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 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
126 * @list: linked into migrate_nodes, pending placement in the proper node tree
127 * @hlist: hlist head of rmap_items using this ksm page
128 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
129 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
133 struct rb_node node; /* when node of stable tree */
134 struct { /* when listed for migration */
135 struct list_head *head;
136 struct list_head list;
139 struct hlist_head hlist;
147 * struct rmap_item - reverse mapping item for virtual addresses
148 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
149 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
150 * @mm: the memory structure this rmap_item is pointing into
151 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
152 * @oldchecksum: previous checksum of the page at that virtual address
153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
154 * @node: rb node of this rmap_item in the unstable tree
155 * @head: pointer to stable_node heading this list in the stable tree
156 * @hlist: link into hlist of rmap_items hanging off that stable_node
159 struct rmap_item *rmap_list;
160 struct anon_vma *anon_vma; /* when stable */
161 struct mm_struct *mm;
162 unsigned long address; /* + low bits used for flags below */
163 unsigned int oldchecksum; /* when unstable */
168 struct rb_node node; /* when node of unstable tree */
169 struct { /* when listed from stable tree */
170 struct stable_node *head;
171 struct hlist_node hlist;
176 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
177 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
178 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
180 /* The stable and unstable tree heads */
181 static struct rb_root root_unstable_tree[MAX_NUMNODES];
182 static struct rb_root root_stable_tree[MAX_NUMNODES];
184 /* Recently migrated nodes of stable tree, pending proper placement */
185 static LIST_HEAD(migrate_nodes);
187 #define MM_SLOTS_HASH_BITS 10
188 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
190 static struct mm_slot ksm_mm_head = {
191 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
193 static struct ksm_scan ksm_scan = {
194 .mm_slot = &ksm_mm_head,
197 static struct kmem_cache *rmap_item_cache;
198 static struct kmem_cache *stable_node_cache;
199 static struct kmem_cache *mm_slot_cache;
201 /* The number of nodes in the stable tree */
202 static unsigned long ksm_pages_shared;
204 /* The number of page slots additionally sharing those nodes */
205 static unsigned long ksm_pages_sharing;
207 /* The number of nodes in the unstable tree */
208 static unsigned long ksm_pages_unshared;
210 /* The number of rmap_items in use: to calculate pages_volatile */
211 static unsigned long ksm_rmap_items;
213 /* Number of pages ksmd should scan in one batch */
214 static unsigned int ksm_thread_pages_to_scan = 100;
216 /* Milliseconds ksmd should sleep between batches */
217 static unsigned int ksm_thread_sleep_millisecs = 20;
220 /* Zeroed when merging across nodes is not allowed */
221 static unsigned int ksm_merge_across_nodes = 1;
223 #define ksm_merge_across_nodes 1U
226 #define KSM_RUN_STOP 0
227 #define KSM_RUN_MERGE 1
228 #define KSM_RUN_UNMERGE 2
229 static unsigned int ksm_run = KSM_RUN_STOP;
231 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
232 static DEFINE_MUTEX(ksm_thread_mutex);
233 static DEFINE_SPINLOCK(ksm_mmlist_lock);
235 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
236 sizeof(struct __struct), __alignof__(struct __struct),\
239 static int __init ksm_slab_init(void)
241 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
242 if (!rmap_item_cache)
245 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
246 if (!stable_node_cache)
249 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
256 kmem_cache_destroy(stable_node_cache);
258 kmem_cache_destroy(rmap_item_cache);
263 static void __init ksm_slab_free(void)
265 kmem_cache_destroy(mm_slot_cache);
266 kmem_cache_destroy(stable_node_cache);
267 kmem_cache_destroy(rmap_item_cache);
268 mm_slot_cache = NULL;
271 static inline struct rmap_item *alloc_rmap_item(void)
273 struct rmap_item *rmap_item;
275 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
281 static inline void free_rmap_item(struct rmap_item *rmap_item)
284 rmap_item->mm = NULL; /* debug safety */
285 kmem_cache_free(rmap_item_cache, rmap_item);
288 static inline struct stable_node *alloc_stable_node(void)
290 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
293 static inline void free_stable_node(struct stable_node *stable_node)
295 kmem_cache_free(stable_node_cache, stable_node);
298 static inline struct mm_slot *alloc_mm_slot(void)
300 if (!mm_slot_cache) /* initialization failed */
302 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
305 static inline void free_mm_slot(struct mm_slot *mm_slot)
307 kmem_cache_free(mm_slot_cache, mm_slot);
310 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
312 struct hlist_node *node;
313 struct mm_slot *slot;
315 hash_for_each_possible(mm_slots_hash, slot, node, link, (unsigned long)mm)
322 static void insert_to_mm_slots_hash(struct mm_struct *mm,
323 struct mm_slot *mm_slot)
326 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
330 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
331 * page tables after it has passed through ksm_exit() - which, if necessary,
332 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
333 * a special flag: they can just back out as soon as mm_users goes to zero.
334 * ksm_test_exit() is used throughout to make this test for exit: in some
335 * places for correctness, in some places just to avoid unnecessary work.
337 static inline bool ksm_test_exit(struct mm_struct *mm)
339 return atomic_read(&mm->mm_users) == 0;
343 * We use break_ksm to break COW on a ksm page: it's a stripped down
345 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
348 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
349 * in case the application has unmapped and remapped mm,addr meanwhile.
350 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
351 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
353 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
360 page = follow_page(vma, addr, FOLL_GET);
361 if (IS_ERR_OR_NULL(page))
364 ret = handle_mm_fault(vma->vm_mm, vma, addr,
367 ret = VM_FAULT_WRITE;
369 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
371 * We must loop because handle_mm_fault() may back out if there's
372 * any difficulty e.g. if pte accessed bit gets updated concurrently.
374 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
375 * COW has been broken, even if the vma does not permit VM_WRITE;
376 * but note that a concurrent fault might break PageKsm for us.
378 * VM_FAULT_SIGBUS could occur if we race with truncation of the
379 * backing file, which also invalidates anonymous pages: that's
380 * okay, that truncation will have unmapped the PageKsm for us.
382 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
383 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
384 * current task has TIF_MEMDIE set, and will be OOM killed on return
385 * to user; and ksmd, having no mm, would never be chosen for that.
387 * But if the mm is in a limited mem_cgroup, then the fault may fail
388 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
389 * even ksmd can fail in this way - though it's usually breaking ksm
390 * just to undo a merge it made a moment before, so unlikely to oom.
392 * That's a pity: we might therefore have more kernel pages allocated
393 * than we're counting as nodes in the stable tree; but ksm_do_scan
394 * will retry to break_cow on each pass, so should recover the page
395 * in due course. The important thing is to not let VM_MERGEABLE
396 * be cleared while any such pages might remain in the area.
398 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
401 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
404 struct vm_area_struct *vma;
405 if (ksm_test_exit(mm))
407 vma = find_vma(mm, addr);
408 if (!vma || vma->vm_start > addr)
410 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
415 static void break_cow(struct rmap_item *rmap_item)
417 struct mm_struct *mm = rmap_item->mm;
418 unsigned long addr = rmap_item->address;
419 struct vm_area_struct *vma;
422 * It is not an accident that whenever we want to break COW
423 * to undo, we also need to drop a reference to the anon_vma.
425 put_anon_vma(rmap_item->anon_vma);
427 down_read(&mm->mmap_sem);
428 vma = find_mergeable_vma(mm, addr);
430 break_ksm(vma, addr);
431 up_read(&mm->mmap_sem);
434 static struct page *page_trans_compound_anon(struct page *page)
436 if (PageTransCompound(page)) {
437 struct page *head = compound_trans_head(page);
439 * head may actually be splitted and freed from under
440 * us but it's ok here.
448 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
450 struct mm_struct *mm = rmap_item->mm;
451 unsigned long addr = rmap_item->address;
452 struct vm_area_struct *vma;
455 down_read(&mm->mmap_sem);
456 vma = find_mergeable_vma(mm, addr);
460 page = follow_page(vma, addr, FOLL_GET);
461 if (IS_ERR_OR_NULL(page))
463 if (PageAnon(page) || page_trans_compound_anon(page)) {
464 flush_anon_page(vma, page, addr);
465 flush_dcache_page(page);
470 up_read(&mm->mmap_sem);
475 * This helper is used for getting right index into array of tree roots.
476 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
477 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
478 * every node has its own stable and unstable tree.
480 static inline int get_kpfn_nid(unsigned long kpfn)
482 return ksm_merge_across_nodes ? 0 : pfn_to_nid(kpfn);
485 static void remove_node_from_stable_tree(struct stable_node *stable_node)
487 struct rmap_item *rmap_item;
488 struct hlist_node *hlist;
490 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
491 if (rmap_item->hlist.next)
495 put_anon_vma(rmap_item->anon_vma);
496 rmap_item->address &= PAGE_MASK;
500 if (stable_node->head == &migrate_nodes)
501 list_del(&stable_node->list);
503 rb_erase(&stable_node->node,
504 &root_stable_tree[NUMA(stable_node->nid)]);
505 free_stable_node(stable_node);
509 * get_ksm_page: checks if the page indicated by the stable node
510 * is still its ksm page, despite having held no reference to it.
511 * In which case we can trust the content of the page, and it
512 * returns the gotten page; but if the page has now been zapped,
513 * remove the stale node from the stable tree and return NULL.
514 * But beware, the stable node's page might be being migrated.
516 * You would expect the stable_node to hold a reference to the ksm page.
517 * But if it increments the page's count, swapping out has to wait for
518 * ksmd to come around again before it can free the page, which may take
519 * seconds or even minutes: much too unresponsive. So instead we use a
520 * "keyhole reference": access to the ksm page from the stable node peeps
521 * out through its keyhole to see if that page still holds the right key,
522 * pointing back to this stable node. This relies on freeing a PageAnon
523 * page to reset its page->mapping to NULL, and relies on no other use of
524 * a page to put something that might look like our key in page->mapping.
525 * is on its way to being freed; but it is an anomaly to bear in mind.
527 static struct page *get_ksm_page(struct stable_node *stable_node, bool locked)
530 void *expected_mapping;
533 expected_mapping = (void *)stable_node +
534 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
536 kpfn = ACCESS_ONCE(stable_node->kpfn);
537 page = pfn_to_page(kpfn);
540 * page is computed from kpfn, so on most architectures reading
541 * page->mapping is naturally ordered after reading node->kpfn,
542 * but on Alpha we need to be more careful.
544 smp_read_barrier_depends();
545 if (ACCESS_ONCE(page->mapping) != expected_mapping)
549 * We cannot do anything with the page while its refcount is 0.
550 * Usually 0 means free, or tail of a higher-order page: in which
551 * case this node is no longer referenced, and should be freed;
552 * however, it might mean that the page is under page_freeze_refs().
553 * The __remove_mapping() case is easy, again the node is now stale;
554 * but if page is swapcache in migrate_page_move_mapping(), it might
555 * still be our page, in which case it's essential to keep the node.
557 while (!get_page_unless_zero(page)) {
559 * Another check for page->mapping != expected_mapping would
560 * work here too. We have chosen the !PageSwapCache test to
561 * optimize the common case, when the page is or is about to
562 * be freed: PageSwapCache is cleared (under spin_lock_irq)
563 * in the freeze_refs section of __remove_mapping(); but Anon
564 * page->mapping reset to NULL later, in free_pages_prepare().
566 if (!PageSwapCache(page))
571 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
578 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
588 * We come here from above when page->mapping or !PageSwapCache
589 * suggests that the node is stale; but it might be under migration.
590 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
591 * before checking whether node->kpfn has been changed.
594 if (ACCESS_ONCE(stable_node->kpfn) != kpfn)
596 remove_node_from_stable_tree(stable_node);
601 * Removing rmap_item from stable or unstable tree.
602 * This function will clean the information from the stable/unstable tree.
604 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
606 if (rmap_item->address & STABLE_FLAG) {
607 struct stable_node *stable_node;
610 stable_node = rmap_item->head;
611 page = get_ksm_page(stable_node, true);
615 hlist_del(&rmap_item->hlist);
619 if (stable_node->hlist.first)
624 put_anon_vma(rmap_item->anon_vma);
625 rmap_item->address &= PAGE_MASK;
627 } else if (rmap_item->address & UNSTABLE_FLAG) {
630 * Usually ksmd can and must skip the rb_erase, because
631 * root_unstable_tree was already reset to RB_ROOT.
632 * But be careful when an mm is exiting: do the rb_erase
633 * if this rmap_item was inserted by this scan, rather
634 * than left over from before.
636 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
639 rb_erase(&rmap_item->node,
640 &root_unstable_tree[NUMA(rmap_item->nid)]);
641 ksm_pages_unshared--;
642 rmap_item->address &= PAGE_MASK;
645 cond_resched(); /* we're called from many long loops */
648 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
649 struct rmap_item **rmap_list)
652 struct rmap_item *rmap_item = *rmap_list;
653 *rmap_list = rmap_item->rmap_list;
654 remove_rmap_item_from_tree(rmap_item);
655 free_rmap_item(rmap_item);
660 * Though it's very tempting to unmerge rmap_items from stable tree rather
661 * than check every pte of a given vma, the locking doesn't quite work for
662 * that - an rmap_item is assigned to the stable tree after inserting ksm
663 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
664 * rmap_items from parent to child at fork time (so as not to waste time
665 * if exit comes before the next scan reaches it).
667 * Similarly, although we'd like to remove rmap_items (so updating counts
668 * and freeing memory) when unmerging an area, it's easier to leave that
669 * to the next pass of ksmd - consider, for example, how ksmd might be
670 * in cmp_and_merge_page on one of the rmap_items we would be removing.
672 static int unmerge_ksm_pages(struct vm_area_struct *vma,
673 unsigned long start, unsigned long end)
678 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
679 if (ksm_test_exit(vma->vm_mm))
681 if (signal_pending(current))
684 err = break_ksm(vma, addr);
691 * Only called through the sysfs control interface:
693 static int remove_stable_node(struct stable_node *stable_node)
698 page = get_ksm_page(stable_node, true);
701 * get_ksm_page did remove_node_from_stable_tree itself.
706 if (WARN_ON_ONCE(page_mapped(page)))
710 * This page might be in a pagevec waiting to be freed,
711 * or it might be PageSwapCache (perhaps under writeback),
712 * or it might have been removed from swapcache a moment ago.
714 set_page_stable_node(page, NULL);
715 remove_node_from_stable_tree(stable_node);
724 static int remove_all_stable_nodes(void)
726 struct stable_node *stable_node;
727 struct list_head *this, *next;
731 for (nid = 0; nid < nr_node_ids; nid++) {
732 while (root_stable_tree[nid].rb_node) {
733 stable_node = rb_entry(root_stable_tree[nid].rb_node,
734 struct stable_node, node);
735 if (remove_stable_node(stable_node)) {
737 break; /* proceed to next nid */
742 list_for_each_safe(this, next, &migrate_nodes) {
743 stable_node = list_entry(this, struct stable_node, list);
744 if (remove_stable_node(stable_node))
751 static int unmerge_and_remove_all_rmap_items(void)
753 struct mm_slot *mm_slot;
754 struct mm_struct *mm;
755 struct vm_area_struct *vma;
758 spin_lock(&ksm_mmlist_lock);
759 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
760 struct mm_slot, mm_list);
761 spin_unlock(&ksm_mmlist_lock);
763 for (mm_slot = ksm_scan.mm_slot;
764 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
766 down_read(&mm->mmap_sem);
767 for (vma = mm->mmap; vma; vma = vma->vm_next) {
768 if (ksm_test_exit(mm))
770 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
772 err = unmerge_ksm_pages(vma,
773 vma->vm_start, vma->vm_end);
778 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
780 spin_lock(&ksm_mmlist_lock);
781 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
782 struct mm_slot, mm_list);
783 if (ksm_test_exit(mm)) {
784 hash_del(&mm_slot->link);
785 list_del(&mm_slot->mm_list);
786 spin_unlock(&ksm_mmlist_lock);
788 free_mm_slot(mm_slot);
789 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
790 up_read(&mm->mmap_sem);
793 spin_unlock(&ksm_mmlist_lock);
794 up_read(&mm->mmap_sem);
798 /* Clean up stable nodes, but don't worry if some are still busy */
799 remove_all_stable_nodes();
804 up_read(&mm->mmap_sem);
805 spin_lock(&ksm_mmlist_lock);
806 ksm_scan.mm_slot = &ksm_mm_head;
807 spin_unlock(&ksm_mmlist_lock);
810 #endif /* CONFIG_SYSFS */
812 static u32 calc_checksum(struct page *page)
815 void *addr = kmap_atomic(page);
816 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
821 static int memcmp_pages(struct page *page1, struct page *page2)
826 addr1 = kmap_atomic(page1);
827 addr2 = kmap_atomic(page2);
828 ret = memcmp(addr1, addr2, PAGE_SIZE);
829 kunmap_atomic(addr2);
830 kunmap_atomic(addr1);
834 static inline int pages_identical(struct page *page1, struct page *page2)
836 return !memcmp_pages(page1, page2);
839 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
842 struct mm_struct *mm = vma->vm_mm;
848 unsigned long mmun_start; /* For mmu_notifiers */
849 unsigned long mmun_end; /* For mmu_notifiers */
851 addr = page_address_in_vma(page, vma);
855 BUG_ON(PageTransCompound(page));
858 mmun_end = addr + PAGE_SIZE;
859 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
861 ptep = page_check_address(page, mm, addr, &ptl, 0);
865 if (pte_write(*ptep) || pte_dirty(*ptep)) {
868 swapped = PageSwapCache(page);
869 flush_cache_page(vma, addr, page_to_pfn(page));
871 * Ok this is tricky, when get_user_pages_fast() run it doesn't
872 * take any lock, therefore the check that we are going to make
873 * with the pagecount against the mapcount is racey and
874 * O_DIRECT can happen right after the check.
875 * So we clear the pte and flush the tlb before the check
876 * this assure us that no O_DIRECT can happen after the check
877 * or in the middle of the check.
879 entry = ptep_clear_flush(vma, addr, ptep);
881 * Check that no O_DIRECT or similar I/O is in progress on the
884 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
885 set_pte_at(mm, addr, ptep, entry);
888 if (pte_dirty(entry))
889 set_page_dirty(page);
890 entry = pte_mkclean(pte_wrprotect(entry));
891 set_pte_at_notify(mm, addr, ptep, entry);
897 pte_unmap_unlock(ptep, ptl);
899 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
905 * replace_page - replace page in vma by new ksm page
906 * @vma: vma that holds the pte pointing to page
907 * @page: the page we are replacing by kpage
908 * @kpage: the ksm page we replace page by
909 * @orig_pte: the original value of the pte
911 * Returns 0 on success, -EFAULT on failure.
913 static int replace_page(struct vm_area_struct *vma, struct page *page,
914 struct page *kpage, pte_t orig_pte)
916 struct mm_struct *mm = vma->vm_mm;
922 unsigned long mmun_start; /* For mmu_notifiers */
923 unsigned long mmun_end; /* For mmu_notifiers */
925 addr = page_address_in_vma(page, vma);
929 pmd = mm_find_pmd(mm, addr);
932 BUG_ON(pmd_trans_huge(*pmd));
935 mmun_end = addr + PAGE_SIZE;
936 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
938 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
939 if (!pte_same(*ptep, orig_pte)) {
940 pte_unmap_unlock(ptep, ptl);
945 page_add_anon_rmap(kpage, vma, addr);
947 flush_cache_page(vma, addr, pte_pfn(*ptep));
948 ptep_clear_flush(vma, addr, ptep);
949 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
951 page_remove_rmap(page);
952 if (!page_mapped(page))
953 try_to_free_swap(page);
956 pte_unmap_unlock(ptep, ptl);
959 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
964 static int page_trans_compound_anon_split(struct page *page)
967 struct page *transhuge_head = page_trans_compound_anon(page);
968 if (transhuge_head) {
969 /* Get the reference on the head to split it. */
970 if (get_page_unless_zero(transhuge_head)) {
972 * Recheck we got the reference while the head
973 * was still anonymous.
975 if (PageAnon(transhuge_head))
976 ret = split_huge_page(transhuge_head);
979 * Retry later if split_huge_page run
983 put_page(transhuge_head);
985 /* Retry later if split_huge_page run from under us. */
992 * try_to_merge_one_page - take two pages and merge them into one
993 * @vma: the vma that holds the pte pointing to page
994 * @page: the PageAnon page that we want to replace with kpage
995 * @kpage: the PageKsm page that we want to map instead of page,
996 * or NULL the first time when we want to use page as kpage.
998 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1000 static int try_to_merge_one_page(struct vm_area_struct *vma,
1001 struct page *page, struct page *kpage)
1003 pte_t orig_pte = __pte(0);
1006 if (page == kpage) /* ksm page forked */
1009 if (!(vma->vm_flags & VM_MERGEABLE))
1011 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1013 BUG_ON(PageTransCompound(page));
1014 if (!PageAnon(page))
1018 * We need the page lock to read a stable PageSwapCache in
1019 * write_protect_page(). We use trylock_page() instead of
1020 * lock_page() because we don't want to wait here - we
1021 * prefer to continue scanning and merging different pages,
1022 * then come back to this page when it is unlocked.
1024 if (!trylock_page(page))
1027 * If this anonymous page is mapped only here, its pte may need
1028 * to be write-protected. If it's mapped elsewhere, all of its
1029 * ptes are necessarily already write-protected. But in either
1030 * case, we need to lock and check page_count is not raised.
1032 if (write_protect_page(vma, page, &orig_pte) == 0) {
1035 * While we hold page lock, upgrade page from
1036 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1037 * stable_tree_insert() will update stable_node.
1039 set_page_stable_node(page, NULL);
1040 mark_page_accessed(page);
1042 } else if (pages_identical(page, kpage))
1043 err = replace_page(vma, page, kpage, orig_pte);
1046 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1047 munlock_vma_page(page);
1048 if (!PageMlocked(kpage)) {
1051 mlock_vma_page(kpage);
1052 page = kpage; /* for final unlock */
1062 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1063 * but no new kernel page is allocated: kpage must already be a ksm page.
1065 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1067 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1068 struct page *page, struct page *kpage)
1070 struct mm_struct *mm = rmap_item->mm;
1071 struct vm_area_struct *vma;
1074 down_read(&mm->mmap_sem);
1075 if (ksm_test_exit(mm))
1077 vma = find_vma(mm, rmap_item->address);
1078 if (!vma || vma->vm_start > rmap_item->address)
1081 err = try_to_merge_one_page(vma, page, kpage);
1085 /* Must get reference to anon_vma while still holding mmap_sem */
1086 rmap_item->anon_vma = vma->anon_vma;
1087 get_anon_vma(vma->anon_vma);
1089 up_read(&mm->mmap_sem);
1094 * try_to_merge_two_pages - take two identical pages and prepare them
1095 * to be merged into one page.
1097 * This function returns the kpage if we successfully merged two identical
1098 * pages into one ksm page, NULL otherwise.
1100 * Note that this function upgrades page to ksm page: if one of the pages
1101 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1103 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1105 struct rmap_item *tree_rmap_item,
1106 struct page *tree_page)
1110 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1112 err = try_to_merge_with_ksm_page(tree_rmap_item,
1115 * If that fails, we have a ksm page with only one pte
1116 * pointing to it: so break it.
1119 break_cow(rmap_item);
1121 return err ? NULL : page;
1125 * stable_tree_search - search for page inside the stable tree
1127 * This function checks if there is a page inside the stable tree
1128 * with identical content to the page that we are scanning right now.
1130 * This function returns the stable tree node of identical content if found,
1133 static struct page *stable_tree_search(struct page *page)
1136 struct rb_node **new;
1137 struct rb_node *parent;
1138 struct stable_node *stable_node;
1139 struct stable_node *page_node;
1141 page_node = page_stable_node(page);
1142 if (page_node && page_node->head != &migrate_nodes) {
1143 /* ksm page forked */
1148 nid = get_kpfn_nid(page_to_pfn(page));
1150 new = &root_stable_tree[nid].rb_node;
1154 struct page *tree_page;
1158 stable_node = rb_entry(*new, struct stable_node, node);
1159 tree_page = get_ksm_page(stable_node, false);
1163 ret = memcmp_pages(page, tree_page);
1164 put_page(tree_page);
1168 new = &parent->rb_left;
1170 new = &parent->rb_right;
1173 * Lock and unlock the stable_node's page (which
1174 * might already have been migrated) so that page
1175 * migration is sure to notice its raised count.
1176 * It would be more elegant to return stable_node
1177 * than kpage, but that involves more changes.
1179 tree_page = get_ksm_page(stable_node, true);
1181 unlock_page(tree_page);
1182 if (get_kpfn_nid(stable_node->kpfn) !=
1183 NUMA(stable_node->nid)) {
1184 put_page(tree_page);
1190 * There is now a place for page_node, but the tree may
1191 * have been rebalanced, so re-evaluate parent and new.
1202 list_del(&page_node->list);
1203 DO_NUMA(page_node->nid = nid);
1204 rb_link_node(&page_node->node, parent, new);
1205 rb_insert_color(&page_node->node, &root_stable_tree[nid]);
1211 list_del(&page_node->list);
1212 DO_NUMA(page_node->nid = nid);
1213 rb_replace_node(&stable_node->node,
1214 &page_node->node, &root_stable_tree[nid]);
1217 rb_erase(&stable_node->node, &root_stable_tree[nid]);
1220 stable_node->head = &migrate_nodes;
1221 list_add(&stable_node->list, stable_node->head);
1226 * stable_tree_insert - insert stable tree node pointing to new ksm page
1227 * into the stable tree.
1229 * This function returns the stable tree node just allocated on success,
1232 static struct stable_node *stable_tree_insert(struct page *kpage)
1236 struct rb_node **new;
1237 struct rb_node *parent = NULL;
1238 struct stable_node *stable_node;
1240 kpfn = page_to_pfn(kpage);
1241 nid = get_kpfn_nid(kpfn);
1242 new = &root_stable_tree[nid].rb_node;
1245 struct page *tree_page;
1249 stable_node = rb_entry(*new, struct stable_node, node);
1250 tree_page = get_ksm_page(stable_node, false);
1254 ret = memcmp_pages(kpage, tree_page);
1255 put_page(tree_page);
1259 new = &parent->rb_left;
1261 new = &parent->rb_right;
1264 * It is not a bug that stable_tree_search() didn't
1265 * find this node: because at that time our page was
1266 * not yet write-protected, so may have changed since.
1272 stable_node = alloc_stable_node();
1276 INIT_HLIST_HEAD(&stable_node->hlist);
1277 stable_node->kpfn = kpfn;
1278 set_page_stable_node(kpage, stable_node);
1279 DO_NUMA(stable_node->nid = nid);
1280 rb_link_node(&stable_node->node, parent, new);
1281 rb_insert_color(&stable_node->node, &root_stable_tree[nid]);
1287 * unstable_tree_search_insert - search for identical page,
1288 * else insert rmap_item into the unstable tree.
1290 * This function searches for a page in the unstable tree identical to the
1291 * page currently being scanned; and if no identical page is found in the
1292 * tree, we insert rmap_item as a new object into the unstable tree.
1294 * This function returns pointer to rmap_item found to be identical
1295 * to the currently scanned page, NULL otherwise.
1297 * This function does both searching and inserting, because they share
1298 * the same walking algorithm in an rbtree.
1301 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1303 struct page **tree_pagep)
1305 struct rb_node **new;
1306 struct rb_root *root;
1307 struct rb_node *parent = NULL;
1310 nid = get_kpfn_nid(page_to_pfn(page));
1311 root = &root_unstable_tree[nid];
1312 new = &root->rb_node;
1315 struct rmap_item *tree_rmap_item;
1316 struct page *tree_page;
1320 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1321 tree_page = get_mergeable_page(tree_rmap_item);
1322 if (IS_ERR_OR_NULL(tree_page))
1326 * Don't substitute a ksm page for a forked page.
1328 if (page == tree_page) {
1329 put_page(tree_page);
1334 * If tree_page has been migrated to another NUMA node, it
1335 * will be flushed out and put into the right unstable tree
1336 * next time: only merge with it if merge_across_nodes.
1338 if (!ksm_merge_across_nodes && page_to_nid(tree_page) != nid) {
1339 put_page(tree_page);
1343 ret = memcmp_pages(page, tree_page);
1347 put_page(tree_page);
1348 new = &parent->rb_left;
1349 } else if (ret > 0) {
1350 put_page(tree_page);
1351 new = &parent->rb_right;
1353 *tree_pagep = tree_page;
1354 return tree_rmap_item;
1358 rmap_item->address |= UNSTABLE_FLAG;
1359 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1360 DO_NUMA(rmap_item->nid = nid);
1361 rb_link_node(&rmap_item->node, parent, new);
1362 rb_insert_color(&rmap_item->node, root);
1364 ksm_pages_unshared++;
1369 * stable_tree_append - add another rmap_item to the linked list of
1370 * rmap_items hanging off a given node of the stable tree, all sharing
1371 * the same ksm page.
1373 static void stable_tree_append(struct rmap_item *rmap_item,
1374 struct stable_node *stable_node)
1376 rmap_item->head = stable_node;
1377 rmap_item->address |= STABLE_FLAG;
1378 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1380 if (rmap_item->hlist.next)
1381 ksm_pages_sharing++;
1387 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1388 * if not, compare checksum to previous and if it's the same, see if page can
1389 * be inserted into the unstable tree, or merged with a page already there and
1390 * both transferred to the stable tree.
1392 * @page: the page that we are searching identical page to.
1393 * @rmap_item: the reverse mapping into the virtual address of this page
1395 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1397 struct rmap_item *tree_rmap_item;
1398 struct page *tree_page = NULL;
1399 struct stable_node *stable_node;
1401 unsigned int checksum;
1404 stable_node = page_stable_node(page);
1406 if (stable_node->head != &migrate_nodes &&
1407 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1408 rb_erase(&stable_node->node,
1409 &root_stable_tree[NUMA(stable_node->nid)]);
1410 stable_node->head = &migrate_nodes;
1411 list_add(&stable_node->list, stable_node->head);
1413 if (stable_node->head != &migrate_nodes &&
1414 rmap_item->head == stable_node)
1418 /* We first start with searching the page inside the stable tree */
1419 kpage = stable_tree_search(page);
1420 if (kpage == page && rmap_item->head == stable_node) {
1425 remove_rmap_item_from_tree(rmap_item);
1428 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1431 * The page was successfully merged:
1432 * add its rmap_item to the stable tree.
1435 stable_tree_append(rmap_item, page_stable_node(kpage));
1443 * If the hash value of the page has changed from the last time
1444 * we calculated it, this page is changing frequently: therefore we
1445 * don't want to insert it in the unstable tree, and we don't want
1446 * to waste our time searching for something identical to it there.
1448 checksum = calc_checksum(page);
1449 if (rmap_item->oldchecksum != checksum) {
1450 rmap_item->oldchecksum = checksum;
1455 unstable_tree_search_insert(rmap_item, page, &tree_page);
1456 if (tree_rmap_item) {
1457 kpage = try_to_merge_two_pages(rmap_item, page,
1458 tree_rmap_item, tree_page);
1459 put_page(tree_page);
1461 * As soon as we merge this page, we want to remove the
1462 * rmap_item of the page we have merged with from the unstable
1463 * tree, and insert it instead as new node in the stable tree.
1466 remove_rmap_item_from_tree(tree_rmap_item);
1469 stable_node = stable_tree_insert(kpage);
1471 stable_tree_append(tree_rmap_item, stable_node);
1472 stable_tree_append(rmap_item, stable_node);
1477 * If we fail to insert the page into the stable tree,
1478 * we will have 2 virtual addresses that are pointing
1479 * to a ksm page left outside the stable tree,
1480 * in which case we need to break_cow on both.
1483 break_cow(tree_rmap_item);
1484 break_cow(rmap_item);
1490 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1491 struct rmap_item **rmap_list,
1494 struct rmap_item *rmap_item;
1496 while (*rmap_list) {
1497 rmap_item = *rmap_list;
1498 if ((rmap_item->address & PAGE_MASK) == addr)
1500 if (rmap_item->address > addr)
1502 *rmap_list = rmap_item->rmap_list;
1503 remove_rmap_item_from_tree(rmap_item);
1504 free_rmap_item(rmap_item);
1507 rmap_item = alloc_rmap_item();
1509 /* It has already been zeroed */
1510 rmap_item->mm = mm_slot->mm;
1511 rmap_item->address = addr;
1512 rmap_item->rmap_list = *rmap_list;
1513 *rmap_list = rmap_item;
1518 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1520 struct mm_struct *mm;
1521 struct mm_slot *slot;
1522 struct vm_area_struct *vma;
1523 struct rmap_item *rmap_item;
1526 if (list_empty(&ksm_mm_head.mm_list))
1529 slot = ksm_scan.mm_slot;
1530 if (slot == &ksm_mm_head) {
1532 * A number of pages can hang around indefinitely on per-cpu
1533 * pagevecs, raised page count preventing write_protect_page
1534 * from merging them. Though it doesn't really matter much,
1535 * it is puzzling to see some stuck in pages_volatile until
1536 * other activity jostles them out, and they also prevented
1537 * LTP's KSM test from succeeding deterministically; so drain
1538 * them here (here rather than on entry to ksm_do_scan(),
1539 * so we don't IPI too often when pages_to_scan is set low).
1541 lru_add_drain_all();
1544 * Whereas stale stable_nodes on the stable_tree itself
1545 * get pruned in the regular course of stable_tree_search(),
1546 * those moved out to the migrate_nodes list can accumulate:
1547 * so prune them once before each full scan.
1549 if (!ksm_merge_across_nodes) {
1550 struct stable_node *stable_node;
1551 struct list_head *this, *next;
1554 list_for_each_safe(this, next, &migrate_nodes) {
1555 stable_node = list_entry(this,
1556 struct stable_node, list);
1557 page = get_ksm_page(stable_node, false);
1564 for (nid = 0; nid < nr_node_ids; nid++)
1565 root_unstable_tree[nid] = RB_ROOT;
1567 spin_lock(&ksm_mmlist_lock);
1568 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1569 ksm_scan.mm_slot = slot;
1570 spin_unlock(&ksm_mmlist_lock);
1572 * Although we tested list_empty() above, a racing __ksm_exit
1573 * of the last mm on the list may have removed it since then.
1575 if (slot == &ksm_mm_head)
1578 ksm_scan.address = 0;
1579 ksm_scan.rmap_list = &slot->rmap_list;
1583 down_read(&mm->mmap_sem);
1584 if (ksm_test_exit(mm))
1587 vma = find_vma(mm, ksm_scan.address);
1589 for (; vma; vma = vma->vm_next) {
1590 if (!(vma->vm_flags & VM_MERGEABLE))
1592 if (ksm_scan.address < vma->vm_start)
1593 ksm_scan.address = vma->vm_start;
1595 ksm_scan.address = vma->vm_end;
1597 while (ksm_scan.address < vma->vm_end) {
1598 if (ksm_test_exit(mm))
1600 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1601 if (IS_ERR_OR_NULL(*page)) {
1602 ksm_scan.address += PAGE_SIZE;
1606 if (PageAnon(*page) ||
1607 page_trans_compound_anon(*page)) {
1608 flush_anon_page(vma, *page, ksm_scan.address);
1609 flush_dcache_page(*page);
1610 rmap_item = get_next_rmap_item(slot,
1611 ksm_scan.rmap_list, ksm_scan.address);
1613 ksm_scan.rmap_list =
1614 &rmap_item->rmap_list;
1615 ksm_scan.address += PAGE_SIZE;
1618 up_read(&mm->mmap_sem);
1622 ksm_scan.address += PAGE_SIZE;
1627 if (ksm_test_exit(mm)) {
1628 ksm_scan.address = 0;
1629 ksm_scan.rmap_list = &slot->rmap_list;
1632 * Nuke all the rmap_items that are above this current rmap:
1633 * because there were no VM_MERGEABLE vmas with such addresses.
1635 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1637 spin_lock(&ksm_mmlist_lock);
1638 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1639 struct mm_slot, mm_list);
1640 if (ksm_scan.address == 0) {
1642 * We've completed a full scan of all vmas, holding mmap_sem
1643 * throughout, and found no VM_MERGEABLE: so do the same as
1644 * __ksm_exit does to remove this mm from all our lists now.
1645 * This applies either when cleaning up after __ksm_exit
1646 * (but beware: we can reach here even before __ksm_exit),
1647 * or when all VM_MERGEABLE areas have been unmapped (and
1648 * mmap_sem then protects against race with MADV_MERGEABLE).
1650 hash_del(&slot->link);
1651 list_del(&slot->mm_list);
1652 spin_unlock(&ksm_mmlist_lock);
1655 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1656 up_read(&mm->mmap_sem);
1659 spin_unlock(&ksm_mmlist_lock);
1660 up_read(&mm->mmap_sem);
1663 /* Repeat until we've completed scanning the whole list */
1664 slot = ksm_scan.mm_slot;
1665 if (slot != &ksm_mm_head)
1673 * ksm_do_scan - the ksm scanner main worker function.
1674 * @scan_npages - number of pages we want to scan before we return.
1676 static void ksm_do_scan(unsigned int scan_npages)
1678 struct rmap_item *rmap_item;
1679 struct page *uninitialized_var(page);
1681 while (scan_npages-- && likely(!freezing(current))) {
1683 rmap_item = scan_get_next_rmap_item(&page);
1686 cmp_and_merge_page(page, rmap_item);
1691 static int ksmd_should_run(void)
1693 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1696 static int ksm_scan_thread(void *nothing)
1699 set_user_nice(current, 5);
1701 while (!kthread_should_stop()) {
1702 mutex_lock(&ksm_thread_mutex);
1703 if (ksmd_should_run())
1704 ksm_do_scan(ksm_thread_pages_to_scan);
1705 mutex_unlock(&ksm_thread_mutex);
1709 if (ksmd_should_run()) {
1710 schedule_timeout_interruptible(
1711 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1713 wait_event_freezable(ksm_thread_wait,
1714 ksmd_should_run() || kthread_should_stop());
1720 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1721 unsigned long end, int advice, unsigned long *vm_flags)
1723 struct mm_struct *mm = vma->vm_mm;
1727 case MADV_MERGEABLE:
1729 * Be somewhat over-protective for now!
1731 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1732 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1733 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1734 return 0; /* just ignore the advice */
1737 if (*vm_flags & VM_SAO)
1741 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1742 err = __ksm_enter(mm);
1747 *vm_flags |= VM_MERGEABLE;
1750 case MADV_UNMERGEABLE:
1751 if (!(*vm_flags & VM_MERGEABLE))
1752 return 0; /* just ignore the advice */
1754 if (vma->anon_vma) {
1755 err = unmerge_ksm_pages(vma, start, end);
1760 *vm_flags &= ~VM_MERGEABLE;
1767 int __ksm_enter(struct mm_struct *mm)
1769 struct mm_slot *mm_slot;
1772 mm_slot = alloc_mm_slot();
1776 /* Check ksm_run too? Would need tighter locking */
1777 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1779 spin_lock(&ksm_mmlist_lock);
1780 insert_to_mm_slots_hash(mm, mm_slot);
1782 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1783 * insert just behind the scanning cursor, to let the area settle
1784 * down a little; when fork is followed by immediate exec, we don't
1785 * want ksmd to waste time setting up and tearing down an rmap_list.
1787 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1788 * scanning cursor, otherwise KSM pages in newly forked mms will be
1789 * missed: then we might as well insert at the end of the list.
1791 if (ksm_run & KSM_RUN_UNMERGE)
1792 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1794 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1795 spin_unlock(&ksm_mmlist_lock);
1797 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1798 atomic_inc(&mm->mm_count);
1801 wake_up_interruptible(&ksm_thread_wait);
1806 void __ksm_exit(struct mm_struct *mm)
1808 struct mm_slot *mm_slot;
1809 int easy_to_free = 0;
1812 * This process is exiting: if it's straightforward (as is the
1813 * case when ksmd was never running), free mm_slot immediately.
1814 * But if it's at the cursor or has rmap_items linked to it, use
1815 * mmap_sem to synchronize with any break_cows before pagetables
1816 * are freed, and leave the mm_slot on the list for ksmd to free.
1817 * Beware: ksm may already have noticed it exiting and freed the slot.
1820 spin_lock(&ksm_mmlist_lock);
1821 mm_slot = get_mm_slot(mm);
1822 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1823 if (!mm_slot->rmap_list) {
1824 hash_del(&mm_slot->link);
1825 list_del(&mm_slot->mm_list);
1828 list_move(&mm_slot->mm_list,
1829 &ksm_scan.mm_slot->mm_list);
1832 spin_unlock(&ksm_mmlist_lock);
1835 free_mm_slot(mm_slot);
1836 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1838 } else if (mm_slot) {
1839 down_write(&mm->mmap_sem);
1840 up_write(&mm->mmap_sem);
1844 struct page *ksm_might_need_to_copy(struct page *page,
1845 struct vm_area_struct *vma, unsigned long address)
1847 struct anon_vma *anon_vma = page_anon_vma(page);
1848 struct page *new_page;
1850 if (PageKsm(page)) {
1851 if (page_stable_node(page) &&
1852 !(ksm_run & KSM_RUN_UNMERGE))
1853 return page; /* no need to copy it */
1854 } else if (!anon_vma) {
1855 return page; /* no need to copy it */
1856 } else if (anon_vma->root == vma->anon_vma->root &&
1857 page->index == linear_page_index(vma, address)) {
1858 return page; /* still no need to copy it */
1860 if (!PageUptodate(page))
1861 return page; /* let do_swap_page report the error */
1863 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1865 copy_user_highpage(new_page, page, address, vma);
1867 SetPageDirty(new_page);
1868 __SetPageUptodate(new_page);
1869 __set_page_locked(new_page);
1875 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1876 unsigned long *vm_flags)
1878 struct stable_node *stable_node;
1879 struct rmap_item *rmap_item;
1880 struct hlist_node *hlist;
1881 unsigned int mapcount = page_mapcount(page);
1883 int search_new_forks = 0;
1885 VM_BUG_ON(!PageKsm(page));
1886 VM_BUG_ON(!PageLocked(page));
1888 stable_node = page_stable_node(page);
1892 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1893 struct anon_vma *anon_vma = rmap_item->anon_vma;
1894 struct anon_vma_chain *vmac;
1895 struct vm_area_struct *vma;
1897 anon_vma_lock_read(anon_vma);
1898 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1901 if (rmap_item->address < vma->vm_start ||
1902 rmap_item->address >= vma->vm_end)
1905 * Initially we examine only the vma which covers this
1906 * rmap_item; but later, if there is still work to do,
1907 * we examine covering vmas in other mms: in case they
1908 * were forked from the original since ksmd passed.
1910 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1913 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1916 referenced += page_referenced_one(page, vma,
1917 rmap_item->address, &mapcount, vm_flags);
1918 if (!search_new_forks || !mapcount)
1921 anon_vma_unlock_read(anon_vma);
1925 if (!search_new_forks++)
1931 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1933 struct stable_node *stable_node;
1934 struct hlist_node *hlist;
1935 struct rmap_item *rmap_item;
1936 int ret = SWAP_AGAIN;
1937 int search_new_forks = 0;
1939 VM_BUG_ON(!PageKsm(page));
1940 VM_BUG_ON(!PageLocked(page));
1942 stable_node = page_stable_node(page);
1946 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1947 struct anon_vma *anon_vma = rmap_item->anon_vma;
1948 struct anon_vma_chain *vmac;
1949 struct vm_area_struct *vma;
1951 anon_vma_lock_read(anon_vma);
1952 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1955 if (rmap_item->address < vma->vm_start ||
1956 rmap_item->address >= vma->vm_end)
1959 * Initially we examine only the vma which covers this
1960 * rmap_item; but later, if there is still work to do,
1961 * we examine covering vmas in other mms: in case they
1962 * were forked from the original since ksmd passed.
1964 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1967 ret = try_to_unmap_one(page, vma,
1968 rmap_item->address, flags);
1969 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1970 anon_vma_unlock_read(anon_vma);
1974 anon_vma_unlock_read(anon_vma);
1976 if (!search_new_forks++)
1982 #ifdef CONFIG_MIGRATION
1983 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1984 struct vm_area_struct *, unsigned long, void *), void *arg)
1986 struct stable_node *stable_node;
1987 struct hlist_node *hlist;
1988 struct rmap_item *rmap_item;
1989 int ret = SWAP_AGAIN;
1990 int search_new_forks = 0;
1992 VM_BUG_ON(!PageKsm(page));
1993 VM_BUG_ON(!PageLocked(page));
1995 stable_node = page_stable_node(page);
1999 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
2000 struct anon_vma *anon_vma = rmap_item->anon_vma;
2001 struct anon_vma_chain *vmac;
2002 struct vm_area_struct *vma;
2004 anon_vma_lock_read(anon_vma);
2005 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2008 if (rmap_item->address < vma->vm_start ||
2009 rmap_item->address >= vma->vm_end)
2012 * Initially we examine only the vma which covers this
2013 * rmap_item; but later, if there is still work to do,
2014 * we examine covering vmas in other mms: in case they
2015 * were forked from the original since ksmd passed.
2017 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2020 ret = rmap_one(page, vma, rmap_item->address, arg);
2021 if (ret != SWAP_AGAIN) {
2022 anon_vma_unlock_read(anon_vma);
2026 anon_vma_unlock_read(anon_vma);
2028 if (!search_new_forks++)
2034 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2036 struct stable_node *stable_node;
2038 VM_BUG_ON(!PageLocked(oldpage));
2039 VM_BUG_ON(!PageLocked(newpage));
2040 VM_BUG_ON(newpage->mapping != oldpage->mapping);
2042 stable_node = page_stable_node(newpage);
2044 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
2045 stable_node->kpfn = page_to_pfn(newpage);
2047 * newpage->mapping was set in advance; now we need smp_wmb()
2048 * to make sure that the new stable_node->kpfn is visible
2049 * to get_ksm_page() before it can see that oldpage->mapping
2050 * has gone stale (or that PageSwapCache has been cleared).
2053 set_page_stable_node(oldpage, NULL);
2056 #endif /* CONFIG_MIGRATION */
2058 #ifdef CONFIG_MEMORY_HOTREMOVE
2059 static void ksm_check_stable_tree(unsigned long start_pfn,
2060 unsigned long end_pfn)
2062 struct stable_node *stable_node;
2063 struct list_head *this, *next;
2064 struct rb_node *node;
2067 for (nid = 0; nid < nr_node_ids; nid++) {
2068 node = rb_first(&root_stable_tree[nid]);
2070 stable_node = rb_entry(node, struct stable_node, node);
2071 if (stable_node->kpfn >= start_pfn &&
2072 stable_node->kpfn < end_pfn) {
2074 * Don't get_ksm_page, page has already gone:
2075 * which is why we keep kpfn instead of page*
2077 remove_node_from_stable_tree(stable_node);
2078 node = rb_first(&root_stable_tree[nid]);
2080 node = rb_next(node);
2084 list_for_each_safe(this, next, &migrate_nodes) {
2085 stable_node = list_entry(this, struct stable_node, list);
2086 if (stable_node->kpfn >= start_pfn &&
2087 stable_node->kpfn < end_pfn)
2088 remove_node_from_stable_tree(stable_node);
2093 static int ksm_memory_callback(struct notifier_block *self,
2094 unsigned long action, void *arg)
2096 struct memory_notify *mn = arg;
2099 case MEM_GOING_OFFLINE:
2101 * Keep it very simple for now: just lock out ksmd and
2102 * MADV_UNMERGEABLE while any memory is going offline.
2103 * mutex_lock_nested() is necessary because lockdep was alarmed
2104 * that here we take ksm_thread_mutex inside notifier chain
2105 * mutex, and later take notifier chain mutex inside
2106 * ksm_thread_mutex to unlock it. But that's safe because both
2107 * are inside mem_hotplug_mutex.
2109 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
2114 * Most of the work is done by page migration; but there might
2115 * be a few stable_nodes left over, still pointing to struct
2116 * pages which have been offlined: prune those from the tree,
2117 * otherwise get_ksm_page() might later try to access a
2118 * non-existent struct page.
2120 ksm_check_stable_tree(mn->start_pfn,
2121 mn->start_pfn + mn->nr_pages);
2124 case MEM_CANCEL_OFFLINE:
2125 mutex_unlock(&ksm_thread_mutex);
2130 #endif /* CONFIG_MEMORY_HOTREMOVE */
2134 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2137 #define KSM_ATTR_RO(_name) \
2138 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2139 #define KSM_ATTR(_name) \
2140 static struct kobj_attribute _name##_attr = \
2141 __ATTR(_name, 0644, _name##_show, _name##_store)
2143 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2144 struct kobj_attribute *attr, char *buf)
2146 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2149 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2150 struct kobj_attribute *attr,
2151 const char *buf, size_t count)
2153 unsigned long msecs;
2156 err = strict_strtoul(buf, 10, &msecs);
2157 if (err || msecs > UINT_MAX)
2160 ksm_thread_sleep_millisecs = msecs;
2164 KSM_ATTR(sleep_millisecs);
2166 static ssize_t pages_to_scan_show(struct kobject *kobj,
2167 struct kobj_attribute *attr, char *buf)
2169 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2172 static ssize_t pages_to_scan_store(struct kobject *kobj,
2173 struct kobj_attribute *attr,
2174 const char *buf, size_t count)
2177 unsigned long nr_pages;
2179 err = strict_strtoul(buf, 10, &nr_pages);
2180 if (err || nr_pages > UINT_MAX)
2183 ksm_thread_pages_to_scan = nr_pages;
2187 KSM_ATTR(pages_to_scan);
2189 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2192 return sprintf(buf, "%u\n", ksm_run);
2195 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2196 const char *buf, size_t count)
2199 unsigned long flags;
2201 err = strict_strtoul(buf, 10, &flags);
2202 if (err || flags > UINT_MAX)
2204 if (flags > KSM_RUN_UNMERGE)
2208 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2209 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2210 * breaking COW to free the pages_shared (but leaves mm_slots
2211 * on the list for when ksmd may be set running again).
2214 mutex_lock(&ksm_thread_mutex);
2215 if (ksm_run != flags) {
2217 if (flags & KSM_RUN_UNMERGE) {
2218 set_current_oom_origin();
2219 err = unmerge_and_remove_all_rmap_items();
2220 clear_current_oom_origin();
2222 ksm_run = KSM_RUN_STOP;
2227 mutex_unlock(&ksm_thread_mutex);
2229 if (flags & KSM_RUN_MERGE)
2230 wake_up_interruptible(&ksm_thread_wait);
2237 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2238 struct kobj_attribute *attr, char *buf)
2240 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2243 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2244 struct kobj_attribute *attr,
2245 const char *buf, size_t count)
2250 err = kstrtoul(buf, 10, &knob);
2256 mutex_lock(&ksm_thread_mutex);
2257 if (ksm_merge_across_nodes != knob) {
2258 if (ksm_pages_shared || remove_all_stable_nodes())
2261 ksm_merge_across_nodes = knob;
2263 mutex_unlock(&ksm_thread_mutex);
2265 return err ? err : count;
2267 KSM_ATTR(merge_across_nodes);
2270 static ssize_t pages_shared_show(struct kobject *kobj,
2271 struct kobj_attribute *attr, char *buf)
2273 return sprintf(buf, "%lu\n", ksm_pages_shared);
2275 KSM_ATTR_RO(pages_shared);
2277 static ssize_t pages_sharing_show(struct kobject *kobj,
2278 struct kobj_attribute *attr, char *buf)
2280 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2282 KSM_ATTR_RO(pages_sharing);
2284 static ssize_t pages_unshared_show(struct kobject *kobj,
2285 struct kobj_attribute *attr, char *buf)
2287 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2289 KSM_ATTR_RO(pages_unshared);
2291 static ssize_t pages_volatile_show(struct kobject *kobj,
2292 struct kobj_attribute *attr, char *buf)
2294 long ksm_pages_volatile;
2296 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2297 - ksm_pages_sharing - ksm_pages_unshared;
2299 * It was not worth any locking to calculate that statistic,
2300 * but it might therefore sometimes be negative: conceal that.
2302 if (ksm_pages_volatile < 0)
2303 ksm_pages_volatile = 0;
2304 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2306 KSM_ATTR_RO(pages_volatile);
2308 static ssize_t full_scans_show(struct kobject *kobj,
2309 struct kobj_attribute *attr, char *buf)
2311 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2313 KSM_ATTR_RO(full_scans);
2315 static struct attribute *ksm_attrs[] = {
2316 &sleep_millisecs_attr.attr,
2317 &pages_to_scan_attr.attr,
2319 &pages_shared_attr.attr,
2320 &pages_sharing_attr.attr,
2321 &pages_unshared_attr.attr,
2322 &pages_volatile_attr.attr,
2323 &full_scans_attr.attr,
2325 &merge_across_nodes_attr.attr,
2330 static struct attribute_group ksm_attr_group = {
2334 #endif /* CONFIG_SYSFS */
2336 static int __init ksm_init(void)
2338 struct task_struct *ksm_thread;
2342 err = ksm_slab_init();
2346 for (nid = 0; nid < nr_node_ids; nid++)
2347 root_stable_tree[nid] = RB_ROOT;
2349 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2350 if (IS_ERR(ksm_thread)) {
2351 printk(KERN_ERR "ksm: creating kthread failed\n");
2352 err = PTR_ERR(ksm_thread);
2357 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2359 printk(KERN_ERR "ksm: register sysfs failed\n");
2360 kthread_stop(ksm_thread);
2364 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2366 #endif /* CONFIG_SYSFS */
2368 #ifdef CONFIG_MEMORY_HOTREMOVE
2370 * Choose a high priority since the callback takes ksm_thread_mutex:
2371 * later callbacks could only be taking locks which nest within that.
2373 hotplug_memory_notifier(ksm_memory_callback, 100);
2382 module_init(ksm_init)