac;
1799 struct vm_area_struct *vma;
1801 anon_vma_lock_read(anon_vma);
1802 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1805 if (rmap_item->address < vma->vm_start ||
1806 rmap_item->address >= vma->vm_end)
1809 * Initially we examine only the vma which covers this
1810 * rmap_item; but later, if there is still work to do,
1811 * we examine covering vmas in other mms: in case they
1812 * were forked from the original since ksmd passed.
1814 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1817 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1820 referenced += page_referenced_one(page, vma,
1821 rmap_item->address, &mapcount, vm_flags);
1822 if (!search_new_forks || !mapcount)
1825 anon_vma_unlock_read(anon_vma);
1829 if (!search_new_forks++)
1835 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1837 struct stable_node *stable_node;
1838 struct hlist_node *hlist;
1839 struct rmap_item *rmap_item;
1840 int ret = SWAP_AGAIN;
1841 int search_new_forks = 0;
1843 VM_BUG_ON(!PageKsm(page));
1844 VM_BUG_ON(!PageLocked(page));
1846 stable_node = page_stable_node(page);
1850 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1851 struct anon_vma *anon_vma = rmap_item->anon_vma;
1852 struct anon_vma_chain *vmac;
1853 struct vm_area_struct *vma;
1855 anon_vma_lock_read(anon_vma);
1856 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1859 if (rmap_item->address < vma->vm_start ||
1860 rmap_item->address >= vma->vm_end)
1863 * Initially we examine only the vma which covers this
1864 * rmap_item; but later, if there is still work to do,
1865 * we examine covering vmas in other mms: in case they
1866 * were forked from the original since ksmd passed.
1868 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1871 ret = try_to_unmap_one(page, vma,
1872 rmap_item->address, flags);
1873 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1874 anon_vma_unlock_read(anon_vma);
1878 anon_vma_unlock_read(anon_vma);
1880 if (!search_new_forks++)
1886 #ifdef CONFIG_MIGRATION
1887 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1888 struct vm_area_struct *, unsigned long, void *), void *arg)
1890 struct stable_node *stable_node;
1891 struct hlist_node *hlist;
1892 struct rmap_item *rmap_item;
1893 int ret = SWAP_AGAIN;
1894 int search_new_forks = 0;
1896 VM_BUG_ON(!PageKsm(page));
1897 VM_BUG_ON(!PageLocked(page));
1899 stable_node = page_stable_node(page);
1903 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1904 struct anon_vma *anon_vma = rmap_item->anon_vma;
1905 struct anon_vma_chain *vmac;
1906 struct vm_area_struct *vma;
1908 anon_vma_lock_read(anon_vma);
1909 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1912 if (rmap_item->address < vma->vm_start ||
1913 rmap_item->address >= vma->vm_end)
1916 * Initially we examine only the vma which covers this
1917 * rmap_item; but later, if there is still work to do,
1918 * we examine covering vmas in other mms: in case they
1919 * were forked from the original since ksmd passed.
1921 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1924 ret = rmap_one(page, vma, rmap_item->address, arg);
1925 if (ret != SWAP_AGAIN) {
1926 anon_vma_unlock_read(anon_vma);
1930 anon_vma_unlock_read(anon_vma);
1932 if (!search_new_forks++)
1938 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1940 struct stable_node *stable_node;
1942 VM_BUG_ON(!PageLocked(oldpage));
1943 VM_BUG_ON(!PageLocked(newpage));
1944 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1946 stable_node = page_stable_node(newpage);
1948 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1949 stable_node->kpfn = page_to_pfn(newpage);
1951 * newpage->mapping was set in advance; now we need smp_wmb()
1952 * to make sure that the new stable_node->kpfn is visible
1953 * to get_ksm_page() before it can see that oldpage->mapping
1954 * has gone stale (or that PageSwapCache has been cleared).
1957 set_page_stable_node(oldpage, NULL);
1960 #endif /* CONFIG_MIGRATION */
1962 #ifdef CONFIG_MEMORY_HOTREMOVE
1963 static void ksm_check_stable_tree(unsigned long start_pfn,
1964 unsigned long end_pfn)
1966 struct stable_node *stable_node;
1967 struct rb_node *node;
1970 for (nid = 0; nid < nr_node_ids; nid++) {
1971 node = rb_first(&root_stable_tree[nid]);
1973 stable_node = rb_entry(node, struct stable_node, node);
1974 if (stable_node->kpfn >= start_pfn &&
1975 stable_node->kpfn < end_pfn) {
1977 * Don't get_ksm_page, page has already gone:
1978 * which is why we keep kpfn instead of page*
1980 remove_node_from_stable_tree(stable_node);
1981 node = rb_first(&root_stable_tree[nid]);
1983 node = rb_next(node);
1989 static int ksm_memory_callback(struct notifier_block *self,
1990 unsigned long action, void *arg)
1992 struct memory_notify *mn = arg;
1995 case MEM_GOING_OFFLINE:
1997 * Keep it very simple for now: just lock out ksmd and
1998 * MADV_UNMERGEABLE while any memory is going offline.
1999 * mutex_lock_nested() is necessary because lockdep was alarmed
2000 * that here we take ksm_thread_mutex inside notifier chain
2001 * mutex, and later take notifier chain mutex inside
2002 * ksm_thread_mutex to unlock it. But that's safe because both
2003 * are inside mem_hotplug_mutex.
2005 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
2010 * Most of the work is done by page migration; but there might
2011 * be a few stable_nodes left over, still pointing to struct
2012 * pages which have been offlined: prune those from the tree,
2013 * otherwise get_ksm_page() might later try to access a
2014 * non-existent struct page.
2016 ksm_check_stable_tree(mn->start_pfn,
2017 mn->start_pfn + mn->nr_pages);
2020 case MEM_CANCEL_OFFLINE:
2021 mutex_unlock(&ksm_thread_mutex);
2026 #endif /* CONFIG_MEMORY_HOTREMOVE */
2030 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2033 #define KSM_ATTR_RO(_name) \
2034 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2035 #define KSM_ATTR(_name) \
2036 static struct kobj_attribute _name##_attr = \
2037 __ATTR(_name, 0644, _name##_show, _name##_store)
2039 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2040 struct kobj_attribute *attr, char *buf)
2042 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2045 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2046 struct kobj_attribute *attr,
2047 const char *buf, size_t count)
2049 unsigned long msecs;
2052 err = strict_strtoul(buf, 10, &msecs);
2053 if (err || msecs > UINT_MAX)
2056 ksm_thread_sleep_millisecs = msecs;
2060 KSM_ATTR(sleep_millisecs);
2062 static ssize_t pages_to_scan_show(struct kobject *kobj,
2063 struct kobj_attribute *attr, char *buf)
2065 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2068 static ssize_t pages_to_scan_store(struct kobject *kobj,
2069 struct kobj_attribute *attr,
2070 const char *buf, size_t count)
2073 unsigned long nr_pages;
2075 err = strict_strtoul(buf, 10, &nr_pages);
2076 if (err || nr_pages > UINT_MAX)
2079 ksm_thread_pages_to_scan = nr_pages;
2083 KSM_ATTR(pages_to_scan);
2085 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2088 return sprintf(buf, "%u\n", ksm_run);
2091 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2092 const char *buf, size_t count)
2095 unsigned long flags;
2097 err = strict_strtoul(buf, 10, &flags);
2098 if (err || flags > UINT_MAX)
2100 if (flags > KSM_RUN_UNMERGE)
2104 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2105 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2106 * breaking COW to free the pages_shared (but leaves mm_slots
2107 * on the list for when ksmd may be set running again).
2110 mutex_lock(&ksm_thread_mutex);
2111 if (ksm_run != flags) {
2113 if (flags & KSM_RUN_UNMERGE) {
2114 set_current_oom_origin();
2115 err = unmerge_and_remove_all_rmap_items();
2116 clear_current_oom_origin();
2118 ksm_run = KSM_RUN_STOP;
2123 mutex_unlock(&ksm_thread_mutex);
2125 if (flags & KSM_RUN_MERGE)
2126 wake_up_interruptible(&ksm_thread_wait);
2133 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2134 struct kobj_attribute *attr, char *buf)
2136 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2139 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2140 struct kobj_attribute *attr,
2141 const char *buf, size_t count)
2146 err = kstrtoul(buf, 10, &knob);
2152 mutex_lock(&ksm_thread_mutex);
2153 if (ksm_merge_across_nodes != knob) {
2154 if (ksm_pages_shared || remove_all_stable_nodes())
2157 ksm_merge_across_nodes = knob;
2159 mutex_unlock(&ksm_thread_mutex);
2161 return err ? err : count;
2163 KSM_ATTR(merge_across_nodes);
2166 static ssize_t pages_shared_show(struct kobject *kobj,
2167 struct kobj_attribute *attr, char *buf)
2169 return sprintf(buf, "%lu\n", ksm_pages_shared);
2171 KSM_ATTR_RO(pages_shared);
2173 static ssize_t pages_sharing_show(struct kobject *kobj,
2174 struct kobj_attribute *attr, char *buf)
2176 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2178 KSM_ATTR_RO(pages_sharing);
2180 static ssize_t pages_unshared_show(struct kobject *kobj,
2181 struct kobj_attribute *attr, char *buf)
2183 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2185 KSM_ATTR_RO(pages_unshared);
2187 static ssize_t pages_volatile_show(struct kobject *kobj,
2188 struct kobj_attribute *attr, char *buf)
2190 long ksm_pages_volatile;
2192 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2193 - ksm_pages_sharing - ksm_pages_unshared;
2195 * It was not worth any locking to calculate that statistic,
2196 * but it might therefore sometimes be negative: conceal that.
2198 if (ksm_pages_volatile < 0)
2199 ksm_pages_volatile = 0;
2200 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2202 KSM_ATTR_RO(pages_volatile);
2204 static ssize_t full_scans_show(struct kobject *kobj,
2205 struct kobj_attribute *attr, char *buf)
2207 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2209 KSM_ATTR_RO(full_scans);
2211 static struct attribute *ksm_attrs[] = {
2212 &sleep_millisecs_attr.attr,
2213 &pages_to_scan_attr.attr,
2215 &pages_shared_attr.attr,
2216 &pages_sharing_attr.attr,
2217 &pages_unshared_attr.attr,
2218 &pages_volatile_attr.attr,
2219 &full_scans_attr.attr,
2221 &merge_across_nodes_attr.attr,
2226 static struct attribute_group ksm_attr_group = {
2230 #endif /* CONFIG_SYSFS */
2232 static int __init ksm_init(void)
2234 struct task_struct *ksm_thread;
2238 err = ksm_slab_init();
2242 for (nid = 0; nid < nr_node_ids; nid++)
2243 root_stable_tree[nid] = RB_ROOT;
2245 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2246 if (IS_ERR(ksm_thread)) {
2247 printk(KERN_ERR "ksm: creating kthread failed\n");
2248 err = PTR_ERR(ksm_thread);
2253 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2255 printk(KERN_ERR "ksm: register sysfs failed\n");
2256 kthread_stop(ksm_thread);
2260 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2262 #endif /* CONFIG_SYSFS */
2264 #ifdef CONFIG_MEMORY_HOTREMOVE
2266 * Choose a high priority since the callback takes ksm_thread_mutex:
2267 * later callbacks could only be taking locks which nest within that.
2269 hotplug_memory_notifier(ksm_memory_callback, 100);
2278 module_init(ksm_init)